chore: import upstream snapshot with attribution

python/sglang/srt/layers/quantization/__init__.py

@@ -0,0 +1,170 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Adapted from https://raw.githubusercontent.com/vllm-project/vllm/v0.5.5/vllm/model_executor/layers/quantization/__init__.py
+from __future__ import annotations
+
+import builtins
+import inspect
+from typing import TYPE_CHECKING, Dict, Optional, Type
+
+import torch
+
+
+# Define empty classes as placeholders when vllm is not available
+class DummyConfig:
+    def override_quantization_method(self, *args, **kwargs):
+        return None
+
+
+CompressedTensorsConfig = DummyConfig
+
+from sglang.srt.layers.quantization.auto_round import AutoRoundConfig
+from sglang.srt.layers.quantization.awq import AWQConfig, AWQCPUConfig, AWQMarlinConfig
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.quantization.bitsandbytes import BitsAndBytesConfig
+from sglang.srt.layers.quantization.blockwise_int8 import BlockInt8Config
+from sglang.srt.layers.quantization.compressed_tensors.compressed_tensors import (
+    CompressedTensorsConfig,
+)
+from sglang.srt.layers.quantization.fp8 import Fp8Config
+from sglang.srt.layers.quantization.fpgemm_fp8 import FBGEMMFp8Config
+from sglang.srt.layers.quantization.gguf import GGUFConfig
+from sglang.srt.layers.quantization.gptq import (
+    CPUGPTQConfig,
+    GPTQAscendConfig,
+    GPTQConfig,
+    GPTQMarlinConfig,
+)
+from sglang.srt.layers.quantization.mlx import MlxQuantizationConfig
+from sglang.srt.layers.quantization.modelopt_quant import (
+    ModelOptFp4Config,
+    ModelOptFp8Config,
+    ModelOptMixedPrecisionConfig,
+)
+from sglang.srt.layers.quantization.modelslim.modelslim import ModelSlimConfig
+from sglang.srt.layers.quantization.moe_wna16 import MoeWNA16Config
+from sglang.srt.layers.quantization.mxfp4 import Mxfp4Config
+from sglang.srt.layers.quantization.npu_mxfp4 import Mxfp4W4A8Config
+from sglang.srt.layers.quantization.nvfp4_online import NvFp4OnlineConfig
+from sglang.srt.layers.quantization.petit import PetitNvFp4Config
+from sglang.srt.layers.quantization.qoq import QoQConfig
+from sglang.srt.layers.quantization.quark.quark import QuarkConfig
+from sglang.srt.layers.quantization.quark_int4fp8_moe import QuarkInt4Fp8Config
+from sglang.srt.layers.quantization.w4afp8 import W4AFp8Config
+from sglang.srt.layers.quantization.w8a8_fp8 import W8A8Fp8Config
+from sglang.srt.layers.quantization.w8a8_int8 import W8A8Int8Config
+from sglang.srt.platforms import current_platform
+from sglang.srt.utils import (
+    cpu_has_amx_support,
+    is_cpu,
+    is_cuda,
+    is_hip,
+    is_mps,
+    is_npu,
+    mxfp_supported,
+)
+
+_is_mxfp_supported = mxfp_supported()
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.topk import TopKOutput
+
+# Base quantization methods
+BASE_QUANTIZATION_METHODS: Dict[str, Type[QuantizationConfig]] = {
+    "fp8": Fp8Config,
+    "mxfp8": Fp8Config,
+    "blockwise_int8": BlockInt8Config,
+    "modelopt": ModelOptFp8Config,  # Auto-detect, defaults to FP8
+    "modelopt_fp8": ModelOptFp8Config,
+    "modelopt_fp4": ModelOptFp4Config,
+    "nvfp4_online": NvFp4OnlineConfig,
+    "modelopt_mixed": ModelOptMixedPrecisionConfig,
+    "w8a8_int8": W8A8Int8Config,
+    "w8a8_fp8": W8A8Fp8Config,
+    "awq": AWQConfig,
+    "awq_marlin": AWQMarlinConfig,
+    "bitsandbytes": BitsAndBytesConfig,
+    "gguf": GGUFConfig,
+    "gptq": GPTQConfig,
+    "gptq_marlin": GPTQMarlinConfig,
+    "moe_wna16": MoeWNA16Config,
+    "compressed-tensors": CompressedTensorsConfig,
+    "qoq": QoQConfig,
+    "w4afp8": W4AFp8Config,
+    "petit_nvfp4": PetitNvFp4Config,
+    "fbgemm_fp8": FBGEMMFp8Config,
+    "quark": QuarkConfig,
+    "quark_mxfp4": QuarkConfig,
+    "auto-round": AutoRoundConfig,
+    "auto-round-int8": W8A8Int8Config,
+    "modelslim": ModelSlimConfig,
+    "quark_int4fp8_moe": QuarkInt4Fp8Config,
+    "mxfp_w4a8": Mxfp4W4A8Config,
+}
+
+
+if is_cpu() or is_cuda() or (_is_mxfp_supported and is_hip()):
+    BASE_QUANTIZATION_METHODS.update(
+        {
+            "mxfp4": Mxfp4Config,
+        }
+    )
+
+
+if is_npu():
+    BASE_QUANTIZATION_METHODS.update(
+        {
+            "gptq": GPTQAscendConfig,
+        }
+    )
+
+
+if is_mps():
+    BASE_QUANTIZATION_METHODS.update(
+        {
+            "mlx_q4": MlxQuantizationConfig,
+            "mlx_q8": MlxQuantizationConfig,
+        }
+    )
+
+# subset of above quant methods, supported on CPU
+CPU_QUANTIZATION_METHODS = {
+    "fp8": Fp8Config,
+    "w8a8_int8": W8A8Int8Config,
+    "compressed-tensors": CompressedTensorsConfig,
+    "awq": AWQCPUConfig,
+    "gptq": CPUGPTQConfig,
+    "mxfp4": Mxfp4Config,
+}
+
+QUANTIZATION_METHODS = {**BASE_QUANTIZATION_METHODS}
+
+
+def get_quantization_config(quantization: str) -> Type[QuantizationConfig]:
+    if quantization not in QUANTIZATION_METHODS:
+        raise ValueError(
+            f"Invalid quantization method: {quantization}. "
+            f"Available methods: {list(QUANTIZATION_METHODS.keys())}"
+        )
+    from sglang.srt.utils import is_cpu
+
+    if is_cpu() and cpu_has_amx_support():
+        if quantization not in CPU_QUANTIZATION_METHODS:
+            raise ValueError(
+                f"Invalid quantization method on CPU: {quantization}. "
+                f"Available methods on CPU: {list(QUANTIZATION_METHODS.keys())}"
+            )
+        else:
+            return CPU_QUANTIZATION_METHODS[quantization]
+
+    if current_platform.is_out_of_tree():
+        config = current_platform.get_quantization_config(quantization)
+
+        # If the platform has a quantization config, use it else use the default
+        if config is not None:
+            return config
+
+    return QUANTIZATION_METHODS[quantization]
+
+
+original_isinstance = builtins.isinstance

python/sglang/srt/layers/quantization/auto_round.py

@@ -0,0 +1,427 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import logging
+import re
+from fractions import Fraction
+from typing import Any, Optional, Union
+
+import torch
+
+logger = logging.getLogger(__name__)
+
+from sglang.srt.layers.quantization.utils import get_scalar_types
+
+ScalarType, scalar_types = get_scalar_types()
+
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.utils import is_npu
+
+_is_npu = is_npu()
+
+
+class AutoRoundConfig(QuantizationConfig):
+    """Config class for AutoRound.
+    Reference: https://arxiv.org/pdf/2309.05516
+    """
+
+    SUPPORTED_BITS = {2, 3, 4, 8}
+    SUPPORTED_DTYPES = {"int"}
+    SUPPORTED_FORMATS = {"auto_round:auto_gptq", "auto_round:auto_awq"}
+    SUPPORTED_BACKENDS = {"auto", "gptq", "gptq:marlin", "awq", "awq:marlin", "marlin"}
+
+    def __init__(
+        self,
+        weight_bits: int,
+        group_size: int,
+        sym: bool = True,
+        packing_format: str = "auto_round:auto_gptq",
+        block_name_to_quantize: Optional[Union[str, list[str]]] = None,
+        extra_config: Optional[dict[str, Any]] = None,
+        data_type: str = "int",
+        backend: str = "auto",
+    ) -> None:
+        super().__init__()
+        if weight_bits not in self.SUPPORTED_BITS:
+            raise ValueError(
+                f"Unsupported weight_bits: {weight_bits}, "
+                f"currently only support  {self.SUPPORTED_BITS}"
+            )
+        if data_type not in self.SUPPORTED_DTYPES:
+            raise ValueError(
+                f"Unsupported data_type: {data_type},"
+                f" currently only support  {self.SUPPORTED_DTYPES}"
+            )
+        if packing_format not in self.SUPPORTED_FORMATS:
+            raise ValueError(
+                f"Unsupported packing_format: {packing_format}, "
+                f"currently only support  {self.SUPPORTED_FORMATS}"
+            )
+        if backend not in self.SUPPORTED_BACKENDS:
+            raise ValueError(
+                f"Unsupported backend: {backend},  "
+                f"currently only support  {self.SUPPORTED_BACKENDS}"
+            )
+
+        self.weight_bits = weight_bits
+        self.group_size = group_size
+        self.sym = sym
+        self.packing_format = packing_format
+        self.block_name_to_quantize = (
+            block_name_to_quantize.split(",")
+            if isinstance(block_name_to_quantize, str)
+            else block_name_to_quantize
+        )
+        self.extra_config = extra_config
+        self.data_type = data_type
+        self.backend = backend
+        self.pack_factor = Fraction(32, weight_bits)
+
+    def __repr__(self) -> str:
+        return (
+            f"AutoRoundConfig(weight_bits={self.weight_bits}, "
+            f"group_size={self.group_size}, sym={self.sym})"
+        )
+
+    @classmethod
+    def get_name(cls):
+        return "auto-round"
+
+    @classmethod
+    def get_supported_act_dtypes(cls) -> list[torch.dtype]:
+        return [torch.half, torch.bfloat16]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 60
+
+    @classmethod
+    def get_config_filenames(cls) -> list[str]:
+        return ["quantization_config.json"]
+
+    @classmethod
+    def from_config(cls, config: dict[str, Any]) -> "AutoRoundConfig":
+        return cls(
+            weight_bits=cls.get_from_keys(config, ["bits"]),
+            group_size=cls.get_from_keys(config, ["group_size"]),
+            sym=cls.get_from_keys(config, ["sym"]),
+            packing_format=cls.get_from_keys_or(
+                config,
+                ["packing_format"],
+                "auto_round:auto_gptq",
+            ),
+            block_name_to_quantize=cls.get_from_keys_or(
+                config, ["block_name_to_quantize", "to_quant_block_names"], None
+            ),
+            extra_config=cls.get_from_keys_or(config, ["extra_config"], None),
+            data_type=cls.get_from_keys_or(config, ["data_type"], "int"),
+            backend=cls.get_from_keys_or(
+                config, ["backend", "vllm_backend", "sglang_backend"], "auto"
+            ),
+        )
+
+    def get_scaled_act_names(self) -> list[str]:
+        """Returns the activation function names that should be post-scaled.
+
+        For now, this is only used by AWQ.
+        """
+        raise NotImplementedError
+
+    def get_layer_config(self, layer, layer_name: str):
+        from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
+
+        def get_config(name: str, quantized: bool = True):
+            if not self.extra_config:
+                return (
+                    self.weight_bits if quantized else 16,
+                    self.group_size if quantized else -1,
+                    self.sym if quantized else True,
+                )
+
+            # Exact match first
+            if name in self.extra_config:
+                cfg = self.extra_config[name]
+                return (
+                    cfg.get("bits", self.weight_bits if quantized else 16),
+                    cfg.get("group_size", self.group_size if quantized else -1),
+                    cfg.get("sym", self.sym if quantized else True),
+                )
+
+            REGEX_SPECIAL_CHARS = set(r"*+?^$()[]{}|\\")
+            for pattern, cfg in self.extra_config.items():
+                if not isinstance(pattern, str) or not any(
+                    c in REGEX_SPECIAL_CHARS for c in pattern
+                ):
+                    continue
+
+                try:
+                    if re.fullmatch(pattern, name):
+                        return (
+                            cfg.get("bits", self.weight_bits if quantized else 16),
+                            cfg.get("group_size", self.group_size if quantized else -1),
+                            cfg.get("sym", self.sym if quantized else True),
+                        )
+                except re.error:
+                    # Invalid regex, ignore.
+                    continue
+
+            return (
+                self.weight_bits if quantized else 16,
+                self.group_size if quantized else -1,
+                self.sym if quantized else True,
+            )
+
+        # 1. Exact match from config
+        if self.extra_config and layer_name in self.extra_config:
+            return get_config(layer_name)
+
+        # 2. Determine whether layer should be quantized
+        quantized = not isinstance(layer, ParallelLMHead)
+        if self.block_name_to_quantize:
+            quantized = any(
+                layer_name.startswith(name) for name in self.block_name_to_quantize
+            )
+
+        # 3. Handle fused MoE
+        if self.extra_config and "fusedmoe" in layer.__class__.__name__.lower():
+            moe_configs = [
+                get_config(name, quantized)
+                for name in self.extra_config
+                if name.startswith(layer_name)
+            ]
+            if moe_configs:
+                if len(set(moe_configs)) == 1:
+                    return moe_configs[0]
+                raise ValueError(
+                    f"Fused MoE layer '{layer_name}' requires "
+                    f"consistent quant config for all sub-layers"
+                )
+
+        # 4. Handle fused QKV or other patterns
+        if self.extra_config:
+            for fusion_key, sub_keys in self.packed_modules_mapping.items():
+                if fusion_key in layer_name and layer_name.count(fusion_key) == 1:
+                    sub_names = [
+                        layer_name.replace(fusion_key, sub_key) for sub_key in sub_keys
+                    ]
+                    sub_configs = [get_config(name, quantized) for name in sub_names]
+                    if len(set(sub_configs)) == 1:
+                        return sub_configs[0]
+                    raise ValueError(
+                        f"Fused module '{layer_name}' requires "
+                        f"consistent quant config for {sub_names}"
+                    )
+
+        # 5. Fallback or try a regular expression match
+        return get_config(layer_name, quantized)
+
+    def check_quantized(self, weight_bits: int) -> bool:
+        return weight_bits < 16
+
+    def apply_awq_quant_layer(self, layer, prefix: str, backend: str = "auto"):
+        from sglang.srt.layers.linear import LinearBase
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+        from sglang.srt.layers.quantization.marlin_utils import (
+            check_marlin_supported,
+            check_moe_marlin_supports_layer,
+        )
+        from sglang.srt.layers.quantization.unquant import UnquantizedLinearMethod
+        from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
+
+        weight_bits, group_size, sym = self.get_layer_config(layer, prefix)
+        if not self.check_quantized(weight_bits):
+            if isinstance(layer, (LinearBase, ParallelLMHead)):
+                return UnquantizedLinearMethod()
+            else:
+                return None
+        logger.debug(
+            "[%s] Type: %s, Bits: %s, Group Size: %s, Sym: %s",
+            prefix,
+            layer.__class__.__name__,
+            weight_bits,
+            group_size,
+            sym,
+        )
+        if backend == "auto" or "marlin" in backend:
+            AWQ_TYPE_MAP = {
+                4: scalar_types.uint4,
+                8: scalar_types.uint8,
+            }
+            use_marlin = (weight_bits in AWQ_TYPE_MAP) and check_marlin_supported(
+                AWQ_TYPE_MAP[weight_bits], group_size, not sym
+            )
+            if isinstance(layer, FusedMoE):
+                use_marlin = use_marlin and check_moe_marlin_supports_layer(
+                    layer, group_size
+                )
+
+        else:
+            use_marlin = False
+        if use_marlin:
+            from sglang.srt.layers.quantization.awq import (
+                AWQLinearMethod,
+                AWQMarlinConfig,
+                AWQMoEMethod,
+            )
+
+            quant_args_marlin = AWQMarlinConfig(
+                weight_bits=weight_bits,
+                group_size=group_size,
+                zero_point=not sym,
+                lm_head_quantized=False,
+                full_config={},
+                modules_to_not_convert=[],
+            )
+        else:
+            from sglang.srt.layers.quantization.awq import AWQConfig, AWQLinearMethod
+
+            quant_args = AWQConfig(
+                weight_bits=weight_bits,
+                group_size=group_size,
+                zero_point=not sym,
+            )
+
+        if isinstance(layer, FusedMoE):
+            if use_marlin:
+                layer.scheme = quant_args_marlin.get_moe_scheme(layer)
+                return AWQMoEMethod(quant_args_marlin)
+            from sglang.srt.layers.quantization.moe_wna16 import MoeWNA16Config
+
+            config = {
+                "quant_method": "awq",
+                "bits": weight_bits,
+                "group_size": group_size,
+                "zero_point": not sym,
+                "lm_head": False,
+            }
+            return MoeWNA16Config.from_config(config).get_quant_method(layer, prefix)
+
+        if isinstance(layer, (LinearBase, ParallelLMHead)):
+            if use_marlin:
+                layer.scheme = quant_args_marlin.get_linear_scheme(layer)
+                return AWQLinearMethod(quant_args_marlin)
+            else:
+                layer.scheme = quant_args.get_linear_scheme(layer)
+                return AWQLinearMethod(quant_args)
+        return None
+
+    def apply_gptq_quant_layer(self, layer, prefix: str, backend: str = "auto"):
+        from sglang.srt.layers.linear import LinearBase
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+        from sglang.srt.layers.quantization.gptq import (
+            GPTQAscendConfig,
+            GPTQLinearMethod,
+            GPTQMoEMethod,
+        )
+        from sglang.srt.layers.quantization.marlin_utils import (
+            check_marlin_supported,
+            check_moe_marlin_supports_layer,
+        )
+        from sglang.srt.layers.quantization.unquant import UnquantizedLinearMethod
+        from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
+
+        weight_bits, group_size, sym = self.get_layer_config(layer, prefix)
+        if not self.check_quantized(weight_bits):
+            if isinstance(layer, (LinearBase, ParallelLMHead)):
+                return UnquantizedLinearMethod()
+            else:
+                return None
+
+        logger.debug(
+            "[%s] Type: %s, Bits: %s, Group Size: %s, Sym: %s",
+            prefix,
+            layer.__class__.__name__,
+            weight_bits,
+            group_size,
+            sym,
+        )
+        if _is_npu:
+            quant_args = GPTQAscendConfig(
+                weight_bits=weight_bits,
+                group_size=group_size,
+                lm_head_quantized=False,
+                desc_act=False,
+                dynamic={},
+            )
+            quant_args.sym = sym
+
+            if isinstance(layer, FusedMoE):
+                layer.scheme = quant_args.get_moe_scheme(layer)
+                return GPTQMoEMethod(quant_args)
+
+            if isinstance(layer, (LinearBase, ParallelLMHead)):
+                layer.scheme = quant_args.get_linear_scheme(layer)
+                return GPTQLinearMethod(quant_args)
+
+            return None
+
+        if backend == "auto" or "marlin" in backend:
+            GPTQ_TYPE_MAP = {
+                (4, True): scalar_types.uint4b8,
+                (8, True): scalar_types.uint8b128,
+            }
+            use_marlin = (weight_bits, sym) in GPTQ_TYPE_MAP and check_marlin_supported(
+                GPTQ_TYPE_MAP[(weight_bits, sym)], group_size, has_zp=not sym
+            )
+            if isinstance(layer, FusedMoE):
+                use_marlin = use_marlin and check_moe_marlin_supports_layer(
+                    layer, group_size
+                )
+        else:
+            use_marlin = False
+        if use_marlin:
+            from sglang.srt.layers.quantization.gptq import (
+                GPTQMarlinConfig,
+                GPTQMarlinLinearMethod,
+                GPTQMarlinMoEMethod,
+            )
+
+            quant_args_marlin = GPTQMarlinConfig(
+                weight_bits=weight_bits,
+                group_size=group_size,
+                is_sym=sym,
+                lm_head_quantized=False,
+                desc_act=False,
+                dynamic={},
+                full_config={},
+            )
+        else:
+            from sglang.srt.layers.quantization.gptq import GPTQConfig, GPTQLinearMethod
+
+            quant_args = GPTQConfig(
+                weight_bits=weight_bits,
+                group_size=group_size,
+                lm_head_quantized=False,
+                desc_act=False,
+                dynamic={},
+            )
+
+        if isinstance(layer, FusedMoE):
+            if use_marlin:
+                from sglang.srt.layers.quantization.moe_wna16 import MoeWNA16Config
+
+                config = {
+                    "quant_method": "gptq",
+                    "bits": weight_bits,
+                    "group_size": group_size,
+                    "sym": sym,
+                    "lm_head": False,
+                }
+                return MoeWNA16Config.from_config(config).get_quant_method(
+                    layer, prefix
+                )
+            return GPTQMarlinMoEMethod(quant_args_marlin)
+
+        if isinstance(layer, (LinearBase, ParallelLMHead)):
+            if use_marlin:
+                return GPTQMarlinLinearMethod(quant_args_marlin)
+            else:
+                return GPTQLinearMethod(quant_args)
+
+        return None
+
+    def get_quant_method(self, layer: torch.nn.Module, prefix: str):
+        # TODO enable CPU quant method later
+        if "gptq" in self.packing_format or "gptq" in self.backend:
+            return self.apply_gptq_quant_layer(layer, prefix)
+        if "awq" in self.packing_format or "awq" in self.backend:
+            return self.apply_awq_quant_layer(layer, prefix)

python/sglang/srt/layers/quantization/awq/__init__.py

@@ -0,0 +1,32 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from .awq import (
+    AWQConfig,
+    AWQCPUConfig,
+    AWQLinearMethod,
+    AWQMarlinConfig,
+    AWQMoEMethod,
+)
+from .awq_triton import awq_dequantize_decomposition, awq_dequantize_triton
+from .schemes import (
+    AWQAscendLinearScheme,
+    AWQAscendMoEScheme,
+    AWQLinearScheme,
+    AWQMarlinLinearScheme,
+    AWQMoEScheme,
+)
+
+__all__ = [
+    "AWQConfig",
+    "AWQCPUConfig",
+    "AWQMarlinConfig",
+    "AWQLinearMethod",
+    "AWQMoEMethod",
+    "AWQLinearScheme",
+    "AWQMarlinLinearScheme",
+    "AWQAscendLinearScheme",
+    "AWQMoEScheme",
+    "AWQAscendMoEScheme",
+    "awq_dequantize_triton",
+    "awq_dequantize_decomposition",
+]

python/sglang/srt/layers/quantization/awq/awq.py

@@ -0,0 +1,488 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+import logging
+import warnings
+from typing import TYPE_CHECKING, Any, Dict, List, Optional
+
+import torch
+
+from sglang.srt.layers.linear import LinearBase
+from sglang.srt.layers.moe import MoeRunnerConfig
+from sglang.srt.layers.quantization.base_config import (
+    FusedMoEMethodBase,
+    LinearMethodBase,
+    QuantizationConfig,
+    QuantizeMethodBase,
+)
+from sglang.srt.layers.quantization.marlin_utils import (
+    check_marlin_supported,
+    check_marlin_supports_layer,
+    check_moe_marlin_supports_layer,
+    verify_marlin_supported,
+)
+from sglang.srt.layers.quantization.unquant import UnquantizedLinearMethod
+from sglang.srt.layers.quantization.utils import get_scalar_types
+from sglang.srt.utils.patch_torch import register_fake_if_exists
+
+from .schemes import (
+    AWQAscendLinearScheme,
+    AWQAscendMoEScheme,
+    AWQIntelAMXLinearScheme,
+    AWQIntelAMXMoEScheme,
+    AWQLinearScheme,
+    AWQMarlinLinearScheme,
+    AWQMoEScheme,
+)
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
+
+from sglang.srt.utils import is_cuda, is_hip, is_npu, is_xpu
+
+_is_cuda = is_cuda()
+_is_hip = is_hip()
+_is_xpu = is_xpu()
+_is_npu = is_npu()
+
+if not (_is_cuda or _is_hip or _is_xpu or _is_npu):
+    warnings.warn(f"Only CUDA, HIP and XPU support AWQ currently.")
+
+logger = logging.getLogger(__name__)
+
+
+ScalarType, scalar_types = get_scalar_types()
+
+
+def is_layer_skipped_awq(prefix: str, modules_to_not_convert: List[str]):
+    return any(module_name in prefix for module_name in modules_to_not_convert)
+
+
+class AWQConfig(QuantizationConfig):
+    """Config class for AWQ.
+
+    Reference: https://arxiv.org/abs/2306.00978
+    """
+
+    def __init__(
+        self,
+        weight_bits: int,
+        group_size: int,
+        zero_point: bool,
+        modules_to_not_convert: Optional[List[str]] = None,
+    ) -> None:
+        super().__init__()
+        self.weight_bits = weight_bits
+        self.group_size = group_size
+        self.zero_point = zero_point
+        self.modules_to_not_convert = modules_to_not_convert or []
+
+        if self.weight_bits != 4:
+            raise ValueError(
+                "Currently, only 4-bit weight quantization is supported for "
+                f"AWQ, but got {self.weight_bits} bits."
+            )
+        self.pack_factor = 32 // self.weight_bits
+
+    def __repr__(self) -> str:
+        return (
+            f"AWQConfig(weight_bits={self.weight_bits}, "
+            f"group_size={self.group_size}, "
+            f"zero_point={self.zero_point}, "
+            f"modules_to_not_convert={self.modules_to_not_convert})"
+        )
+
+    def get_scaled_act_names(self) -> List[str]:
+        return []
+
+    def get_name(self) -> str:
+        return "awq"
+
+    def get_supported_act_dtypes(self) -> List[torch.dtype]:
+        return [torch.float16] if not _is_npu else [torch.float16, torch.bfloat16]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # The AWQ kernel only supports Turing or newer GPUs.
+        if _is_npu:
+            raise NotImplementedError(
+                'NPU hardware does not support "get_min_capability" feature.'
+            )
+        else:
+            return 75
+
+    @staticmethod
+    def get_config_filenames() -> List[str]:
+        return [
+            "quant_config.json",  # E.g., casperhansen/vicuna-7b-v1.5-awq
+            # E.g., abhinavkulkarni/mosaicml-mpt-7b-instruct-w4-g128-awq
+            "quantize_config.json",
+        ]
+
+    @classmethod
+    def from_config(cls, config: Dict[str, Any]) -> AWQConfig:
+        weight_bits = cls.get_from_keys(config, ["w_bit", "bits"])
+        group_size = cls.get_from_keys(config, ["q_group_size", "group_size"])
+        zero_point = cls.get_from_keys(config, ["zero_point"])
+        modules_to_not_convert = cls.get_from_keys_or(
+            config, ["modules_to_not_convert"], None
+        )
+        return cls(weight_bits, group_size, zero_point, modules_to_not_convert)
+
+    def get_quant_method(
+        self, layer: torch.nn.Module, prefix: str
+    ) -> Optional[LinearMethodBase]:
+        from sglang.srt.layers.linear import LinearBase
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+
+        if _is_npu:
+            if isinstance(layer, LinearBase):
+                if is_layer_skipped_awq(prefix, self.modules_to_not_convert):
+                    return UnquantizedLinearMethod()
+                layer.scheme = self.get_linear_scheme(layer)
+                return AWQLinearMethod(self)
+            elif isinstance(layer, FusedMoE):
+                if is_layer_skipped_awq(prefix, self.modules_to_not_convert):
+                    return None
+                layer.scheme = self.get_moe_scheme(layer)
+                return AWQMoEMethod(self)
+            return None
+
+        if isinstance(layer, LinearBase):
+            if is_layer_skipped_awq(prefix, self.modules_to_not_convert):
+                return UnquantizedLinearMethod()
+            layer.scheme = self.get_linear_scheme(layer)
+            return AWQLinearMethod(self)
+        return None
+
+    def get_linear_scheme(self, layer: torch.nn.Module):
+        assert isinstance(layer, LinearBase)
+        # TODO: move platform-specific AWQ scheme selection into the platform
+        # plugin factory once quantization hooks are available there.
+        if _is_npu:
+            return AWQAscendLinearScheme(self)
+        return AWQLinearScheme(self)
+
+    def get_moe_scheme(self, layer: torch.nn.Module):
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+
+        assert isinstance(layer, FusedMoE)
+        # This is currently only reached by the NPU path in get_quant_method.
+        if _is_npu:
+            return AWQAscendMoEScheme(self)
+        raise NotImplementedError("AWQConfig only supports MoE scheme on NPU.")
+
+
+class AWQCPUConfig(AWQConfig):
+    """CPU Config class for AWQ, inherit from AWQConfig"""
+
+    def get_supported_act_dtypes(self) -> List[torch.dtype]:
+        return [torch.float16, torch.bfloat16]
+
+    def get_quant_method(
+        self, layer: torch.nn.Module, prefix: str
+    ) -> Optional[LinearMethodBase]:
+        from sglang.srt.layers.linear import LinearBase
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+
+        if isinstance(layer, LinearBase):
+            if is_layer_skipped_awq(prefix, self.modules_to_not_convert):
+                return UnquantizedLinearMethod()
+            layer.scheme = self.get_linear_scheme(layer)
+            return AWQLinearMethod(self)
+        elif isinstance(layer, FusedMoE):
+            layer.scheme = self.get_moe_scheme(layer)
+            return AWQMoEMethod(self)
+        return None
+
+    def get_linear_scheme(self, layer: torch.nn.Module):
+        from sglang.srt.layers.linear import LinearBase
+
+        assert isinstance(layer, LinearBase)
+        return AWQIntelAMXLinearScheme(self)
+
+    def get_moe_scheme(self, layer: torch.nn.Module):
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+
+        assert isinstance(layer, FusedMoE)
+        return AWQIntelAMXMoEScheme(self)
+
+
+class AWQMarlinConfig(QuantizationConfig):
+    """Config class for AWQ Marlin"""
+
+    # num_bits -> type
+    TYPE_MAP = {
+        4: scalar_types.uint4,
+        8: scalar_types.uint8,
+    }
+
+    def __init__(
+        self,
+        weight_bits: int,
+        group_size: int,
+        zero_point: bool,
+        lm_head_quantized: bool,
+        modules_to_not_convert: Optional[list[str]],
+        full_config: dict[str, Any],
+    ) -> None:
+        super().__init__()
+        if _is_hip:
+            warnings.warn(f"HIP does not support fused_marlin_moe currently.")
+        self.pack_factor = 32 // weight_bits  # packed into int32
+        self.group_size = group_size
+        self.zero_point = zero_point
+        self.lm_head_quantized = lm_head_quantized
+        self.weight_bits = weight_bits
+        self.modules_to_not_convert = modules_to_not_convert or []
+        self.full_config = full_config
+
+        if self.weight_bits not in self.TYPE_MAP:
+            raise ValueError(
+                f"Unsupported num_bits = {self.weight_bits}. "
+                f"Supported num_bits = {self.TYPE_MAP.keys()}"
+            )
+
+        self.quant_type = self.TYPE_MAP[self.weight_bits]
+
+        verify_marlin_supported(
+            self.quant_type, group_size=self.group_size, has_zp=self.zero_point
+        )
+
+    def __repr__(self) -> str:
+        return (
+            f"AWQMarlinConfig(quant_type={self.quant_type}, "
+            f"group_size={self.group_size}, "
+            f"zero_point={self.zero_point}, "
+            f"lm_head_quantized={self.lm_head_quantized}, "
+            f"modules_to_not_convert={self.modules_to_not_convert})"
+        )
+
+    def get_scaled_act_names(self) -> List[str]:
+        return []
+
+    @classmethod
+    def get_name(cls) -> str:
+        return "awq_marlin"
+
+    @classmethod
+    def get_supported_act_dtypes(cls) -> list[torch.dtype]:
+        return [torch.half, torch.bfloat16]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 80
+
+    @classmethod
+    def get_config_filenames(cls) -> list[str]:
+        return ["quantize_config.json"]
+
+    @classmethod
+    def from_config(cls, config: dict[str, Any]) -> AWQMarlinConfig:
+        weight_bits = cls.get_from_keys(config, ["bits"])
+        group_size = cls.get_from_keys(config, ["group_size"])
+        zero_point = cls.get_from_keys(config, ["zero_point"])
+        lm_head_quantized = cls.get_from_keys_or(config, ["lm_head"], default=False)
+        modules_to_not_convert = cls.get_from_keys_or(
+            config, ["modules_to_not_convert"], None
+        )
+        return cls(
+            weight_bits,
+            group_size,
+            zero_point,
+            lm_head_quantized,
+            modules_to_not_convert,
+            config,
+        )
+
+    @classmethod
+    def override_quantization_method(cls, hf_quant_cfg, user_quant) -> Optional[str]:
+        can_convert = cls.is_awq_marlin_compatible(hf_quant_cfg)
+        is_valid_user_quant = (
+            user_quant is None or user_quant == "marlin" or user_quant == "awq_marlin"
+        )
+
+        if can_convert and is_valid_user_quant:
+            msg = (
+                "The model is convertible to {} during runtime."
+                " Using {} kernel.".format(cls.get_name(), cls.get_name())
+            )
+            logger.info(msg)
+            return cls.get_name()
+
+        if can_convert and user_quant == "awq":
+            logger.info(
+                "Detected that the model can run with awq_marlin"
+                ", however you specified quantization=awq explicitly,"
+                " so forcing awq. Use quantization=awq_marlin for"
+                " faster inference"
+            )
+        return None
+
+    def get_quant_method(
+        self, layer: torch.nn.Module, prefix: str
+    ) -> Optional[QuantizeMethodBase]:
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+        from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
+
+        if isinstance(layer, LinearBase) or (
+            isinstance(layer, ParallelLMHead) and self.lm_head_quantized
+        ):
+            if is_layer_skipped_awq(prefix, self.modules_to_not_convert):
+                return UnquantizedLinearMethod()
+            # Check if the layer is supported by AWQMarlin.
+            if not check_marlin_supports_layer(layer, self.group_size):
+                logger.warning_once(
+                    "Layer '%s' is not supported by AWQMarlin. Falling back to unoptimized AWQ kernels.",  # noqa: E501
+                    prefix,
+                )
+                return AWQConfig.from_config(self.full_config).get_quant_method(
+                    layer, prefix
+                )
+            layer.scheme = self.get_linear_scheme(layer)
+            return AWQLinearMethod(self)
+        elif isinstance(layer, FusedMoE):
+            if is_layer_skipped_awq(prefix, self.modules_to_not_convert):
+                return None
+            from sglang.srt.layers.quantization.moe_wna16 import MoeWNA16Config
+
+            if not check_moe_marlin_supports_layer(layer, self.group_size):
+                logger.warning_once(
+                    f"Layer '{prefix}' is not supported by AWQMoeMarlin. "
+                    "Falling back to Moe WNA16 kernels."
+                )
+                return MoeWNA16Config.from_config(self.full_config).get_quant_method(
+                    layer, prefix
+                )
+            layer.scheme = self.get_moe_scheme(layer)
+            return AWQMoEMethod(self)
+        return None
+
+    def get_linear_scheme(self, layer: torch.nn.Module):
+        return AWQMarlinLinearScheme(self)
+
+    def get_moe_scheme(self, layer: torch.nn.Module):
+        return AWQMoEScheme(self)
+
+    @classmethod
+    def is_awq_marlin_compatible(cls, quant_config: dict[str, Any]):
+        # Extract data from quant config.
+        quant_method = quant_config.get("quant_method", "").lower()
+        num_bits = quant_config.get("bits")
+        group_size = quant_config.get("group_size")
+        zero_point = quant_config.get("zero_point")
+
+        if not _is_cuda:
+            return False
+
+        if quant_method != "awq":
+            return False
+
+        # If we cannot find the info needed in the config, cannot convert.
+        if num_bits is None or group_size is None or zero_point is None:
+            return False
+
+        if num_bits not in cls.TYPE_MAP:
+            return False
+
+        return check_marlin_supported(
+            quant_type=cls.TYPE_MAP[num_bits], group_size=group_size, has_zp=zero_point
+        )
+
+
+class AWQLinearMethod(LinearMethodBase):
+    """Linear method for AWQ.
+
+    Args:
+        quant_config: The AWQ quantization config.
+    """
+
+    def __init__(self, quant_config: AWQConfig):
+        self.quant_config = quant_config
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        input_size_per_partition: int,
+        output_partition_sizes: List[int],
+        input_size: int,
+        output_size: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        weight_loader = extra_weight_attrs.get("weight_loader")
+        layer.scheme.create_weights(
+            layer=layer,
+            input_size_per_partition=input_size_per_partition,
+            output_partition_sizes=output_partition_sizes,
+            input_size=input_size,
+            output_size=output_size,
+            params_dtype=params_dtype,
+            weight_loader=weight_loader,
+        )
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        layer.scheme.process_weights_after_loading(layer)
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        return layer.scheme.apply_weights(layer, x, bias)
+
+
+class AWQMoEMethod(FusedMoEMethodBase):
+
+    def __init__(self, quant_config: AWQMarlinConfig):
+        self.quant_config = quant_config
+        self.quant_type = scalar_types.uint4
+        if self.quant_config.weight_bits != 4:
+            raise ValueError("AWQMoEMethod only supports 4bit now.")
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        layer.scheme.create_weights(
+            layer=layer,
+            num_experts=num_experts,
+            hidden_size=hidden_size,
+            intermediate_size_per_partition=intermediate_size_per_partition,
+            params_dtype=params_dtype,
+            **extra_weight_attrs,
+        )
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        layer.scheme.process_weights_after_loading(layer)
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        layer.scheme.create_moe_runner(layer, moe_runner_config)
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+        return layer.scheme.apply_weights(layer, dispatch_output)
+
+
+# Register fake implementations for torch.compile support
+if _is_cuda:
+
+    @register_fake_if_exists("sgl_kernel::awq_marlin_repack")
+    def _(b_q_weight, size_k, size_n, num_bits):
+        return b_q_weight.new_empty(
+            (size_k // 16, size_n * (num_bits // 2)), dtype=b_q_weight.dtype
+        )

python/sglang/srt/layers/quantization/awq/awq_triton.py

@@ -0,0 +1,368 @@
+# Adapted from https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/quantization/awq_triton.py
+
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import torch
+import triton
+import triton.language as tl
+
+AWQ_TRITON_SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128]
+
+
+@triton.jit
+def awq_dequantize_kernel(
+    qweight_ptr,  # quantized matrix
+    scales_ptr,  # scales, per group
+    zeros_ptr,  # zeros, per group
+    group_size,  # Should always be one of the supported group sizes
+    result_ptr,  # Output matrix
+    num_cols,  # input num cols in qweight
+    num_rows,  # input num rows in qweight
+    BLOCK_SIZE_X: tl.constexpr,
+    BLOCK_SIZE_Y: tl.constexpr,
+):
+    # Setup the pids.
+    pid_x = tl.program_id(axis=0)
+    pid_y = tl.program_id(axis=1)
+
+    # Compute offsets and masks for qweight_ptr.
+    offsets_y = pid_y * BLOCK_SIZE_Y + tl.arange(0, BLOCK_SIZE_Y)
+    offsets_x = pid_x * BLOCK_SIZE_X + tl.arange(0, BLOCK_SIZE_X)
+    offsets = num_cols * offsets_y[:, None] + offsets_x[None, :]
+
+    masks_y = offsets_y < num_rows
+    masks_x = offsets_x < num_cols
+
+    masks = masks_y[:, None] & masks_x[None, :]
+
+    # Compute offsets and masks for result output ptr.
+    result_offsets_y = pid_y * BLOCK_SIZE_Y + tl.arange(0, BLOCK_SIZE_Y)
+    result_offsets_x = pid_x * BLOCK_SIZE_X * 8 + tl.arange(0, BLOCK_SIZE_X * 8)
+    result_offsets = (
+        8 * num_cols * result_offsets_y[:, None] + result_offsets_x[None, :]
+    )
+
+    result_masks_y = result_offsets_y < num_rows
+    result_masks_x = result_offsets_x < num_cols * 8
+    result_masks = result_masks_y[:, None] & result_masks_x[None, :]
+
+    # Load the weights.
+    iweights = tl.load(qweight_ptr + offsets, masks, 0.0)
+    iweights = tl.interleave(iweights, iweights)
+    iweights = tl.interleave(iweights, iweights)
+    iweights = tl.interleave(iweights, iweights)
+
+    # Create reverse AWQ order as tensor: [0, 4, 1, 5, 2, 6, 3, 7]
+    # that will map given indices to the correct order.
+    reverse_awq_order_tensor = (
+        (tl.arange(0, 2) * 4)[None, :] + tl.arange(0, 4)[:, None]
+    ).reshape(8)
+
+    # Use this to compute a set of shifts that can be used to unpack and
+    # reorder the values in iweights and zeros.
+    shifts = reverse_awq_order_tensor * 4
+    shifts = tl.broadcast_to(shifts[None, :], (BLOCK_SIZE_Y * BLOCK_SIZE_X, 8))
+    shifts = tl.reshape(shifts, (BLOCK_SIZE_Y, BLOCK_SIZE_X * 8))
+
+    # Unpack and reorder: shift out the correct 4-bit value and mask.
+    iweights = (iweights >> shifts) & 0xF
+
+    # Compute zero offsets and masks.
+    zero_offsets_y = pid_y * BLOCK_SIZE_Y // group_size + tl.arange(0, 1)
+    zero_offsets_x = pid_x * BLOCK_SIZE_X + tl.arange(0, BLOCK_SIZE_X)
+    zero_offsets = num_cols * zero_offsets_y[:, None] + zero_offsets_x[None, :]
+
+    zero_masks_y = zero_offsets_y < num_rows // group_size
+    zero_masks_x = zero_offsets_x < num_cols
+    zero_masks = zero_masks_y[:, None] & zero_masks_x[None, :]
+
+    # Load the zeros.
+    zeros = tl.load(zeros_ptr + zero_offsets, zero_masks, 0.0)
+    zeros = tl.interleave(zeros, zeros)
+    zeros = tl.interleave(zeros, zeros)
+    zeros = tl.interleave(zeros, zeros)
+    zeros = tl.broadcast_to(zeros, (BLOCK_SIZE_Y, BLOCK_SIZE_X * 8))
+
+    # Unpack and reorder: shift out the correct 4-bit value and mask.
+    zeros = (zeros >> shifts) & 0xF
+
+    # Compute scale offsets and masks.
+    scale_offsets_y = pid_y * BLOCK_SIZE_Y // group_size + tl.arange(0, 1)
+    scale_offsets_x = pid_x * BLOCK_SIZE_X * 8 + tl.arange(0, BLOCK_SIZE_X * 8)
+    scale_offsets = num_cols * 8 * scale_offsets_y[:, None] + scale_offsets_x[None, :]
+    scale_masks_y = scale_offsets_y < num_rows // group_size
+    scale_masks_x = scale_offsets_x < num_cols * 8
+    scale_masks = scale_masks_y[:, None] & scale_masks_x[None, :]
+
+    # Load the scales.
+    scales = tl.load(scales_ptr + scale_offsets, scale_masks, 0.0)
+    scales = tl.broadcast_to(scales, (BLOCK_SIZE_Y, BLOCK_SIZE_X * 8))
+
+    # Dequantize.
+    iweights = (iweights - zeros) * scales
+    iweights = iweights.to(result_ptr.type.element_ty)
+
+    # Finally, store.
+    tl.store(result_ptr + result_offsets, iweights, result_masks)
+
+
+@triton.jit
+def awq_gemm_kernel(
+    a_ptr,
+    b_ptr,
+    c_ptr,
+    zeros_ptr,
+    scales_ptr,
+    M,
+    N,
+    K,
+    group_size,
+    BLOCK_SIZE_M: tl.constexpr,
+    BLOCK_SIZE_N: tl.constexpr,
+    BLOCK_SIZE_K: tl.constexpr,
+    SPLIT_K: tl.constexpr,
+):
+    pid = tl.program_id(axis=0)
+    pid_z = tl.program_id(1)
+
+    # NOTE: This doesn't work in TRITON_INTERPRET=1 mode.  Use below instead.
+    # num_pid_n = (N + BLOCK_SIZE_N - 1) // BLOCK_SIZE_N
+    num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
+
+    pid_m = pid // num_pid_n
+    pid_n = pid % num_pid_n
+
+    accumulator_dtype = c_ptr.type.element_ty
+
+    # NOTE: This doesn't work in TRITON_INTERPRET=1 mode.  Use below instead.
+    # accumulator = tl.arange(0, BLOCK_SIZE_N)
+    # accumulator = tl.broadcast_to(accumulator[None, :],
+    # (BLOCK_SIZE_M, BLOCK_SIZE_N))
+    # accumulator = accumulator & 0x0
+    # accumulator = accumulator.to(accumulator_dtype)
+    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=accumulator_dtype)
+
+    # Create reverse AWQ order as tensor: [0, 4, 1, 5, 2, 6, 3, 7]
+    # that will map given indices to the correct order.
+    reverse_awq_order_tensor = (
+        (tl.arange(0, 2) * 4)[None, :] + tl.arange(0, 4)[:, None]
+    ).reshape(8)
+
+    # Create the necessary shifts to use to unpack.
+    shifts = reverse_awq_order_tensor * 4
+    shifts = tl.broadcast_to(shifts[None, :], (BLOCK_SIZE_K * (BLOCK_SIZE_N // 8), 8))
+    shifts = tl.reshape(shifts, (BLOCK_SIZE_K, BLOCK_SIZE_N))
+
+    # Offsets and masks.
+    offsets_am = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+    masks_am = offsets_am < M
+
+    offsets_bn = pid_n * (BLOCK_SIZE_N // 8) + tl.arange(0, BLOCK_SIZE_N // 8)
+    masks_bn = offsets_bn < N // 8
+
+    offsets_zn = pid_n * (BLOCK_SIZE_N // 8) + tl.arange(0, BLOCK_SIZE_N // 8)
+    masks_zn = offsets_zn < N // 8
+
+    offsets_sn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    masks_sn = offsets_sn < N
+
+    offsets_k = pid_z * BLOCK_SIZE_K + tl.arange(0, BLOCK_SIZE_K)
+    offsets_a = K * offsets_am[:, None] + offsets_k[None, :]
+    offsets_b = (N // 8) * offsets_k[:, None] + offsets_bn[None, :]
+
+    a_ptrs = a_ptr + offsets_a
+    b_ptrs = b_ptr + offsets_b
+
+    # NOTE: Use this in TRITON_INTERPRET=1 mode instead of tl.cdiv
+    # block_offset = BLOCK_SIZE_K * SPLIT_K
+    # for k in range(0, (K + block_offset - 1) // (block_offset)):
+    for k in range(0, tl.cdiv(K, BLOCK_SIZE_K * SPLIT_K)):
+        masks_k = offsets_k < K
+        masks_a = masks_am[:, None] & masks_k[None, :]
+        a = tl.load(a_ptrs, mask=masks_a, other=0.0)
+
+        masks_b = masks_k[:, None] & masks_bn[None, :]
+        b = tl.load(b_ptrs, mask=masks_b, other=0.0)
+        b = tl.interleave(b, b)
+        b = tl.interleave(b, b)
+        b = tl.interleave(b, b)
+
+        # Dequantize b.
+        offsets_szk = (
+            BLOCK_SIZE_K * SPLIT_K * k + pid_z * BLOCK_SIZE_K
+        ) // group_size + tl.arange(0, 1)
+        offsets_z = (N // 8) * offsets_szk[:, None] + offsets_zn[None, :]
+        masks_zk = offsets_szk < K // group_size
+        masks_z = masks_zk[:, None] & masks_zn[None, :]
+        zeros_ptrs = zeros_ptr + offsets_z
+        zeros = tl.load(zeros_ptrs, mask=masks_z, other=0.0)
+        zeros = tl.interleave(zeros, zeros)
+        zeros = tl.interleave(zeros, zeros)
+        zeros = tl.interleave(zeros, zeros)
+        zeros = tl.broadcast_to(zeros, (BLOCK_SIZE_K, BLOCK_SIZE_N))
+
+        offsets_s = N * offsets_szk[:, None] + offsets_sn[None, :]
+        masks_sk = offsets_szk < K // group_size
+        masks_s = masks_sk[:, None] & masks_sn[None, :]
+        scales_ptrs = scales_ptr + offsets_s
+        scales = tl.load(scales_ptrs, mask=masks_s, other=0.0)
+        scales = tl.broadcast_to(scales, (BLOCK_SIZE_K, BLOCK_SIZE_N))
+
+        b = (b >> shifts) & 0xF
+        zeros = (zeros >> shifts) & 0xF
+        b = (b - zeros) * scales
+        b = b.to(c_ptr.type.element_ty)
+
+        # Accumulate results.
+        accumulator = tl.dot(a, b, accumulator, out_dtype=accumulator_dtype)
+
+        offsets_k += BLOCK_SIZE_K * SPLIT_K
+        a_ptrs += BLOCK_SIZE_K * SPLIT_K
+        b_ptrs += BLOCK_SIZE_K * SPLIT_K * (N // 8)
+
+    c = accumulator.to(c_ptr.type.element_ty)
+    offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    c_ptrs = c_ptr + pid_z * N * M + N * offs_cm[:, None] + offs_cn[None, :]
+    c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
+    tl.store(c_ptrs, c, mask=c_mask)
+
+
+# qweights - [K     , M // 8], int32
+# scales   - [K // G, M     ], float16
+# zeros    - [K // G, M // 8], int32
+def awq_dequantize_triton(
+    qweight: torch.Tensor,
+    scales: torch.Tensor,
+    zeros: torch.Tensor,
+    block_size_x: int = 32,
+    block_size_y: int = 32,
+) -> torch.Tensor:
+    K = qweight.shape[0]
+    M = scales.shape[1]
+    group_size = qweight.shape[0] // scales.shape[0]
+
+    assert K > 0 and M > 0
+    assert scales.shape[0] == K // group_size and scales.shape[1] == M
+    assert zeros.shape[0] == K // group_size and zeros.shape[1] == M // 8
+    assert group_size <= K
+    assert group_size in AWQ_TRITON_SUPPORTED_GROUP_SIZES or group_size == K
+
+    # Result tensor:
+    # number of rows = same as input tensor
+    # number of cols = 8 x input tensor num cols
+    result = torch.empty(
+        qweight.shape[0],
+        qweight.shape[1] * 8,
+        device=qweight.device,
+        dtype=scales.dtype,
+    )
+
+    Y = qweight.shape[0]  # num rows
+    X = qweight.shape[1]  # num cols
+
+    grid = lambda META: (
+        triton.cdiv(X, META["BLOCK_SIZE_X"]),
+        triton.cdiv(Y, META["BLOCK_SIZE_Y"]),
+    )
+    awq_dequantize_kernel[grid](
+        qweight,
+        scales,
+        zeros,
+        group_size,
+        result,
+        X,
+        Y,
+        BLOCK_SIZE_X=block_size_x,
+        BLOCK_SIZE_Y=block_size_y,
+    )
+
+    return result
+
+
+# input   - [M, K]
+# qweight - [K, N // 8]
+# qzeros  - [K // G, N // 8]
+# scales  - [K // G, N]
+# split_k_iters - parallelism along K-dimension, int, power of 2.
+def awq_gemm_triton(
+    input: torch.Tensor,
+    qweight: torch.Tensor,
+    scales: torch.Tensor,
+    qzeros: torch.Tensor,
+    split_k_iters: int,
+    block_size_m: int = 32,
+    block_size_n: int = 32,
+    block_size_k: int = 32,
+) -> torch.Tensor:
+    M, K = input.shape
+    N = qweight.shape[1] * 8
+    group_size = qweight.shape[0] // qzeros.shape[0]
+
+    assert N > 0 and K > 0 and M > 0
+    assert qweight.shape[0] == K and qweight.shape[1] == N // 8
+    assert qzeros.shape[0] == K // group_size and qzeros.shape[1] == N // 8
+    assert scales.shape[0] == K // group_size and scales.shape[1] == N
+    assert split_k_iters & (split_k_iters - 1) == 0 and split_k_iters != 0
+    assert split_k_iters <= 32
+    assert group_size <= K
+    assert group_size in AWQ_TRITON_SUPPORTED_GROUP_SIZES or group_size == K
+
+    grid = lambda META: (
+        triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]),
+        split_k_iters,
+    )
+
+    result = torch.zeros((split_k_iters, M, N), dtype=scales.dtype, device=input.device)
+
+    # A = input, B = qweight, C = result
+    # A = M x K, B = K x N, C = M x N
+    awq_gemm_kernel[grid](
+        input,
+        qweight,
+        result,
+        qzeros,
+        scales,
+        M,
+        N,
+        K,
+        group_size,
+        BLOCK_SIZE_M=block_size_m,
+        BLOCK_SIZE_N=block_size_n,
+        BLOCK_SIZE_K=block_size_k,
+        SPLIT_K=split_k_iters,
+    )
+
+    result = result.sum(0)
+
+    return result
+
+
+def awq_dequantize_decomposition(
+    qweight: torch.Tensor,
+    scales: torch.Tensor,
+    zeros: torch.Tensor,
+) -> torch.Tensor:
+    qweight_tmp = qweight
+    qzeros_tmp = zeros
+    qweight_list = []
+    qzeros_list = []
+    shifts = [0, 4, 1, 5, 2, 6, 3, 7]
+    for i in range(0, 8):
+        shift_num = shifts[i] * 4
+        qzeros_list.append((qzeros_tmp.reshape(-1, 1) >> shift_num) & 0xF)
+        qweight_list.append((qweight_tmp.reshape(-1, 1) >> shift_num) & 0xF)
+    qzeros_tmp = (
+        torch.cat(qzeros_list, dim=-1).reshape(qzeros_tmp.shape[0], -1).to(scales.dtype)
+    )
+    qweight_tmp = (
+        torch.cat(qweight_list, dim=-1)
+        .reshape(qweight_tmp.shape[0], -1)
+        .to(scales.dtype)
+    )
+    res = (
+        qweight_tmp.reshape(qzeros_tmp.shape[0], -1, qzeros_tmp.shape[1])
+        - qzeros_tmp.unsqueeze(1)
+    ) * scales.unsqueeze(1)
+    return res.reshape(qweight_tmp.shape[0], -1)

python/sglang/srt/layers/quantization/awq/schemes/__init__.py

@@ -0,0 +1,19 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from .awq_cpu import AWQIntelAMXLinearScheme, AWQIntelAMXMoEScheme
+from .awq_linear import AWQAscendLinearScheme, AWQLinearScheme
+from .awq_marlin import AWQMarlinLinearScheme
+from .awq_moe import AWQAscendMoEScheme, AWQMoEScheme
+from .awq_scheme import AWQLinearSchemeBase, AWQMoESchemeBase
+
+__all__ = [
+    "AWQLinearSchemeBase",
+    "AWQMoESchemeBase",
+    "AWQLinearScheme",
+    "AWQAscendLinearScheme",
+    "AWQIntelAMXLinearScheme",
+    "AWQMarlinLinearScheme",
+    "AWQMoEScheme",
+    "AWQAscendMoEScheme",
+    "AWQIntelAMXMoEScheme",
+]

python/sglang/srt/layers/quantization/awq/schemes/awq_cpu.py

@@ -0,0 +1,40 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import torch
+
+from sglang.srt.hardware_backend.cpu.quantization.awq_kernels import (
+    AWQIntelAMXLinearKernel,
+    AWQIntelAMXMoEKernel,
+)
+from sglang.srt.layers.moe import MoeRunnerConfig
+
+from .awq_linear import AWQLinearScheme
+from .awq_moe import AWQMoEScheme
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.quantization.awq.awq import AWQConfig
+
+__all__ = ["AWQIntelAMXLinearScheme", "AWQIntelAMXMoEScheme"]
+
+
+class AWQIntelAMXLinearScheme(AWQLinearScheme):
+    """Linear scheme for AWQ on Intel CPU with AMX."""
+
+    def _init_kernel(self, quant_config: AWQConfig):
+        return AWQIntelAMXLinearKernel(quant_config)
+
+
+class AWQIntelAMXMoEScheme(AWQMoEScheme):
+    """MoE scheme for AWQ on Intel CPU with AMX."""
+
+    def _init_kernel(self, quant_config: AWQConfig):
+        return AWQIntelAMXMoEKernel(quant_config)
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+        self.kernel.create_moe_runner(layer, moe_runner_config)

python/sglang/srt/layers/quantization/awq/schemes/awq_linear.py

@@ -0,0 +1,110 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, List, Optional
+
+import torch
+
+from sglang.srt.layers.parameter import GroupQuantScaleParameter, PackedvLLMParameter
+
+from .awq_scheme import AWQLinearSchemeBase
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.quantization.awq.awq import AWQConfig
+
+__all__ = ["AWQLinearScheme", "AWQAscendLinearScheme"]
+
+
+class AWQLinearScheme(AWQLinearSchemeBase):
+    def __init__(self, quant_config: AWQConfig):
+        self.quant_config = quant_config
+        self.kernel = self._init_kernel(quant_config)
+
+    def _init_kernel(self, quant_config: AWQConfig):
+        from sglang.srt.hardware_backend.gpu.quantization.awq_kernels import (
+            AWQLinearKernel,
+        )
+
+        return AWQLinearKernel(quant_config)
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        input_size_per_partition: int,
+        output_partition_sizes: List[int],
+        params_dtype: torch.dtype,
+        weight_loader,
+        **kwargs,
+    ):
+        if input_size_per_partition % self.quant_config.group_size != 0:
+            raise ValueError(
+                "The input size is not aligned with the quantized "
+                "weight shape. This can be caused by too large "
+                "tensor parallel size."
+            )
+
+        output_size_per_partition = sum(output_partition_sizes)
+        if output_size_per_partition % self.quant_config.pack_factor != 0:
+            raise ValueError(
+                "The output size is not aligned with the quantized "
+                "weight shape. This can be caused by too large "
+                "tensor parallel size."
+            )
+
+        qweight = PackedvLLMParameter(
+            data=torch.empty(
+                input_size_per_partition,
+                output_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            input_dim=0,
+            output_dim=1,
+            packed_dim=1,
+            packed_factor=self.quant_config.pack_factor,
+            weight_loader=weight_loader,
+        )
+
+        qzeros = PackedvLLMParameter(
+            data=torch.empty(
+                input_size_per_partition // self.quant_config.group_size,
+                output_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            input_dim=0,
+            output_dim=1,
+            packed_dim=1,
+            packed_factor=self.quant_config.pack_factor,
+            weight_loader=weight_loader,
+        )
+
+        scales = GroupQuantScaleParameter(
+            data=torch.empty(
+                input_size_per_partition // self.quant_config.group_size,
+                output_size_per_partition,
+                dtype=params_dtype,
+            ),
+            input_dim=0,
+            output_dim=1,
+            weight_loader=weight_loader,
+        )
+
+        layer.register_parameter("qweight", qweight)
+        layer.register_parameter("qzeros", qzeros)
+        layer.register_parameter("scales", scales)
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        self.kernel.process_weights_after_loading(layer)
+
+    def apply_weights(
+        self, layer: torch.nn.Module, x: torch.Tensor, bias: Optional[torch.Tensor]
+    ):
+        return self.kernel.apply(layer, x, bias)
+
+
+class AWQAscendLinearScheme(AWQLinearScheme):
+    def _init_kernel(self, quant_config: AWQConfig):
+        from sglang.srt.hardware_backend.npu.quantization.awq_kernels import (
+            AWQAscendLinearKernel,
+        )
+
+        return AWQAscendLinearKernel(quant_config)

python/sglang/srt/layers/quantization/awq/schemes/awq_marlin.py

@@ -0,0 +1,109 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Optional
+
+import torch
+
+from sglang.srt.layers.parameter import GroupQuantScaleParameter, PackedvLLMParameter
+from sglang.srt.layers.quantization.marlin_utils import verify_marlin_supports_shape
+
+from .awq_scheme import AWQLinearSchemeBase
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.quantization.awq.awq import AWQMarlinConfig
+
+__all__ = ["AWQMarlinLinearScheme"]
+
+
+class AWQMarlinLinearScheme(AWQLinearSchemeBase):
+    def __init__(self, quant_config: AWQMarlinConfig):
+        self.quant_config = quant_config
+        self.kernel = self._init_kernel(quant_config)
+
+    def _init_kernel(self, quant_config: AWQMarlinConfig):
+        from sglang.srt.hardware_backend.gpu.quantization.awq_kernels import (
+            AWQMarlinLinearKernel,
+        )
+
+        return AWQMarlinLinearKernel(quant_config)
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        input_size_per_partition: int,
+        output_partition_sizes: list[int],
+        input_size: int,
+        params_dtype: torch.dtype,
+        weight_loader,
+        **kwargs,
+    ) -> None:
+        output_size_per_partition = sum(output_partition_sizes)
+
+        group_size = (
+            self.quant_config.group_size
+            if self.quant_config.group_size != -1
+            else input_size
+        )
+
+        verify_marlin_supports_shape(
+            output_size_per_partition=output_size_per_partition,
+            input_size_per_partition=input_size_per_partition,
+            input_size=input_size,
+            group_size=group_size,
+        )
+
+        qweight = PackedvLLMParameter(
+            data=torch.empty(
+                input_size_per_partition,
+                output_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            input_dim=0,
+            output_dim=1,
+            packed_dim=1,
+            packed_factor=self.quant_config.pack_factor,
+            weight_loader=weight_loader,
+        )
+
+        num_groups = input_size_per_partition // group_size
+
+        qzeros = PackedvLLMParameter(
+            data=torch.empty(
+                num_groups,
+                output_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            input_dim=0,
+            output_dim=1,
+            packed_dim=1,
+            packed_factor=self.quant_config.pack_factor,
+            weight_loader=weight_loader,
+        )
+
+        scales = GroupQuantScaleParameter(
+            data=torch.empty(
+                num_groups,
+                output_size_per_partition,
+                dtype=params_dtype,
+            ),
+            input_dim=0,
+            output_dim=1,
+            weight_loader=weight_loader,
+        )
+
+        layer.register_parameter("qweight", qweight)
+        layer.register_parameter("qzeros", qzeros)
+        layer.register_parameter("scales", scales)
+
+        layer.input_size_per_partition = input_size_per_partition
+        layer.output_size_per_partition = output_size_per_partition
+        layer.num_groups = num_groups
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        self.kernel.process_weights_after_loading(layer)
+
+    def apply_weights(
+        self, layer: torch.nn.Module, x: torch.Tensor, bias: Optional[torch.Tensor]
+    ):
+        return self.kernel.apply(layer, x, bias)

python/sglang/srt/layers/quantization/awq/schemes/awq_moe.py

@@ -0,0 +1,156 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import torch
+
+from sglang.srt.layers.linear import set_weight_attrs
+from sglang.srt.layers.moe import (
+    MoeRunner,
+    MoeRunnerBackend,
+    MoeRunnerConfig,
+    get_moe_runner_backend,
+)
+
+from .awq_scheme import AWQMoESchemeBase
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.token_dispatcher import StandardDispatchOutput
+    from sglang.srt.layers.quantization.awq.awq import AWQConfig, AWQMarlinConfig
+
+__all__ = ["AWQMoEScheme", "AWQAscendMoEScheme"]
+
+
+class AWQMoEScheme(AWQMoESchemeBase):
+    def __init__(self, quant_config: AWQMarlinConfig):
+        self.quant_config = quant_config
+        if self.quant_config.weight_bits != 4:
+            raise ValueError("AWQMoEScheme only supports 4bit now.")
+        self.kernel = self._init_kernel(quant_config)
+
+    def _init_kernel(self, quant_config: AWQMarlinConfig):
+        from sglang.srt.hardware_backend.gpu.quantization.awq_kernels import (
+            AWQMoEKernel,
+        )
+
+        return AWQMoEKernel(quant_config)
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoeWeightScaleSupported
+
+        extra_weight_attrs.update(
+            {
+                "is_transposed": True,
+                "quant_method": FusedMoeWeightScaleSupported.GROUP.value,
+            }
+        )
+
+        w13_qweight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                2 * intermediate_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_qweight", w13_qweight)
+        set_weight_attrs(w13_qweight, extra_weight_attrs)
+
+        w2_qweight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                intermediate_size_per_partition,
+                hidden_size // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_qweight", w2_qweight)
+        set_weight_attrs(w2_qweight, extra_weight_attrs)
+
+        num_groups_w13 = hidden_size // self.quant_config.group_size
+        num_groups_w2 = intermediate_size_per_partition // self.quant_config.group_size
+
+        w13_scales = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                num_groups_w13,
+                intermediate_size_per_partition * 2,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_scales", w13_scales)
+        set_weight_attrs(w13_scales, extra_weight_attrs)
+
+        w2_scales = torch.nn.Parameter(
+            torch.empty(num_experts, num_groups_w2, hidden_size, dtype=params_dtype),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_scales", w2_scales)
+        set_weight_attrs(w2_scales, extra_weight_attrs)
+
+        w13_qzeros = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                num_groups_w13,
+                2 * intermediate_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_qzeros", w13_qzeros)
+        set_weight_attrs(w13_qzeros, extra_weight_attrs)
+
+        w2_qzeros = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                num_groups_w2,
+                hidden_size // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_qzeros", w2_qzeros)
+        set_weight_attrs(w2_qzeros, extra_weight_attrs)
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        self.kernel.process_weights_after_loading(layer)
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        assert get_moe_runner_backend().is_auto()
+        self.moe_runner_config = moe_runner_config
+        self.kernel.runner = MoeRunner(MoeRunnerBackend.MARLIN, moe_runner_config)
+
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: StandardDispatchOutput,
+    ):
+        return self.kernel.apply(layer, dispatch_output)
+
+
+class AWQAscendMoEScheme(AWQMoEScheme):
+    def _init_kernel(self, quant_config: AWQConfig):
+        from sglang.srt.hardware_backend.npu.quantization.awq_kernels import (
+            AWQAscendMoEKernel,
+        )
+
+        return AWQAscendMoEKernel(quant_config)
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config

python/sglang/srt/layers/quantization/awq/schemes/awq_scheme.py

@@ -0,0 +1,54 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from abc import abstractmethod
+from typing import TYPE_CHECKING, Optional
+
+import torch
+
+from sglang.srt.layers.moe import MoeRunnerConfig
+from sglang.srt.layers.quantization.base_scheme import BaseLinearScheme, BaseMoEScheme
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.token_dispatcher import StandardDispatchOutput
+
+__all__ = ["AWQLinearSchemeBase", "AWQMoESchemeBase"]
+
+
+class AWQLinearSchemeBase(BaseLinearScheme):
+    @abstractmethod
+    def create_weights(self, *args, **kwargs):
+        raise NotImplementedError
+
+    @abstractmethod
+    def process_weights_after_loading(self, layer: torch.nn.Module):
+        raise NotImplementedError
+
+    @abstractmethod
+    def apply_weights(
+        self, layer: torch.nn.Module, x: torch.Tensor, bias: Optional[torch.Tensor]
+    ):
+        raise NotImplementedError
+
+
+class AWQMoESchemeBase(BaseMoEScheme):
+    @abstractmethod
+    def create_weights(self, *args, **kwargs):
+        raise NotImplementedError
+
+    @abstractmethod
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        raise NotImplementedError
+
+    @abstractmethod
+    def process_weights_after_loading(self, layer: torch.nn.Module):
+        raise NotImplementedError
+
+    @abstractmethod
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: "StandardDispatchOutput",
+    ):
+        raise NotImplementedError

python/sglang/srt/layers/quantization/base_config.py

@@ -0,0 +1,269 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Adapted from https://raw.githubusercontent.com/vllm-project/vllm/v0.5.5/vllm/model_executor/layers/quantization/base_config.py
+from __future__ import annotations
+
+import inspect
+from abc import ABC, abstractmethod
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Type
+
+import torch
+from torch import nn
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.moe_runner import MoeRunnerConfig
+    from sglang.srt.layers.moe.moe_runner.triton import TritonMoeQuantInfo
+    from sglang.srt.layers.moe.token_dispatcher import CombineInput, DispatchOutput
+    from sglang.srt.models.utils import WeightsMapper
+
+
+class QuantizeMethodBase(ABC):
+    """Base class for different quantized methods."""
+
+    def create_weights(
+        self, layer: torch.nn.Module, *weight_args, **extra_weight_attrs
+    ):
+        """Create weights for a layer.
+
+        The weights will be set as attributes of the layer."""
+        raise NotImplementedError()
+
+    @abstractmethod
+    def apply(self, layer: torch.nn.Module, *args, **kwargs) -> torch.Tensor:
+        """Apply the weights in layer to the input tensor.
+
+        Expects create_weights to have been called before on the layer."""
+        raise NotImplementedError()
+
+    def process_weights_after_loading(self, layer: nn.Module) -> None:
+        """Process the weight after loading.
+
+        This can be used for example, to transpose weights for computation.
+        """
+        return
+
+
+class LinearMethodBase(QuantizeMethodBase):
+    """Base class for different (maybe quantized) linear methods."""
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        input_size_per_partition: int,
+        output_partition_sizes: List[int],
+        input_size: int,
+        output_size: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        """Create weights for a linear layer.
+           The weights will be set as attributes of the layer.
+
+        Args:
+            layer: The layer that is using the LinearMethodBase factory.
+            input_size_per_partition: Size of the weight input dim on rank X.
+            output_partition_sizes: Sizes of the output dim of each logical
+                weight on rank X. E.g., output_partition_sizes for QKVLinear
+                is a list contains the width of Wq, Wk, Wv on rank X.
+            input_size: Size of the input dim of the weight across all ranks.
+            output_size: Size of the output dim of the weight across all ranks.
+            params_dtype: Datatype of the parameters.
+        """
+        raise NotImplementedError()
+
+    @abstractmethod
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """Apply the weights in layer to the input tensor.
+        Expects create_weights to have been called before on the layer."""
+        raise NotImplementedError()
+
+
+class FusedMoEMethodBase(QuantizeMethodBase):
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        raise NotImplementedError
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        raise NotImplementedError
+
+    @abstractmethod
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: DispatchOutput,
+    ) -> CombineInput:
+        raise NotImplementedError
+
+    def get_triton_quant_info(self, layer: torch.nn.Module) -> TritonMoeQuantInfo:
+        """Return a ``TritonMoeQuantInfo`` describing the quantisation state
+        stored on *layer*.
+
+        The LoRA MoE runner calls this so that ``invoke_fused_moe_kernel``
+        receives the correct flags / scales / block-shape for the base
+        weights.  Each quantisation method must override this with the
+        same construction it already uses inside ``apply()``.
+        """
+        raise NotImplementedError(
+            f"{type(self).__name__} must implement get_triton_quant_info()"
+        )
+
+
+class QuantizationConfig(ABC):
+    """Base class for quantization configs."""
+
+    def __init__(self):
+        super().__init__()
+        # mapping is updated by models as they initialize
+        self.packed_modules_mapping: Dict[str, List[str]] = dict()
+
+    def update_packed_modules_mapping(self, mapping: Dict[str, List[str]]) -> None:
+        self.packed_modules_mapping = mapping
+
+    @abstractmethod
+    def get_name(self) -> str:
+        """Name of the quantization method."""
+        raise NotImplementedError()
+
+    @abstractmethod
+    def get_supported_act_dtypes(self) -> List[torch.dtype]:
+        """List of supported activation dtypes."""
+        raise NotImplementedError()
+
+    @classmethod
+    @abstractmethod
+    def get_min_capability(cls) -> int:
+        """Minimum GPU capability to support the quantization method.
+
+        E.g., 70 for Volta, 75 for Turing, 80 for Ampere.
+        This requirement is due to the custom CUDA kernels used by the
+        quantization method.
+        """
+        raise NotImplementedError()
+
+    @staticmethod
+    @abstractmethod
+    def get_config_filenames() -> List[str]:
+        """List of filenames to search for in the model directory."""
+        raise NotImplementedError()
+
+    @classmethod
+    @abstractmethod
+    def from_config(cls, config: Dict[str, Any]) -> QuantizationConfig:
+        """Create a config class from the model's quantization config."""
+        raise NotImplementedError()
+
+    @classmethod
+    def override_quantization_method(cls, hf_quant_cfg, user_quant) -> Optional[str]:
+        """
+        Detects if this quantization method can support a given checkpoint
+        format by overriding the user specified quantization method --
+        this method should only be overwritten by subclasses in exceptional
+        circumstances
+        """
+        return None
+
+    @classmethod
+    def _modelopt_override_quantization_method(
+        cls, hf_quant_config, user_quant
+    ) -> Optional[str]:
+        """Shared ModelOpt quantization method override logic."""
+        if hf_quant_config is None:
+            return None
+
+        # Check if this is a ModelOpt config
+        quant_algo = hf_quant_config.get("quant_algo", "").upper()
+
+        # If user specified generic "modelopt", auto-detect the specific method
+        if user_quant == "modelopt":
+            if "FP8" in quant_algo:
+                return "modelopt_fp8"
+            elif "NVFP4" in quant_algo or "FP4" in quant_algo:
+                return "modelopt_fp4"
+
+        # The hf_quant_config may be a parsed quant config, so we need to check the
+        # quant_method.
+        if hf_quant_config.get("quant_method", "") == "modelopt_fp8":
+            return "modelopt_fp8"
+        elif hf_quant_config.get("quant_method", "") == "modelopt_fp4":
+            return "modelopt_fp4"
+
+        return None
+
+    @staticmethod
+    def get_from_keys(config: Dict[str, Any], keys: List[str]) -> Any:
+        """Get a value from the model's quantization config."""
+        for key in keys:
+            if key in config:
+                return config[key]
+        raise ValueError(
+            f"Cannot find any of {keys} in the model's " "quantization config."
+        )
+
+    @staticmethod
+    def get_from_keys_or(config: Dict[str, Any], keys: List[str], default: Any) -> Any:
+        """Get a optional value from the model's quantization config."""
+        try:
+            return QuantizationConfig.get_from_keys(config, keys)
+        except ValueError:
+            return default
+
+    @abstractmethod
+    def get_quant_method(
+        self, layer: torch.nn.Module, prefix: str
+    ) -> Optional[QuantizeMethodBase]:
+        """Get the quantize method to use for the quantized layer.
+
+        Args:
+            layer: The layer for the quant method.
+            prefix: The full name of the layer in the state dict
+        Returns:
+            The quantize method. None if the given layer doesn't support quant
+            method.
+        """
+        raise NotImplementedError()
+
+    @abstractmethod
+    def get_scaled_act_names(self) -> List[str]:
+        """Returns the activation function names that should be post-scaled.
+
+        For now, this is only used by AWQ.
+        """
+        raise NotImplementedError()
+
+    def apply_weight_name_mapper(
+        self, hf_to_sglang_mapper: WeightsMapper
+    ):  # noqa: B027
+        """
+        Interface for models to update module names referenced in
+        quantization configs in order to reflect the sglang model structure
+        :param hf_to_sglang_mapper: maps from hf model structure (the assumed
+            structure of the qconfig) to sglang model structure
+        """
+        pass
+
+
+def method_has_implemented_embedding(method_class: Type[QuantizeMethodBase]) -> bool:
+    """
+    Not all quant methods have embedding implemented, so we need to check that
+    it exists for our given method. We check this by making sure the function
+    has been changed from the base implementation.
+    """
+    base_embedding = inspect.getattr_static(QuantizeMethodBase, "embedding", None)
+    class_embedding = inspect.getattr_static(method_class, "embedding", None)
+
+    return class_embedding is not None and class_embedding is not base_embedding

python/sglang/srt/layers/quantization/base_scheme.py

@@ -0,0 +1,99 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from abc import ABC, abstractmethod
+from typing import TYPE_CHECKING, Optional
+
+import torch
+
+from sglang.srt.layers.moe import MoeRunnerConfig
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.token_dispatcher import StandardDispatchOutput
+
+__all__ = ["BaseLinearScheme", "BaseMoEScheme"]
+
+
+class BaseLinearScheme(ABC):
+    """
+    Abstract class used to describe the weight creation and forward pass
+    of different quantization schemes.
+    """
+
+    @abstractmethod
+    def create_weights(self, *args, **kwargs):
+        """
+        Weight creation for the particular scheme. Inputs to this function
+
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def process_weights_after_loading(self, layer: torch.nn.Module):
+        """
+        Called after weight loading is complete for any cleanup that
+        needs to occur.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def apply_weights(
+        self, layer: torch.nn.Module, x: torch.Tensor, bias: Optional[torch.Tensor]
+    ):
+        """
+        Run the forward pass for the particular scheme. This is where
+        scheme-specific dequant/quant steps/kernels should be applied.
+
+        :param layer: torch.nn.Module with the registered weights and
+            other parameters relevant to the particular scheme.
+        :param x: input to the layer
+        :param bias: bias parameter
+
+        """
+        raise NotImplementedError
+
+
+class BaseMoEScheme(ABC):
+    """
+    Abstract class used to describe the weight creation and forward pass
+    of different quantization schemes.
+    """
+
+    @abstractmethod
+    def create_weights(self, *args, **kwargs):
+        """
+        Weight creation for the particular scheme. Inputs to this function
+
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        raise NotImplementedError
+
+    @abstractmethod
+    def process_weights_after_loading(self, layer: torch.nn.Module):
+        """
+        Called after weight loading is complete for any cleanup that
+        needs to occur.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: "StandardDispatchOutput",
+    ):
+        """
+        Run the forward pass for the particular scheme. This is where
+        scheme-specific dequant/quant steps/kernels should be applied.
+
+        :param layer: torch.nn.Module with the registered weights and
+            other parameters relevant to the particular scheme.
+        :param x: input to the layer
+        :param bias: bias parameter
+
+        """
+        raise NotImplementedError

python/sglang/srt/layers/quantization/bitsandbytes.py

@@ -0,0 +1,601 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Adapted from: https://github.com/vllm-project/vllm/blob/d4d2751732c3ccae162a5a0160c7d4fe05d2779a/vllm/model_executor/layers/quantization/bitsandbytes.py
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any, Optional
+
+import torch
+from packaging import version
+
+from sglang.srt.layers.linear import LinearBase
+from sglang.srt.layers.quantization.base_config import (
+    FusedMoEMethodBase,
+    LinearMethodBase,
+    QuantizationConfig,
+    QuantizeMethodBase,
+)
+from sglang.srt.layers.quantization.unquant import UnquantizedLinearMethod
+from sglang.srt.utils import set_weight_attrs
+from sglang.srt.utils.custom_op import register_custom_op
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
+
+
+class BitsAndBytesConfig(QuantizationConfig):
+    """Config class for BitsAndBytes Quantization.
+
+    Reference: https://arxiv.org/abs/2305.14314
+    """
+
+    def __init__(
+        self,
+        load_in_8bit: bool = False,
+        load_in_4bit: bool = True,
+        bnb_4bit_compute_dtype: str = "float32",
+        bnb_4bit_quant_storage: str = "uint8",
+        bnb_4bit_quant_type: str = "fp4",
+        bnb_4bit_use_double_quant: bool = False,
+        llm_int8_enable_fp32_cpu_offload: bool = False,
+        llm_int8_has_fp16_weight: bool = False,
+        llm_int8_skip_modules: list[str] | None = None,
+        llm_int8_threshold: float = 6.0,
+    ) -> None:
+        super().__init__()
+        self.load_in_8bit = load_in_8bit
+        self.load_in_4bit = load_in_4bit
+        self.bnb_4bit_compute_dtype = bnb_4bit_compute_dtype
+        self.bnb_4bit_quant_storage = bnb_4bit_quant_storage
+        self.bnb_4bit_quant_type = bnb_4bit_quant_type
+        self.bnb_4bit_use_double_quant = bnb_4bit_use_double_quant
+        self.llm_int8_enable_fp32_cpu_offload = llm_int8_enable_fp32_cpu_offload
+        self.llm_int8_has_fp16_weight = llm_int8_has_fp16_weight
+        self.llm_int8_skip_modules = llm_int8_skip_modules or []
+        self.llm_int8_threshold = llm_int8_threshold
+
+        if self.bnb_4bit_quant_storage not in ["uint8"]:
+            raise ValueError(
+                f"Unsupported bnb_4bit_quant_storage: {self.bnb_4bit_quant_storage}"
+            )
+
+    def __repr__(self) -> str:
+        return (
+            f"BitsAndBytesConfig(load_in_8bit={self.load_in_8bit}, "
+            f"load_in_4bit={self.load_in_4bit}, "
+            f"bnb_4bit_compute_dtype={self.bnb_4bit_compute_dtype}, "
+            f"bnb_4bit_quant_storage={self.bnb_4bit_quant_storage}, "
+            f"bnb_4bit_quant_type={self.bnb_4bit_quant_type}, "
+            f"llm_int8_skip_modules={self.llm_int8_skip_modules})"
+        )
+
+    def get_name(self) -> str:
+        return "bitsandbytes"
+
+    def get_scaled_act_names(self) -> list[str]:
+        return []
+
+    def get_supported_act_dtypes(self) -> list[torch.dtype]:
+        return [torch.float32, torch.float16, torch.bfloat16]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 70
+
+    @staticmethod
+    def get_config_filenames() -> list[str]:
+        return []
+
+    @classmethod
+    def from_config(cls, config: dict[str, Any]) -> BitsAndBytesConfig:
+        def get_safe_value(config, keys, default_value=None):
+            try:
+                value = QuantizationConfig.get_from_keys(config, keys)
+                return value if value is not None else default_value
+            except ValueError:
+                return default_value
+
+        load_in_8bit = get_safe_value(config, ["load_in_8bit"], default_value=False)
+        load_in_4bit = get_safe_value(config, ["load_in_4bit"], default_value=True)
+        bnb_4bit_compute_dtype = get_safe_value(
+            config, ["bnb_4bit_compute_dtype"], default_value="float32"
+        )
+        bnb_4bit_quant_storage = get_safe_value(
+            config, ["bnb_4bit_quant_storage"], default_value="uint8"
+        )
+        bnb_4bit_quant_type = get_safe_value(
+            config, ["bnb_4bit_quant_type"], default_value="fp4"
+        )
+        bnb_4bit_use_double_quant = get_safe_value(
+            config, ["bnb_4bit_use_double_quant"], default_value=False
+        )
+        llm_int8_enable_fp32_cpu_offload = get_safe_value(
+            config, ["llm_int8_enable_fp32_cpu_offload"], default_value=False
+        )
+        llm_int8_has_fp16_weight = get_safe_value(
+            config, ["llm_int8_has_fp16_weight"], default_value=False
+        )
+        llm_int8_skip_modules = get_safe_value(
+            config, ["llm_int8_skip_modules"], default_value=[]
+        )
+        llm_int8_threshold = get_safe_value(
+            config, ["llm_int8_threshold"], default_value=6.0
+        )
+
+        return cls(
+            load_in_8bit=load_in_8bit,
+            load_in_4bit=load_in_4bit,
+            bnb_4bit_compute_dtype=bnb_4bit_compute_dtype,
+            bnb_4bit_quant_storage=bnb_4bit_quant_storage,
+            bnb_4bit_quant_type=bnb_4bit_quant_type,
+            bnb_4bit_use_double_quant=bnb_4bit_use_double_quant,
+            llm_int8_enable_fp32_cpu_offload=llm_int8_enable_fp32_cpu_offload,
+            llm_int8_has_fp16_weight=llm_int8_has_fp16_weight,
+            llm_int8_skip_modules=llm_int8_skip_modules,
+            llm_int8_threshold=llm_int8_threshold,
+        )
+
+    def get_quant_method(
+        self, layer: torch.nn.Module, prefix: str
+    ) -> Optional[QuantizeMethodBase]:
+        from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
+
+        if isinstance(layer, LinearBase):
+            if is_layer_skipped_bnb(prefix, self.llm_int8_skip_modules):
+                return UnquantizedLinearMethod()
+            return BitsAndBytesLinearMethod(self)
+        elif isinstance(layer, FusedMoE):
+            return BitsAndBytesMoEMethod(self)
+        return None
+
+
+def is_layer_skipped_bnb(prefix: str, llm_int8_skip_modules: list[str]):
+    # Split the prefix into its dot-separated components
+    components = prefix.split(".")
+
+    # Check if any of the skip modules exactly matches any component
+    substr_check = any(
+        module_name in components for module_name in llm_int8_skip_modules
+    )
+
+    # Allow certain layers to not be quantized
+    set_components = set(".".join(components[: i + 1]) for i in range(len(components)))
+    set_llm_int8_skip_modules = set(llm_int8_skip_modules)
+    prefix_check = len(set_llm_int8_skip_modules & set_components) != 0
+
+    return substr_check or prefix_check
+
+
+def calculate_quant_ratio(dtype):
+    if dtype.is_floating_point:
+        return torch.finfo(dtype).bits // torch.iinfo(torch.uint8).bits
+    else:
+        return torch.iinfo(dtype).bits // torch.iinfo(torch.uint8).bits
+
+
+class BitsAndBytesLinearMethod(LinearMethodBase):
+    """Linear method for BitsAndBytes.
+
+    Args:
+       quant_config: The BitsAndBytes quantization config.
+    """
+
+    def __init__(self, quant_config: BitsAndBytesConfig):
+        try:
+            import bitsandbytes
+
+            if version.parse(bitsandbytes.__version__) < version.parse("0.46.1"):
+                raise ImportError(
+                    "bitsandbytes version is wrong. Please "
+                    "install bitsandbytes>=0.46.1."
+                )
+        except ImportError as err:
+            raise ImportError(
+                "Please install bitsandbytes>=0.46.1 via "
+                "`pip install bitsandbytes>=0.46.1` to use "
+                "bitsandbytes quantizer."
+            ) from err
+
+        self.quant_config = quant_config
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        input_size_per_partition: int,
+        output_partition_sizes: list[int],
+        input_size: int,
+        output_size: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        from bitsandbytes.nn import Int8Params
+
+        def create_qweight_for_8bit():
+            qweight = Int8Params(
+                data=torch.empty(
+                    sum(output_partition_sizes),
+                    input_size_per_partition,
+                    dtype=torch.int8,
+                ),
+                has_fp16_weights=self.quant_config.llm_int8_has_fp16_weight,
+                requires_grad=False,
+            )
+            set_weight_attrs(
+                qweight,
+                {
+                    "input_dim": 0,
+                    "output_dim": 0,
+                    "pack_factor": 1,
+                    "use_bitsandbytes_8bit": True,
+                    "generation": 0,
+                },
+            )
+            return qweight
+
+        def create_qweight_for_4bit():
+            quant_ratio = calculate_quant_ratio(params_dtype)
+
+            total_size = input_size_per_partition * sum(output_partition_sizes)
+            if total_size % quant_ratio != 0:
+                raise ValueError(
+                    "The input size is not aligned with the quantized weight shape."
+                )
+
+            qweight = torch.nn.Parameter(
+                torch.empty(total_size // quant_ratio, 1, dtype=torch.uint8),
+                requires_grad=False,
+            )
+            set_weight_attrs(
+                qweight,
+                {
+                    "input_dim": 0,
+                    "output_dim": 0,
+                    "pack_factor": quant_ratio,
+                    "use_bitsandbytes_4bit": True,
+                },
+            )
+            return qweight
+
+        if self.quant_config.load_in_8bit:
+            qweight = create_qweight_for_8bit()
+        else:
+            qweight = create_qweight_for_4bit()
+        # Enable parameters to have the same name as in the BNB
+        # checkpoint format.
+        layer.register_parameter("weight", qweight)
+        set_weight_attrs(qweight, extra_weight_attrs)
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        if self.quant_config.load_in_8bit:
+            return self._apply_8bit_weight(layer, x, bias)
+        else:
+            return self._apply_4bit_weight(layer, x, bias)
+
+    def _apply_8bit_weight(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        # only load the bitsandbytes module when needed
+        from bitsandbytes import MatmulLtState, matmul
+
+        original_type = x.dtype
+        original_shape = x.shape
+        reshape_after_matmul = False
+        if x.ndim > 2:
+            x = x.reshape(-1, x.size(-1))
+            reshape_after_matmul = True
+        bf_x = x.to(torch.bfloat16)
+
+        qweight = layer.weight
+        offsets = qweight.bnb_shard_offsets
+        quant_states = qweight.bnb_quant_state
+        matmul_states = qweight.matmul_state
+        generation = qweight.generation
+
+        out_dim_0 = x.shape[0]
+        out_dim_1 = sum(
+            [quant_state[1].shape[0] for quant_state in quant_states.items()]
+        )
+        out = torch.empty(out_dim_0, out_dim_1, dtype=torch.float16, device=x.device)
+
+        current_index = 0
+        for i in range(len(quant_states)):
+            output_size = quant_states[i].shape[0]
+
+            # in profile_run or the first generation of inference,
+            # create new matmul_states
+            if generation == 0 or generation == 1:
+                matmul_states[i] = MatmulLtState()
+                matmul_states[i].CB = qweight[offsets[i] : offsets[i + 1]]
+                matmul_states[i].SCB = quant_states[i].to(x.device)
+                matmul_states[i].threshold = self.quant_config.llm_int8_threshold
+                matmul_states[i].has_fp16_weights = (
+                    self.quant_config.llm_int8_has_fp16_weight
+                )
+                matmul_states[i].is_training = False
+                if (
+                    matmul_states[i].threshold > 0.0
+                    and not matmul_states[i].has_fp16_weights
+                ):
+                    matmul_states[i].use_pool = True
+
+            new_x = bf_x.unsqueeze(0)
+
+            out[:, current_index : current_index + output_size] = matmul(
+                new_x, qweight[offsets[i] : offsets[i + 1]], state=matmul_states[i]
+            )
+
+            current_index += output_size
+
+            # only update the matmul_states if it is not profile_run
+            if (
+                generation > 0
+                and not self.quant_config.llm_int8_has_fp16_weight
+                and matmul_states[i].CB is not None
+                and matmul_states[i].CxB is not None
+            ):
+                del matmul_states[i].CB
+                qweight[offsets[i] : offsets[i + 1]] = matmul_states[i].CxB
+
+        out = out.to(original_type)
+
+        if reshape_after_matmul:
+            out = out.view(*original_shape[:-1], out.size(-1))
+
+        if bias is not None:
+            out += bias
+
+        qweight.generation += 1
+
+        return out
+
+    def _apply_4bit_weight(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: torch.Tensor | None = None,
+    ) -> torch.Tensor:
+        original_type = x.dtype
+        original_shape = x.shape
+        reshape_after_matmul = False
+        if x.ndim > 2:
+            x = x.reshape(-1, x.size(-1))
+            reshape_after_matmul = True
+        bf_x = x.to(torch.bfloat16)
+
+        qweight = layer.weight
+        quant_states = qweight.bnb_quant_state
+        offsets = qweight.bnb_shard_offsets
+
+        out_dim_0 = x.shape[0]
+        out_dim_1 = sum(
+            [quant_state[1].shape[0] for quant_state in quant_states.items()]
+        )
+        out = torch.empty(out_dim_0, out_dim_1, dtype=torch.bfloat16, device=x.device)
+        apply_bnb_4bit(bf_x, qweight, offsets, out)
+        out = out.to(original_type)
+
+        if reshape_after_matmul:
+            out = out.view(*original_shape[:-1], out.size(-1))
+
+        if bias is not None:
+            out += bias
+
+        return out
+
+
+@register_custom_op(mutates_args=["out"])
+def apply_bnb_4bit(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    offsets: torch.Tensor,
+    out: torch.Tensor,
+) -> None:
+    # only load the bitsandbytes module when needed
+    from bitsandbytes import matmul_4bit
+
+    quant_states = weight.bnb_quant_state
+    current_index = 0
+    for i in range(len(quant_states)):
+        output_size = quant_states[i].shape[0]
+        # It is more efficient to use out kwarg like
+        # matmul_4bit(..., out = ...).  Infeasible now due to the bug
+        # https://github.com/TimDettmers/bitsandbytes/issues/1235.
+        # Need to change  after the bug is fixed.
+        out[:, current_index : current_index + output_size] = matmul_4bit(
+            x, weight[offsets[i] : offsets[i + 1]].t(), quant_states[i]
+        )
+        current_index += output_size
+
+
+class BitsAndBytesMoEMethod(FusedMoEMethodBase):
+    """MoE method for BitsAndBytes.
+
+    Args:
+       quant_config: The BitsAndBytes quantization config.
+    """
+
+    def __init__(
+        self,
+        quant_config: BitsAndBytesConfig,
+    ):
+        super().__init__()
+        try:
+            import bitsandbytes
+
+            if version.parse(bitsandbytes.__version__) < version.parse("0.46.1"):
+                raise ImportError(
+                    "bitsandbytes version is wrong. Please "
+                    "install bitsandbytes>=0.46.1."
+                )
+        except ImportError as err:
+            raise ImportError(
+                "Please install bitsandbytes>=0.46.1 via "
+                "`pip install bitsandbytes>=0.46.1` to use "
+                "bitsandbytes quantizer."
+            ) from err
+        self.quant_config = quant_config
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        if self.quant_config.load_in_8bit:
+            call_fun = self._create_weights_8bit
+        else:
+            call_fun = self._create_weights_4bit
+        call_fun(
+            layer,
+            num_experts,
+            hidden_size,
+            intermediate_size_per_partition,
+            params_dtype,
+            **extra_weight_attrs,
+        )
+
+    def create_moe_runner(self, layer: torch.nn.Module, moe_runner_config):
+        self.moe_runner_config = moe_runner_config
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+        from sglang.srt.layers.moe.moe_runner.triton_utils.fused_moe import fused_moe
+        from sglang.srt.layers.moe.token_dispatcher import StandardCombineInput
+
+        x = dispatch_output.hidden_states
+        topk_output = dispatch_output.topk_output
+
+        # TODO(bnell): Do these need to be called on the hot path?
+        if self.quant_config.load_in_8bit:
+            w13, w2 = self._apply_8bit_dequant(layer)
+        else:
+            w13, w2 = self._apply_4bit_dequant(layer)
+
+        moe_runner_config = self.moe_runner_config
+        output = fused_moe(
+            hidden_states=x,
+            w1=w13,
+            w2=w2,
+            topk_output=topk_output,
+            moe_runner_config=moe_runner_config,
+        )
+        return StandardCombineInput(hidden_states=output)
+
+    def _create_weights_4bit(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        quant_ratio = calculate_quant_ratio(params_dtype)
+        # Fused gate_up_proj (column parallel)
+        w13_total_size = (
+            hidden_size * 2 * intermediate_size_per_partition
+        ) // quant_ratio
+        w13_qweight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                w13_total_size,
+                1,
+                dtype=torch.uint8,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight", w13_qweight)
+        set_weight_attrs(w13_qweight, extra_weight_attrs)
+        set_weight_attrs(
+            w13_qweight,
+            {
+                "num_experts": num_experts,
+                "input_dim": hidden_size,
+                "output_dim": 2 * intermediate_size_per_partition,
+                "experts_shape": (
+                    num_experts,
+                    intermediate_size_per_partition * 2,
+                    hidden_size,
+                ),
+                "pack_factor": quant_ratio,
+                "use_bitsandbytes_4bit": True,
+            },
+        )
+        # down_proj (row parallel)
+        w2_total_size = (hidden_size * intermediate_size_per_partition) // quant_ratio
+        w2_qweight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                w2_total_size,
+                1,
+                dtype=torch.uint8,
+            ),
+            requires_grad=False,
+        )
+        set_weight_attrs(
+            w2_qweight,
+            {
+                "num_experts": num_experts,
+                "input_dim": intermediate_size_per_partition,
+                "output_dim": hidden_size,
+                "experts_shape": (
+                    num_experts,
+                    hidden_size,
+                    intermediate_size_per_partition,
+                ),
+                "pack_factor": quant_ratio,
+                "use_bitsandbytes_4bit": True,
+            },
+        )
+        layer.register_parameter("w2_weight", w2_qweight)
+        set_weight_attrs(w2_qweight, extra_weight_attrs)
+
+    def _create_weights_8bit(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        raise NotImplementedError
+
+    def _apply_4bit_dequant(
+        self, layer: torch.nn.Module
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        from bitsandbytes.functional import dequantize_4bit
+
+        w13 = dequantize_4bit(
+            layer.w13_weight.reshape(-1, 1),
+            layer.w13_weight.bnb_quant_state,
+        )
+        w2 = dequantize_4bit(
+            layer.w2_weight.reshape(-1, 1),
+            layer.w2_weight.bnb_quant_state,
+        )
+        w13 = w13.reshape(layer.w13_weight.experts_shape)
+        w2 = w2.reshape(layer.w2_weight.experts_shape)
+        return w13, w2
+
+    def _apply_8bit_dequant(
+        self, layer: torch.nn.Module
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        raise NotImplementedError

python/sglang/srt/layers/quantization/blockwise_int8.py

@@ -0,0 +1,385 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Adapted from https://github.com/vllm-project/vllm/blob/v0.6.4.post1/vllm/model_executor/layers/quantization/fp8.py
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING, Any, Dict, List, Optional
+
+import torch
+from torch.nn import Module
+
+from sglang.srt.layers.moe import MoeRunner, MoeRunnerBackend, MoeRunnerConfig
+from sglang.srt.layers.moe.moe_runner.triton import TritonMoeQuantInfo
+from sglang.srt.layers.parameter import BlockQuantScaleParameter, ModelWeightParameter
+from sglang.srt.layers.quantization.base_config import (
+    FusedMoEMethodBase,
+    LinearMethodBase,
+    QuantizationConfig,
+    QuantizeMethodBase,
+)
+from sglang.srt.layers.quantization.int8_utils import apply_w8a8_block_int8_linear
+from sglang.srt.layers.quantization.unquant import UnquantizedLinearMethod
+from sglang.srt.layers.quantization.utils import is_layer_skipped
+from sglang.srt.runtime_context import get_parallel
+from sglang.srt.utils import set_weight_attrs
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
+
+ACTIVATION_SCHEMES = ["static", "dynamic"]
+
+logger = logging.getLogger(__name__)
+
+
+class BlockInt8Config(QuantizationConfig):
+    """Config class for INT8."""
+
+    def __init__(
+        self,
+        is_checkpoint_int8_serialized: bool = False,
+        activation_scheme: str = "dynamic",
+        ignored_layers: Optional[List[str]] = None,
+        weight_block_size: List[int] = None,
+    ) -> None:
+        self.is_checkpoint_int8_serialized = is_checkpoint_int8_serialized
+        if is_checkpoint_int8_serialized:
+            logger.warning(
+                "Detected int8 checkpoint. Please note that the "
+                "format is experimental and subject to change."
+            )
+        if activation_scheme not in ACTIVATION_SCHEMES:
+            raise ValueError(f"Unsupported activation scheme {activation_scheme}")
+        self.activation_scheme = activation_scheme
+        self.ignored_layers = ignored_layers or []
+        if weight_block_size is not None:
+            if not is_checkpoint_int8_serialized:
+                raise ValueError(
+                    f"The block-wise quantization only supports int8-serialized checkpoint for now."
+                )
+            if len(weight_block_size) != 2:
+                raise ValueError(
+                    f"The quantization block size of weight must have 2 dimensions, but got {len(weight_block_size)} dimensions."
+                )
+            if activation_scheme != "dynamic":
+                raise ValueError(
+                    f"The block-wise quantization only supports dynamic activation scheme for now, but got {activation_scheme} activation scheme."
+                )
+        self.weight_block_size = weight_block_size
+
+    @classmethod
+    def get_name(cls) -> str:
+        return "blockwise_int8"
+
+    @classmethod
+    def get_supported_act_dtypes(cls) -> List[torch.dtype]:
+        return [torch.bfloat16, torch.half]
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 80
+
+    @classmethod
+    def get_config_filenames(cls) -> List[str]:
+        return []
+
+    @classmethod
+    def from_config(cls, config: Dict[str, Any]) -> BlockInt8Config:
+        quant_method = cls.get_from_keys(config, ["quant_method"])
+        is_checkpoint_int8_serialized = "int8" in quant_method
+        activation_scheme = cls.get_from_keys(config, ["activation_scheme"])
+        ignored_layers = cls.get_from_keys_or(config, ["ignored_layers"], None)
+        weight_block_size = cls.get_from_keys_or(config, ["weight_block_size"], None)
+        return cls(
+            is_checkpoint_int8_serialized=is_checkpoint_int8_serialized,
+            activation_scheme=activation_scheme,
+            ignored_layers=ignored_layers,
+            weight_block_size=weight_block_size,
+        )
+
+    def get_quant_method(
+        self, layer: torch.nn.Module, prefix: str
+    ) -> Optional[QuantizeMethodBase]:
+        from sglang.srt.layers.linear import LinearBase
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+
+        if isinstance(layer, LinearBase):
+            if is_layer_skipped(prefix, self.ignored_layers):
+                return UnquantizedLinearMethod()
+            return BlockInt8LinearMethod(self)
+        elif isinstance(layer, FusedMoE):
+            return BlockInt8MoEMethod(self)
+        return None
+
+    def get_scaled_act_names(self) -> List[str]:
+        return []
+
+
+class BlockInt8LinearMethod(LinearMethodBase):
+    """Linear method for INT8.
+    Supports loading INT8 checkpoints with static weight scale and
+    dynamic activation scale.
+
+    Limitations:
+    Only support block-wise int8 quantization and int8 checkpoint
+
+    Args:
+        quant_config: The quantization config.
+    """
+
+    def __init__(self, quant_config: BlockInt8Config):
+        self.quant_config = quant_config
+        assert self.quant_config.weight_block_size is not None
+        assert self.quant_config.is_checkpoint_int8_serialized
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        input_size_per_partition: int,
+        output_partition_sizes: List[int],
+        input_size: int,
+        output_size: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        output_size_per_partition = sum(output_partition_sizes)
+        weight_loader = extra_weight_attrs.get("weight_loader")
+
+        tp_size = get_parallel().tp_size
+
+        block_n, block_k = (
+            self.quant_config.weight_block_size[0],
+            self.quant_config.weight_block_size[1],
+        )
+        # Required by row parallel
+        if tp_size > 1 and input_size // input_size_per_partition == tp_size:
+            if input_size_per_partition % block_k != 0:
+                raise ValueError(
+                    f"Weight input_size_per_partition = "
+                    f"{input_size_per_partition} is not divisible by "
+                    f"weight quantization block_k = {block_k}."
+                )
+        # Required by column parallel or enabling merged weights
+        if (tp_size > 1 and output_size // output_size_per_partition == tp_size) or len(
+            output_partition_sizes
+        ) > 1:
+            for output_partition_size in output_partition_sizes:
+                if output_partition_size % block_n != 0:
+                    raise ValueError(
+                        f"Weight output_partition_size = "
+                        f"{output_partition_size} is not divisible by "
+                        f"weight quantization block_n = {block_n}."
+                    )
+
+        layer.logical_widths = output_partition_sizes
+
+        layer.input_size_per_partition = input_size_per_partition
+        layer.output_size_per_partition = output_size_per_partition
+        layer.orig_dtype = params_dtype
+
+        # WEIGHT
+        weight_dtype = (
+            torch.int8
+            if self.quant_config.is_checkpoint_int8_serialized
+            else params_dtype
+        )
+
+        weight = ModelWeightParameter(
+            data=torch.empty(
+                output_size_per_partition, input_size_per_partition, dtype=weight_dtype
+            ),
+            input_dim=1,
+            output_dim=0,
+            weight_loader=weight_loader,
+        )
+        layer.register_parameter("weight", weight)
+
+        # WEIGHT SCALE
+
+        scale = BlockQuantScaleParameter(
+            data=torch.empty(
+                (output_size_per_partition + block_n - 1) // block_n,
+                (input_size_per_partition + block_k - 1) // block_k,
+                dtype=torch.float32,
+            ),
+            input_dim=1,
+            output_dim=0,
+            weight_loader=weight_loader,
+        )
+        scale[:] = torch.finfo(torch.float32).min
+        layer.register_parameter("weight_scale_inv", scale)
+
+        # INPUT ACTIVATION SCALE
+        assert self.quant_config.activation_scheme == "dynamic"
+        layer.register_parameter("input_scale", None)
+
+    def process_weights_after_loading(self, layer: Module) -> None:
+        # Block quant doesn't need to process weights after loading
+        # Use torch Parameter to avoid cuda graph capturing issue
+        layer.weight = torch.nn.Parameter(layer.weight.data, requires_grad=False)
+        layer.weight_scale_inv = torch.nn.Parameter(
+            layer.weight_scale_inv.data, requires_grad=False
+        )
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        return apply_w8a8_block_int8_linear(
+            input=x,
+            weight=layer.weight,
+            block_size=self.quant_config.weight_block_size,
+            weight_scale=layer.weight_scale_inv,
+            input_scale=None,
+            bias=bias,
+        )
+
+
+class BlockInt8MoEMethod(FusedMoEMethodBase):
+    """MoE method for INT8.
+    Supports loading INT8 checkpoints with static weight scale and
+    dynamic activation scale.
+
+    Limitations:
+    Only support block-wise int8 quantization and int8 checkpoint
+
+    Args:
+        quant_config: The quantization config.
+    """
+
+    def __init__(self, quant_config: BlockInt8Config):
+        self.quant_config = quant_config
+        assert self.quant_config.weight_block_size is not None
+        assert self.quant_config.is_checkpoint_int8_serialized
+
+    def create_weights(
+        self,
+        layer: Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoeWeightScaleSupported
+
+        if self.quant_config.is_checkpoint_int8_serialized:
+            params_dtype = torch.int8
+        tp_size = get_parallel().tp_size
+
+        block_n, block_k = (
+            self.quant_config.weight_block_size[0],
+            self.quant_config.weight_block_size[1],
+        )
+        # NOTE(HandH1998): To ensure proper alignment of the block-wise quantization scales, the output_size of the weights for both the gate and up layers must be divisible by block_n.
+        # Required by column parallel or enabling merged weights
+        if intermediate_size_per_partition % block_n != 0:
+            raise ValueError(
+                f"The output_size of gate's and up's weight = "
+                f"{intermediate_size_per_partition} is not divisible by "
+                f"weight quantization block_n = {block_n}."
+            )
+        if tp_size > 1:
+            # Required by row parallel
+            if intermediate_size_per_partition % block_k != 0:
+                raise ValueError(
+                    f"The input_size of down's weight = "
+                    f"{intermediate_size_per_partition} is not divisible by "
+                    f"weight quantization block_k = {block_k}."
+                )
+
+        # WEIGHTS
+        w13_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                2 * intermediate_size_per_partition,
+                hidden_size,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight", w13_weight)
+        set_weight_attrs(w13_weight, extra_weight_attrs)
+
+        w2_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                intermediate_size_per_partition,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight", w2_weight)
+        set_weight_attrs(w2_weight, extra_weight_attrs)
+
+        # WEIGHT_SCALES
+        w13_weight_scale = torch.nn.Parameter(
+            torch.ones(
+                num_experts,
+                2 * ((intermediate_size_per_partition + block_n - 1) // block_n),
+                (hidden_size + block_k - 1) // block_k,
+                dtype=torch.float32,
+            ),
+            requires_grad=False,
+        )
+        w2_weight_scale = torch.nn.Parameter(
+            torch.ones(
+                num_experts,
+                (hidden_size + block_n - 1) // block_n,
+                (intermediate_size_per_partition + block_k - 1) // block_k,
+                dtype=torch.float32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_scale_inv", w13_weight_scale)
+        layer.register_parameter("w2_weight_scale_inv", w2_weight_scale)
+
+        extra_weight_attrs.update(
+            {"quant_method": FusedMoeWeightScaleSupported.BLOCK.value}
+        )
+        set_weight_attrs(w13_weight_scale, extra_weight_attrs)
+        set_weight_attrs(w2_weight_scale, extra_weight_attrs)
+
+        # INPUT_SCALES
+        assert self.quant_config.activation_scheme == "dynamic"
+        layer.w13_input_scale = None
+        layer.w2_input_scale = None
+
+    def process_weights_after_loading(self, layer: Module) -> None:
+        # Block quant doesn't need to process weights after loading
+        return
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+        self.runner = MoeRunner(MoeRunnerBackend.TRITON, moe_runner_config)
+
+    def get_triton_quant_info(self, layer: torch.nn.Module) -> TritonMoeQuantInfo:
+        return TritonMoeQuantInfo(
+            w13_weight=layer.w13_weight,
+            w2_weight=layer.w2_weight,
+            use_int8_w8a8=True,
+            w13_scale=layer.w13_weight_scale_inv,
+            w2_scale=layer.w2_weight_scale_inv,
+            a13_scale=layer.w13_input_scale,
+            a2_scale=layer.w2_input_scale,
+            block_shape=self.quant_config.weight_block_size,
+        )
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+
+        quant_info = self.get_triton_quant_info(layer)
+
+        return self.runner.run(dispatch_output, quant_info)

python/sglang/srt/layers/quantization/compressed_tensors/README.md

@@ -0,0 +1,6 @@
+# quantization compressed_tensors module
+
+To support compressed_tensors format quantization models, we adapted https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/quantization/compressed_tensors into SGLang.
+
+
+For practical purposes, we have only applied the compressed_tensors format of `w8a8_fp8`. If you have requirements for other formats, you can submit an issue through this [link](https://github.com/sgl-project/sglang/issues).

python/sglang/srt/layers/quantization/compressed_tensors/schemes/__init__.py

@@ -0,0 +1,44 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from .compressed_tensors_scheme import (
+    CompressedTensorsLinearScheme,
+    CompressedTensorsMoEScheme,
+)
+from .compressed_tensors_w4a4_mxint4_moe import CompressedTensorsMxInt4MoE
+from .compressed_tensors_w4a4_nvfp4 import CompressedTensorsW4A4Fp4
+from .compressed_tensors_w4a4_nvfp4_moe import CompressedTensorsW4A4Nvfp4MoE
+from .compressed_tensors_w4a8_int8_moe import NPUCompressedTensorsW4A8Int8DynamicMoE
+from .compressed_tensors_w8a8_fp8 import CompressedTensorsW8A8Fp8
+from .compressed_tensors_w8a8_fp8_moe import CompressedTensorsW8A8Fp8MoE
+from .compressed_tensors_w8a8_int8 import (
+    CompressedTensorsW8A8Int8,
+    NPUCompressedTensorsW8A8Int8,
+)
+from .compressed_tensors_w8a8_int8_moe import NPUCompressedTensorsW8A8Int8DynamicMoE
+from .compressed_tensors_w8a16_fp8 import CompressedTensorsW8A16Fp8
+from .compressed_tensors_wNa16 import WNA16_SUPPORTED_BITS, CompressedTensorsWNA16
+from .compressed_tensors_wNa16_moe import (
+    CompressedTensorsWNA16MoE,
+    CompressedTensorsWNA16TritonMoE,
+    NPUCompressedTensorsW4A16Int4DynamicMoE,
+)
+
+__all__ = [
+    "CompressedTensorsLinearScheme",
+    "CompressedTensorsMoEScheme",
+    "CompressedTensorsW8A8Fp8",
+    "CompressedTensorsW8A8Fp8MoE",
+    "CompressedTensorsW8A16Fp8",
+    "CompressedTensorsW8A8Int8",
+    "NPUCompressedTensorsW8A8Int8",
+    "NPUCompressedTensorsW8A8Int8DynamicMoE",
+    "CompressedTensorsWNA16",
+    "CompressedTensorsWNA16MoE",
+    "CompressedTensorsWNA16TritonMoE",
+    "NPUCompressedTensorsW4A16Int4DynamicMoE",
+    "WNA16_SUPPORTED_BITS",
+    "CompressedTensorsW4A4Fp4",
+    "CompressedTensorsW4A4Nvfp4MoE",
+    "NPUCompressedTensorsW4A8Int8DynamicMoE",
+    "CompressedTensorsMxInt4MoE",
+]

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_scheme.py

@@ -0,0 +1,116 @@
+# Adapted from https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/quantization/compressed_tensors
+# SPDX-License-Identifier: Apache-2.0
+
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from abc import abstractmethod
+from typing import TYPE_CHECKING, Optional
+
+import torch
+
+from sglang.srt.layers.moe import MoeRunnerConfig
+from sglang.srt.layers.quantization.base_scheme import BaseLinearScheme, BaseMoEScheme
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.token_dispatcher import StandardDispatchOutput
+
+__all__ = ["CompressedTensorsLinearScheme", "CompressedTensorsMoEScheme"]
+
+
+class CompressedTensorsLinearScheme(BaseLinearScheme):
+    """
+    Abstract class used to describe the weight creation and forward pass
+    of different quantization schemes supported by CompressedTensors.
+    """
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        """
+        Get minimum device capability.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def create_weights(self, *args, **kwargs):
+        """
+        Weight creation for the particular scheme. Inputs to this function
+
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def apply_weights(
+        self, layer: torch.nn.Module, x: torch.Tensor, bias: Optional[torch.Tensor]
+    ):
+        """
+        Run the forward pass for the particular scheme. This is where
+        scheme-specific dequant/quant steps/kernels should be applied.
+
+        :param layer: torch.nn.Module with the registered weights and
+            other parameters relevant to the particular scheme.
+        :param x: input to the layer
+        :param bias: bias parameter
+
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def process_weights_after_loading(self, layer: torch.nn.Module):
+        """
+        Called after weight loading is complete for any cleanup that
+        needs to occur.
+        """
+        raise NotImplementedError
+
+
+class CompressedTensorsMoEScheme(BaseMoEScheme):
+    """
+    Abstract class used to describe the weight creation and forward pass
+    of different quantization schemes supported by CompressedTensors.
+    """
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        """
+        Get minimum device capability.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def create_weights(self, *args, **kwargs):
+        """
+        Weight creation for the particular scheme. Inputs to this function
+
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        raise NotImplementedError
+
+    @abstractmethod
+    def process_weights_after_loading(self, layer: torch.nn.Module):
+        """
+        Called after weight loading is complete for any cleanup that
+        needs to occur.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: "StandardDispatchOutput",
+    ):
+        """
+        Run the forward pass for the particular scheme. This is where
+        scheme-specific dequant/quant steps/kernels should be applied.
+
+        :param layer: torch.nn.Module with the registered weights and
+            other parameters relevant to the particular scheme.
+        :param x: input to the layer
+        :param bias: bias parameter
+
+        """
+        raise NotImplementedError

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_mxint4_moe.py

@@ -0,0 +1,364 @@
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING
+
+import torch
+from compressed_tensors import CompressionFormat
+
+from sglang.srt.distributed import get_tp_group
+from sglang.srt.distributed.device_communicators.pynccl_allocator import (
+    use_symmetric_memory,
+)
+from sglang.srt.layers.dp_attention import is_allocation_symmetric
+from sglang.srt.layers.moe import MoeRunnerConfig
+from sglang.srt.layers.moe.utils import RoutingMethodType, get_moe_runner_backend
+from sglang.srt.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsMoEScheme,
+)
+from sglang.srt.layers.quantization.utils import replace_parameter
+from sglang.srt.runtime_context import get_parallel
+from sglang.srt.utils import is_flashinfer_available, next_power_of_2, set_weight_attrs
+
+logger = logging.getLogger(__name__)
+
+__all__ = ["CompressedTensorsMxInt4MoE"]
+
+if TYPE_CHECKING:
+    from compressed_tensors.quantization import QuantizationArgs
+
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
+    from sglang.srt.layers.quantization.compressed_tensors.compressed_tensors import (
+        CompressedTensorsConfig,
+    )
+
+if is_flashinfer_available():
+    from flashinfer.fp4_quantization import block_scale_interleave
+    from flashinfer.fused_moe import (
+        convert_to_block_layout,
+        trtllm_mxint4_block_scale_moe,
+    )
+    from flashinfer.fused_moe.core import (
+        _maybe_get_cached_w3_w1_permute_indices,
+        get_w2_permute_indices_with_cache,
+    )
+
+
+class CompressedTensorsMxInt4MoE(CompressedTensorsMoEScheme):
+    def __init__(
+        self, quant_config: CompressedTensorsConfig, weight_quant: QuantizationArgs
+    ):
+        self.quant_config = quant_config
+        # Per-layer scheme already resolved by get_moe_scheme(); reuse it directly
+        # (mixed-precision MoE has no "Linear" config group to fall back on).
+        config = weight_quant
+        self.num_bits = config.num_bits
+        self.packed_factor = 32 // config.num_bits
+        self.strategy = config.strategy
+        self.group_size = config.group_size
+        self.actorder = config.actorder
+        assert (
+            config.strategy == "group"
+            and config.group_size == 32
+            and config.num_bits == 4
+        ), "MxInt4 only supports group strategy with group size 32"
+        assert config.symmetric, "Only symmetric quantization is supported for MoE"
+        assert (
+            get_moe_runner_backend().is_flashinfer_trtllm()
+        ), "MxInt4 only supports flashinfer_trtllm backend"
+        assert (
+            not config.actorder
+        ), "Actorder is not supported by flashinfer_trtllm backend"
+        self.moe_ep_rank = get_parallel().moe_ep_rank
+
+        if self.quant_config.quant_format != CompressionFormat.pack_quantized.value:
+            raise ValueError(
+                f"For Fused MoE layers, only {CompressionFormat.pack_quantized.value} "
+                "is supported for the mxint4"
+            )
+        self._cache_permute_indices = {}
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # Requires sm100(blackwell) architecture
+        return 100
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        assert (
+            params_dtype == torch.bfloat16
+        ), f"Params dtype should be torch.bfloat16, but got: {params_dtype}"
+
+        extra_weight_attrs.update({"quant_method": self.strategy})
+        w13_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                2 * intermediate_size_per_partition,
+                hidden_size // self.packed_factor,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_packed", w13_weight)
+        set_weight_attrs(w13_weight, extra_weight_attrs)
+
+        w2_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                intermediate_size_per_partition // self.packed_factor,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_packed", w2_weight)
+        set_weight_attrs(w2_weight, extra_weight_attrs)
+
+        w2_scales_size = intermediate_size_per_partition
+        num_groups_w2 = w2_scales_size // self.group_size
+        num_groups_w13 = hidden_size // self.group_size
+
+        w13_scale = torch.nn.Parameter(
+            torch.ones(
+                num_experts,
+                2 * intermediate_size_per_partition,
+                num_groups_w13,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_scale", w13_scale)
+        set_weight_attrs(w13_scale, extra_weight_attrs)
+
+        w2_scale = torch.nn.Parameter(
+            torch.ones(num_experts, hidden_size, num_groups_w2, dtype=params_dtype),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_scale", w2_scale)
+        set_weight_attrs(w2_scale, extra_weight_attrs)
+
+        w13_weight_shape = torch.nn.Parameter(
+            torch.empty(num_experts, 2), requires_grad=False
+        )
+
+        layer.register_parameter("w13_weight_shape", w13_weight_shape)
+        set_weight_attrs(w13_weight_shape, extra_weight_attrs)
+
+        w2_weight_shape = torch.nn.Parameter(
+            torch.empty(num_experts, 2), requires_grad=False
+        )
+        layer.register_parameter("w2_weight_shape", w2_weight_shape)
+        set_weight_attrs(w2_weight_shape, extra_weight_attrs)
+
+        layer.a13_scale = None
+        layer.a2_scale = None
+
+    # Adapted from https://github.com/flashinfer-ai/flashinfer/blob/main/tests/moe/test_trtllm_gen_fused_moe.py
+    def prepare_static_weights_for_kernel(
+        self,
+        gemm1_weights,
+        gemm2_weights,
+        gemm1_scales,
+        gemm2_scales,
+        num_experts,
+    ):
+        """Prepare quantized weights for kernel (done offline with weights)."""
+
+        epilogue_tile_m = 128
+        gemm1_weights_mxint4_shuffled = []
+        gemm1_scales_shuffled = []
+        gemm2_weights_mxint4_shuffled = []
+        gemm2_scales_shuffled = []
+
+        def repack(w):
+            assert w.dim() == 2 and w.dtype == torch.int32
+            shifts = torch.arange(0, 32, 4, dtype=torch.int32, device=w.device)
+            w = (w.unsqueeze(2) >> shifts) & 0x0F
+            w = (w - 8).to(torch.int8).reshape(w.shape[0], -1, 2)
+            w = (w[..., 0] & 0x0F) | ((w[..., 1] & 0x0F) << 4)
+            w = w.to(torch.uint8)
+            return w
+
+        for i in range(num_experts):
+            # NOTE(HandH1998):
+            # the huggingface weight format follows (w/s + 8) to pack,
+            # however, trtllm requires (w/s) to pack
+            # we need to convert the weight to trtllm's format first
+            cur_expert_gemm1_weight = repack(gemm1_weights[i])
+            cur_expert_gemm2_weight = repack(gemm2_weights[i])
+
+            # Calculate the permute indices for the following:
+            # 1. Reorder rows of W1 and scales for fused gated activation
+            # 2. Shuffle weights and scaling factors for transposed mma output
+            # for both w3_w1 and w2 weights and scale factors
+            permute_indices = _maybe_get_cached_w3_w1_permute_indices(
+                self._cache_permute_indices,
+                cur_expert_gemm1_weight,
+                epilogue_tile_m,
+            )
+            gemm1_weights_shuffled = cur_expert_gemm1_weight[
+                permute_indices.to(gemm1_weights.device)
+            ].contiguous()
+            permute_sf_indices = _maybe_get_cached_w3_w1_permute_indices(
+                self._cache_permute_indices,
+                gemm1_scales[i].to(torch.bfloat16),
+                epilogue_tile_m,
+                num_elts_per_sf=32,
+            )
+            gemm1_scales_shuffled.append(
+                block_scale_interleave(
+                    gemm1_scales[i]
+                    .to(torch.bfloat16)[permute_sf_indices.to(gemm1_scales.device)]
+                    .contiguous()
+                )
+            )
+
+            permute_indices = get_w2_permute_indices_with_cache(
+                self._cache_permute_indices,
+                cur_expert_gemm2_weight,
+                epilogue_tile_m,
+            )
+            gemm2_weights_shuffled = cur_expert_gemm2_weight[
+                permute_indices.to(gemm2_weights.device)
+            ].contiguous()
+
+            permute_sf_indices = get_w2_permute_indices_with_cache(
+                self._cache_permute_indices,
+                gemm2_scales[i].to(torch.bfloat16),
+                epilogue_tile_m,
+                num_elts_per_sf=16,
+            )
+            gemm2_scales_shuffled.append(
+                block_scale_interleave(
+                    gemm2_scales[i]
+                    .to(torch.bfloat16)[permute_sf_indices.to(gemm2_scales.device)]
+                    .contiguous()
+                )
+            )
+
+            block_k = 128
+            gemm1_weights_shuffled = convert_to_block_layout(
+                gemm1_weights_shuffled.view(torch.uint8), block_k
+            )
+            gemm2_weights_shuffled = convert_to_block_layout(
+                gemm2_weights_shuffled.view(torch.uint8), block_k
+            )
+
+            gemm1_weights_mxint4_shuffled.append(gemm1_weights_shuffled)
+            gemm2_weights_mxint4_shuffled.append(gemm2_weights_shuffled)
+
+        gemm1_weights_mxint4_shuffled = torch.stack(gemm1_weights_mxint4_shuffled)
+        gemm2_weights_mxint4_shuffled = torch.stack(gemm2_weights_mxint4_shuffled)
+        gemm1_scales_shuffled = torch.stack(gemm1_scales_shuffled).view(torch.bfloat16)
+        gemm2_scales_shuffled = torch.stack(gemm2_scales_shuffled).view(torch.bfloat16)
+
+        return (
+            gemm1_weights_mxint4_shuffled,
+            gemm1_scales_shuffled,
+            gemm2_weights_mxint4_shuffled,
+            gemm2_scales_shuffled,
+        )
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+
+        num_experts = layer.w13_weight_packed.shape[0]
+        (
+            gemm1_weights_mxint4_shuffled,
+            gemm1_scales_shuffled,
+            gemm2_weights_mxint4_shuffled,
+            gemm2_scales_shuffled,
+        ) = self.prepare_static_weights_for_kernel(
+            layer.w13_weight_packed,
+            layer.w2_weight_packed,
+            layer.w13_weight_scale,
+            layer.w2_weight_scale,
+            num_experts=num_experts,
+        )
+        replace_parameter(layer, "w13_weight_packed", gemm1_weights_mxint4_shuffled)
+        replace_parameter(layer, "w2_weight_packed", gemm2_weights_mxint4_shuffled)
+        replace_parameter(layer, "w13_weight_scale", gemm1_scales_shuffled)
+        replace_parameter(layer, "w2_weight_scale", gemm2_scales_shuffled)
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+        from sglang.srt.layers.moe.token_dispatcher import StandardCombineInput
+
+        assert (
+            self.moe_runner_config.is_gated
+        ), "Only gated MoEs are supported for flashinfer mxint4"
+
+        x = dispatch_output.hidden_states
+        topk_output = dispatch_output.topk_output
+
+        router_logits = topk_output.router_logits
+        topk_config = topk_output.topk_config
+        correction_bias = (
+            None
+            if topk_config.correction_bias is None
+            else topk_config.correction_bias.to(x.dtype)
+        )
+
+        local_num_experts = self.moe_runner_config.num_local_experts
+        routing_method_type = layer.routing_method_type
+        assert routing_method_type is not None
+        # DeepSeekV3 style routing requires float32 router logits,
+        # see this PR for details: https://github.com/flashinfer-ai/flashinfer/commit/d84e1d560da0a27961c19ca788d96c19cb9dcfb6
+        if routing_method_type == RoutingMethodType.DeepSeekV3:
+            router_logits = router_logits.to(torch.float32)
+        routed_scaling_factor = self.moe_runner_config.routed_scaling_factor
+        routed_scaling_factor = (
+            routed_scaling_factor if routed_scaling_factor is not None else 1.0
+        )
+
+        with use_symmetric_memory(
+            get_tp_group(), disabled=not is_allocation_symmetric()
+        ):
+            num_tokens = x.shape[0]
+            hidden_size = x.shape[-1]
+            symm_output = torch.empty(
+                num_tokens, hidden_size, dtype=torch.bfloat16, device=x.device
+            )
+
+        trtllm_mxint4_block_scale_moe(
+            routing_logits=router_logits,  # float
+            routing_bias=correction_bias,
+            hidden_states=x,
+            gemm1_weights=layer.w13_weight_packed,
+            gemm1_weights_scale=layer.w13_weight_scale,
+            gemm1_alpha=self.moe_runner_config.gemm1_alpha,
+            gemm1_beta=None,
+            gemm1_clamp_limit=self.moe_runner_config.gemm1_clamp_limit,
+            gemm2_weights=layer.w2_weight_packed,
+            gemm2_weights_scale=layer.w2_weight_scale,
+            num_experts=self.moe_runner_config.num_experts,
+            top_k=topk_config.top_k,
+            n_group=topk_config.num_expert_group,
+            topk_group=topk_config.topk_group,
+            intermediate_size=self.moe_runner_config.intermediate_size_per_partition,
+            local_expert_offset=self.moe_ep_rank * local_num_experts,
+            local_num_experts=local_num_experts,
+            routed_scaling_factor=routed_scaling_factor,
+            routing_method_type=routing_method_type,
+            tune_max_num_tokens=next_power_of_2(x.shape[0]),
+            output=symm_output,
+        )
+
+        return StandardCombineInput(hidden_states=symm_output)

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_nvfp4.py

@@ -0,0 +1,172 @@
+# Adapted from https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/quantization/compressed_tensors
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import logging
+from collections.abc import Callable
+from typing import Optional
+
+import torch
+from torch.nn.parameter import Parameter
+
+from sglang.srt.layers.parameter import (
+    GroupQuantScaleParameter,
+    ModelWeightParameter,
+    PerTensorScaleParameter,
+)
+from sglang.srt.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsLinearScheme,
+)
+from sglang.srt.layers.quantization.fp4_utils import get_fp4_gemm_runner_backend
+from sglang.srt.layers.quantization.modelopt_quant import (
+    enable_flashinfer_fp4_gemm,
+    fp4_gemm,
+    fp4_quantize,
+)
+from sglang.srt.layers.quantization.utils import swizzle_blockscale
+
+logger = logging.getLogger(__name__)
+
+__all__ = ["CompressedTensorsW4A4Fp4"]
+
+
+class CompressedTensorsW4A4Fp4(CompressedTensorsLinearScheme):
+    def __init__(self):
+        self.group_size = 16
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return 100
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        output_partition_sizes: list[int],
+        input_size_per_partition: int,
+        params_dtype: torch.dtype,
+        weight_loader: Callable,
+        **kwargs,
+    ):
+        output_size_per_partition = sum(output_partition_sizes)
+        layer.logical_widths = output_partition_sizes
+        layer.input_size_per_partition = input_size_per_partition
+        layer.output_size_per_partition = output_size_per_partition
+
+        # Weight
+        weight = ModelWeightParameter(
+            data=torch.empty(
+                sum(output_partition_sizes),
+                input_size_per_partition // 2,
+                dtype=torch.uint8,
+            ),
+            input_dim=1,
+            output_dim=0,
+            weight_loader=weight_loader,
+        )
+        layer.register_parameter("weight_packed", weight)
+
+        # Global Weight Scale
+        weight_global_scale = PerTensorScaleParameter(
+            data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
+            weight_loader=weight_loader,
+        )
+        layer.register_parameter("weight_global_scale", weight_global_scale)
+
+        # Per Group Weight Scale
+        weight_scale = GroupQuantScaleParameter(
+            data=torch.empty(
+                sum(output_partition_sizes),
+                input_size_per_partition // self.group_size,
+                dtype=torch.float8_e4m3fn,
+            ),
+            input_dim=1,
+            output_dim=0,
+            weight_loader=weight_loader,
+        )
+
+        layer.register_parameter("weight_scale", weight_scale)
+
+        input_global_scale = PerTensorScaleParameter(
+            data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
+            weight_loader=weight_loader,
+        )
+        layer.register_parameter("input_global_scale", input_global_scale)
+
+    def process_weights_after_loading(self, layer) -> None:
+        global_input_scale = layer.input_global_scale.max().to(torch.float32)
+        layer.input_global_scale = Parameter(global_input_scale, requires_grad=False)
+
+        layer.weight_global_scale = Parameter(
+            layer.weight_global_scale.max().to(torch.float32), requires_grad=False
+        )
+
+        if get_fp4_gemm_runner_backend().is_flashinfer_trtllm():
+            # FlashInfer TRTLLM FP4 GEMM requires a different weight layout.
+            # FlashInfer provides nvfp4_quantize to quantize + shuffle the
+            # layout but we use our own quantization so we have to call
+            # shuffles ourselves.
+            from flashinfer import shuffle_matrix_a, shuffle_matrix_sf_a
+
+            weight = layer.weight_packed.data
+            weight_scale = layer.weight_scale.data
+
+            epilogue_tile_m = 128
+            weight = shuffle_matrix_a(weight.view(torch.uint8), epilogue_tile_m)
+            weight_scale = (
+                shuffle_matrix_sf_a(weight_scale.view(torch.uint8), epilogue_tile_m)
+                .reshape(weight_scale.shape)
+                .view(torch.float8_e4m3fn)
+            )
+
+            layer.weight_scale = Parameter(weight_scale, requires_grad=False)
+            layer.weight_packed = Parameter(weight, requires_grad=False)
+        else:
+            swizzled_weight_scale = swizzle_blockscale(layer.weight_scale)
+            layer.weight_scale = Parameter(swizzled_weight_scale, requires_grad=False)
+            layer.weight_packed = Parameter(
+                layer.weight_packed.data, requires_grad=False
+            )
+
+        layer.alpha = Parameter(
+            1 / (layer.input_global_scale * layer.weight_global_scale),
+            requires_grad=False,
+        )
+
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        output_dtype = x.dtype
+        w_n, _ = layer.weight_packed.shape
+        output_shape = [x.shape[0], w_n]
+
+        # quantize BF16 or FP16 to (FP4 and interleaved block scale)
+        x_fp4, x_blockscale = fp4_quantize(x, layer.input_global_scale)
+
+        assert x_fp4.dtype == torch.uint8
+        assert layer.weight_packed.dtype == torch.uint8
+        assert layer.weight_scale.dtype == torch.float8_e4m3fn
+        assert layer.alpha.dtype == torch.float32
+
+        w = layer.weight_packed
+        w_blockscale = layer.weight_scale
+        if (
+            enable_flashinfer_fp4_gemm
+            and not get_fp4_gemm_runner_backend().is_cutlass()
+        ):
+            w = layer.weight_packed.T
+            w_blockscale = layer.weight_scale.T
+
+        out = fp4_gemm(
+            x_fp4,
+            w,
+            x_blockscale,
+            w_blockscale,
+            layer.alpha,
+            output_dtype,
+            w_n,
+        )
+        if bias is not None:
+            out = out + bias
+        return out.view(*output_shape)

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_nvfp4_moe.py

@@ -0,0 +1,408 @@
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING
+
+import torch
+
+from sglang.srt.distributed import get_tp_group
+from sglang.srt.distributed.device_communicators.pynccl_allocator import (
+    use_symmetric_memory,
+)
+from sglang.srt.layers.dp_attention import is_allocation_symmetric
+from sglang.srt.layers.moe import MoeRunner, MoeRunnerBackend, MoeRunnerConfig
+from sglang.srt.layers.moe.cutlass_moe_params import CutlassMoEParams, CutlassMoEType
+from sglang.srt.layers.moe.utils import RoutingMethodType, get_moe_runner_backend
+from sglang.srt.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsMoEScheme,
+)
+from sglang.srt.layers.quantization.fp8_utils import is_blackwell_supported
+from sglang.srt.layers.quantization.utils import (
+    prepare_static_weights_for_trtllm_fp4_moe,
+    reorder_w1w3_to_w3w1,
+    replace_parameter,
+    swizzle_blockscale,
+)
+from sglang.srt.utils import next_power_of_2, set_weight_attrs
+
+logger = logging.getLogger(__name__)
+
+__all__ = ["CompressedTensorsW4A4Nvfp4MoE"]
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
+
+
+class CompressedTensorsW4A4Nvfp4MoE(CompressedTensorsMoEScheme):
+
+    def __init__(self):
+        if not is_blackwell_supported():
+            raise ValueError(
+                "Current platform does not support NVFP4"
+                " quantization. Please use Blackwell and"
+                " above."
+            )
+        self.group_size = 16
+        self.use_flashinfer_trtllm = get_moe_runner_backend().is_flashinfer_trtllm()
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # Requires sm100(blackwell) architecture
+        return 100
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoeWeightScaleSupported
+
+        layer.params_dtype = params_dtype
+
+        w13_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                2 * intermediate_size_per_partition,
+                # 2 fp4 items are packed in the input dimension
+                hidden_size // 2,
+                requires_grad=False,
+                dtype=torch.uint8,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_packed", w13_weight)
+        set_weight_attrs(w13_weight, extra_weight_attrs)
+
+        w2_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                # 2 fp4 items are packed in the input dimension
+                intermediate_size_per_partition // 2,
+                dtype=torch.uint8,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_packed", w2_weight)
+        set_weight_attrs(w2_weight, extra_weight_attrs)
+
+        # Weight Scales
+        w13_weight_scale = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                2 * intermediate_size_per_partition,
+                # 2 fp4 items are packed in the input dimension
+                hidden_size // self.group_size,
+                dtype=torch.float8_e4m3fn,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_scale", w13_weight_scale)
+        extra_weight_attrs.update(
+            {"quant_method": FusedMoeWeightScaleSupported.GROUP.value}
+        )
+        set_weight_attrs(w13_weight_scale, extra_weight_attrs)
+
+        w2_weight_scale = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                # 2 fp4 items are packed in the input dimension
+                intermediate_size_per_partition // self.group_size,
+                dtype=torch.float8_e4m3fn,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_scale", w2_weight_scale)
+        extra_weight_attrs.update(
+            {"quant_method": FusedMoeWeightScaleSupported.GROUP.value}
+        )
+        set_weight_attrs(w2_weight_scale, extra_weight_attrs)
+
+        # Weight Global Scales
+        w13_weight_scale_2 = torch.nn.Parameter(
+            torch.empty(num_experts, 2, dtype=torch.float32), requires_grad=False
+        )
+        layer.register_parameter("w13_weight_global_scale", w13_weight_scale_2)
+        extra_weight_attrs.update(
+            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
+        )
+        set_weight_attrs(w13_weight_scale_2, extra_weight_attrs)
+
+        w2_weight_scale_2 = torch.nn.Parameter(
+            torch.empty(num_experts, dtype=torch.float32), requires_grad=False
+        )
+        layer.register_parameter("w2_weight_global_scale", w2_weight_scale_2)
+        extra_weight_attrs.update(
+            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
+        )
+        set_weight_attrs(w2_weight_scale_2, extra_weight_attrs)
+
+        # Input Global Scales
+        w13_input_scale = torch.nn.Parameter(
+            torch.empty(num_experts, 2, dtype=torch.float32), requires_grad=False
+        )
+        layer.register_parameter("w13_input_global_scale", w13_input_scale)
+        extra_weight_attrs.update(
+            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
+        )
+        set_weight_attrs(w13_input_scale, extra_weight_attrs)
+
+        w2_input_scale = torch.nn.Parameter(
+            torch.empty(num_experts, dtype=torch.float32), requires_grad=False
+        )
+        layer.register_parameter("w2_input_global_scale", w2_input_scale)
+        extra_weight_attrs.update(
+            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
+        )
+        set_weight_attrs(w2_input_scale, extra_weight_attrs)
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        # From packed to weight
+        layer.w13_weight = torch.nn.Parameter(
+            layer.w13_weight_packed.data, requires_grad=False
+        )
+        delattr(layer, "w13_weight_packed")
+
+        layer.w2_weight = torch.nn.Parameter(
+            layer.w2_weight_packed.data, requires_grad=False
+        )
+        delattr(layer, "w2_weight_packed")
+
+        if self.use_flashinfer_trtllm:
+            w, s = reorder_w1w3_to_w3w1(
+                layer.w13_weight.data, layer.w13_weight_scale.data, dim=-2
+            )
+            layer.w13_weight = torch.nn.Parameter(w, requires_grad=False)
+            layer.w13_weight_scale = torch.nn.Parameter(s, requires_grad=False)
+
+        if not torch.allclose(
+            layer.w13_weight_global_scale[:, 0], layer.w13_weight_global_scale[:, 1]
+        ):
+            logger.warning_once(
+                "w1_weight_global_scale must match w3_weight_global_scale. "
+                "Accuracy may be affected."
+            )
+
+        # Take inverse of global scale saved to disk
+        layer.w13_weight_scale_2 = torch.nn.Parameter(
+            1 / layer.w13_weight_global_scale[:, 0], requires_grad=False
+        )
+
+        layer.w2_weight_scale_2 = torch.nn.Parameter(
+            1 / layer.w2_weight_global_scale.data, requires_grad=False
+        )
+
+        # w13
+        if self.use_flashinfer_trtllm:
+            w13_input_global_scale = (
+                layer.w13_input_global_scale.min()
+                .to(torch.float32)
+                .expand(layer.num_local_experts)
+            )
+        else:
+            w13_input_global_scale = layer.w13_input_global_scale.min(dim=1).values.to(
+                torch.float32
+            )
+        layer.g1_alphas = torch.nn.Parameter(
+            ((1 / w13_input_global_scale) * layer.w13_weight_scale_2),
+            requires_grad=False,
+        )
+
+        layer.w13_input_scale_quant = torch.nn.Parameter(
+            (w13_input_global_scale), requires_grad=False
+        )
+
+        # w2
+        if self.use_flashinfer_trtllm:
+            w2_input_global_scale = (
+                layer.w2_input_global_scale.min()
+                .to(torch.float32)
+                .expand(layer.num_local_experts)
+            )
+        else:
+            w2_input_global_scale = layer.w2_input_global_scale
+
+        layer.g2_alphas = torch.nn.Parameter(
+            ((1 / w2_input_global_scale) * layer.w2_weight_scale_2).to(torch.float32),
+            requires_grad=False,
+        )
+
+        layer.w2_input_scale_quant = torch.nn.Parameter(
+            (w2_input_global_scale), requires_grad=False
+        )
+
+        # TensorRT-LLM specific processing
+        if self.use_flashinfer_trtllm:
+            # Prepare static weights for TRT-LLM kernel
+            (
+                gemm1_weights_fp4_shuffled,
+                gemm1_scales_fp4_shuffled,
+                gemm2_weights_fp4_shuffled,
+                gemm2_scales_fp4_shuffled,
+            ) = prepare_static_weights_for_trtllm_fp4_moe(
+                layer.w13_weight,
+                layer.w2_weight,
+                layer.w13_weight_scale,
+                layer.w2_weight_scale,
+                layer.w2_weight.size(-2),  # hidden_size
+                layer.w13_weight.size(-2) // 2,  # intermediate_size
+                layer.w13_weight.size(0),  # num_experts
+            )
+            logger.debug("Finished shuffling weights for TRT-LLM MOE")
+
+            replace_parameter(layer, "w13_weight", gemm1_weights_fp4_shuffled)
+            replace_parameter(layer, "w2_weight", gemm2_weights_fp4_shuffled)
+            replace_parameter(layer, "w13_weight_scale", gemm1_scales_fp4_shuffled)
+            replace_parameter(layer, "w2_weight_scale", gemm2_scales_fp4_shuffled)
+
+            # Additional parameter needed for TRT-LLM
+            layer.g1_scale_c = torch.nn.Parameter(
+                (layer.w2_input_scale_quant * layer.g1_alphas).to(torch.float32),
+                requires_grad=False,
+            )
+        else:
+            # swizzle weight scales
+            layer.w13_weight_scale = torch.nn.Parameter(
+                swizzle_blockscale(layer.w13_weight_scale), requires_grad=False
+            )
+
+            layer.w2_weight_scale = torch.nn.Parameter(
+                swizzle_blockscale(layer.w2_weight_scale), requires_grad=False
+            )
+
+            layer.cutlass_moe_params = CutlassMoEParams(
+                CutlassMoEType.BlockscaledFP4,
+                layer.w13_weight.device,
+                num_experts=layer.num_experts,
+                intermediate_size_per_partition=layer.w2_weight.shape[2] * 2,
+                hidden_size=layer.w13_weight.shape[2] * 2,
+            )
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+        self.runner = MoeRunner(MoeRunnerBackend.TRITON, moe_runner_config)
+
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+
+        from sglang.srt.layers.moe.token_dispatcher import StandardCombineInput
+
+        x = dispatch_output.hidden_states
+        topk_output = dispatch_output.topk_output
+
+        if self.use_flashinfer_trtllm:
+            from flashinfer import trtllm_fp4_block_scale_moe
+
+            from sglang.srt.layers.quantization.fp4_utils import fp4_quantize
+
+            router_logits = topk_output.router_logits
+            topk_config = topk_output.topk_config
+
+            # global_scale must be shape [1] (strict in cute-dsl backend).
+            hs_fp4_bytes, hs_sf_bytes = fp4_quantize(
+                x,
+                layer.w13_input_scale_quant[:1],
+                self.group_size,  # sf_vec_size
+                False,  # use_ue8m0
+                False,  # is_sf_swizzled_layout
+            )
+            hs_fp4 = hs_fp4_bytes.reshape(x.shape[0], x.shape[1] // 2)
+            hs_scale = hs_sf_bytes.view(torch.float8_e4m3fn).reshape(
+                *hs_sf_bytes.shape[:-1], -1
+            )
+
+            correction_bias = (
+                None
+                if topk_config.correction_bias is None
+                else topk_config.correction_bias.to(x.dtype)
+            )
+
+            assert layer.routing_method_type is not None
+
+            # DeepSeekV3 style routing requires float32 router logits
+            if layer.routing_method_type == RoutingMethodType.DeepSeekV3:
+                router_logits = router_logits.to(torch.float32)
+
+            routed_scaling_factor = self.moe_runner_config.routed_scaling_factor
+            routed_scaling_factor = (
+                routed_scaling_factor if routed_scaling_factor is not None else 1.0
+            )
+
+            with use_symmetric_memory(
+                get_tp_group(), disabled=not is_allocation_symmetric()
+            ):
+                num_tokens = hs_fp4.shape[0]
+                hidden_size = (
+                    hs_fp4.shape[-1] * 2
+                    if hs_fp4.dtype == torch.uint8
+                    else hs_fp4.shape[-1]
+                )
+                symm_output = torch.empty(
+                    num_tokens, hidden_size, dtype=torch.bfloat16, device=hs_fp4.device
+                )
+
+            output = trtllm_fp4_block_scale_moe(
+                routing_logits=router_logits,
+                routing_bias=correction_bias,
+                hidden_states=hs_fp4,
+                hidden_states_scale=hs_scale,
+                gemm1_weights=layer.w13_weight,
+                gemm1_weights_scale=layer.w13_weight_scale.view(torch.float8_e4m3fn),
+                gemm1_bias=None,
+                gemm1_alpha=None,
+                gemm1_beta=None,
+                gemm1_clamp_limit=None,
+                gemm2_weights=layer.w2_weight,
+                gemm2_weights_scale=layer.w2_weight_scale.view(torch.float8_e4m3fn),
+                gemm2_bias=None,
+                output1_scale_scalar=layer.g1_scale_c,
+                output1_scale_gate_scalar=layer.g1_alphas,
+                output2_scale_scalar=layer.g2_alphas,
+                num_experts=layer.num_experts,
+                top_k=topk_config.top_k,
+                n_group=topk_config.num_expert_group,
+                topk_group=topk_config.topk_group,
+                intermediate_size=layer.intermediate_size_per_partition,
+                local_expert_offset=layer.moe_ep_rank * layer.num_local_experts,
+                local_num_experts=layer.num_local_experts,
+                routed_scaling_factor=routed_scaling_factor,
+                routing_method_type=layer.routing_method_type,
+                do_finalize=True,
+                tune_max_num_tokens=next_power_of_2(hs_fp4.shape[0]),
+                output=symm_output,
+            )[0]
+        else:
+            from sglang.srt.layers.moe.cutlass_moe import cutlass_moe_fp4
+
+            topk_weights, topk_ids = topk_output.topk_weights, topk_output.topk_ids
+
+            output = cutlass_moe_fp4(
+                a=x,
+                a1_gscale=layer.w13_input_scale_quant,
+                w1_fp4=layer.w13_weight,
+                w1_blockscale=layer.w13_weight_scale,
+                w1_alphas=layer.g1_alphas,
+                a2_gscale=layer.w2_input_scale_quant,
+                w2_fp4=layer.w2_weight,
+                w2_blockscale=layer.w2_weight_scale,
+                w2_alphas=layer.g2_alphas,
+                topk_weights=topk_weights,
+                topk_ids=topk_ids,
+                params=layer.cutlass_moe_params,
+                apply_router_weight_on_input=self.moe_runner_config.apply_router_weight_on_input,
+            ).to(x.dtype)
+
+        return StandardCombineInput(hidden_states=output)

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a8_int8_moe.py

@@ -0,0 +1,293 @@
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING
+
+import torch
+
+from sglang.srt.hardware_backend.npu.quantization.fused_moe_method_npu import (
+    NPUW4A8Int8DynamicMoEMethod,
+)
+from sglang.srt.layers.moe import MoeRunnerConfig
+from sglang.srt.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsMoEScheme,
+)
+from sglang.srt.utils import set_weight_attrs
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
+
+__all__ = ["NPUCompressedTensorsW4A8Int8DynamicMoE"]
+
+
+logger = logging.getLogger(__name__)
+
+
+class NPUCompressedTensorsW4A8Int8DynamicMoE(CompressedTensorsMoEScheme):
+
+    ### TODO: Get rid of code duplication with python/sglang/srt/modelslim/modelslim_moe.py @OrangeRedeng @TamirBaydasov
+    def __init__(self, quantization_config) -> None:
+        self.group_size = 0
+        self.is_per_channel_weight = self.group_size == 0
+        self.tp_size = 1
+        self.activation_use_clip = (
+            quantization_config.get("config_groups", {})
+            .get("group_1", {})
+            .get("activation_use_clip", False)
+        )
+        self.kernel = NPUW4A8Int8DynamicMoEMethod()
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ) -> None:
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoeWeightScaleSupported
+
+        self.num_experts = num_experts
+        extra_weight_attrs.update(
+            {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value}
+        )
+
+        # >> weight
+        w13_output_size = intermediate_size_per_partition
+        w2_output_size = hidden_size // 2
+        w13_weight = torch.nn.Parameter(
+            torch.empty(num_experts, w13_output_size, hidden_size, dtype=torch.int8),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight", w13_weight)
+        set_weight_attrs(w13_weight, extra_weight_attrs)
+        w2_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                w2_output_size,
+                intermediate_size_per_partition,
+                dtype=torch.int8,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight", w2_weight)
+        set_weight_attrs(w2_weight, extra_weight_attrs)
+
+        # >> scale
+        weight_scale_dtype = torch.int64 if self.activation_use_clip else torch.float32
+        w13_weight_scale = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                2 * intermediate_size_per_partition,
+                1,
+                dtype=weight_scale_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_scale", w13_weight_scale)
+        set_weight_attrs(w13_weight_scale, extra_weight_attrs)
+
+        w2_weight_scale = torch.nn.Parameter(
+            torch.empty(num_experts, hidden_size, 1, dtype=weight_scale_dtype),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_scale", w2_weight_scale)
+        set_weight_attrs(w2_weight_scale, extra_weight_attrs)
+
+        # >> offset
+        w13_weight_offset = torch.nn.Parameter(
+            torch.empty(
+                num_experts, 2 * intermediate_size_per_partition, 1, dtype=torch.float32
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_offset", w13_weight_offset)
+        set_weight_attrs(w13_weight_offset, extra_weight_attrs)
+
+        w2_weight_offset = torch.nn.Parameter(
+            torch.empty(num_experts, hidden_size, 1, dtype=torch.float32),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_offset", w2_weight_offset)
+        set_weight_attrs(w2_weight_offset, extra_weight_attrs)
+
+        # >>> special param for w4a8
+        if self.activation_use_clip:
+            self._init_activation_clip_params(
+                layer,
+                num_experts,
+                hidden_size,
+                intermediate_size_per_partition,
+                extra_weight_attrs,
+            )
+        else:
+            self._init_extra_scale_params(
+                layer,
+                num_experts,
+                hidden_size,
+                intermediate_size_per_partition,
+                extra_weight_attrs,
+            )
+
+    def _init_activation_clip_params(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        extra_weight_attrs: dict,
+    ) -> None:
+        """
+        Initializes bias and alpha parameters for quantization schemes that use activation clipping.
+
+        This helper registers `w13_bias`, `w2_bias`, and `w2_alpha`, which are required to
+        shift and scale the activations or outputs to compensate for the precision loss
+        introduced by clamping activations.
+        """
+        w13_bias = torch.nn.Parameter(
+            torch.ones(
+                num_experts, 2 * intermediate_size_per_partition, dtype=torch.float
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_bias", w13_bias)
+        set_weight_attrs(w13_bias, extra_weight_attrs)
+
+        w2_bias = torch.nn.Parameter(
+            torch.ones(num_experts, hidden_size, dtype=torch.float),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_bias", w2_bias)
+        set_weight_attrs(w2_bias, extra_weight_attrs)
+
+        w2_alpha = torch.nn.Parameter(
+            torch.ones(num_experts, dtype=torch.float), requires_grad=False
+        )
+        layer.register_parameter("w2_alpha", w2_alpha)
+        set_weight_attrs(w2_alpha, extra_weight_attrs)
+
+    def _init_extra_scale_params(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        extra_weight_attrs: dict,
+    ) -> None:
+        """
+        Initializes additional scaling, offset, and bias parameters for quantization schemes without activation clipping.
+
+        This method registers the following parameters:
+        1. Scale Biases: `w13_scale_bias` and `w2_scale_bias`.
+        2. Secondary Quantization Params (initialized only for grouped quantization):
+            `w13_weight_scale_second`, `w13_weight_offset_second`,
+            `w2_weight_scale_second`, and `w2_weight_offset_second`.
+        """
+        if not self.is_per_channel_weight:
+            w13_weight_scale_second = torch.nn.Parameter(
+                torch.empty(
+                    num_experts,
+                    2 * intermediate_size_per_partition,
+                    hidden_size // self.group_size,
+                    dtype=torch.float32,
+                ),
+                requires_grad=False,
+            )
+            layer.register_parameter("w13_weight_scale_second", w13_weight_scale_second)
+            set_weight_attrs(w13_weight_scale_second, extra_weight_attrs)
+
+            w13_weight_offset_second = torch.nn.Parameter(
+                torch.empty(
+                    num_experts,
+                    2 * intermediate_size_per_partition,
+                    hidden_size // self.group_size,
+                    dtype=torch.float32,
+                ),
+                requires_grad=False,
+            )
+            layer.register_parameter(
+                "w13_weight_offset_second", w13_weight_offset_second
+            )
+            set_weight_attrs(w13_weight_offset_second, extra_weight_attrs)
+
+            w2_weight_scale_second = torch.nn.Parameter(
+                torch.empty(
+                    num_experts,
+                    hidden_size,
+                    intermediate_size_per_partition // self.group_size,
+                    dtype=torch.float32,
+                ),
+                requires_grad=False,
+            )
+            layer.register_parameter("w2_weight_scale_second", w2_weight_scale_second)
+            set_weight_attrs(w2_weight_scale_second, extra_weight_attrs)
+
+            w2_weight_offset_second = torch.nn.Parameter(
+                torch.empty(
+                    num_experts,
+                    hidden_size,
+                    intermediate_size_per_partition // self.group_size,
+                    dtype=torch.float32,
+                ),
+                requires_grad=False,
+            )
+            layer.register_parameter("w2_weight_offset_second", w2_weight_offset_second)
+            set_weight_attrs(w2_weight_offset_second, extra_weight_attrs)
+
+        w13_scale_bias = torch.nn.Parameter(
+            torch.empty(
+                num_experts, 2 * intermediate_size_per_partition, 1, dtype=torch.float32
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_scale_bias", w13_scale_bias)
+        set_weight_attrs(w13_scale_bias, extra_weight_attrs)
+
+        w2_scale_bias = torch.nn.Parameter(
+            torch.empty(
+                num_experts, hidden_size, 16 // self.tp_size, dtype=torch.float32
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_scale_bias", w2_scale_bias)
+        set_weight_attrs(w2_scale_bias, extra_weight_attrs)
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        self.kernel.process_weights_after_loading(
+            layer, self.is_per_channel_weight, self.activation_use_clip
+        )
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+
+        return self.kernel.apply(layer, dispatch_output)
+
+    def apply_weights_with_router_logits(
+        self,
+        layer,
+        hidden_states,
+        hidden_states_scale,
+        group_list_type,
+        group_list,
+        output_dtype,
+    ):
+        return self.kernel.apply_without_routing_weights(
+            layer,
+            hidden_states,
+            hidden_states_scale,
+            group_list_type,
+            group_list,
+            output_dtype,
+        )

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a16_fp8.py

@@ -0,0 +1,136 @@
+# Adapted from https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/quantization/compressed_tensors
+# SPDX-License-Identifier: Apache-2.0
+
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Callable, List, Optional
+
+import torch
+from compressed_tensors.quantization import QuantizationStrategy
+
+from sglang.srt.layers.parameter import (
+    ChannelQuantScaleParameter,
+    ModelWeightParameter,
+    PerTensorScaleParameter,
+)
+from sglang.srt.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsLinearScheme,
+)
+from sglang.srt.layers.quantization.marlin_utils_fp8 import (
+    apply_fp8_marlin_linear,
+    prepare_fp8_layer_for_marlin,
+)
+from sglang.srt.layers.quantization.utils import convert_to_channelwise
+
+__all__ = ["CompressedTensorsW8A16Fp8"]
+
+SUPPORTED_STRATEGIES = [QuantizationStrategy.CHANNEL, QuantizationStrategy.TENSOR]
+
+
+class CompressedTensorsW8A16Fp8(CompressedTensorsLinearScheme):
+    def __init__(self, strategy: str, is_static_input_scheme: bool):
+        self.strategy = strategy
+        self.is_static_input_scheme = is_static_input_scheme
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # ampere and up
+        return 80
+
+    # W8A8-Fp8 kernels support only per-tensor and per-channel cases.
+    # So if we have a fused module (QKV, MLP) with per tensor scales,
+    # we expand each scale to its shard's channels.
+    def process_weights_after_loading(self, layer) -> None:
+        if self.strategy == QuantizationStrategy.TENSOR:
+            ws_channelwise = convert_to_channelwise(
+                layer.weight_scale, layer.logical_widths
+            )
+            layer.weight_scale = torch.nn.Parameter(ws_channelwise, requires_grad=False)
+        else:
+            # required by torch.compile to be torch.nn.Parameter
+            layer.weight_scale = torch.nn.Parameter(
+                layer.weight_scale.data, requires_grad=False
+            )
+
+        # Weights must be transposed for marlin
+        layer.weight = torch.nn.Parameter(layer.weight.t(), requires_grad=False)
+
+        if self.is_static_input_scheme:
+            # required by torch.compile to be torch.nn.Parameter
+            layer.input_scale = torch.nn.Parameter(
+                layer.input_scale.data, requires_grad=False
+            )
+        prepare_fp8_layer_for_marlin(layer, size_k_first=True)
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        input_size: int,
+        output_partition_sizes: List[int],
+        input_size_per_partition: int,
+        params_dtype: torch.dtype,
+        weight_loader: Callable,
+        **kwargs,
+    ):
+        output_size_per_partition = sum(output_partition_sizes)
+        layer.logical_widths = output_partition_sizes
+        layer.input_size_per_partition = input_size_per_partition
+        layer.output_size_per_partition = output_size_per_partition
+        layer.orig_dtype = params_dtype
+
+        # WEIGHT
+        weight = ModelWeightParameter(
+            data=torch.empty(
+                output_size_per_partition,
+                input_size_per_partition,
+                dtype=torch.float8_e4m3fn,
+            ),
+            input_dim=1,
+            output_dim=0,
+            weight_loader=weight_loader,
+        )
+        layer.register_parameter("weight", weight)
+
+        # WEIGHT SCALE
+        if self.strategy == QuantizationStrategy.CHANNEL:
+            weight_scale = ChannelQuantScaleParameter(
+                data=torch.empty((sum(output_partition_sizes), 1), dtype=torch.float32),
+                output_dim=0,
+                weight_loader=weight_loader,
+            )
+        elif self.strategy == QuantizationStrategy.TENSOR:
+            weight_scale = PerTensorScaleParameter(
+                data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
+                weight_loader=weight_loader,
+            )
+        else:
+            raise ValueError(
+                f"Unsupported weight strategy={self.strategy}, "
+                f"supported strategies are {SUPPORTED_STRATEGIES}"
+            )
+
+        weight_scale[:] = torch.finfo(torch.float32).min
+        layer.register_parameter("weight_scale", weight_scale)
+
+        # INPUT SCALE (to deal with converted checkpoints)
+        if self.is_static_input_scheme:
+            input_scale = PerTensorScaleParameter(
+                data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
+                weight_loader=weight_loader,
+            )
+            layer.register_parameter("input_scale", input_scale)
+
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        return apply_fp8_marlin_linear(
+            input=x,
+            weight=layer.weight,
+            weight_scale=layer.weight_scale,
+            workspace=layer.workspace,
+            size_n=layer.output_size_per_partition,
+            size_k=layer.input_size_per_partition,
+            bias=bias,
+        )

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py

@@ -0,0 +1,263 @@
+# Adapted from https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/quantization/compressed_tensors
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from typing import Callable, Optional
+
+import torch
+from compressed_tensors.quantization import QuantizationArgs, QuantizationStrategy
+from torch.nn import Parameter
+
+from sglang.srt.layers.parameter import (
+    BlockQuantScaleParameter,
+    ChannelQuantScaleParameter,
+    ModelWeightParameter,
+    PerTensorScaleParameter,
+)
+from sglang.srt.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsLinearScheme,
+)
+from sglang.srt.layers.quantization.fp8_kernel import is_fp8_fnuz
+from sglang.srt.layers.quantization.fp8_utils import (
+    apply_fp8_linear,
+    apply_fp8_ptpc_linear,
+    deepgemm_w8a8_block_fp8_linear_with_fallback,
+    dispatch_w8a8_block_fp8_linear,
+    normalize_e4m3fn_to_e4m3fnuz,
+    requant_block_scale_ue8m0_for_deepgemm,
+    validate_fp8_block_shape,
+)
+from sglang.srt.layers.quantization.utils import requantize_with_max_scale
+from sglang.srt.utils import get_bool_env_var, is_hip
+
+__all__ = ["CompressedTensorsW8A8Fp8"]
+
+_is_hip = is_hip()
+_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
+if _use_aiter:
+    from aiter.ops.shuffle import shuffle_weight
+
+
+strategy_to_parameter_type = {
+    QuantizationStrategy.BLOCK: BlockQuantScaleParameter,
+    QuantizationStrategy.CHANNEL: ChannelQuantScaleParameter,
+    QuantizationStrategy.TENSOR: PerTensorScaleParameter,
+}
+
+
+class CompressedTensorsW8A8Fp8(CompressedTensorsLinearScheme):
+    def __init__(self, weight_quant: QuantizationArgs, is_static_input_scheme: bool):
+        self.weight_quant = weight_quant
+        self.strategy = self.weight_quant.strategy
+        self.is_static_input_scheme = is_static_input_scheme
+        self.weight_block_size = self.weight_quant.block_structure
+        if self.weight_block_size is not None:
+            self.w8a8_block_fp8_linear = dispatch_w8a8_block_fp8_linear()
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # lovelace and up
+        return 89
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        input_size_per_partition: int,
+        output_partition_sizes: list[int],
+        input_size: int,
+        output_size: int,
+        params_dtype: torch.dtype,
+        weight_loader: Callable,
+        **kwargs,
+    ):
+        output_size_per_partition = sum(output_partition_sizes)
+        layer.logical_widths = output_partition_sizes
+        layer.weight_block_size = None
+        layer.orig_dtype = params_dtype
+
+        if self.strategy == QuantizationStrategy.BLOCK:
+            assert self.weight_block_size is not None
+            layer.weight_block_size = self.weight_block_size
+            # Validate block quantization shapes
+            validate_fp8_block_shape(
+                layer,
+                input_size,
+                output_size,
+                input_size_per_partition,
+                output_partition_sizes,
+                self.weight_block_size,
+            )
+
+        # WEIGHT
+        weight = ModelWeightParameter(
+            data=torch.empty(
+                output_size_per_partition,
+                input_size_per_partition,
+                dtype=torch.float8_e4m3fn,
+            ),
+            input_dim=1,
+            output_dim=0,
+            weight_loader=weight_loader,
+        )
+        layer.register_parameter("weight", weight)
+
+        # WEIGHT SCALE
+        if self.strategy == QuantizationStrategy.CHANNEL:
+            weight_scale = ChannelQuantScaleParameter(
+                data=torch.empty((sum(output_partition_sizes), 1), dtype=torch.float32),
+                output_dim=0,
+                weight_loader=weight_loader,
+            )
+            weight_scale[:] = torch.finfo(torch.float32).min
+        elif self.strategy == QuantizationStrategy.TENSOR:
+            weight_scale = PerTensorScaleParameter(
+                data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
+                weight_loader=weight_loader,
+            )
+            weight_scale[:] = torch.finfo(torch.float32).min
+        elif self.strategy == QuantizationStrategy.BLOCK:
+            assert layer.weight_block_size is not None
+            block_n, block_k = layer.weight_block_size[0], layer.weight_block_size[1]
+            output_size_per_partition = sum(output_partition_sizes)
+            weight_scale = BlockQuantScaleParameter(
+                data=torch.empty(
+                    (output_size_per_partition + block_n - 1) // block_n,
+                    (input_size_per_partition + block_k - 1) // block_k,
+                    dtype=torch.float32,
+                ),
+                input_dim=1,
+                output_dim=0,
+                weight_loader=weight_loader,
+            )
+            weight_scale.format_ue8m0 = False
+            weight_scale[:] = torch.finfo(torch.float32).min
+
+        layer.register_parameter("weight_scale", weight_scale)
+        # INPUT SCALE
+        if self.is_static_input_scheme:
+            input_scale = PerTensorScaleParameter(
+                data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
+                weight_loader=weight_loader,
+            )
+            input_scale[:] = torch.finfo(torch.float32).min
+            layer.register_parameter("input_scale", input_scale)
+
+    def process_weights_after_loading(self, layer) -> None:
+        if self.strategy == QuantizationStrategy.TENSOR:
+            max_w_scale, weight = requantize_with_max_scale(
+                weight=layer.weight,
+                weight_scale=layer.weight_scale,
+                logical_widths=layer.logical_widths,
+            )
+
+            if is_fp8_fnuz():
+                input_scale = getattr(layer, "input_scale", None)
+
+                weight, max_w_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz(
+                    weight=weight, weight_scale=max_w_scale, input_scale=input_scale
+                )
+                if input_scale is not None:
+                    layer.input_scale = Parameter(input_scale, requires_grad=False)
+            layer.weight = Parameter(weight.t(), requires_grad=False)
+            layer.weight_scale = Parameter(max_w_scale, requires_grad=False)
+
+        elif self.strategy == QuantizationStrategy.CHANNEL:
+            weight = layer.weight
+
+            if is_fp8_fnuz():
+                input_scale = getattr(layer, "input_scale", None)
+
+                weight, weight_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz(
+                    weight=weight,
+                    weight_scale=layer.weight_scale,
+                    input_scale=input_scale,
+                )
+                if input_scale is not None:
+                    layer.input_scale = Parameter(input_scale, requires_grad=False)
+            else:
+                weight_scale = layer.weight_scale.data
+
+            if _use_aiter:
+                # keep the weight as (N, K)
+                layer.weight = Parameter(
+                    shuffle_weight(weight, (16, 16)), requires_grad=False
+                )
+            else:
+                layer.weight = Parameter(weight.t(), requires_grad=False)
+
+            # required by torch.compile to be torch.nn.Parameter
+            layer.weight_scale = Parameter(weight_scale, requires_grad=False)
+
+        elif self.strategy == QuantizationStrategy.BLOCK:
+            assert self.is_static_input_scheme is False
+            if is_fp8_fnuz():
+                weight, weight_scale, _ = normalize_e4m3fn_to_e4m3fnuz(
+                    weight=layer.weight, weight_scale=layer.weight_scale
+                )
+                layer.weight = Parameter(weight.data, requires_grad=False)
+                layer.weight_scale = Parameter(weight_scale.data, requires_grad=False)
+                layer.weight_scale.format_ue8m0 = False
+            else:
+                layer.weight.requires_grad_(False)
+                layer.weight_scale.requires_grad_(False)
+
+            # On Blackwell, block-FP8 dispatches to DeepGEMM, which needs the
+            # weight scales UE8M0-packed to match its UE8M0 activation scales.
+            use_deepgemm_runner = (
+                self.w8a8_block_fp8_linear
+                is deepgemm_w8a8_block_fp8_linear_with_fallback
+            )
+            requant_block_scale_ue8m0_for_deepgemm(
+                layer.weight,
+                layer.weight_scale,
+                self.weight_block_size,
+                use_deepgemm_runner=use_deepgemm_runner,
+                output_dtype=getattr(layer, "orig_dtype", None),
+                weight_shape=layer.weight.shape,
+            )
+
+        else:
+            raise ValueError(f"Unknown quantization strategy {self.strategy}")
+
+        # INPUT SCALE
+        if self.is_static_input_scheme and hasattr(layer, "input_scale"):
+            layer.input_scale = Parameter(layer.input_scale.max(), requires_grad=False)
+        else:
+            layer.input_scale = None
+
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if self.weight_block_size is not None:
+            return self.w8a8_block_fp8_linear(
+                input=x,
+                weight=layer.weight,
+                block_size=self.weight_block_size,
+                weight_scale=layer.weight_scale,
+                input_scale=layer.input_scale,
+                bias=bias,
+            )
+
+        if _use_aiter and self.strategy == QuantizationStrategy.CHANNEL:
+            return apply_fp8_ptpc_linear(
+                input=x,
+                weight=layer.weight,
+                weight_scale=layer.weight_scale,
+                input_scale=layer.input_scale,
+                bias=bias,
+                use_per_token_if_dynamic=True,
+                compressed_tensor_quant=True,
+            )
+        else:
+            return apply_fp8_linear(
+                input=x,
+                weight=layer.weight,
+                weight_scale=layer.weight_scale,
+                input_scale=layer.input_scale,
+                bias=bias,
+                use_per_token_if_dynamic=True,
+                compressed_tensor_quant=True,
+            )

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8_moe.py

@@ -0,0 +1,445 @@
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING
+
+import torch
+from compressed_tensors.quantization import QuantizationStrategy
+
+from sglang.srt.layers.moe import MoeRunner, MoeRunnerBackend, MoeRunnerConfig
+from sglang.srt.layers.moe.moe_runner.flashinfer_trtllm import (
+    FlashInferTrtllmFp8MoeQuantInfo,
+)
+from sglang.srt.layers.moe.moe_runner.triton import TritonMoeQuantInfo
+from sglang.srt.layers.moe.utils import (
+    get_moe_a2a_backend,
+    get_moe_runner_backend,
+    get_moe_weight_sizes,
+)
+from sglang.srt.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsMoEScheme,
+)
+from sglang.srt.layers.quantization.fp8_kernel import is_fp8_fnuz, scaled_fp8_quant
+from sglang.srt.layers.quantization.fp8_utils import normalize_e4m3fn_to_e4m3fnuz
+from sglang.srt.layers.quantization.utils import (
+    all_close_1d,
+    per_tensor_dequantize,
+    swap_w13_to_w31,
+)
+from sglang.srt.runtime_context import get_parallel
+from sglang.srt.utils import get_bool_env_var, is_hip, set_weight_attrs
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
+
+__all__ = ["CompressedTensorsW8A8Fp8MoE"]
+
+_is_hip = is_hip()
+_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
+
+if _use_aiter:
+    from aiter.ops.shuffle import shuffle_weight
+
+
+logger = logging.getLogger(__name__)
+
+
+class CompressedTensorsW8A8Fp8MoE(CompressedTensorsMoEScheme):
+
+    def __init__(self, weight_quant, input_quant):
+        self.weight_quant = weight_quant
+        self.input_quant = input_quant
+        self.use_flashinfer_trtllm = get_moe_runner_backend().is_flashinfer_trtllm()
+
+        per_tensor = (
+            self.weight_quant.strategy == QuantizationStrategy.TENSOR
+            and self.input_quant.strategy == QuantizationStrategy.TENSOR
+        )
+        per_channel = (
+            self.weight_quant.strategy == QuantizationStrategy.CHANNEL
+            and self.input_quant.strategy == QuantizationStrategy.TOKEN
+        )
+        if not (per_tensor or per_channel):
+            assert self.weight_quant.strategy == QuantizationStrategy.BLOCK
+            self.weight_block_size = self.weight_quant.block_structure
+            assert self.weight_quant.dynamic is not None
+        else:
+            self.weight_block_size = None
+        self.block_quant = self.weight_block_size is not None
+
+        self.static_input_scales = not self.input_quant.dynamic
+        if self.static_input_scales and per_channel:
+            raise ValueError(
+                "For FP8 Fused MoE layer, we require either per tensor or "
+                "channelwise, dynamic per token quantization."
+            )
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # ampere and up
+        return 80
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoeWeightScaleSupported
+
+        params_dtype = torch.float8_e4m3fn
+
+        if self.block_quant:
+            assert self.weight_block_size is not None
+            layer.weight_block_size = self.weight_block_size
+            tp_size = get_parallel().tp_size
+            block_n, block_k = (
+                self.weight_block_size[0],
+                self.weight_block_size[1],
+            )
+            # NOTE: To ensure proper alignment of the block-wise quantization
+            # scales, the output_size of the weights for both the gate and up
+            # layers must be divisible by block_n.
+            # Required by column parallel or enabling merged weights
+            if intermediate_size_per_partition % block_n != 0:
+                raise ValueError(
+                    f"The output_size of gate's and up's weight = "
+                    f"{intermediate_size_per_partition} is not divisible by "
+                    f"weight quantization block_n = {block_n}."
+                )
+            if tp_size > 1 and intermediate_size_per_partition % block_k != 0:
+                # Required by row parallel
+                raise ValueError(
+                    f"The input_size of down's weight = "
+                    f"{intermediate_size_per_partition} is not divisible by "
+                    f"weight quantization block_k = {block_k}."
+                )
+
+        w13_up_dim, w2_down_dim, weight_padded = get_moe_weight_sizes(
+            intermediate_size_per_partition,
+            is_aiter_moe=_use_aiter,
+            is_concat=True,
+            is_packed=False,
+        )
+
+        extra_weight_attrs.update(
+            {"weight_padded": weight_padded},
+        )
+
+        # WEIGHTS
+        w13_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                w13_up_dim,
+                hidden_size,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight", w13_weight)
+        set_weight_attrs(w13_weight, extra_weight_attrs)
+
+        w2_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                w2_down_dim,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight", w2_weight)
+        set_weight_attrs(w2_weight, extra_weight_attrs)
+
+        # WEIGHT_SCALES
+        # per-tensor quantization
+        if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
+            # Allocate 2 scales for w1 and w3 respectively.
+            # They will be combined to a single scale after weight loading.
+            w13_weight_scale = torch.nn.Parameter(
+                torch.ones(num_experts, 2, dtype=torch.float32), requires_grad=False
+            )
+            w2_weight_scale = torch.nn.Parameter(
+                torch.ones(num_experts, dtype=torch.float32), requires_grad=False
+            )
+            weight_quant_method = FusedMoeWeightScaleSupported.TENSOR.value
+        elif self.weight_quant.strategy == QuantizationStrategy.CHANNEL:
+            w13_weight_scale = torch.nn.Parameter(
+                torch.ones(
+                    num_experts,
+                    w13_up_dim,
+                    1,
+                    dtype=torch.float32,
+                ),
+                requires_grad=False,
+            )
+            w2_weight_scale = torch.nn.Parameter(
+                torch.ones(num_experts, hidden_size, 1, dtype=torch.float32),
+                requires_grad=False,
+            )
+            weight_quant_method = FusedMoeWeightScaleSupported.CHANNEL.value
+        elif self.weight_quant.strategy == QuantizationStrategy.BLOCK:
+            w13_weight_scale = torch.nn.Parameter(
+                torch.ones(
+                    num_experts,
+                    2 * ((intermediate_size_per_partition + block_n - 1) // block_n),
+                    (hidden_size + block_k - 1) // block_k,
+                    dtype=torch.float32,
+                ),
+                requires_grad=False,
+            )
+            w2_weight_scale = torch.nn.Parameter(
+                torch.ones(
+                    num_experts,
+                    (hidden_size + block_n - 1) // block_n,
+                    (intermediate_size_per_partition + block_k - 1) // block_k,
+                    dtype=torch.float32,
+                ),
+                requires_grad=False,
+            )
+            weight_quant_method = FusedMoeWeightScaleSupported.BLOCK.value
+        else:
+            raise ValueError(
+                f"Unsupported weight quantization strategy: {self.weight_quant.strategy}"
+            )
+
+        layer.register_parameter("w13_weight_scale", w13_weight_scale)
+        layer.register_parameter("w2_weight_scale", w2_weight_scale)
+        # Add the quantization method used (per tensor/grouped/channel)
+        # to ensure the weight scales are loaded in properly
+        extra_weight_attrs.update({"quant_method": weight_quant_method})
+        set_weight_attrs(w13_weight_scale, extra_weight_attrs)
+        set_weight_attrs(w2_weight_scale, extra_weight_attrs)
+
+        # INPUT_SCALES
+        if self.static_input_scales:
+            assert (
+                self.input_quant.strategy == QuantizationStrategy.TENSOR
+            ), "Only per-tensor quantization is supported for static input scales"
+            w13_input_scale = torch.nn.Parameter(
+                torch.ones(num_experts, dtype=torch.float32), requires_grad=False
+            )
+            layer.register_parameter("w13_input_scale", w13_input_scale)
+            set_weight_attrs(w13_input_scale, extra_weight_attrs)
+
+            w2_input_scale = torch.nn.Parameter(
+                torch.ones(num_experts, dtype=torch.float32), requires_grad=False
+            )
+            layer.register_parameter("w2_input_scale", w2_input_scale)
+            set_weight_attrs(w2_input_scale, extra_weight_attrs)
+        else:
+            layer.w13_input_scale = None
+            layer.w2_input_scale = None
+
+    def process_weights_after_loading(self, layer: torch.nn.Module | FusedMoE) -> None:
+        # Fp8 moe kernels require a single activation scale.
+        # We take the max of all the scales in case they differ.
+        if self.static_input_scales:
+            if layer.w13_input_scale is None or layer.w2_input_scale is None:
+                raise ValueError(
+                    "QuantConfig has static quantization, but found "
+                    "activation scales are None."
+                )
+            if not all_close_1d(layer.w13_input_scale) or not all_close_1d(
+                layer.w2_input_scale
+            ):
+                logger.warning(
+                    "Found input_scales that are not equal for "
+                    "fp8 MoE layer. Using the maximum across experts "
+                    "for each layer."
+                )
+            layer.w13_input_scale = torch.nn.Parameter(
+                layer.w13_input_scale.max(), requires_grad=False
+            )
+            layer.w2_input_scale = torch.nn.Parameter(
+                layer.w2_input_scale.max(), requires_grad=False
+            )
+
+        if is_fp8_fnuz():
+            # Normalize the weights and scales
+            w13_weight, w13_weight_scale, w13_input_scale = (
+                normalize_e4m3fn_to_e4m3fnuz(
+                    layer.w13_weight, layer.w13_weight_scale, layer.w13_input_scale
+                )
+            )
+            w2_weight, w2_weight_scale, w2_input_scale = normalize_e4m3fn_to_e4m3fnuz(
+                layer.w2_weight, layer.w2_weight_scale, layer.w2_input_scale
+            )
+            # Reset the parameter
+            layer.w13_weight = torch.nn.Parameter(w13_weight, requires_grad=False)
+            layer.w13_weight_scale = torch.nn.Parameter(
+                w13_weight_scale, requires_grad=False
+            )
+            if w13_input_scale is not None:
+                layer.w13_input_scale = torch.nn.Parameter(
+                    w13_input_scale, requires_grad=False
+                )
+            layer.w2_weight = torch.nn.Parameter(w2_weight, requires_grad=False)
+            layer.w2_weight_scale = torch.nn.Parameter(
+                w2_weight_scale, requires_grad=False
+            )
+            if w2_input_scale is not None:
+                layer.w2_input_scale = torch.nn.Parameter(
+                    w2_input_scale, requires_grad=False
+                )
+        if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
+            # Fp8 moe kernel needs single weight scale for w13 per expert.
+            # We take the max then dequant and requant each expert.
+            assert layer.w13_weight_scale is not None
+            shard_size = layer.intermediate_size_per_partition
+            max_w13_scales = layer.w13_weight_scale.max(dim=1).values
+            for expert_id in range(layer.num_local_experts):
+                start = 0
+                for shard_id in range(2):
+                    dq_weight = per_tensor_dequantize(
+                        layer.w13_weight[expert_id][start : start + shard_size, :],
+                        layer.w13_weight_scale[expert_id][shard_id],
+                    )
+                    (
+                        layer.w13_weight[expert_id][start : start + shard_size, :],
+                        _,
+                    ) = scaled_fp8_quant(dq_weight, max_w13_scales[expert_id])
+
+                    start += shard_size
+
+            layer.w13_weight_scale = torch.nn.Parameter(
+                max_w13_scales, requires_grad=False
+            )
+
+        if self.weight_quant.strategy == QuantizationStrategy.CHANNEL and _use_aiter:
+            with torch.no_grad():
+                # Pre-shuffle weights
+                layer.w13_weight = torch.nn.Parameter(
+                    shuffle_weight(layer.w13_weight.data, (16, 16)),
+                    requires_grad=False,
+                )
+                torch.cuda.empty_cache()
+                layer.w2_weight = torch.nn.Parameter(
+                    shuffle_weight(layer.w2_weight.data, (16, 16)),
+                    requires_grad=False,
+                )
+                torch.cuda.empty_cache()
+
+        if (
+            self.weight_quant.strategy == QuantizationStrategy.BLOCK
+            and self.use_flashinfer_trtllm
+        ):
+            layer.w13_weight = torch.nn.Parameter(
+                swap_w13_to_w31(layer.w13_weight.data),
+                requires_grad=False,
+            )
+            layer.w13_weight_scale = torch.nn.Parameter(
+                swap_w13_to_w31(layer.w13_weight_scale.data),
+                requires_grad=False,
+            )
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+        moe_runner_backend = get_moe_runner_backend()
+        if moe_runner_backend.is_auto():
+            if (
+                _use_aiter
+                and self.weight_quant.strategy == QuantizationStrategy.CHANNEL
+                and get_moe_a2a_backend().supports_aiter()
+            ):
+                moe_runner_backend = MoeRunnerBackend.AITER
+            else:
+                moe_runner_backend = MoeRunnerBackend.TRITON
+
+        if (
+            moe_runner_backend.is_aiter()
+            or moe_runner_backend.is_triton()
+            or moe_runner_backend.is_flashinfer_trtllm()
+            or moe_runner_backend.is_flashinfer_trtllm_routed()
+        ):
+            self.runner = MoeRunner(moe_runner_backend, moe_runner_config)
+        else:
+            # TODO(cwan): refactor other backends
+            pass
+
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+
+        x = dispatch_output.hidden_states
+        topk_output = dispatch_output.topk_output
+
+        moe_runner_config = self.moe_runner_config
+
+        if self.runner.runner_backend.is_aiter():
+            from sglang.srt.layers.moe.moe_runner.aiter import (
+                AiterMoeQuantInfo,
+                AiterQuantType,
+            )
+
+            assert not moe_runner_config.no_combine, "unsupported"
+            quant_info = AiterMoeQuantInfo(
+                w13_weight=layer.w13_weight,
+                w2_weight=layer.w2_weight,
+                quant_type=AiterQuantType.PER_TOKEN,
+                w13_scale=layer.w13_weight_scale,
+                w2_scale=layer.w2_weight_scale,
+                a13_scale=layer.w13_input_scale,
+                a2_scale=layer.w2_input_scale,
+            )
+            return self.runner.run(dispatch_output, quant_info)
+        elif self.weight_quant.strategy == QuantizationStrategy.BLOCK:
+            if self.use_flashinfer_trtllm:
+                from sglang.srt.layers.moe.moe_runner.flashinfer_trtllm import (
+                    get_activation_type,
+                )
+
+                activation_type = get_activation_type(
+                    moe_runner_config.activation,
+                    is_gated=moe_runner_config.is_gated,
+                )
+                quant_info = FlashInferTrtllmFp8MoeQuantInfo(
+                    w13_weight=layer.w13_weight,
+                    w2_weight=layer.w2_weight,
+                    global_num_experts=layer.num_experts,
+                    local_expert_offset=layer.moe_ep_rank * layer.num_local_experts,
+                    local_num_experts=layer.num_local_experts,
+                    intermediate_size=layer.w2_weight.shape[2],
+                    routing_method_type=layer.routing_method_type,
+                    block_quant=self.block_quant,
+                    weight_block_k=self.weight_block_size[1],
+                    w13_weight_scale_inv=layer.w13_weight_scale,
+                    w2_weight_scale_inv=layer.w2_weight_scale,
+                    activation_type=activation_type,
+                )
+            else:
+                quant_info = TritonMoeQuantInfo(
+                    w13_weight=layer.w13_weight,
+                    w2_weight=layer.w2_weight,
+                    use_fp8_w8a8=True,
+                    w13_scale=layer.w13_weight_scale,
+                    w2_scale=layer.w2_weight_scale,
+                    a13_scale=layer.w13_input_scale,
+                    a2_scale=layer.w2_input_scale,
+                    block_shape=self.weight_block_size,
+                )
+            return self.runner.run(dispatch_output, quant_info)
+        else:
+            quant_info = TritonMoeQuantInfo(
+                w13_weight=layer.w13_weight,
+                w2_weight=layer.w2_weight,
+                use_fp8_w8a8=True,
+                per_channel_quant=self.weight_quant.strategy
+                == QuantizationStrategy.CHANNEL,
+                w13_scale=layer.w13_weight_scale,
+                w2_scale=layer.w2_weight_scale,
+                a13_scale=layer.w13_input_scale,
+                a2_scale=layer.w2_input_scale,
+            )
+            return self.runner.run(dispatch_output, quant_info)

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_int8.py

@@ -0,0 +1,204 @@
+# Adapted from https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/quantization/compressed_tensors
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from typing import Callable, Optional
+
+import torch
+from compressed_tensors.quantization import QuantizationStrategy
+from torch.nn import Parameter
+
+from sglang.srt.hardware_backend.npu.quantization.linear_method_npu import (
+    NPUW8A8Int8DynamicLinearMethod,
+)
+from sglang.srt.layers.parameter import (
+    ChannelQuantScaleParameter,
+    ModelWeightParameter,
+    PerTensorScaleParameter,
+)
+from sglang.srt.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsLinearScheme,
+)
+from sglang.srt.layers.quantization.int8_kernel import per_token_quant_int8
+from sglang.srt.layers.quantization.utils import requantize_with_max_scale
+from sglang.srt.utils import is_cuda
+
+__all__ = ["CompressedTensorsW8A8Int8", "NPUCompressedTensorsW8A8Int8"]
+
+_is_cuda = is_cuda()
+if _is_cuda:
+    from sgl_kernel import int8_scaled_mm
+
+
+class CompressedTensorsW8A8Int8(CompressedTensorsLinearScheme):
+
+    def __init__(
+        self, strategy: str, is_static_input_scheme: bool, input_symmetric: bool
+    ):
+        self.strategy = strategy
+        self.is_static_input_scheme = is_static_input_scheme
+        self.input_symmetric = input_symmetric
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        output_partition_sizes: list[int],
+        input_size_per_partition: int,
+        params_dtype: torch.dtype,
+        weight_loader: Callable,
+        **kwargs,
+    ):
+        output_size_per_partition = sum(output_partition_sizes)
+        layer.logical_widths = output_partition_sizes
+
+        # WEIGHT
+        weight = ModelWeightParameter(
+            data=torch.empty(
+                output_size_per_partition, input_size_per_partition, dtype=torch.int8
+            ),
+            input_dim=1,
+            output_dim=0,
+            weight_loader=weight_loader,
+        )
+
+        layer.register_parameter("weight", weight)
+
+        # WEIGHT SCALE
+        if self.strategy == QuantizationStrategy.CHANNEL:
+            weight_scale = ChannelQuantScaleParameter(
+                data=torch.empty((sum(output_partition_sizes), 1), dtype=torch.float32),
+                output_dim=0,
+                weight_loader=weight_loader,
+            )
+        else:
+            assert self.strategy == QuantizationStrategy.TENSOR
+            weight_scale = PerTensorScaleParameter(
+                data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
+                weight_loader=weight_loader,
+            )
+        layer.register_parameter("weight_scale", weight_scale)
+
+        # INPUT SCALE
+        if self.is_static_input_scheme:
+            input_scale = PerTensorScaleParameter(
+                data=torch.empty(1, dtype=torch.float32), weight_loader=weight_loader
+            )
+            layer.register_parameter("input_scale", input_scale)
+
+            if not self.input_symmetric:
+                # Note: compressed-tensors stores the zp using the same dtype
+                # as the weights
+                # AZP loaded as int8 but used as int32
+                input_zero_point = PerTensorScaleParameter(
+                    data=torch.empty(1, dtype=torch.int8), weight_loader=weight_loader
+                )
+                layer.register_parameter("input_zero_point", input_zero_point)
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # ampere and up
+        return 80
+
+    def process_weights_after_loading(self, layer) -> None:
+        # If per tensor, when we have a fused module (e.g. QKV) with per
+        # tensor scales (thus N scales being passed to the kernel),
+        # requantize so we can always run per channel
+        if self.strategy == QuantizationStrategy.TENSOR:
+            max_w_scale, weight = requantize_with_max_scale(
+                weight=layer.weight,
+                weight_scale=layer.weight_scale,
+                logical_widths=layer.logical_widths,
+            )
+
+            layer.weight = Parameter(weight.t(), requires_grad=False)
+            layer.weight_scale = Parameter(max_w_scale, requires_grad=False)
+
+        # If channelwise, scales are already lined up, so just transpose.
+        elif self.strategy == QuantizationStrategy.CHANNEL:
+            weight = layer.weight
+            weight_scale = layer.weight_scale.data
+
+            layer.weight = Parameter(weight.t(), requires_grad=False)
+            # required by torch.compile to be torch.nn.Parameter
+            layer.weight_scale = Parameter(weight_scale, requires_grad=False)
+
+        else:
+            raise ValueError(f"Unknown quantization strategy {self.strategy}")
+
+        # INPUT SCALE
+        if self.is_static_input_scheme and hasattr(layer, "input_scale"):
+            if self.input_symmetric:
+                layer.input_scale = Parameter(
+                    layer.input_scale.max(), requires_grad=False
+                )
+            else:
+                input_scale = layer.input_scale
+                input_zero_point = layer.input_zero_point
+
+                # reconstruct the ranges
+                int8_traits = torch.iinfo(torch.int8)
+                azps = input_zero_point.to(dtype=torch.int32)
+                range_max = (input_scale * (int8_traits.max - azps)).max()
+                range_min = (input_scale * (int8_traits.min - azps)).min()
+
+                scale = (range_max - range_min) / (int8_traits.max - int8_traits.min)
+
+                # AZP loaded as int8 but used as int32
+                azp = (int8_traits.min - range_min / scale).to(dtype=torch.int32)
+
+                layer.input_scale = Parameter(scale, requires_grad=False)
+                layer.input_zero_point = Parameter(azp, requires_grad=False)
+        else:
+            layer.input_scale = None
+            layer.input_zero_point = None
+
+        # azp_adj is the AZP adjustment term, used to account for weights.
+        # It does not depend on scales or azp, so it is the same for
+        # static and dynamic quantization.
+        # For more details, see csrc/quantization/cutlass_w8a8/Epilogues.md
+        # https://github.com/vllm-project/vllm/blob/8d59dbb00044a588cab96bcdc028006ed922eb06/csrc/quantization/cutlass_w8a8/Epilogues.md
+        if not self.input_symmetric:
+            weight = layer.weight
+            azp_adj = weight.sum(dim=0, keepdim=True, dtype=torch.int32)
+            if self.is_static_input_scheme:
+                # cutlass_w8a8 requires azp to be folded into azp_adj
+                # in the per-tensor case
+                azp_adj = layer.input_zero_point * azp_adj
+            layer.azp_adj = Parameter(azp_adj, requires_grad=False)
+        else:
+            layer.azp_adj = None
+
+    def apply_weights(
+        self, layer: torch.nn.Module, x: torch.Tensor, bias: Optional[torch.Tensor]
+    ) -> torch.Tensor:
+        # TODO: add cutlass_scaled_mm_azp support
+        x_q, x_scale = per_token_quant_int8(x)
+
+        return int8_scaled_mm(
+            x_q, layer.weight, x_scale, layer.weight_scale, out_dtype=x.dtype, bias=bias
+        )
+
+
+class NPUCompressedTensorsW8A8Int8(CompressedTensorsW8A8Int8):
+
+    def __init__(
+        self, strategy: str, is_static_input_scheme: bool, input_symmetric: bool
+    ):
+        super().__init__(strategy, is_static_input_scheme, input_symmetric)
+        # TODO: Currently, NPU kernel for static quant requires quant_bias field,
+        # which can't be replicated in compressed-tensors.
+        if self.is_static_input_scheme:
+            raise NotImplementedError(
+                "Static compressed-tensors scheme is not yet supported on NPU."
+            )
+        self.kernel = NPUW8A8Int8DynamicLinearMethod()
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        return NotImplementedError
+
+    def process_weights_after_loading(self, layer):
+        return self.kernel.process_weights_after_loading(layer)
+
+    def apply_weights(self, layer, x, bias):
+        return self.kernel.apply(layer, x, bias)

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_int8_moe.py

@@ -0,0 +1,154 @@
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING
+
+import torch
+from compressed_tensors.quantization import QuantizationStrategy
+
+from sglang.srt.hardware_backend.npu.quantization.fused_moe_method_npu import (
+    NPUW8A8Int8DynamicMoEMethod,
+)
+from sglang.srt.layers.moe import MoeRunnerConfig
+from sglang.srt.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsMoEScheme,
+)
+from sglang.srt.utils import set_weight_attrs
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
+
+__all__ = ["NPUCompressedTensorsW8A8Int8DynamicMoE"]
+
+logger = logging.getLogger(__name__)
+
+
+class NPUCompressedTensorsW8A8Int8DynamicMoE(CompressedTensorsMoEScheme):
+
+    def __init__(self, weight_quant, input_quant):
+        self.weight_quant = weight_quant
+        self.input_quant = input_quant
+        self.kernel = NPUW8A8Int8DynamicMoEMethod()
+
+        self.static_input_scales = not self.input_quant.dynamic
+        per_channel = (
+            self.weight_quant.strategy == QuantizationStrategy.CHANNEL
+            and self.input_quant.strategy == QuantizationStrategy.TOKEN
+        )
+        if not per_channel:
+            raise ValueError(
+                "For INT8 Fused MoE layers, we require channelwise, "
+                "dynamic per token quantization. Found "
+                f"{self.weight_quant}, {self.input_quant}"
+            )
+
+        self.static_input_scales = not self.input_quant.dynamic
+        if self.static_input_scales:
+            raise ValueError(
+                "For INT8 Fused MoE layers, we require channelwise, "
+                "dynamic per token quantization. Found static input scales."
+            )
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoeWeightScaleSupported
+
+        params_dtype = torch.int8
+
+        # WEIGHTS
+        w13_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                2 * intermediate_size_per_partition,
+                hidden_size,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight", w13_weight)
+        set_weight_attrs(w13_weight, extra_weight_attrs)
+
+        w2_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                intermediate_size_per_partition,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight", w2_weight)
+        set_weight_attrs(w2_weight, extra_weight_attrs)
+
+        # WEIGHT_SCALES
+        assert self.weight_quant.strategy == QuantizationStrategy.CHANNEL
+        w13_weight_scale = torch.nn.Parameter(
+            torch.ones(
+                num_experts, 2 * intermediate_size_per_partition, 1, dtype=torch.float32
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_scale", w13_weight_scale)
+        w2_weight_scale = torch.nn.Parameter(
+            torch.ones(num_experts, hidden_size, 1, dtype=torch.float32),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_scale", w2_weight_scale)
+        # Add PER-CHANNEL quantization for FusedMoE.weight_loader.
+        extra_weight_attrs.update(
+            {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value}
+        )
+        set_weight_attrs(w13_weight_scale, extra_weight_attrs)
+        set_weight_attrs(w2_weight_scale, extra_weight_attrs)
+
+        # INPUT_SCALES
+        assert not self.static_input_scales
+        layer.w13_input_scale = None
+        layer.w2_input_scale = None
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        self.kernel.process_weights_after_loading(layer)
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+
+        return self.kernel.apply(layer, dispatch_output)
+
+    def apply_without_routing_weights(
+        self,
+        layer,
+        hidden_states,
+        hidden_states_scale,
+        group_list_type,
+        group_list,
+        output_dtype,
+    ):
+        # NPU MoE bypasses MoeRunner: expose the kernel's existing
+        # apply_without_routing_weights directly through the scheme.
+        return self.kernel.apply_without_routing_weights(
+            layer,
+            hidden_states,
+            hidden_states_scale,
+            group_list_type,
+            group_list,
+            output_dtype,
+        )

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16.py

@@ -0,0 +1,340 @@
+# Adapted from https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/quantization/compressed_tensors
+# SPDX-License-Identifier: Apache-2.0
+
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import logging
+from typing import Callable, Optional
+
+import torch
+from compressed_tensors.quantization import ActivationOrdering
+
+# yapf conflicts with isort for this block
+# yapf: disable
+from sglang.srt.layers.parameter import (
+    BasevLLMParameter,
+    ChannelQuantScaleParameter,
+    GroupQuantScaleParameter,
+    PackedColumnParameter,
+    PackedvLLMParameter,
+    RowvLLMParameter,
+    permute_param_layout_,
+)
+from sglang.srt.layers.quantization.compressed_tensors.schemes import (
+    CompressedTensorsLinearScheme,
+)
+from sglang.srt.layers.quantization.marlin_utils import (
+    MarlinLinearLayerConfig,
+    apply_gptq_marlin_linear,
+    check_marlin_supports_shape,
+    marlin_is_k_full,
+    marlin_make_empty_g_idx,
+    marlin_make_workspace,
+    marlin_permute_scales,
+    marlin_repeat_scales_on_all_ranks,
+    marlin_sort_g_idx,
+    marlin_zero_points,
+)
+from sglang.srt.layers.quantization.utils import (
+    get_scalar_types,
+    replace_parameter,
+    unpack_cols,
+)
+from sglang.srt.utils import is_cuda
+
+_is_cuda = is_cuda()
+
+if _is_cuda:
+    from sglang.jit_kernel.gptq_marlin_repack import gptq_marlin_repack
+
+
+ScalarType, scalar_types = get_scalar_types()
+
+logger = logging.getLogger(__name__)
+
+__all__ = ["CompressedTensorsWNA16"]
+WNA16_SUPPORTED_TYPES_MAP = {
+    4: scalar_types.uint4b8,
+    8: scalar_types.uint8b128
+}
+WNA16_ZP_SUPPORTED_TYPES_MAP = {4: scalar_types.uint4, 8: scalar_types.uint8}
+WNA16_SUPPORTED_BITS = list(WNA16_SUPPORTED_TYPES_MAP.keys())
+
+
+class CompressedTensorsWNA16(CompressedTensorsLinearScheme):
+    _kernel_backends_being_used: set[str] = set()
+
+    def __init__(self,
+                 strategy: str,
+                 num_bits: int,
+                 group_size: Optional[int] = None,
+                 symmetric: Optional[bool] = True,
+                 actorder: Optional[ActivationOrdering] = None):
+
+        self.pack_factor = 32 // num_bits
+        self.strategy = strategy
+        self.symmetric = symmetric
+        self.group_size = -1 if group_size is None else group_size
+        self.has_g_idx = actorder == ActivationOrdering.GROUP
+
+        if self.group_size == -1 and self.strategy != "channel":
+            raise ValueError("Marlin kernels require group quantization or "
+                             "channelwise quantization, but found no group "
+                             "size and strategy is not channelwise.")
+
+        if num_bits not in WNA16_SUPPORTED_TYPES_MAP:
+            raise ValueError(
+                f"Unsupported num_bits = {num_bits}. "
+                f"Supported num_bits = {WNA16_SUPPORTED_TYPES_MAP.keys()}")
+
+        self.quant_type = (WNA16_ZP_SUPPORTED_TYPES_MAP[num_bits]
+                           if not self.symmetric else
+                           WNA16_SUPPORTED_TYPES_MAP[num_bits])
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # ampere and up
+        return 80
+
+    def create_weights(self, layer: torch.nn.Module, output_size: int,
+                       input_size: int, output_partition_sizes: list[int],
+                       input_size_per_partition: int,
+                       params_dtype: torch.dtype, weight_loader: Callable,
+                       **kwargs):
+
+        output_size_per_partition = sum(output_partition_sizes)
+
+        self.kernel_config = MarlinLinearLayerConfig(
+            full_weight_shape=(input_size, output_size),
+            partition_weight_shape=(
+                input_size_per_partition,
+                output_size_per_partition,
+            ),
+            weight_type=self.quant_type,
+            act_type=params_dtype,
+            group_size=self.group_size,
+            zero_points=not self.symmetric,
+            has_g_idx=self.has_g_idx
+        )
+
+        # If group_size is -1, we are in channelwise case.
+        group_size = self.group_size if self.group_size != -1 else input_size
+        row_parallel = (input_size != input_size_per_partition)
+        partition_scales = not marlin_repeat_scales_on_all_ranks(
+            self.has_g_idx, self.group_size, row_parallel)
+
+        scales_and_zp_size = input_size // group_size
+
+        if partition_scales:
+            assert input_size_per_partition % group_size == 0
+            scales_and_zp_size = input_size_per_partition // group_size
+
+        weight = PackedvLLMParameter(input_dim=1,
+                                     output_dim=0,
+                                     weight_loader=weight_loader,
+                                     packed_factor=self.pack_factor,
+                                     packed_dim=1,
+                                     data=torch.empty(
+                                         output_size_per_partition,
+                                         input_size_per_partition //
+                                         self.pack_factor,
+                                         dtype=torch.int32,
+                                     ))
+
+        weight_scale_args = {
+            "weight_loader":
+            weight_loader,
+            "data":
+            torch.empty(
+                output_size_per_partition,
+                scales_and_zp_size,
+                dtype=params_dtype,
+            )
+        }
+
+        zeros_args = {
+            "weight_loader":
+            weight_loader,
+            "data":
+            torch.zeros(
+                output_size_per_partition // self.pack_factor,
+                scales_and_zp_size,
+                dtype=torch.int32,
+            )
+        }
+
+        if not partition_scales:
+            weight_scale = ChannelQuantScaleParameter(output_dim=0,
+                                                      **weight_scale_args)
+
+            if not self.symmetric:
+                qzeros = PackedColumnParameter(output_dim=0,
+                                               packed_dim=0,
+                                               packed_factor=self.pack_factor,
+                                               **zeros_args)
+        else:
+            weight_scale = GroupQuantScaleParameter(output_dim=0,
+                                                    input_dim=1,
+                                                    **weight_scale_args)
+            if not self.symmetric:
+                qzeros = PackedvLLMParameter(input_dim=1,
+                                             output_dim=0,
+                                             packed_dim=0,
+                                             packed_factor=self.pack_factor,
+                                             **zeros_args)
+
+        # A 2D array defining the original shape of the weights
+        # before packing
+        weight_shape = BasevLLMParameter(data=torch.empty(2,
+                                                          dtype=torch.int64),
+                                         weight_loader=weight_loader)
+
+        layer.register_parameter("weight_packed", weight)
+        layer.register_parameter("weight_scale", weight_scale)
+        layer.register_parameter("weight_shape", weight_shape)
+
+        if not self.symmetric:
+            layer.register_parameter("weight_zero_point", qzeros)
+
+        # group index (for activation reordering)
+        if self.has_g_idx:
+            weight_g_idx = RowvLLMParameter(data=torch.empty(
+                input_size_per_partition,
+                dtype=torch.int32,
+            ),
+                                            input_dim=0,
+                                            weight_loader=weight_loader)
+            layer.register_parameter("weight_g_idx", weight_g_idx)
+
+    # Checkpoints are serialized in compressed-tensors format, which is
+    # different from the format the kernel may want. Handle repacking here.
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        # Default names since marlin requires empty parameters for these,
+        # TODO: remove this requirement from marlin (allow optional tensors)
+        self.w_q_name = "weight_packed"
+        self.w_s_name = "weight_scale"
+        self.w_zp_name = "weight_zero_point"
+        self.w_gidx_name = "weight_g_idx"
+
+        device = getattr(layer, self.w_q_name).device
+        c = self.kernel_config
+
+        check_marlin_supports_shape(
+            c.partition_weight_shape[1],  # out_features
+            c.partition_weight_shape[0],  # in_features
+            c.full_weight_shape[0],  # in_features
+            c.group_size,
+        )
+
+        row_parallel = c.partition_weight_shape[0] != c.full_weight_shape[0]
+        self.is_k_full = marlin_is_k_full(c.has_g_idx, row_parallel)
+
+        # Allocate marlin workspace.
+        self.workspace = marlin_make_workspace(device)
+
+        def _transform_param(
+            layer: torch.nn.Module, name: Optional[str], fn: Callable
+        ) -> None:
+            if name is not None and getattr(layer, name, None) is not None:
+
+                old_param = getattr(layer, name)
+                new_param = fn(old_param)
+                # replace the parameter with torch.nn.Parameter for TorchDynamo
+                # compatibility
+                replace_parameter(
+                    layer, name, torch.nn.Parameter(new_param.data, requires_grad=False)
+                )
+
+        def transform_w_q(x):
+            assert isinstance(x, BasevLLMParameter)
+            permute_param_layout_(x, input_dim=0, output_dim=1, packed_dim=0)
+            x.data = gptq_marlin_repack(
+                x.data.contiguous(),
+                perm=layer.g_idx_sort_indices,
+                size_k=c.partition_weight_shape[0],
+                size_n=c.partition_weight_shape[1],
+                num_bits=c.weight_type.size_bits,
+            )
+            return x
+
+        def transform_w_s(x):
+            assert isinstance(x, BasevLLMParameter)
+            permute_param_layout_(x, input_dim=0, output_dim=1)
+            x.data = marlin_permute_scales(
+                x.data.contiguous(),
+                size_k=c.partition_weight_shape[0],
+                size_n=c.partition_weight_shape[1],
+                group_size=c.group_size,
+            )
+            return x
+
+        if c.has_g_idx:
+            g_idx, g_idx_sort_indices = marlin_sort_g_idx(
+                getattr(layer, self.w_gidx_name)
+            )
+            _transform_param(layer, self.w_gidx_name, lambda _: g_idx)
+            layer.g_idx_sort_indices = g_idx_sort_indices
+        else:
+            setattr(layer, self.w_gidx_name, marlin_make_empty_g_idx(device))
+            layer.g_idx_sort_indices = marlin_make_empty_g_idx(device)
+
+        if c.zero_points:
+            grouped_k = (
+                c.partition_weight_shape[0] // c.group_size if c.group_size != -1 else 1
+            )
+            _transform_param(
+                layer,
+                self.w_zp_name,
+                lambda x: marlin_zero_points(
+                    unpack_cols(
+                        x.t(),
+                        c.weight_type.size_bits,
+                        grouped_k,
+                        c.partition_weight_shape[1],
+                    ),
+                    size_k=grouped_k,
+                    size_n=c.partition_weight_shape[1],
+                    num_bits=c.weight_type.size_bits,
+                ),
+            )
+        else:
+            setattr(layer, self.w_zp_name, marlin_make_empty_g_idx(device))
+        _transform_param(layer, self.w_q_name, transform_w_q)
+        _transform_param(layer, self.w_s_name, transform_w_s)
+
+    def apply_weights(self, layer: torch.nn.Module, x: torch.Tensor,
+                      bias: Optional[torch.Tensor]) -> torch.Tensor:
+        c = self.kernel_config
+
+        def _get_weight_params(
+            layer: torch.nn.Module,
+        ) -> tuple[
+            torch.Tensor,  # w_q
+            torch.Tensor,  # w_s
+            Optional[torch.Tensor],  # w_zp,
+            Optional[torch.Tensor],  # w_gidx
+        ]:
+            return (
+                getattr(layer, self.w_q_name),
+                getattr(layer, self.w_s_name),
+                getattr(layer, self.w_zp_name or "", None),
+                getattr(layer, self.w_gidx_name or "", None),
+            )
+
+        w_q, w_s, w_zp, w_gidx = _get_weight_params(layer)
+
+        # `process_weights_after_loading` will ensure w_zp and w_gidx are not
+        #  None for marlin
+        return apply_gptq_marlin_linear(
+            input=x,
+            weight=w_q,
+            weight_scale=w_s,
+            weight_zp=w_zp,  # type: ignore
+            g_idx=w_gidx,  # type: ignore
+            g_idx_sort_indices=layer.g_idx_sort_indices,
+            workspace=self.workspace,
+            wtype=c.weight_type,
+            input_size_per_partition=c.partition_weight_shape[0],
+            output_size_per_partition=c.partition_weight_shape[1],
+            is_k_full=self.is_k_full,
+            bias=bias,
+        )

python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16_moe.py

@@ -0,0 +1,727 @@
+from __future__ import annotations
+
+import enum
+import logging
+from enum import Enum
+from typing import TYPE_CHECKING
+
+import torch
+from compressed_tensors import CompressionFormat
+
+from sglang.srt.hardware_backend.gpu.quantization.gptq_kernels import (
+    gptq_marlin_moe_repack,
+)
+from sglang.srt.hardware_backend.npu.quantization.fused_moe_method_npu import (
+    NPUW4A16Int4DynamicMoEMethod,
+)
+from sglang.srt.layers.moe import MoeRunner, MoeRunnerBackend, MoeRunnerConfig
+from sglang.srt.layers.quantization.compressed_tensors.schemes import (
+    WNA16_SUPPORTED_BITS,
+    CompressedTensorsMoEScheme,
+)
+from sglang.srt.layers.quantization.marlin_utils import (
+    marlin_make_workspace,
+    marlin_moe_permute_scales,
+    moe_awq_to_marlin_zero_points,
+)
+from sglang.srt.layers.quantization.utils import replace_parameter
+from sglang.srt.utils import get_bool_env_var, is_cuda, is_hip, set_weight_attrs
+
+if TYPE_CHECKING:
+    from compressed_tensors.quantization import QuantizationArgs
+
+    from sglang.srt.layers.moe.token_dispatcher import (
+        CombineInput,
+        StandardDispatchOutput,
+    )
+    from sglang.srt.layers.quantization.compressed_tensors.compressed_tensors import (
+        CompressedTensorsConfig,
+    )
+
+
+__all__ = [
+    "CompressedTensorsWNA16MoE",
+    "CompressedTensorsWNA16TritonMoE",
+    "NPUCompressedTensorsW4A16Int4DynamicMoE",
+]
+
+_is_hip = is_hip()
+_is_cuda = is_cuda()
+
+_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
+
+if _use_aiter:
+    pass
+
+
+logger = logging.getLogger(__name__)
+
+
+class GPTQMarlinState(Enum):
+    REPACK = enum.auto()
+    READY = enum.auto()
+
+
+class CompressedTensorsWNA16MoE(CompressedTensorsMoEScheme):
+
+    def __init__(
+        self,
+        quant_config: CompressedTensorsConfig,
+        weight_quant: QuantizationArgs,
+        num_gpu_experts: int = -1,
+    ):
+        self.quant_config = quant_config
+        # Per-layer scheme already resolved by get_moe_scheme(); reuse it directly
+        # (mixed-precision MoE has no "Linear" config group to fall back on).
+        config = weight_quant
+        self.num_bits = config.num_bits
+        self.packed_factor = 32 // config.num_bits
+        self.strategy = config.strategy
+        self.group_size = config.group_size
+        self.actorder = config.actorder
+        self.sym = config.symmetric
+
+        if not (
+            self.quant_config.quant_format == CompressionFormat.pack_quantized.value
+            and self.num_bits in WNA16_SUPPORTED_BITS
+        ):
+            raise ValueError(
+                "For Fused MoE layers, only ",
+                f"{CompressionFormat.pack_quantized.value} ",
+                "is supported for the following bits: ",
+                f"{WNA16_SUPPORTED_BITS}",
+            )
+        self.num_gpu_experts = num_gpu_experts
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # ampere and up
+        return 80
+
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ):
+        # Will transpose the loaded weight along the
+        # intermediate and hidden dim sizes. Will
+        # shard for TP along the transposed dims
+        extra_weight_attrs.update(
+            {"is_transposed": True, "quant_method": self.strategy}
+        )
+        w13_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size // self.packed_factor,
+                2 * intermediate_size_per_partition,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_packed", w13_weight)
+        set_weight_attrs(w13_weight, extra_weight_attrs)
+
+        w2_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                intermediate_size_per_partition // self.packed_factor,
+                hidden_size,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_packed", w2_weight)
+        set_weight_attrs(w2_weight, extra_weight_attrs)
+
+        # In the case where we have actorder/g_idx,
+        # we do not partition the w2 scales
+        load_full_w2 = (
+            self.actorder is not None
+            and self.actorder != "static"
+            and self.group_size != -1
+        )
+
+        if load_full_w2:
+            w2_scales_size = intermediate_size_per_partition * layer.moe_tp_size
+        else:
+            w2_scales_size = intermediate_size_per_partition
+
+        self.is_k_full = (not self.actorder) or layer.moe_tp_size == 1
+
+        if self.strategy == "channel":
+            num_groups_w2 = num_groups_w13 = 1
+            self.group_size = -1
+        else:
+            num_groups_w2 = w2_scales_size // self.group_size
+            num_groups_w13 = hidden_size // self.group_size
+
+        w13_scale = torch.nn.Parameter(
+            torch.ones(
+                num_experts,
+                num_groups_w13,
+                2 * intermediate_size_per_partition,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_scale", w13_scale)
+        set_weight_attrs(w13_scale, extra_weight_attrs)
+
+        w2_scale = torch.nn.Parameter(
+            torch.ones(num_experts, num_groups_w2, hidden_size, dtype=params_dtype),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_scale", w2_scale)
+        set_weight_attrs(w2_scale, extra_weight_attrs)
+        set_weight_attrs(w2_scale, {"load_full_w2": load_full_w2})
+
+        w2_weight_shape = torch.nn.Parameter(
+            torch.empty(num_experts, 2), requires_grad=False
+        )
+        layer.register_parameter("w2_weight_shape", w2_weight_shape)
+        set_weight_attrs(w2_weight_shape, extra_weight_attrs)
+        w13_weight_shape = torch.nn.Parameter(
+            torch.empty(num_experts, 2), requires_grad=False
+        )
+
+        layer.register_parameter("w13_weight_shape", w13_weight_shape)
+        set_weight_attrs(w13_weight_shape, extra_weight_attrs)
+
+        # add zero param
+        if not self.sym:
+            w13_qzeros = torch.nn.Parameter(
+                torch.empty(
+                    num_experts,
+                    num_groups_w13,
+                    2 * intermediate_size_per_partition // self.packed_factor,
+                    dtype=torch.int32,
+                ),
+                requires_grad=False,
+            )
+            layer.register_parameter("w13_weight_zero_point", w13_qzeros)
+            set_weight_attrs(w13_qzeros, extra_weight_attrs)
+
+            w2_qzeros = torch.nn.Parameter(
+                torch.empty(
+                    num_experts,
+                    num_groups_w2,
+                    hidden_size // self.packed_factor,
+                    dtype=torch.int32,
+                ),
+                requires_grad=False,
+            )
+            layer.register_parameter("w2_weight_zero_point", w2_qzeros)
+            set_weight_attrs(w2_qzeros, extra_weight_attrs)
+
+        w13_g_idx = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_g_idx", w13_g_idx)
+        set_weight_attrs(w13_g_idx, extra_weight_attrs)
+
+        w2_g_idx = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                intermediate_size_per_partition,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_g_idx", w2_g_idx)
+        set_weight_attrs(w2_g_idx, extra_weight_attrs)
+
+        w13_g_idx_sort_indices = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_g_idx_sort_indices", w13_g_idx_sort_indices)
+        set_weight_attrs(w13_g_idx_sort_indices, extra_weight_attrs)
+
+        w2_g_idx_sort_indices = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                intermediate_size_per_partition,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_g_idx_sort_indices", w2_g_idx_sort_indices)
+        set_weight_attrs(w2_g_idx_sort_indices, extra_weight_attrs)
+
+        layer.a13_scale = None
+        layer.a2_scale = None
+        layer.marlin_state = GPTQMarlinState.REPACK
+
+        if not hasattr(layer, "_original_shapes"):
+            layer._original_shapes = {}
+
+        # Force record: these are the target GPTQ shapes for rollback.
+        layer._original_shapes["w13_weight_packed"] = tuple(w13_weight.shape)
+        layer._original_shapes["w2_weight_packed"] = tuple(w2_weight.shape)
+
+        # Also record the shapes of the scales.
+        layer._original_shapes["w2_weight_scale"] = tuple(w2_scale.shape)
+        layer._original_shapes["w13_weight_scale"] = tuple(w13_scale.shape)
+
+        if not self.sym:
+            layer._original_shapes["w13_weight_zero_point"] = w13_qzeros.shape
+            layer._original_shapes["w2_weight_zero_point"] = tuple(w2_qzeros.shape)
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+
+        # Skip if the layer is already converted to Marlin format to prevent double-packing.
+        if getattr(layer, "is_marlin_converted", False):
+            return
+
+        if not hasattr(layer, "_original_shapes"):
+            layer._original_shapes = {}
+
+        def replace_tensor(name, new_t):
+            target_attr = getattr(layer, name)
+
+            # Only save if the key doesn't exist to prevent overwriting with Marlin shapes.
+            if name not in layer._original_shapes:
+                # This is a safety check; `create_weights` usually handles this already.
+                layer._original_shapes[name] = tuple(target_attr.shape)
+
+            # It is important to use resize_() here since it ensures
+            # the same buffer is reused
+            target_attr.resize_(new_t.shape)
+            target_attr.copy_(new_t)
+            del new_t
+
+        num_experts = layer.w13_weight_g_idx.shape[0]
+        device = layer.w13_weight_g_idx.device
+
+        # when running models with grouped act order,
+        # resort to g_idx values provided in checkpoint
+        if self.actorder == "group":
+            w13_g_idx_sort_indices = torch.empty_like(layer.w13_weight_g_idx)
+            w2_g_idx_sort_indices = torch.empty_like(layer.w2_weight_g_idx)
+            w13_sorted_g_idx = torch.empty_like(layer.w13_weight_g_idx)
+            w2_sorted_g_idx = torch.empty_like(layer.w2_weight_g_idx)
+
+            for e in range(num_experts):
+                w13_g_idx_sort_indices[e] = torch.argsort(layer.w13_weight_g_idx[e]).to(
+                    torch.int32
+                )
+                w2_g_idx_sort_indices[e] = torch.argsort(layer.w2_weight_g_idx[e]).to(
+                    torch.int32
+                )
+                w13_sorted_g_idx[e] = layer.w13_weight_g_idx[e][
+                    w13_g_idx_sort_indices[e]
+                ]
+                w2_sorted_g_idx[e] = layer.w2_weight_g_idx[e][w2_g_idx_sort_indices[e]]
+
+            replace_parameter(layer, "w13_weight_g_idx", w13_sorted_g_idx)
+            replace_parameter(layer, "w2_weight_g_idx", w2_sorted_g_idx)
+            replace_parameter(layer, "w13_g_idx_sort_indices", w13_g_idx_sort_indices)
+            replace_parameter(layer, "w2_g_idx_sort_indices", w2_g_idx_sort_indices)
+
+        else:
+            layer.w13_weight_g_idx = torch.nn.Parameter(
+                torch.empty((num_experts, 0), dtype=torch.int32, device=device),
+                requires_grad=False,
+            )
+            layer.w2_weight_g_idx = torch.nn.Parameter(
+                torch.empty((num_experts, 0), dtype=torch.int32, device=device),
+                requires_grad=False,
+            )
+            layer.w13_g_idx_sort_indices = torch.nn.Parameter(
+                torch.empty((num_experts, 0), dtype=torch.int32, device=device),
+                requires_grad=False,
+            )
+            layer.w2_g_idx_sort_indices = torch.nn.Parameter(
+                torch.empty((num_experts, 0), dtype=torch.int32, device=device),
+                requires_grad=False,
+            )
+
+        marlin_w13_qweight = gptq_marlin_moe_repack(
+            layer.w13_weight_packed,
+            layer.w13_g_idx_sort_indices,
+            layer.w13_weight_packed.shape[1] * self.packed_factor,
+            layer.w13_weight_packed.shape[2],
+            self.num_bits,
+        )
+        replace_tensor("w13_weight_packed", marlin_w13_qweight)
+        marlin_w2_qweight = gptq_marlin_moe_repack(
+            layer.w2_weight_packed,
+            layer.w2_g_idx_sort_indices,
+            layer.w2_weight_packed.shape[1] * self.packed_factor,
+            layer.w2_weight_packed.shape[2],
+            self.num_bits,
+        )
+        replace_tensor("w2_weight_packed", marlin_w2_qweight)
+        # Repack scales
+        marlin_w13_scales = marlin_moe_permute_scales(
+            layer.w13_weight_scale,
+            layer.w13_weight_packed.shape[2],
+            layer.w13_weight_scale.shape[2],
+            self.group_size,
+        )
+        replace_tensor("w13_weight_scale", marlin_w13_scales)
+
+        marlin_w2_scales = marlin_moe_permute_scales(
+            layer.w2_weight_scale,
+            layer.w2_weight_scale.shape[1]
+            * (self.group_size if self.group_size != -1 else self.packed_factor),
+            layer.w2_weight_scale.shape[2],
+            self.group_size,
+        )
+        replace_tensor("w2_weight_scale", marlin_w2_scales)
+
+        # Repack zero
+        if not self.sym:
+            marlin_w13_zp = moe_awq_to_marlin_zero_points(
+                layer.w13_weight_zero_point,
+                size_k=layer.w13_weight_zero_point.shape[1],
+                size_n=layer.w13_weight_zero_point.shape[2] * self.packed_factor,
+                num_bits=self.num_bits,
+            )
+            replace_tensor("w13_weight_zero_point", marlin_w13_zp)
+
+            marlin_w2_zp = moe_awq_to_marlin_zero_points(
+                layer.w2_weight_zero_point,
+                size_k=layer.w2_weight_zero_point.shape[1],
+                size_n=layer.w2_weight_zero_point.shape[2] * self.packed_factor,
+                num_bits=self.num_bits,
+            )
+            replace_tensor("w2_weight_zero_point", marlin_w2_zp)
+
+        layer.workspace = marlin_make_workspace(layer.w13_weight_packed.device, 4)
+        layer.is_marlin_converted = True
+
+    def restore_weights_before_loading(self, layer: torch.nn.Module):
+        """Forcibly resize parameters back to their original shapes (e.g., GPTQ format) before loading weights."""
+
+        if not hasattr(layer, "_original_shapes"):
+            return
+
+        for name, orig_shape in layer._original_shapes.items():
+            param = getattr(layer, name, None)
+
+            if param is not None and param.shape != orig_shape:
+                param.resize_(orig_shape)
+
+        layer.is_marlin_converted = False
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+        self.runner = MoeRunner(MoeRunnerBackend.MARLIN, moe_runner_config)
+
+    def get_marlin_quant_info(self, layer):
+        from sglang.srt.layers.moe.moe_runner.marlin import MarlinMoeQuantInfo
+
+        return MarlinMoeQuantInfo(
+            w13_qweight=layer.w13_weight_packed,
+            w2_qweight=layer.w2_weight_packed,
+            w13_scales=layer.w13_weight_scale,
+            w2_scales=layer.w2_weight_scale,
+            w13_g_idx_sort_indices=getattr(layer, "w13_g_idx_sort_indices", None),
+            w2_g_idx_sort_indices=getattr(layer, "w2_g_idx_sort_indices", None),
+            weight_bits=self.num_bits,
+            w13_g_idx=getattr(layer, "w13_weight_g_idx", None),
+            w2_g_idx=getattr(layer, "w2_weight_g_idx", None),
+            is_k_full=self.is_k_full,
+            w13_qzeros=layer.w13_weight_zero_point if not self.sym else None,
+            w2_qzeros=layer.w2_weight_zero_point if not self.sym else None,
+        )
+
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+        from sglang.srt.layers.moe.fused_moe_triton.fused_marlin_moe import (
+            fused_marlin_moe,
+        )
+        from sglang.srt.layers.moe.token_dispatcher import StandardCombineInput
+
+        assert (
+            self.moe_runner_config.activation == "silu"
+        ), "Only SiLU activation is supported."
+
+        x = dispatch_output.hidden_states
+        topk_output = dispatch_output.topk_output
+
+        topk_weights, topk_ids, router_logits = topk_output
+
+        # Get expert_map for EP support
+        expert_map = None
+        global_num_experts = -1
+        if hasattr(layer, "dispatcher") and hasattr(
+            layer.dispatcher, "local_expert_mapping"
+        ):
+            expert_map = layer.dispatcher.local_expert_mapping
+            if expert_map is not None:
+                global_num_experts = self.moe_runner_config.num_experts
+
+        output = fused_marlin_moe(
+            x,
+            layer.w13_weight_packed,
+            layer.w2_weight_packed,
+            layer.w13_weight_scale,
+            layer.w2_weight_scale,
+            router_logits,
+            topk_weights,
+            topk_ids,
+            global_num_experts=global_num_experts,
+            expert_map=expert_map,
+            g_idx1=layer.w13_weight_g_idx,
+            g_idx2=layer.w2_weight_g_idx,
+            sort_indices1=layer.w13_g_idx_sort_indices,
+            sort_indices2=layer.w2_g_idx_sort_indices,
+            w1_zeros=layer.w13_weight_zero_point if not self.sym else None,
+            w2_zeros=layer.w2_weight_zero_point if not self.sym else None,
+            num_bits=self.num_bits,
+            is_k_full=self.is_k_full,
+            routed_scaling_factor=self.moe_runner_config.routed_scaling_factor,
+            clamp_limit=self.moe_runner_config.swiglu_limit,
+            workspace=layer.workspace,
+        )
+        return StandardCombineInput(hidden_states=output)
+
+
+class CompressedTensorsWNA16TritonMoE(CompressedTensorsWNA16MoE):
+    """ROCm/HIP-compatible W4A16 MoE method using Triton kernels instead of Marlin.
+
+    Inherits weight creation from CompressedTensorsWNA16MoE but converts
+    weights to the uint8-packed format expected by the Triton fused MoE kernel
+    instead of the Marlin-specific format.
+    """
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        if getattr(layer, "is_triton_converted", False):
+            return
+
+        num_experts = layer.w13_weight_packed.shape[0]
+
+        # Convert w13 weights: [E, K//8, N] int32 -> [E, N, K//2] uint8
+        w13 = layer.w13_weight_packed.data
+        w13 = w13.transpose(1, 2).contiguous().view(torch.uint8)
+        layer.w13_weight_packed = torch.nn.Parameter(w13, requires_grad=False)
+
+        # Convert w2 weights: [E, K//8, N] int32 -> [E, N, K//2] uint8
+        w2 = layer.w2_weight_packed.data
+        w2 = w2.transpose(1, 2).contiguous().view(torch.uint8)
+        layer.w2_weight_packed = torch.nn.Parameter(w2, requires_grad=False)
+
+        # Convert w13 scales: [E, K//group_size, N] -> [E, N, K//group_size]
+        w13_scale = layer.w13_weight_scale.data
+        w13_scale = w13_scale.transpose(1, 2).contiguous()
+        layer.w13_weight_scale = torch.nn.Parameter(w13_scale, requires_grad=False)
+
+        # Convert w2 scales: [E, K//group_size, N] -> [E, N, K//group_size]
+        w2_scale = layer.w2_weight_scale.data
+        w2_scale = w2_scale.transpose(1, 2).contiguous()
+        layer.w2_weight_scale = torch.nn.Parameter(w2_scale, requires_grad=False)
+
+        layer.is_triton_converted = True
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+        self.runner = MoeRunner(MoeRunnerBackend.TRITON, moe_runner_config)
+
+    def get_triton_quant_info(self, layer):
+        from sglang.srt.layers.moe.moe_runner.triton import TritonMoeQuantInfo
+
+        return TritonMoeQuantInfo(
+            w13_weight=layer.w13_weight_packed,
+            w2_weight=layer.w2_weight_packed,
+            use_int4_w4a16=True,
+            w13_scale=layer.w13_weight_scale,
+            w2_scale=layer.w2_weight_scale,
+            block_shape=[0, self.group_size],
+        )
+
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+        assert (
+            self.moe_runner_config.activation == "silu"
+        ), "Only SiLU activation is supported."
+
+        quant_info = self.get_triton_quant_info(layer)
+        return self.runner.run(dispatch_output, quant_info)
+
+
+class NPUCompressedTensorsW4A16Int4DynamicMoE(CompressedTensorsMoEScheme):
+
+    def __init__(self, quantization_config) -> None:
+        self.pack_factor = 8  # weight dtype is int4,  but use int32 to create
+        target = (
+            "MoEGMM" if "MoEGMM" in quantization_config.target_scheme_map else "Linear"
+        )
+        if target in quantization_config.target_scheme_map:
+            self.group_size = quantization_config.target_scheme_map[target][
+                "weights"
+            ].group_size
+        else:
+            self.group_size = 128
+
+        self.kernel = NPUW4A16Int4DynamicMoEMethod()
+
+    # TODO: See if we can merge this method's logic
+    # with CompressedTensorsWNA16MoE. Need more models and tests.
+    # @OrangeRedeng @TamirBaydasov
+    def create_weights(
+        self,
+        layer: torch.nn.Module,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size_per_partition: int,
+        params_dtype: torch.dtype,
+        **extra_weight_attrs,
+    ) -> None:
+        from sglang.srt.layers.moe.fused_moe_triton import FusedMoeWeightScaleSupported
+
+        self.num_experts = num_experts
+        if (
+            extra_weight_attrs.get(
+                "moe_intermediate_size", intermediate_size_per_partition
+            )
+            // intermediate_size_per_partition
+            > 1
+        ):
+            quant_method = FusedMoeWeightScaleSupported.GROUP.value
+        else:
+            quant_method = FusedMoeWeightScaleSupported.CHANNEL.value
+        extra_weight_attrs.update({"quant_method": quant_method})
+        # weight
+        w13_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                2 * intermediate_size_per_partition,
+                hidden_size // self.pack_factor,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight", w13_weight)
+        set_weight_attrs(w13_weight, extra_weight_attrs)
+        w2_weight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                intermediate_size_per_partition // self.pack_factor,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight", w2_weight)
+        set_weight_attrs(w2_weight, extra_weight_attrs)
+
+        # scale
+        weight_scale_dtype = torch.bfloat16
+        w13_weight_scale = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                2 * intermediate_size_per_partition,
+                hidden_size // self.group_size,
+                dtype=weight_scale_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_scale", w13_weight_scale)
+        set_weight_attrs(w13_weight_scale, extra_weight_attrs)
+        w2_weight_scale = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                intermediate_size_per_partition // self.group_size,
+                dtype=weight_scale_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_scale", w2_weight_scale)
+        set_weight_attrs(w2_weight_scale, extra_weight_attrs)
+
+        # offset
+        w13_weight_offset = torch.nn.Parameter(
+            torch.zeros(
+                num_experts,
+                2 * intermediate_size_per_partition,
+                hidden_size // self.group_size,
+                dtype=weight_scale_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_weight_offset", w13_weight_offset)
+        set_weight_attrs(w13_weight_offset, extra_weight_attrs)
+
+        w2_weight_offset = torch.nn.Parameter(
+            torch.zeros(
+                num_experts,
+                hidden_size,
+                intermediate_size_per_partition // self.group_size,
+                dtype=weight_scale_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_weight_offset", w2_weight_offset)
+        set_weight_attrs(w2_weight_offset, extra_weight_attrs)
+
+        w13_weight_shape = torch.nn.Parameter(
+            torch.empty(num_experts, 2), requires_grad=False
+        )
+        layer.register_parameter("w13_weight_shape", w13_weight_shape)
+        set_weight_attrs(w13_weight_shape, extra_weight_attrs)
+
+        w2_weight_shape = torch.nn.Parameter(
+            torch.empty(num_experts, 2), requires_grad=False
+        )
+        layer.register_parameter("w2_weight_shape", w2_weight_shape)
+        set_weight_attrs(w2_weight_shape, extra_weight_attrs)
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        self.kernel.process_weights_after_loading(layer)
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.moe_runner_config = moe_runner_config
+
+    def apply_weights(
+        self,
+        layer: torch.nn.Module,
+        dispatch_output: StandardDispatchOutput,
+    ) -> CombineInput:
+
+        return self.kernel.apply(layer, dispatch_output)
+
+    def apply_without_routing_weights(
+        self,
+        layer,
+        hidden_states,
+        hidden_states_scale,
+        group_list_type,
+        group_list,
+        output_dtype,
+    ):
+        return self.kernel.apply_without_routing_weights(
+            layer,
+            hidden_states,
+            hidden_states_scale,
+            group_list_type,
+            group_list,
+            output_dtype,
+        )

python/sglang/srt/layers/quantization/compressed_tensors/utils.py

@@ -0,0 +1,220 @@
+# Adapted from https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/layers/quantization/compressed_tensors
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import re
+from types import MappingProxyType
+from typing import Iterable, List, Mapping, Optional
+
+from compressed_tensors import CompressionFormat
+from torch.nn import Module
+
+
+def is_activation_quantization_format(format: str) -> bool:
+    _ACTIVATION_QUANTIZATION_FORMATS = [
+        CompressionFormat.naive_quantized.value,
+        CompressionFormat.int_quantized.value,
+        CompressionFormat.float_quantized.value,
+        CompressionFormat.nvfp4_pack_quantized.value,
+    ]
+    return format in _ACTIVATION_QUANTIZATION_FORMATS
+
+
+def should_ignore_layer(
+    layer_name: Optional[str],
+    ignore: Iterable[str] = tuple(),
+    fused_mapping: Mapping[str, List[str]] = MappingProxyType({}),
+) -> bool:
+    if layer_name is None:
+        return False
+
+    # layer_name = model.layers.0.self_attn.qkv_proj
+    # proj_name = qkv_proj
+    proj_name = layer_name.split(".")[-1]
+
+    # Fused layers like gate_up_proj or qkv_proj will not be fused
+    # in the safetensors checkpoint. So, we convert the name
+    # from the fused version to unfused + check to make sure that
+    # each shard of the fused layer has the same scheme.
+    if proj_name in fused_mapping and layer_name not in ignore:
+        shard_proj_names = fused_mapping[proj_name]
+
+        # Convert fused_name --> [shard_names]
+        shard_names = [
+            layer_name.replace(proj_name, shard_proj_name)
+            for shard_proj_name in shard_proj_names
+        ]
+
+        # Layer should be ignored if shards are ignored.
+        should_ignore_layer = None
+        for shard_name in shard_names:
+            should_ignore_shard = check_equal_or_regex_match(
+                layer_name=shard_name, targets=ignore
+            )
+
+            # If shard_idx=0, set layer ignore to match shard.
+            if should_ignore_layer is None:
+                should_ignore_layer = should_ignore_shard
+
+            # If shard_idx=1+ confirm scheme matches prior shards.
+            elif should_ignore_shard != should_ignore_layer:
+                raise ValueError(
+                    f"Found different quantization schemes for "
+                    f"{shard_proj_names} in {layer_name}. SGLang "
+                    "requires all to use the same scheme."
+                )
+
+    # Unfused layers like down_proj and o_proj will match
+    # the safetensors checkpoint already.
+    else:
+        should_ignore_layer = check_equal_or_regex_match(
+            layer_name=layer_name, targets=ignore
+        )
+
+    assert should_ignore_layer is not None
+    return should_ignore_layer
+
+
+def check_equal_or_regex_match(layer_name: str, targets: Iterable[str]) -> bool:
+    """
+    Checks whether a layer_name is exactly equal or a regex match for
+    if target starts with 're:' to any target in list.
+    """
+    for target in targets:
+        if _is_equal_or_regex_match(layer_name, target, check_contains=True):
+            return True
+    return False
+
+
+def find_matched_target(
+    layer_name: Optional[str],
+    module: Module,
+    targets: Iterable[str],
+    fused_mapping: Mapping[str, List[str]] = MappingProxyType({}),
+) -> str:
+    """
+    Helper function to look up which "target" in the compressed-tensors
+    config that a layer corresponds to.
+
+    Recall that a compressed-tensors configs has a concept of
+    config_groups, where each layer can be quantized with with a different
+    scheme.
+
+    targets in each config_group will be a list of either layer names
+    (or regexes corresponding to layer names) or names of torch Modules.
+
+    First, we try to match the layer_name with a target
+    Second, we try to match the module's name with a target
+    Third, we try to map the layer_name to a list of fused module names.
+        *All* component module names must match in order for a match to be
+        successful. A successful match returns the first component target
+
+    :param layer_name: layer name
+    :param module: torch.nn.Module
+    :param targets: list of targets to match the layer against
+    :param fused_mapping: map from fused layer names to its components
+    :param fused_strategy: either "all" or "any". If using "all", fused
+        layers match if "all" of its components match
+    """
+
+    if layer_name is None:
+        layer_name = ""
+
+    matched_target = (
+        _find_first_match(layer_name, targets)
+        or _find_first_match(module.__class__.__name__, targets, True)
+        or _match_fused_layer(layer_name, targets, fused_mapping)
+    )
+
+    if matched_target is None:
+        raise ValueError(
+            f"Unable to find matching target for {layer_name} in the "
+            "compressed-tensors config."
+        )
+
+    return matched_target
+
+
+def _find_first_match(
+    value: str, targets: Iterable[str], check_contains: bool = False
+) -> Optional[str]:
+    """
+    Returns first element of target that matches value either
+    exactly or as a regex after 're:'. If check_contains is set to True,
+    additionally checks if the target string is contained within the value.
+
+    :param value: string to compare the list of targets against
+    :param targets: list of targets to match the layer against
+    :param check_contains: whether or not to do a substring match
+    """
+
+    for target in targets:
+        if _is_equal_or_regex_match(value, target, check_contains=check_contains):
+            return target
+    return None
+
+
+def _is_equal_or_regex_match(
+    value: str, target: str, check_contains: bool = False
+) -> bool:
+    """
+    Checks whether a value is exactly equal or a regex match for target
+    if target starts with 're:'. If check_contains is set to True,
+    additionally checks if the target string is contained within the value.
+    """
+
+    if target.startswith("re:"):
+        pattern = target[3:]
+        if re.match(pattern, value):
+            return True
+    elif check_contains:
+        if target.lower() in value.lower():
+            return True
+    elif target == value:
+        return True
+    return False
+
+
+def _match_fused_layer(
+    layer_name: str,
+    target_layers: Iterable[str],
+    fused_mapping: Mapping[str, List[str]],
+) -> Optional[str]:
+    """
+    Match a fused layer name to its corresponding individual layer in
+    target_layers. Returns first value in fused_mapping which matches targets
+
+    Implements an "all" matching strategy where a fused layer matches iff
+    "all" of its components match
+
+    :param layer_name: layer name
+    :param target_layers: list of targets to match the layer against
+    :param fused_mapping: map from fused layer names to its components
+
+    Examples:
+        layer_name = "model.layers.0.self_attn.qkv_proj"
+        target_layers = ["model.layers.0.self_attn.q_proj",
+                        "model.layers.0.self_attn.k_proj",
+                        "model.layers.0.self_attn.v_proj"]
+    """
+    # find layer_name in mapping
+    fused = next((key for key in fused_mapping if layer_name.endswith(key)), None)
+    if fused is None:
+        return None
+
+    # expand path of unfused components
+    unfused_paths = [
+        layer_name.replace(fused, unfused) for unfused in fused_mapping[fused]
+    ]
+
+    # for each unfused component, find a match in targets
+    unfused_matches: List[Optional[str]] = []
+    for unfused in unfused_paths:
+        for target in target_layers:
+            if _is_equal_or_regex_match(unfused, target):
+                unfused_matches.append(target)
+                break
+        else:
+            unfused_matches.append(None)
+
+    return unfused_matches[0] if all(unfused_matches) else None

python/sglang/srt/layers/quantization/configs/N=1280,K=5120,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,26 @@
+{
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=1536,K=1536,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=1536,K=1536,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=1536,K=1536,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=1536,K=1536,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,26 @@
+{
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

python/sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

python/sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

python/sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,26 @@
+{
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=2048,K=4096,device_name=NVIDIA_L40,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    }
+}

python/sglang/srt/layers/quantization/configs/N=2048,K=512,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=2048,K=512,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "24": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=2048,K=512,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "96": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "128": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "256": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    }
+}

python/sglang/srt/layers/quantization/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=2048,K=512,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,26 @@
+{
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=2304,K=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=2304,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=2304,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=2304,K=7168,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,26 @@
+{
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=24576,K=1536,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

python/sglang/srt/layers/quantization/configs/N=24576,K=1536,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

python/sglang/srt/layers/quantization/configs/N=24576,K=1536,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

python/sglang/srt/layers/quantization/configs/N=24576,K=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=24576,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=24576,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=24576,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=24576,K=7168,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,26 @@
+{
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=256,K=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    }
+}

python/sglang/srt/layers/quantization/configs/N=256,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    }
+}

python/sglang/srt/layers/quantization/configs/N=256,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=256,K=7168,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,26 @@
+{
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 5
+    }
+}

python/sglang/srt/layers/quantization/configs/N=3072,K=1536,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

python/sglang/srt/layers/quantization/configs/N=3072,K=1536,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

python/sglang/srt/layers/quantization/configs/N=3072,K=1536,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 8,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

python/sglang/srt/layers/quantization/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=3072,K=1536,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=3072,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

python/sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

python/sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

python/sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "4": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "128": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "256": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 5
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "128": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    }
+}

python/sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "24": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "32": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "96": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 4
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,26 @@
+{
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,146 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 5
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "512": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 3
+    }
+}

python/sglang/srt/layers/quantization/configs/N=4096,K=512,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}

python/sglang/srt/layers/quantization/configs/N=4096,K=512,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json

@@ -0,0 +1,164 @@
+{
+    "1": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "8": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "16": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "24": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "32": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "48": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "64": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "96": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "128": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "256": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "512": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 4,
+        "num_warps": 4,
+        "num_stages": 2,
+        "waves_per_eu": 0
+    }
+}