chore: import upstream snapshot with attribution

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

@@ -0,0 +1,19 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from .gptq_cpu import GPTQIntelAMXLinearScheme, GPTQIntelAMXMoEScheme
+from .gptq_linear import GPTQAscendLinearScheme, GPTQLinearScheme
+from .gptq_marlin import GPTQMarlinLinearScheme
+from .gptq_moe import GPTQMarlinMoEScheme, GPTQMoEAscendScheme
+from .gptq_scheme import GPTQLinearSchemeBase, GPTQMoESchemeBase
+
+__all__ = [
+    "GPTQLinearSchemeBase",
+    "GPTQMoESchemeBase",
+    "GPTQLinearScheme",
+    "GPTQAscendLinearScheme",
+    "GPTQIntelAMXLinearScheme",
+    "GPTQMarlinLinearScheme",
+    "GPTQMoEAscendScheme",
+    "GPTQIntelAMXMoEScheme",
+    "GPTQMarlinMoEScheme",
+]

python/sglang/srt/layers/quantization/gptq/schemes/gptq_cpu.py

@@ -0,0 +1,285 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import torch
+
+from sglang.srt.hardware_backend.cpu.quantization.gptq_kernels import (
+    GPTQIntelAMXLinearKernel,
+    GPTQIntelAMXMoEKernel,
+)
+from sglang.srt.layers.linear import set_weight_attrs
+from sglang.srt.layers.moe import MoeRunnerConfig
+from sglang.srt.layers.parameter import (
+    ChannelQuantScaleParameter,
+    GroupQuantScaleParameter,
+    PackedColumnParameter,
+    PackedvLLMParameter,
+    RowvLLMParameter,
+)
+
+from .gptq_linear import GPTQLinearScheme
+from .gptq_scheme import GPTQMoESchemeBase
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.token_dispatcher import StandardDispatchOutput
+    from sglang.srt.layers.quantization.gptq.gptq import GPTQConfig
+
+__all__ = ["GPTQIntelAMXLinearScheme", "GPTQIntelAMXMoEScheme"]
+
+
+def _check_cpu_amx_support(quant_config: GPTQConfig) -> None:
+    if quant_config.desc_act and not (
+        quant_config.true_sequential and quant_config.static_groups
+    ):
+        raise ValueError(
+            "Currently, desc_act (True) is only supported with sequential "
+            "and static group on CPU with AMX."
+        )
+    if quant_config.weight_bits != 4:
+        raise ValueError("Currently, only 4bits is supported on CPU with AMX.")
+    if quant_config.checkpoint_format == "gptq_v2":
+        raise ValueError("Currently, gptq_v2 is not supported on CPU with AMX.")
+
+
+class GPTQIntelAMXLinearScheme(GPTQLinearScheme):
+    """Linear scheme for GPTQ on Intel CPU with AMX."""
+
+    def _init_kernel(self, quant_config: GPTQConfig):
+        return GPTQIntelAMXLinearKernel(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,
+    ):
+        _check_cpu_amx_support(self.quant_config)
+
+        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.numerator != 0:
+            raise ValueError(
+                "The output size is not aligned with the quantized "
+                "weight shape. This can be caused by too large "
+                "tensor parallel size."
+            )
+
+        if self.quant_config.group_size != -1:
+            group_size = self.quant_config.group_size
+        else:
+            group_size = input_size
+
+        scale_and_zero_size = input_size_per_partition // group_size
+        scale_and_zero_input_dim = 0
+
+        qweight = PackedvLLMParameter(
+            data=torch.empty(
+                input_size_per_partition // self.quant_config.pack_factor,
+                output_size_per_partition,
+                dtype=torch.int32,
+            ),
+            input_dim=0,
+            output_dim=1,
+            packed_dim=0,
+            packed_factor=self.quant_config.pack_factor,
+            weight_loader=weight_loader,
+        )
+
+        g_idx = RowvLLMParameter(
+            data=torch.tensor(
+                [
+                    i // self.quant_config.group_size
+                    for i in range(input_size_per_partition)
+                ],
+                dtype=torch.int32,
+            ),
+            input_dim=0,
+            weight_loader=weight_loader,
+        )
+        qzeros_args = {
+            "data": torch.empty(
+                scale_and_zero_size,
+                output_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            "weight_loader": weight_loader,
+        }
+        weight_scale_args = {
+            "data": torch.empty(
+                scale_and_zero_size,
+                output_size_per_partition,
+                dtype=params_dtype,
+            ),
+            "weight_loader": weight_loader,
+        }
+        if scale_and_zero_input_dim is None:
+            scales = ChannelQuantScaleParameter(output_dim=1, **weight_scale_args)
+            qzeros = PackedColumnParameter(
+                output_dim=1,
+                packed_dim=1,
+                packed_factor=self.quant_config.pack_factor,
+                **qzeros_args,
+            )
+        else:
+            scales = GroupQuantScaleParameter(
+                output_dim=1, input_dim=0, **weight_scale_args
+            )
+            qzeros = PackedvLLMParameter(
+                input_dim=0,
+                output_dim=1,
+                packed_dim=1,
+                packed_factor=self.quant_config.pack_factor,
+                **qzeros_args,
+            )
+
+        layer.register_parameter("qweight", qweight)
+        layer.register_parameter("g_idx", g_idx)
+        layer.register_parameter("qzeros", qzeros)
+        layer.register_parameter("scales", scales)
+
+
+class GPTQIntelAMXMoEScheme(GPTQMoESchemeBase):
+    """MoE scheme for GPTQ on Intel CPU with AMX."""
+
+    def __init__(self, quant_config: GPTQConfig):
+        self.quant_config = quant_config
+        self.kernel = GPTQIntelAMXMoEKernel(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
+
+        _check_cpu_amx_support(self.quant_config)
+        pack_factor = self.quant_config.pack_factor
+
+        if self.quant_config.group_size != -1:
+            scales_size13 = hidden_size // self.quant_config.group_size
+            w2_scales_size = intermediate_size_per_partition
+            scales_size2 = w2_scales_size // self.quant_config.group_size
+            strategy = FusedMoeWeightScaleSupported.GROUP.value
+        else:
+            scales_size13 = 1
+            scales_size2 = 1
+            strategy = FusedMoeWeightScaleSupported.CHANNEL.value
+
+        extra_weight_attrs.update({"quant_method": strategy, "is_transposed": True})
+
+        w13_qweight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size // pack_factor,
+                2 * intermediate_size_per_partition,
+                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 // pack_factor,
+                hidden_size,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_qweight", w2_qweight)
+        set_weight_attrs(w2_qweight, extra_weight_attrs)
+
+        w13_scales = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                scales_size13,
+                2 * intermediate_size_per_partition,
+                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, scales_size2, hidden_size, dtype=params_dtype),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_scales", w2_scales)
+        set_weight_attrs(w2_scales, extra_weight_attrs)
+        set_weight_attrs(w2_scales, {"load_full_w2": self.quant_config.desc_act})
+
+        w13_qzeros = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                scales_size13,
+                2 * intermediate_size_per_partition // 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,
+                scales_size2,
+                hidden_size // pack_factor,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_qzeros", w2_qzeros)
+        set_weight_attrs(w2_qzeros, extra_weight_attrs)
+        set_weight_attrs(w2_qzeros, {"load_full_w2": self.quant_config.desc_act})
+
+        w13_g_idx = torch.nn.Parameter(
+            torch.empty(num_experts, hidden_size, dtype=torch.int32),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_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_g_idx", w2_g_idx)
+        set_weight_attrs(w2_g_idx, extra_weight_attrs)
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.kernel.create_moe_runner(layer, moe_runner_config)
+
+    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,
+        dispatch_output: StandardDispatchOutput,
+    ):
+        return self.kernel.apply(layer, dispatch_output)

python/sglang/srt/layers/quantization/gptq/schemes/gptq_linear.py

@@ -0,0 +1,171 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Optional
+
+import torch
+
+from sglang.srt.layers.parameter import (
+    ChannelQuantScaleParameter,
+    GroupQuantScaleParameter,
+    PackedColumnParameter,
+    PackedvLLMParameter,
+    RowvLLMParameter,
+)
+from sglang.srt.utils import set_weight_attrs
+
+from .gptq_scheme import GPTQLinearSchemeBase
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.quantization.gptq.gptq import GPTQConfig
+
+__all__ = ["GPTQLinearScheme", "GPTQAscendLinearScheme"]
+
+
+class GPTQLinearScheme(GPTQLinearSchemeBase):
+    def __init__(self, quant_config: GPTQConfig):
+        self.quant_config = quant_config
+        self.use_v2_format = quant_config.checkpoint_format == "gptq_v2"
+        self.kernel = self._init_kernel(quant_config)
+
+    def _init_kernel(self, quant_config: GPTQConfig):
+        from sglang.srt.hardware_backend.gpu.quantization.gptq_kernels import (
+            GPTQLinearKernel,
+        )
+
+        return GPTQLinearKernel(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,
+    ):
+        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.numerator != 0:
+            raise ValueError(
+                "The output size is not aligned with the quantized "
+                "weight shape. This can be caused by too large "
+                "tensor parallel size."
+            )
+
+        group_size = (
+            self.quant_config.group_size
+            if self.quant_config.group_size != -1
+            else input_size
+        )
+        self.kernel.use_shuffle = True
+        scale_and_zero_size = input_size // group_size
+        scale_and_zero_input_dim = None
+        if (
+            input_size != input_size_per_partition
+            and self.quant_config.group_size != -1
+        ):
+            if self.quant_config.desc_act:
+                self.kernel.use_shuffle = False
+            else:
+                scale_and_zero_size = input_size_per_partition // group_size
+                scale_and_zero_input_dim = 0
+
+        qweight = PackedvLLMParameter(
+            data=torch.empty(
+                input_size_per_partition // self.quant_config.pack_factor,
+                output_size_per_partition,
+                dtype=torch.int32,
+            ),
+            input_dim=0,
+            output_dim=1,
+            packed_dim=0,
+            packed_factor=self.quant_config.pack_factor,
+            weight_loader=weight_loader,
+        )
+
+        g_idx = RowvLLMParameter(
+            data=torch.tensor(
+                [
+                    i // self.quant_config.group_size
+                    for i in range(input_size_per_partition)
+                ],
+                dtype=torch.int32,
+            ),
+            input_dim=0,
+            weight_loader=weight_loader,
+        )
+        qzeros_args = {
+            "data": torch.empty(
+                scale_and_zero_size,
+                output_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            "weight_loader": weight_loader,
+        }
+        weight_scale_args = {
+            "data": torch.empty(
+                scale_and_zero_size,
+                output_size_per_partition,
+                dtype=params_dtype,
+            ),
+            "weight_loader": weight_loader,
+        }
+        if scale_and_zero_input_dim is None:
+            scales = ChannelQuantScaleParameter(output_dim=1, **weight_scale_args)
+            qzeros = PackedColumnParameter(
+                output_dim=1,
+                packed_dim=1,
+                packed_factor=self.quant_config.pack_factor,
+                **qzeros_args,
+            )
+        else:
+            scales = GroupQuantScaleParameter(
+                output_dim=1, input_dim=0, **weight_scale_args
+            )
+            qzeros = PackedvLLMParameter(
+                input_dim=0,
+                output_dim=1,
+                packed_dim=1,
+                packed_factor=self.quant_config.pack_factor,
+                **qzeros_args,
+            )
+
+        layer.register_parameter("qweight", qweight)
+        layer.register_parameter("g_idx", g_idx)
+        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 GPTQAscendLinearScheme(GPTQLinearScheme):
+    def _init_kernel(self, quant_config: GPTQConfig):
+        from sglang.srt.hardware_backend.npu.quantization.gptq_kernels import (
+            GPTQLinearAscendKernel,
+        )
+
+        return GPTQLinearAscendKernel(quant_config)
+
+    def create_weights(self, layer: torch.nn.Module, **kwargs):
+        if self.quant_config.desc_act:
+            raise ValueError(
+                "Currently, desc_act (True) is not supported by GPTQ "
+                "quantization on npu."
+            )
+
+        super().create_weights(layer=layer, **kwargs)
+        set_weight_attrs(layer.qzeros, {"pack_factor": self.quant_config.pack_factor})
+        set_weight_attrs(layer.qweight, {"pack_factor": self.quant_config.pack_factor})

python/sglang/srt/layers/quantization/gptq/schemes/gptq_marlin.py

@@ -0,0 +1,158 @@
+# SPDX-License-Identifier: Apache-2.0
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Optional
+
+import torch
+
+from sglang.srt.layers.parameter import (
+    ChannelQuantScaleParameter,
+    GroupQuantScaleParameter,
+    PackedColumnParameter,
+    PackedvLLMParameter,
+    RowvLLMParameter,
+)
+from sglang.srt.layers.quantization.marlin_utils import (
+    MarlinLinearLayerConfig,
+    marlin_repeat_scales_on_all_ranks,
+    verify_marlin_supported,
+)
+
+from .gptq_scheme import GPTQLinearSchemeBase
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.quantization.gptq.gptq import GPTQMarlinConfig
+
+__all__ = ["GPTQMarlinLinearScheme"]
+
+
+class GPTQMarlinLinearScheme(GPTQLinearSchemeBase):
+    def __init__(self, quant_config: GPTQMarlinConfig):
+        self.quant_config = quant_config
+        self.kernel = self._init_kernel(quant_config)
+
+        verify_marlin_supported(
+            quant_type=self.quant_config.quant_type,
+            group_size=self.quant_config.group_size,
+        )
+
+    def _init_kernel(self, quant_config: GPTQMarlinConfig):
+        from sglang.srt.hardware_backend.gpu.quantization.gptq_kernels import (
+            GPTQMarlinLinearKernel,
+        )
+
+        return GPTQMarlinLinearKernel(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,
+        weight_loader,
+        **kwargs,
+    ) -> None:
+        output_size_per_partition = sum(output_partition_sizes)
+        is_row_parallel = input_size != input_size_per_partition
+
+        self.kernel.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_config.quant_type,
+            act_type=params_dtype,
+            group_size=self.quant_config.group_size,
+            zero_points=False,
+            has_g_idx=self.quant_config.desc_act,
+        )
+
+        group_size = (
+            self.quant_config.group_size
+            if self.quant_config.group_size != -1
+            else input_size
+        )
+
+        if marlin_repeat_scales_on_all_ranks(
+            self.quant_config.desc_act, self.quant_config.group_size, is_row_parallel
+        ):
+            scales_and_zp_input_dim = None
+            scales_and_zp_size = input_size // group_size
+        else:
+            scales_and_zp_input_dim = 0
+            scales_and_zp_size = input_size_per_partition // group_size
+
+        qweight = PackedvLLMParameter(
+            data=torch.empty(
+                input_size_per_partition // self.quant_config.pack_factor,
+                output_size_per_partition,
+                dtype=torch.int32,
+            ),
+            input_dim=0,
+            output_dim=1,
+            packed_dim=0,
+            packed_factor=self.quant_config.pack_factor,
+            weight_loader=weight_loader,
+        )
+
+        g_idx = RowvLLMParameter(
+            data=torch.empty(
+                input_size_per_partition,
+                dtype=torch.int32,
+            ),
+            input_dim=0,
+            weight_loader=weight_loader,
+        )
+
+        qzeros_args = {
+            "data": torch.empty(
+                scales_and_zp_size,
+                output_size_per_partition // self.quant_config.pack_factor,
+                dtype=torch.int32,
+            ),
+            "weight_loader": weight_loader,
+        }
+        weight_scale_args = {
+            "data": torch.empty(
+                scales_and_zp_size,
+                output_size_per_partition,
+                dtype=params_dtype,
+            ),
+            "weight_loader": weight_loader,
+        }
+
+        if scales_and_zp_input_dim is None:
+            scales = ChannelQuantScaleParameter(output_dim=1, **weight_scale_args)
+            qzeros = PackedColumnParameter(
+                output_dim=1,
+                packed_dim=1,
+                packed_factor=self.quant_config.pack_factor,
+                **qzeros_args,
+            )
+        else:
+            scales = GroupQuantScaleParameter(
+                output_dim=1, input_dim=0, **weight_scale_args
+            )
+            qzeros = PackedvLLMParameter(
+                input_dim=0,
+                output_dim=1,
+                packed_dim=1,
+                packed_factor=self.quant_config.pack_factor,
+                **qzeros_args,
+            )
+
+        layer.register_parameter("qweight", qweight)
+        layer.register_parameter("g_idx", g_idx)
+        layer.register_parameter("scales", scales)
+        layer.register_parameter("qzeros", qzeros)
+
+    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/gptq/schemes/gptq_moe.py

@@ -0,0 +1,305 @@
+# 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 MoeRunnerConfig
+
+from .gptq_scheme import GPTQMoESchemeBase
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.moe.token_dispatcher import StandardDispatchOutput
+    from sglang.srt.layers.quantization.gptq.gptq import GPTQConfig, GPTQMarlinConfig
+
+__all__ = ["GPTQMoEAscendScheme", "GPTQMarlinMoEScheme"]
+
+
+class GPTQMoEAscendScheme(GPTQMoESchemeBase):
+    def __init__(self, quant_config: GPTQConfig):
+        self.quant_config = quant_config
+        from sglang.srt.hardware_backend.npu.quantization.gptq_kernels import (
+            GPTQMoEAscendKernel,
+        )
+
+        self.kernel = GPTQMoEAscendKernel(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
+
+        pack_factor = self.quant_config.pack_factor
+
+        num_groups_w13 = hidden_size // self.quant_config.group_size
+        num_groups_w2 = intermediate_size_per_partition // self.quant_config.group_size
+
+        extra_weight_attrs.update(
+            {
+                "is_transposed": True,
+                "quant_method": FusedMoeWeightScaleSupported.GROUP.value,
+            }
+        )
+
+        w13_qweight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size // pack_factor,
+                2 * intermediate_size_per_partition,
+                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 // pack_factor,
+                hidden_size,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_qweight", w2_qweight)
+        set_weight_attrs(w2_qweight, extra_weight_attrs)
+
+        w13_scales = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                num_groups_w13,
+                2 * intermediate_size_per_partition,
+                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 // 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 // pack_factor,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_qzeros", w2_qzeros)
+        set_weight_attrs(w2_qzeros, extra_weight_attrs)
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.kernel.create_moe_runner(layer, moe_runner_config)
+
+    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,
+        dispatch_output: StandardDispatchOutput,
+    ):
+        return self.kernel.apply(layer, dispatch_output)
+
+
+class GPTQMarlinMoEScheme(GPTQMoESchemeBase):
+    def __init__(self, quant_config: GPTQMarlinConfig):
+        self.quant_config = quant_config
+        from sglang.srt.hardware_backend.gpu.quantization.gptq_kernels import (
+            GPTQMarlinMoEKernel,
+        )
+
+        self.kernel = GPTQMarlinMoEKernel(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
+
+        self.kernel.is_k_full = (
+            not self.quant_config.desc_act
+        ) or layer.moe_tp_size == 1
+
+        if self.quant_config.group_size != -1:
+            scales_size13 = hidden_size // self.quant_config.group_size
+            if self.quant_config.desc_act:
+                w2_scales_size = intermediate_size_per_partition
+            else:
+                w2_scales_size = intermediate_size_per_partition * layer.moe_tp_size
+            scales_size2 = w2_scales_size // self.quant_config.group_size
+            strategy = FusedMoeWeightScaleSupported.GROUP.value
+        else:
+            scales_size13 = 1
+            scales_size2 = 1
+            strategy = FusedMoeWeightScaleSupported.CHANNEL.value
+
+        extra_weight_attrs.update({"quant_method": strategy, "is_transposed": True})
+
+        w13_qweight = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size // self.quant_config.pack_factor,
+                2 * intermediate_size_per_partition,
+                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 // self.quant_config.pack_factor,
+                hidden_size,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_qweight", w2_qweight)
+        set_weight_attrs(w2_qweight, extra_weight_attrs)
+
+        w13_scales = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                scales_size13,
+                2 * intermediate_size_per_partition,
+                dtype=torch.half,
+            ),
+            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, scales_size2, hidden_size, dtype=torch.half),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_scales", w2_scales)
+        set_weight_attrs(w2_scales, extra_weight_attrs)
+        set_weight_attrs(w2_scales, {"load_full_w2": self.quant_config.desc_act})
+
+        w13_qzeros = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                scales_size13,
+                2 * intermediate_size_per_partition // self.quant_config.pack_factor,
+                dtype=params_dtype,
+            ),
+            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,
+                scales_size2,
+                hidden_size // self.quant_config.pack_factor,
+                dtype=params_dtype,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w2_qzeros", w2_qzeros)
+        set_weight_attrs(w2_qzeros, extra_weight_attrs)
+        set_weight_attrs(w2_qzeros, {"load_full_w2": self.quant_config.desc_act})
+
+        w13_g_idx = torch.nn.Parameter(
+            torch.empty(
+                num_experts,
+                hidden_size,
+                dtype=torch.int32,
+            ),
+            requires_grad=False,
+        )
+        layer.register_parameter("w13_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_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)
+
+    def create_moe_runner(
+        self, layer: torch.nn.Module, moe_runner_config: MoeRunnerConfig
+    ):
+        self.kernel.create_moe_runner(layer, moe_runner_config)
+
+    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,
+        dispatch_output: StandardDispatchOutput,
+    ):
+        return self.kernel.apply(layer, dispatch_output)

python/sglang/srt/layers/quantization/gptq/schemes/gptq_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__ = ["GPTQLinearSchemeBase", "GPTQMoESchemeBase"]
+
+
+class GPTQLinearSchemeBase(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 GPTQMoESchemeBase(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