# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project # Adapted from https://github.com/vllm-project/vllm/blob/v0.6.3.post1/vllm/model_executor/model_loader/loader.py from __future__ import annotations # ruff: noqa: SIM117 import collections import dataclasses import fnmatch import gc import glob import json import logging import math import os import re import socket import tempfile import threading import time from abc import ABC, abstractmethod from contextlib import contextmanager, suppress from typing import ( TYPE_CHECKING, Any, Dict, Generator, Iterable, List, Optional, Tuple, Union, cast, ) import huggingface_hub import numpy as np import torch from sglang.srt.constants import GIB_BYTES from sglang.srt.model_loader.remote_instance_weight_loader_utils import ( RemoteInstanceWeightLoaderBackend, get_remote_instance_transfer_engine_info_per_rank, register_memory_region, ) from sglang.srt.runtime_context import get_server_args from sglang.srt.utils import get_available_gpu_memory # Try to import accelerate (optional dependency) try: from accelerate import infer_auto_device_map, init_empty_weights from accelerate.utils import get_max_memory HAS_ACCELERATE = True except ImportError: HAS_ACCELERATE = False infer_auto_device_map = None init_empty_weights = None get_max_memory = None from huggingface_hub import HfApi, hf_hub_download from torch import nn from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer from transformers.utils import SAFE_WEIGHTS_INDEX_NAME from sglang.srt.configs.load_config import LoadConfig, LoadFormat from sglang.srt.connector import ( ConnectorType, create_remote_connector, get_connector_type, ) from sglang.srt.connector.utils import parse_model_name from sglang.srt.distributed import ( model_parallel_is_initialized, ) from sglang.srt.layers.modelopt_utils import QUANT_CFG_CHOICES from sglang.srt.layers.quantization.base_config import QuantizationConfig from sglang.srt.model_loader.remote_instance_weight_loader_utils import ( trigger_transferring_weights_request, ) from sglang.srt.model_loader.utils import ( get_model_architecture, set_default_torch_dtype, ) from sglang.srt.utils.common import is_cuda_alike # Constants for memory management DEFAULT_GPU_MEMORY_FRACTION_FOR_CALIBRATION = ( 0.8 # Reserve 20% GPU memory headroom for ModelOpt calibration ) from sglang.srt.environ import envs from sglang.srt.model_loader.weight_utils import ( buffered_multi_thread_safetensors_weights_iterator, download_safetensors_index_file_from_hf, download_weights_from_hf, fastsafetensors_weights_iterator, filter_duplicate_safetensors_files, filter_files_not_needed_for_inference, get_gguf_extra_tensor_names, get_quant_config, gguf_quant_weights_iterator, initialize_dummy_weights, maybe_add_mtp_safetensors, multi_thread_pt_weights_iterator, np_cache_weights_iterator, pt_weights_iterator, safetensors_weights_iterator, set_runai_streamer_env, ) from sglang.srt.platforms import current_platform from sglang.srt.runtime_context import get_parallel from sglang.srt.utils import ( get_bool_env_var, get_device_capability, is_npu, is_pin_memory_available, rank0_log, set_weight_attrs, ) from sglang.srt.utils.common import temp_set_env if TYPE_CHECKING: from sglang.srt.configs.device_config import DeviceConfig from sglang.srt.configs.model_config import ModelConfig from sglang.srt.layers.quantization.base_config import QuantizationConfig _is_npu = is_npu() # ModelOpt: QUANT_CFG_CHOICES is imported from modelopt_utils.py # which contains the complete mapping of quantization config choices logger = logging.getLogger(__name__) @contextmanager def device_loading_context(module: torch.nn.Module, target_device: torch.device): if target_device.type == "cpu": # If target is CPU, no need to move anything yield module return original_infos: Dict[str, Dict] = {} # Store original device states and move parameters to GPU if they're on CPU for name, p in module.named_parameters(): if p.device.type == "cpu": original_data = p.data device_data = p.data.to(target_device) original_infos[name] = dict( device=p.device, original_data=original_data, device_data=device_data, ) p.data = device_data # Parameters already on target device are not touched try: yield module finally: # Restore parameters to their original devices, ignoring new parameters pin_memory = is_pin_memory_available() for name, p in module.named_parameters(): if name in original_infos: original_info = original_infos[name] device_data = original_info["device_data"] original_data = original_info["original_data"] original_device: torch.device = original_info["device"] if ( (device_data.device == p.data.device) and (device_data.data_ptr() == p.data.data_ptr()) and (device_data.shape == p.data.shape) and (device_data.dtype == p.data.dtype) ): original_data.copy_(p.data.to(original_data.device)) p.data = original_data elif original_device.type == "cpu": # `torch.empty_like` does not support `pin_memory` argument cpu_data = torch.empty_strided( size=p.data.size(), stride=p.data.stride(), dtype=p.data.dtype, layout=p.data.layout, device="cpu", pin_memory=pin_memory, ) cpu_data.copy_(p.data) p.data = cpu_data else: p.data = p.data.to(original_device) # New parameters or parameters already on target device are untouched logger = logging.getLogger(__name__) def _get_quantization_config( model_config: ModelConfig, load_config: LoadConfig, ) -> Optional[QuantizationConfig]: """Get the quantization config.""" model_class, _ = get_model_architecture(model_config) packed_modules_mapping = getattr(model_class, "packed_modules_mapping", {}) remap_prefix = getattr(model_class, "remap_prefix", None) # TODO: we should remove this code and switch to the packed_modules_mapping declared inside the modeling files if model_config.quantization == "quark": packed_modules_mapping.update( { "gate_up_proj": ["gate_proj", "up_proj"], "fused_qkv_a_proj_with_mqa": ["q_a_proj", "kv_a_proj_with_mqa"], } ) if _is_npu: packed_modules_mapping.update( { "visual": { "qkv_proj": ["qkv"], "gate_up_proj": ["gate_proj", "up_proj"], }, "vision_model": { "qkv_proj": ["q_proj", "k_proj", "v_proj"], "proj": ["out_proj"], }, "model": { "qkv_proj": ["q_proj", "k_proj", "v_proj"], "gate_up_proj": ["gate_proj", "up_proj"], "fused_qkv_a_proj_with_mqa": [ "q_a_proj", "kv_a_proj_with_mqa", ], }, } ) if model_config.quantization is not None: quant_config = get_quant_config( model_config, load_config, packed_modules_mapping, remap_prefix ) # (yizhang2077) workaround for nvidia/Llama-4-Maverick-17B-128E-Eagle3 if quant_config is None: return None # Carry DSV4 expert layout into Fp8Config so downstream readers don't read env. from sglang.srt.layers.quantization.fp8 import Fp8Config if isinstance(quant_config, Fp8Config): quant_config.is_fp4_experts = model_config.is_fp4_experts quant_config.dequant_fp4_to_fp8 = envs.SGLANG_DSV4_FP4_DEQUANT.get() # Handle hybrid NVFP4 moe (nvidia/DeepSeek-V4-Pro-NVFP4) nvfp4_meta = model_config.nvfp4_moe_meta if nvfp4_meta is not None: from sglang.srt.layers.quantization.modelopt_quant import ( HybridFp8NvFp4Config, ModelOptFp4Config, ) # MTP MoE layers (model.decoder.*) are not NVFP4 quantized. nvfp4_exclude_modules = list( nvfp4_meta.get("exclude_modules") or [] ) + ["model.decoder.*"] nvfp4_config = ModelOptFp4Config( is_checkpoint_nvfp4_serialized=True, group_size=int(nvfp4_meta["group_size"]), exclude_modules=nvfp4_exclude_modules, packed_modules_mapping=quant_config.packed_modules_mapping, ) quant_config = HybridFp8NvFp4Config( fp8_config=quant_config, nvfp4_config=nvfp4_config ) if not _is_npu: major, minor = get_device_capability() if major is not None and minor is not None: assert 0 <= minor < 10 capability = major * 10 + minor if capability < quant_config.get_min_capability(): raise ValueError( f"The quantization method {model_config.quantization} " "is not supported for the current GPU. " f"Minimum capability: {quant_config.get_min_capability()}. " f"Current capability: {capability}." ) supported_dtypes = quant_config.get_supported_act_dtypes() if model_config.dtype not in supported_dtypes: raise ValueError( f"{model_config.dtype} is not supported for quantization " f"method {model_config.quantization}. Supported dtypes: " f"{supported_dtypes}" ) hf_to_sglang_mapper = getattr(model_class, "hf_to_sglang_mapper", None) # pass mappings by reference to quant_config if hf_to_sglang_mapper is not None and quant_config is not None: quant_config.apply_weight_name_mapper(hf_to_sglang_mapper) return quant_config return None def _initialize_model( model_config: ModelConfig, load_config: LoadConfig, quant_config: Optional[QuantizationConfig] = None, ) -> nn.Module: """Initialize a model with the given configurations.""" model_class, _ = get_model_architecture(model_config) kwargs = { "config": model_config.hf_config, "quant_config": quant_config, } # Only add sparse head kwargs if envs.SGLANG_EMBEDDINGS_SPARSE_HEAD.is_set() if envs.SGLANG_EMBEDDINGS_SPARSE_HEAD.is_set(): kwargs["sparse_head"] = envs.SGLANG_EMBEDDINGS_SPARSE_HEAD.get() kwargs["model_path"] = model_config.model_path if load_config.draft_model_idx is not None: kwargs["draft_model_idx"] = load_config.draft_model_idx return model_class(**kwargs) def _post_load_weights(model: nn.Module) -> None: # Loaders that bypass `model.load_weights()` (dummy / sharded state / remote instance / # remote fs) must trigger the model's post-load fixup explicitly; `model.load_weights()` # would normally do it internally. NextN subclasses override the method to fill in # `is_nextn=True`, so the loader doesn't need to know. if hasattr(model, "post_load_weights"): model.post_load_weights() class BaseModelLoader(ABC): """Base class for model loaders.""" def __init__(self, load_config: LoadConfig): self.load_config = load_config @abstractmethod def download_model(self, model_config: ModelConfig) -> None: """Download a model so that it can be immediately loaded.""" raise NotImplementedError @abstractmethod def load_model( self, *, model_config: ModelConfig, device_config: DeviceConfig, ) -> nn.Module: """Load a model with the given configurations.""" raise NotImplementedError class DefaultModelLoader(BaseModelLoader): """Model loader that can load different file types from disk.""" # default number of thread when enable multithread weight loading DEFAULT_NUM_THREADS = 8 _MTP_PATTERN = re.compile(r"model\.mtp\.layers\.(\d+)\.") @dataclasses.dataclass class Source: """A source for weights.""" model_or_path: str """The model ID or path.""" revision: Optional[str] """The optional model revision.""" prefix: str = "" """A prefix to prepend to all weights.""" fall_back_to_pt: bool = True """Whether .pt weights can be used.""" model_config: Optional[ModelConfig] = None """The model configuration (for checking architecture, etc).""" @classmethod def init_new(cls, model_config: ModelConfig, model): return cls( model_config.model_path, model_config.revision, prefix="", fall_back_to_pt=getattr(model, "fall_back_to_pt_during_load", True), model_config=model_config, ) counter_before_loading_weights: float = 0.0 counter_after_loading_weights: float = 0.0 def __init__(self, load_config: LoadConfig): super().__init__(load_config) extra_config = load_config.model_loader_extra_config allowed_keys = {"enable_multithread_load", "num_threads"} unexpected_keys = set(extra_config.keys()) - allowed_keys if unexpected_keys: raise ValueError( f"Unexpected extra config keys for load format " f"{load_config.load_format}: " f"{unexpected_keys}" ) def _maybe_download_from_modelscope( self, model: str, revision: Optional[str] ) -> str: """Download model from ModelScope hub if SGLANG_USE_MODELSCOPE is True. Returns the path to the downloaded model, or the original model path if not downloaded from ModelScope.""" if get_bool_env_var("SGLANG_USE_MODELSCOPE"): # download model from ModelScope hub, # lazy import so that modelscope is not required for normal use. # pylint: disable=C. from modelscope.hub.snapshot_download import snapshot_download if not os.path.exists(model): model_path = snapshot_download( model_id=model, cache_dir=self.load_config.download_dir, local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE, revision=revision, ignore_file_pattern=self.load_config.ignore_patterns, ) else: model_path = model return model_path return model def _prepare_weights( self, model_name_or_path: str, revision: Optional[str], fall_back_to_pt: bool ) -> Tuple[str, List[str], bool]: """Prepare weights for the model. If the model is not local, it will be downloaded.""" model_name_or_path = self._maybe_download_from_modelscope( model_name_or_path, revision ) is_local = os.path.isdir(model_name_or_path) load_format = self.load_config.load_format use_safetensors = False index_file = SAFE_WEIGHTS_INDEX_NAME # Some quantized models use .pt files for storing the weights. if load_format == LoadFormat.AUTO: allow_patterns = ["*.safetensors", "*.bin"] elif ( load_format == LoadFormat.SAFETENSORS or load_format == LoadFormat.FASTSAFETENSORS ): use_safetensors = True allow_patterns = ["*.safetensors"] elif load_format == LoadFormat.MISTRAL: use_safetensors = True allow_patterns = ["consolidated*.safetensors"] index_file = "consolidated.safetensors.index.json" elif load_format == LoadFormat.PT: allow_patterns = ["*.pt"] elif load_format == LoadFormat.NPCACHE: allow_patterns = ["*.bin"] elif load_format == LoadFormat.DUMMY: raise ValueError( f"DUMMY load_format should use DummyModelLoader and not call _prepare_weights" ) else: raise ValueError(f"Unknown load_format: {load_format}") if fall_back_to_pt: allow_patterns += ["*.pt"] if not is_local: hf_folder = download_weights_from_hf( model_name_or_path, self.load_config.download_dir, allow_patterns, revision, ignore_patterns=self.load_config.ignore_patterns, ) else: hf_folder = model_name_or_path server_args = get_server_args() if server_args and server_args.model_checksum is not None: from sglang.srt.utils.model_file_verifier import verify checksums_source = server_args.model_checksum or model_name_or_path verify(model_path=hf_folder, checksums_source=checksums_source) hf_weights_files: List[str] = [] for pattern in allow_patterns: hf_weights_files += glob.glob(os.path.join(hf_folder, pattern)) if len(hf_weights_files) > 0: if pattern == "*.safetensors": use_safetensors = True break if use_safetensors: # For models like Mistral-7B-Instruct-v0.3 # there are both sharded safetensors files and a consolidated # safetensors file. Using both breaks. # Here, we download the `model.safetensors.index.json` and filter # any files not found in the index. if not is_local: download_safetensors_index_file_from_hf( model_name_or_path, index_file, self.load_config.download_dir, revision, ) hf_weights_files = filter_duplicate_safetensors_files( hf_weights_files, hf_folder, index_file ) else: hf_weights_files = filter_files_not_needed_for_inference(hf_weights_files) if len(hf_weights_files) == 0: raise RuntimeError( f"Cannot find any model weights with `{model_name_or_path}`" ) # Sort and optionally stagger weight files (see SGLANG_SORT_WEIGHT_FILES). # k=-1: no sort; k=0: sort only; k>0: sort + stagger by (tp_rank*k). k = envs.SGLANG_SORT_WEIGHT_FILES.get() if k >= 0: hf_weights_files.sort() if k > 0: tp_size = get_parallel().tp_size if tp_size > 1: tp_rank = get_parallel().tp_rank group_size = tp_size * k staggered: List[str] = [] for i in range(0, len(hf_weights_files), group_size): group = hf_weights_files[i : i + group_size] n = len(group) staggered.extend(group[(j + tp_rank * k) % n] for j in range(n)) hf_weights_files = staggered return hf_folder, hf_weights_files, use_safetensors def _get_weights_iterator( self, source: Source ) -> Generator[Tuple[str, torch.Tensor], None, None]: """Get an iterator for the model weights based on the load format.""" extra_config = self.load_config.model_loader_extra_config use_multithread = extra_config.get("enable_multithread_load", True) hf_folder, hf_weights_files, use_safetensors = self._prepare_weights( source.model_or_path, source.revision, source.fall_back_to_pt ) if use_safetensors and source.model_config is not None: hf_weights_files = maybe_add_mtp_safetensors( hf_weights_files, hf_folder, "model.safetensors.index.json", source.model_config.hf_config, ) if self.load_config.load_format == LoadFormat.NPCACHE: # Currently np_cache only support *.bin checkpoints assert use_safetensors is False weights_iterator = np_cache_weights_iterator( source.model_or_path, self.load_config.download_dir, hf_folder, hf_weights_files, ) elif use_safetensors: server_args = get_server_args() weight_loader_disable_mmap = server_args.weight_loader_disable_mmap weight_loader_prefetch = server_args.weight_loader_prefetch_checkpoints prefetch_num_threads = server_args.weight_loader_prefetch_num_threads weight_loader_drop_cache_after_load = ( server_args.weight_loader_drop_cache_after_load ) # Prefetch and multi-threaded loading both read the same shards, # competing for I/O on shared/network storage. When prefetch is # active (mmap path, not FASTSAFETENSORS) and the user didn't # explicitly request multi-threaded loading, fall back to the # single-threaded loader and let prefetch feed the page cache. # Setting enable_multithread_load or num_threads in # --model-loader-extra-config opts out (the latter is consumed # only by the multi-threaded iterator, so it signals intent); # e.g. local NVMe, where prefetch is a no-op and multi-threading # helps. if ( weight_loader_prefetch and not weight_loader_disable_mmap and self.load_config.load_format != LoadFormat.FASTSAFETENSORS and use_multithread and not ( {"enable_multithread_load", "num_threads"} & extra_config.keys() ) ): logger.warning( "--weight-loader-prefetch-checkpoints is enabled; falling " "back to single-threaded weight loading to avoid I/O " "oversubscription with the prefetch threads. Set " "enable_multithread_load=true in --model-loader-extra-config " "to keep multi-threaded loading." ) use_multithread = False if self.load_config.load_format == LoadFormat.FASTSAFETENSORS: weights_iterator = fastsafetensors_weights_iterator( hf_weights_files, ) elif use_multithread: weights_iterator = buffered_multi_thread_safetensors_weights_iterator( hf_weights_files, max_workers=extra_config.get( "num_threads", self.DEFAULT_NUM_THREADS ), disable_mmap=weight_loader_disable_mmap, prefetch=weight_loader_prefetch, prefetch_num_threads=prefetch_num_threads, drop_cache_after_load=weight_loader_drop_cache_after_load, ) else: weights_iterator = safetensors_weights_iterator( hf_weights_files, disable_mmap=weight_loader_disable_mmap, prefetch=weight_loader_prefetch, prefetch_num_threads=prefetch_num_threads, drop_cache_after_load=weight_loader_drop_cache_after_load, ) else: if use_multithread: weights_iterator = multi_thread_pt_weights_iterator( hf_weights_files, max_workers=extra_config.get( "num_threads", self.DEFAULT_NUM_THREADS ), ) else: weights_iterator = pt_weights_iterator(hf_weights_files) if self.load_config.draft_model_idx is not None: return self._filter_mtp_weights( weights_iterator, source.prefix, self.load_config.draft_model_idx ) if self.counter_before_loading_weights == 0.0: self.counter_before_loading_weights = time.perf_counter() # Apply the prefix. return ((source.prefix + name, tensor) for (name, tensor) in weights_iterator) @classmethod def _filter_mtp_weights( cls, weights_iterator, prefix: str, draft_model_idx: int ) -> Generator[Tuple[str, torch.Tensor], None, None]: """Filter MTP weights to keep only the specified draft model layer and remap it to layer 0. Yields lazily so the upstream buffered iterator's sliding window actually bounds CPU memory — eager materialization caused page-reclaim hangs on large MoE checkpoints with multi-layer EAGLE.""" for name, tensor in weights_iterator: match = cls._MTP_PATTERN.match(name) if match is not None: idx = int(match.group(1)) if idx != draft_model_idx: continue new_name = name.replace(match.group(), "model.mtp.layers.0.") else: new_name = name yield (prefix + new_name, tensor) def _get_all_weights( self, model_config: ModelConfig, model: nn.Module, ) -> Generator[Tuple[str, torch.Tensor], None, None]: primary_weights = DefaultModelLoader.Source.init_new(model_config, model) yield from self._get_weights_iterator(primary_weights) secondary_weights = cast( Iterable[DefaultModelLoader.Source], getattr(model, "secondary_weights", ()) ) for source in secondary_weights: yield from self._get_weights_iterator(source) def download_model(self, model_config: ModelConfig) -> None: self._prepare_weights( model_config.model_path, model_config.revision, fall_back_to_pt=True ) def _load_modelopt_base_model(self, model_config: ModelConfig) -> nn.Module: """Load and prepare the base model for ModelOpt quantization. This method handles the common model loading logic shared between DefaultModelLoader (conditional) and ModelOptModelLoader (dedicated). """ if not HAS_ACCELERATE: raise ImportError( "accelerate is required for ModelOpt quantization. " "Please install it with: pip install accelerate" ) try: hf_config = AutoConfig.from_pretrained( model_config.model_path, trust_remote_code=True, local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE, ) except (KeyError, ValueError): from sglang.srt.utils.hf_transformers_utils import get_config hf_config = get_config( model_config.model_path, trust_remote_code=True, ) with init_empty_weights(): torch_dtype = getattr(hf_config, "torch_dtype", torch.float16) model = AutoModelForCausalLM.from_config( hf_config, torch_dtype=torch_dtype, trust_remote_code=True ) max_memory = get_max_memory() inferred_device_map = infer_auto_device_map(model, max_memory=max_memory) on_cpu = "cpu" in inferred_device_map.values() model_kwargs = {"torch_dtype": "auto"} device_map = "auto" if on_cpu: for device in max_memory.keys(): if isinstance(device, int): max_memory[device] *= DEFAULT_GPU_MEMORY_FRACTION_FOR_CALIBRATION logger.warning( "Model does not fit to the GPU mem. " f"We apply the following memory limit for calibration: \n{max_memory}\n" f"If you hit GPU OOM issue, please adjust the memory fraction " f"(currently {DEFAULT_GPU_MEMORY_FRACTION_FOR_CALIBRATION}) or " "reduce the calibration `batch_size` manually." ) model_kwargs["max_memory"] = max_memory model = AutoModelForCausalLM.from_pretrained( model_config.model_path, config=hf_config, device_map=device_map, **model_kwargs, trust_remote_code=True, local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE, ) # Handle both legacy modelopt_quant and unified quantization flags if hasattr(model_config, "modelopt_quant") and model_config.modelopt_quant: # Legacy approach quant_choice_str = model_config.modelopt_quant rank0_log(f"ModelOpt quantization requested (legacy): {quant_choice_str}") else: # Unified approach - extract quantization type quant_choice_str = model_config._get_modelopt_quant_type() rank0_log( f"ModelOpt quantization requested (unified): {model_config.quantization} -> {quant_choice_str}" ) if not isinstance(quant_choice_str, str): raise TypeError( f"Quantization type must be a string (e.g., 'fp8'), " f"got {type(quant_choice_str)}" ) return model def load_model( self, *, model_config: ModelConfig, device_config: DeviceConfig, ) -> nn.Module: if hasattr(model_config, "modelopt_quant") and model_config.modelopt_quant: # Load base model using shared method model = self._load_modelopt_base_model(model_config) # Note: DefaultModelLoader doesn't do additional quantization processing # For full ModelOpt quantization, use ModelOptModelLoader return model.eval() target_device = torch.device(device_config.device) quant_config = _get_quantization_config(model_config, self.load_config) with set_default_torch_dtype(model_config.dtype): with target_device: model = _initialize_model( model_config, self.load_config, quant_config, ) self.load_weights_and_postprocess( model, self._get_all_weights(model_config, model), target_device ) self.counter_after_loading_weights = time.perf_counter() return model.eval() @staticmethod def load_weights_and_postprocess(model, weights, target_device): # Used in tests to verify memory savings when using online quantization. if is_cuda_alike(): peak_memory = torch.cuda.max_memory_allocated() logger.debug( "Peak GPU memory before loading weights: %s GiB", f"{peak_memory / GIB_BYTES:.3f}", ) memory_start = get_available_gpu_memory( target_device.type, gpu_id=torch.cuda.current_device() ) quant_config = getattr(model, "quant_config", None) is_nvfp4_online = getattr(quant_config, "is_nvfp4_online", False) if is_nvfp4_online: # Scope exact FP4 quantization math to load-time conversion only; # restore the original environment before serving starts. with temp_set_env( TRTLLM_DISABLE_FP4_QUANT_FAST_MATH="1", FLASHINFER_DISABLE_FP4_QUANT_FAST_MATH="1", ): model.load_weights(weights) if target_device.type == "cuda": torch.cuda.synchronize() torch.cuda.empty_cache() else: model.load_weights(weights) # Used in tests to verify memory savings when using online quantization. if is_cuda_alike(): memory_end = get_available_gpu_memory( target_device.type, gpu_id=torch.cuda.current_device() ) logger.debug( "Memory increase during load_weights: %s GiB", f"{memory_start - memory_end:.3f}", ) for _, module in model.named_modules(): quant_method = getattr(module, "quant_method", None) if quant_method is not None: # When quant methods need to process weights after loading # (for repacking, quantizing, etc), they expect parameters # to be on the global target device. This scope is for the # case where cpu offloading is used, where we will move the # parameters onto device for processing and back off after. with device_loading_context(module, target_device): quant_method.process_weights_after_loading(module) class LayeredModelLoader(DefaultModelLoader): """Model loader that loads weights layer by layer so that one can quantize a layer before loading another to make the peak memory envelope smaller.""" def __init__(self, load_config: LoadConfig): # Back to the default load format load_config.load_format = LoadFormat.AUTO super().__init__(load_config) def load_model( self, *, model_config: ModelConfig, device_config: DeviceConfig, ) -> nn.Module: from sglang.srt.layers.torchao_utils import apply_torchao_config_to_model from sglang.srt.runtime_context import get_server_args torchao_config = get_server_args().torchao_config target_device = torch.device(device_config.device) quant_config = _get_quantization_config(model_config, self.load_config) with set_default_torch_dtype(model_config.dtype): # Create model on meta device with torch.device("meta"): model = _initialize_model( model_config, self.load_config, quant_config, ) # Check model's layered load support if not hasattr(model, "load_weights_to_module"): raise ValueError( "LayeredModelLoader requires the model to have a " "`load_weights_to_module` method. " f"{model_config.model_path} does not support it." ) # Get all weights from disk weights = self._get_all_weights(model_config, model) # Helper function to recursively fill the weights of a module def fill_module(module, fqn: List[str], weights): """ fqn: list of strings representing the fully qualified name of `module`. """ # Layer by layer for name, submod in module.named_children(): fill_module(submod, fqn + [name], weights) # First materialize on target device module.to_empty(device=target_device, recurse=False) fqn_path = ".".join(fqn) # Fill weights model.load_weights_to_module( fqn_path, weights, ) # Quantize weights if applicable if torchao_config and "proj" in fqn_path: # Note: `None` here is needed to indicate no filter, see # `apply_torchao_config_to_model` for details. apply_torchao_config_to_model(module, torchao_config, None) # Start calling on root module fill_module(model, [], weights) if torchao_config: model.torchao_applied = True return model.eval() class QuantizedRLModelLoader(DefaultModelLoader): """ Model loader for RL training with FP8 quantization (profile-free, native SGLang). Workflow: 1. Initial load: Load base model → Record state → Apply FP8 quantization 2. Training Actor in full precision 3. Reload: Trainer sends full precision weights → Quantize to FP8 → Copy to original memory 4. Use torch.as_strided to preserve memory locations across reloads Usage: --model-path Qwen/Qwen2.5-7B --quantization fp8 --load-format flash_rl """ # Parameter attributes to record for weight reloading RECORDED_LOADER_KEYS = [ "weight_loader", "load_qkv_weight", "load_column_parallel_weight", "load_row_parallel_weight", "load_merged_column_weight", "output_dim", "input_dim", "_assert_and_load", ] # Parameters to skip during FP8 quantization (matches FlashRL's exclude_list) SKIP_QUANTIZATION_PARAMS = [ "weight_scale", "input_scale", "output_scale", ".bias", "lm_head.weight", "model.norm.weight", "embed_tokens", # BF16 params "rotary_emb.inv_freq", "rotary_emb.cos_cached", "rotary_emb.sin_cached", "projector", "input_layernorm.weight", "post_attention_layernorm.weight", # LayerNorms ] # Stacked parameters (Qwen2): shards loaded separately, then combined STACKED_PARAMS_MAPPING = [ ("qkv_proj", ["q_proj", "k_proj", "v_proj"]), ("gate_up_proj", ["gate_proj", "up_proj"]), ] _QKV_SHARD_ALIASES = { "q_proj": "q", "k_proj": "k", "v_proj": "v", } def __init__(self, load_config: LoadConfig): super().__init__(load_config) logger.info("[QuantizedRL] Profile-free FP8 quantization enabled") self._initial_load_complete = False def _prepare_weights( self, model_name_or_path: str, revision: Optional[str], fall_back_to_pt: bool ): """Standard weight preparation using base model path.""" logger.info(f"[QuantizedRL] Loading from base model: {model_name_or_path}") temp_config = LoadConfig(load_format=LoadFormat.AUTO) temp_loader = DefaultModelLoader(temp_config) return temp_loader._prepare_weights( model_name_or_path, revision, fall_back_to_pt ) @staticmethod def _bind_method_to_cls(func, obj): """Bind function to object instance (for weight_loader methods).""" import types if hasattr(func, "__self__") or not callable(func): return func return types.MethodType(func, obj) def load_weights_and_postprocess(self, model, weights, target_device): """ Initial load: Load BF16 → Record state → Apply FP8 quantization. Called ONCE during model initialization. """ logger.info("[QuantizedRL] Initial load with FP8 quantization") original_load_weights = model.load_weights def load_weights_proxy(weights): if QuantizedRLModelLoader.is_reload_scenario(model): logger.info("[QuantizedRL] Using fast path reload in load_weights") QuantizedRLModelLoader.rebinding_and_load_weights( model, original_load_weights, weights ) else: original_load_weights(weights) model.load_weights = load_weights_proxy model.load_weights(weights) original_weights = dict(model.named_parameters()) # Record pre-quantization state (shape/stride) for torch.as_strided reset model.original_weights_rebuild_keys = {} for name, p in original_weights.items(): model.original_weights_rebuild_keys[name] = { "shape": p.shape, "stride": p.stride(), "dtype": p.dtype, "nbytes": p.untyped_storage().nbytes(), } # Record parameter attributes (weight_loader, etc.) before quantization recorded_loader = { k: dict() for k in QuantizedRLModelLoader.RECORDED_LOADER_KEYS } for name, p in original_weights.items(): for key in QuantizedRLModelLoader.RECORDED_LOADER_KEYS: if hasattr(p, key): attr = getattr(p, key) if not callable(attr): recorded_loader[key][name] = attr elif hasattr(attr, "__self__") and p is attr.__self__: recorded_loader[key][name] = attr.__func__ # Store unbound else: recorded_loader[key][name] = attr model.recorded_loader = recorded_loader # Apply FP8 quantization (creates new Parameters, loses attributes) for _, module in model.named_modules(): quant_method = getattr(module, "quant_method", None) if quant_method is not None: with device_loading_context(module, target_device): quant_method.process_weights_after_loading(module) model.flash_rl_initial_load_complete = True self._initial_load_complete = True logger.info("[QuantizedRL] Initial load complete") @staticmethod def is_reload_scenario(model): """Check if model is ready for reloading (initial load completed).""" return ( hasattr(model, "original_weights_rebuild_keys") and hasattr(model, "recorded_loader") and getattr(model, "flash_rl_initial_load_complete", False) ) @staticmethod def _is_stacked_param(name): """Check if parameter is stacked (qkv_proj, gate_up_proj).""" for stacked_name, _ in QuantizedRLModelLoader.STACKED_PARAMS_MAPPING: if stacked_name in name: return True return False @staticmethod def _resolve_stacked_info(name: str) -> Tuple[str, Optional[str], Optional[Any]]: for target, shard_names in QuantizedRLModelLoader.STACKED_PARAMS_MAPPING: for idx, shard in enumerate(shard_names): if shard in name: shard_id = ( QuantizedRLModelLoader._QKV_SHARD_ALIASES.get(shard, shard) if target == "qkv_proj" else idx ) return name.replace(shard, target), target, shard_id return name, None, None @staticmethod def _store_quantized_scale( scale_store: Dict[str, Union[torch.Tensor, Dict[Any, torch.Tensor]]], name: str, scale: torch.Tensor, ) -> None: param_name, stacked_key, shard_id = ( QuantizedRLModelLoader._resolve_stacked_info(name) ) if stacked_key is None: scale_store[param_name] = scale else: shard_dict = scale_store.setdefault(param_name, {}) assert isinstance(shard_dict, dict) shard_dict[shard_id] = scale @staticmethod def _apply_scale_update( all_params: Dict[str, torch.nn.Parameter], param_name: str, scale_info: Union[torch.Tensor, Dict[Any, torch.Tensor], None], ) -> None: if scale_info is None: return # Get tp rank and size tp_rank = get_parallel().tp_rank tp_size = get_parallel().tp_size def _get_tp_sharded_scale(full_scale_tensor): """Get tp sharded scale from full scale tensor""" if tp_size == 1: return full_scale_tensor full_dim = full_scale_tensor.shape[0] shard_dim = full_dim // tp_size start_idx = tp_rank * shard_dim end_idx = start_idx + shard_dim return full_scale_tensor[start_idx:end_idx] if param_name.endswith(".weight"): scale_param_name = f"{param_name[:-7]}.weight_scale" else: scale_param_name = f"{param_name}.weight_scale" scale_param = all_params.get(scale_param_name) if scale_param is None: logger.warning( "[QuantizedRL] Scale parameter not found: %s", scale_param_name ) return if isinstance(scale_info, torch.Tensor): new_scale = scale_info.t().contiguous() if scale_param.data.shape == new_scale.shape: scale_param.data.copy_(new_scale) else: logger.warning( "[QuantizedRL] Scale shape mismatch for %s: expected %s, got %s", scale_param_name, scale_param.data.shape, new_scale.shape, ) else: stacked_key = next( ( target for target, _ in QuantizedRLModelLoader.STACKED_PARAMS_MAPPING if target in param_name ), None, ) shard_names = next( ( names for target, names in QuantizedRLModelLoader.STACKED_PARAMS_MAPPING if target == stacked_key ), [], ) rows_per_shard = scale_param.data.shape[-1] // max(len(shard_names), 1) if rows_per_shard * len(shard_names) != scale_param.data.shape[-1]: logger.warning( f"Scale param shape {scale_param.data.shape[-1]} not divisible by {len(shard_names)}" ) offset = 0 for idx, shard in enumerate(shard_names): shard_id = ( QuantizedRLModelLoader._QKV_SHARD_ALIASES.get(shard, shard) if stacked_key == "qkv_proj" else idx ) shard_scale = scale_info.get(shard_id) shard_scale = _get_tp_sharded_scale(shard_scale) if shard_scale is None: offset += rows_per_shard continue shard_rows = shard_scale.shape[0] start = offset end = start + shard_rows scale_param.data[..., start:end] = shard_scale.t().contiguous() offset = end @staticmethod def rebinding_and_load_weights(model, first_time_load_weights, weights): """ Reload: VERL sends BF16 → Quantize to FP8 → Copy to original memory. Flow: Reset params → Restore attributes → Quantize in iterator → Load → Copy back """ logger.info("[QuantizedRL] Reload: Updating weights with FP8 quantization") weights_list = list(weights) updated_param_names, is_last_update = ( QuantizedRLModelLoader._get_updated_params(weights_list, model) ) # Save current FP8 parameter data pointers existing_params = dict(model.named_parameters()) current_param_data = {} for name in updated_param_names: if name in existing_params: current_param_data[name] = existing_params[name].data # Reset to pre-quantization shape using torch.as_strided # Keeps same storage, just changes view - critical for memory preservation for name, rebuild_info in model.original_weights_rebuild_keys.items(): if name in updated_param_names and name in existing_params: existing_params[name].data = torch.as_strided( # Note: avoid clone here existing_params[name].data.clone(), rebuild_info["shape"], rebuild_info["stride"], ) # Restore weight loader attributes (only if missing) for k, loader_dict in model.recorded_loader.items(): for param_name, loader in loader_dict.items(): if param_name in updated_param_names and param_name in existing_params: param = existing_params[param_name] if not hasattr(param, k): if callable(loader): if hasattr(loader, "__self__"): setattr(param, k, loader) else: setattr( param, k, QuantizedRLModelLoader._bind_method_to_cls( loader, param ), ) else: setattr(param, k, loader) del existing_params # Quantize BF16 weights to FP8 in iterator (before weight_loader) # Store scales for later update quantized_scales: Dict[str, Union[torch.Tensor, Dict[Any, torch.Tensor]]] = {} def quantize_weights_iterator(weights_iter): """Quantize individual shards before weight_loader stacks them.""" from sglang.srt.layers.quantization.fp8_kernel import ( per_token_group_quant_fp8, ) for name, weight in weights_iter: if any( skip in name for skip in QuantizedRLModelLoader.SKIP_QUANTIZATION_PARAMS ): logger.info(f"[QuantizedRL] Skip: {name} ({weight.dtype})") yield (name, weight) elif weight.dtype in [torch.bfloat16, torch.float32, torch.float16]: qweight, scale = per_token_group_quant_fp8(weight, weight.shape[-1]) logger.info(f"[QuantizedRL] Quantize: {name} {weight.dtype}→FP8") QuantizedRLModelLoader._store_quantized_scale( quantized_scales, name, scale ) yield (name, qweight) else: logger.info(f"[QuantizedRL] Keep: {name} ({weight.dtype})") yield (name, weight) # Load quantized weights (weight_loader stacks FP8 shards) first_time_load_weights(quantize_weights_iterator(iter(weights_list))) # Copy back to original FP8 memory locations and update scales all_params = dict(model.named_parameters()) for name in updated_param_names: if name not in all_params or name not in current_param_data: continue if any( skip in name for skip in QuantizedRLModelLoader.SKIP_QUANTIZATION_PARAMS ): continue new_param = all_params[name] old_fp8_data = current_param_data[name] # Handle embeddings/lm_head (BF16) and quantized weights (FP8) if "embed_tokens" in name or "lm_head" in name: old_fp8_data.copy_(new_param.data) new_param.data = old_fp8_data elif ( new_param.dtype == torch.float8_e4m3fn and old_fp8_data.dtype == torch.float8_e4m3fn ): # FP8: Use strided view for transposed storage strided_data = torch.as_strided( new_param.data, old_fp8_data.shape, old_fp8_data.stride() ) old_fp8_data.copy_(strided_data) new_param.data = old_fp8_data QuantizedRLModelLoader._apply_scale_update( all_params, name, quantized_scales.get(name), ) elif new_param.dtype == old_fp8_data.dtype: # Same dtype (LayerNorm, etc.): Direct copy old_fp8_data.copy_(new_param.data) new_param.data = old_fp8_data else: raise RuntimeError( f"Unexpected dtype mismatch for {name}: " f"new={new_param.dtype}, old={old_fp8_data.dtype}" ) # Cleanup del current_param_data if is_last_update: gc.collect() current_platform.empty_cache() logger.info("[QuantizedRL] Reload complete") return updated_param_names, is_last_update @staticmethod def _get_updated_params(weights_list, model): """Identify which parameters need updating from incoming weights.""" stacked_params_mapping = [ ("qkv_proj", "q_proj", "q"), ("qkv_proj", "k_proj", "k"), ("qkv_proj", "v_proj", "v"), ("gate_up_proj", "gate_proj", 0), ("gate_up_proj", "up_proj", 1), ] params_dict = dict(model.named_parameters()) updated_params = set() is_last_update = False for name, _ in weights_list: if name == "lm_head.weight": is_last_update = True if any( skip in name for skip in QuantizedRLModelLoader.SKIP_QUANTIZATION_PARAMS ): continue from sglang.srt.layers.utils import get_layer_id # Skip params outside layer range (for pipeline parallelism) layer_id = get_layer_id(name) if ( layer_id is not None and hasattr(model, "start_layer") and (layer_id < model.start_layer or layer_id >= model.end_layer) ): continue # Skip tied embeddings and vision tower params if ( hasattr(model, "config") and model.config.tie_word_embeddings and "lm_head.weight" in name ): continue if name.startswith("model.vision_tower") and name not in params_dict: continue # Map stacked param shards (q/k/v_proj → qkv_proj) mapped = False for param_name, weight_name, shard_id in stacked_params_mapping: if weight_name in name: name = name.replace(weight_name, param_name) if name.endswith(".bias") and name not in params_dict: continue updated_params.add(name) mapped = True break if not mapped: if name.endswith(".bias") and name not in params_dict: continue if name in params_dict: updated_params.add(name) return list(updated_params), is_last_update class DummyModelLoader(BaseModelLoader): """Model loader that will set model weights to random values.""" def __init__(self, load_config: LoadConfig): super().__init__(load_config) if load_config.model_loader_extra_config: raise ValueError( f"Model loader extra config is not supported for " f"load format {load_config.load_format}" ) def download_model(self, model_config: ModelConfig) -> None: pass # Nothing to download def load_model( self, *, model_config: ModelConfig, device_config: DeviceConfig, ) -> nn.Module: if get_bool_env_var("SGL_CPU_QUANTIZATION"): return load_model_with_cpu_quantization( self, model_config=model_config, device_config=device_config ) quant_config = _get_quantization_config(model_config, self.load_config) with set_default_torch_dtype(model_config.dtype): with torch.device(device_config.device): model = _initialize_model( model_config, self.load_config, quant_config, ) # NOTE(woosuk): For accurate performance evaluation, we assign # random values to the weights. initialize_dummy_weights(model) _post_load_weights(model) for _, module in model.named_modules(): quant_method = getattr(module, "quant_method", None) if quant_method is not None: # Skip FusedMoE layers already quantized during init (FP8 or FP4) if ( hasattr(module, "is_weights_quantized") and module.is_weights_quantized() ): continue quant_method.process_weights_after_loading(module) return model.eval() class ShardedStateLoader(BaseModelLoader): """ Model loader that directly loads each worker's model state dict, which enables a fast load path for large tensor-parallel models where each worker only needs to read its own shard rather than the entire checkpoint. See `examples/runtime/engine/save_sharded_state.py` for creating a sharded checkpoint. """ DEFAULT_PATTERN = "model-rank-{rank}-part-{part}.safetensors" def __init__(self, load_config: LoadConfig): super().__init__(load_config) extra_config = ( {} if load_config.model_loader_extra_config is None else load_config.model_loader_extra_config.copy() ) self.pattern = extra_config.pop("pattern", self.DEFAULT_PATTERN) if extra_config: raise ValueError( f"Unexpected extra config keys for load format " f"{load_config.load_format}: " f"{load_config.model_loader_extra_config.keys()}" ) @staticmethod def _filter_subtensors(tensors: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]: """ Filter out all tensors that share the same memory or a subset of the memory of another tensor. """ same_storage_groups: Dict[Any, List[Tuple[str, torch.Tensor]]] = ( collections.defaultdict(list) ) for key, tensor in tensors.items(): if tensor.numel(): ptr = tensor.untyped_storage().data_ptr() same_storage_groups[tensor.device, ptr].append((key, tensor)) def get_end_ptr(tensor: torch.Tensor) -> int: return tensor.view(-1)[-1].data_ptr() + tensor.element_size() result: Dict[str, torch.Tensor] = {} for group in same_storage_groups.values(): for k, t in group: a, b = t.data_ptr(), get_end_ptr(t) for k2, t2 in group: if not t2.is_contiguous(): continue a2, b2 = t2.data_ptr(), get_end_ptr(t2) if a < a2 or b2 < b: continue if a2 < a or b < b2 or not t.is_contiguous(): break # t2 covers strictly more memory than t. if k2 < k: # Same tensors, keep the one with the smaller key. break else: result[k] = t return result def _prepare_weights(self, model_name_or_path: str, revision: Optional[str]): if os.path.isdir(model_name_or_path): return model_name_or_path else: allow_patterns = ["*.safetensors"] return download_weights_from_hf( model_name_or_path, self.load_config.download_dir, allow_patterns, revision, ignore_patterns=self.load_config.ignore_patterns, ) def download_model(self, model_config: ModelConfig) -> None: self._prepare_weights(model_config.model_path, model_config.revision) def load_model( self, *, model_config: ModelConfig, device_config: DeviceConfig, ) -> nn.Module: from safetensors.torch import safe_open local_model_path = self._prepare_weights( model_config.model_path, model_config.revision ) quant_config = _get_quantization_config(model_config, self.load_config) with set_default_torch_dtype(model_config.dtype): with torch.device(device_config.device): model = _initialize_model(model_config, self.load_config, quant_config) for _, module in model.named_modules(): quant_method = getattr(module, "quant_method", None) if quant_method is not None: quant_method.process_weights_after_loading(module) rank = get_parallel().tp_rank pattern = os.path.join( local_model_path, self.pattern.format(rank=rank, part="*"), ) filepaths = glob.glob(pattern) if not filepaths: # TODO: support un-sharded checkpoints too raise ValueError( f"Could not find checkpoint files '{pattern}', only " f"pre-sharded checkpoints are currently supported!" ) state_dict = self._filter_subtensors(model.state_dict()) for path in filepaths: with safe_open(path, framework="pt") as f: for key in f.keys(): # noqa: SIM118 tensor = f.get_tensor(key) # If loading with LoRA enabled, additional padding may # be added to certain parameters. We only load into a # narrowed view of the parameter data. param_data = state_dict[key].data param_shape = state_dict[key].shape for dim, size in enumerate(tensor.shape): if size < param_shape[dim]: param_data = param_data.narrow(dim, 0, size) if tensor.shape != param_shape: logger.warning( "loading tensor of shape %s into " "parameter '%s' of shape %s", tensor.shape, key, param_shape, ) param_data.copy_(tensor) state_dict.pop(key) if state_dict: raise ValueError(f"Missing keys {tuple(state_dict)} in loaded state!") _post_load_weights(model) return model.eval() @staticmethod def save_model( model: torch.nn.Module, path: str, pattern: Optional[str] = None, max_size: Optional[int] = None, ) -> None: from safetensors.torch import save_file if pattern is None: pattern = ShardedStateLoader.DEFAULT_PATTERN rank = get_parallel().tp_rank part_idx = 0 total_size = 0 state_dict = ShardedStateLoader._filter_subtensors(model.state_dict()) state_dict_part: Dict[str, torch.Tensor] = {} for key, tensor in state_dict.items(): param_size = tensor.nelement() * tensor.element_size() if max_size is not None and total_size + param_size > max_size: filename = pattern.format(rank=rank, part=part_idx) save_file( state_dict_part, os.path.join(path, filename), ) part_idx += 1 total_size = 0 state_dict_part = {} state_dict_part[key] = tensor total_size += param_size if len(state_dict_part) > 0: filename = pattern.format(rank=rank, part=part_idx) save_file( state_dict_part, os.path.join(path, filename), ) class BitsAndBytesModelLoader(BaseModelLoader): """Model loader to load model weights with BitAndBytes quantization.""" possible_config_file_names = ["adapter_config.json"] default_target_modules = [ ".gate_proj.", ".down_proj.", ".up_proj.", ".q_proj.", ".k_proj.", ".v_proj.", ".o_proj.", ".fc1.", ".fc2.", ".dense.", ".query_key_value.", ".qkv_proj.", ".dense_h_to_4h.", ".dense_4h_to_h.", ".out_proj.", ] def __init__(self, load_config: LoadConfig): super().__init__(load_config) # we don't need to quantize the whole model, only the target modules # that are specified in the adapter config file. If the adapter config # file is not provided, we will quantize the default modules. if ( not load_config.model_loader_extra_config or "qlora_adapter_name_or_path" not in load_config.model_loader_extra_config ): self.target_modules = [] return qlora_adapter = load_config.model_loader_extra_config[ "qlora_adapter_name_or_path" ] config_file_path = self._get_config_file(qlora_adapter) with open(config_file_path, "r") as f: config = json.load(f) self.target_modules = config["target_modules"] def _get_config_file(self, qlora_adapter: str) -> str: is_local = os.path.isdir(qlora_adapter) config_file_path = None if is_local: for file in self.possible_config_file_names: config_file_path = os.path.join(qlora_adapter, file) if os.path.exists(config_file_path): break else: hf_api = HfApi() repo_files = hf_api.list_repo_files(repo_id=qlora_adapter) for file in self.possible_config_file_names: if file in repo_files: config_file_path = hf_hub_download( repo_id=qlora_adapter, filename=file ) break if not config_file_path: raise ValueError(f"Cannot find adapter config file in {qlora_adapter}") return config_file_path def _get_weight_files( self, model_name_or_path: str, allowed_patterns: List[str], revision: Optional[str] = None, ) -> Tuple[List[str], str]: """Retrieve weight files. Download the files if necessary. Return the weight files and the file pattern.""" is_local = os.path.isdir(model_name_or_path) if is_local: for pattern in allowed_patterns: weight_files = glob.glob(os.path.join(model_name_or_path, pattern)) if weight_files: return weight_files, pattern else: hf_api = HfApi() repo_files = hf_api.list_repo_files(repo_id=model_name_or_path) for pattern in allowed_patterns: matching_files = fnmatch.filter(repo_files, pattern) if matching_files: hf_folder = download_weights_from_hf( model_name_or_path, self.load_config.download_dir, [pattern], revision, ignore_patterns=self.load_config.ignore_patterns, ) return glob.glob(os.path.join(hf_folder, pattern)), pattern raise RuntimeError(f"No model weights found in: `{model_name_or_path}`") def _prepare_weights( self, model_name_or_path: str, revision: Optional[str] ) -> Tuple[List[str], bool]: """Prepare weight files for the model.""" allowed_patterns = ["*.safetensors", "*.bin", "*.pt"] hf_weights_files, matched_pattern = self._get_weight_files( model_name_or_path, allowed_patterns, revision ) if matched_pattern != "*.safetensors": hf_weights_files = filter_files_not_needed_for_inference(hf_weights_files) if len(hf_weights_files) == 0: raise RuntimeError( f"Cannot find any model weights with `{model_name_or_path}`" ) return hf_weights_files, matched_pattern == "*.safetensors" def _hf_weight_iter(self, hf_weights_files, use_safetensors: bool): if use_safetensors: return safetensors_weights_iterator(hf_weights_files) else: return pt_weights_iterator(hf_weights_files) def _get_quantized_weights_iterator( self, model_name_or_path: str, revision: Optional[str], pre_quant: bool, load_8bit: bool, ) -> Tuple[Generator[Tuple[str, torch.Tensor], None, None], Dict[str, Any]]: """Get an iterator to the model weights with bitsandbytes quantization, as well as the quantization state dictionary.""" # only load the bitsandbytes module when needed try: import bitsandbytes if bitsandbytes.__version__ < "0.44.0": raise ImportError( "bitsandbytes version is wrong. Please " "install bitsandbytes>=0.44.0." ) except ImportError as err: raise ImportError( "Please install bitsandbytes>=0.44.0 via " "`pip install bitsandbytes>=0.44.0` to use " "bitsandbytes quantizer." ) from err hf_weights_files, use_safetensors = self._prepare_weights( model_name_or_path, revision ) quant_state_dict: Dict[str, Any] = {} if pre_quant: if load_8bit: return ( self._quantized_8bit_generator( hf_weights_files, use_safetensors, quant_state_dict ), quant_state_dict, ) else: return ( self._quantized_4bit_generator( hf_weights_files, use_safetensors, quant_state_dict ), quant_state_dict, ) return ( self._unquantized_generator( hf_weights_files, use_safetensors, quant_state_dict ), quant_state_dict, ) def _is_8bit_weight_name(self, weight_name: str): quantized_suffix = {".scb", ".weight_format"} return any(weight_name.lower().endswith(suffix) for suffix in quantized_suffix) def _is_4bit_weight_name(self, weight_name: str): quantized_suffix = { "absmax", "quant_map", "nested_absmax", "nested_quant_map", "bitsandbytes", } suffix = weight_name.split(".")[-1] return any(q_suffix in suffix for q_suffix in quantized_suffix) def _quantized_8bit_generator( self, hf_weights_files, use_safetensors, quant_state_dict ) -> Generator: for weight_name, weight_tensor in self._hf_weight_iter( hf_weights_files, use_safetensors ): if not weight_name.lower().endswith(".scb"): continue weight_key = weight_name.lower().replace(".scb", ".weight") quant_state_dict[weight_key] = weight_tensor for weight_name, weight_tensor in self._hf_weight_iter( hf_weights_files, use_safetensors ): if self._is_8bit_weight_name(weight_name): continue if weight_name in quant_state_dict: set_weight_attrs(weight_tensor, {"load_in_8bit": True}) yield weight_name, weight_tensor else: yield weight_name, weight_tensor def _quantized_4bit_generator( self, hf_weights_files, use_safetensors, quant_state_dict ) -> Generator: from bitsandbytes.functional import QuantState # First iterate over all quant state weights weight_iterator = self._hf_weight_iter(hf_weights_files, use_safetensors) temp_state_dict = {} for weight_name, weight_tensor in weight_iterator: if not self._is_4bit_weight_name(weight_name): continue # bitsandbytes library requires # weight.quant_state.bitsandbytes__* in CPU if "quant_state.bitsandbytes" in weight_name: temp_state_dict[weight_name] = weight_tensor.cpu().data else: temp_state_dict[weight_name] = weight_tensor # Closure to parse quant_state for each prequant weight def _parse_quant_state(param_name: str, temp_state_dict: Dict) -> QuantState: quant_state = {} for k in temp_state_dict: if param_name + "." in k: quant_state[k] = temp_state_dict[k] return QuantState.from_dict(quant_state, device="cuda") # Second iterate over all prequant and normal weights # pre quantized weights would have a quant_state for weight_name, weight_tensor in self._hf_weight_iter( hf_weights_files, use_safetensors ): if self._is_4bit_weight_name(weight_name): continue if (f"{weight_name}.quant_state.bitsandbytes__nf4" in temp_state_dict) or ( f"{weight_name}.quant_state.bitsandbytes__fp4" in temp_state_dict ): quant_state = _parse_quant_state(weight_name, temp_state_dict) quant_state_dict[weight_name] = quant_state yield weight_name, weight_tensor else: yield weight_name, weight_tensor def _unquantized_generator( self, hf_weights_files, use_safetensors, quant_state_dict ) -> Generator: from bitsandbytes.functional import quantize_4bit tp_size = get_parallel().tp_size tp_rank = get_parallel().tp_rank for weight_name, weight_tensor in self._hf_weight_iter( hf_weights_files, use_safetensors ): if any( target_module in weight_name for target_module in self.target_modules ) and weight_name.endswith(".weight"): weight_name = weight_name.replace(".weight", ".qweight") if any( module in weight_name for module in self.column_parallel_weights_modules ): total_size = weight_tensor.size(-1) start_index = total_size // tp_size * tp_rank end_index = total_size // tp_size * (tp_rank + 1) weight_sub_tensor = weight_tensor[..., start_index:end_index] else: total_size = weight_tensor.size(0) start_index = total_size // tp_size * tp_rank end_index = total_size // tp_size * (tp_rank + 1) weight_sub_tensor = weight_tensor[start_index:end_index, ...] # bitsandbytes requires data in GPU if weight_sub_tensor.is_cuda: loaded_weight = weight_sub_tensor else: loaded_weight = weight_sub_tensor.cuda() # remove the following after the issue is fixed: # https://github.com/bitsandbytes-foundation/bitsandbytes/issues/1342 if loaded_weight.is_contiguous() is False: loaded_weight = loaded_weight.contiguous() with set_default_torch_dtype(torch.float32): processed_weight, quant_state = quantize_4bit( loaded_weight, compress_statistics=True, quant_type="nf4" ) quant_state_dict[weight_name] = quant_state else: processed_weight = weight_tensor yield weight_name, processed_weight def _load_weights(self, model_config: ModelConfig, model: nn.Module) -> None: if not hasattr(model, "load_weights"): raise AttributeError( "The required method 'load_weights' is not defined in class" f" {type(model).__name__}." ) if not hasattr(model, "bitsandbytes_stacked_params_mapping"): raise AttributeError( f"Model {type(model).__name__} does not support BitsAndBytes " "quantization yet." ) if len(self.target_modules) == 0: if hasattr(model, "default_bitsandbytes_target_modules"): self.target_modules = model.default_bitsandbytes_target_modules else: self.target_modules = self.default_target_modules if hasattr(model, "column_parallel_weights_modules"): self.column_parallel_weights_modules = model.column_parallel_weights_modules else: self.column_parallel_weights_modules = [] self.model_type = type(model).__name__ logger.info( "Loading weights with BitsAndBytes quantization. " " May take a while ..." ) quant_config = getattr(model_config.hf_config, "quantization_config", None) pre_quant = False if quant_config is not None: quant_method = quant_config.get("quant_method") if quant_method == "bitsandbytes": pre_quant = True else: raise ValueError( f"BitsAndBytes loader does not support {quant_method} " "quantization" ) # The quant_states in pre_quantized models cannot work with a split # weight tensor. So TP does not work with pre_quantized bnb models. if pre_quant and get_parallel().tp_size > 1: raise ValueError( "Prequant BitsAndBytes models with TP is not supported." "Please try with PP." ) load_8bit = False if pre_quant: load_8bit = quant_config.get("load_in_8bit", False) qweight_iterator, quant_state_dict = self._get_quantized_weights_iterator( model_config.model_path, model_config.revision, pre_quant, load_8bit ) model.load_weights(qweight_iterator) current_platform.empty_cache() param_dict = dict(model.named_parameters()) stacked_quant_state_dict: Dict[str, Dict[int, Any]] = {} model_type = model_config.hf_config.model_type for quant_param_name in quant_state_dict: non_stacked_param_name = quant_param_name if model_type == "mllama" and "vision_model" in quant_param_name: # adapt to VisionAttention quant_param_name = quant_param_name.replace( "self_attn.o_proj", "self_attn.proj" ) shard_index = 0 for shard_name, ( weight_name, index, ) in model.bitsandbytes_stacked_params_mapping.items(): if ( model_type in ["qwen2_vl", "qwen2_5_vl"] and "visual" in quant_param_name ): break if shard_name in quant_param_name: shard_index = index quant_param_name = quant_param_name.replace(shard_name, weight_name) break if ( model_type in ["qwen2_vl", "qwen2_5_vl"] and "visual" in quant_param_name ): quant_param_name = quant_param_name.replace( r"attn.qkv.", r"attn.qkv_proj." ) if quant_param_name not in param_dict: raise ValueError( f"Parameter {quant_param_name} not found in the model." ) if quant_param_name not in stacked_quant_state_dict: stacked_quant_state_dict[quant_param_name] = {} stacked_quant_state_dict[quant_param_name][shard_index] = quant_state_dict[ non_stacked_param_name ] # save quant_states and offsets as the attributes of the parameters for param_name, param in param_dict.items(): if param_name in stacked_quant_state_dict: quant_states = stacked_quant_state_dict[param_name] set_weight_attrs(param, {"bnb_quant_state": quant_states}) pack_ratio = getattr(param, "pack_factor", -1) if pack_ratio == -1: raise ValueError(f"pack_factor not set for parameter {param_name}.") num_elements = [0] * len(quant_states) for seq, quant_state in quant_states.items(): num_elements[seq] = math.prod(quant_state.shape) // pack_ratio offsets = np.concatenate(([0], np.cumsum(num_elements))) # Make torch infer_schema happy(Compatible with vLLM) offsets = torch.tensor(offsets).cpu() set_weight_attrs(param, {"bnb_shard_offsets": offsets}) if load_8bit: set_weight_attrs( param, {"matmul_state": [None] * len(quant_states)} ) def download_model(self, model_config: ModelConfig) -> None: self._prepare_weights(model_config.model_path, model_config.revision) def load_model( self, *, model_config: ModelConfig, device_config: DeviceConfig, ) -> nn.Module: quant_config = _get_quantization_config(model_config, self.load_config) with set_default_torch_dtype(model_config.dtype): with torch.device(device_config.device): model = _initialize_model( model_config, self.load_config, quant_config, ) self._load_weights(model_config, model) return model.eval() class GGUFModelLoader(BaseModelLoader): """ Model loader that can load GGUF files. This is useful for loading models that are quantized with GGUF and saved in the GGUF format. This loader supports loading both full models and sharded models. """ def __init__(self, load_config: LoadConfig): super().__init__(load_config) if load_config.model_loader_extra_config: raise ValueError( f"Model loader extra config is not supported for " f"load format {load_config.load_format}" ) def _prepare_weights(self, model_name_or_path: str): if os.path.isfile(model_name_or_path): return model_name_or_path else: raise ValueError(f"{model_name_or_path} is not a file.") def _get_gguf_weights_map(self, model_config: ModelConfig): """ GGUF uses this naming convention for their tensors from HF checkpoint: `blk.N.BB.weight` and `blk.N.BB.bias` where N signifies the block number of a layer, and BB signifies the attention/mlp layer components. See "Standardized tensor names" in https://github.com/ggerganov/ggml/blob/master/docs/gguf.md for details. """ # only load the gguf module when needed try: import gguf # FIXME: add version check for gguf except ImportError as err: raise ImportError( "Please install gguf via `pip install gguf` to use gguf quantizer." ) from err config = model_config.hf_config model_type = config.model_type # hack: ggufs have a different name than transformers if model_type == "cohere": model_type = "command-r" elif model_type == "qwen3_moe": model_type = "qwen3moe" arch = None for key, value in gguf.MODEL_ARCH_NAMES.items(): if value == model_type: arch = key break if arch is None: raise RuntimeError(f"Unknown gguf model_type: {model_type}") num_layers = config.num_hidden_layers name_map = gguf.get_tensor_name_map(arch, num_layers) with torch.device("meta"): dummy_model = AutoModelForCausalLM.from_config(config) state_dict = dummy_model.state_dict() gguf_to_hf_name_map = {} for hf_name in state_dict: name, suffix = hf_name.rsplit(".", 1) gguf_name = name_map.get_name(name) gguf_to_hf_name_map[f"{gguf_name}.{suffix}"] = hf_name return gguf_to_hf_name_map def _get_weights_iterator( self, model_name_or_path: str, gguf_to_hf_name_map: Dict[str, str] ) -> Generator[Tuple[str, torch.Tensor], None, None]: return gguf_quant_weights_iterator(model_name_or_path, gguf_to_hf_name_map) def download_model(self, model_config: ModelConfig) -> None: self._prepare_weights(model_config.model_path) def load_model( self, *, model_config: ModelConfig, device_config: DeviceConfig, ) -> nn.Module: local_model_path = self._prepare_weights(model_config.model_path) gguf_weights_map = self._get_gguf_weights_map(model_config) # we can only know if tie word embeddings after mapping weights if "lm_head.weight" in get_gguf_extra_tensor_names( local_model_path, gguf_weights_map ): model_config.hf_config.update({"tie_word_embeddings": True}) target_device = torch.device(device_config.device) quant_config = _get_quantization_config(model_config, self.load_config) with set_default_torch_dtype(model_config.dtype): with target_device: model = _initialize_model(model_config, self.load_config, quant_config) model.load_weights( self._get_weights_iterator(local_model_path, gguf_weights_map) ) for _, module in model.named_modules(): quant_method = getattr(module, "quant_method", None) if quant_method is not None: with device_loading_context(module, target_device): quant_method.process_weights_after_loading(module) return model class RemoteInstanceModelLoader(BaseModelLoader): """Model loader that can load Tensors from remote sglang instance.""" def __init__(self, load_config: LoadConfig): super().__init__(load_config) if load_config.model_loader_extra_config: raise ValueError( f"Model loader extra config is not supported for " f"load format {load_config.load_format}" ) self.remote_instance_transfer_engine_weight_info = None def download_model(self, model_config: ModelConfig) -> None: raise NotImplementedError def load_model( self, *, model_config: ModelConfig, device_config: DeviceConfig, ) -> nn.Module: logger.info("Loading weights from remote instance ...") load_config = self.load_config assert load_config.load_format == LoadFormat.REMOTE_INSTANCE, ( f"Model loader {self.load_config.load_format} is not supported for " f"load format {load_config.load_format}" ) quant_config = _get_quantization_config(model_config, self.load_config) with set_default_torch_dtype(model_config.dtype): with torch.device(device_config.device): model = _initialize_model(model_config, self.load_config, quant_config) if ( load_config.remote_instance_weight_loader_backend == RemoteInstanceWeightLoaderBackend.NCCL ): model_weights = f"instance://{load_config.remote_instance_weight_loader_seed_instance_ip}:{load_config.remote_instance_weight_loader_send_weights_group_ports[load_config.tp_rank]}" with create_remote_connector(model_weights, device_config.device) as client: connector_type = get_connector_type(client) if connector_type == ConnectorType.INSTANCE: self.load_model_from_remote_instance_by_nccl( model, client, model_config, device_config ) else: raise ValueError( f"Unsupported connector type {connector_type} for " f"remote tensor model loading." ) elif ( load_config.remote_instance_weight_loader_backend == RemoteInstanceWeightLoaderBackend.TRANSFER_ENGINE ): if load_config.remote_instance_weight_loader_transfer_engine is None: raise RuntimeError( "Transfer engine is not initialized for remote instance " "model loader with `transfer_engine` backend. " ) logger.info( "TransferEngine registering memory regions (this may take a few seconds)..." ) # register memory region self.remote_instance_transfer_engine_weight_info = register_memory_region( model, load_config.remote_instance_weight_loader_transfer_engine ) logger.info( "TransferEngine memory regions have been successfully registered." ) # transfer weights success = self.load_model_from_remote_instance_by_transfer_engine( model, load_config.remote_instance_weight_loader_transfer_engine, f"http://{load_config.remote_instance_weight_loader_seed_instance_ip}:{load_config.remote_instance_weight_loader_seed_instance_service_port}", load_config.tp_rank, ) if not success: raise RuntimeError( "Failed to load weights from remote instance via transfer engine." ) elif ( load_config.remote_instance_weight_loader_backend == RemoteInstanceWeightLoaderBackend.MODELEXPRESS ): try: from modelexpress.engines.sglang.loader import MxModelLoader except ImportError as exc: raise ImportError( "ModelExpress support requires the 'modelexpress' " "package. Install it in the SGLang image." ) from exc model = MxModelLoader(load_config).load_model( model=model, model_config=model_config, device_config=device_config, ) else: raise ValueError("Invalid remote instance weight loader backend.") return model.eval() def load_model_from_remote_instance_by_nccl( self, model, client, model_config: ModelConfig, device_config: DeviceConfig ) -> nn.Module: load_config = self.load_config instance_ip = socket.gethostbyname(socket.gethostname()) start_build_group_tic = time.time() client.build_group( gpu_id=device_config.gpu_id, tp_rank=load_config.tp_rank, instance_ip=instance_ip, ) current_platform.synchronize() end_build_group_tic = time.time() logger.debug( f"finish building group for remote instance, time used: {(end_build_group_tic - start_build_group_tic):.4f}s" ) if load_config.tp_rank == 0: t = threading.Thread( target=trigger_transferring_weights_request, args=( load_config.remote_instance_weight_loader_seed_instance_ip, load_config.remote_instance_weight_loader_seed_instance_service_port, load_config.remote_instance_weight_loader_send_weights_group_ports, instance_ip, ), ) t.start() start_get_weights_tic = time.time() with set_default_torch_dtype(model_config.dtype): for _, tensor in model.named_parameters(): torch.distributed.broadcast( tensor.data, src=0, group=client._model_update_group, ) current_platform.synchronize() _post_load_weights(model) end_get_weights_tic = time.time() logger.debug( f"finish getting all weights from remote instance, time used: {(end_get_weights_tic - start_get_weights_tic):.4f}s" ) # destroy the process group after loading weights torch.distributed.distributed_c10d.destroy_process_group( client._model_update_group ) current_platform.empty_cache() def load_model_from_remote_instance_by_transfer_engine( self, model, transfer_engine, seed_url, tp_rank ) -> bool: # get remote weights metadata from source instance seed_transfer_engine_session_id, seed_transfer_engine_weight_info = ( get_remote_instance_transfer_engine_info_per_rank(seed_url, tp_rank) ) if ( seed_transfer_engine_session_id is None or seed_transfer_engine_weight_info is None ): logger.error("Cannot get transfer engine session or weight info.") return False # prepare local/remote RDMA keys seed_ptr_list = [] client_ptr_list = [] client_len_list = [] for name, tensor in model.named_parameters(): weight_info = seed_transfer_engine_weight_info.get(name, None) if weight_info is None: logger.error(f"Cannot find weight info for {name}.") return False seed_ptr, seed_numel, seed_element_size = weight_info if ( seed_numel != tensor.numel() or seed_element_size != tensor.element_size() ): logger.error( f"Weight info does not match for {name}, " f"expected ({seed_numel}, {seed_element_size}), " f"got ({tensor.numel()}, {tensor.element_size()})" ) return False client_ptr = tensor.data_ptr() client_len = tensor.numel() * tensor.element_size() seed_ptr_list.append(seed_ptr) client_ptr_list.append(client_ptr) client_len_list.append(client_len) # load weights from source instance through TransferEngine ret = transfer_engine.batch_transfer_sync_read( seed_transfer_engine_session_id, client_ptr_list, seed_ptr_list, client_len_list, ) if ret < 0: logger.error(f"batch transfer failed, error: {ret}") return False _post_load_weights(model) return True class RemoteModelLoader(BaseModelLoader): """Model loader that can load Tensors from remote database.""" def __init__(self, load_config: LoadConfig): super().__init__(load_config) # TODO @DellCurry: move to s3 connector only set_runai_streamer_env(load_config) def _get_weights_iterator_kv( self, client, ) -> Generator[Tuple[str, torch.Tensor], None, None]: """Get an iterator for the model weights from remote storage.""" assert get_connector_type(client) == ConnectorType.KV rank = get_parallel().tp_rank return client.weight_iterator(rank) def _get_weights_iterator_fs( self, client, ) -> Generator[Tuple[str, torch.Tensor], None, None]: """Get an iterator for the model weights from remote storage.""" assert get_connector_type(client) == ConnectorType.FS return client.weight_iterator() def download_model(self, model_config: ModelConfig) -> None: pass @staticmethod def save_model( model: torch.nn.Module, model_path: str, url: str, ) -> None: with create_remote_connector(url) as client: assert get_connector_type(client) == ConnectorType.KV model_name = parse_model_name(url) rank = get_parallel().tp_rank state_dict = ShardedStateLoader._filter_subtensors(model.state_dict()) for key, tensor in state_dict.items(): r_key = f"{model_name}/keys/rank_{rank}/{key}" client.set(r_key, tensor) for root, _, files in os.walk(model_path): for file_name in files: # ignore hidden files if file_name.startswith("."): continue if os.path.splitext(file_name)[1] in (".json", ".py"): file_path = os.path.join(root, file_name) with open(file_path, encoding="utf-8") as file: file_content = file.read() f_key = f"{model_name}/files/{file_name}" client.setstr(f_key, file_content) def _load_model_from_remote_kv( self, model: nn.Module, model_config: ModelConfig, client ): for _, module in model.named_modules(): quant_method = getattr(module, "quant_method", None) if quant_method is not None: quant_method.process_weights_after_loading(module) weights_iterator = self._get_weights_iterator_kv(client) state_dict = ShardedStateLoader._filter_subtensors(model.state_dict()) for key, tensor in weights_iterator: # If loading with LoRA enabled, additional padding may # be added to certain parameters. We only load into a # narrowed view of the parameter data. param_data = state_dict[key].data param_shape = state_dict[key].shape for dim, size in enumerate(tensor.shape): if size < param_shape[dim]: param_data = param_data.narrow(dim, 0, size) if tensor.shape != param_shape: logger.warning( "loading tensor of shape %s into " "parameter '%s' of shape %s", tensor.shape, key, param_shape, ) param_data.copy_(tensor) state_dict.pop(key) if state_dict: raise ValueError(f"Missing keys {tuple(state_dict)} in loaded state!") _post_load_weights(model) def _load_model_from_remote_fs( self, model, client, model_config: ModelConfig, device_config: DeviceConfig ) -> nn.Module: target_device = torch.device(device_config.device) with set_default_torch_dtype(model_config.dtype): model.load_weights(self._get_weights_iterator_fs(client)) for _, module in model.named_modules(): quant_method = getattr(module, "quant_method", None) if quant_method is not None: # When quant methods need to process weights after loading # (for repacking, quantizing, etc), they expect parameters # to be on the global target device. This scope is for the # case where cpu offloading is used, where we will move the # parameters onto device for processing and back off after. with device_loading_context(module, target_device): quant_method.process_weights_after_loading(module) def load_model( self, *, model_config: ModelConfig, device_config: DeviceConfig, ) -> nn.Module: logger.info("Loading weights from remote storage ...") start = time.perf_counter() load_config = self.load_config assert load_config.load_format == LoadFormat.REMOTE, ( f"Model loader {self.load_config.load_format} is not supported for " f"load format {load_config.load_format}" ) model_weights = model_config.model_path if hasattr(model_config, "model_weights"): model_weights = model_config.model_weights quant_config = _get_quantization_config(model_config, self.load_config) with set_default_torch_dtype(model_config.dtype): with torch.device(device_config.device): model = _initialize_model(model_config, self.load_config, quant_config) with create_remote_connector( model_weights, device=device_config.device ) as client: connector_type = get_connector_type(client) if connector_type == ConnectorType.KV: self._load_model_from_remote_kv(model, model_config, client) elif connector_type == ConnectorType.FS: self._load_model_from_remote_fs( model, client, model_config, device_config ) end = time.perf_counter() logger.info("Loaded weights from remote storage in %.2f seconds.", end - start) return model.eval() def load_model_with_cpu_quantization( self, *, model_config: ModelConfig, device_config: DeviceConfig, ) -> nn.Module: target_device = torch.device(device_config.device) quant_config = _get_quantization_config(model_config, self.load_config) with set_default_torch_dtype(model_config.dtype): model = _initialize_model( model_config, self.load_config, quant_config, ) if not isinstance(self, DummyModelLoader): model.load_weights(self._get_all_weights(model_config, model)) for _, module in model.named_modules(): quant_method = getattr(module, "quant_method", None) if quant_method is not None: # When quant methods need to process weights after loading # (for repacking, quantizing, etc), they expect parameters # to be on the global target device. This scope is for the # case where cpu offloading is used, where we will move the # parameters onto device for processing and back off after. with device_loading_context(module, target_device): quant_method.process_weights_after_loading(module) model.to(target_device) return model.eval() class IncModelLoader(DefaultModelLoader): """ Model loader that applies Intel AutoRound quantization """ def __init__(self, load_config: LoadConfig): super().__init__(load_config) def load_model( self, *, model_config: ModelConfig, device_config: DeviceConfig, ) -> nn.Module: logger.info("IncModelLoader: Loading model...") # Check if model is already quantized if model_config._is_already_quantized(): logger.info("Model is already quantized, loading directly...") # Use default loading for pre-quantized models return super().load_model( model_config=model_config, device_config=device_config ) quant_model = self._autoround_quantization_workflow(model_config, device_config) target_device = torch.device(device_config.device) # Return autoround model for offline quantization mode if self.load_config.inc_save_path is not None: quant_model.to(target_device) return quant_model.eval() model_config.hf_config = quant_model.config quant_config = _get_quantization_config(model_config, self.load_config) with set_default_torch_dtype(model_config.dtype): with target_device: model = _initialize_model( model_config, self.load_config, quant_config, ) self.load_weights_and_postprocess( model, iter(quant_model.state_dict().items()), target_device ) return model.eval() def _parse_quantization(self, quantization: str): """Map quantization to AutoRound's scheme and format.""" AR_QUANT_CFG_CHOICES = { "auto-round-int8": ("INT8", "llm_compressor"), } quant_cfg = AR_QUANT_CFG_CHOICES.get(quantization) if not quant_cfg: raise ValueError( f"Invalid quantization choice: '{quantization}'. " f"Available choices: {list(AR_QUANT_CFG_CHOICES.keys())}" ) return quant_cfg def _autoround_quantization_workflow( self, model_config: ModelConfig, device_config: DeviceConfig ) -> nn.Module: """Auto-round quantization workflow: quantize, save checkpoint, then return model.""" try: from auto_round import AutoRound except ImportError: logger.error( "auto-round library not found. " "Please install it using `pip install auto-round` to use AutoRound quantization." ) raise scheme, format = self._parse_quantization(model_config.quantization) try: autoround = AutoRound( model_config.model_path, scheme=scheme, iters=self.load_config.inc_tuning_iters, disable_opt_rtn=self.load_config.inc_disable_opt_rtn, low_cpu_mem_usage=False, ) if self.load_config.inc_save_path is not None: logger.info("Offline quantization mode: Will quantize and save") model, _ = autoround.quantize_and_save( output_dir=self.load_config.inc_save_path, format=format ) return model else: logger.info("Online quantization mode: Will quantize and skip saving") # Use a temporary directory and discard it so nothing is persisted in online mode. with tempfile.TemporaryDirectory() as tmp_save_dir: model, _ = autoround.quantize_and_save( output_dir=tmp_save_dir, format=format ) return model except Exception as e: raise ValueError(f"AutoRound quantization failed: {e}") class ModelOptModelLoader(DefaultModelLoader): """ Model loader that applies NVIDIA Model Optimizer quantization """ def __init__(self, load_config: LoadConfig): super().__init__(load_config) # Any ModelOpt specific initialization if needed def _setup_modelopt_quantization( self, model, tokenizer, quant_cfg, quantized_ckpt_restore_path: str | None = None, quantized_ckpt_save_path: str | None = None, export_path: str | None = None, ) -> None: """ Set up ModelOpt quantization for the given model. Args: model: The model to quantize tokenizer: The tokenizer associated with the model quant_cfg: The quantization configuration quantized_ckpt_restore_path: Path to restore quantized checkpoint from quantized_ckpt_save_path: Path to save quantized checkpoint to export_path: Path to export the quantized model in HuggingFace format Raises: ImportError: If ModelOpt is not available Exception: If quantization setup fails """ try: import modelopt.torch.opt as mto import modelopt.torch.quantization as mtq from modelopt.torch.quantization.utils import is_quantized except ImportError as e: raise ImportError( "ModelOpt is not available. Please install modelopt." ) from e if is_quantized(model): rank0_log("Model is already quantized, skipping quantization setup.") return # Restore from checkpoint if provided if quantized_ckpt_restore_path: try: mto.restore(model, quantized_ckpt_restore_path) rank0_log( f"Restored quantized model from {quantized_ckpt_restore_path}" ) # Export model if path provided (even when restoring from checkpoint) self._maybe_export_modelopt(model, export_path) return except Exception as e: logger.warning( f"Failed to restore from {quantized_ckpt_restore_path}: {e}" ) rank0_log("Proceeding with calibration-based quantization...") # Set up calibration-based quantization try: # Left padding tends to work better for batched generation with decoder-only LMs with suppress(Exception): tokenizer.padding_side = "left" from modelopt.torch.utils.dataset_utils import ( create_forward_loop, get_dataset_dataloader, ) # Create calibration dataloader calib_dataloader = get_dataset_dataloader( dataset_name="cnn_dailymail", # TODO: Consider making this configurable tokenizer=tokenizer, batch_size=36, # TODO: Consider making this configurable num_samples=512, # TODO: Consider making this configurable device=model.device, include_labels=False, ) calibrate_loop = create_forward_loop(dataloader=calib_dataloader) # Apply quantization mtq.quantize(model, quant_cfg, forward_loop=calibrate_loop) if not model_parallel_is_initialized() or get_parallel().tp_rank == 0: mtq.print_quant_summary(model) # Save checkpoint if path provided if quantized_ckpt_save_path: try: mto.save(model, quantized_ckpt_save_path) rank0_log(f"Quantized model saved to {quantized_ckpt_save_path}") except Exception as e: logger.warning( f"Failed to save quantized checkpoint to {quantized_ckpt_save_path}: {e}" ) # Export model if path provided self._maybe_export_modelopt(model, export_path) except Exception as e: raise Exception(f"Failed to set up ModelOpt quantization: {e}") from e def _maybe_export_modelopt(self, model, export_path: str | None) -> None: """Export model to HuggingFace format if export_path is provided.""" if export_path: try: # Get the original model path from the model config original_model_path = getattr(self, "_original_model_path", None) self._export_modelopt_checkpoint( model, export_path, original_model_path ) rank0_log( f"Quantized model exported to HuggingFace format at {export_path}" ) except Exception as e: rank0_log( f"Warning: Failed to export quantized model to {export_path}: {e}" ) def _export_modelopt_checkpoint( self, model, export_path: str, model_path: str = None, trust_remote_code: bool = True, ) -> None: """ Export the quantized model to HuggingFace format using ModelOpt export API. Args: model: The quantized model to export export_path: Directory path to export the model to model_path: Path to the original model (for tokenizer export) trust_remote_code: Whether to trust remote code for tokenizer loading Raises: ImportError: If ModelOpt export functionality is not available Exception: If export fails """ try: from modelopt.torch.export import export_hf_checkpoint from transformers import AutoTokenizer except ImportError as e: raise ImportError( "ModelOpt export functionality is not available. " "Please ensure you have the latest version of modelopt installed." ) from e # Create export directory if it doesn't exist os.makedirs(export_path, exist_ok=True) # Export the quantized model export_hf_checkpoint(model, export_dir=export_path) # Export the tokenizer if model_path is provided if model_path: try: tokenizer = AutoTokenizer.from_pretrained( model_path, trust_remote_code=trust_remote_code ) tokenizer.save_pretrained(export_path) rank0_log(f"Tokenizer exported to {export_path}") except Exception as e: rank0_log(f"Warning: Failed to export tokenizer: {e}") def load_model( self, *, model_config: ModelConfig, device_config: DeviceConfig, ) -> nn.Module: logger.info("ModelOptModelLoader: Loading base model...") # Store the original model path for tokenizer export self._original_model_path = model_config.model_path # Check if model is already quantized if model_config._is_already_quantized(): logger.info("Model is already quantized, loading directly...") # Use default loading for pre-quantized models return super().load_model( model_config=model_config, device_config=device_config ) # TODO: Quantize-and-serve mode has been disabled at the ModelConfig level # All quantization now uses the standard workflow (quantize + export/save) logger.info("Standard quantization mode: Will quantize and export/save") return self._standard_quantization_workflow(model_config, device_config) def _standard_quantization_workflow( self, model_config: ModelConfig, device_config: DeviceConfig ) -> nn.Module: """Standard quantization workflow: quantize, save checkpoint, export, then return model.""" # Use shared method from parent class to load base model for quantization model = self._load_modelopt_base_model(model_config) # Import ModelOpt modules try: import modelopt.torch.quantization as mtq except ImportError: logger.error( "NVIDIA Model Optimizer (modelopt) library not found. " "Please install it to use ModelOpt quantization." ) raise # Handle both old modelopt_quant and new unified quantization flags if hasattr(model_config, "modelopt_quant") and model_config.modelopt_quant: # Legacy modelopt_quant flag quant_choice_str = model_config.modelopt_quant else: # Unified quantization flag - extract the type (fp8/fp4) quant_choice_str = model_config._get_modelopt_quant_type() quant_cfg_name = QUANT_CFG_CHOICES.get(quant_choice_str) if not quant_cfg_name: raise ValueError( f"Invalid quantization choice: '{quant_choice_str}'. " f"Available choices: {list(QUANT_CFG_CHOICES.keys())}" ) try: # getattr will fetch the config object, e.g., mtq.FP8_DEFAULT_CFG quant_cfg = getattr(mtq, quant_cfg_name) except AttributeError: raise AttributeError( f"ModelOpt quantization config '{quant_cfg_name}' not found. " "Please verify the ModelOpt library installation." ) logger.info( f"Quantizing model with ModelOpt using config: mtq.{quant_cfg_name}" ) # Get ModelOpt configuration from LoadConfig modelopt_config = self.load_config.modelopt_config quantized_ckpt_restore_path = ( modelopt_config.checkpoint_restore_path if modelopt_config else None ) quantized_ckpt_save_path = ( modelopt_config.checkpoint_save_path if modelopt_config else None ) export_path = modelopt_config.export_path if modelopt_config else None tokenizer = AutoTokenizer.from_pretrained( model_config.model_path, use_fast=True ) try: self._setup_modelopt_quantization( model, tokenizer, quant_cfg, quantized_ckpt_restore_path=quantized_ckpt_restore_path, quantized_ckpt_save_path=quantized_ckpt_save_path, export_path=export_path, ) except Exception as e: logger.warning(f"ModelOpt quantization failed: {e}") rank0_log("Proceeding without quantization...") return model.eval() class RunaiModelStreamerLoader(BaseModelLoader): """ Model loader that uses Runai Model Streamer to load a model. Supports fast model loading from SSDs, shared filesystems and object storage (S3, GCS, Azure blob) with weight streaming. Configuration (via load_config.model_loader_extra_config): - distributed (bool): Enable distributed streaming - True by default for url paths (object storage) - concurrency (int): Number of concurrent downloads - memory_limit (int): Memory limit for streaming buffer Note: Metadata files must be pre-downloaded via ObjectStorageModel.download_and_get_path() before instantiation. """ @dataclasses.dataclass class Source: """A source for weights.""" model_or_path: str """The model ID or path.""" revision: Optional[str] """The optional model revision.""" prefix: str = "" """A prefix to prepend to all weights.""" fall_back_to_pt: bool = True """Whether .pt weights can be used.""" model_config: Optional[ModelConfig] = None """The model configuration (for checking architecture, etc).""" @classmethod def init_new(cls, model_config: ModelConfig, model): model_weights = model_config.model_path if hasattr(model_config, "model_weights"): model_weights = model_config.model_weights return cls( model_weights, model_config.revision, prefix="", fall_back_to_pt=getattr(model, "fall_back_to_pt_during_load", True), model_config=model_config, ) def __init__(self, load_config: LoadConfig): super().__init__(load_config) extra_config = load_config.model_loader_extra_config allowed_keys = {"distributed", "concurrency", "memory_limit"} unexpected_keys = set(extra_config.keys()) - allowed_keys if unexpected_keys: raise ValueError( f"Unexpected extra config keys for load format " f"{load_config.load_format}: " f"{unexpected_keys}" ) set_runai_streamer_env(load_config) self._is_distributed = None if load_config.model_loader_extra_config: extra_config = load_config.model_loader_extra_config if "distributed" in extra_config and isinstance( extra_config.get("distributed"), bool ): self._is_distributed = extra_config.get("distributed") def _prepare_weights( self, model_name_or_path: str, revision: Optional[str] ) -> Tuple[str, List[str]]: """Prepare weights for the model. If the model is not local, it will be downloaded.""" from sglang.srt.utils.runai_utils import is_runai_obj_uri, list_safetensors is_object_storage_path = is_runai_obj_uri(model_name_or_path) if self._is_distributed is None: self._is_distributed = is_object_storage_path is_local = os.path.isdir(model_name_or_path) safetensors_pattern = "*.safetensors" index_file = SAFE_WEIGHTS_INDEX_NAME hf_folder = ( model_name_or_path if (is_local or is_object_storage_path) else download_weights_from_hf( model_name_or_path, self.load_config.download_dir, [safetensors_pattern], revision, ignore_patterns=self.load_config.ignore_patterns, ) ) server_args = get_server_args() if server_args and server_args.model_checksum is not None: from sglang.srt.utils.model_file_verifier import verify checksums_source = server_args.model_checksum or model_name_or_path verify(model_path=hf_folder, checksums_source=checksums_source) hf_weights_files = list_safetensors(path=hf_folder) # For models like Mistral-7B-Instruct-v0.3 # there are both sharded safetensors files and a consolidated # safetensors file. Using both breaks. # Here, we download the `model.safetensors.index.json` and filter # any files not found in the index. if not is_local and not is_object_storage_path: download_safetensors_index_file_from_hf( model_name_or_path, index_file, self.load_config.download_dir, revision, ) hf_weights_files = filter_duplicate_safetensors_files( hf_weights_files, hf_folder, index_file ) if len(hf_weights_files) == 0: raise RuntimeError( f"Cannot find any model weights with `{model_name_or_path}`" ) return hf_folder, hf_weights_files def _get_weights_iterator( self, source: Source ) -> Generator[Tuple[str, torch.Tensor], None, None]: """Get an iterator for the model weights based on the load format.""" from sglang.srt.model_loader.weight_utils import ( runai_safetensors_weights_iterator, ) hf_folder, hf_weights_files = self._prepare_weights( source.model_or_path, source.revision ) if source.model_config is not None: hf_weights_files = maybe_add_mtp_safetensors( hf_weights_files, hf_folder, "model.safetensors.index.json", source.model_config.hf_config, ) weights_iterator = runai_safetensors_weights_iterator( hf_weights_files, self._is_distributed, self.target_device_str ) if self.load_config.draft_model_idx is not None: import re def filter_weights(original_weights_iterator): pattern = r"model.mtp.layers.(\d+)." for name, tensor in original_weights_iterator: group = re.match(pattern, name) if group is not None: idx = int(group.group(1)) if idx != self.load_config.draft_model_idx: continue new_name = name.replace(group.group(), "model.mtp.layers.0.") else: new_name = name yield (new_name, tensor) weights_iterator = filter_weights(weights_iterator) def apply_prefix(original_weights_iterator): yield from ( (source.prefix + name, tensor) for (name, tensor) in original_weights_iterator ) return apply_prefix(weights_iterator) def _get_all_weights( self, model_config: ModelConfig, model: nn.Module, ) -> Generator[Tuple[str, torch.Tensor], None, None]: primary_weights = RunaiModelStreamerLoader.Source.init_new(model_config, model) yield from self._get_weights_iterator(primary_weights) secondary_weights = cast( Iterable[RunaiModelStreamerLoader.Source], getattr(model, "secondary_weights", ()), ) for source in secondary_weights: yield from self._get_weights_iterator(source) def download_model(self, model_config: ModelConfig) -> None: self._prepare_weights(model_config.model_path, model_config.revision) def load_model( self, *, model_config: ModelConfig, device_config: DeviceConfig, ) -> nn.Module: if hasattr(model_config, "modelopt_quant") and model_config.modelopt_quant: # Load base model using shared method raise NotImplementedError( "Runai Model Streamer Loader does not support ModelOpt quantization yet" ) assert device_config.device_type in ("cuda", "cpu"), ( f"Runai Model Streamer only supports CUDA and CPU, " f"got {device_config.device_type}" ) if device_config.device_type == "cuda": self.target_device_str = ( device_config.device_type + ":" + str(device_config.gpu_id) ) else: self.target_device_str = "cpu" target_device = torch.device(device_config.device) quant_config = _get_quantization_config(model_config, self.load_config) with set_default_torch_dtype(model_config.dtype): with target_device: model = _initialize_model( model_config, self.load_config, quant_config, ) DefaultModelLoader.load_weights_and_postprocess( model, self._get_all_weights(model_config, model), target_device ) return model.eval() def get_model_loader( load_config: LoadConfig, model_config: Optional[ModelConfig] = None ) -> BaseModelLoader: """Get a model loader based on the load format.""" if load_config.load_format == LoadFormat.DUMMY: return DummyModelLoader(load_config) if model_config and model_config.quantization in ["auto-round-int8"]: logger.info("Using IncModelLoader due to AutoRound quantization config.") return IncModelLoader(load_config) # ModelOptModelLoader's local-copy quantize-and-export workflow doesn't apply # to non-local loaders. These loaders own their weight transport path and still # initialize the model with ModelOpt quantization config where applicable. model_optloader_allowed = model_config and load_config.load_format not in ( LoadFormat.RUNAI_STREAMER, LoadFormat.REMOTE_INSTANCE, ) if model_optloader_allowed and ( (hasattr(model_config, "modelopt_quant") and model_config.modelopt_quant) or model_config.quantization in ["modelopt_fp8", "modelopt_fp4", "modelopt_mixed", "modelopt"] ): logger.info("Using ModelOptModelLoader due to ModelOpt quantization config.") return ModelOptModelLoader(load_config) # Use ModelOptModelLoader for unified quantization flags if ( model_optloader_allowed and hasattr(model_config, "quantization") and model_config.quantization in ["modelopt_fp8", "modelopt_fp4", "modelopt_mixed"] ): if model_config._is_already_quantized(): logger.info( f"Using ModelOptModelLoader for pre-quantized model: {model_config.quantization}" ) else: logger.info( f"Using ModelOptModelLoader for quantization: {model_config.quantization}" ) return ModelOptModelLoader(load_config) if isinstance(load_config.load_format, type): return load_config.load_format(load_config) if load_config.load_format == LoadFormat.SHARDED_STATE: return ShardedStateLoader(load_config) if load_config.load_format == LoadFormat.BITSANDBYTES: return BitsAndBytesModelLoader(load_config) if load_config.load_format == LoadFormat.GGUF: return GGUFModelLoader(load_config) if load_config.load_format == LoadFormat.LAYERED: return LayeredModelLoader(load_config) # Check for FLASH_RL format early # FP8 approach: BF16/FP16 model with native FP8 quantization if load_config.load_format == LoadFormat.FLASH_RL: logger.info( "Using QuantizedRLModelLoader for RL training with native FP8 quantization." ) logger.info( "FP8 approach: Model loads with native SGLang FP8 quantization. " "Same model path for both training and inference." ) # Set quantization to FP8 for native SGLang support if model_config and not model_config.quantization: logger.info( "QuantizedRL: Setting quantization to fp8 (native SGLang support). " "Model will be loaded with FP8 infrastructure" ) model_config.quantization = "fp8" return QuantizedRLModelLoader(load_config) if load_config.load_format == LoadFormat.REMOTE: return RemoteModelLoader(load_config) if load_config.load_format == LoadFormat.REMOTE_INSTANCE: return RemoteInstanceModelLoader(load_config) if load_config.load_format == LoadFormat.PRIVATE: import importlib try: module = importlib.import_module("sglang.private.private_model_loader") return module.PrivateModelLoader(load_config) except ImportError: raise ValueError("Failed to import sglang.private.private_model_loader") if load_config.load_format == LoadFormat.RUNAI_STREAMER: return RunaiModelStreamerLoader(load_config) return DefaultModelLoader(load_config)