| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from copy import deepcopy |
| 5 | from typing import Any |
| 6 | |
| 7 | import torch |
| 8 | from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE |
| 9 | from transformers import PreTrainedConfig |
| 10 | |
| 11 | import vllm.model_executor.layers.fused_moe # noqa |
| 12 | from vllm.logger import init_logger |
| 13 | from vllm.model_executor.kernels.linear import ( |
| 14 | MPLinearLayerConfig, |
| 15 | choose_mp_linear_kernel, |
| 16 | ) |
| 17 | from vllm.model_executor.layers.fused_moe import ( |
| 18 | FusedMoEConfig, |
| 19 | FusedMoEMethodBase, |
| 20 | FusedMoEQuantConfig, |
| 21 | FusedMoeWeightScaleSupported, |
| 22 | RoutedExperts, |
| 23 | SharedExperts, |
| 24 | UnquantizedFusedMoEMethod, |
| 25 | ) |
| 26 | from vllm.model_executor.layers.fused_moe.oracle.int_wna16 import ( |
| 27 | WNA16MoEBackend, |
| 28 | convert_to_wna16_moe_kernel_format, |
| 29 | make_wna16_moe_kernel, |
| 30 | select_wna16_moe_backend, |
| 31 | ) |
| 32 | from vllm.model_executor.layers.linear import LinearMethodBase, set_weight_attrs |
| 33 | from vllm.model_executor.layers.quantization import QuantizationMethods |
| 34 | from vllm.model_executor.layers.quantization.base_config import ( |
| 35 | QuantizationConfig, |
| 36 | QuantizeMethodBase, |
| 37 | ) |
| 38 | from vllm.model_executor.layers.quantization.utils import replace_parameter |
| 39 | from vllm.model_executor.layers.quantization.utils.gptq_utils import ( |
| 40 | get_dynamic_override, |
| 41 | get_linear_quant_method, |
| 42 | normalize_and_validate_gptq_desc_act, |
| 43 | override_config, |
| 44 | ) |
| 45 | from vllm.model_executor.layers.quantization.utils.marlin_utils import ( |
| 46 | check_moe_marlin_supports_layer, |
| 47 | get_marlin_input_dtype, |
| 48 | marlin_repeat_scales_on_all_ranks, |
| 49 | verify_marlin_supported, |
| 50 | ) |
| 51 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 52 | QuantKey, |
| 53 | kInt4StaticGroupScale, |
| 54 | kInt8StaticGroupScale, |
| 55 | ) |
| 56 | from vllm.model_executor.parameter import ( |
| 57 | ChannelQuantScaleParameter, |
| 58 | GroupQuantScaleParameter, |
| 59 | PackedColumnParameter, |
| 60 | PackedvLLMParameter, |
| 61 | ) |
| 62 | from vllm.scalar_type import scalar_types |
| 63 | from vllm.transformers_utils.config import get_safetensors_params_metadata |
| 64 | from vllm.utils.collection_utils import is_list_of |
| 65 | |
| 66 | logger = init_logger(__name__) |
| 67 | |
| 68 | |
| 69 | def get_moe_quant_method( |
| 70 | config: "AutoGPTQConfig", |
| 71 | layer: RoutedExperts, |
| 72 | prefix: str, |
| 73 | moe_method_cls: type, |
| 74 | ): |
| 75 | cloned_config = deepcopy(config) |
| 76 | |
| 77 | assert isinstance(layer, RoutedExperts) |
| 78 | # False = skip module, None = no override, else = Positive match |
| 79 | if ( |
| 80 | get_dynamic_override( # noqa: E712 |
| 81 | cloned_config, # noqa: E712 |
| 82 | layer_name=prefix, |
| 83 | ) |
| 84 | == False |
| 85 | ): # noqa: E712 |
| 86 | return UnquantizedFusedMoEMethod(layer.moe_config) |
| 87 | |
| 88 | if prefix: |
| 89 | # Dynamic per module/layer rules may override base config |
| 90 | override_config(cloned_config, prefix=prefix) |
| 91 | |
| 92 | return moe_method_cls(cloned_config, layer.moe_config) |
| 93 | |
| 94 | |
| 95 | class AutoGPTQConfig(QuantizationConfig): |
| 96 | """Config class for AutoGPTQ quantization using Marlin kernels.""" |
| 97 | |
| 98 | # (num_bits, is_sym) -> quant_type |
| 99 | TYPE_MAP = { |
| 100 | (4, True): scalar_types.uint4b8, |
| 101 | (8, True): scalar_types.uint8b128, |
| 102 | } |
| 103 | |
| 104 | def __init__( |
| 105 | self, |
| 106 | weight_bits: int, |
| 107 | group_size: int, |
| 108 | desc_act: bool, |
| 109 | is_sym: bool, |
| 110 | lm_head_quantized: bool, |
| 111 | dynamic: dict[str, dict[str, int | bool]], |
| 112 | full_config: dict[str, Any], |
| 113 | modules_in_block_to_quantize: list[str] | None = None, |
| 114 | ) -> None: |
| 115 | super().__init__() |
| 116 | desc_act = normalize_and_validate_gptq_desc_act(desc_act, group_size, dynamic) |
| 117 | |
| 118 | # GPTQModel use `dynamic` config property to allow per module |
| 119 | # quantization config so each module can be individually optimized. |
| 120 | # Format is dict[str, dict] where key is a regex string that can |
| 121 | # perform both positive ("+:" prefixed) or negative ("-:" prefixed) |
| 122 | # matching of a module. |
| 123 | # Default to positive match, override base quant config mode, if no |
| 124 | # prefix is used. Value is in dict format of field key and override |
| 125 | # value. |
| 126 | # Negative matching will skip quantization init for this module |
| 127 | # entirely: |
| 128 | # non-quantized inference. More details and quantization examples can be |
| 129 | # found at: https://github.com/ModelCloud/GPTQModel |
| 130 | # Example: |
| 131 | # # last 1/2 of the layers 10-21 has 8bit vs 4bit for 0-9 |
| 132 | # # last 1/4 of the layers 16-21 has 8bit and group_size 64 |
| 133 | # dynamic = { |
| 134 | # #`.*\.` matches the layers_node prefix |
| 135 | # # positive match layer 10-15 |
| 136 | # r"+:.*\.(?:1[0-5])\..*": {"bits": 8,}, |
| 137 | # # positive match layer 16-21 |
| 138 | # r"+:.*\.(?:1[6-9]|20|21)\..*": {"bits": 8, "group_size": 64,}, |
| 139 | # r"-:.*\.moe\..*": {}, # negative match (skip) all `moe` layers |
| 140 | # } |
| 141 | self.dynamic = dynamic |
| 142 | |
| 143 | self.weight_bits = weight_bits |
| 144 | self.is_sym = is_sym |
| 145 | |
| 146 | self.pack_factor = 32 // weight_bits # packed into int32 |
| 147 | self.group_size = group_size |
| 148 | self.desc_act: bool = desc_act |
| 149 | self.lm_head_quantized = lm_head_quantized |
| 150 | self.full_config = full_config |
| 151 | |
| 152 | if (weight_bits, is_sym) not in self.TYPE_MAP: |
| 153 | raise ValueError( |
| 154 | f"Unsupported quantization config: bits={weight_bits}, sym={is_sym}" |
| 155 | ) |
| 156 | |
| 157 | self.quant_type = self.TYPE_MAP[(weight_bits, is_sym)] |
| 158 | |
| 159 | self.modules_in_block_to_quantize = modules_in_block_to_quantize or [] |
| 160 | # used to identify GPTQ model quantized by autoround |
| 161 | self.autoround_version = full_config.get("autoround_version", "") |
| 162 | |
| 163 | def __repr__(self) -> str: |
| 164 | return ( |
| 165 | f"AutoGPTQConfig(quant_type={self.quant_type}, " |
| 166 | f"group_size={self.group_size}, " |
| 167 | f"desc_act={self.desc_act}, " |
| 168 | f"lm_head_quantized={self.lm_head_quantized}, " |
| 169 | f"dynamic={self.dynamic}, " |
| 170 | f"modules_in_block_to_quantize={self.modules_in_block_to_quantize})" |
| 171 | ) |
| 172 | |
| 173 | @classmethod |
| 174 | def get_name(cls) -> QuantizationMethods: |
| 175 | return "auto_gptq" |
| 176 | |
| 177 | @classmethod |
| 178 | def get_supported_act_dtypes(cls) -> list[torch.dtype]: |
| 179 | return [torch.half, torch.bfloat16] |
| 180 | |
| 181 | @classmethod |
| 182 | def get_min_capability(cls) -> int: |
| 183 | return 60 |
| 184 | |
| 185 | @classmethod |
| 186 | def get_config_filenames(cls) -> list[str]: |
| 187 | return ["quantize_config.json"] |
| 188 | |
| 189 | @classmethod |
| 190 | def from_config(cls, config: dict[str, Any]) -> "AutoGPTQConfig": |
| 191 | dynamic = cls.get_from_keys_or(config, ["dynamic"], default={}) |
| 192 | dynamic = {} if dynamic is None else dynamic |
| 193 | |
| 194 | weight_bits = cls.get_from_keys(config, ["bits"]) |
| 195 | group_size = cls.get_from_keys(config, ["group_size"]) |
| 196 | desc_act = cls.get_from_keys(config, ["desc_act"]) |
| 197 | is_sym = cls.get_from_keys(config, ["sym"]) |
| 198 | lm_head_quantized = cls.get_from_keys_or(config, ["lm_head"], default=False) |
| 199 | modules_in_block_to_quantize = cls.get_from_keys_or( |
| 200 | config, ["modules_in_block_to_quantize"], default=None |
| 201 | ) |
| 202 | return cls( |
| 203 | weight_bits, |
| 204 | group_size, |
| 205 | desc_act, |
| 206 | is_sym, |
| 207 | lm_head_quantized, |
| 208 | dynamic, |
| 209 | config, |
| 210 | modules_in_block_to_quantize, |
| 211 | ) |
| 212 | |
| 213 | @classmethod |
| 214 | def override_quantization_method( |
| 215 | cls, hf_quant_cfg, user_quant, hf_config=None |
| 216 | ) -> QuantizationMethods | None: |
| 217 | """Override to use AutoGPTQ for compatible GPTQ models.""" |
| 218 | quant_method = hf_quant_cfg.get("quant_method", "").lower() |
| 219 | |
| 220 | if quant_method != "gptq": |
| 221 | return None |
| 222 | |
| 223 | is_valid_user_quant = user_quant is None or user_quant in ( |
| 224 | "gptq", |
| 225 | "gptq_marlin", |
| 226 | "auto_gptq", |
| 227 | "marlin", |
| 228 | ) |
| 229 | |
| 230 | if is_valid_user_quant: |
| 231 | return cls.get_name() |
| 232 | |
| 233 | return None |
| 234 | |
| 235 | def get_quant_method( |
| 236 | self, layer: torch.nn.Module, prefix: str |
| 237 | ) -> "QuantizeMethodBase | None": |
| 238 | if isinstance(layer, RoutedExperts): |
| 239 | from vllm.model_executor.layers.quantization.moe_wna16 import MoeWNA16Config |
| 240 | |
| 241 | if not check_moe_marlin_supports_layer( |
| 242 | layer, self.group_size, allow_tile_padding=True |
| 243 | ): |
| 244 | logger.warning_once( |
| 245 | f"Layer '{prefix}' is not supported by GPTQMoeMarlin. " |
| 246 | "Falling back to Moe WNA16 kernels." |
| 247 | ) |
| 248 | return MoeWNA16Config.from_config(self.full_config).get_quant_method( |
| 249 | layer, prefix |
| 250 | ) |
| 251 | moe_quant_method = get_moe_quant_method( |
| 252 | self, layer, prefix, AutoGPTQMoEMethod |
| 253 | ) |
| 254 | if moe_quant_method is None: |
| 255 | return None |
| 256 | moe_quant_method.input_dtype = get_marlin_input_dtype(prefix) |
| 257 | return moe_quant_method |
| 258 | |
| 259 | quant_method = get_linear_quant_method( |
| 260 | self, layer, prefix, AutoGPTQLinearMethod |
| 261 | ) |
| 262 | if quant_method is None: |
| 263 | return None |
| 264 | quant_method.input_dtype = get_marlin_input_dtype(prefix) |
| 265 | return quant_method |
| 266 | |
| 267 | def apply_vllm_mapper(self, hf_to_vllm_mapper): |
| 268 | if self.modules_in_block_to_quantize is not None: |
| 269 | self.modules_in_block_to_quantize = hf_to_vllm_mapper.apply_list( |
| 270 | self.modules_in_block_to_quantize |
| 271 | ) |
| 272 | |
| 273 | def maybe_update_config( |
| 274 | self, |
| 275 | model_name: str, |
| 276 | hf_config: PreTrainedConfig | None = None, |
| 277 | revision: str | None = None, |
| 278 | ): |
| 279 | if self.modules_in_block_to_quantize: |
| 280 | if is_list_of(self.modules_in_block_to_quantize, list): |
| 281 | # original modules_in_block_to_quantize: list[list[str]] |
| 282 | # flatten original modules_in_block_to_quantize |
| 283 | self.modules_in_block_to_quantize = [ |
| 284 | item |
| 285 | for sublist in self.modules_in_block_to_quantize |
| 286 | for item in sublist |
| 287 | ] |
| 288 | return |
| 289 | |
| 290 | unquant_dtypes = [torch.float16, torch.bfloat16, torch.float32] |
| 291 | metadata = get_safetensors_params_metadata(model_name, revision=revision) |
| 292 | quant_layers: set[str] = { |
| 293 | param_name.rsplit(".", 1)[0] |
| 294 | for param_name, info in metadata.items() |
| 295 | if (dtype := info.get("dtype", None)) |
| 296 | and _SAFETENSORS_TO_TORCH_DTYPE[dtype] not in unquant_dtypes |
| 297 | } |
| 298 | self.modules_in_block_to_quantize = list(quant_layers) |
| 299 | |
| 300 | |
| 301 | class AutoGPTQLinearMethod(LinearMethodBase): |
| 302 | """Linear method for AutoGPTQ using Marlin kernels. |
| 303 | |
| 304 | Args: |
| 305 | quant_config: The AutoGPTQ quantization config. |
| 306 | |
| 307 | """ |
| 308 | |
| 309 | _kernel_backends_being_used: set[str] = set() |
| 310 | |
| 311 | def __init__(self, quant_config: AutoGPTQConfig) -> None: |
| 312 | self.quant_config = quant_config |
| 313 | self.input_dtype = None |
| 314 | self.quant_type = self.quant_config.quant_type |
| 315 | |
| 316 | # Verify supported on platform. |
| 317 | verify_marlin_supported( |
| 318 | quant_type=self.quant_config.quant_type, |
| 319 | group_size=self.quant_config.group_size, |
| 320 | ) |
| 321 | |
| 322 | def create_weights( |
| 323 | self, |
| 324 | layer: torch.nn.Module, |
| 325 | input_size_per_partition: int, |
| 326 | output_partition_sizes: list[int], |
| 327 | input_size: int, |
| 328 | output_size: int, |
| 329 | params_dtype: torch.dtype, |
| 330 | **extra_weight_attrs, |
| 331 | ) -> None: |
| 332 | output_size_per_partition = sum(output_partition_sizes) |
| 333 | is_row_parallel = input_size != input_size_per_partition |
| 334 | weight_loader = extra_weight_attrs.get("weight_loader") |
| 335 | input_dtype = self.input_dtype |
| 336 | |
| 337 | mp_linear_kernel_config = MPLinearLayerConfig( |
| 338 | full_weight_shape=(input_size, output_size), |
| 339 | partition_weight_shape=( |
| 340 | input_size_per_partition, |
| 341 | output_size_per_partition, |
| 342 | ), |
| 343 | weight_type=self.quant_config.quant_type, |
| 344 | act_type=params_dtype if input_dtype is None else input_dtype, |
| 345 | group_size=self.quant_config.group_size, |
| 346 | zero_points=False, |
| 347 | ) |
| 348 | |
| 349 | kernel_type = choose_mp_linear_kernel(mp_linear_kernel_config) |
| 350 | |
| 351 | if kernel_type.__name__ not in self._kernel_backends_being_used: |
| 352 | logger.info("Using %s for AutoGPTQLinearMethod", kernel_type.__name__) |
| 353 | self._kernel_backends_being_used.add(kernel_type.__name__) |
| 354 | |
| 355 | # Normalize group_size |
| 356 | if self.quant_config.group_size != -1: |
| 357 | group_size = self.quant_config.group_size |
| 358 | else: |
| 359 | group_size = input_size |
| 360 | |
| 361 | # Determine sharding |
| 362 | if marlin_repeat_scales_on_all_ranks( |
| 363 | self.quant_config.group_size, is_row_parallel |
| 364 | ): |
| 365 | # By setting scale_dim == None, weight_loader will |
| 366 | # repeat the scales on each GPU in TP>1 case. |
| 367 | scales_and_zp_input_dim = None |
| 368 | scales_and_zp_size = input_size // group_size |
| 369 | else: |
| 370 | # By setting scale_dim == 0, weight_loader will |
| 371 | # shard the scales in TP>1 case. |
| 372 | scales_and_zp_input_dim = 0 |
| 373 | scales_and_zp_size = input_size_per_partition // group_size |
| 374 | |
| 375 | # Quantized weights |
| 376 | qweight = PackedvLLMParameter( |
| 377 | data=torch.empty( |
| 378 | input_size_per_partition // self.quant_config.pack_factor, |
| 379 | output_size_per_partition, |
| 380 | dtype=torch.int32, |
| 381 | ), |
| 382 | input_dim=0, |
| 383 | output_dim=1, |
| 384 | packed_dim=0, |
| 385 | packed_factor=self.quant_config.pack_factor, |
| 386 | weight_loader=weight_loader, |
| 387 | ) |
| 388 | |
| 389 | qzeros_args = { |
| 390 | "data": torch.empty( |
| 391 | scales_and_zp_size, |
| 392 | output_size_per_partition // self.quant_config.pack_factor, |
| 393 | dtype=torch.int32, |
| 394 | ), |
| 395 | "weight_loader": weight_loader, |
| 396 | } |
| 397 | weight_scale_args = { |
| 398 | "data": torch.empty( |
| 399 | scales_and_zp_size, |
| 400 | output_size_per_partition, |
| 401 | dtype=params_dtype, |
| 402 | ), |
| 403 | "weight_loader": weight_loader, |
| 404 | } |
| 405 | |
| 406 | if scales_and_zp_input_dim is None: |
| 407 | scales = ChannelQuantScaleParameter(output_dim=1, **weight_scale_args) |
| 408 | qzeros = PackedColumnParameter( |
| 409 | output_dim=1, |
| 410 | packed_dim=1, |
| 411 | packed_factor=self.quant_config.pack_factor, |
| 412 | **qzeros_args, |
| 413 | ) |
| 414 | |
| 415 | else: |
| 416 | scales = GroupQuantScaleParameter( |
| 417 | output_dim=1, input_dim=0, **weight_scale_args |
| 418 | ) |
| 419 | qzeros = PackedvLLMParameter( |
| 420 | input_dim=0, |
| 421 | output_dim=1, |
| 422 | packed_dim=1, |
| 423 | packed_factor=self.quant_config.pack_factor, |
| 424 | **qzeros_args, |
| 425 | ) |
| 426 | |
| 427 | layer.register_parameter("qweight", qweight) |
| 428 | layer.register_parameter("scales", scales) |
| 429 | layer.register_parameter("qzeros", qzeros) |
| 430 | |
| 431 | self.kernel = kernel_type( |
| 432 | mp_linear_kernel_config, |
| 433 | w_q_param_name="qweight", |
| 434 | w_s_param_name="scales", |
| 435 | w_zp_param_name="qzeros", |
| 436 | ) |
| 437 | |
| 438 | def process_weights_after_loading(self, layer: torch.nn.Module) -> None: |
| 439 | self.kernel.process_weights_after_loading(layer) |
| 440 | |
| 441 | def apply( |
| 442 | self, |
| 443 | layer: torch.nn.Module, |
| 444 | x: torch.Tensor, |
| 445 | bias: torch.Tensor | None = None, |
| 446 | ) -> torch.Tensor: |
| 447 | return self.kernel.apply_weights(layer, x, bias) |
| 448 | |
| 449 | |
| 450 | class AutoGPTQMoEMethod(FusedMoEMethodBase): |
| 451 | """MoE Marlin method with quantization.""" |
| 452 | |
| 453 | def __init__( |
| 454 | self, |
| 455 | quant_config: AutoGPTQConfig, |
| 456 | moe: FusedMoEConfig, |
| 457 | ) -> None: |
| 458 | super().__init__(moe) |
| 459 | self.quant_config = quant_config |
| 460 | if self.quant_config.quant_type.size_bits == 4: |
| 461 | quant_type = scalar_types.uint4b8 |
| 462 | scale = kInt4StaticGroupScale |
| 463 | elif self.quant_config.quant_type.size_bits == 8: |
| 464 | quant_type = scalar_types.uint8b128 |
| 465 | scale = kInt8StaticGroupScale |
| 466 | else: |
| 467 | raise ValueError("AutoGPTQMoEMethod only supports int4 and int8 now.") |
| 468 | self.input_dtype = None |
| 469 | self.use_marlin = True |
| 470 | weight_key = QuantKey(quant_type, scale) |
| 471 | |
| 472 | self.wna16_moe_backend, self.experts_cls = select_wna16_moe_backend( |
| 473 | moe, |
| 474 | weight_key, |
| 475 | quant_config=self.quant_config, |
| 476 | may_have_zp=not self.quant_config.is_sym, |
| 477 | may_have_bias=True, |
| 478 | allow_tile_padding=True, |
| 479 | ) |
| 480 | |
| 481 | def create_weights( |
| 482 | self, |
| 483 | layer: RoutedExperts, |
| 484 | num_experts: int, |
| 485 | hidden_size: int, |
| 486 | intermediate_size_per_partition: int, |
| 487 | params_dtype: torch.dtype, |
| 488 | **extra_weight_attrs, |
| 489 | ): |
| 490 | layer.input_dtype = self.input_dtype |
| 491 | is_a_8bit = self.input_dtype is not None and self.input_dtype.itemsize == 1 |
| 492 | |
| 493 | if is_a_8bit: |
| 494 | assert self.quant_config.quant_type.size_bits == 8, ( |
| 495 | "W8A8-INT8 is not supported by marlin kernel." |
| 496 | ) |
| 497 | |
| 498 | if self.quant_config.group_size != -1: |
| 499 | scales_size13 = hidden_size // self.quant_config.group_size |
| 500 | w2_scales_size = intermediate_size_per_partition |
| 501 | scales_size2 = w2_scales_size // self.quant_config.group_size |
| 502 | strategy = FusedMoeWeightScaleSupported.GROUP.value |
| 503 | else: |
| 504 | scales_size13 = 1 |
| 505 | scales_size2 = 1 |
| 506 | strategy = FusedMoeWeightScaleSupported.CHANNEL.value |
| 507 | |
| 508 | layer.num_groups_w13 = scales_size13 |
| 509 | layer.num_groups_w2 = scales_size2 |
| 510 | |
| 511 | extra_weight_attrs.update({"quant_method": strategy, "is_transposed": True}) |
| 512 | # Fused gate_up_proj (column parallel) |
| 513 | w13_qweight = torch.nn.Parameter( |
| 514 | torch.empty( |
| 515 | num_experts, |
| 516 | hidden_size // self.quant_config.pack_factor, |
| 517 | self.moe.w13_num_shards * intermediate_size_per_partition, |
| 518 | dtype=torch.int32, |
| 519 | ), |
| 520 | requires_grad=False, |
| 521 | ) |
| 522 | layer.register_parameter("w13_qweight", w13_qweight) |
| 523 | set_weight_attrs(w13_qweight, extra_weight_attrs) |
| 524 | # down_proj (row parallel) |
| 525 | w2_qweight = torch.nn.Parameter( |
| 526 | torch.empty( |
| 527 | num_experts, |
| 528 | intermediate_size_per_partition // self.quant_config.pack_factor, |
| 529 | hidden_size, |
| 530 | dtype=torch.int32, |
| 531 | ), |
| 532 | requires_grad=False, |
| 533 | ) |
| 534 | layer.register_parameter("w2_qweight", w2_qweight) |
| 535 | set_weight_attrs(w2_qweight, extra_weight_attrs) |
| 536 | # up_proj scales |
| 537 | w13_scales = torch.nn.Parameter( |
| 538 | torch.empty( |
| 539 | num_experts, |
| 540 | scales_size13, |
| 541 | self.moe.w13_num_shards * intermediate_size_per_partition, |
| 542 | dtype=params_dtype, |
| 543 | ), |
| 544 | requires_grad=False, |
| 545 | ) |
| 546 | layer.register_parameter("w13_scales", w13_scales) |
| 547 | set_weight_attrs(w13_scales, extra_weight_attrs) |
| 548 | # down_proj scales |
| 549 | w2_scales = torch.nn.Parameter( |
| 550 | torch.empty(num_experts, scales_size2, hidden_size, dtype=params_dtype), |
| 551 | requires_grad=False, |
| 552 | ) |
| 553 | layer.register_parameter("w2_scales", w2_scales) |
| 554 | set_weight_attrs(w2_scales, extra_weight_attrs) |
| 555 | # up_proj zero points |
| 556 | w13_qzeros = torch.nn.Parameter( |
| 557 | torch.empty( |
| 558 | num_experts, |
| 559 | scales_size13, |
| 560 | self.moe.w13_num_shards |
| 561 | * intermediate_size_per_partition |
| 562 | // self.quant_config.pack_factor, |
| 563 | dtype=torch.int32, |
| 564 | ), |
| 565 | requires_grad=False, |
| 566 | ) |
| 567 | layer.register_parameter("w13_qzeros", w13_qzeros) |
| 568 | set_weight_attrs(w13_qzeros, extra_weight_attrs) |
| 569 | # down_proj zero points |
| 570 | w2_qzeros = torch.nn.Parameter( |
| 571 | torch.empty( |
| 572 | num_experts, |
| 573 | scales_size2, |
| 574 | hidden_size // self.quant_config.pack_factor, |
| 575 | dtype=torch.int32, |
| 576 | ), |
| 577 | requires_grad=False, |
| 578 | ) |
| 579 | layer.register_parameter("w2_qzeros", w2_qzeros) |
| 580 | set_weight_attrs(w2_qzeros, extra_weight_attrs) |
| 581 | |
| 582 | # Some GPTQ checkpoints contain expert biases even when the model |
| 583 | # architecture does not declare them. Zero initialization keeps |
| 584 | # checkpoints without biases equivalent to the bias-free path. |
| 585 | w13_bias = torch.nn.Parameter( |
| 586 | torch.zeros( |
| 587 | num_experts, |
| 588 | self.moe.w13_num_shards * intermediate_size_per_partition, |
| 589 | dtype=params_dtype, |
| 590 | ), |
| 591 | requires_grad=False, |
| 592 | ) |
| 593 | layer.register_parameter("w13_bias", w13_bias) |
| 594 | set_weight_attrs(w13_bias, extra_weight_attrs) |
| 595 | w2_bias = torch.nn.Parameter( |
| 596 | torch.zeros(num_experts, hidden_size, dtype=params_dtype), |
| 597 | requires_grad=False, |
| 598 | ) |
| 599 | layer.register_parameter("w2_bias", w2_bias) |
| 600 | set_weight_attrs(w2_bias, extra_weight_attrs) |
| 601 | |
| 602 | def process_weights_after_loading(self, layer: RoutedExperts) -> None: |
| 603 | def replace_or_register(name: str, val: torch.Tensor | None): |
| 604 | if val is None: |
| 605 | return |
| 606 | |
| 607 | if hasattr(layer, name): |
| 608 | replace_parameter(layer, name, val) |
| 609 | else: |
| 610 | layer.register_parameter( |
| 611 | name, torch.nn.Parameter(val, requires_grad=False) |
| 612 | ) |
| 613 | |
| 614 | is_a_8bit = self.input_dtype is not None and self.input_dtype.itemsize == 1 |
| 615 | |
| 616 | assert not is_a_8bit or self.quant_config.quant_type.size_bits == 8, ( |
| 617 | "W8A8-INT8 is not supported by marlin kernel." |
| 618 | ) |
| 619 | |
| 620 | w13_bias = getattr(layer, "w13_bias", None) |
| 621 | if "w13_bias" not in layer._loaded_expert_biases: |
| 622 | layer.register_parameter("w13_bias", None) |
| 623 | w13_bias = None |
| 624 | w2_bias = getattr(layer, "w2_bias", None) |
| 625 | if "w2_bias" not in layer._loaded_expert_biases: |
| 626 | layer.register_parameter("w2_bias", None) |
| 627 | w2_bias = None |
| 628 | |
| 629 | # Normalize GPTQ K-first layout to canonical N-first format. |
| 630 | w13 = layer.w13_qweight.data.transpose(1, 2).contiguous() |
| 631 | w2 = layer.w2_qweight.data.transpose(1, 2).contiguous() |
| 632 | w13_scale = layer.w13_scales.data.transpose(1, 2).contiguous() |
| 633 | w2_scale = layer.w2_scales.data.transpose(1, 2).contiguous() |
| 634 | w13_qzeros = getattr(layer, "w13_qzeros", None) |
| 635 | w2_qzeros = getattr(layer, "w2_qzeros", None) |
| 636 | if w13_qzeros is not None: |
| 637 | w13_qzeros = w13_qzeros.data.transpose(1, 2).contiguous() |
| 638 | if w2_qzeros is not None: |
| 639 | w2_qzeros = w2_qzeros.data.transpose(1, 2).contiguous() |
| 640 | |
| 641 | converted = convert_to_wna16_moe_kernel_format( |
| 642 | backend=self.wna16_moe_backend, |
| 643 | layer=layer, |
| 644 | quant_config=self.quant_config, |
| 645 | input_dtype=self.input_dtype, |
| 646 | w13=w13, |
| 647 | w2=w2, |
| 648 | w13_scale=w13_scale, |
| 649 | w2_scale=w2_scale, |
| 650 | w13_bias=w13_bias, |
| 651 | w2_bias=w2_bias, |
| 652 | w13_qzeros=w13_qzeros, |
| 653 | w2_qzeros=w2_qzeros, |
| 654 | ) |
| 655 | |
| 656 | if converted is None: |
| 657 | # Backend rewrote the layer's params in place (e.g. Humming). |
| 658 | self._setup_kernel(layer) |
| 659 | return |
| 660 | |
| 661 | ( |
| 662 | w13, |
| 663 | w2, |
| 664 | w13_scale, |
| 665 | w2_scale, |
| 666 | w13_qzeros, |
| 667 | w2_qzeros, |
| 668 | w13_input_global_scale, |
| 669 | w2_input_global_scale, |
| 670 | w13_bias, |
| 671 | w2_bias, |
| 672 | ) = converted |
| 673 | |
| 674 | replace_parameter(layer, "w13_qweight", w13) |
| 675 | replace_parameter(layer, "w2_qweight", w2) |
| 676 | replace_parameter(layer, "w13_scales", w13_scale) |
| 677 | replace_parameter(layer, "w2_scales", w2_scale) |
| 678 | replace_or_register("w13_input_global_scale", w13_input_global_scale) |
| 679 | replace_or_register("w2_input_global_scale", w2_input_global_scale) |
| 680 | replace_or_register("w13_bias", w13_bias) |
| 681 | replace_or_register("w2_bias", w2_bias) |
| 682 | replace_or_register("w13_qzeros", w13_qzeros) |
| 683 | replace_or_register("w2_qzeros", w2_qzeros) |
| 684 | |
| 685 | # The modular kernel reads w13_weight/w2_weight; marlin keeps *_qweight. |
| 686 | layer.w13_weight = layer.w13_qweight |
| 687 | layer.w2_weight = layer.w2_qweight |
| 688 | |
| 689 | self._setup_kernel(layer) |
| 690 | |
| 691 | def _setup_kernel(self, layer: RoutedExperts) -> None: |
| 692 | """Build the FusedMoEKernel for this layer.""" |
| 693 | self.moe_quant_config = self.get_fused_moe_quant_config(layer) |
| 694 | self.moe_kernel = make_wna16_moe_kernel( |
| 695 | moe_quant_config=self.moe_quant_config, |
| 696 | moe_config=self.moe, |
| 697 | experts_cls=self.experts_cls, |
| 698 | backend=self.wna16_moe_backend, |
| 699 | routing_tables=layer._expert_routing_tables(), |
| 700 | ) |
| 701 | |
| 702 | def get_fused_moe_quant_config(self, layer: RoutedExperts) -> FusedMoEQuantConfig: |
| 703 | if self.wna16_moe_backend == WNA16MoEBackend.HUMMING: |
| 704 | from vllm.model_executor.layers.quantization.utils.humming import ( |
| 705 | get_humming_moe_quant_config, |
| 706 | ) |
| 707 | |
| 708 | return get_humming_moe_quant_config( |
| 709 | layer, |
| 710 | gemm1_alpha=getattr(layer, "swiglu_alpha", None), |
| 711 | gemm1_beta=getattr(layer, "swiglu_beta", None), |
| 712 | gemm1_clamp_limit=getattr(layer, "swiglu_limit", None), |
| 713 | ) |
| 714 | |
| 715 | from vllm.model_executor.layers.fused_moe.config import ( |
| 716 | gptq_marlin_moe_quant_config, |
| 717 | ) |
| 718 | |
| 719 | # CPU fused_experts_cpu requires zero points even for symmetric quant |
| 720 | use_zp = ( |
| 721 | not self.quant_config.is_sym |
| 722 | or self.wna16_moe_backend == WNA16MoEBackend.CPU |
| 723 | ) |
| 724 | return gptq_marlin_moe_quant_config( |
| 725 | w1_scale=layer.w13_scales, |
| 726 | w2_scale=layer.w2_scales, |
| 727 | weight_bits=self.quant_config.weight_bits, |
| 728 | group_size=self.quant_config.group_size, |
| 729 | w1_zp=getattr(layer, "w13_qzeros", None) if use_zp else None, |
| 730 | w2_zp=getattr(layer, "w2_qzeros", None) if use_zp else None, |
| 731 | w1_bias=getattr(layer, "w13_bias", None), |
| 732 | w2_bias=getattr(layer, "w2_bias", None), |
| 733 | ) |
| 734 | |
| 735 | def apply( |
| 736 | self, |
| 737 | layer: RoutedExperts, |
| 738 | x: torch.Tensor, |
| 739 | topk_weights: torch.Tensor, |
| 740 | topk_ids: torch.Tensor, |
| 741 | shared_experts: SharedExperts | None, |
| 742 | shared_experts_input: torch.Tensor | None, |
| 743 | ) -> torch.Tensor: |
| 744 | assert not self.is_monolithic |
| 745 | assert self.moe_kernel is not None |
| 746 | return self.moe_kernel.apply( |
| 747 | hidden_states=x, |
| 748 | w1=layer.w13_weight, |
| 749 | w2=layer.w2_weight, |
| 750 | topk_weights=topk_weights, |
| 751 | topk_ids=topk_ids, |
| 752 | activation=layer.activation, |
| 753 | global_num_experts=layer.global_num_experts, |
| 754 | apply_router_weight_on_input=layer.apply_router_weight_on_input, |
| 755 | expert_map=layer.expert_map, |
| 756 | shared_experts=shared_experts, |
| 757 | shared_experts_input=shared_experts_input, |
| 758 | ) |
| 759 | |
| 760 | def apply_monolithic( |
| 761 | self, |
| 762 | layer: RoutedExperts, |
| 763 | x: torch.Tensor, |
| 764 | router_logits: torch.Tensor, |
| 765 | input_ids: torch.Tensor | None = None, |
| 766 | ) -> torch.Tensor: |
| 767 | assert self.is_monolithic |
| 768 | assert self.moe_kernel is not None |
| 769 | return self.moe_kernel.apply_monolithic( |
| 770 | hidden_states=x, |
| 771 | w1=layer.w13_weight, |
| 772 | w2=layer.w2_weight, |
| 773 | router_logits=router_logits, |
| 774 | activation=layer.activation, |
| 775 | global_num_experts=layer.global_num_experts, |
| 776 | expert_map=layer.expert_map, |
| 777 | apply_router_weight_on_input=layer.apply_router_weight_on_input, |
| 778 | num_expert_group=layer.num_expert_group, |
| 779 | topk_group=layer.topk_group, |
| 780 | e_score_correction_bias=layer.e_score_correction_bias, |
| 781 | routed_scaling_factor=layer.routed_scaling_factor, |
| 782 | ) |