| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Copyright 2025 The ZhipuAI Team. |
| 5 | # Copyright 2023 The vLLM team. |
| 6 | # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. |
| 7 | # |
| 8 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX |
| 9 | # and OPT implementations in this library. It has been modified from its |
| 10 | # original forms to accommodate minor architectural differences compared |
| 11 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. |
| 12 | # |
| 13 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 14 | # you may not use this file except in compliance with the License. |
| 15 | # You may obtain a copy of the License at |
| 16 | # |
| 17 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 18 | # |
| 19 | # Unless required by applicable law or agreed to in writing, software |
| 20 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 21 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 22 | # See the License for the specific language governing permissions and |
| 23 | # limitations under the License. |
| 24 | """Inference-only GLM-4.7-Flash model compatible with HuggingFace weights.""" |
| 25 | |
| 26 | import typing |
| 27 | from collections.abc import Callable, Iterable |
| 28 | from itertools import islice |
| 29 | from typing import TYPE_CHECKING |
| 30 | |
| 31 | import torch |
| 32 | from torch import nn |
| 33 | |
| 34 | if TYPE_CHECKING: |
| 35 | from transformers.models.glm4_moe_lite import Glm4MoeLiteConfig |
| 36 | |
| 37 | from vllm.compilation.decorators import support_torch_compile |
| 38 | from vllm.config import VllmConfig |
| 39 | from vllm.distributed import ( |
| 40 | get_pp_group, |
| 41 | ) |
| 42 | from vllm.logger import init_logger |
| 43 | from vllm.model_executor.layers.fused_moe import ( |
| 44 | fused_moe_make_expert_params_mapping, |
| 45 | ) |
| 46 | from vllm.model_executor.layers.fused_moe.utils import ( |
| 47 | is_model_fused_shared_expert_compatible, |
| 48 | ) |
| 49 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 50 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 51 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 52 | ParallelLMHead, |
| 53 | VocabParallelEmbedding, |
| 54 | ) |
| 55 | from vllm.model_executor.model_loader.weight_utils import ( |
| 56 | default_weight_loader, |
| 57 | maybe_remap_kv_scale_name, |
| 58 | ) |
| 59 | from vllm.model_executor.models.deepseek_v2 import ( |
| 60 | DeepseekV2Attention, |
| 61 | DeepseekV2MLAAttention, |
| 62 | ) |
| 63 | from vllm.model_executor.models.glm4_moe import ( |
| 64 | Glm4MixtureOfExperts, |
| 65 | Glm4MoE, |
| 66 | Glm4MoeMLP, |
| 67 | ) |
| 68 | from vllm.platforms import current_platform |
| 69 | from vllm.sequence import IntermediateTensors |
| 70 | from vllm.v1.attention.backends.mla.index_group import ( |
| 71 | SparseMLAIndexGroupBuilder, |
| 72 | get_sparse_mla_index_group_max_rows, |
| 73 | ) |
| 74 | |
| 75 | from .interfaces import SupportsLoRA, SupportsPP |
| 76 | from .utils import ( |
| 77 | AutoWeightsLoader, |
| 78 | PPMissingLayer, |
| 79 | get_spec_layer_idx_from_weight_name, |
| 80 | is_pp_missing_parameter, |
| 81 | make_layers, |
| 82 | maybe_prefix, |
| 83 | ) |
| 84 | |
| 85 | logger = init_logger(__name__) |
| 86 | |
| 87 | |
| 88 | class Glm4MoeLiteMLP(Glm4MoeMLP): |
| 89 | pass |
| 90 | |
| 91 | |
| 92 | class Glm4MoeLite(Glm4MoE): |
| 93 | pass |
| 94 | |
| 95 | |
| 96 | class Glm4LiteMixtureOfExperts(Glm4MixtureOfExperts): |
| 97 | pass |
| 98 | |
| 99 | |
| 100 | class Glm4MoeLiteAttention(DeepseekV2Attention): |
| 101 | pass |
| 102 | |
| 103 | |
| 104 | class Glm4MoeLiteMLAAttention(DeepseekV2MLAAttention): |
| 105 | pass |
| 106 | |
| 107 | |
| 108 | class Glm4MoeLiteDecoderLayer(nn.Module): |
| 109 | def __init__( |
| 110 | self, |
| 111 | vllm_config: VllmConfig, |
| 112 | prefix: str, |
| 113 | config: "Glm4MoeLiteConfig | None" = None, |
| 114 | topk_indices_buffer: torch.Tensor | None = None, |
| 115 | index_group_builder: SparseMLAIndexGroupBuilder | None = None, |
| 116 | ) -> None: |
| 117 | super().__init__() |
| 118 | |
| 119 | if config is None: |
| 120 | config = vllm_config.model_config.hf_config |
| 121 | model_config = vllm_config.model_config |
| 122 | cache_config = vllm_config.cache_config |
| 123 | quant_config = vllm_config.quant_config |
| 124 | |
| 125 | self.hidden_size = config.hidden_size |
| 126 | max_position_embeddings = getattr(config, "max_position_embeddings", 8192) |
| 127 | moe_layer_freq = getattr(config, "moe_layer_freq", 1) |
| 128 | # DecoderLayers are created with `make_layers` which passes the prefix |
| 129 | # with the layer's index. |
| 130 | layer_idx = int(prefix.split(sep=".")[-1]) |
| 131 | self.layer_idx = layer_idx |
| 132 | |
| 133 | # verify MLA attention specific fields |
| 134 | qk_nope_head_dim = getattr(config, "qk_nope_head_dim", 0) |
| 135 | qk_rope_head_dim = getattr(config, "qk_rope_head_dim", 0) |
| 136 | v_head_dim = getattr(config, "v_head_dim", 0) |
| 137 | kv_lora_rank = getattr(config, "kv_lora_rank", 0) |
| 138 | |
| 139 | attn_cls: Callable[..., Glm4MoeLiteMLAAttention | Glm4MoeLiteAttention] |
| 140 | if model_config.use_mla: |
| 141 | attn_cls = Glm4MoeLiteMLAAttention |
| 142 | else: |
| 143 | attn_cls = Glm4MoeLiteAttention |
| 144 | attn_kwargs: dict[str, typing.Any] = ( |
| 145 | {"index_group_builder": index_group_builder} |
| 146 | if attn_cls is Glm4MoeLiteMLAAttention |
| 147 | else {} |
| 148 | ) |
| 149 | |
| 150 | self.self_attn = attn_cls( |
| 151 | vllm_config=vllm_config, |
| 152 | config=config, |
| 153 | hidden_size=self.hidden_size, |
| 154 | num_heads=config.num_attention_heads, |
| 155 | qk_nope_head_dim=qk_nope_head_dim, |
| 156 | qk_rope_head_dim=qk_rope_head_dim, |
| 157 | v_head_dim=v_head_dim, |
| 158 | q_lora_rank=getattr(config, "q_lora_rank", None), |
| 159 | kv_lora_rank=kv_lora_rank, |
| 160 | max_position_embeddings=max_position_embeddings, |
| 161 | cache_config=cache_config, |
| 162 | quant_config=quant_config, |
| 163 | prefix=f"{prefix}.self_attn", |
| 164 | topk_indices_buffer=topk_indices_buffer, |
| 165 | **attn_kwargs, |
| 166 | ) |
| 167 | |
| 168 | if ( |
| 169 | config.n_routed_experts is not None |
| 170 | and layer_idx >= config.first_k_dense_replace |
| 171 | and layer_idx % moe_layer_freq == 0 |
| 172 | ): |
| 173 | self.mlp = Glm4MoeLite( |
| 174 | config=config, |
| 175 | quant_config=quant_config, |
| 176 | prefix=f"{prefix}.mlp", |
| 177 | ) |
| 178 | else: |
| 179 | self.mlp = Glm4MoeLiteMLP( |
| 180 | hidden_size=config.hidden_size, |
| 181 | intermediate_size=config.intermediate_size, |
| 182 | hidden_act=config.hidden_act, |
| 183 | quant_config=quant_config, |
| 184 | prefix=f"{prefix}.mlp", |
| 185 | ) |
| 186 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 187 | self.post_attention_layernorm = RMSNorm( |
| 188 | config.hidden_size, eps=config.rms_norm_eps |
| 189 | ) |
| 190 | self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0) |
| 191 | |
| 192 | def forward( |
| 193 | self, |
| 194 | positions: torch.Tensor, |
| 195 | hidden_states: torch.Tensor, |
| 196 | residual: torch.Tensor | None, |
| 197 | llama_4_scaling: torch.Tensor | None = None, |
| 198 | ) -> torch.Tensor: |
| 199 | # Self Attention |
| 200 | if residual is None: |
| 201 | residual = hidden_states |
| 202 | hidden_states = self.input_layernorm(hidden_states) |
| 203 | else: |
| 204 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 205 | |
| 206 | attn_kwargs = { |
| 207 | "positions": positions, |
| 208 | "hidden_states": hidden_states, |
| 209 | } |
| 210 | attn_kwargs["llama_4_scaling"] = llama_4_scaling |
| 211 | hidden_states = self.self_attn(**attn_kwargs) |
| 212 | |
| 213 | hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) |
| 214 | hidden_states = self.mlp(hidden_states) |
| 215 | |
| 216 | return hidden_states, residual |
| 217 | |
| 218 | |
| 219 | @support_torch_compile( |
| 220 | dynamic_arg_dims={ |
| 221 | "input_ids": 0, |
| 222 | "positions": -1, |
| 223 | "intermediate_tensors": 0, |
| 224 | "inputs_embeds": 0, |
| 225 | } |
| 226 | ) |
| 227 | class Glm4MoeLiteModel(nn.Module): |
| 228 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 229 | super().__init__() |
| 230 | |
| 231 | config = vllm_config.model_config.hf_config |
| 232 | quant_config = vllm_config.quant_config |
| 233 | self.config = config |
| 234 | self.device = current_platform.device_type |
| 235 | |
| 236 | self.vocab_size = config.vocab_size |
| 237 | self.is_v32 = hasattr(config, "index_topk") |
| 238 | if self.is_v32: |
| 239 | topk_tokens = config.index_topk |
| 240 | topk_indices_buffer = torch.empty( |
| 241 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 242 | topk_tokens, |
| 243 | dtype=torch.int32, |
| 244 | device=self.device, |
| 245 | ) |
| 246 | else: |
| 247 | topk_indices_buffer = None |
| 248 | index_group_builder = ( |
| 249 | SparseMLAIndexGroupBuilder( |
| 250 | topk_indices_buffer, |
| 251 | get_sparse_mla_index_group_max_rows(vllm_config), |
| 252 | ) |
| 253 | if topk_indices_buffer is not None |
| 254 | else None |
| 255 | ) |
| 256 | |
| 257 | if get_pp_group().is_first_rank: |
| 258 | self.embed_tokens = VocabParallelEmbedding( |
| 259 | config.vocab_size, |
| 260 | config.hidden_size, |
| 261 | quant_config=quant_config, |
| 262 | prefix=f"{prefix}.embed_tokens", |
| 263 | ) |
| 264 | else: |
| 265 | self.embed_tokens = PPMissingLayer() |
| 266 | |
| 267 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 268 | config.num_hidden_layers, |
| 269 | lambda prefix: Glm4MoeLiteDecoderLayer( |
| 270 | vllm_config=vllm_config, |
| 271 | config=config, |
| 272 | prefix=prefix, |
| 273 | topk_indices_buffer=topk_indices_buffer, |
| 274 | index_group_builder=index_group_builder, |
| 275 | ), |
| 276 | prefix=f"{prefix}.layers", |
| 277 | ) |
| 278 | |
| 279 | self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible( |
| 280 | self.layers, |
| 281 | Glm4MoeLite, |
| 282 | "mlp", |
| 283 | ) |
| 284 | |
| 285 | if get_pp_group().is_last_rank: |
| 286 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 287 | else: |
| 288 | self.norm = PPMissingLayer() |
| 289 | |
| 290 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 291 | return self.embed_tokens(input_ids) |
| 292 | |
| 293 | def forward( |
| 294 | self, |
| 295 | input_ids: torch.Tensor | None, |
| 296 | positions: torch.Tensor, |
| 297 | intermediate_tensors: IntermediateTensors | None = None, |
| 298 | inputs_embeds: torch.Tensor | None = None, |
| 299 | ) -> torch.Tensor | IntermediateTensors: |
| 300 | if get_pp_group().is_first_rank: |
| 301 | if inputs_embeds is not None: |
| 302 | hidden_states = inputs_embeds |
| 303 | else: |
| 304 | hidden_states = self.embed_input_ids(input_ids) |
| 305 | residual = None |
| 306 | else: |
| 307 | assert intermediate_tensors is not None |
| 308 | hidden_states = intermediate_tensors["hidden_states"] |
| 309 | residual = intermediate_tensors["residual"] |
| 310 | |
| 311 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 312 | hidden_states, residual = layer(positions, hidden_states, residual) |
| 313 | |
| 314 | if not get_pp_group().is_last_rank: |
| 315 | return IntermediateTensors( |
| 316 | {"hidden_states": hidden_states, "residual": residual} |
| 317 | ) |
| 318 | |
| 319 | hidden_states, _ = self.norm(hidden_states, residual) |
| 320 | return hidden_states |
| 321 | |
| 322 | def make_empty_intermediate_tensors( |
| 323 | self, batch_size: int, dtype: torch.dtype, device: torch.device |
| 324 | ) -> IntermediateTensors: |
| 325 | return IntermediateTensors( |
| 326 | { |
| 327 | "hidden_states": torch.zeros( |
| 328 | (batch_size, self.config.hidden_size), dtype=dtype, device=device |
| 329 | ), |
| 330 | "residual": torch.zeros( |
| 331 | (batch_size, self.config.hidden_size), dtype=dtype, device=device |
| 332 | ), |
| 333 | } |
| 334 | ) |
| 335 | |
| 336 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 337 | # Params for weights, fp8 weight scales, fp8 activation scales |
| 338 | # (param_name, weight_name, expert_id, shard_id) |
| 339 | return fused_moe_make_expert_params_mapping( |
| 340 | self, |
| 341 | ckpt_gate_proj_name="gate_proj", |
| 342 | ckpt_down_proj_name="down_proj", |
| 343 | ckpt_up_proj_name="up_proj", |
| 344 | num_experts=self.config.n_routed_experts, |
| 345 | ) |
| 346 | |
| 347 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 348 | stacked_params_mapping = [ |
| 349 | # (param_name, shard_name, shard_id) |
| 350 | ("gate_up_proj", "gate_proj", 0), |
| 351 | ("gate_up_proj", "up_proj", 1), |
| 352 | ] |
| 353 | mla_params_mapping = [ |
| 354 | ("fused_qkv_a_proj", "q_a_proj", 0), |
| 355 | ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1), |
| 356 | ] |
| 357 | |
| 358 | stacked_params_mapping.extend(mla_params_mapping) |
| 359 | |
| 360 | # Params for weights, fp8 weight scales, fp8 activation scales |
| 361 | # (param_name, weight_name, expert_id, shard_id) |
| 362 | expert_params_mapping = fused_moe_make_expert_params_mapping( |
| 363 | self, |
| 364 | ckpt_gate_proj_name="gate_proj", |
| 365 | ckpt_down_proj_name="down_proj", |
| 366 | ckpt_up_proj_name="up_proj", |
| 367 | num_experts=self.config.n_routed_experts |
| 368 | + ( |
| 369 | self.config.n_shared_experts |
| 370 | if self.is_fused_shared_expert_enabled |
| 371 | else 0 |
| 372 | ), |
| 373 | ) |
| 374 | |
| 375 | params_dict = dict(self.named_parameters()) |
| 376 | loaded_params: set[str] = set() |
| 377 | for name, loaded_weight in weights: |
| 378 | if "rotary_emb.inv_freq" in name: |
| 379 | continue |
| 380 | |
| 381 | spec_layer = get_spec_layer_idx_from_weight_name(self.config, name) |
| 382 | if spec_layer is not None: |
| 383 | continue # skip spec decode layers for main model |
| 384 | |
| 385 | is_fusion_moe_shared_experts_layer = ( |
| 386 | self.is_fused_shared_expert_enabled and ("mlp.shared_experts" in name) |
| 387 | ) |
| 388 | |
| 389 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 390 | # Skip non-stacked layers and experts (experts handled below). |
| 391 | if weight_name not in name: |
| 392 | continue |
| 393 | # We have mlp.experts[0].gate_proj in the checkpoint. |
| 394 | # Since we handle the experts below in expert_params_mapping, |
| 395 | # we need to skip here BEFORE we update the name, otherwise |
| 396 | # name will be updated to mlp.experts[0].gate_up_proj, which |
| 397 | # will then be updated below in expert_params_mapping |
| 398 | # for mlp.experts[0].gate_gate_up_proj, which breaks load. |
| 399 | if ("mlp.experts." in name) and name not in params_dict: |
| 400 | continue |
| 401 | if is_fusion_moe_shared_experts_layer: |
| 402 | continue |
| 403 | name_mapped = name.replace(weight_name, param_name) |
| 404 | |
| 405 | # QKV fusion is optional, fall back to normal |
| 406 | # weight loading if it's not enabled |
| 407 | # if go with fusion option, then update name |
| 408 | if ( |
| 409 | param_name == "fused_qkv_a_proj" |
| 410 | ) and name_mapped not in params_dict: |
| 411 | continue |
| 412 | else: |
| 413 | name = name_mapped |
| 414 | # Skip loading extra bias for GPTQ models. |
| 415 | if name.endswith(".bias") and name not in params_dict: |
| 416 | continue |
| 417 | |
| 418 | if is_pp_missing_parameter(name, self): |
| 419 | continue |
| 420 | |
| 421 | param = params_dict[name] |
| 422 | weight_loader = param.weight_loader |
| 423 | weight_loader(param, loaded_weight, shard_id) |
| 424 | break |
| 425 | else: |
| 426 | is_expert_weight = False |
| 427 | |
| 428 | # Special handling: when AITER fusion_shared_experts is enabled, |
| 429 | # checkpoints may provide a single widened shared_experts tensor |
| 430 | # without explicit expert indices |
| 431 | # (e.g. ...mlp.shared_experts.gate_proj.weight). |
| 432 | # For models with multiple shared experts, split that tensor |
| 433 | # evenly into per-shared-expert slices and load them into |
| 434 | # appended expert slots mlp.experts.{n_routed_experts + j}.* |
| 435 | # accordingly. |
| 436 | num_chunks = 1 |
| 437 | if is_fusion_moe_shared_experts_layer: |
| 438 | num_chunks = getattr(self.config, "n_shared_experts", 1) or 1 |
| 439 | # Determine split axis based on op type |
| 440 | # gate/up: ColumnParallel → split along dim 0 |
| 441 | # down: RowParallel → split along dim 1 |
| 442 | split_dim = 1 if "down_proj.weight" in name else 0 |
| 443 | total = loaded_weight.shape[split_dim] |
| 444 | assert total % num_chunks == 0, ( |
| 445 | f"Shared expert weight dim {total} " |
| 446 | f"not divisible by num_chunks {num_chunks}" |
| 447 | ) |
| 448 | chunk_size = total // num_chunks |
| 449 | |
| 450 | for j in range(num_chunks): |
| 451 | chunk_name = name |
| 452 | weight_to_load = loaded_weight |
| 453 | |
| 454 | if is_fusion_moe_shared_experts_layer: |
| 455 | if split_dim == 0: |
| 456 | weight_to_load = loaded_weight[ |
| 457 | j * chunk_size : (j + 1) * chunk_size, : |
| 458 | ] |
| 459 | else: |
| 460 | weight_to_load = loaded_weight[ |
| 461 | :, j * chunk_size : (j + 1) * chunk_size |
| 462 | ] |
| 463 | # Synthesize an expert-style name so expert mapping |
| 464 | # can route it |
| 465 | chunk_name = name.replace( |
| 466 | "mlp.shared_experts", |
| 467 | f"mlp.experts.{self.config.n_routed_experts + j}", |
| 468 | ) |
| 469 | |
| 470 | # Use expert_params_mapping to locate the destination |
| 471 | # param and delegate to its expert-aware weight_loader |
| 472 | # with expert_id. |
| 473 | for mapping in expert_params_mapping: |
| 474 | param_name, weight_name, mapping_expert_id, mapping_shard_id = ( |
| 475 | mapping |
| 476 | ) |
| 477 | if weight_name not in chunk_name: |
| 478 | continue |
| 479 | |
| 480 | # Anyway, this is an expert weight and should not be |
| 481 | # attempted to load as other weights later |
| 482 | is_expert_weight = True |
| 483 | |
| 484 | # Do not modify `name` since the loop may continue here |
| 485 | # Instead, create a new variable |
| 486 | name_mapped = chunk_name.replace(weight_name, param_name) |
| 487 | |
| 488 | if is_pp_missing_parameter(name_mapped, self): |
| 489 | continue |
| 490 | |
| 491 | param = params_dict[name_mapped] |
| 492 | # We should ask the weight loader to return success or |
| 493 | # not here since otherwise we may skip experts with |
| 494 | # other available replicas. |
| 495 | weight_loader = typing.cast( |
| 496 | Callable[..., bool], param.weight_loader |
| 497 | ) |
| 498 | success = weight_loader( |
| 499 | param, |
| 500 | weight_to_load, |
| 501 | name_mapped, |
| 502 | shard_id=mapping_shard_id, |
| 503 | expert_id=mapping_expert_id, |
| 504 | return_success=True, |
| 505 | ) |
| 506 | if success: |
| 507 | if not is_fusion_moe_shared_experts_layer: |
| 508 | name = name_mapped |
| 509 | else: |
| 510 | loaded_params.add(name_mapped) |
| 511 | break |
| 512 | else: |
| 513 | if is_expert_weight: |
| 514 | # We've checked that this is an expert weight |
| 515 | # However it's not mapped locally to this rank |
| 516 | # So we simply skip it |
| 517 | continue |
| 518 | |
| 519 | # Skip loading extra bias for GPTQ models. |
| 520 | if name.endswith(".bias") and name not in params_dict: |
| 521 | continue |
| 522 | |
| 523 | # Remapping the name of FP8 kv-scale. |
| 524 | remapped_name = maybe_remap_kv_scale_name(name, params_dict) |
| 525 | if remapped_name is None: |
| 526 | continue |
| 527 | name = remapped_name |
| 528 | |
| 529 | if is_pp_missing_parameter(name, self): |
| 530 | continue |
| 531 | |
| 532 | param = params_dict[name] |
| 533 | weight_loader = getattr( |
| 534 | param, "weight_loader", default_weight_loader |
| 535 | ) |
| 536 | weight_loader(param, loaded_weight) |
| 537 | if not is_fusion_moe_shared_experts_layer: |
| 538 | loaded_params.add(name) |
| 539 | |
| 540 | return loaded_params |
| 541 | |
| 542 | |
| 543 | class Glm4MoeLiteForCausalLM( |
| 544 | nn.Module, SupportsPP, SupportsLoRA, Glm4LiteMixtureOfExperts |
| 545 | ): |
| 546 | packed_modules_mapping = { |
| 547 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 548 | } |
| 549 | |
| 550 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 551 | super().__init__() |
| 552 | config = vllm_config.model_config.hf_config |
| 553 | quant_config = vllm_config.quant_config |
| 554 | self.config = config |
| 555 | self.quant_config = quant_config |
| 556 | |
| 557 | qk_nope_head_dim = getattr(config, "qk_nope_head_dim", 0) |
| 558 | qk_rope_head_dim = getattr(config, "qk_rope_head_dim", 0) |
| 559 | self.use_mha = config.model_type == "deepseek" or all( |
| 560 | dim == 0 for dim in (qk_nope_head_dim, qk_rope_head_dim) |
| 561 | ) |
| 562 | |
| 563 | if self.use_mha: |
| 564 | self.packed_modules_mapping["qkv_proj"] = ["q_proj", "k_proj", "v_proj"] |
| 565 | |
| 566 | # `packed_modules_mapping` needs to be modified before |
| 567 | # initializing DeepseekV2Model, as it is passed inplace to |
| 568 | # quantization config init and may be used to select the |
| 569 | # quant_method for relevant layers during initialization. |
| 570 | self.fuse_qkv_a_proj = ( |
| 571 | hasattr(config, "q_lora_rank") and config.q_lora_rank is not None |
| 572 | ) |
| 573 | if self.fuse_qkv_a_proj: |
| 574 | self.packed_modules_mapping["fused_qkv_a_proj"] = [ |
| 575 | "q_a_proj", |
| 576 | "kv_a_proj_with_mqa", |
| 577 | ] |
| 578 | |
| 579 | self.model = Glm4MoeLiteModel( |
| 580 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 581 | ) |
| 582 | if get_pp_group().is_last_rank: |
| 583 | self.lm_head = ParallelLMHead( |
| 584 | config.vocab_size, |
| 585 | config.hidden_size, |
| 586 | quant_config=quant_config, |
| 587 | prefix=maybe_prefix(prefix, "lm_head"), |
| 588 | ) |
| 589 | else: |
| 590 | self.lm_head = PPMissingLayer() |
| 591 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 592 | self.make_empty_intermediate_tensors = ( |
| 593 | self.model.make_empty_intermediate_tensors |
| 594 | ) |
| 595 | # Set MoE hyperparameters |
| 596 | self.num_moe_layers = ( |
| 597 | self.config.num_hidden_layers - self.config.first_k_dense_replace |
| 598 | ) |
| 599 | self.set_moe_parameters() |
| 600 | |
| 601 | def set_moe_parameters(self): |
| 602 | self.num_expert_groups = getattr(self.config, "n_group", 1) |
| 603 | |
| 604 | self.moe_layers = [] |
| 605 | self.moe_mlp_layers = [] |
| 606 | example_moe = None |
| 607 | for layer in self.model.layers: |
| 608 | if isinstance(layer, PPMissingLayer): |
| 609 | continue |
| 610 | |
| 611 | assert isinstance(layer, Glm4MoeLiteDecoderLayer) |
| 612 | if isinstance(layer.mlp, Glm4MoeLite): |
| 613 | # Pick last one layer since the first ones may be dense layers. |
| 614 | example_moe = layer.mlp |
| 615 | self.moe_mlp_layers.append(layer.mlp) |
| 616 | self.moe_layers.append(layer.mlp.experts) |
| 617 | |
| 618 | self.extract_moe_parameters(example_moe) |
| 619 | |
| 620 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 621 | return self.model.embed_input_ids(input_ids) |
| 622 | |
| 623 | def forward( |
| 624 | self, |
| 625 | input_ids: torch.Tensor | None, |
| 626 | positions: torch.Tensor, |
| 627 | intermediate_tensors: IntermediateTensors | None = None, |
| 628 | inputs_embeds: torch.Tensor | None = None, |
| 629 | ) -> torch.Tensor | IntermediateTensors: |
| 630 | hidden_states = self.model( |
| 631 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 632 | ) |
| 633 | return hidden_states |
| 634 | |
| 635 | def compute_logits( |
| 636 | self, |
| 637 | hidden_states: torch.Tensor, |
| 638 | ) -> torch.Tensor | None: |
| 639 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 640 | return logits |
| 641 | |
| 642 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 643 | # Params for weights, fp8 weight scales, fp8 activation scales |
| 644 | # (param_name, weight_name, expert_id, shard_id) |
| 645 | return fused_moe_make_expert_params_mapping( |
| 646 | self, |
| 647 | ckpt_gate_proj_name="gate_proj", |
| 648 | ckpt_down_proj_name="down_proj", |
| 649 | ckpt_up_proj_name="up_proj", |
| 650 | num_experts=self.config.n_routed_experts, |
| 651 | num_redundant_experts=0, |
| 652 | ) |
| 653 | |
| 654 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 655 | loader = AutoWeightsLoader(self) |
| 656 | return loader.load_weights(weights) |