| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved. |
| 4 | # Copyright 2023 The vLLM team. |
| 5 | # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. |
| 6 | # |
| 7 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX |
| 8 | # and OPT implementations in this library. It has been modified from its |
| 9 | # original forms to accommodate minor architectural differences compared |
| 10 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. |
| 11 | # |
| 12 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 13 | # you may not use this file except in compliance with the License. |
| 14 | # You may obtain a copy of the License at |
| 15 | # |
| 16 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 17 | # |
| 18 | # Unless required by applicable law or agreed to in writing, software |
| 19 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 20 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 21 | # See the License for the specific language governing permissions and |
| 22 | # limitations under the License. |
| 23 | |
| 24 | from collections.abc import Iterable |
| 25 | from typing import Any |
| 26 | |
| 27 | import torch |
| 28 | from torch import nn |
| 29 | from transformers import PreTrainedConfig |
| 30 | |
| 31 | from vllm.compilation.decorators import support_torch_compile |
| 32 | from vllm.config import CacheConfig, ParallelConfig, VllmConfig |
| 33 | from vllm.distributed import ( |
| 34 | get_ep_group, |
| 35 | get_pp_group, |
| 36 | get_tensor_model_parallel_rank, |
| 37 | get_tensor_model_parallel_world_size, |
| 38 | get_tp_group, |
| 39 | tensor_model_parallel_all_gather, |
| 40 | ) |
| 41 | from vllm.model_executor.layers.activation import SiluAndMul |
| 42 | from vllm.model_executor.layers.attention import ( |
| 43 | Attention, |
| 44 | StaticSinkAttention, |
| 45 | ) |
| 46 | from vllm.model_executor.layers.fused_moe import ( |
| 47 | FusedMoEFactory, |
| 48 | GateLinear, |
| 49 | ) |
| 50 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 51 | from vllm.model_executor.layers.linear import ( |
| 52 | ColumnParallelLinear, |
| 53 | MergedColumnParallelLinear, |
| 54 | QKVParallelLinear, |
| 55 | ReplicatedLinear, |
| 56 | RowParallelLinear, |
| 57 | ) |
| 58 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 59 | from vllm.model_executor.layers.mla import MLAModules, MultiHeadLatentAttentionWrapper |
| 60 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 61 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 62 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 63 | ParallelLMHead, |
| 64 | VocabParallelEmbedding, |
| 65 | ) |
| 66 | from vllm.model_executor.models.interfaces import ( |
| 67 | MixtureOfExperts, |
| 68 | SupportsLoRA, |
| 69 | SupportsPP, |
| 70 | ) |
| 71 | from vllm.model_executor.models.utils import ( |
| 72 | AutoWeightsLoader, |
| 73 | PPMissingLayer, |
| 74 | WeightsMapper, |
| 75 | extract_layer_index, |
| 76 | make_empty_intermediate_tensors_factory, |
| 77 | make_layers, |
| 78 | maybe_prefix, |
| 79 | sequence_parallel_chunk, |
| 80 | ) |
| 81 | from vllm.model_executor.utils import set_weight_attrs |
| 82 | from vllm.platforms import current_platform |
| 83 | from vllm.sequence import IntermediateTensors |
| 84 | from vllm.transformers_utils.config import set_default_rope_theta |
| 85 | from vllm.v1.attention.backend import AttentionType |
| 86 | from vllm.v1.attention.backends.flash_attn_diffkv import FlashAttentionDiffKVBackend |
| 87 | |
| 88 | |
| 89 | def check_ffn_act_fn(act_fn: str): |
| 90 | if act_fn != "silu": |
| 91 | raise ValueError( |
| 92 | f"Unsupported activation: {act_fn}. Only silu is supported for now." |
| 93 | ) |
| 94 | |
| 95 | |
| 96 | class OpenPanguMLP(nn.Module): |
| 97 | def __init__( |
| 98 | self, |
| 99 | hidden_size: int, |
| 100 | intermediate_size: int, |
| 101 | hidden_act: str, |
| 102 | quant_config: QuantizationConfig | None = None, |
| 103 | bias: bool = False, |
| 104 | reduce_results: bool = True, |
| 105 | is_sequence_parallel=False, |
| 106 | prefix: str = "", |
| 107 | ) -> None: |
| 108 | super().__init__() |
| 109 | self.gate_up_proj = MergedColumnParallelLinear( |
| 110 | hidden_size, |
| 111 | [intermediate_size] * 2, |
| 112 | bias=bias, |
| 113 | quant_config=quant_config, |
| 114 | disable_tp=is_sequence_parallel, |
| 115 | prefix=f"{prefix}.gate_up_proj", |
| 116 | ) |
| 117 | self.down_proj = RowParallelLinear( |
| 118 | intermediate_size, |
| 119 | hidden_size, |
| 120 | bias=bias, |
| 121 | quant_config=quant_config, |
| 122 | reduce_results=reduce_results, |
| 123 | disable_tp=is_sequence_parallel, |
| 124 | prefix=f"{prefix}.down_proj", |
| 125 | ) |
| 126 | |
| 127 | check_ffn_act_fn(hidden_act) |
| 128 | self.act_fn = SiluAndMul() |
| 129 | |
| 130 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 131 | return self.down_proj(self.act_fn(self.gate_up_proj(x)[0]))[0] |
| 132 | |
| 133 | |
| 134 | class OpenPanguMoE(nn.Module): |
| 135 | def __init__( |
| 136 | self, |
| 137 | config: PreTrainedConfig, |
| 138 | parallel_config: ParallelConfig, |
| 139 | quant_config: QuantizationConfig | None = None, |
| 140 | prefix: str = "", |
| 141 | ): |
| 142 | super().__init__() |
| 143 | self.tp_size = get_tensor_model_parallel_world_size() |
| 144 | self.tp_rank = get_tp_group().rank_in_group |
| 145 | |
| 146 | self.routed_scaling_factor = config.routed_scaling_factor |
| 147 | self.ep_group = get_ep_group().device_group |
| 148 | self.ep_size = self.ep_group.size() |
| 149 | self.n_routed_experts: int = config.n_routed_experts |
| 150 | self.n_shared_experts: int = config.n_shared_experts |
| 151 | |
| 152 | self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe |
| 153 | check_ffn_act_fn(config.hidden_act) |
| 154 | |
| 155 | self.gate = GateLinear( |
| 156 | config.hidden_size, |
| 157 | config.n_routed_experts, |
| 158 | prefix=f"{prefix}.gate", |
| 159 | ) |
| 160 | if ( |
| 161 | hasattr(config, "router_enable_expert_bias") |
| 162 | and config.router_enable_expert_bias |
| 163 | ): |
| 164 | self.gate.e_score_correction_bias = nn.Parameter( |
| 165 | torch.empty(self.n_routed_experts, dtype=torch.float32) |
| 166 | ) |
| 167 | else: |
| 168 | self.gate.e_score_correction_bias = None |
| 169 | |
| 170 | # Load balancing settings. |
| 171 | eplb_config = parallel_config.eplb_config |
| 172 | self.enable_eplb = parallel_config.enable_eplb |
| 173 | |
| 174 | self.n_redundant_experts = eplb_config.num_redundant_experts |
| 175 | self.n_logical_experts = self.n_routed_experts |
| 176 | self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts |
| 177 | self.n_local_physical_experts = self.n_physical_experts // self.ep_size |
| 178 | |
| 179 | self.shared_experts: OpenPanguMLP | None |
| 180 | if config.n_shared_experts is not None: |
| 181 | intermediate_size = config.moe_intermediate_size * config.n_shared_experts |
| 182 | self.shared_experts = OpenPanguMLP( |
| 183 | hidden_size=config.hidden_size, |
| 184 | intermediate_size=intermediate_size, |
| 185 | hidden_act=config.hidden_act, |
| 186 | quant_config=quant_config, |
| 187 | is_sequence_parallel=self.is_sequence_parallel, |
| 188 | reduce_results=False, |
| 189 | prefix=f"{prefix}.shared_experts", |
| 190 | ) |
| 191 | else: |
| 192 | self.shared_experts = None |
| 193 | |
| 194 | self.experts = FusedMoEFactory( |
| 195 | shared_experts=self.shared_experts, |
| 196 | num_experts=config.n_routed_experts, |
| 197 | top_k=config.num_experts_per_tok, |
| 198 | hidden_size=config.hidden_size, |
| 199 | intermediate_size=config.moe_intermediate_size, |
| 200 | renormalize=config.norm_topk_prob, |
| 201 | quant_config=quant_config, |
| 202 | use_grouped_topk=True, |
| 203 | num_expert_group=1, |
| 204 | topk_group=1, |
| 205 | prefix=f"{prefix}.experts", |
| 206 | scoring_func="sigmoid", |
| 207 | routed_scaling_factor=self.routed_scaling_factor, |
| 208 | apply_routed_scale_to_output=True, |
| 209 | e_score_correction_bias=self.gate.e_score_correction_bias, |
| 210 | enable_eplb=self.enable_eplb, |
| 211 | num_redundant_experts=self.n_redundant_experts, |
| 212 | is_sequence_parallel=self.is_sequence_parallel, |
| 213 | ) |
| 214 | |
| 215 | def forward( |
| 216 | self, |
| 217 | hidden_states: torch.Tensor, |
| 218 | ) -> torch.Tensor: |
| 219 | num_tokens, hidden_dim = hidden_states.shape |
| 220 | hidden_states = hidden_states.view(-1, hidden_dim) |
| 221 | |
| 222 | if self.is_sequence_parallel: |
| 223 | hidden_states = sequence_parallel_chunk(hidden_states) |
| 224 | |
| 225 | router_logits, _ = self.gate(hidden_states) |
| 226 | |
| 227 | final_hidden_states = self.experts( |
| 228 | hidden_states=hidden_states, router_logits=router_logits |
| 229 | ) |
| 230 | |
| 231 | if self.is_sequence_parallel: |
| 232 | final_hidden_states = tensor_model_parallel_all_gather( |
| 233 | final_hidden_states, 0 |
| 234 | ) |
| 235 | final_hidden_states = final_hidden_states[:num_tokens] |
| 236 | |
| 237 | return final_hidden_states.view(num_tokens, hidden_dim) |
| 238 | |
| 239 | |
| 240 | class OpenPanguMLAAttention(nn.Module): |
| 241 | def __init__( |
| 242 | self, |
| 243 | config: PreTrainedConfig, |
| 244 | hidden_size: int, |
| 245 | num_heads: int, |
| 246 | qk_nope_head_dim: int, |
| 247 | qk_rope_head_dim: int, |
| 248 | v_head_dim: int, |
| 249 | q_lora_rank: int | None, |
| 250 | kv_lora_rank: int, |
| 251 | max_position_embeddings: int = 8192, |
| 252 | cache_config: CacheConfig | None = None, |
| 253 | quant_config: QuantizationConfig | None = None, |
| 254 | prefix: str = "", |
| 255 | ) -> None: |
| 256 | super().__init__() |
| 257 | self.hidden_size = hidden_size |
| 258 | self.num_heads = num_heads |
| 259 | self.qk_nope_head_dim = qk_nope_head_dim |
| 260 | self.qk_rope_head_dim = qk_rope_head_dim |
| 261 | self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim |
| 262 | self.v_head_dim = v_head_dim |
| 263 | self.q_lora_rank = q_lora_rank |
| 264 | self.kv_lora_rank = kv_lora_rank |
| 265 | self.tp_size = get_tensor_model_parallel_world_size() |
| 266 | if num_heads % self.tp_size != 0: |
| 267 | raise ValueError( |
| 268 | f"num_heads {num_heads} is not divisible by tp_size {self.tp_size}." |
| 269 | ) |
| 270 | self.num_local_heads = num_heads // self.tp_size |
| 271 | |
| 272 | self.scaling = self.qk_head_dim**-0.5 |
| 273 | self.max_position_embeddings = max_position_embeddings |
| 274 | |
| 275 | self.prefix = prefix |
| 276 | |
| 277 | if self.q_lora_rank is not None: |
| 278 | self.fused_qkv_a_proj = MergedColumnParallelLinear( |
| 279 | self.hidden_size, |
| 280 | [self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim], |
| 281 | bias=False, |
| 282 | quant_config=quant_config, |
| 283 | prefix=f"{prefix}.fused_qkv_a_proj", |
| 284 | disable_tp=True, |
| 285 | ) |
| 286 | self.q_a_layernorm = RMSNorm(self.q_lora_rank, eps=config.rms_norm_eps) |
| 287 | self.q_b_proj = ColumnParallelLinear( |
| 288 | self.q_lora_rank, |
| 289 | self.num_heads * self.qk_head_dim, |
| 290 | bias=False, |
| 291 | quant_config=quant_config, |
| 292 | prefix=f"{prefix}.q_b_proj", |
| 293 | ) |
| 294 | else: |
| 295 | self.q_proj = ColumnParallelLinear( |
| 296 | self.hidden_size, |
| 297 | self.num_heads * self.qk_head_dim, |
| 298 | bias=False, |
| 299 | quant_config=quant_config, |
| 300 | prefix=f"{prefix}.q_proj", |
| 301 | ) |
| 302 | self.kv_a_proj_with_mqa = ReplicatedLinear( |
| 303 | self.hidden_size, |
| 304 | self.kv_lora_rank + self.qk_rope_head_dim, |
| 305 | bias=False, |
| 306 | quant_config=quant_config, |
| 307 | prefix=f"{prefix}.kv_a_proj_with_mqa", |
| 308 | ) |
| 309 | |
| 310 | self.kv_a_layernorm = RMSNorm(self.kv_lora_rank, eps=config.rms_norm_eps) |
| 311 | self.kv_b_proj = ColumnParallelLinear( |
| 312 | self.kv_lora_rank, |
| 313 | self.num_heads * (self.qk_nope_head_dim + self.v_head_dim), |
| 314 | bias=False, |
| 315 | quant_config=quant_config, |
| 316 | prefix=f"{prefix}.kv_b_proj", |
| 317 | ) |
| 318 | |
| 319 | self.o_proj = RowParallelLinear( |
| 320 | self.num_heads * self.v_head_dim, |
| 321 | self.hidden_size, |
| 322 | bias=False, |
| 323 | quant_config=quant_config, |
| 324 | prefix=f"{prefix}.o_proj", |
| 325 | ) |
| 326 | |
| 327 | # TODO: remove hard coding |
| 328 | set_default_rope_theta(config, default_theta=10000) |
| 329 | rope_parameters = { |
| 330 | "rope_theta": config.rope_parameters["rope_theta"], |
| 331 | "beta_fast": 32, |
| 332 | "beta_slow": 1, |
| 333 | "factor": 1, |
| 334 | "mscale": 1.0, |
| 335 | "mscale_all_dim": 1.0, |
| 336 | "original_max_position_embeddings": max_position_embeddings, |
| 337 | "type": "yarn", |
| 338 | "rope_type": "deepseek_yarn", |
| 339 | } |
| 340 | self.rotary_emb = get_rope( |
| 341 | qk_rope_head_dim, |
| 342 | max_position=max_position_embeddings, |
| 343 | rope_parameters=rope_parameters, |
| 344 | is_neox_style=False, |
| 345 | ) |
| 346 | |
| 347 | mla_modules = MLAModules( |
| 348 | kv_a_layernorm=self.kv_a_layernorm, |
| 349 | kv_b_proj=self.kv_b_proj, |
| 350 | rotary_emb=self.rotary_emb, |
| 351 | o_proj=self.o_proj, |
| 352 | fused_qkv_a_proj=self.fused_qkv_a_proj |
| 353 | if self.q_lora_rank is not None |
| 354 | else None, |
| 355 | kv_a_proj_with_mqa=self.kv_a_proj_with_mqa |
| 356 | if self.q_lora_rank is None |
| 357 | else None, |
| 358 | q_a_layernorm=self.q_a_layernorm if self.q_lora_rank is not None else None, |
| 359 | q_b_proj=self.q_b_proj if self.q_lora_rank is not None else None, |
| 360 | q_proj=self.q_proj if self.q_lora_rank is None else None, |
| 361 | indexer=None, |
| 362 | is_sparse=False, |
| 363 | topk_indices_buffer=None, |
| 364 | ) |
| 365 | |
| 366 | self.mla_attn = MultiHeadLatentAttentionWrapper( |
| 367 | self.hidden_size, |
| 368 | self.num_local_heads, |
| 369 | self.scaling, |
| 370 | self.qk_nope_head_dim, |
| 371 | self.qk_rope_head_dim, |
| 372 | self.v_head_dim, |
| 373 | self.q_lora_rank, |
| 374 | self.kv_lora_rank, |
| 375 | mla_modules, |
| 376 | cache_config, |
| 377 | quant_config, |
| 378 | prefix, |
| 379 | ) |
| 380 | |
| 381 | def forward( |
| 382 | self, |
| 383 | positions: torch.Tensor, |
| 384 | hidden_states: torch.Tensor, |
| 385 | ) -> torch.Tensor: |
| 386 | return self.mla_attn(positions, hidden_states) |
| 387 | |
| 388 | |
| 389 | class OpenPanguEmbeddedAttention(nn.Module): |
| 390 | def __init__( |
| 391 | self, |
| 392 | config: PreTrainedConfig, |
| 393 | hidden_size: int, |
| 394 | num_heads: int, |
| 395 | num_kv_heads: int, |
| 396 | max_position_embeddings: int = 8192, |
| 397 | quant_config: QuantizationConfig | None = None, |
| 398 | bias: bool = False, |
| 399 | bias_o_proj: bool = False, |
| 400 | cache_config: CacheConfig | None = None, |
| 401 | prefix: str = "", |
| 402 | attn_type: str = AttentionType.DECODER, |
| 403 | ) -> None: |
| 404 | super().__init__() |
| 405 | layer_idx = extract_layer_index(prefix) |
| 406 | self.hidden_size = hidden_size |
| 407 | tp_size = get_tensor_model_parallel_world_size() |
| 408 | self.total_num_heads = num_heads |
| 409 | if self.total_num_heads % tp_size != 0: |
| 410 | raise ValueError( |
| 411 | f"total_num_heads {self.total_num_heads} " |
| 412 | f"is not divisible by tp_size {tp_size}." |
| 413 | ) |
| 414 | self.num_heads = self.total_num_heads // tp_size |
| 415 | self.total_num_kv_heads = num_kv_heads |
| 416 | if self.total_num_kv_heads > tp_size and self.total_num_kv_heads % tp_size != 0: |
| 417 | # Number of KV heads is greater than TP size, so we partition |
| 418 | # the KV heads across multiple tensor parallel ranks. |
| 419 | raise ValueError( |
| 420 | "Number of KV heads is greater than TP size, " |
| 421 | f"but total_num_kv_heads {self.total_num_kv_heads} " |
| 422 | f"is not divisible by tp_size {tp_size}." |
| 423 | ) |
| 424 | elif ( |
| 425 | self.total_num_kv_heads < tp_size and tp_size % self.total_num_kv_heads != 0 |
| 426 | ): |
| 427 | # Number of KV heads is less than TP size, so we replicate |
| 428 | # the KV heads across multiple tensor parallel ranks. |
| 429 | raise ValueError( |
| 430 | f"Number of KV heads is less than TP size, but tp_size {tp_size} " |
| 431 | f"is not divisible by total_num_kv_heads {self.total_num_kv_heads}." |
| 432 | ) |
| 433 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 434 | head_dim = getattr(config, "head_dim", None) |
| 435 | if head_dim is None: |
| 436 | head_dim = self.hidden_size // self.total_num_heads |
| 437 | self.head_dim = head_dim |
| 438 | self.q_size = self.num_heads * self.head_dim |
| 439 | self.kv_size = self.num_kv_heads * self.head_dim |
| 440 | self.scaling = self.head_dim**-0.5 |
| 441 | self.max_position_embeddings = max_position_embeddings |
| 442 | |
| 443 | self.qkv_proj = QKVParallelLinear( |
| 444 | hidden_size=hidden_size, |
| 445 | head_size=self.head_dim, |
| 446 | total_num_heads=self.total_num_heads, |
| 447 | total_num_kv_heads=self.total_num_kv_heads, |
| 448 | bias=bias, |
| 449 | quant_config=quant_config, |
| 450 | prefix=f"{prefix}.qkv_proj", |
| 451 | ) |
| 452 | |
| 453 | self.o_proj = RowParallelLinear( |
| 454 | input_size=self.total_num_heads * self.head_dim, |
| 455 | output_size=hidden_size, |
| 456 | bias=bias_o_proj, |
| 457 | quant_config=quant_config, |
| 458 | prefix=f"{prefix}.o_proj", |
| 459 | ) |
| 460 | |
| 461 | self._init_rotary_emb(config, quant_config=quant_config) |
| 462 | |
| 463 | if hasattr(config, "interleaved_sliding_window"): |
| 464 | interleaved_sliding_window = config.interleaved_sliding_window |
| 465 | if isinstance(interleaved_sliding_window, int): |
| 466 | sliding_window = interleaved_sliding_window |
| 467 | elif isinstance(interleaved_sliding_window, list): |
| 468 | sw_idx = layer_idx % len(interleaved_sliding_window) |
| 469 | sliding_window = interleaved_sliding_window[sw_idx] |
| 470 | else: |
| 471 | raise ValueError( |
| 472 | f"{type(interleaved_sliding_window)} " |
| 473 | "for interleaved_sliding_window is not supported." |
| 474 | ) |
| 475 | else: |
| 476 | sliding_window = None |
| 477 | |
| 478 | self.attn = Attention( |
| 479 | self.num_heads, |
| 480 | self.head_dim, |
| 481 | self.scaling, |
| 482 | num_kv_heads=self.num_kv_heads, |
| 483 | cache_config=cache_config, |
| 484 | quant_config=quant_config, |
| 485 | per_layer_sliding_window=sliding_window, |
| 486 | attn_type=attn_type, |
| 487 | prefix=f"{prefix}.attn", |
| 488 | ) |
| 489 | |
| 490 | def forward( |
| 491 | self, |
| 492 | positions: torch.Tensor, |
| 493 | hidden_states: torch.Tensor, |
| 494 | ) -> torch.Tensor: |
| 495 | qkv, _ = self.qkv_proj(hidden_states) |
| 496 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 497 | q, k = self.rotary_emb(positions, q, k) |
| 498 | attn_output = self.attn(q, k, v) |
| 499 | output, _ = self.o_proj(attn_output) |
| 500 | return output |
| 501 | |
| 502 | def _init_rotary_emb( |
| 503 | self, |
| 504 | config: PreTrainedConfig, |
| 505 | quant_config: QuantizationConfig | None, |
| 506 | ) -> None: |
| 507 | is_neox_style = True |
| 508 | rope_parameters = config.rope_parameters or {} |
| 509 | if rope_parameters is not None and rope_parameters.get( |
| 510 | "mrope_interleaved", False |
| 511 | ): |
| 512 | rope_parameters["rope_type"] = "openpangu" |
| 513 | |
| 514 | self.rotary_emb = get_rope( |
| 515 | self.head_dim, |
| 516 | max_position=self.max_position_embeddings, |
| 517 | rope_parameters=rope_parameters, |
| 518 | is_neox_style=is_neox_style, |
| 519 | ) |
| 520 | |
| 521 | |
| 522 | class OpenPanguSinkAttention(nn.Module): |
| 523 | def __init__( |
| 524 | self, |
| 525 | config: PreTrainedConfig, |
| 526 | hidden_size: int, |
| 527 | num_heads: int, |
| 528 | num_kv_heads: int, |
| 529 | rope_parameters: dict[str, Any] | None = None, |
| 530 | max_position_embeddings: int = 8192, |
| 531 | quant_config: QuantizationConfig | None = None, |
| 532 | bias: bool = False, |
| 533 | bias_o_proj: bool = False, |
| 534 | cache_config: CacheConfig | None = None, |
| 535 | prefix: str = "", |
| 536 | attn_type: str = AttentionType.DECODER, |
| 537 | ) -> None: |
| 538 | super().__init__() |
| 539 | layer_idx = extract_layer_index(prefix) |
| 540 | self.hidden_size = hidden_size |
| 541 | self.tp_size = get_tensor_model_parallel_world_size() |
| 542 | self.tp_rank = get_tensor_model_parallel_rank() |
| 543 | self.total_num_heads = num_heads |
| 544 | if self.total_num_heads % self.tp_size != 0: |
| 545 | raise ValueError( |
| 546 | f"total_num_heads {self.total_num_heads} " |
| 547 | f"is not divisible by tp_size {self.tp_size}." |
| 548 | ) |
| 549 | self.num_heads = self.total_num_heads // self.tp_size |
| 550 | self.total_num_kv_heads = num_kv_heads |
| 551 | if ( |
| 552 | self.total_num_kv_heads > self.tp_size |
| 553 | and self.total_num_kv_heads % self.tp_size != 0 |
| 554 | ): |
| 555 | # Number of KV heads is greater than TP size, so we partition |
| 556 | # the KV heads across multiple tensor parallel ranks. |
| 557 | raise ValueError( |
| 558 | "Number of KV heads is greater than TP size, " |
| 559 | f"but total_num_kv_heads {self.total_num_kv_heads} " |
| 560 | f"is not divisible by tp_size {self.tp_size}." |
| 561 | ) |
| 562 | elif self.total_num_kv_heads < self.tp_size: |
| 563 | # TODO: Number of KV heads is less than TP size, so we replicate |
| 564 | # the KV heads across multiple tensor parallel ranks. |
| 565 | raise ValueError( |
| 566 | f"Number of KV heads {self.total_num_kv_heads} is less than " |
| 567 | f"TP size {self.tp_size}, KV heads replication is not support yet." |
| 568 | ) |
| 569 | self.num_kv_heads = max(1, self.total_num_kv_heads // self.tp_size) |
| 570 | self.qk_nope_dim = getattr(config, "qk_nope_dim", None) |
| 571 | self.qk_rope_dim = getattr(config, "qk_rope_dim", None) |
| 572 | self.v_channels = getattr(config, "v_channels", None) |
| 573 | assert self.qk_nope_dim is not None |
| 574 | assert self.qk_rope_dim is not None |
| 575 | assert self.v_channels is not None |
| 576 | self.head_dim = self.qk_rope_dim + self.qk_nope_dim |
| 577 | self.q_size = self.num_heads * self.head_dim |
| 578 | self.k_size = self.num_kv_heads * self.head_dim |
| 579 | self.v_size = self.num_kv_heads * self.v_channels |
| 580 | self.scaling = self.head_dim**-0.5 |
| 581 | self.max_position_embeddings = max_position_embeddings |
| 582 | |
| 583 | self.param_sink_number = getattr(config, "param_sink_number", 0) |
| 584 | self.param_sink_with_value = getattr(config, "param_sink_with_value", False) |
| 585 | self.param_sink_scalar = getattr(config, "param_sink_scalar", None) |
| 586 | self.param_sink_of_head_num = getattr(config, "param_sink_of_head_dim", False) |
| 587 | |
| 588 | self.qkv_proj = MergedColumnParallelLinear( |
| 589 | input_size=hidden_size, |
| 590 | output_sizes=[ |
| 591 | self.q_size * self.tp_size, |
| 592 | self.k_size * self.tp_size, |
| 593 | self.v_size * self.tp_size, |
| 594 | ], |
| 595 | bias=bias, |
| 596 | quant_config=quant_config, |
| 597 | prefix=f"{prefix}.qkv_proj", |
| 598 | ) |
| 599 | |
| 600 | self.o_proj = RowParallelLinear( |
| 601 | input_size=self.total_num_heads * self.v_channels, |
| 602 | output_size=hidden_size, |
| 603 | bias=bias_o_proj, |
| 604 | quant_config=quant_config, |
| 605 | prefix=f"{prefix}.o_proj", |
| 606 | ) |
| 607 | |
| 608 | self.k_layernorm = RMSNorm(self.head_dim, eps=config.rms_norm_eps) |
| 609 | |
| 610 | self._init_rotary_emb( |
| 611 | config, rope_parameters=rope_parameters, quant_config=quant_config |
| 612 | ) |
| 613 | |
| 614 | if hasattr(config, "interleaved_sliding_window"): |
| 615 | interleaved_sliding_window = config.interleaved_sliding_window |
| 616 | if isinstance(interleaved_sliding_window, int): |
| 617 | sliding_window = interleaved_sliding_window |
| 618 | elif isinstance(interleaved_sliding_window, list): |
| 619 | sw_idx = layer_idx % len(interleaved_sliding_window) |
| 620 | sliding_window = interleaved_sliding_window[sw_idx] |
| 621 | else: |
| 622 | raise ValueError( |
| 623 | f"{type(interleaved_sliding_window)} " |
| 624 | "for interleaved_sliding_window is not supported." |
| 625 | ) |
| 626 | else: |
| 627 | sliding_window = None |
| 628 | |
| 629 | FlashAttentionDiffKVBackend.set_head_size_v(self.v_channels) |
| 630 | self.attn = StaticSinkAttention( |
| 631 | self.num_heads, |
| 632 | self.head_dim, |
| 633 | self.scaling, |
| 634 | sink_len=self.param_sink_number, |
| 635 | num_kv_heads=self.num_kv_heads, |
| 636 | cache_config=cache_config, |
| 637 | quant_config=quant_config, |
| 638 | per_layer_sliding_window=sliding_window, |
| 639 | attn_type=attn_type, |
| 640 | prefix=f"{prefix}.attn", |
| 641 | attn_backend=FlashAttentionDiffKVBackend, |
| 642 | head_size_v=self.v_channels, |
| 643 | ) |
| 644 | |
| 645 | if self.param_sink_number > 0: |
| 646 | self.param_sink_key = torch.nn.Parameter( |
| 647 | torch.empty( |
| 648 | ( |
| 649 | self.param_sink_number, |
| 650 | self.num_kv_heads, |
| 651 | self.head_dim, |
| 652 | ), |
| 653 | device=current_platform.current_device(), |
| 654 | dtype=config.torch_dtype, |
| 655 | ) |
| 656 | ) |
| 657 | set_weight_attrs( |
| 658 | self.param_sink_key, |
| 659 | { |
| 660 | "output_dim": 1, |
| 661 | "weight_loader": self.weight_loader, |
| 662 | }, |
| 663 | ) |
| 664 | |
| 665 | if self.param_sink_with_value: |
| 666 | self.param_sink_value = torch.nn.Parameter( |
| 667 | torch.empty( |
| 668 | ( |
| 669 | self.param_sink_number, |
| 670 | self.num_kv_heads, |
| 671 | self.v_channels, |
| 672 | ), |
| 673 | device=current_platform.current_device(), |
| 674 | dtype=config.torch_dtype, |
| 675 | ) |
| 676 | ) |
| 677 | set_weight_attrs( |
| 678 | self.param_sink_value, |
| 679 | { |
| 680 | "output_dim": 1, |
| 681 | "weight_loader": self.weight_loader, |
| 682 | }, |
| 683 | ) |
| 684 | else: |
| 685 | self.register_buffer( |
| 686 | "param_sink_value", |
| 687 | torch.zeros( |
| 688 | ( |
| 689 | self.param_sink_number, |
| 690 | self.num_kv_heads, |
| 691 | self.v_channels, |
| 692 | ), |
| 693 | device=current_platform.current_device(), |
| 694 | dtype=config.torch_dtype, |
| 695 | ), |
| 696 | persistent=False, |
| 697 | ) |
| 698 | # To enable dummy run with out weight |
| 699 | self.post_weight_load() |
| 700 | |
| 701 | def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor): |
| 702 | output_dim = getattr(param, "output_dim", None) |
| 703 | |
| 704 | is_sharded_weight = getattr(param, "is_sharded_weight", False) |
| 705 | |
| 706 | param_data = param.data |
| 707 | if output_dim is not None and not is_sharded_weight: |
| 708 | shard_size = param_data.shape[output_dim] |
| 709 | start_idx = self.tp_rank * shard_size |
| 710 | loaded_weight = loaded_weight.narrow(output_dim, start_idx, shard_size) |
| 711 | |
| 712 | # Special case for loading scales off disk, which often do not |
| 713 | # have a shape (such as in the case of AutoFP8). |
| 714 | if len(loaded_weight.shape) == 0: |
| 715 | loaded_weight = loaded_weight.reshape(1) |
| 716 | |
| 717 | assert param_data.shape == loaded_weight.shape |
| 718 | param_data.copy_(loaded_weight) |
| 719 | |
| 720 | def forward( |
| 721 | self, |
| 722 | positions: torch.Tensor, |
| 723 | hidden_states: torch.Tensor, |
| 724 | ) -> torch.Tensor: |
| 725 | qkv, _ = self.qkv_proj(hidden_states) |
| 726 | q, k, v = qkv.split([self.q_size, self.k_size, self.v_size], dim=-1) |
| 727 | k = self.k_layernorm(k.view(-1, self.num_kv_heads, self.head_dim)) |
| 728 | q, k = self.rotary_emb(positions, q, k) |
| 729 | |
| 730 | q = q.view(-1, self.q_size) |
| 731 | k = k.view(-1, self.k_size) |
| 732 | |
| 733 | attn_output = self.attn( |
| 734 | q, |
| 735 | k, |
| 736 | v, |
| 737 | output_shape=torch.Size( |
| 738 | [q.shape[0], q.shape[1] // self.head_dim * self.v_channels] |
| 739 | ), |
| 740 | ) |
| 741 | output, _ = self.o_proj(attn_output) |
| 742 | return output |
| 743 | |
| 744 | def _init_rotary_emb( |
| 745 | self, |
| 746 | config: PreTrainedConfig, |
| 747 | rope_parameters: dict[str, Any] | None, |
| 748 | quant_config: QuantizationConfig | None, |
| 749 | ) -> None: |
| 750 | is_neox_style = False |
| 751 | rope_parameters = {"partial_rotary_factor": self.qk_rope_dim / self.head_dim} |
| 752 | |
| 753 | self.rotary_emb = get_rope( |
| 754 | self.head_dim, |
| 755 | max_position=self.max_position_embeddings, |
| 756 | rope_parameters=rope_parameters, |
| 757 | is_neox_style=is_neox_style, |
| 758 | ) |
| 759 | |
| 760 | def post_weight_load(self) -> None: |
| 761 | if hasattr(self, "k_layernorm") and self.k_layernorm is not None: |
| 762 | param_sink_key = self.k_layernorm(self.param_sink_key) |
| 763 | else: |
| 764 | param_sink_key = self.param_sink_key |
| 765 | |
| 766 | self.attn.update_sink_kv(param_sink_key, self.param_sink_value) |
| 767 | |
| 768 | |
| 769 | class OpenPanguDecoderLayer(nn.Module): |
| 770 | def __init__( |
| 771 | self, |
| 772 | config: PreTrainedConfig, |
| 773 | prefix: str, |
| 774 | vllm_config: VllmConfig, |
| 775 | ) -> None: |
| 776 | super().__init__() |
| 777 | |
| 778 | if config is None: |
| 779 | config = vllm_config.model_config.hf_config |
| 780 | cache_config = vllm_config.cache_config |
| 781 | quant_config = vllm_config.quant_config |
| 782 | parallel_config = vllm_config.parallel_config |
| 783 | |
| 784 | self.hidden_size = config.hidden_size |
| 785 | max_position_embeddings = getattr(config, "max_position_embeddings", 8192) |
| 786 | |
| 787 | layer_idx = int(prefix.split(sep=".")[-1]) |
| 788 | self.layer_idx = layer_idx |
| 789 | |
| 790 | self.use_mla = ( |
| 791 | hasattr(config, "qk_nope_head_dim") |
| 792 | and hasattr(config, "qk_rope_head_dim") |
| 793 | and hasattr(config, "v_head_dim") |
| 794 | and hasattr(config, "kv_lora_rank") |
| 795 | ) |
| 796 | self.use_sink_attention = ( |
| 797 | hasattr(config, "param_sink_number") and config.param_sink_number > 0 |
| 798 | ) |
| 799 | if self.use_mla: |
| 800 | self.self_attn = OpenPanguMLAAttention( |
| 801 | config=config, |
| 802 | hidden_size=self.hidden_size, |
| 803 | num_heads=config.num_attention_heads, |
| 804 | qk_nope_head_dim=config.qk_nope_head_dim, |
| 805 | qk_rope_head_dim=config.qk_rope_head_dim, |
| 806 | v_head_dim=config.v_head_dim, |
| 807 | q_lora_rank=( |
| 808 | config.q_lora_rank if hasattr(config, "q_lora_rank") else None |
| 809 | ), |
| 810 | kv_lora_rank=config.kv_lora_rank, |
| 811 | max_position_embeddings=max_position_embeddings, |
| 812 | cache_config=cache_config, |
| 813 | quant_config=quant_config, |
| 814 | prefix=f"{prefix}.self_attn", |
| 815 | ) |
| 816 | elif self.use_sink_attention: |
| 817 | attention_bias = getattr(config, "attention_bias", False) or getattr( |
| 818 | config, "bias", False |
| 819 | ) |
| 820 | bias_o_proj = attention_bias |
| 821 | if hasattr(config, "qkv_bias"): |
| 822 | attention_bias = config.qkv_bias |
| 823 | if getattr(config, "is_causal", True): |
| 824 | attn_type = AttentionType.DECODER |
| 825 | else: |
| 826 | raise ValueError( |
| 827 | f"is_causal={config.is_causal} is not support " |
| 828 | "for attention with sink" |
| 829 | ) |
| 830 | rope_parameters = getattr(config, "rope_scaling", None) |
| 831 | if rope_parameters is None: |
| 832 | rope_parameters = { |
| 833 | "rope_type": "default", |
| 834 | "rope_theta": config.rope_theta, |
| 835 | } |
| 836 | self.self_attn = OpenPanguSinkAttention( |
| 837 | config=config, |
| 838 | hidden_size=self.hidden_size, |
| 839 | num_heads=config.num_attention_heads, |
| 840 | num_kv_heads=getattr( |
| 841 | config, "num_key_value_heads", config.num_attention_heads |
| 842 | ), |
| 843 | rope_parameters=rope_parameters, |
| 844 | max_position_embeddings=max_position_embeddings, |
| 845 | quant_config=quant_config, |
| 846 | bias=attention_bias, |
| 847 | bias_o_proj=bias_o_proj, |
| 848 | cache_config=cache_config, |
| 849 | prefix=f"{prefix}.self_attn", |
| 850 | attn_type=attn_type, |
| 851 | ) |
| 852 | else: |
| 853 | attention_bias = getattr(config, "attention_bias", False) or getattr( |
| 854 | config, "bias", False |
| 855 | ) |
| 856 | bias_o_proj = attention_bias |
| 857 | if hasattr(config, "qkv_bias"): |
| 858 | attention_bias = config.qkv_bias |
| 859 | # By default, PanguEmbedded uses causal attention |
| 860 | # as it is a decoder-only model. |
| 861 | # You can override the HF config with `is_causal=False` to enable |
| 862 | # bidirectional attention, which is used in some embedding models |
| 863 | if getattr(config, "is_causal", True): |
| 864 | attn_type = AttentionType.DECODER |
| 865 | else: |
| 866 | attn_type = AttentionType.ENCODER_ONLY |
| 867 | self.self_attn = OpenPanguEmbeddedAttention( |
| 868 | config=config, |
| 869 | hidden_size=self.hidden_size, |
| 870 | num_heads=config.num_attention_heads, |
| 871 | num_kv_heads=getattr( |
| 872 | config, "num_key_value_heads", config.num_attention_heads |
| 873 | ), |
| 874 | max_position_embeddings=max_position_embeddings, |
| 875 | quant_config=quant_config, |
| 876 | bias=attention_bias, |
| 877 | bias_o_proj=bias_o_proj, |
| 878 | cache_config=cache_config, |
| 879 | prefix=f"{prefix}.self_attn", |
| 880 | attn_type=attn_type, |
| 881 | ) |
| 882 | |
| 883 | if ( |
| 884 | getattr(config, "n_routed_experts", None) is not None |
| 885 | and layer_idx >= config.first_k_dense_replace |
| 886 | ): |
| 887 | self.mlp = OpenPanguMoE( |
| 888 | config=config, |
| 889 | parallel_config=parallel_config, |
| 890 | quant_config=quant_config, |
| 891 | prefix=f"{prefix}.mlp", |
| 892 | ) |
| 893 | else: |
| 894 | self.mlp = OpenPanguMLP( |
| 895 | hidden_size=self.hidden_size, |
| 896 | intermediate_size=config.intermediate_size, |
| 897 | hidden_act=config.hidden_act, |
| 898 | quant_config=quant_config, |
| 899 | bias=getattr(config, "mlp_bias", False), |
| 900 | prefix=f"{prefix}.mlp", |
| 901 | ) |
| 902 | self.routed_scaling_factor = getattr(config, "routed_scaling_factor", 1.0) |
| 903 | self.num_hidden_layers = config.num_hidden_layers |
| 904 | self.first_k_dense_replace = getattr( |
| 905 | config, "first_k_dense_replace", self.num_hidden_layers |
| 906 | ) |
| 907 | |
| 908 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 909 | self.post_attention_layernorm = RMSNorm( |
| 910 | config.hidden_size, eps=config.rms_norm_eps |
| 911 | ) |
| 912 | self.tp_group = get_tp_group().device_group |
| 913 | self.sandwich_norm = getattr(config, "sandwich_norm", False) |
| 914 | if self.sandwich_norm: |
| 915 | self.pre_mlp_layernorm = RMSNorm( |
| 916 | config.hidden_size, eps=config.rms_norm_eps |
| 917 | ) |
| 918 | self.post_mlp_layernorm = RMSNorm( |
| 919 | config.hidden_size, eps=config.rms_norm_eps |
| 920 | ) |
| 921 | |
| 922 | def forward( |
| 923 | self, |
| 924 | positions: torch.Tensor, |
| 925 | hidden_states: torch.Tensor, |
| 926 | residual: torch.Tensor | None, |
| 927 | ) -> torch.Tensor: |
| 928 | if residual is None: |
| 929 | residual = hidden_states |
| 930 | hidden_states = self.input_layernorm(hidden_states) |
| 931 | else: |
| 932 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 933 | |
| 934 | hidden_states = self.self_attn( |
| 935 | positions=positions, |
| 936 | hidden_states=hidden_states, |
| 937 | ) |
| 938 | |
| 939 | if ( |
| 940 | self.routed_scaling_factor is not None |
| 941 | and hidden_states.dtype == torch.float16 |
| 942 | ): |
| 943 | # Fix FP16 overflow |
| 944 | # We scale both hidden_states and residual before |
| 945 | # rmsnorm, and rmsnorm result would not affect by scale. |
| 946 | hidden_states *= 1.0 / self.routed_scaling_factor |
| 947 | if self.layer_idx == 0: |
| 948 | # The residual is shared by all layers, we only scale it on |
| 949 | # first layer. |
| 950 | residual *= 1.0 / self.routed_scaling_factor |
| 951 | |
| 952 | if self.sandwich_norm: |
| 953 | hidden_states = self.post_attention_layernorm(hidden_states) |
| 954 | hidden_states, residual = self.pre_mlp_layernorm(hidden_states, residual) |
| 955 | else: |
| 956 | hidden_states, residual = self.post_attention_layernorm( |
| 957 | hidden_states, residual |
| 958 | ) |
| 959 | |
| 960 | # Fully Connected |
| 961 | hidden_states = self.mlp(hidden_states) |
| 962 | |
| 963 | if ( |
| 964 | self.routed_scaling_factor is not None |
| 965 | and isinstance(self.mlp, OpenPanguMLP) |
| 966 | and hidden_states.dtype == torch.float16 |
| 967 | ): |
| 968 | hidden_states *= 1.0 / self.routed_scaling_factor |
| 969 | |
| 970 | if self.sandwich_norm: |
| 971 | hidden_states = self.post_mlp_layernorm(hidden_states) |
| 972 | |
| 973 | return hidden_states, residual |
| 974 | |
| 975 | |
| 976 | @support_torch_compile |
| 977 | class OpenPanguModel(nn.Module): |
| 978 | fall_back_to_pt_during_load = False |
| 979 | |
| 980 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 981 | super().__init__() |
| 982 | |
| 983 | config = vllm_config.model_config.hf_config |
| 984 | quant_config = vllm_config.quant_config |
| 985 | eplb_config = vllm_config.parallel_config.eplb_config |
| 986 | self.config = config |
| 987 | self.num_redundant_experts = eplb_config.num_redundant_experts |
| 988 | |
| 989 | self.vocab_size = config.vocab_size |
| 990 | |
| 991 | if get_pp_group().is_first_rank or ( |
| 992 | config.tie_word_embeddings and get_pp_group().is_last_rank |
| 993 | ): |
| 994 | self.embed_tokens = VocabParallelEmbedding( |
| 995 | config.vocab_size, |
| 996 | config.hidden_size, |
| 997 | quant_config=quant_config, |
| 998 | prefix=f"{prefix}.embed_tokens", |
| 999 | ) |
| 1000 | else: |
| 1001 | self.embed_tokens = PPMissingLayer() |
| 1002 | |
| 1003 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 1004 | config.num_hidden_layers, |
| 1005 | lambda prefix: OpenPanguDecoderLayer(config, prefix, vllm_config), |
| 1006 | prefix=f"{prefix}.layers", |
| 1007 | ) |
| 1008 | |
| 1009 | if get_pp_group().is_last_rank: |
| 1010 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 1011 | else: |
| 1012 | self.norm = PPMissingLayer() |
| 1013 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 1014 | ["hidden_states", "residual"], config.hidden_size |
| 1015 | ) |
| 1016 | |
| 1017 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1018 | return self.embed_tokens(input_ids) |
| 1019 | |
| 1020 | def forward( |
| 1021 | self, |
| 1022 | input_ids: torch.Tensor | None, |
| 1023 | positions: torch.Tensor, |
| 1024 | intermediate_tensors: IntermediateTensors | None, |
| 1025 | inputs_embeds: torch.Tensor | None = None, |
| 1026 | ) -> torch.Tensor | IntermediateTensors: |
| 1027 | if get_pp_group().is_first_rank: |
| 1028 | if inputs_embeds is not None: |
| 1029 | hidden_states = inputs_embeds |
| 1030 | else: |
| 1031 | hidden_states = self.embed_input_ids(input_ids) |
| 1032 | residual = None |
| 1033 | else: |
| 1034 | assert intermediate_tensors is not None |
| 1035 | hidden_states = intermediate_tensors["hidden_states"] |
| 1036 | residual = intermediate_tensors["residual"] |
| 1037 | |
| 1038 | for i in range(self.start_layer, self.end_layer): |
| 1039 | layer = self.layers[i] |
| 1040 | hidden_states, residual = layer(positions, hidden_states, residual) |
| 1041 | |
| 1042 | if not get_pp_group().is_last_rank: |
| 1043 | return IntermediateTensors( |
| 1044 | {"hidden_states": hidden_states, "residual": residual} |
| 1045 | ) |
| 1046 | |
| 1047 | hidden_states, _ = self.norm(hidden_states, residual) |
| 1048 | return hidden_states |
| 1049 | |
| 1050 | def _filter_spec_layers( |
| 1051 | self, weights: Iterable[tuple[str, torch.Tensor]] |
| 1052 | ) -> Iterable[tuple[str, torch.Tensor]]: |
| 1053 | # Skip the MTP (spec-decode) layers that follow the main model's layers. |
| 1054 | num_mtp = getattr(self.config, "num_nextn_predict_layers", 0) |
| 1055 | for name, loaded_weight in weights: |
| 1056 | if "layers" in name and num_mtp > 0: |
| 1057 | layer_idx = int(name.split("layers.")[-1].split(".")[0]) |
| 1058 | mtp_idx = layer_idx - self.config.num_hidden_layers |
| 1059 | if 0 <= mtp_idx < num_mtp: |
| 1060 | continue |
| 1061 | yield name, loaded_weight |
| 1062 | |
| 1063 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1064 | # .experts.gate_up_proj must be handled by MoERunner.load_weights for EP |
| 1065 | stacked: dict[str, tuple[str, int | str]] = { |
| 1066 | ".mlp.gate_proj": (".mlp.gate_up_proj", 0), |
| 1067 | ".mlp.up_proj": (".mlp.gate_up_proj", 1), |
| 1068 | ".shared_experts.gate_proj": (".shared_experts.gate_up_proj", 0), |
| 1069 | ".shared_experts.up_proj": (".shared_experts.gate_up_proj", 1), |
| 1070 | } |
| 1071 | # The attention layout is config-dependent: MLA with a fused low-rank |
| 1072 | # projection, MLA without it, or standard qkv. Only add the stacked |
| 1073 | # entries whose fused target actually exists as a parameter. |
| 1074 | param_names = list(dict(self.named_parameters())) |
| 1075 | if any(".fused_qkv_a_proj." in n for n in param_names): |
| 1076 | stacked[".q_a_proj"] = (".fused_qkv_a_proj", 0) |
| 1077 | stacked[".kv_a_proj_with_mqa"] = (".fused_qkv_a_proj", 1) |
| 1078 | if any(".qkv_proj." in n for n in param_names): |
| 1079 | stacked[".q_proj"] = (".qkv_proj", "q") |
| 1080 | stacked[".k_proj"] = (".qkv_proj", "k") |
| 1081 | stacked[".v_proj"] = (".qkv_proj", "v") |
| 1082 | mapper = WeightsMapper( |
| 1083 | orig_to_new_substr={ |
| 1084 | "e_score_correction_bias": "gate.e_score_correction_bias", |
| 1085 | }, |
| 1086 | orig_to_new_stacked=stacked, |
| 1087 | ) |
| 1088 | loader = AutoWeightsLoader(self) |
| 1089 | loaded = loader.load_weights(self._filter_spec_layers(weights), mapper=mapper) |
| 1090 | self.post_weight_load() |
| 1091 | return loaded |
| 1092 | |
| 1093 | def post_weight_load(self) -> None: |
| 1094 | for name, module in self.named_modules(): |
| 1095 | if module is self: |
| 1096 | continue |
| 1097 | if hasattr(module, "post_weight_load"): |
| 1098 | module.post_weight_load() |
| 1099 | |
| 1100 | |
| 1101 | class OpenPanguModelBase(nn.Module, SupportsPP, SupportsLoRA): |
| 1102 | packed_modules_mapping = { |
| 1103 | "qkv_proj": ["q_proj", "k_proj", "v_proj"], |
| 1104 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 1105 | } |
| 1106 | |
| 1107 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1108 | super().__init__() |
| 1109 | config = vllm_config.model_config.hf_config |
| 1110 | quant_config = vllm_config.quant_config |
| 1111 | self.config = config |
| 1112 | self.quant_config = quant_config |
| 1113 | |
| 1114 | self.fuse_qkv_a_proj = ( |
| 1115 | hasattr(config, "q_lora_rank") and config.q_lora_rank is not None |
| 1116 | ) |
| 1117 | if self.fuse_qkv_a_proj: |
| 1118 | self.packed_modules_mapping["fused_qkv_a_proj"] = [ |
| 1119 | "q_a_proj", |
| 1120 | "kv_a_proj_with_mqa", |
| 1121 | ] |
| 1122 | |
| 1123 | self.model = OpenPanguModel( |
| 1124 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 1125 | ) |
| 1126 | if get_pp_group().is_last_rank: |
| 1127 | self.lm_head = ParallelLMHead( |
| 1128 | config.vocab_size, |
| 1129 | config.hidden_size, |
| 1130 | quant_config=quant_config, |
| 1131 | prefix=maybe_prefix(prefix, "lm_head"), |
| 1132 | ) |
| 1133 | if config.tie_word_embeddings: |
| 1134 | self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) |
| 1135 | else: |
| 1136 | self.lm_head = PPMissingLayer() |
| 1137 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 1138 | self.make_empty_intermediate_tensors = ( |
| 1139 | self.model.make_empty_intermediate_tensors |
| 1140 | ) |
| 1141 | |
| 1142 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1143 | return self.model.embed_input_ids(input_ids) |
| 1144 | |
| 1145 | def forward( |
| 1146 | self, |
| 1147 | input_ids: torch.Tensor | None, |
| 1148 | positions: torch.Tensor, |
| 1149 | intermediate_tensors: IntermediateTensors | None = None, |
| 1150 | inputs_embeds: torch.Tensor | None = None, |
| 1151 | ) -> torch.Tensor | IntermediateTensors: |
| 1152 | hidden_states = self.model( |
| 1153 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 1154 | ) |
| 1155 | return hidden_states |
| 1156 | |
| 1157 | def compute_logits( |
| 1158 | self, |
| 1159 | hidden_states: torch.Tensor, |
| 1160 | ) -> torch.Tensor | None: |
| 1161 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 1162 | return logits |
| 1163 | |
| 1164 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1165 | loader = AutoWeightsLoader(self) |
| 1166 | return loader.load_weights(weights) |
| 1167 | |
| 1168 | |
| 1169 | class OpenPanguMoEModel(OpenPanguModelBase, MixtureOfExperts): |
| 1170 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1171 | super().__init__(vllm_config=vllm_config, prefix=prefix) |
| 1172 | config = vllm_config.model_config.hf_config |
| 1173 | |
| 1174 | # Set MoE hyperparameters |
| 1175 | self.num_moe_layers = config.num_hidden_layers - config.first_k_dense_replace |
| 1176 | self.num_expert_groups = 1 |
| 1177 | |
| 1178 | self.moe_layers = [] |
| 1179 | example_moe = None |
| 1180 | for layer in self.model.layers: |
| 1181 | if isinstance(layer, PPMissingLayer): |
| 1182 | continue |
| 1183 | |
| 1184 | assert isinstance(layer, OpenPanguDecoderLayer) |
| 1185 | if isinstance(layer.mlp, OpenPanguMoE): |
| 1186 | # Pick last one layer since the first ones may be dense layers. |
| 1187 | example_moe = layer.mlp |
| 1188 | self.moe_layers.append(layer.mlp.experts) |
| 1189 | |
| 1190 | if example_moe is None: |
| 1191 | raise RuntimeError("No MOE layer found in model.layers.") |
| 1192 | |
| 1193 | self.num_logical_experts = example_moe.n_logical_experts |
| 1194 | self.num_physical_experts = example_moe.n_physical_experts |
| 1195 | self.num_local_physical_experts = example_moe.n_local_physical_experts |
| 1196 | self.n_routed_experts = example_moe.n_routed_experts |
| 1197 | self.n_shared_experts = example_moe.n_shared_experts |
| 1198 | self.num_redundant_experts = example_moe.n_redundant_experts |
| 1199 | |
| 1200 | def update_physical_experts_metadata( |
| 1201 | self, |
| 1202 | num_physical_experts: int, |
| 1203 | num_local_physical_experts: int, |
| 1204 | ) -> None: |
| 1205 | assert self.num_local_physical_experts == num_local_physical_experts |
| 1206 | self.num_physical_experts = num_physical_experts |
| 1207 | self.num_local_physical_experts = num_local_physical_experts |
| 1208 | self.num_redundant_experts = num_physical_experts - self.num_logical_experts |
| 1209 | for layer in self.model.layers: |
| 1210 | if isinstance(layer.mlp, OpenPanguMoE): |
| 1211 | moe = layer.mlp |
| 1212 | moe.n_local_physical_experts = num_local_physical_experts |
| 1213 | moe.n_physical_experts = num_physical_experts |
| 1214 | moe.n_redundant_experts = self.num_redundant_experts |
| 1215 | moe.experts.update_expert_map() |
| 1216 | |
| 1217 | |
| 1218 | class OpenPanguEmbeddedModel(OpenPanguModelBase): |
| 1219 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1220 | super().__init__(vllm_config=vllm_config, prefix=prefix) |
| 1221 | |
| 1222 | |
| 1223 | class PanguEmbeddedForCausalLM(OpenPanguEmbeddedModel): |
| 1224 | pass |
| 1225 | |
| 1226 | |
| 1227 | class PanguUltraMoEForCausalLM(OpenPanguMoEModel): |
| 1228 | pass |
| 1229 | |
| 1230 | |
| 1231 | class PanguProMoEV2ForCausalLM(OpenPanguMoEModel): |
| 1232 | pass |