| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # Apache License, Version 2.0: |
| 4 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | # you may not use this file except in compliance with the License. |
| 6 | # You may obtain a copy of the License at |
| 7 | # |
| 8 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | # |
| 10 | # Unless required by applicable law or agreed to in writing, software |
| 11 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | # See the License for the specific language governing permissions and |
| 14 | # limitations under the License. |
| 15 | # |
| 16 | # MIT License: |
| 17 | # Permission is hereby granted, free of charge, to any person obtaining a copy |
| 18 | # of this software and associated documentation files (the "Software"), to deal |
| 19 | # in the Software without restriction, including without limitation the rights |
| 20 | # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
| 21 | # copies of the Software, and to permit persons to whom the Software is |
| 22 | # furnished to do so, subject to the following conditions: |
| 23 | # |
| 24 | # The above copyright notice and this permission notice shall be included in all |
| 25 | # copies or substantial portions of the Software. |
| 26 | # |
| 27 | # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 28 | # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 29 | # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 30 | # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 31 | # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 32 | # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 33 | # SOFTWARE. |
| 34 | """Inference-only Flash model compatible with HuggingFace weights.""" |
| 35 | |
| 36 | import typing |
| 37 | from collections.abc import Callable, Iterable |
| 38 | from itertools import islice |
| 39 | |
| 40 | import torch |
| 41 | from torch import nn |
| 42 | from transformers import PreTrainedConfig |
| 43 | |
| 44 | from vllm.compilation.decorators import support_torch_compile |
| 45 | from vllm.config import CacheConfig, VllmConfig |
| 46 | from vllm.distributed import get_pp_group |
| 47 | from vllm.logger import init_logger |
| 48 | from vllm.model_executor.layers.activation import SiluAndMul |
| 49 | from vllm.model_executor.layers.fused_moe import ( |
| 50 | FusedMoEFactory, |
| 51 | GateLinear, |
| 52 | fused_moe_make_expert_params_mapping, |
| 53 | ) |
| 54 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 55 | from vllm.model_executor.layers.linear import ( |
| 56 | MergedColumnParallelLinear, |
| 57 | RowParallelLinear, |
| 58 | ) |
| 59 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 60 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 61 | from vllm.model_executor.layers.quantization.utils.int8_utils import block_dequant |
| 62 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 63 | ParallelLMHead, |
| 64 | VocabParallelEmbedding, |
| 65 | ) |
| 66 | from vllm.model_executor.model_loader.weight_utils import default_weight_loader |
| 67 | from vllm.model_executor.models.deepseek_v2 import DeepseekV2MLAAttention |
| 68 | from vllm.sequence import IntermediateTensors |
| 69 | |
| 70 | from .interfaces import SupportsLoRA, SupportsPP |
| 71 | from .utils import ( |
| 72 | AutoWeightsLoader, |
| 73 | PPMissingLayer, |
| 74 | is_pp_missing_parameter, |
| 75 | make_empty_intermediate_tensors_factory, |
| 76 | make_layers, |
| 77 | maybe_prefix, |
| 78 | ) |
| 79 | |
| 80 | logger = init_logger(__name__) |
| 81 | |
| 82 | |
| 83 | class FlashConfig(PreTrainedConfig): |
| 84 | """Flash model configuration.""" |
| 85 | |
| 86 | moe_intermediate_size: int |
| 87 | |
| 88 | model_type = "longcat_flash" |
| 89 | keys_to_ignore_at_inference = ["past_key_values"] |
| 90 | |
| 91 | def __init__( |
| 92 | self, |
| 93 | vocab_size=131072, |
| 94 | hidden_size=4096, |
| 95 | intermediate_size=8192, |
| 96 | num_layers=28, |
| 97 | num_hidden_layers=None, |
| 98 | num_attention_heads=96, |
| 99 | num_key_value_heads=128, |
| 100 | ep_size=1, |
| 101 | kv_lora_rank=512, |
| 102 | q_lora_rank=1536, |
| 103 | qk_rope_head_dim=64, |
| 104 | v_head_dim=128, |
| 105 | qk_nope_head_dim=128, |
| 106 | num_experts_per_tok=None, |
| 107 | norm_topk_prob=False, |
| 108 | max_position_embeddings=8192, |
| 109 | initializer_range=0.02, |
| 110 | rms_norm_eps=1e-05, |
| 111 | use_cache=True, |
| 112 | pad_token_id=None, |
| 113 | bos_token_id=100000, |
| 114 | eos_token_id=100001, |
| 115 | pretraining_tp=1, |
| 116 | tie_word_embeddings=False, |
| 117 | rope_parameters=None, |
| 118 | attention_bias=False, |
| 119 | attention_dropout=0.0, |
| 120 | mla_scale_q_lora=False, |
| 121 | mla_scale_kv_lora=False, |
| 122 | dtype="bfloat16", |
| 123 | params_dtype="bfloat16", |
| 124 | router_dtype="float32", |
| 125 | router_bias=False, |
| 126 | topk_method=None, |
| 127 | routed_scaling_factor=1.0, |
| 128 | zero_expert_num=0, |
| 129 | zero_expert_type=None, |
| 130 | nextn_use_scmoe=False, |
| 131 | **kwargs, |
| 132 | ): |
| 133 | super().__init__( |
| 134 | pad_token_id=pad_token_id, |
| 135 | bos_token_id=bos_token_id, |
| 136 | eos_token_id=eos_token_id, |
| 137 | tie_word_embeddings=tie_word_embeddings, |
| 138 | dtype=dtype, |
| 139 | params_dtype=params_dtype, |
| 140 | router_dtype=router_dtype, |
| 141 | topk_method=topk_method, |
| 142 | router_bias=router_bias, |
| 143 | nextn_use_scmoe=nextn_use_scmoe, |
| 144 | **kwargs, |
| 145 | ) |
| 146 | self.vocab_size = vocab_size |
| 147 | self.max_position_embeddings = max_position_embeddings |
| 148 | self.hidden_size = hidden_size |
| 149 | self.num_hidden_layers = ( |
| 150 | num_hidden_layers if num_hidden_layers is not None else num_layers |
| 151 | ) |
| 152 | self.num_attention_heads = num_attention_heads |
| 153 | self.ep_size = ep_size |
| 154 | self.kv_lora_rank = kv_lora_rank |
| 155 | self.q_lora_rank = q_lora_rank |
| 156 | self.qk_rope_head_dim = qk_rope_head_dim |
| 157 | self.v_head_dim = v_head_dim |
| 158 | self.qk_nope_head_dim = qk_nope_head_dim |
| 159 | self.num_experts_per_tok = num_experts_per_tok |
| 160 | self.norm_topk_prob = norm_topk_prob |
| 161 | # for backward compatibility |
| 162 | if num_key_value_heads is None: |
| 163 | num_key_value_heads = num_attention_heads |
| 164 | |
| 165 | self.num_key_value_heads = num_key_value_heads |
| 166 | self.initializer_range = initializer_range |
| 167 | self.rms_norm_eps = rms_norm_eps |
| 168 | self.pretraining_tp = pretraining_tp |
| 169 | self.use_cache = use_cache |
| 170 | # Try to set `rope_scaling` if available, otherwise use `rope_parameters` |
| 171 | rope_scaling = kwargs.pop("rope_scaling", None) |
| 172 | rope_parameters = rope_scaling or rope_parameters or {"rope_type": "default"} |
| 173 | rope_theta = kwargs.pop("rope_theta", 1000000.0) |
| 174 | if "rope_theta" not in rope_parameters: |
| 175 | rope_parameters["rope_theta"] = rope_theta |
| 176 | self.rope_parameters = rope_parameters |
| 177 | self.attention_bias = attention_bias |
| 178 | self.attention_dropout = attention_dropout |
| 179 | self.mla_scale_q_lora = mla_scale_q_lora |
| 180 | self.mla_scale_kv_lora = mla_scale_kv_lora |
| 181 | self.zero_expert_num = zero_expert_num |
| 182 | self.zero_expert_type = zero_expert_type |
| 183 | self.routed_scaling_factor = routed_scaling_factor |
| 184 | self.hidden_act = "silu" |
| 185 | self.intermediate_size = ( |
| 186 | self.ffn_hidden_size |
| 187 | if hasattr(self, "ffn_hidden_size") |
| 188 | else intermediate_size |
| 189 | ) |
| 190 | if hasattr(self, "moe_intermediate_size"): |
| 191 | self.moe_intermediate_size = self.moe_intermediate_size |
| 192 | elif hasattr(self, "expert_ffn_hidden_size"): |
| 193 | self.moe_intermediate_size = self.expert_ffn_hidden_size |
| 194 | else: |
| 195 | self.moe_intermediate_size = self.intermediate_size |
| 196 | |
| 197 | |
| 198 | class FlashMLP(nn.Module): |
| 199 | """Flash MLP layer.""" |
| 200 | |
| 201 | def __init__( |
| 202 | self, |
| 203 | hidden_size: int, |
| 204 | intermediate_size: int, |
| 205 | hidden_act: str, |
| 206 | quant_config: QuantizationConfig | None = None, |
| 207 | reduce_results: bool = True, |
| 208 | prefix: str = "", |
| 209 | ) -> None: |
| 210 | super().__init__() |
| 211 | self.gate_up_proj = MergedColumnParallelLinear( |
| 212 | hidden_size, |
| 213 | [intermediate_size] * 2, |
| 214 | bias=False, |
| 215 | quant_config=quant_config, |
| 216 | prefix=f"{prefix}.gate_up_proj", |
| 217 | ) |
| 218 | self.down_proj = RowParallelLinear( |
| 219 | intermediate_size, |
| 220 | hidden_size, |
| 221 | bias=False, |
| 222 | quant_config=quant_config, |
| 223 | reduce_results=reduce_results, |
| 224 | prefix=f"{prefix}.down_proj", |
| 225 | ) |
| 226 | if hidden_act != "silu": |
| 227 | raise ValueError( |
| 228 | f"Unsupported activation: {hidden_act}. Only silu is supported for now." |
| 229 | ) |
| 230 | self.act_fn = SiluAndMul() |
| 231 | |
| 232 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 233 | if x.numel() == 0: |
| 234 | return x |
| 235 | |
| 236 | gate_up, _ = self.gate_up_proj(x) |
| 237 | x = self.act_fn(gate_up) |
| 238 | x, _ = self.down_proj(x) |
| 239 | return x |
| 240 | |
| 241 | |
| 242 | class LongcatRouter(nn.Module): |
| 243 | def __init__( |
| 244 | self, |
| 245 | config: FlashConfig, |
| 246 | zero_expert_num: int, |
| 247 | router_params_dtype: torch.dtype, |
| 248 | prefix: str = "", |
| 249 | ): |
| 250 | super().__init__() |
| 251 | self.n_routed_experts = ( |
| 252 | config.n_routed_experts |
| 253 | if hasattr(config, "n_routed_experts") |
| 254 | else config.num_experts[0] |
| 255 | ) |
| 256 | self.n_routed_experts = self.n_routed_experts + zero_expert_num |
| 257 | self.classifier = GateLinear( |
| 258 | config.hidden_size, |
| 259 | self.n_routed_experts, |
| 260 | bias=config.router_bias, |
| 261 | out_dtype=router_params_dtype, |
| 262 | params_dtype=router_params_dtype, |
| 263 | prefix=f"{prefix}.classifier", |
| 264 | ) |
| 265 | self.e_score_correction_bias = nn.Parameter( |
| 266 | torch.zeros((self.n_routed_experts), dtype=router_params_dtype) |
| 267 | ) |
| 268 | |
| 269 | def forward(self, hidden_states): |
| 270 | logits, _ = self.classifier(hidden_states) |
| 271 | return logits |
| 272 | |
| 273 | |
| 274 | class LongcatMoe(nn.Module): |
| 275 | def __init__( |
| 276 | self, |
| 277 | config: FlashConfig, |
| 278 | num_experts: int, |
| 279 | top_k: int, |
| 280 | hidden_size: int, |
| 281 | intermediate_size: int, |
| 282 | params_dtype: torch.dtype | None = None, |
| 283 | quant_config: QuantizationConfig | None = None, |
| 284 | prefix: str = "", |
| 285 | enable_eplb: bool = False, |
| 286 | ): |
| 287 | super().__init__() |
| 288 | self.hidden_size = hidden_size |
| 289 | # Gate always runs at half / full precision for now. |
| 290 | self.router_params_dtype = params_dtype |
| 291 | if config.router_dtype == "float32": |
| 292 | self.router_params_dtype = torch.float32 |
| 293 | |
| 294 | self.router = LongcatRouter( |
| 295 | config=config, |
| 296 | zero_expert_num=config.zero_expert_num, |
| 297 | router_params_dtype=self.router_params_dtype, |
| 298 | prefix=f"{prefix}.gate", |
| 299 | ) |
| 300 | |
| 301 | assert config.zero_expert_type is not None |
| 302 | self.experts = FusedMoEFactory( |
| 303 | zero_expert_type=config.zero_expert_type, |
| 304 | e_score_correction_bias=self.router.e_score_correction_bias, |
| 305 | num_experts=num_experts, |
| 306 | top_k=top_k, |
| 307 | hidden_size=hidden_size, |
| 308 | intermediate_size=intermediate_size, |
| 309 | params_dtype=params_dtype, |
| 310 | renormalize=False, |
| 311 | quant_config=quant_config, |
| 312 | prefix=f"{prefix}.experts", |
| 313 | enable_eplb=enable_eplb, |
| 314 | routed_scaling_factor=config.routed_scaling_factor, |
| 315 | router_logits_dtype=self.router_params_dtype, |
| 316 | ) |
| 317 | |
| 318 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 319 | num_tokens, hidden_dim = hidden_states.shape |
| 320 | hidden_states = hidden_states.view(-1, hidden_dim) |
| 321 | |
| 322 | # Align to MoERunner padded hidden size to avoid dim mismatch |
| 323 | padded_hidden = self.experts.moe_config.hidden_dim |
| 324 | if hidden_dim < padded_hidden: |
| 325 | hidden_states_padded = torch.nn.functional.pad( |
| 326 | hidden_states, |
| 327 | (0, padded_hidden - hidden_dim), |
| 328 | mode="constant", |
| 329 | value=0.0, |
| 330 | ) |
| 331 | else: |
| 332 | hidden_states_padded = hidden_states |
| 333 | |
| 334 | router_logits_full = self.router(hidden_states_padded) |
| 335 | |
| 336 | # MoERunner handles routing memoization and zero expert computation |
| 337 | # internally. Pass full router_logits (including zero experts) so that |
| 338 | # zero experts can be properly identified in routing. |
| 339 | final_hidden_states = self.experts( |
| 340 | hidden_states=hidden_states_padded, |
| 341 | router_logits=router_logits_full, # Full logits (includes zero experts) |
| 342 | ) |
| 343 | |
| 344 | # Crop back to original hidden dimension if padded earlier |
| 345 | if padded_hidden != hidden_dim: |
| 346 | final_hidden_states = final_hidden_states[..., :hidden_dim] |
| 347 | |
| 348 | return final_hidden_states.view(num_tokens, hidden_dim) |
| 349 | |
| 350 | |
| 351 | class FlashDecoderLayer(nn.Module): |
| 352 | """Flash decoder layer with dual attention and MLP structure.""" |
| 353 | |
| 354 | def __init__( |
| 355 | self, |
| 356 | vllm_config: VllmConfig, |
| 357 | config: FlashConfig, |
| 358 | cache_config: CacheConfig | None = None, |
| 359 | quant_config: QuantizationConfig | None = None, |
| 360 | prefix: str = "", |
| 361 | enable_eplb: bool = False, |
| 362 | ) -> None: |
| 363 | super().__init__() |
| 364 | self.layer_idx = int(prefix.split(sep=".")[-1]) |
| 365 | self.hidden_size = config.hidden_size |
| 366 | max_position_embeddings = getattr(config, "max_position_embeddings", 8192) |
| 367 | |
| 368 | # Dual attention structure |
| 369 | self.self_attn = nn.ModuleList( |
| 370 | [ |
| 371 | DeepseekV2MLAAttention( |
| 372 | vllm_config=vllm_config, |
| 373 | config=config, |
| 374 | hidden_size=self.hidden_size, |
| 375 | num_heads=config.num_attention_heads, |
| 376 | qk_nope_head_dim=config.qk_nope_head_dim, |
| 377 | qk_rope_head_dim=config.qk_rope_head_dim, |
| 378 | v_head_dim=config.v_head_dim, |
| 379 | q_lora_rank=( |
| 380 | config.q_lora_rank if hasattr(config, "q_lora_rank") else None |
| 381 | ), |
| 382 | kv_lora_rank=config.kv_lora_rank, |
| 383 | max_position_embeddings=max_position_embeddings, |
| 384 | cache_config=cache_config, |
| 385 | quant_config=None |
| 386 | if "self_attn" in getattr(config, "disable_quant_module", []) |
| 387 | else quant_config, |
| 388 | prefix=f"{prefix}.self_attn.{i}", |
| 389 | ) |
| 390 | for i in range(2) |
| 391 | ] |
| 392 | ) |
| 393 | self.input_layernorm = nn.ModuleList( |
| 394 | [RMSNorm(config.hidden_size, eps=config.rms_norm_eps) for i in range(2)] |
| 395 | ) |
| 396 | self.post_attention_layernorm = nn.ModuleList( |
| 397 | [RMSNorm(config.hidden_size, eps=config.rms_norm_eps) for i in range(2)] |
| 398 | ) |
| 399 | |
| 400 | # Dual MLP structure |
| 401 | self.mlps = nn.ModuleList( |
| 402 | [ |
| 403 | FlashMLP( |
| 404 | hidden_size=self.hidden_size, |
| 405 | intermediate_size=config.intermediate_size, |
| 406 | hidden_act=config.hidden_act, |
| 407 | quant_config=None |
| 408 | if "mlps" in getattr(config, "disable_quant_module", []) |
| 409 | else quant_config, |
| 410 | prefix=f"{prefix}.mlps.{i}", |
| 411 | ) |
| 412 | for i in range(2) |
| 413 | ] |
| 414 | ) |
| 415 | |
| 416 | self.mlp = LongcatMoe( |
| 417 | config=config, |
| 418 | num_experts=config.n_routed_experts |
| 419 | if hasattr(config, "n_routed_experts") |
| 420 | else config.num_experts[self.layer_idx], |
| 421 | top_k=config.moe_topk |
| 422 | if hasattr(config, "moe_topk") |
| 423 | else config.num_experts_per_tok, |
| 424 | hidden_size=config.hidden_size, |
| 425 | intermediate_size=config.moe_intermediate_size, |
| 426 | quant_config=quant_config, |
| 427 | prefix=(f"{prefix}.mlp"), |
| 428 | ) |
| 429 | |
| 430 | def forward( |
| 431 | self, |
| 432 | positions: torch.Tensor, |
| 433 | hidden_states: torch.Tensor, |
| 434 | residual: torch.Tensor | None, |
| 435 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 436 | if residual is None: |
| 437 | residual = hidden_states |
| 438 | hidden_states = self.input_layernorm[0](hidden_states) |
| 439 | else: |
| 440 | hidden_states, residual = self.input_layernorm[0](hidden_states, residual) |
| 441 | |
| 442 | hidden_states = self.self_attn[0]( |
| 443 | positions=positions, |
| 444 | hidden_states=hidden_states, |
| 445 | llama_4_scaling=None, |
| 446 | ) |
| 447 | |
| 448 | hidden_states, residual = self.post_attention_layernorm[0]( |
| 449 | hidden_states, residual |
| 450 | ) |
| 451 | |
| 452 | # moe |
| 453 | hidden_states_copy = hidden_states.clone() |
| 454 | moe_hidden_states = self.mlp(hidden_states_copy) |
| 455 | |
| 456 | # first mlp |
| 457 | hidden_states = self.mlps[0](hidden_states) |
| 458 | |
| 459 | hidden_states, residual = self.input_layernorm[1](hidden_states, residual) |
| 460 | |
| 461 | # second_attn |
| 462 | hidden_states = self.self_attn[1]( |
| 463 | positions=positions, |
| 464 | hidden_states=hidden_states, |
| 465 | llama_4_scaling=None, |
| 466 | ) |
| 467 | hidden_states, residual = self.post_attention_layernorm[1]( |
| 468 | hidden_states, residual |
| 469 | ) |
| 470 | |
| 471 | # second_mlp |
| 472 | hidden_states = self.mlps[1](hidden_states) |
| 473 | |
| 474 | hidden_states = hidden_states + moe_hidden_states |
| 475 | |
| 476 | return hidden_states, residual |
| 477 | |
| 478 | |
| 479 | @support_torch_compile |
| 480 | class FlashModel(nn.Module): |
| 481 | """Flash model.""" |
| 482 | |
| 483 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 484 | super().__init__() |
| 485 | config = FlashConfig(**vllm_config.model_config.hf_config.__dict__) |
| 486 | cache_config = vllm_config.cache_config |
| 487 | quant_config = vllm_config.quant_config |
| 488 | self.config = config |
| 489 | self.quant_config = quant_config |
| 490 | |
| 491 | self.vocab_size = config.vocab_size |
| 492 | |
| 493 | if get_pp_group().is_first_rank: |
| 494 | self.embed_tokens = VocabParallelEmbedding( |
| 495 | config.vocab_size, |
| 496 | config.hidden_size, |
| 497 | prefix=maybe_prefix(prefix, "embed_tokens"), |
| 498 | ) |
| 499 | else: |
| 500 | self.embed_tokens = PPMissingLayer() |
| 501 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 502 | config.num_hidden_layers, |
| 503 | lambda prefix: FlashDecoderLayer( |
| 504 | vllm_config, |
| 505 | config, |
| 506 | cache_config=cache_config, |
| 507 | quant_config=quant_config, |
| 508 | prefix=prefix, |
| 509 | ), |
| 510 | prefix=f"{prefix}.layers", |
| 511 | ) |
| 512 | if get_pp_group().is_last_rank: |
| 513 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 514 | else: |
| 515 | self.norm = PPMissingLayer() |
| 516 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 517 | ["hidden_states", "residual"], config.hidden_size |
| 518 | ) |
| 519 | |
| 520 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 521 | return self.embed_tokens(input_ids) |
| 522 | |
| 523 | def forward( |
| 524 | self, |
| 525 | input_ids: torch.Tensor | None, |
| 526 | positions: torch.Tensor, |
| 527 | intermediate_tensors: IntermediateTensors | None = None, |
| 528 | inputs_embeds: torch.Tensor | None = None, |
| 529 | ) -> torch.Tensor | IntermediateTensors: |
| 530 | if get_pp_group().is_first_rank: |
| 531 | if inputs_embeds is not None: |
| 532 | hidden_states = inputs_embeds |
| 533 | else: |
| 534 | hidden_states = self.embed_input_ids(input_ids) |
| 535 | residual = None |
| 536 | else: |
| 537 | assert intermediate_tensors is not None |
| 538 | hidden_states = intermediate_tensors["hidden_states"] |
| 539 | residual = intermediate_tensors["residual"] |
| 540 | |
| 541 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 542 | hidden_states, residual = layer( |
| 543 | positions, |
| 544 | hidden_states, |
| 545 | residual, |
| 546 | ) |
| 547 | |
| 548 | if not get_pp_group().is_last_rank: |
| 549 | return IntermediateTensors( |
| 550 | {"hidden_states": hidden_states, "residual": residual} |
| 551 | ) |
| 552 | |
| 553 | hidden_states, _ = self.norm(hidden_states, residual) |
| 554 | return hidden_states |
| 555 | |
| 556 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 557 | # Params for weights, fp8 weight scales, fp8 activation scales |
| 558 | # (param_name, weight_name, expert_id, shard_id) |
| 559 | return fused_moe_make_expert_params_mapping( |
| 560 | self, |
| 561 | ckpt_gate_proj_name="gate_proj", |
| 562 | ckpt_down_proj_name="down_proj", |
| 563 | ckpt_up_proj_name="up_proj", |
| 564 | num_experts=self.config.n_routed_experts |
| 565 | if hasattr(self.config, "n_routed_experts") |
| 566 | else self.config.num_experts[0], |
| 567 | ) |
| 568 | |
| 569 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 570 | stacked_params_mapping = [ |
| 571 | ("fused_qkv_a_proj", "q_a_proj", 0), |
| 572 | ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1), |
| 573 | (".gate_up_proj", ".gate_proj", 0), |
| 574 | (".gate_up_proj", ".up_proj", 1), |
| 575 | ] |
| 576 | |
| 577 | expert_params_mapping = self.get_expert_mapping() |
| 578 | loaded_params: set[str] = set() |
| 579 | |
| 580 | params_dict = dict(self.named_parameters()) |
| 581 | for name, loaded_weight in weights: |
| 582 | if "rotary_emb.inv_freq" in name: |
| 583 | continue |
| 584 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 585 | if weight_name not in name: |
| 586 | continue |
| 587 | if "mlp" in name and "mlps" not in name: |
| 588 | continue |
| 589 | name = name.replace(weight_name, param_name) |
| 590 | # Skip loading extra bias for GPTQ models. |
| 591 | if ( |
| 592 | name.endswith(".bias") or name.endswith("_bias") |
| 593 | ) and name not in params_dict: |
| 594 | continue |
| 595 | # Skip mtp |
| 596 | if ".mtp." in name: |
| 597 | continue |
| 598 | if is_pp_missing_parameter(name, self): |
| 599 | continue |
| 600 | param = params_dict[name] |
| 601 | weight_loader = param.weight_loader |
| 602 | weight_loader(param, loaded_weight, shard_id) |
| 603 | break |
| 604 | else: |
| 605 | is_expert_weight = False |
| 606 | for mapping in expert_params_mapping: |
| 607 | param_name, weight_name, expert_id, expert_shard_id = mapping |
| 608 | if weight_name not in name: |
| 609 | continue |
| 610 | is_expert_weight = True |
| 611 | name_mapped = name.replace(weight_name, param_name) |
| 612 | # Skip mtp |
| 613 | if ".mtp." in name_mapped: |
| 614 | continue |
| 615 | if ( |
| 616 | name_mapped.endswith(".bias") or name_mapped.endswith("_bias") |
| 617 | ) and name not in params_dict: |
| 618 | continue |
| 619 | if is_pp_missing_parameter(name, self): |
| 620 | continue |
| 621 | param = params_dict[name_mapped] |
| 622 | weight_loader = param.weight_loader |
| 623 | weight_loader = typing.cast( |
| 624 | Callable[..., bool], param.weight_loader |
| 625 | ) |
| 626 | success = weight_loader( |
| 627 | param, |
| 628 | loaded_weight, |
| 629 | name_mapped, |
| 630 | shard_id=expert_shard_id, |
| 631 | expert_id=expert_id, |
| 632 | return_success=True, |
| 633 | ) |
| 634 | if success: |
| 635 | name = name_mapped |
| 636 | break |
| 637 | else: |
| 638 | if is_expert_weight: |
| 639 | # We've checked that this is an expert weight |
| 640 | # However it's not mapped locally to this rank |
| 641 | # So we simply skip it |
| 642 | continue |
| 643 | # Skip loading extra bias for GPTQ models. |
| 644 | if name.endswith(".bias") and name not in params_dict: |
| 645 | continue |
| 646 | # Skip loading kv_scale from ckpts towards new design. |
| 647 | if name.endswith(".kv_scale") and name not in params_dict: |
| 648 | continue |
| 649 | # Skip mtp |
| 650 | if ".mtp." in name: |
| 651 | continue |
| 652 | if name is None: |
| 653 | continue |
| 654 | if is_pp_missing_parameter(name, self): |
| 655 | continue |
| 656 | param = params_dict[name] |
| 657 | weight_loader = getattr( |
| 658 | param, "weight_loader", default_weight_loader |
| 659 | ) |
| 660 | weight_loader(param, loaded_weight) |
| 661 | loaded_params.add(name) |
| 662 | for layer_id in range(self.config.num_hidden_layers): |
| 663 | for i in range(2): |
| 664 | if isinstance(self.layers[layer_id], PPMissingLayer): |
| 665 | continue |
| 666 | self_attn = self.layers[layer_id].self_attn[i] |
| 667 | if ( |
| 668 | self.quant_config is not None |
| 669 | and hasattr(self.quant_config, "weight_block_size") |
| 670 | and self_attn.kv_b_proj.weight.dtype |
| 671 | in ( |
| 672 | torch.float8_e4m3fn, |
| 673 | torch.float8_e4m3fnuz, |
| 674 | ) |
| 675 | ): |
| 676 | weight_block_size = self.quant_config.weight_block_size |
| 677 | if weight_block_size is not None: |
| 678 | assert hasattr(self_attn.kv_b_proj, "weight_scale_inv") |
| 679 | dtype = torch.get_default_dtype() |
| 680 | w = block_dequant( |
| 681 | self_attn.kv_b_proj.weight, |
| 682 | self_attn.kv_b_proj.weight_scale_inv, |
| 683 | weight_block_size, |
| 684 | ).to(dtype) |
| 685 | else: |
| 686 | w = self_attn.kv_b_proj.weight |
| 687 | |
| 688 | w_kc, w_vc = w.unflatten( |
| 689 | 0, (-1, self_attn.qk_nope_head_dim + self_attn.v_head_dim) |
| 690 | ).split([self_attn.qk_nope_head_dim, self_attn.v_head_dim], dim=1) |
| 691 | self_attn.w_kc = w_kc.transpose(1, 2).contiguous().transpose(1, 2) |
| 692 | self_attn.w_vc = w_vc.contiguous().transpose(1, 2) |
| 693 | # Guard against compounding on incremental load_weights calls: |
| 694 | # the in-place ``*=`` would otherwise re-apply the MLA LoRA |
| 695 | # scaling to the layernorm weights on each pass. |
| 696 | if self.config.mla_scale_q_lora and not getattr( |
| 697 | self_attn, "_mla_q_lora_scaled", False |
| 698 | ): |
| 699 | self_attn.q_a_layernorm.weight.data *= ( |
| 700 | self.config.hidden_size / self.config.q_lora_rank |
| 701 | ) ** 0.5 |
| 702 | self_attn._mla_q_lora_scaled = True |
| 703 | if self.config.mla_scale_kv_lora and not getattr( |
| 704 | self_attn, "_mla_kv_lora_scaled", False |
| 705 | ): |
| 706 | self_attn.kv_a_layernorm.weight.data *= ( |
| 707 | self.config.hidden_size / self.config.kv_lora_rank |
| 708 | ) ** 0.5 |
| 709 | self_attn._mla_kv_lora_scaled = True |
| 710 | return loaded_params |
| 711 | |
| 712 | |
| 713 | class LongcatFlashForCausalLM(nn.Module, SupportsLoRA, SupportsPP): |
| 714 | """Flash model for causal language modeling.""" |
| 715 | |
| 716 | packed_modules_mapping = { |
| 717 | "qkv_proj": [ |
| 718 | "q_proj", |
| 719 | "k_proj", |
| 720 | "v_proj", |
| 721 | ], |
| 722 | "gate_up_proj": [ |
| 723 | "gate_proj", |
| 724 | "up_proj", |
| 725 | ], |
| 726 | } |
| 727 | |
| 728 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 729 | super().__init__() |
| 730 | config = FlashConfig(**vllm_config.model_config.hf_config.__dict__) |
| 731 | quant_config = vllm_config.quant_config |
| 732 | |
| 733 | self.config = config |
| 734 | config.intermediate_size = ( |
| 735 | config.ffn_hidden_size |
| 736 | if hasattr(config, "ffn_hidden_size") |
| 737 | else config.intermediate_size |
| 738 | ) |
| 739 | |
| 740 | self.quant_config = quant_config |
| 741 | |
| 742 | self.model = FlashModel( |
| 743 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 744 | ) |
| 745 | |
| 746 | if get_pp_group().is_last_rank: |
| 747 | self.lm_head = ParallelLMHead( |
| 748 | config.vocab_size, |
| 749 | config.hidden_size, |
| 750 | quant_config=quant_config, |
| 751 | prefix=maybe_prefix(prefix, "lm_head"), |
| 752 | ) |
| 753 | else: |
| 754 | self.lm_head = PPMissingLayer() |
| 755 | |
| 756 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 757 | self.make_empty_intermediate_tensors = ( |
| 758 | self.model.make_empty_intermediate_tensors |
| 759 | ) |
| 760 | |
| 761 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 762 | return self.model.embed_input_ids(input_ids) |
| 763 | |
| 764 | def forward( |
| 765 | self, |
| 766 | input_ids: torch.Tensor | None, |
| 767 | positions: torch.Tensor, |
| 768 | intermediate_tensors: IntermediateTensors | None = None, |
| 769 | inputs_embeds: torch.Tensor | None = None, |
| 770 | ) -> torch.Tensor | IntermediateTensors: |
| 771 | hidden_states = self.model( |
| 772 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 773 | ) |
| 774 | return hidden_states |
| 775 | |
| 776 | def compute_logits( |
| 777 | self, |
| 778 | hidden_states: torch.Tensor, |
| 779 | ) -> torch.Tensor | None: |
| 780 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 781 | return logits |
| 782 | |
| 783 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 784 | return self.model.get_expert_mapping() |
| 785 | |
| 786 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 787 | loader = AutoWeightsLoader(self) |
| 788 | return loader.load_weights(weights) |