| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # coding=utf-8 |
| 5 | # Copyright 2026 The HY team. |
| 6 | # Copyright 2023 The vLLM team. |
| 7 | # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. |
| 8 | # |
| 9 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX |
| 10 | # and OPT implementations in this library. It has been modified from its |
| 11 | # original forms to accommodate minor architectural differences compared |
| 12 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. |
| 13 | # |
| 14 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 15 | # you may not use this file except in compliance with the License. |
| 16 | # You may obtain a copy of the License at |
| 17 | # |
| 18 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 19 | # |
| 20 | # Unless required by applicable law or agreed to in writing, software |
| 21 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 22 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 23 | # See the License for the specific language governing permissions and |
| 24 | # limitations under the License. |
| 25 | """Inference-only HY V3 MTP model compatible with HuggingFace weights.""" |
| 26 | |
| 27 | from collections.abc import Callable, Iterable |
| 28 | |
| 29 | import regex as re |
| 30 | import torch |
| 31 | from torch import nn |
| 32 | from transformers import PreTrainedConfig |
| 33 | |
| 34 | from vllm.config import CacheConfig, ModelConfig, VllmConfig |
| 35 | from vllm.model_executor.layers.fused_moe import ( |
| 36 | fused_moe_make_expert_params_mapping, |
| 37 | ) |
| 38 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 39 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 40 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 41 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 42 | ParallelLMHead, |
| 43 | VocabParallelEmbedding, |
| 44 | ) |
| 45 | from vllm.model_executor.model_loader.weight_utils import ( |
| 46 | default_weight_loader, |
| 47 | maybe_remap_kv_scale_name, |
| 48 | ) |
| 49 | from vllm.sequence import IntermediateTensors |
| 50 | from vllm.v1.outputs import SamplerOutput |
| 51 | from vllm.v1.sample.metadata import SamplingMetadata |
| 52 | from vllm.v1.sample.sampler import Sampler |
| 53 | |
| 54 | from .hy_v3 import HYV3DecoderLayer |
| 55 | from .utils import ( |
| 56 | get_spec_layer_idx_from_weight_name, |
| 57 | is_pp_missing_parameter, |
| 58 | maybe_prefix, |
| 59 | ) |
| 60 | |
| 61 | |
| 62 | def _is_moe(config: PreTrainedConfig) -> bool: |
| 63 | return bool( |
| 64 | getattr(config, "num_experts", None) |
| 65 | and ( |
| 66 | (isinstance(config.num_experts, int) and config.num_experts > 1) |
| 67 | or (isinstance(config.num_experts, list) and max(config.num_experts) > 1) |
| 68 | ) |
| 69 | ) |
| 70 | |
| 71 | |
| 72 | def _get_cla_factor(config: PreTrainedConfig) -> int: |
| 73 | if not getattr(config, "use_cla", False): |
| 74 | return 1 |
| 75 | return getattr(config, "cla_share_factor", 1) |
| 76 | |
| 77 | |
| 78 | class HYV3SharedHead(nn.Module): |
| 79 | def __init__( |
| 80 | self, |
| 81 | config: PreTrainedConfig, |
| 82 | quant_config: QuantizationConfig | None = None, |
| 83 | ) -> None: |
| 84 | super().__init__() |
| 85 | self.head = ParallelLMHead( |
| 86 | config.vocab_size, config.hidden_size, quant_config=quant_config |
| 87 | ) |
| 88 | |
| 89 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 90 | return hidden_states |
| 91 | |
| 92 | |
| 93 | class HYV3MultiTokenPredictorLayer(nn.Module): |
| 94 | def __init__( |
| 95 | self, |
| 96 | config: PreTrainedConfig, |
| 97 | prefix: str, |
| 98 | model_config: ModelConfig, |
| 99 | cache_config: CacheConfig | None = None, |
| 100 | quant_config: QuantizationConfig | None = None, |
| 101 | ) -> None: |
| 102 | super().__init__() |
| 103 | |
| 104 | self.enorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 105 | self.hnorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 106 | self.eh_proj = nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False) |
| 107 | self.shared_head = HYV3SharedHead(config=config, quant_config=quant_config) |
| 108 | self.mtp_block = HYV3DecoderLayer( |
| 109 | config=config, |
| 110 | cache_config=cache_config, |
| 111 | quant_config=quant_config, |
| 112 | prefix=prefix, |
| 113 | ) |
| 114 | # Final layernorm applied after transformer block, before logits |
| 115 | # projection (matches HF HYV3MTPDecoderLayer.final_layernorm) |
| 116 | self.final_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 117 | |
| 118 | def forward( |
| 119 | self, |
| 120 | input_ids: torch.Tensor, |
| 121 | positions: torch.Tensor, |
| 122 | previous_hidden_states: torch.Tensor, |
| 123 | inputs_embeds: torch.Tensor | None = None, |
| 124 | spec_step_index: int = 0, |
| 125 | ) -> torch.Tensor: |
| 126 | assert inputs_embeds is not None |
| 127 | # masking inputs at position 0, as not needed by MTP |
| 128 | inputs_embeds[positions == 0] = 0 |
| 129 | inputs_embeds = self.enorm(inputs_embeds) |
| 130 | previous_hidden_states = self.hnorm(previous_hidden_states) |
| 131 | |
| 132 | hidden_states = self.eh_proj( |
| 133 | torch.cat([inputs_embeds, previous_hidden_states], dim=-1) |
| 134 | ) |
| 135 | |
| 136 | # HYV3DecoderLayer returns (hidden_states, residual) |
| 137 | hidden_states, residual = self.mtp_block( |
| 138 | positions=positions, hidden_states=hidden_states, residual=None |
| 139 | ) |
| 140 | hidden_states = residual + hidden_states |
| 141 | hidden_states = self.final_layernorm(hidden_states) |
| 142 | return hidden_states |
| 143 | |
| 144 | |
| 145 | class HYV3MultiTokenPredictor(nn.Module): |
| 146 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 147 | super().__init__() |
| 148 | config = vllm_config.model_config.hf_config |
| 149 | self.mtp_start_layer_idx = config.num_hidden_layers |
| 150 | self.num_mtp_layers = config.num_nextn_predict_layers |
| 151 | |
| 152 | # to map the exact layer index from weights |
| 153 | self.layers = torch.nn.ModuleDict( |
| 154 | { |
| 155 | str(idx): HYV3MultiTokenPredictorLayer( |
| 156 | config, |
| 157 | f"{prefix}.layers.{idx}", |
| 158 | model_config=vllm_config.model_config, |
| 159 | cache_config=vllm_config.cache_config, |
| 160 | quant_config=vllm_config.quant_config, |
| 161 | ) |
| 162 | for idx in range( |
| 163 | self.mtp_start_layer_idx, |
| 164 | self.mtp_start_layer_idx + self.num_mtp_layers, |
| 165 | ) |
| 166 | } |
| 167 | ) |
| 168 | |
| 169 | self.embed_tokens = VocabParallelEmbedding( |
| 170 | config.vocab_size, |
| 171 | config.hidden_size, |
| 172 | ) |
| 173 | |
| 174 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 175 | |
| 176 | def forward( |
| 177 | self, |
| 178 | input_ids: torch.Tensor, |
| 179 | positions: torch.Tensor, |
| 180 | previous_hidden_states: torch.Tensor, |
| 181 | inputs_embeds: torch.Tensor | None = None, |
| 182 | spec_step_idx: int = 0, |
| 183 | ) -> torch.Tensor: |
| 184 | if inputs_embeds is None: |
| 185 | inputs_embeds = self.embed_tokens(input_ids) |
| 186 | current_step_idx = spec_step_idx % self.num_mtp_layers |
| 187 | return self.layers[str(self.mtp_start_layer_idx + current_step_idx)]( |
| 188 | input_ids, |
| 189 | positions, |
| 190 | previous_hidden_states, |
| 191 | inputs_embeds, |
| 192 | current_step_idx, |
| 193 | ) |
| 194 | |
| 195 | def compute_logits( |
| 196 | self, |
| 197 | hidden_states: torch.Tensor, |
| 198 | spec_step_idx: int = 0, |
| 199 | ) -> torch.Tensor: |
| 200 | current_step_idx = spec_step_idx % self.num_mtp_layers |
| 201 | mtp_layer = self.layers[str(self.mtp_start_layer_idx + current_step_idx)] |
| 202 | logits = self.logits_processor( |
| 203 | mtp_layer.shared_head.head, mtp_layer.shared_head(hidden_states) |
| 204 | ) |
| 205 | return logits |
| 206 | |
| 207 | |
| 208 | class HYV3MTP(nn.Module): |
| 209 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 210 | super().__init__() |
| 211 | self.config = vllm_config.model_config.hf_config |
| 212 | self.quant_config = vllm_config.quant_config |
| 213 | self.model = HYV3MultiTokenPredictor( |
| 214 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 215 | ) |
| 216 | |
| 217 | self.sampler = Sampler() |
| 218 | |
| 219 | def forward( |
| 220 | self, |
| 221 | input_ids: torch.Tensor, |
| 222 | positions: torch.Tensor, |
| 223 | hidden_states: torch.Tensor, |
| 224 | intermediate_tensors: IntermediateTensors | None = None, |
| 225 | inputs_embeds: torch.Tensor | None = None, |
| 226 | spec_step_idx: int = 0, |
| 227 | ) -> torch.Tensor: |
| 228 | hidden_states = self.model( |
| 229 | input_ids, positions, hidden_states, inputs_embeds, spec_step_idx |
| 230 | ) |
| 231 | return hidden_states |
| 232 | |
| 233 | def compute_logits( |
| 234 | self, |
| 235 | hidden_states: torch.Tensor, |
| 236 | spec_step_idx: int = 0, |
| 237 | ) -> torch.Tensor | None: |
| 238 | return self.model.compute_logits(hidden_states, spec_step_idx) |
| 239 | |
| 240 | def sample( |
| 241 | self, |
| 242 | logits: torch.Tensor, |
| 243 | sampling_metadata: SamplingMetadata, |
| 244 | ) -> SamplerOutput | None: |
| 245 | next_tokens = self.sampler(logits, sampling_metadata) |
| 246 | return next_tokens |
| 247 | |
| 248 | def _split_qkv_weight(self, qkv: torch.Tensor): |
| 249 | num_attention_heads = self.config.num_attention_heads |
| 250 | num_kv_heads = getattr( |
| 251 | self.config, "num_key_value_heads", self.config.num_attention_heads |
| 252 | ) |
| 253 | num_key_value_groups = num_attention_heads // num_kv_heads |
| 254 | hidden_size = self.config.hidden_size |
| 255 | |
| 256 | if hasattr(self.config, "head_dim"): |
| 257 | attention_head_dim = self.config.head_dim |
| 258 | elif hasattr(self.config, "attention_head_dim"): |
| 259 | attention_head_dim = self.config.attention_head_dim |
| 260 | else: |
| 261 | attention_head_dim = self.config.hidden_size // num_attention_heads |
| 262 | |
| 263 | qkv = qkv.reshape( |
| 264 | num_kv_heads, num_key_value_groups + 2, attention_head_dim, hidden_size |
| 265 | ) |
| 266 | q, k, v = torch.split(qkv, (num_key_value_groups, 1, 1), dim=1) |
| 267 | q = q.reshape(-1, hidden_size) |
| 268 | k = k.reshape(-1, hidden_size) |
| 269 | v = v.reshape(-1, hidden_size) |
| 270 | return torch.concat((q, k, v)) |
| 271 | |
| 272 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): |
| 273 | if self.quant_config is not None and ( |
| 274 | cache_scale_mapper := self.quant_config.get_cache_scale_mapper() |
| 275 | ): |
| 276 | weights = cache_scale_mapper.apply(weights) |
| 277 | cla_factor = _get_cla_factor(self.config) |
| 278 | stacked_params_mapping = [ |
| 279 | # (param_name, shard_name, shard_id) |
| 280 | (".qkv_proj", ".q_proj", "q"), |
| 281 | (".qkv_proj", ".k_proj", "k"), |
| 282 | (".qkv_proj", ".v_proj", "v"), |
| 283 | (".gate_up_proj", ".gate_proj", 0), |
| 284 | (".gate_up_proj", ".up_proj", 1), |
| 285 | ] |
| 286 | |
| 287 | num_attention_heads = self.config.num_attention_heads |
| 288 | num_kv_heads = getattr( |
| 289 | self.config, "num_key_value_heads", self.config.num_attention_heads |
| 290 | ) |
| 291 | split_params_mapping: list[ |
| 292 | tuple[ |
| 293 | str, |
| 294 | str, |
| 295 | int, |
| 296 | list[tuple[str | int, int]], |
| 297 | Callable[[torch.Tensor], torch.Tensor] | None, |
| 298 | ] |
| 299 | ] = [ |
| 300 | (".gate_up_proj", ".gate_and_up_proj", 2, [(1, 1), (0, 1)], None), |
| 301 | ( |
| 302 | ".qkv_proj", |
| 303 | ".qkv_proj", |
| 304 | num_attention_heads + num_kv_heads * 2, |
| 305 | [("q", num_attention_heads), ("k", num_kv_heads), ("v", num_kv_heads)], |
| 306 | self._split_qkv_weight, |
| 307 | ), |
| 308 | ] |
| 309 | |
| 310 | if _is_moe(self.config): |
| 311 | expert_params_mapping = fused_moe_make_expert_params_mapping( |
| 312 | self, |
| 313 | ckpt_gate_proj_name="gate_proj", |
| 314 | ckpt_down_proj_name="down_proj", |
| 315 | ckpt_up_proj_name="up_proj", |
| 316 | num_experts=self.config.num_experts, |
| 317 | ) |
| 318 | else: |
| 319 | expert_params_mapping = [] |
| 320 | |
| 321 | params_dict = dict(self.named_parameters()) |
| 322 | |
| 323 | # V3 shared weights mapping: |
| 324 | # - embed_tokens: from main model's model.embed_tokens.weight |
| 325 | # - lm_head: from main model's lm_head.weight → MTP shared_head.head |
| 326 | # (HF infer_mtp uses head_weight=self.lm_head.weight, not the |
| 327 | # checkpoint's model.layers.<N>.shared_head.weight) |
| 328 | # - No norm mapping (V3 MTP has no intermediate norm before lm_head) |
| 329 | mtp_start = self.config.num_hidden_layers |
| 330 | v3_shared_weights = { |
| 331 | "model.embed_tokens.weight": "model.embed_tokens.weight", |
| 332 | "lm_head.weight": f"model.layers.{mtp_start}.shared_head.head.weight", |
| 333 | } |
| 334 | |
| 335 | for name, loaded_weight in weights: |
| 336 | # Intercept shared weights before any other processing |
| 337 | if name in v3_shared_weights: |
| 338 | target_name = v3_shared_weights[name] |
| 339 | if target_name in params_dict: |
| 340 | param = params_dict[target_name] |
| 341 | weight_loader = getattr( |
| 342 | param, "weight_loader", default_weight_loader |
| 343 | ) |
| 344 | weight_loader(param, loaded_weight) |
| 345 | continue |
| 346 | |
| 347 | if "rotary_emb.inv_freq" in name: |
| 348 | continue |
| 349 | if "gate_proj_bias" in name: |
| 350 | name = name.replace("gate_proj_bias", "gate_proj.bias") |
| 351 | if "up_proj_bias" in name: |
| 352 | name = name.replace("up_proj_bias", "up_proj.bias") |
| 353 | if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name: |
| 354 | continue |
| 355 | if self.config.tie_word_embeddings and "lm_head.weight" in name: |
| 356 | continue |
| 357 | spec_layer = get_spec_layer_idx_from_weight_name(self.config, name) |
| 358 | if spec_layer is None: |
| 359 | continue |
| 360 | name = self._rewrite_spec_layer_name(spec_layer, name) |
| 361 | # Skip weights that _rewrite_spec_layer_name marked for skipping |
| 362 | if name == "__skip__": |
| 363 | continue |
| 364 | if "scale" in name: |
| 365 | remapped_name = maybe_remap_kv_scale_name(name, params_dict) |
| 366 | if remapped_name is None: |
| 367 | continue |
| 368 | name = remapped_name |
| 369 | is_found = False |
| 370 | |
| 371 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 372 | if weight_name not in name: |
| 373 | continue |
| 374 | if "mlp.experts" in name: |
| 375 | continue |
| 376 | if weight_name == ".q_proj": |
| 377 | match = re.search(r"layers\.\d+", name) |
| 378 | if match: |
| 379 | layer_id = int(match.group(0).split(".")[-1]) |
| 380 | if cla_factor > 1 and layer_id % cla_factor != 0: |
| 381 | continue |
| 382 | name = name.replace(weight_name, param_name) |
| 383 | if name.endswith(".bias") and name not in params_dict: |
| 384 | continue |
| 385 | |
| 386 | if is_pp_missing_parameter(name, self): |
| 387 | continue |
| 388 | |
| 389 | param = params_dict[name] |
| 390 | weight_loader = param.weight_loader |
| 391 | weight_loader(param, loaded_weight, shard_id) |
| 392 | |
| 393 | is_found = True |
| 394 | break |
| 395 | if is_found: |
| 396 | continue |
| 397 | |
| 398 | for param_name, weight_name, den, split_param, func in split_params_mapping: |
| 399 | if weight_name not in name: |
| 400 | continue |
| 401 | name = name.replace(weight_name, param_name) |
| 402 | if name.endswith(".bias") and name not in params_dict: |
| 403 | continue |
| 404 | |
| 405 | if is_pp_missing_parameter(name, self): |
| 406 | continue |
| 407 | |
| 408 | assert loaded_weight.shape[0] % den == 0 |
| 409 | units = loaded_weight.shape[0] // den |
| 410 | |
| 411 | param = params_dict[name] |
| 412 | weight_loader = param.weight_loader |
| 413 | offset = 0 |
| 414 | for shard_id, num in split_param: |
| 415 | new_offset = offset + num * units |
| 416 | if func: |
| 417 | weight_loader( |
| 418 | param, func(loaded_weight)[offset:new_offset], shard_id |
| 419 | ) |
| 420 | else: |
| 421 | weight_loader(param, loaded_weight[offset:new_offset], shard_id) |
| 422 | offset = new_offset |
| 423 | |
| 424 | break |
| 425 | else: |
| 426 | if name.endswith(".bias") and name not in params_dict: |
| 427 | continue |
| 428 | for mapping in expert_params_mapping: |
| 429 | param_name, weight_name, expert_id, shard_id = mapping |
| 430 | if weight_name not in name: |
| 431 | continue |
| 432 | name = name.replace(weight_name, param_name) |
| 433 | if is_pp_missing_parameter(name, self): |
| 434 | continue |
| 435 | param = params_dict[name] |
| 436 | weight_loader = param.weight_loader |
| 437 | weight_loader( |
| 438 | param, |
| 439 | loaded_weight, |
| 440 | name, |
| 441 | shard_id=shard_id, |
| 442 | expert_id=expert_id, |
| 443 | ) |
| 444 | break |
| 445 | else: |
| 446 | if is_pp_missing_parameter(name, self): |
| 447 | continue |
| 448 | |
| 449 | if "mlp.gate.wg." in name: |
| 450 | name = name.replace("wg.", "") |
| 451 | # V3 checkpoint: mlp.router.gate -> mlp.gate |
| 452 | if "mlp.router.gate." in name: |
| 453 | name = name.replace("router.gate.", "gate.") |
| 454 | |
| 455 | param = params_dict[name] |
| 456 | weight_loader = getattr( |
| 457 | param, "weight_loader", default_weight_loader |
| 458 | ) |
| 459 | weight_loader(param, loaded_weight) |
| 460 | |
| 461 | def _rewrite_spec_layer_name(self, spec_layer: int, name: str) -> str: |
| 462 | """Rewrite spec layer weight names to match vLLM module structure.""" |
| 463 | # Skip embed_tokens (doesn't exist in V3 MTP checkpoint under spec |
| 464 | # layer) and shared_head (we use main model's lm_head instead) |
| 465 | if f"model.layers.{spec_layer}.embed_tokens" in name: |
| 466 | return "__skip__" |
| 467 | if f"model.layers.{spec_layer}.shared_head" in name: |
| 468 | return "__skip__" |
| 469 | |
| 470 | spec_layer_weight_names = ["enorm", "hnorm", "eh_proj", "final_layernorm"] |
| 471 | spec_layer_weight = False |
| 472 | for weight_name in spec_layer_weight_names: |
| 473 | if weight_name in name: |
| 474 | spec_layer_weight = True |
| 475 | break |
| 476 | if not spec_layer_weight: |
| 477 | # Transformer block weights go under .mtp_block |
| 478 | name = name.replace( |
| 479 | f"model.layers.{spec_layer}.", f"model.layers.{spec_layer}.mtp_block." |
| 480 | ) |
| 481 | return name |