| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from typing import final |
| 5 | |
| 6 | import torch |
| 7 | from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE |
| 8 | from transformers import PreTrainedConfig |
| 9 | |
| 10 | from vllm import envs |
| 11 | from vllm.config.model_arch import ( |
| 12 | ModelArchitectureConfig, |
| 13 | ) |
| 14 | from vllm.config.utils import getattr_iter |
| 15 | from vllm.logger import init_logger |
| 16 | from vllm.transformers_utils.config import ( |
| 17 | ConfigFormat, |
| 18 | get_safetensors_params_metadata, |
| 19 | ) |
| 20 | from vllm.transformers_utils.configs.gemma4 import gemma4_layer_config |
| 21 | from vllm.utils.torch_utils import common_broadcastable_dtype |
| 22 | |
| 23 | logger = init_logger(__name__) |
| 24 | |
| 25 | |
| 26 | class ModelArchConfigConvertorBase: |
| 27 | def __init__( |
| 28 | self, |
| 29 | hf_config: PreTrainedConfig, |
| 30 | hf_text_config: PreTrainedConfig, |
| 31 | revision: str | None = None, |
| 32 | ): |
| 33 | self.hf_config = hf_config |
| 34 | self.hf_text_config = hf_text_config |
| 35 | self.revision = revision |
| 36 | |
| 37 | def get_per_layer_hf_configs( |
| 38 | self, |
| 39 | ) -> list[tuple[PreTrainedConfig, PreTrainedConfig]] | None: |
| 40 | """`(hf_config, hf_text_config)` per layer, or `None` if homogeneous. |
| 41 | |
| 42 | This is the only place that decides whether a checkpoint is heterogeneous |
| 43 | and how many layers it has, so a convertor can declare variation that |
| 44 | Transformers does not express (see `Gemma4ModelArchConfigConvertor`). |
| 45 | |
| 46 | Each pair is converted independently and the results are diffed, so a |
| 47 | convertor that overrides this must not also collapse a varying field in |
| 48 | its getters: the collapse would then be applied per layer and every layer |
| 49 | would report the same value. |
| 50 | """ |
| 51 | config_het = getattr(self.hf_config, "is_heterogeneous", False) |
| 52 | text_het = getattr(self.hf_text_config, "is_heterogeneous", False) |
| 53 | if not (config_het or text_het): |
| 54 | return None |
| 55 | # The text config decides the layer count when it is the heterogeneous one: |
| 56 | # its stack is what vLLM builds attention layers for. |
| 57 | source = self.hf_text_config if text_het else self.hf_config |
| 58 | return [ |
| 59 | ( |
| 60 | self.hf_config.per_layer_config[i] if config_het else self.hf_config, |
| 61 | self.hf_text_config.per_layer_config[i] |
| 62 | if text_het |
| 63 | else self.hf_text_config, |
| 64 | ) |
| 65 | for i in range(len(source.per_layer_config)) |
| 66 | ] |
| 67 | |
| 68 | def get_architectures(self) -> list[str]: |
| 69 | # Sometimes we get here from `vllm_config.with_hf_config(text_config)` where |
| 70 | # `text_config` is a sub-config from a multi-modal model. If this is the case, |
| 71 | # the sub-config will not have `architectures` and it will explicitly be `None` |
| 72 | return getattr(self.hf_config, "architectures", None) or [] |
| 73 | |
| 74 | def get_num_hidden_layers(self) -> int: |
| 75 | return getattr(self.hf_text_config, "num_hidden_layers", 0) |
| 76 | |
| 77 | def get_total_num_attention_heads(self) -> int: |
| 78 | return getattr(self.hf_text_config, "num_attention_heads", 0) |
| 79 | |
| 80 | def get_vocab_size(self) -> int: |
| 81 | return getattr(self.hf_text_config, "vocab_size", 0) |
| 82 | |
| 83 | def get_hidden_size(self) -> int: |
| 84 | return getattr(self.hf_text_config, "hidden_size", 0) |
| 85 | |
| 86 | def get_head_size(self) -> int: |
| 87 | if self.is_deepseek_mla(): |
| 88 | # special case for deepseek_v4 |
| 89 | if hasattr(self.hf_text_config, "compress_ratios"): |
| 90 | return self.hf_text_config.head_dim |
| 91 | qk_rope_head_dim = self._get_qk_rope_head_dim() |
| 92 | if not envs.VLLM_MLA_DISABLE: |
| 93 | return self.hf_text_config.kv_lora_rank + qk_rope_head_dim |
| 94 | else: |
| 95 | qk_nope_head_dim = getattr(self.hf_text_config, "qk_nope_head_dim", 0) |
| 96 | if qk_rope_head_dim and qk_nope_head_dim: |
| 97 | return qk_rope_head_dim + qk_nope_head_dim |
| 98 | |
| 99 | # NOTE: Some config classes may set head_dim=None or materialize a missing |
| 100 | # head_dim as 0 (for example, DeepseekVLV2TextConfig). |
| 101 | if ( |
| 102 | head_dim := getattr(self.hf_text_config, "head_dim", None) |
| 103 | ) is not None and head_dim > 0: |
| 104 | return head_dim |
| 105 | |
| 106 | # NOTE: Some models (such as PLaMo2.1) use `hidden_size_per_head` |
| 107 | if getattr(self.hf_text_config, "hidden_size_per_head", None) is not None: |
| 108 | return self.hf_text_config.hidden_size_per_head |
| 109 | |
| 110 | if (total_num_attention_heads := self.get_total_num_attention_heads()) == 0: |
| 111 | return 0 |
| 112 | # FIXME(woosuk): This may not be true for all models. |
| 113 | return self.get_hidden_size() // total_num_attention_heads |
| 114 | |
| 115 | def _get_qk_rope_head_dim(self) -> int: |
| 116 | """Get qk_rope_head_dim, fixing the transformers v5.4+ attribute_map bug.""" |
| 117 | cfg = self.hf_text_config |
| 118 | qk_rope_head_dim = getattr(cfg, "qk_rope_head_dim", 0) |
| 119 | qk_nope_head_dim = getattr(cfg, "qk_nope_head_dim", 0) |
| 120 | |
| 121 | # In valid MLA configs, qk_rope_head_dim != qk_nope_head_dim. |
| 122 | if qk_rope_head_dim == 0 or qk_rope_head_dim != qk_nope_head_dim: |
| 123 | return qk_rope_head_dim # not corrupted |
| 124 | |
| 125 | # Read the correct value from raw config.json. |
| 126 | from vllm.transformers_utils.repo_utils import get_hf_file_to_dict |
| 127 | |
| 128 | model_path = self.hf_config.name_or_path |
| 129 | if not model_path: |
| 130 | return qk_rope_head_dim |
| 131 | raw = get_hf_file_to_dict("config.json", model_path, self.revision) |
| 132 | if raw and "qk_rope_head_dim" in raw: |
| 133 | correct = raw["qk_rope_head_dim"] |
| 134 | if correct != qk_rope_head_dim: |
| 135 | logger.info( |
| 136 | "Fixing qk_rope_head_dim: %d -> %d " |
| 137 | "(transformers v5.4+ attribute_map bug)", |
| 138 | qk_rope_head_dim, |
| 139 | correct, |
| 140 | ) |
| 141 | # Patch the config so downstream model layers also get |
| 142 | # the correct value. |
| 143 | cfg.qk_rope_head_dim = correct |
| 144 | return correct |
| 145 | return qk_rope_head_dim |
| 146 | |
| 147 | def get_total_num_kv_heads(self) -> int: |
| 148 | attributes = [ |
| 149 | # For Falcon: |
| 150 | "n_head_kv", |
| 151 | "num_kv_heads", |
| 152 | # For LLaMA-2: |
| 153 | "num_key_value_heads", |
| 154 | # For ChatGLM: |
| 155 | "multi_query_group_num", |
| 156 | # For Step3p5: |
| 157 | "num_attention_groups", |
| 158 | ] |
| 159 | # For non-grouped-query attention models, the number of KV heads is |
| 160 | # equal to the number of attention heads. |
| 161 | default_factory = self.get_total_num_attention_heads |
| 162 | return getattr_iter( |
| 163 | self.hf_text_config, attributes, default_factory=default_factory |
| 164 | ) |
| 165 | |
| 166 | def get_num_experts_from_block_configs(self) -> int: |
| 167 | """Check block_configs for heterogeneous models (e.g., NemotronH). |
| 168 | |
| 169 | For heterogeneous models with varying expert counts per layer, |
| 170 | returns the MAX to ensure all expert weights can be loaded. |
| 171 | """ |
| 172 | max_experts = 0 |
| 173 | block_configs = getattr(self.hf_text_config, "block_configs", None) |
| 174 | if block_configs: |
| 175 | for block in block_configs: |
| 176 | if isinstance(block, dict): |
| 177 | if block.get("block_type", "") == "moe": |
| 178 | max_experts = max(max_experts, block.get("n_routed_experts", 0)) |
| 179 | else: |
| 180 | if getattr(block, "block_type", "") == "moe": |
| 181 | max_experts = max( |
| 182 | max_experts, getattr(block, "n_routed_experts", 0) |
| 183 | ) |
| 184 | return max_experts |
| 185 | |
| 186 | def get_num_experts(self) -> int: |
| 187 | """Returns the number of experts in the model.""" |
| 188 | num_expert_names = [ |
| 189 | "num_experts", # Jamba |
| 190 | "moe_num_experts", # Dbrx |
| 191 | "n_routed_experts", # DeepSeek |
| 192 | "num_local_experts", # Mixtral |
| 193 | ] |
| 194 | |
| 195 | num_experts = getattr_iter(self.hf_text_config, num_expert_names, 0) |
| 196 | if isinstance(num_experts, list): |
| 197 | # Ernie VL's remote code uses list[int]... |
| 198 | # The values are always the same so we just take the first one. |
| 199 | return num_experts[0] |
| 200 | |
| 201 | if not num_experts: |
| 202 | num_experts = self.get_num_experts_from_block_configs() |
| 203 | return num_experts |
| 204 | |
| 205 | def get_num_experts_per_token(self) -> int: |
| 206 | names = [ |
| 207 | "num_experts_per_tok", |
| 208 | "num_experts_per_token", |
| 209 | "top_k_experts", |
| 210 | "moe_topk", |
| 211 | "moe_top_k", |
| 212 | ] |
| 213 | num_experts_per_token = getattr_iter(self.hf_text_config, names, 0) |
| 214 | if isinstance(num_experts_per_token, list): |
| 215 | return max(num_experts_per_token, default=0) |
| 216 | return num_experts_per_token or 0 |
| 217 | |
| 218 | @final |
| 219 | @classmethod |
| 220 | def get_torch_dtype( |
| 221 | cls, |
| 222 | hf_config: PreTrainedConfig, |
| 223 | model_id: str, |
| 224 | revision: str | None, |
| 225 | config_format: str | ConfigFormat, |
| 226 | ): |
| 227 | # NOTE: getattr(config, "dtype", torch.float32) is not correct |
| 228 | # because config.dtype can be None. |
| 229 | config_dtype = getattr(hf_config, "dtype", None) |
| 230 | |
| 231 | # Fallbacks for multi-modal models if the root config |
| 232 | # does not define dtype |
| 233 | if config_dtype is None: |
| 234 | config_dtype = getattr(hf_config.get_text_config(), "dtype", None) |
| 235 | if config_dtype is None and hasattr(hf_config, "vision_config"): |
| 236 | config_dtype = getattr(hf_config.vision_config, "dtype", None) |
| 237 | if config_dtype is None and hasattr(hf_config, "encoder_config"): |
| 238 | config_dtype = getattr(hf_config.encoder_config, "dtype", None) |
| 239 | |
| 240 | # Try to read the dtype of the weights if they are in safetensors format |
| 241 | if config_dtype is None: |
| 242 | param_mt = get_safetensors_params_metadata(model_id, revision=revision) |
| 243 | |
| 244 | if param_mt: |
| 245 | param_dtypes: set[torch.dtype] = { |
| 246 | _SAFETENSORS_TO_TORCH_DTYPE[dtype] |
| 247 | for info in param_mt.values() |
| 248 | if (dtype := info.get("dtype", None)) |
| 249 | and dtype in _SAFETENSORS_TO_TORCH_DTYPE |
| 250 | } |
| 251 | |
| 252 | if param_dtypes: |
| 253 | return common_broadcastable_dtype(param_dtypes) |
| 254 | |
| 255 | if config_dtype is None: |
| 256 | config_dtype = torch.float32 |
| 257 | |
| 258 | return config_dtype |
| 259 | |
| 260 | def _normalize_quantization_config(self, config: PreTrainedConfig): |
| 261 | quant_cfg = getattr(config, "quantization_config", None) |
| 262 | if quant_cfg is None: |
| 263 | # compressed-tensors uses a "compression_config" key |
| 264 | quant_cfg = getattr(config, "compression_config", None) |
| 265 | |
| 266 | else: |
| 267 | # Set quant_method for ModelOpt models. |
| 268 | producer_name = quant_cfg.get("producer", {}).get("name") |
| 269 | modelopt_quant_cfg = quant_cfg.get("quantization", {}) |
| 270 | is_legacy_modelopt = ( |
| 271 | isinstance(modelopt_quant_cfg, dict) |
| 272 | and "modelopt_quant_config" in modelopt_quant_cfg |
| 273 | ) |
| 274 | if producer_name == "modelopt" or is_legacy_modelopt: |
| 275 | quant_algo = modelopt_quant_cfg.get("quant_algo") |
| 276 | if quant_algo is not None: |
| 277 | quant_algo_upper = str(quant_algo).upper() |
| 278 | if quant_algo_upper in { |
| 279 | "FP8", |
| 280 | "FP8_PER_CHANNEL_PER_TOKEN", |
| 281 | "FP8_PB_WO", |
| 282 | }: |
| 283 | quant_cfg["quant_method"] = "modelopt" |
| 284 | elif quant_algo_upper == "NVFP4": |
| 285 | quant_cfg["quant_method"] = "modelopt_fp4" |
| 286 | else: |
| 287 | raise ValueError(f"Unknown ModelOpt quant algo: {quant_algo}") |
| 288 | |
| 289 | if quant_cfg is not None: |
| 290 | # Use the community standard 'quant_method' |
| 291 | quant_method = quant_cfg.get("quant_method", "").lower() |
| 292 | |
| 293 | # Normalize library names |
| 294 | quant_method = quant_method.replace( |
| 295 | "compressed_tensors", "compressed-tensors" |
| 296 | ) |
| 297 | |
| 298 | quant_cfg["quant_method"] = quant_method |
| 299 | |
| 300 | return quant_cfg |
| 301 | |
| 302 | def get_quantization_config(self): |
| 303 | quant_cfg = self._normalize_quantization_config(self.hf_config) |
| 304 | if quant_cfg is None and ( |
| 305 | text_config := getattr(self.hf_config, "text_config", None) |
| 306 | ): |
| 307 | # Check the text config as well for multi-modal models. |
| 308 | quant_cfg = self._normalize_quantization_config(text_config) |
| 309 | return quant_cfg |
| 310 | |
| 311 | def is_deepseek_mla(self) -> bool: |
| 312 | if not hasattr(self.hf_text_config, "model_type"): |
| 313 | return False |
| 314 | elif self.hf_text_config.model_type in ( |
| 315 | "axk1", |
| 316 | "deepseek_v2", |
| 317 | "deepseek_v3", |
| 318 | "deepseek_v32", |
| 319 | "deepseek_v4", |
| 320 | "deepseek_v41", |
| 321 | "dots3_note", |
| 322 | "deepseek_mtp", |
| 323 | "k3_dspark", |
| 324 | "glm5_next", |
| 325 | "glm5_next_text", |
| 326 | "glm_moe_dsa", |
| 327 | "glm4_moe_lite", |
| 328 | "glm4_moe_lite_mtp", |
| 329 | "kimi_k2", |
| 330 | "kimi_linear", |
| 331 | "longcat_flash", |
| 332 | "longcat_flash_ngram", |
| 333 | "pangu_ultra_moe", |
| 334 | "pangu_ultra_moe_mtp", |
| 335 | "bailing_hybrid", |
| 336 | "bailing_hybrid_mtp", |
| 337 | "bailing_hybrid_v3_mtp", |
| 338 | "hy_v4", |
| 339 | "hy_v4_mtp", |
| 340 | ): |
| 341 | # check is deepseek_v4 model |
| 342 | if hasattr(self.hf_text_config, "compress_ratios"): |
| 343 | return getattr(self.hf_text_config, "head_dim", None) is not None |
| 344 | else: |
| 345 | return getattr(self.hf_text_config, "kv_lora_rank", None) is not None |
| 346 | elif self.hf_text_config.model_type == "eagle": |
| 347 | # if the model is an EAGLE module, check for the |
| 348 | # underlying architecture |
| 349 | return ( |
| 350 | self.hf_text_config.model.model_type |
| 351 | in ( |
| 352 | "axk1", |
| 353 | "deepseek_v2", |
| 354 | "deepseek_v3", |
| 355 | "deepseek_v32", |
| 356 | "deepseek_mtp", |
| 357 | ) |
| 358 | and getattr(self.hf_text_config, "kv_lora_rank", None) is not None |
| 359 | ) |
| 360 | return False |
| 361 | |
| 362 | def is_mm_prefix_lm(self, supports_multimodal: bool = True) -> bool: |
| 363 | """Whether to use bidirectional attention for mm positions. |
| 364 | |
| 365 | ``supports_multimodal`` is False when the deployment is configuration- |
| 366 | disabled for multimodal inputs (text-only serving). In that case |
| 367 | mm_prefix is unnecessary and must stay off so attention backends |
| 368 | without ``supports_mm_prefix()`` remain eligible. |
| 369 | """ |
| 370 | if not supports_multimodal: |
| 371 | return False |
| 372 | if hasattr(self.hf_config, "is_mm_prefix_lm"): |
| 373 | return bool(self.hf_config.is_mm_prefix_lm) |
| 374 | # fallback to list of known models |
| 375 | MM_PREFIX_LM_MODELS = ( |
| 376 | "bagel", |
| 377 | "gemma3", |
| 378 | "molmo2", |
| 379 | "moondream3", |
| 380 | "paligemma", |
| 381 | ) |
| 382 | if not hasattr(self.hf_config, "model_type"): |
| 383 | return False |
| 384 | return self.hf_config.model_type in MM_PREFIX_LM_MODELS |
| 385 | |
| 386 | def rswa_window(self) -> int | None: |
| 387 | value = getattr(self.hf_config, "rswa_window", None) |
| 388 | if value is None: |
| 389 | return None |
| 390 | return int(value) |
| 391 | |
| 392 | def derive_max_model_len_and_key(self) -> tuple[float, str | None]: |
| 393 | derived_max_model_len = float("inf") |
| 394 | possible_keys = [ |
| 395 | # OPT |
| 396 | "max_position_embeddings", |
| 397 | # GPT-2 |
| 398 | "n_positions", |
| 399 | # MPT |
| 400 | "max_seq_len", |
| 401 | # ChatGLM2 |
| 402 | "seq_length", |
| 403 | # Command-R |
| 404 | "model_max_length", |
| 405 | # Whisper |
| 406 | "max_target_positions", |
| 407 | # Others |
| 408 | "max_sequence_length", |
| 409 | "max_seq_length", |
| 410 | "seq_len", |
| 411 | ] |
| 412 | # Choose the smallest "max_length" from the possible keys |
| 413 | max_len_key = None |
| 414 | for key in possible_keys: |
| 415 | max_len = getattr(self.hf_text_config, key, None) |
| 416 | if max_len is not None: |
| 417 | if max_len < derived_max_model_len: |
| 418 | max_len_key = key |
| 419 | derived_max_model_len = min(derived_max_model_len, max_len) |
| 420 | |
| 421 | # For Command-R / Cohere, Cohere2 models |
| 422 | if tmp_max_len := getattr(self.hf_text_config, "model_max_length", None): |
| 423 | max_len_key = "model_max_length" |
| 424 | derived_max_model_len = tmp_max_len |
| 425 | return derived_max_model_len, max_len_key |
| 426 | |
| 427 | def convert(self, supports_multimodal: bool = True) -> ModelArchitectureConfig: |
| 428 | if (per_layer := self.get_per_layer_hf_configs()) is None: |
| 429 | return self.convert_layer(supports_multimodal) |
| 430 | |
| 431 | if self.is_deepseek_mla(): |
| 432 | raise NotImplementedError( |
| 433 | "Heterogeneous MLA models are not supported: `get_head_size` " |
| 434 | "patches the config in place for them, which a per-layer " |
| 435 | "conversion would apply to a throwaway copy." |
| 436 | ) |
| 437 | # Convert each layer and let the result work out which fields differ. |
| 438 | # Reading a varying attribute off the global config would raise, so the |
| 439 | # whole-model config has to be built up from the layers. |
| 440 | return ModelArchitectureConfig.from_layers( |
| 441 | [ |
| 442 | type(self)(*configs).convert_layer(supports_multimodal) |
| 443 | for configs in per_layer |
| 444 | ] |
| 445 | ) |
| 446 | |
| 447 | def convert_layer( |
| 448 | self, supports_multimodal: bool = True |
| 449 | ) -> ModelArchitectureConfig: |
| 450 | """Convert one homogeneous config, without resolving per-layer values. |
| 451 | |
| 452 | `convert` calls this once per layer, so it must not recurse back into |
| 453 | per-layer resolution: the configs it receives are already layer-specific, |
| 454 | and they can still carry the attributes that made the model look |
| 455 | heterogeneous in the first place. |
| 456 | """ |
| 457 | model_arch_config = ModelArchitectureConfig( |
| 458 | architectures=self.get_architectures(), |
| 459 | model_type=self.hf_config.model_type, |
| 460 | text_model_type=getattr(self.hf_text_config, "model_type", None), |
| 461 | hidden_size=self.get_hidden_size(), |
| 462 | total_num_hidden_layers=self.get_num_hidden_layers(), |
| 463 | total_num_attention_heads=self.get_total_num_attention_heads(), |
| 464 | head_size=self.get_head_size(), |
| 465 | vocab_size=self.get_vocab_size(), |
| 466 | total_num_kv_heads=self.get_total_num_kv_heads(), |
| 467 | num_experts=self.get_num_experts(), |
| 468 | num_experts_per_token=self.get_num_experts_per_token(), |
| 469 | quantization_config=self.get_quantization_config(), |
| 470 | is_deepseek_mla=self.is_deepseek_mla(), |
| 471 | is_mm_prefix_lm=self.is_mm_prefix_lm(supports_multimodal), |
| 472 | rswa_window=self.rswa_window(), |
| 473 | derived_max_model_len_and_key=self.derive_max_model_len_and_key(), |
| 474 | ) |
| 475 | |
| 476 | return model_arch_config |
| 477 | |
| 478 | |
| 479 | class CohereAsrModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 480 | def get_total_num_attention_heads(self) -> int: |
| 481 | return self.hf_text_config.transf_decoder["config_dict"]["num_attention_heads"] |
| 482 | |
| 483 | def get_head_size(self) -> int: |
| 484 | hidden_size = self.hf_text_config.transf_decoder["config_dict"]["hidden_size"] |
| 485 | num_attention_heads = self.hf_text_config.transf_decoder["config_dict"][ |
| 486 | "num_attention_heads" |
| 487 | ] |
| 488 | return hidden_size // num_attention_heads |
| 489 | |
| 490 | def get_total_num_kv_heads(self) -> int: |
| 491 | enc_num_kv_heads = self.hf_text_config.encoder["n_heads"] |
| 492 | dec_num_kv_heads = self.hf_text_config.transf_decoder["config_dict"][ |
| 493 | "num_attention_heads" |
| 494 | ] |
| 495 | assert enc_num_kv_heads == dec_num_kv_heads, ( |
| 496 | "Encoder and decoder must have the same number of kv heads" |
| 497 | ) |
| 498 | return enc_num_kv_heads |
| 499 | |
| 500 | def is_mm_prefix_lm(self, supports_multimodal: bool = True) -> bool: |
| 501 | return False |
| 502 | |
| 503 | |
| 504 | class MambaModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 505 | def get_head_size(self) -> int: |
| 506 | return 0 |
| 507 | |
| 508 | def get_total_num_kv_heads(self) -> int: |
| 509 | return 0 |
| 510 | |
| 511 | |
| 512 | class TerratorchModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 513 | def get_head_size(self) -> int: |
| 514 | return 0 |
| 515 | |
| 516 | def get_total_num_kv_heads(self) -> int: |
| 517 | return 0 |
| 518 | |
| 519 | |
| 520 | class MedusaModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 521 | def get_head_size(self) -> int: |
| 522 | return 0 |
| 523 | |
| 524 | def get_total_num_kv_heads(self) -> int: |
| 525 | return 0 |
| 526 | |
| 527 | |
| 528 | class Zamba2ModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 529 | def get_head_size(self) -> int: |
| 530 | return getattr(self.hf_text_config, "attention_head_dim", 0) |
| 531 | |
| 532 | |
| 533 | class FalconModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 534 | def get_total_num_kv_heads(self) -> int: |
| 535 | # NOTE: for falcon, when new_decoder_architecture is True, the |
| 536 | # multi_query flag is ignored and we use n_head_kv for the number of |
| 537 | # KV heads. |
| 538 | new_decoder_arch_falcon = getattr( |
| 539 | self.hf_text_config, "new_decoder_architecture", False |
| 540 | ) |
| 541 | |
| 542 | if not new_decoder_arch_falcon and getattr( |
| 543 | self.hf_text_config, "multi_query", False |
| 544 | ): |
| 545 | # Multi-query attention, only one KV head. |
| 546 | return 1 |
| 547 | |
| 548 | # Use the base implementation which checks n_head_kv, num_kv_heads, etc. |
| 549 | return super().get_total_num_kv_heads() |
| 550 | |
| 551 | |
| 552 | class DbrxModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 553 | def get_total_num_kv_heads(self) -> int: |
| 554 | return getattr( |
| 555 | self.hf_text_config.attn_config, |
| 556 | "kv_n_heads", |
| 557 | self.hf_text_config.num_attention_heads, |
| 558 | ) |
| 559 | |
| 560 | |
| 561 | class NemotronNasModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 562 | def get_total_num_kv_heads(self) -> int: |
| 563 | for block in self.hf_text_config.block_configs: |
| 564 | if not block.attention.no_op: |
| 565 | return ( |
| 566 | self.hf_text_config.num_attention_heads |
| 567 | // block.attention.n_heads_in_group |
| 568 | ) |
| 569 | raise RuntimeError( |
| 570 | "Could not determine the number of key-value attention heads " |
| 571 | "from model configuration. " |
| 572 | f"Architecture: {self.get_architectures()}. " |
| 573 | "This usually indicates an unsupported model architecture or " |
| 574 | "missing configuration. " |
| 575 | "Please check if your model is supported at: " |
| 576 | "https://docs.vllm.ai/en/latest/models/supported_models.html" |
| 577 | ) |
| 578 | |
| 579 | |
| 580 | class DeepSeekMTPModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 581 | def get_num_hidden_layers(self) -> int: |
| 582 | return getattr(self.hf_text_config, "num_nextn_predict_layers", 0) |
| 583 | |
| 584 | |
| 585 | class DeepseekV4ModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 586 | def __init__( |
| 587 | self, |
| 588 | hf_config: PreTrainedConfig, |
| 589 | hf_text_config: PreTrainedConfig, |
| 590 | revision: str | None = None, |
| 591 | ): |
| 592 | # DeepSeek-V4-Flash-Vision-Exp ships the same architectures/model_type |
| 593 | # as the text-only DeepSeek-V4-Flash; route to the VL wrapper class |
| 594 | # when the config carries a vision tower. Mutate (not just override |
| 595 | # get_architectures) because model-class resolution reads the raw |
| 596 | # hf_config.architectures (get_model_architecture). |
| 597 | # Only rewrite the stock text architecture: speculative-draft configs |
| 598 | # (DSparkDraftModel / DeepSeekV4MTPModel) keep their own classes, and |
| 599 | # the VL wrapper marks the config copy it hands to the inner text |
| 600 | # backbone with _dsv4_vl_inner so this rewrite does not recurse. |
| 601 | if ( |
| 602 | getattr(hf_config, "vision_n_layers", 0) > 0 |
| 603 | and getattr(hf_config, "architectures", None) == ["DeepseekV4ForCausalLM"] |
| 604 | and not getattr(hf_config, "_dsv4_vl_inner", False) |
| 605 | ): |
| 606 | hf_config.architectures = ["DeepseekV4ForConditionalGeneration"] |
| 607 | super().__init__(hf_config, hf_text_config, revision) |
| 608 | |
| 609 | def is_mm_prefix_lm(self, supports_multimodal: bool = True) -> bool: |
| 610 | # The vision variant needs the mm-prefix plumbing: it makes the |
| 611 | # scheduler prefill image spans atomically (disable_chunked_mm_input) |
| 612 | # and routes per-request image ranges to the sparse-SWA metadata |
| 613 | # builder, which widens the sliding window bidirectionally in-kernel. |
| 614 | if not supports_multimodal: |
| 615 | return False |
| 616 | return getattr(self.hf_config, "vision_n_layers", 0) > 0 |
| 617 | |
| 618 | |
| 619 | class MimoMTPModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 620 | def get_num_hidden_layers(self) -> int: |
| 621 | return getattr(self.hf_text_config, "num_nextn_predict_layers", 0) |
| 622 | |
| 623 | |
| 624 | def _strip_mimo_v2_attention_chunk_size( |
| 625 | hf_config: PreTrainedConfig, hf_text_config: PreTrainedConfig |
| 626 | ) -> None: |
| 627 | # MiMo-V2-Flash's config.json sets `attention_chunk_size=128` but the |
| 628 | # architecture does not actually use chunked local attention. Leaving it |
| 629 | # set makes vLLM disable the hybrid KV cache manager |
| 630 | for cfg in (hf_text_config, hf_config): |
| 631 | if cfg is not None and hasattr(cfg, "attention_chunk_size"): |
| 632 | delattr(cfg, "attention_chunk_size") |
| 633 | |
| 634 | |
| 635 | class MimoV2ModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 636 | def __init__( |
| 637 | self, |
| 638 | hf_config: PreTrainedConfig, |
| 639 | hf_text_config: PreTrainedConfig, |
| 640 | revision: str | None = None, |
| 641 | ): |
| 642 | if getattr(hf_config, "vision_config", None): |
| 643 | hf_config.architectures = ["MiMoV2OmniForCausalLM"] |
| 644 | super().__init__(hf_config, hf_text_config, revision) |
| 645 | _strip_mimo_v2_attention_chunk_size(hf_config, hf_text_config) |
| 646 | |
| 647 | |
| 648 | class MimoV2MTPModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 649 | def __init__( |
| 650 | self, |
| 651 | hf_config: PreTrainedConfig, |
| 652 | hf_text_config: PreTrainedConfig, |
| 653 | revision: str | None = None, |
| 654 | ): |
| 655 | super().__init__(hf_config, hf_text_config, revision) |
| 656 | _strip_mimo_v2_attention_chunk_size(hf_config, hf_text_config) |
| 657 | |
| 658 | def get_num_hidden_layers(self) -> int: |
| 659 | n = getattr(self.hf_text_config, "num_nextn_predict_layers", None) |
| 660 | if n is not None: |
| 661 | return n |
| 662 | # Fall back to n_predict set by hf_config_override |
| 663 | n = getattr(self.hf_text_config, "n_predict", None) |
| 664 | return n if n is not None else 0 |
| 665 | |
| 666 | |
| 667 | class GLM4MoeMTPModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 668 | def get_num_hidden_layers(self) -> int: |
| 669 | return getattr(self.hf_text_config, "num_nextn_predict_layers", 0) |
| 670 | |
| 671 | |
| 672 | class ErnieMTPModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 673 | def get_num_hidden_layers(self) -> int: |
| 674 | return getattr(self.hf_text_config, "num_nextn_predict_layers", 0) |
| 675 | |
| 676 | |
| 677 | class Qwen3NextMTPModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 678 | def get_num_hidden_layers(self) -> int: |
| 679 | return getattr(self.hf_text_config, "num_nextn_predict_layers", 0) |
| 680 | |
| 681 | |
| 682 | class Qwen4ExpMTPModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 683 | def get_num_hidden_layers(self) -> int: |
| 684 | return getattr( |
| 685 | self.hf_text_config, |
| 686 | "mtp_num_hidden_layers", |
| 687 | getattr(self.hf_text_config, "num_nextn_predict_layers", 0), |
| 688 | ) |
| 689 | |
| 690 | |
| 691 | class BailingHybridMTPModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 692 | def get_num_hidden_layers(self) -> int: |
| 693 | return getattr(self.hf_text_config, "num_nextn_predict_layers", 0) |
| 694 | |
| 695 | |
| 696 | class BailingHybridV3MTPModelArchConfigConvertor( |
| 697 | BailingHybridMTPModelArchConfigConvertor |
| 698 | ): |
| 699 | pass |
| 700 | |
| 701 | |
| 702 | class Qwen3_5MTPModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 703 | def get_num_hidden_layers(self) -> int: |
| 704 | return getattr(self.hf_text_config, "mtp_num_hidden_layers", 0) |
| 705 | |
| 706 | |
| 707 | class Step3p5MTPModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 708 | def get_num_hidden_layers(self) -> int: |
| 709 | return getattr(self.hf_text_config, "num_nextn_predict_layers", 0) |
| 710 | |
| 711 | |
| 712 | class PanguUltraMoeMTPModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 713 | def get_num_hidden_layers(self) -> int: |
| 714 | return getattr(self.hf_text_config, "num_nextn_predict_layers", 0) |
| 715 | |
| 716 | |
| 717 | class LongCatFlashMTPModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 718 | def get_num_hidden_layers(self) -> int: |
| 719 | return getattr(self.hf_text_config, "num_nextn_predict_layers", 1) |
| 720 | |
| 721 | |
| 722 | class Gemma4MTPModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 723 | def get_hidden_size(self) -> int: |
| 724 | # The speculator buffer must match the backbone (target) model's |
| 725 | # hidden dimension, not the draft model's smaller dimension. |
| 726 | return getattr( |
| 727 | self.hf_config, "backbone_hidden_size", super().get_hidden_size() |
| 728 | ) |
| 729 | |
| 730 | def get_num_hidden_layers(self) -> int: |
| 731 | return getattr(self.hf_text_config, "num_hidden_layers", 0) |
| 732 | |
| 733 | |
| 734 | class Gemma4ModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 735 | def is_mm_prefix_lm(self, supports_multimodal: bool = True) -> bool: |
| 736 | if not supports_multimodal: |
| 737 | return False |
| 738 | return ( |
| 739 | getattr(self.hf_text_config, "use_bidirectional_attention", None) |
| 740 | == "vision" |
| 741 | ) |
| 742 | |
| 743 | def get_per_layer_hf_configs( |
| 744 | self, |
| 745 | ) -> list[tuple[PreTrainedConfig, PreTrainedConfig]] | None: |
| 746 | # Gemma4 uses a larger head dimension, and sometimes more KV heads, on its |
| 747 | # full attention layers than on its sliding ones. Transformers >= 5.15.0 |
| 748 | # says so in the config; before that the values are flat attributes picked |
| 749 | # apart by `layer_types`, so build the per-layer configs here instead. Both |
| 750 | # then reach the base convertor as ordinary homogeneous configs. |
| 751 | if getattr(self.hf_text_config, "is_heterogeneous", False): |
| 752 | return super().get_per_layer_hf_configs() |
| 753 | text_config = self.hf_text_config |
| 754 | if not getattr(text_config, "layer_types", None): |
| 755 | return None |
| 756 | # `num_hidden_layers` is what the stack is built from, and it is what |
| 757 | # Transformers sizes `per_layer_config` by. `layer_types` can be longer, |
| 758 | # e.g. when a test truncates the layer count to build a small model. |
| 759 | return [ |
| 760 | (self.hf_config, gemma4_layer_config(text_config, layer_idx)) |
| 761 | for layer_idx in range(self.get_num_hidden_layers()) |
| 762 | ] |
| 763 | |
| 764 | |
| 765 | class MossAudioModelArchConfigConvertor(ModelArchConfigConvertorBase): |
| 766 | def _language_config(self) -> PreTrainedConfig: |
| 767 | return self.hf_config.language_config |
| 768 | |
| 769 | def get_num_hidden_layers(self) -> int: |
| 770 | return getattr(self._language_config(), "num_hidden_layers", 0) |
| 771 | |
| 772 | def get_total_num_attention_heads(self) -> int: |
| 773 | return getattr(self._language_config(), "num_attention_heads", 0) |
| 774 | |
| 775 | def get_vocab_size(self) -> int: |
| 776 | return getattr(self._language_config(), "vocab_size", 0) |
| 777 | |
| 778 | def get_hidden_size(self) -> int: |
| 779 | return getattr(self._language_config(), "hidden_size", 0) |
| 780 | |
| 781 | def get_head_size(self) -> int: |
| 782 | head_dim = getattr(self._language_config(), "head_dim", None) |
| 783 | if head_dim is not None: |
| 784 | return head_dim |
| 785 | total_num_attention_heads = self.get_total_num_attention_heads() |
| 786 | if total_num_attention_heads == 0: |
| 787 | return 0 |
| 788 | return self.get_hidden_size() // total_num_attention_heads |
| 789 | |
| 790 | def get_total_num_kv_heads(self) -> int: |
| 791 | return getattr( |
| 792 | self._language_config(), |
| 793 | "num_key_value_heads", |
| 794 | self.get_total_num_attention_heads(), |
| 795 | ) |
| 796 | |
| 797 | def derive_max_model_len_and_key(self) -> tuple[float, str | None]: |
| 798 | language_config = self._language_config() |
| 799 | max_position_embeddings = getattr( |
| 800 | language_config, |
| 801 | "max_position_embeddings", |
| 802 | None, |
| 803 | ) |
| 804 | if max_position_embeddings is None: |
| 805 | return super().derive_max_model_len_and_key() |
| 806 | return max_position_embeddings, "language_config.max_position_embeddings" |
| 807 | |
| 808 | |
| 809 | # hf_config.model_type -> convertor class |
| 810 | MODEL_ARCH_CONFIG_CONVERTORS = { |
| 811 | "bailing_hybrid_mtp": BailingHybridMTPModelArchConfigConvertor, |
| 812 | "cohere_asr": CohereAsrModelArchConfigConvertor, |
| 813 | "dbrx": DbrxModelArchConfigConvertor, |
| 814 | "deepseek_mtp": DeepSeekMTPModelArchConfigConvertor, |
| 815 | "deepseek_v4": DeepseekV4ModelArchConfigConvertor, |
| 816 | "diffusion_gemma": Gemma4ModelArchConfigConvertor, |
| 817 | "diffusion_gemma_text": Gemma4ModelArchConfigConvertor, |
| 818 | "ernie_mtp": ErnieMTPModelArchConfigConvertor, |
| 819 | "falcon": FalconModelArchConfigConvertor, |
| 820 | "falcon_mamba": MambaModelArchConfigConvertor, |
| 821 | "gemma4": Gemma4ModelArchConfigConvertor, |
| 822 | "gemma4_mtp": Gemma4MTPModelArchConfigConvertor, |
| 823 | "gemma4_text": Gemma4ModelArchConfigConvertor, |
| 824 | "gemma4_unified": Gemma4ModelArchConfigConvertor, |
| 825 | "gemma4_unified_text": Gemma4ModelArchConfigConvertor, |
| 826 | "glm4_moe_mtp": GLM4MoeMTPModelArchConfigConvertor, |
| 827 | "glm_ocr_mtp": GLM4MoeMTPModelArchConfigConvertor, |
| 828 | "longcat_flash_mtp": LongCatFlashMTPModelArchConfigConvertor, |
| 829 | "mamba": MambaModelArchConfigConvertor, |
| 830 | "medusa": MedusaModelArchConfigConvertor, |
| 831 | "mimo_mtp": MimoMTPModelArchConfigConvertor, |
| 832 | "mimo_v2": MimoV2ModelArchConfigConvertor, |
| 833 | "mimo_v2_flash": MimoV2ModelArchConfigConvertor, |
| 834 | "mimo_v2_mtp": MimoV2MTPModelArchConfigConvertor, |
| 835 | "mimo_v2_omni_mtp": MimoV2MTPModelArchConfigConvertor, |
| 836 | "moss_audio": MossAudioModelArchConfigConvertor, |
| 837 | "nemotron-nas": NemotronNasModelArchConfigConvertor, |
| 838 | "bailing_hybrid_v3_mtp": BailingHybridV3MTPModelArchConfigConvertor, |
| 839 | "pangu_ultra_moe_mtp": PanguUltraMoeMTPModelArchConfigConvertor, |
| 840 | "qwen3_5_mtp": Qwen3_5MTPModelArchConfigConvertor, |
| 841 | "qwen3_next_mtp": Qwen3NextMTPModelArchConfigConvertor, |
| 842 | "qwen4_exp_mtp": Qwen4ExpMTPModelArchConfigConvertor, |
| 843 | "RefinedWeb": FalconModelArchConfigConvertor, |
| 844 | "RefinedWebModel": FalconModelArchConfigConvertor, |
| 845 | "step3p5_mtp": Step3p5MTPModelArchConfigConvertor, |
| 846 | "timm_wrapper": TerratorchModelArchConfigConvertor, |
| 847 | "zamba2": Zamba2ModelArchConfigConvertor, |
| 848 | } |