| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Inference-only Qwen3_5 MTP model.""" |
| 4 | |
| 5 | from collections.abc import Iterable |
| 6 | |
| 7 | import torch |
| 8 | from torch import nn |
| 9 | |
| 10 | from vllm.compilation.decorators import support_torch_compile |
| 11 | from vllm.config import VllmConfig |
| 12 | from vllm.distributed import get_pp_group, tensor_model_parallel_all_gather |
| 13 | from vllm.logger import init_logger |
| 14 | from vllm.model_executor.layers.fused_moe.utils import ( |
| 15 | is_model_fused_shared_expert_compatible, |
| 16 | ) |
| 17 | from vllm.model_executor.layers.linear import ColumnParallelLinear |
| 18 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 19 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 20 | ParallelLMHead, |
| 21 | VocabParallelEmbedding, |
| 22 | ) |
| 23 | from vllm.model_executor.models.interfaces import LocalArgmaxMixin |
| 24 | from vllm.model_executor.models.qwen3_5 import ( |
| 25 | Qwen3_5DecoderLayer, |
| 26 | Qwen3_5Model, |
| 27 | Qwen3_5RMSNorm, |
| 28 | ) |
| 29 | from vllm.model_executor.models.qwen3_next import ( |
| 30 | Qwen3NextSparseMoeBlock, |
| 31 | QwenNextMixtureOfExperts, |
| 32 | ) |
| 33 | from vllm.model_executor.models.utils import sequence_parallel_chunk |
| 34 | from vllm.sequence import IntermediateTensors |
| 35 | from vllm.transformers_utils.configs.qwen3_5 import Qwen3_5TextConfig |
| 36 | from vllm.transformers_utils.configs.qwen3_5_moe import Qwen3_5MoeTextConfig |
| 37 | |
| 38 | from .interfaces import ( |
| 39 | MultiModalEmbeddings, |
| 40 | SupportsMultiModal, |
| 41 | SupportsPP, |
| 42 | _require_is_multimodal, |
| 43 | ) |
| 44 | from .utils import ( |
| 45 | AutoWeightsLoader, |
| 46 | PPMissingLayer, |
| 47 | _merge_multimodal_embeddings, |
| 48 | make_empty_intermediate_tensors_factory, |
| 49 | maybe_fuse_shared_experts, |
| 50 | maybe_prefix, |
| 51 | ) |
| 52 | |
| 53 | logger = init_logger(__name__) |
| 54 | |
| 55 | |
| 56 | @support_torch_compile( |
| 57 | dynamic_arg_dims={ |
| 58 | "input_ids": 0, |
| 59 | # positions is of shape (3, seq_len) if mrope is enabled for qwen2-vl, |
| 60 | # otherwise (seq_len, ). |
| 61 | "positions": -1, |
| 62 | "intermediate_tensors": 0, |
| 63 | "inputs_embeds": 0, |
| 64 | "hidden_states": 0, |
| 65 | } |
| 66 | ) |
| 67 | class Qwen3_5MultiTokenPredictor(nn.Module): |
| 68 | hf_to_vllm_mapper = Qwen3_5Model.hf_to_vllm_mapper |
| 69 | |
| 70 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 71 | super().__init__() |
| 72 | |
| 73 | model_config = vllm_config.model_config |
| 74 | quant_config = vllm_config.quant_config |
| 75 | |
| 76 | config: Qwen3_5TextConfig | Qwen3_5MoeTextConfig = model_config.hf_text_config |
| 77 | |
| 78 | self.config = config |
| 79 | |
| 80 | self.vocab_size = config.vocab_size |
| 81 | |
| 82 | self.mtp_start_layer_idx = config.num_hidden_layers |
| 83 | self.num_mtp_layers = getattr(config, "mtp_num_hidden_layers", 1) |
| 84 | |
| 85 | self.embed_tokens = VocabParallelEmbedding( |
| 86 | self.vocab_size, |
| 87 | config.hidden_size, |
| 88 | ) |
| 89 | |
| 90 | # Workaround: mtp.fc is stored as BF16 in NVFP4 checkpoints but is |
| 91 | # missing from hf_quant_config.json exclude_modules. Force unquantized. |
| 92 | # Ref: https://github.com/vllm-project/vllm/pull/38650 |
| 93 | # Ref: https://github.com/NVIDIA/Model-Optimizer/pull/1124 |
| 94 | fc_quant = ( |
| 95 | None |
| 96 | if (quant_config and quant_config.get_name() == "modelopt_fp4") |
| 97 | else quant_config |
| 98 | ) |
| 99 | self.fc = ColumnParallelLinear( |
| 100 | self.config.hidden_size * 2, |
| 101 | self.config.hidden_size, |
| 102 | gather_output=True, |
| 103 | bias=False, |
| 104 | return_bias=False, |
| 105 | quant_config=fc_quant, |
| 106 | prefix=f"{prefix}.fc", |
| 107 | ) |
| 108 | |
| 109 | # GPTQ: quantized checkpoints may exclude MTP from quantization via |
| 110 | # quantization_config.dynamic with "-:pattern" entries. When detected, |
| 111 | # disable quantization for MTP layers so they use unquantized params. |
| 112 | original_quant = vllm_config.quant_config |
| 113 | if quant_config and quant_config.get_name() not in ("modelopt_fp4",): |
| 114 | hf_qc = getattr(model_config.hf_config, "quantization_config", None) |
| 115 | if isinstance(hf_qc, dict): |
| 116 | dynamic = hf_qc.get("dynamic", {}) |
| 117 | if any(k.startswith("-:") and "mtp" in k for k in dynamic): |
| 118 | vllm_config.quant_config = None |
| 119 | self.layers = torch.nn.ModuleList( |
| 120 | Qwen3_5DecoderLayer( |
| 121 | vllm_config, |
| 122 | layer_type="full_attention", |
| 123 | prefix=f"{prefix}.layers.{idx}", |
| 124 | ) |
| 125 | for idx in range(self.num_mtp_layers) |
| 126 | ) |
| 127 | vllm_config.quant_config = original_quant |
| 128 | self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible( |
| 129 | self.layers, |
| 130 | Qwen3NextSparseMoeBlock, |
| 131 | "mlp", |
| 132 | ) |
| 133 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 134 | ["hidden_states", "residual"], config.hidden_size |
| 135 | ) |
| 136 | self.norm = Qwen3_5RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 137 | self.pre_fc_norm_hidden = Qwen3_5RMSNorm( |
| 138 | config.hidden_size, eps=config.rms_norm_eps |
| 139 | ) |
| 140 | self.pre_fc_norm_embedding = Qwen3_5RMSNorm( |
| 141 | config.hidden_size, eps=config.rms_norm_eps |
| 142 | ) |
| 143 | |
| 144 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 145 | return self.embed_tokens(input_ids) |
| 146 | |
| 147 | def forward( |
| 148 | self, |
| 149 | input_ids: torch.Tensor, |
| 150 | positions: torch.Tensor, |
| 151 | hidden_states: torch.Tensor, |
| 152 | intermediate_tensors: IntermediateTensors | None = None, |
| 153 | inputs_embeds: torch.Tensor | None = None, |
| 154 | spec_step_idx: int = 0, |
| 155 | ) -> torch.Tensor: |
| 156 | # Branch on the inputs, not the rank: the drafter is built entirely on |
| 157 | # the last PP stage, where `is_first_rank` is False. |
| 158 | if intermediate_tensors is None: |
| 159 | if inputs_embeds is None: |
| 160 | inputs_embeds = self.embed_input_ids(input_ids) |
| 161 | assert hidden_states.shape[-1] == inputs_embeds.shape[-1] |
| 162 | inputs_embeds = self.pre_fc_norm_embedding(inputs_embeds) |
| 163 | hidden_states = self.pre_fc_norm_hidden(hidden_states) |
| 164 | hidden_states = torch.cat([inputs_embeds, hidden_states], dim=-1) |
| 165 | hidden_states = self.fc(hidden_states) |
| 166 | residual = None |
| 167 | else: |
| 168 | hidden_states = intermediate_tensors["hidden_states"] |
| 169 | residual = intermediate_tensors["residual"] |
| 170 | |
| 171 | current_step_idx = spec_step_idx % self.num_mtp_layers |
| 172 | mtp_layer = self.layers[current_step_idx] |
| 173 | if mtp_layer.use_attn_reduce_scatter_for_moe: |
| 174 | assert hidden_states.shape[0] == positions.shape[-1] |
| 175 | hidden_states = sequence_parallel_chunk(hidden_states) |
| 176 | assert residual is None |
| 177 | hidden_states, residual = mtp_layer( |
| 178 | positions=positions, |
| 179 | hidden_states=hidden_states, |
| 180 | residual=residual, |
| 181 | ) |
| 182 | |
| 183 | if not get_pp_group().is_last_rank: |
| 184 | return IntermediateTensors( |
| 185 | {"hidden_states": hidden_states, "residual": residual} |
| 186 | ) |
| 187 | |
| 188 | hidden_states, _ = self.norm(hidden_states, residual) |
| 189 | if mtp_layer.use_attn_reduce_scatter_for_moe: |
| 190 | hidden_states = tensor_model_parallel_all_gather(hidden_states, 0) |
| 191 | hidden_states = hidden_states[: positions.shape[-1]] |
| 192 | return hidden_states |
| 193 | |
| 194 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 195 | weights = maybe_fuse_shared_experts( |
| 196 | weights, |
| 197 | enabled=self.is_fused_shared_expert_enabled, |
| 198 | n_routed_experts=getattr(self.config, "num_experts", 0), |
| 199 | n_shared_experts=1, |
| 200 | ckpt_prefix="mlp.shared_expert", |
| 201 | ) |
| 202 | loader = AutoWeightsLoader(self) |
| 203 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 204 | |
| 205 | |
| 206 | @support_torch_compile( |
| 207 | dynamic_arg_dims={ |
| 208 | "input_ids": 0, |
| 209 | # positions is of shape (3, seq_len) if mrope is enabled for qwen2-vl, |
| 210 | # otherwise (seq_len, ). |
| 211 | "positions": -1, |
| 212 | "intermediate_tensors": 0, |
| 213 | "inputs_embeds": 0, |
| 214 | "hidden_states": 0, |
| 215 | } |
| 216 | ) |
| 217 | class Qwen3_5MTP(LocalArgmaxMixin, nn.Module, SupportsMultiModal, SupportsPP): |
| 218 | packed_modules_mapping = { |
| 219 | "qkv_proj": [ |
| 220 | "q_proj", |
| 221 | "k_proj", |
| 222 | "v_proj", |
| 223 | ], |
| 224 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 225 | } |
| 226 | |
| 227 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 228 | config = vllm_config.model_config.hf_text_config |
| 229 | self.vllm_config = vllm_config |
| 230 | self.quant_config = vllm_config.quant_config |
| 231 | |
| 232 | super().__init__() |
| 233 | self.config = config |
| 234 | self.model = Qwen3_5MultiTokenPredictor( |
| 235 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "mtp") |
| 236 | ) |
| 237 | |
| 238 | if get_pp_group().is_last_rank: |
| 239 | self.lm_head = ParallelLMHead( |
| 240 | config.vocab_size, |
| 241 | config.hidden_size, |
| 242 | quant_config=self.quant_config, |
| 243 | prefix=maybe_prefix(prefix, "lm_head"), |
| 244 | ) |
| 245 | if config.tie_word_embeddings: |
| 246 | self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) |
| 247 | else: |
| 248 | self.lm_head = PPMissingLayer() |
| 249 | |
| 250 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 251 | |
| 252 | self.make_empty_intermediate_tensors = ( |
| 253 | self.model.make_empty_intermediate_tensors |
| 254 | ) |
| 255 | |
| 256 | def embed_input_ids( |
| 257 | self, |
| 258 | input_ids: torch.Tensor, |
| 259 | multimodal_embeddings: MultiModalEmbeddings | None = None, |
| 260 | *, |
| 261 | is_multimodal: torch.Tensor | None = None, |
| 262 | ) -> torch.Tensor: |
| 263 | inputs_embeds = self._embed_text_input_ids( |
| 264 | input_ids, |
| 265 | self.model.embed_input_ids, |
| 266 | is_multimodal=is_multimodal, |
| 267 | ) |
| 268 | |
| 269 | if multimodal_embeddings is None or len(multimodal_embeddings) == 0: |
| 270 | return inputs_embeds |
| 271 | |
| 272 | is_multimodal = _require_is_multimodal(is_multimodal) |
| 273 | |
| 274 | inputs_embeds = _merge_multimodal_embeddings( |
| 275 | inputs_embeds=inputs_embeds, |
| 276 | multimodal_embeddings=multimodal_embeddings, |
| 277 | is_multimodal=is_multimodal, |
| 278 | ) |
| 279 | |
| 280 | return inputs_embeds |
| 281 | |
| 282 | def forward( # type: ignore[override] |
| 283 | self, |
| 284 | input_ids: torch.Tensor, |
| 285 | positions: torch.Tensor, |
| 286 | hidden_states: torch.Tensor, |
| 287 | intermediate_tensors: IntermediateTensors | None = None, |
| 288 | inputs_embeds: torch.Tensor | None = None, |
| 289 | **kwargs: object, |
| 290 | ): |
| 291 | hidden_states = self.model( |
| 292 | input_ids, positions, hidden_states, intermediate_tensors, inputs_embeds |
| 293 | ) |
| 294 | return hidden_states |
| 295 | |
| 296 | def compute_logits( |
| 297 | self, |
| 298 | hidden_states: torch.Tensor, |
| 299 | spec_step_idx: int = 0, |
| 300 | ) -> torch.Tensor | None: |
| 301 | return self.logits_processor(self.lm_head, hidden_states) |
| 302 | |
| 303 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 304 | def remap_weight_names(weights): |
| 305 | for name, weight in weights: |
| 306 | if name.startswith("mtp."): |
| 307 | name = name.replace("mtp.", "model.") |
| 308 | elif any(key in name for key in ["embed_tokens", "lm_head"]): |
| 309 | if "embed_tokens" in name: |
| 310 | name = name.replace("language_model.", "") |
| 311 | else: |
| 312 | continue |
| 313 | yield name, weight |
| 314 | |
| 315 | loader = AutoWeightsLoader(self) |
| 316 | return loader.load_weights(remap_weight_names(weights)) |
| 317 | |
| 318 | |
| 319 | class Qwen3_5MoeMTP(Qwen3_5MTP, QwenNextMixtureOfExperts): |
| 320 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 321 | super().__init__(vllm_config=vllm_config, prefix=prefix) |
| 322 | self.set_moe_parameters() |