| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Inference-only Qwen3Next model.""" |
| 4 | |
| 5 | from collections.abc import Iterable |
| 6 | from itertools import islice |
| 7 | |
| 8 | import torch |
| 9 | from torch import nn |
| 10 | |
| 11 | from vllm.compilation.decorators import support_torch_compile |
| 12 | from vllm.config import CacheConfig, ModelConfig, VllmConfig |
| 13 | from vllm.distributed import ( |
| 14 | get_ep_group, |
| 15 | get_pp_group, |
| 16 | get_tensor_model_parallel_world_size, |
| 17 | tensor_model_parallel_all_gather, |
| 18 | tensor_model_parallel_reduce_scatter, |
| 19 | ) |
| 20 | from vllm.model_executor.layers.attention import Attention |
| 21 | from vllm.model_executor.layers.fused_moe import ( |
| 22 | FusedMoEFactory, |
| 23 | GateLinear, |
| 24 | ) |
| 25 | from vllm.model_executor.layers.fused_moe.utils import ( |
| 26 | is_model_fused_shared_expert_compatible, |
| 27 | resolve_layer_fused_shared_expert, |
| 28 | ) |
| 29 | from vllm.model_executor.layers.fused_qk_norm_rope import fused_qk_rmsnorm_rope_gate |
| 30 | from vllm.model_executor.layers.layernorm import ( |
| 31 | GemmaRMSNorm as Qwen3NextRMSNorm, |
| 32 | ) |
| 33 | from vllm.model_executor.layers.linear import ( |
| 34 | QKVParallelLinear, |
| 35 | ReplicatedLinear, |
| 36 | RowParallelLinear, |
| 37 | ) |
| 38 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 39 | from vllm.model_executor.layers.mamba.gdn.qwen_gdn_linear_attn import ( |
| 40 | QwenGatedDeltaNetAttention, |
| 41 | ) |
| 42 | from vllm.model_executor.layers.mamba.mamba_utils import ( |
| 43 | MambaStateCopyFunc, |
| 44 | MambaStateCopyFuncCalculator, |
| 45 | MambaStateDtypeCalculator, |
| 46 | MambaStateShapeCalculator, |
| 47 | ) |
| 48 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 49 | from vllm.model_executor.layers.quantization.utils.config_utils import ( |
| 50 | get_quark_ocp_mx_group_size, |
| 51 | ) |
| 52 | from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding, get_rope |
| 53 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 54 | ParallelLMHead, |
| 55 | VocabParallelEmbedding, |
| 56 | ) |
| 57 | from vllm.model_executor.models.qwen2_moe import Qwen2MoeMLP as Qwen3NextMLP |
| 58 | from vllm.model_executor.models.utils import sequence_parallel_chunk |
| 59 | from vllm.platforms import current_platform |
| 60 | from vllm.sequence import IntermediateTensors |
| 61 | from vllm.transformers_utils.configs.qwen3_next import Qwen3NextConfig |
| 62 | from vllm.v1.attention.backend import AttentionType |
| 63 | |
| 64 | from .interfaces import ( |
| 65 | EagleModelMixin, |
| 66 | HasInnerState, |
| 67 | IsHybrid, |
| 68 | MixtureOfExperts, |
| 69 | SupportsEagle3, |
| 70 | SupportsLoRA, |
| 71 | SupportsPP, |
| 72 | ) |
| 73 | from .utils import ( |
| 74 | AutoWeightsLoader, |
| 75 | PPMissingLayer, |
| 76 | WeightsMapper, |
| 77 | extract_layer_index, |
| 78 | make_empty_intermediate_tensors_factory, |
| 79 | make_layers, |
| 80 | maybe_fuse_shared_experts, |
| 81 | maybe_prefix, |
| 82 | ) |
| 83 | |
| 84 | KVCache = tuple[torch.Tensor, torch.Tensor] |
| 85 | |
| 86 | |
| 87 | def _should_use_sequence_parallel(vllm_config: VllmConfig) -> bool: |
| 88 | config = vllm_config.model_config.hf_text_config |
| 89 | parallel_config = vllm_config.parallel_config |
| 90 | return ( |
| 91 | parallel_config.use_sequence_parallel_moe |
| 92 | and parallel_config.pipeline_parallel_size == 1 |
| 93 | and getattr(config, "num_experts", 0) > 0 |
| 94 | and not getattr(config, "mlp_only_layers", []) |
| 95 | and getattr(config, "decoder_sparse_step", 1) == 1 |
| 96 | ) |
| 97 | |
| 98 | |
| 99 | def _should_replicate_misaligned_shared_expert( |
| 100 | intermediate_size: int, |
| 101 | tp_size: int, |
| 102 | group_size: int | None, |
| 103 | enable_expert_parallel: bool, |
| 104 | is_sequence_parallel: bool, |
| 105 | ) -> bool: |
| 106 | if intermediate_size <= 0 or group_size is None: |
| 107 | return False |
| 108 | |
| 109 | partition_size, remainder = divmod(intermediate_size, tp_size) |
| 110 | if remainder == 0 and partition_size % group_size == 0: |
| 111 | return False |
| 112 | |
| 113 | if enable_expert_parallel or is_sequence_parallel: |
| 114 | return True |
| 115 | |
| 116 | if remainder != 0: |
| 117 | raise ValueError( |
| 118 | f"Shared-expert intermediate size {intermediate_size} must be " |
| 119 | f"divisible by tensor-parallel size {tp_size}." |
| 120 | ) |
| 121 | |
| 122 | raise ValueError( |
| 123 | "The Quark OCP MX shared expert cannot be tensor-parallelized: " |
| 124 | f"intermediate size {intermediate_size} with TP size {tp_size} " |
| 125 | f"produces a partition of {partition_size}, which is not divisible by " |
| 126 | f"the OCP MX group size {group_size}. Choose a compatible " |
| 127 | "tensor-parallel size or enable expert parallelism." |
| 128 | ) |
| 129 | |
| 130 | |
| 131 | class Qwen3NextSparseMoeBlock(nn.Module): |
| 132 | def __init__(self, vllm_config: VllmConfig, prefix: str = ""): |
| 133 | super().__init__() |
| 134 | |
| 135 | config = vllm_config.model_config.hf_text_config |
| 136 | parallel_config = vllm_config.parallel_config |
| 137 | quant_config = vllm_config.quant_config |
| 138 | |
| 139 | self.tp_size = get_tensor_model_parallel_world_size() |
| 140 | |
| 141 | self.ep_group = get_ep_group().device_group |
| 142 | self.ep_size = self.ep_group.size() |
| 143 | self.n_routed_experts = config.num_experts |
| 144 | |
| 145 | self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe |
| 146 | |
| 147 | # Resolve shared-expert fusion first (when enabled, TP alignment constraint no |
| 148 | # longer applies) |
| 149 | self.is_fused_shared_expert_enabled = False |
| 150 | if config.shared_expert_intermediate_size > 0: |
| 151 | self.is_fused_shared_expert_enabled = resolve_layer_fused_shared_expert( |
| 152 | quant_config, |
| 153 | prefix, |
| 154 | shared_expert_name="shared_expert", |
| 155 | ) |
| 156 | |
| 157 | if self.is_fused_shared_expert_enabled: |
| 158 | self.replicate_shared_expert = False |
| 159 | else: |
| 160 | shared_expert_group_size = get_quark_ocp_mx_group_size( |
| 161 | quant_config, |
| 162 | f"{prefix}.shared_expert.down_proj", |
| 163 | ) |
| 164 | self.replicate_shared_expert = _should_replicate_misaligned_shared_expert( |
| 165 | config.shared_expert_intermediate_size, |
| 166 | self.tp_size, |
| 167 | shared_expert_group_size, |
| 168 | parallel_config.enable_expert_parallel, |
| 169 | self.is_sequence_parallel, |
| 170 | ) |
| 171 | if self.tp_size > config.num_experts: |
| 172 | raise ValueError( |
| 173 | f"Tensor parallel size {self.tp_size} is greater than " |
| 174 | f"the number of experts {config.num_experts}." |
| 175 | ) |
| 176 | |
| 177 | # Load balancing settings. |
| 178 | eplb_config = vllm_config.parallel_config.eplb_config |
| 179 | self.enable_eplb = parallel_config.enable_eplb |
| 180 | |
| 181 | self.n_logical_experts = self.n_routed_experts |
| 182 | self.n_redundant_experts = eplb_config.num_redundant_experts |
| 183 | self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts |
| 184 | self.n_local_physical_experts = self.n_physical_experts // self.ep_size |
| 185 | |
| 186 | self.gate = GateLinear( |
| 187 | config.hidden_size, |
| 188 | config.num_experts, |
| 189 | prefix=f"{prefix}.gate", |
| 190 | ) |
| 191 | |
| 192 | self.shared_expert_gate = ReplicatedLinear( |
| 193 | config.hidden_size, |
| 194 | 1, |
| 195 | bias=False, |
| 196 | quant_config=None, |
| 197 | prefix=f"{prefix}.shared_expert_gate", |
| 198 | ) |
| 199 | |
| 200 | if ( |
| 201 | self.is_fused_shared_expert_enabled |
| 202 | or config.shared_expert_intermediate_size <= 0 |
| 203 | ): |
| 204 | self.shared_expert = None |
| 205 | else: |
| 206 | self.shared_expert = Qwen3NextMLP( |
| 207 | hidden_size=config.hidden_size, |
| 208 | intermediate_size=config.shared_expert_intermediate_size, |
| 209 | hidden_act=config.hidden_act, |
| 210 | quant_config=quant_config, |
| 211 | reduce_results=False, |
| 212 | expert_gate=self.shared_expert_gate, |
| 213 | is_sequence_parallel=self.is_sequence_parallel, |
| 214 | disable_tp=self.replicate_shared_expert, |
| 215 | prefix=f"{prefix}.shared_expert", |
| 216 | ) |
| 217 | |
| 218 | self.experts = FusedMoEFactory( |
| 219 | shared_experts=( |
| 220 | None if self.replicate_shared_expert else self.shared_expert |
| 221 | ), |
| 222 | gate=self.gate, |
| 223 | num_experts=self.n_routed_experts, |
| 224 | top_k=config.num_experts_per_tok, |
| 225 | hidden_size=config.hidden_size, |
| 226 | intermediate_size=config.moe_intermediate_size, |
| 227 | renormalize=getattr(config, "norm_topk_prob", True), |
| 228 | quant_config=quant_config, |
| 229 | prefix=f"{prefix}.experts", |
| 230 | enable_eplb=self.enable_eplb, |
| 231 | num_redundant_experts=self.n_redundant_experts, |
| 232 | is_sequence_parallel=self.is_sequence_parallel, |
| 233 | n_shared_experts=1 if self.shared_expert is None else None, |
| 234 | fuse_shared_experts=self.is_fused_shared_expert_enabled, |
| 235 | shared_expert_gate=self.shared_expert_gate |
| 236 | if self.shared_expert is None |
| 237 | else None, |
| 238 | ) |
| 239 | |
| 240 | def forward( |
| 241 | self, |
| 242 | hidden_states: torch.Tensor, |
| 243 | already_sequence_parallel: bool = False, |
| 244 | ) -> torch.Tensor: |
| 245 | # NOTE: hidden_states can have either 1D or 2D shape. |
| 246 | orig_shape = hidden_states.shape |
| 247 | num_tokens, hidden_dim = hidden_states.shape |
| 248 | hidden_states = hidden_states.view(-1, hidden_dim) |
| 249 | |
| 250 | if self.is_sequence_parallel and not already_sequence_parallel: |
| 251 | hidden_states = sequence_parallel_chunk(hidden_states) |
| 252 | |
| 253 | replicated_shared_output = ( |
| 254 | self.shared_expert(hidden_states) |
| 255 | if self.replicate_shared_expert and self.shared_expert is not None |
| 256 | else None |
| 257 | ) |
| 258 | final_hidden_states = self.experts( |
| 259 | hidden_states=hidden_states, router_logits=hidden_states |
| 260 | ) |
| 261 | if replicated_shared_output is not None: |
| 262 | final_hidden_states += replicated_shared_output |
| 263 | |
| 264 | if self.is_sequence_parallel and not already_sequence_parallel: |
| 265 | final_hidden_states = tensor_model_parallel_all_gather( |
| 266 | final_hidden_states, 0 |
| 267 | ) |
| 268 | final_hidden_states = final_hidden_states[:num_tokens] |
| 269 | |
| 270 | return final_hidden_states.view(orig_shape) |
| 271 | |
| 272 | |
| 273 | class Qwen3NextAttention(nn.Module): |
| 274 | def __init__( |
| 275 | self, |
| 276 | config: Qwen3NextConfig, |
| 277 | model_config: ModelConfig | None = None, |
| 278 | cache_config: CacheConfig | None = None, |
| 279 | quant_config: QuantizationConfig | None = None, |
| 280 | reduce_results: bool = True, |
| 281 | prefix: str = "", |
| 282 | ) -> None: |
| 283 | super().__init__() |
| 284 | self.config = config |
| 285 | self.hidden_size = config.hidden_size |
| 286 | tp_size = get_tensor_model_parallel_world_size() |
| 287 | self.total_num_heads = config.num_attention_heads |
| 288 | assert self.total_num_heads % tp_size == 0 |
| 289 | self.num_heads = self.total_num_heads // tp_size |
| 290 | self.total_num_kv_heads = config.num_key_value_heads |
| 291 | if self.total_num_kv_heads >= tp_size: |
| 292 | # Number of KV heads is greater than TP size, so we partition |
| 293 | # the KV heads across multiple tensor parallel GPUs. |
| 294 | assert self.total_num_kv_heads % tp_size == 0 |
| 295 | else: |
| 296 | # Number of KV heads is less than TP size, so we replicate |
| 297 | # the KV heads across multiple tensor parallel GPUs. |
| 298 | assert tp_size % self.total_num_kv_heads == 0 |
| 299 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 300 | self.head_dim = config.head_dim or (self.hidden_size // self.num_heads) |
| 301 | self.q_size = self.num_heads * self.head_dim |
| 302 | self.kv_size = self.num_kv_heads * self.head_dim |
| 303 | self.scaling = self.head_dim**-0.5 |
| 304 | self.dual_chunk_attention_config = getattr( |
| 305 | config, "dual_chunk_attention_config", None |
| 306 | ) |
| 307 | self.attn_output_gate = getattr(config, "attn_output_gate", True) |
| 308 | |
| 309 | self.qkv_proj = QKVParallelLinear( |
| 310 | config.hidden_size, |
| 311 | self.head_dim, |
| 312 | self.total_num_heads * (1 + self.attn_output_gate), |
| 313 | self.total_num_kv_heads, |
| 314 | bias=getattr(config, "qkv_bias", False), |
| 315 | quant_config=quant_config, |
| 316 | prefix=f"{prefix}.qkv_proj", |
| 317 | ) |
| 318 | |
| 319 | self.o_proj = RowParallelLinear( |
| 320 | self.total_num_heads * self.head_dim, |
| 321 | config.hidden_size, |
| 322 | bias=False, |
| 323 | reduce_results=reduce_results, |
| 324 | quant_config=quant_config, |
| 325 | prefix=f"{prefix}.o_proj", |
| 326 | ) |
| 327 | |
| 328 | self.rotary_emb = get_rope( |
| 329 | head_size=self.head_dim, |
| 330 | max_position=config.max_position_embeddings, |
| 331 | rope_parameters=config.rope_parameters, |
| 332 | dual_chunk_attention_config=self.dual_chunk_attention_config, |
| 333 | ) |
| 334 | |
| 335 | # Late-interaction retrieval models (e.g. ColQwen3.5) run BIDIRECTIONAL |
| 336 | # attention on the full_attention layers; they set config.is_causal=False |
| 337 | # via a VerifyAndUpdateConfig handler. Generation models leave is_causal |
| 338 | # unset (-> causal/DECODER), so this is a no-op for them. Mirrors qwen3.py. |
| 339 | attn_type = ( |
| 340 | AttentionType.DECODER |
| 341 | if getattr(config, "is_causal", True) |
| 342 | else AttentionType.ENCODER_ONLY |
| 343 | ) |
| 344 | self.attn = Attention( |
| 345 | self.num_heads, |
| 346 | self.head_dim, |
| 347 | self.scaling, |
| 348 | num_kv_heads=self.num_kv_heads, |
| 349 | cache_config=cache_config, |
| 350 | quant_config=quant_config, |
| 351 | prefix=f"{prefix}.attn", |
| 352 | attn_type=attn_type, |
| 353 | **{ |
| 354 | "layer_idx": extract_layer_index(prefix), |
| 355 | "dual_chunk_attention_config": self.dual_chunk_attention_config, |
| 356 | } |
| 357 | if self.dual_chunk_attention_config |
| 358 | else {}, |
| 359 | ) |
| 360 | |
| 361 | self.q_norm = Qwen3NextRMSNorm(self.head_dim, eps=config.rms_norm_eps) |
| 362 | self.k_norm = Qwen3NextRMSNorm(self.head_dim, eps=config.rms_norm_eps) |
| 363 | |
| 364 | # Fuse the gated split + QK-RMSNorm + (partial) NeoX RoPE + gate copy. |
| 365 | mm_config = model_config.multimodal_config if model_config else None |
| 366 | text_only = mm_config is None or mm_config.language_model_only |
| 367 | mrope_section = getattr(self.rotary_emb, "mrope_section", None) |
| 368 | supports_mrope = bool( |
| 369 | type(self.rotary_emb) is MRotaryEmbedding |
| 370 | and mrope_section |
| 371 | and len(mrope_section) == 3 |
| 372 | and sum(mrope_section) == self.rotary_emb.rotary_dim // 2 |
| 373 | and getattr(self.rotary_emb, "mrope_interleaved", False) |
| 374 | ) |
| 375 | supports_dtype = getattr(self.rotary_emb, "dtype", None) in ( |
| 376 | torch.float16, |
| 377 | torch.bfloat16, |
| 378 | ) |
| 379 | self.use_fused_qk_norm_rope_gate = ( |
| 380 | self.attn_output_gate |
| 381 | and getattr(self.rotary_emb, "is_neox_style", False) |
| 382 | and (current_platform.is_cuda() or current_platform.is_xpu()) |
| 383 | and supports_dtype |
| 384 | and (text_only or supports_mrope) |
| 385 | ) |
| 386 | |
| 387 | def _project_qkv_gate( |
| 388 | self, |
| 389 | qkv: torch.Tensor, |
| 390 | positions: torch.Tensor, |
| 391 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor | None]: |
| 392 | """Return post-norm, post-RoPE (q, k, v) and the pre-sigmoid gate. |
| 393 | |
| 394 | Dispatches between the fused Triton kernel and the eager |
| 395 | split + QK-RMSNorm + RoPE path. ``gate`` is ``None`` when output |
| 396 | gating is disabled. |
| 397 | """ |
| 398 | if self.use_fused_qk_norm_rope_gate: |
| 399 | q_gate, k, v = qkv.split( |
| 400 | [self.q_size * 2, self.kv_size, self.kv_size], dim=-1 |
| 401 | ) |
| 402 | if positions.ndim == 2 and not getattr( |
| 403 | self.rotary_emb, "mrope_section", None |
| 404 | ): |
| 405 | positions = positions[0] |
| 406 | q, k, gate = fused_qk_rmsnorm_rope_gate( |
| 407 | q_gate, |
| 408 | k, |
| 409 | self.q_norm.weight, |
| 410 | self.k_norm.weight, |
| 411 | self.rotary_emb.cos_sin_cache, |
| 412 | positions, |
| 413 | self.q_norm.variance_epsilon, |
| 414 | self.num_heads, |
| 415 | self.num_kv_heads, |
| 416 | self.head_dim, |
| 417 | self.rotary_emb.rotary_dim, |
| 418 | norm_beta=1.0, |
| 419 | mrope_section=( |
| 420 | getattr(self.rotary_emb, "mrope_section", None) |
| 421 | if positions.ndim == 2 |
| 422 | else None |
| 423 | ), |
| 424 | ) |
| 425 | return q, k, v, gate |
| 426 | |
| 427 | if self.attn_output_gate: |
| 428 | q_gate, k, v = qkv.split( |
| 429 | [self.q_size * 2, self.kv_size, self.kv_size], dim=-1 |
| 430 | ) |
| 431 | orig_shape = q_gate.shape[:-1] |
| 432 | q_gate = q_gate.view(*orig_shape, self.num_heads, -1) |
| 433 | q, gate = torch.chunk(q_gate, 2, dim=-1) |
| 434 | q = q.reshape(*orig_shape, -1) |
| 435 | gate = gate.reshape(*orig_shape, -1) |
| 436 | else: |
| 437 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 438 | gate = None |
| 439 | |
| 440 | q = self.q_norm(q.view(-1, self.num_heads, self.head_dim)).view( |
| 441 | -1, self.num_heads * self.head_dim |
| 442 | ) |
| 443 | k = self.k_norm(k.view(-1, self.num_kv_heads, self.head_dim)).view( |
| 444 | -1, self.num_kv_heads * self.head_dim |
| 445 | ) |
| 446 | q, k = self.rotary_emb(positions, q, k) |
| 447 | return q, k, v, gate |
| 448 | |
| 449 | def forward( |
| 450 | self, |
| 451 | positions: torch.Tensor, |
| 452 | hidden_states: torch.Tensor, |
| 453 | ) -> torch.Tensor: |
| 454 | qkv, _ = self.qkv_proj(hidden_states) |
| 455 | q, k, v, gate = self._project_qkv_gate(qkv, positions) |
| 456 | attn_output = self.attn(q, k, v) |
| 457 | if gate is not None: |
| 458 | attn_output = attn_output * torch.sigmoid(gate) |
| 459 | output, _ = self.o_proj(attn_output) |
| 460 | return output |
| 461 | |
| 462 | |
| 463 | class Qwen3NextDecoderLayer(nn.Module): |
| 464 | def __init__( |
| 465 | self, |
| 466 | vllm_config: VllmConfig, |
| 467 | layer_type: str, |
| 468 | prefix: str = "", |
| 469 | ) -> None: |
| 470 | super().__init__() |
| 471 | |
| 472 | config = vllm_config.model_config.hf_config |
| 473 | model_config = vllm_config.model_config |
| 474 | cache_config = vllm_config.cache_config |
| 475 | quant_config = vllm_config.quant_config |
| 476 | |
| 477 | self.layer_type = layer_type |
| 478 | self.layer_idx = extract_layer_index(prefix) |
| 479 | |
| 480 | mlp_only_layers = ( |
| 481 | [] if not hasattr(config, "mlp_only_layers") else config.mlp_only_layers |
| 482 | ) |
| 483 | is_moe_layer = (self.layer_idx not in mlp_only_layers) and ( |
| 484 | config.num_experts > 0 |
| 485 | and (self.layer_idx + 1) % config.decoder_sparse_step == 0 |
| 486 | ) |
| 487 | self.use_attn_reduce_scatter_for_moe = _should_use_sequence_parallel( |
| 488 | vllm_config |
| 489 | ) |
| 490 | |
| 491 | if self.layer_type == "linear_attention": |
| 492 | self.linear_attn = QwenGatedDeltaNetAttention( |
| 493 | config, |
| 494 | vllm_config=vllm_config, |
| 495 | prefix=f"{prefix}.linear_attn", |
| 496 | gqa_interleaved_layout=True, |
| 497 | reduce_results=not self.use_attn_reduce_scatter_for_moe, |
| 498 | ) |
| 499 | elif self.layer_type == "full_attention": |
| 500 | self.self_attn = Qwen3NextAttention( |
| 501 | config, |
| 502 | model_config=model_config, |
| 503 | cache_config=cache_config, |
| 504 | quant_config=quant_config, |
| 505 | reduce_results=not self.use_attn_reduce_scatter_for_moe, |
| 506 | prefix=f"{prefix}.self_attn", |
| 507 | ) |
| 508 | else: |
| 509 | raise ValueError(f"Invalid layer_type {self.layer_type}") |
| 510 | |
| 511 | if is_moe_layer: |
| 512 | self.mlp = Qwen3NextSparseMoeBlock( |
| 513 | vllm_config=vllm_config, |
| 514 | prefix=f"{prefix}.mlp", |
| 515 | ) |
| 516 | else: |
| 517 | self.mlp = Qwen3NextMLP( |
| 518 | hidden_size=config.hidden_size, |
| 519 | intermediate_size=config.intermediate_size, |
| 520 | hidden_act=config.hidden_act, |
| 521 | quant_config=quant_config, |
| 522 | prefix=f"{prefix}.mlp", |
| 523 | ) |
| 524 | |
| 525 | self.input_layernorm = Qwen3NextRMSNorm( |
| 526 | config.hidden_size, eps=config.rms_norm_eps |
| 527 | ) |
| 528 | self.post_attention_layernorm = Qwen3NextRMSNorm( |
| 529 | config.hidden_size, eps=config.rms_norm_eps |
| 530 | ) |
| 531 | |
| 532 | self.layer_scale = getattr(config, "layer_scale", False) |
| 533 | if self.layer_scale: |
| 534 | self.attn_layer_scale = torch.nn.Parameter( |
| 535 | torch.zeros( |
| 536 | 1, |
| 537 | 1, |
| 538 | config.hidden_size, |
| 539 | ), |
| 540 | ) |
| 541 | self.ffn_layer_scale = torch.nn.Parameter( |
| 542 | torch.zeros( |
| 543 | 1, |
| 544 | 1, |
| 545 | config.hidden_size, |
| 546 | ), |
| 547 | ) |
| 548 | |
| 549 | def forward( |
| 550 | self, |
| 551 | hidden_states: torch.Tensor, |
| 552 | residual: torch.Tensor | None, |
| 553 | positions: torch.Tensor, |
| 554 | **kwargs: object, |
| 555 | ): |
| 556 | full_num_tokens = positions.shape[-1] |
| 557 | |
| 558 | if residual is None: |
| 559 | residual = hidden_states |
| 560 | hidden_states = self.input_layernorm(hidden_states) |
| 561 | else: |
| 562 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 563 | |
| 564 | if self.use_attn_reduce_scatter_for_moe: |
| 565 | hidden_states = tensor_model_parallel_all_gather(hidden_states, 0) |
| 566 | hidden_states = hidden_states[:full_num_tokens] |
| 567 | |
| 568 | if self.layer_type == "linear_attention": |
| 569 | hidden_states = self.linear_attn(hidden_states=hidden_states) |
| 570 | elif self.layer_type == "full_attention": |
| 571 | hidden_states = self.self_attn( |
| 572 | hidden_states=hidden_states, |
| 573 | positions=positions, |
| 574 | ) |
| 575 | else: |
| 576 | raise ValueError("Invalid layer_type") |
| 577 | |
| 578 | if self.layer_scale: |
| 579 | if len(hidden_states.shape) == 2: |
| 580 | hidden_states = hidden_states * ( |
| 581 | self.attn_layer_scale.to(hidden_states.dtype)[0] + 1 |
| 582 | ) |
| 583 | else: |
| 584 | hidden_states = hidden_states * ( |
| 585 | self.attn_layer_scale.to(hidden_states.dtype) + 1 |
| 586 | ) |
| 587 | |
| 588 | if self.use_attn_reduce_scatter_for_moe: |
| 589 | tp_world_size = get_tensor_model_parallel_world_size() |
| 590 | # small trick using minus, eg. -17 % 8 = 7 |
| 591 | sp_pad = (-hidden_states.shape[0]) % tp_world_size |
| 592 | # pad if not divisible by world size |
| 593 | hidden_states = torch.nn.functional.pad(hidden_states, (0, 0, 0, sp_pad)) |
| 594 | hidden_states = tensor_model_parallel_reduce_scatter(hidden_states, 0) |
| 595 | |
| 596 | # Fully Connected |
| 597 | hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) |
| 598 | if self.use_attn_reduce_scatter_for_moe: |
| 599 | hidden_states = self.mlp( |
| 600 | hidden_states, |
| 601 | already_sequence_parallel=True, |
| 602 | ) |
| 603 | else: |
| 604 | hidden_states = self.mlp(hidden_states) |
| 605 | |
| 606 | if self.layer_scale: |
| 607 | if len(hidden_states.shape) == 2: |
| 608 | hidden_states = hidden_states * ( |
| 609 | self.ffn_layer_scale.to(hidden_states.dtype)[0] + 1 |
| 610 | ) |
| 611 | else: |
| 612 | assert len(hidden_states.shape) == len(self.ffn_layer_scale.shape), ( |
| 613 | f"shape must be the same {len(hidden_states.shape)}, " |
| 614 | f"{len(self.ffn_layer_scale.shape)}" |
| 615 | ) |
| 616 | hidden_states = hidden_states * ( |
| 617 | self.ffn_layer_scale.to(hidden_states.dtype) + 1 |
| 618 | ) |
| 619 | |
| 620 | return hidden_states, residual |
| 621 | |
| 622 | |
| 623 | @support_torch_compile |
| 624 | class Qwen3NextModel(nn.Module, EagleModelMixin): |
| 625 | hf_to_vllm_mapper = WeightsMapper( |
| 626 | orig_to_new_stacked={ |
| 627 | # weight_name: (param_name, shard_id) |
| 628 | ".q_proj": (".qkv_proj", "q"), |
| 629 | ".k_proj": (".qkv_proj", "k"), |
| 630 | ".v_proj": (".qkv_proj", "v"), |
| 631 | ".mlp.gate_proj": (".mlp.gate_up_proj", 0), |
| 632 | ".mlp.up_proj": (".mlp.gate_up_proj", 1), |
| 633 | ".shared_expert.gate_proj": (".shared_expert.gate_up_proj", 0), |
| 634 | ".shared_expert.up_proj": (".shared_expert.gate_up_proj", 1), |
| 635 | } |
| 636 | ) |
| 637 | |
| 638 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 639 | super().__init__() |
| 640 | |
| 641 | config: Qwen3NextConfig = vllm_config.model_config.hf_text_config |
| 642 | parallel_config = vllm_config.parallel_config |
| 643 | |
| 644 | eplb_config = parallel_config.eplb_config |
| 645 | self.num_redundant_experts = eplb_config.num_redundant_experts |
| 646 | |
| 647 | self.config = config |
| 648 | |
| 649 | self.vocab_size = config.vocab_size |
| 650 | |
| 651 | self.embed_tokens = VocabParallelEmbedding( |
| 652 | self.vocab_size, |
| 653 | config.hidden_size, |
| 654 | ) |
| 655 | |
| 656 | def get_layer(prefix: str): |
| 657 | return Qwen3NextDecoderLayer( |
| 658 | vllm_config, |
| 659 | layer_type=config.layer_types[extract_layer_index(prefix)], |
| 660 | prefix=prefix, |
| 661 | ) |
| 662 | |
| 663 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 664 | config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers" |
| 665 | ) |
| 666 | self.is_fused_shared_expert_enabled = is_model_fused_shared_expert_compatible( |
| 667 | self.layers, |
| 668 | Qwen3NextSparseMoeBlock, |
| 669 | "mlp", |
| 670 | ) |
| 671 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 672 | ["hidden_states", "residual"], config.hidden_size |
| 673 | ) |
| 674 | |
| 675 | if get_pp_group().is_last_rank: |
| 676 | self.norm = Qwen3NextRMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 677 | else: |
| 678 | self.norm = PPMissingLayer() |
| 679 | |
| 680 | self.aux_hidden_state_layers: tuple[int, ...] = () |
| 681 | |
| 682 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 683 | return self.embed_tokens(input_ids) |
| 684 | |
| 685 | @property |
| 686 | def use_sequence_parallel(self) -> bool: |
| 687 | return self.layers[self.start_layer].use_attn_reduce_scatter_for_moe |
| 688 | |
| 689 | def forward( |
| 690 | self, |
| 691 | input_ids: torch.Tensor | None, |
| 692 | positions: torch.Tensor, |
| 693 | intermediate_tensors: IntermediateTensors | None = None, |
| 694 | inputs_embeds: torch.Tensor | None = None, |
| 695 | ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]: |
| 696 | if get_pp_group().is_first_rank: |
| 697 | if inputs_embeds is not None: |
| 698 | hidden_states = inputs_embeds |
| 699 | else: |
| 700 | hidden_states = self.embed_input_ids(input_ids) |
| 701 | residual = None |
| 702 | else: |
| 703 | assert intermediate_tensors is not None |
| 704 | hidden_states = intermediate_tensors["hidden_states"] |
| 705 | residual = intermediate_tensors["residual"] |
| 706 | |
| 707 | full_num_tokens = positions.shape[-1] |
| 708 | if self.use_sequence_parallel: |
| 709 | hidden_states = sequence_parallel_chunk(hidden_states) |
| 710 | assert residual is None |
| 711 | |
| 712 | aux_hidden_states = self._maybe_add_hidden_state([], 0, hidden_states, residual) |
| 713 | for layer_idx, layer in enumerate( |
| 714 | islice(self.layers, self.start_layer, self.end_layer), |
| 715 | start=self.start_layer, |
| 716 | ): |
| 717 | hidden_states, residual = layer( |
| 718 | positions=positions, |
| 719 | hidden_states=hidden_states, |
| 720 | residual=residual, |
| 721 | ) |
| 722 | self._maybe_add_hidden_state( |
| 723 | aux_hidden_states, layer_idx + 1, hidden_states, residual |
| 724 | ) |
| 725 | |
| 726 | if not get_pp_group().is_last_rank: |
| 727 | return IntermediateTensors( |
| 728 | {"hidden_states": hidden_states, "residual": residual} |
| 729 | ) |
| 730 | hidden_states, _ = self.norm(hidden_states, residual) |
| 731 | if self.use_sequence_parallel: |
| 732 | if aux_hidden_states: |
| 733 | hidden_size = hidden_states.shape[-1] |
| 734 | hidden_states = torch.cat([hidden_states, *aux_hidden_states], dim=-1) |
| 735 | hidden_states = tensor_model_parallel_all_gather(hidden_states, 0) |
| 736 | hidden_states = hidden_states[:full_num_tokens] |
| 737 | hidden_states, *aux_hidden_states = hidden_states.split( |
| 738 | hidden_size, dim=-1 |
| 739 | ) |
| 740 | else: |
| 741 | hidden_states = tensor_model_parallel_all_gather(hidden_states, 0) |
| 742 | hidden_states = hidden_states[:full_num_tokens] |
| 743 | if aux_hidden_states: |
| 744 | return hidden_states, aux_hidden_states |
| 745 | return hidden_states |
| 746 | |
| 747 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 748 | weights = maybe_fuse_shared_experts( |
| 749 | weights, |
| 750 | enabled=self.is_fused_shared_expert_enabled, |
| 751 | n_routed_experts=getattr(self.config, "num_experts", 0), |
| 752 | n_shared_experts=1, |
| 753 | ckpt_prefix="mlp.shared_expert", |
| 754 | ) |
| 755 | loader = AutoWeightsLoader(self) |
| 756 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 757 | |
| 758 | |
| 759 | class QwenNextMixtureOfExperts(MixtureOfExperts): |
| 760 | model: Qwen3NextModel |
| 761 | num_local_physical_experts: int |
| 762 | |
| 763 | def update_physical_experts_metadata( |
| 764 | self, |
| 765 | num_physical_experts: int, |
| 766 | num_local_physical_experts: int, |
| 767 | ) -> None: |
| 768 | assert self.num_local_physical_experts == num_local_physical_experts |
| 769 | self.num_physical_experts = num_physical_experts |
| 770 | self.num_local_physical_experts = num_local_physical_experts |
| 771 | self.num_redundant_experts = num_physical_experts - self.num_logical_experts |
| 772 | for layer in self.model.layers: |
| 773 | if isinstance(layer.mlp, Qwen3NextSparseMoeBlock): |
| 774 | moe = layer.mlp |
| 775 | moe.n_local_physical_experts = num_local_physical_experts |
| 776 | moe.n_physical_experts = num_physical_experts |
| 777 | moe.n_redundant_experts = self.num_redundant_experts |
| 778 | moe.experts.update_expert_map() |
| 779 | |
| 780 | def set_moe_parameters(self): |
| 781 | self.moe_layers = [] |
| 782 | example_moe = None |
| 783 | for layer in self.model.layers: |
| 784 | if isinstance(layer, Qwen3NextDecoderLayer) and isinstance( |
| 785 | layer.mlp, Qwen3NextSparseMoeBlock |
| 786 | ): |
| 787 | example_moe = layer.mlp |
| 788 | self.moe_layers.append(layer.mlp.experts) |
| 789 | |
| 790 | if example_moe is None: |
| 791 | raise RuntimeError("No Qwen3Next layer found in the model.layers.") |
| 792 | |
| 793 | # Set MoE hyperparameters |
| 794 | self.num_moe_layers = len(self.moe_layers) |
| 795 | self.num_expert_groups = 1 |
| 796 | self.num_shared_experts = 0 |
| 797 | self.num_logical_experts = example_moe.n_logical_experts |
| 798 | self.num_physical_experts = example_moe.n_physical_experts |
| 799 | self.num_local_physical_experts = example_moe.n_local_physical_experts |
| 800 | self.num_routed_experts = example_moe.n_routed_experts |
| 801 | self.num_redundant_experts = example_moe.n_redundant_experts |
| 802 | |
| 803 | |
| 804 | class Qwen3NextForCausalLM( |
| 805 | nn.Module, |
| 806 | HasInnerState, |
| 807 | SupportsLoRA, |
| 808 | SupportsPP, |
| 809 | QwenNextMixtureOfExperts, |
| 810 | IsHybrid, |
| 811 | SupportsEagle3, |
| 812 | ): |
| 813 | # MTP weights are loaded by the draft model, not this one. |
| 814 | hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={"mtp.": None}) |
| 815 | |
| 816 | packed_modules_mapping = { |
| 817 | "qkv_proj": [ |
| 818 | "q_proj", |
| 819 | "k_proj", |
| 820 | "v_proj", |
| 821 | ], |
| 822 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 823 | "in_proj_qkvz": ["in_proj_qkvz"], |
| 824 | "in_proj_ba": ["in_proj_ba"], |
| 825 | } |
| 826 | |
| 827 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 828 | config = vllm_config.model_config.hf_text_config |
| 829 | self.vllm_config = vllm_config |
| 830 | self.model_config = vllm_config.model_config |
| 831 | |
| 832 | scheduler_config = vllm_config.scheduler_config |
| 833 | self.quant_config = vllm_config.quant_config |
| 834 | |
| 835 | super().__init__() |
| 836 | self.config = config |
| 837 | self.scheduler_config = scheduler_config |
| 838 | self.model = Qwen3NextModel( |
| 839 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 840 | ) |
| 841 | |
| 842 | self.lm_head = ParallelLMHead( |
| 843 | config.vocab_size, |
| 844 | config.hidden_size, |
| 845 | prefix=maybe_prefix(prefix, "lm_head"), |
| 846 | ) |
| 847 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 848 | self.make_empty_intermediate_tensors = ( |
| 849 | self.model.make_empty_intermediate_tensors |
| 850 | ) |
| 851 | |
| 852 | # Set MoE hyperparameters |
| 853 | self.set_moe_parameters() |
| 854 | |
| 855 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 856 | return self.model.embed_input_ids(input_ids) |
| 857 | |
| 858 | def forward( |
| 859 | self, |
| 860 | input_ids: torch.Tensor | None, |
| 861 | positions: torch.Tensor, |
| 862 | intermediate_tensors: IntermediateTensors | None = None, |
| 863 | inputs_embeds: torch.Tensor | None = None, |
| 864 | **kwargs: object, |
| 865 | ): |
| 866 | hidden_states = self.model( |
| 867 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 868 | ) |
| 869 | |
| 870 | return hidden_states |
| 871 | |
| 872 | @classmethod |
| 873 | def get_mamba_state_dtype_from_config( |
| 874 | cls, |
| 875 | vllm_config: "VllmConfig", |
| 876 | ) -> tuple[torch.dtype, torch.dtype]: |
| 877 | return MambaStateDtypeCalculator.gated_delta_net_state_dtype( |
| 878 | vllm_config.model_config.dtype, |
| 879 | vllm_config.cache_config.mamba_cache_dtype, |
| 880 | vllm_config.cache_config.mamba_ssm_cache_dtype, |
| 881 | ) |
| 882 | |
| 883 | @classmethod |
| 884 | def get_mamba_state_shape_from_config( |
| 885 | cls, vllm_config: "VllmConfig" |
| 886 | ) -> tuple[tuple[int, int], tuple[int, int]]: |
| 887 | parallel_config = vllm_config.parallel_config |
| 888 | hf_config = vllm_config.model_config.hf_text_config |
| 889 | tp_size = parallel_config.tensor_parallel_size |
| 890 | num_spec = ( |
| 891 | vllm_config.speculative_config.num_speculative_tokens |
| 892 | if vllm_config.speculative_config |
| 893 | else 0 |
| 894 | ) |
| 895 | return MambaStateShapeCalculator.gated_delta_net_state_shape( |
| 896 | tp_size, |
| 897 | hf_config.linear_num_key_heads, |
| 898 | hf_config.linear_num_value_heads, |
| 899 | hf_config.linear_key_head_dim, |
| 900 | hf_config.linear_value_head_dim, |
| 901 | hf_config.linear_conv_kernel_dim, |
| 902 | num_spec, |
| 903 | ) |
| 904 | |
| 905 | @classmethod |
| 906 | def get_mamba_state_copy_func(cls) -> tuple[MambaStateCopyFunc, MambaStateCopyFunc]: |
| 907 | return MambaStateCopyFuncCalculator.gated_delta_net_state_copy_func() |
| 908 | |
| 909 | def compute_logits( |
| 910 | self, |
| 911 | hidden_states: torch.Tensor, |
| 912 | ) -> torch.Tensor | None: |
| 913 | return self.logits_processor(self.lm_head, hidden_states) |
| 914 | |
| 915 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 916 | loader = AutoWeightsLoader(self) |
| 917 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |