| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Inference-only AfMoE model compatible with HuggingFace weights.""" |
| 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, VllmConfig, get_current_vllm_config |
| 13 | from vllm.distributed import ( |
| 14 | get_ep_group, |
| 15 | get_pp_group, |
| 16 | get_tensor_model_parallel_world_size, |
| 17 | ) |
| 18 | from vllm.logger import init_logger |
| 19 | from vllm.model_executor.layers.attention import Attention |
| 20 | from vllm.model_executor.layers.fused_moe import ( |
| 21 | FusedMoEFactory, |
| 22 | GateLinear, |
| 23 | MoERunner, |
| 24 | ) |
| 25 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 26 | from vllm.model_executor.layers.linear import ( |
| 27 | ColumnParallelLinear, |
| 28 | QKVParallelLinear, |
| 29 | RowParallelLinear, |
| 30 | ) |
| 31 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 32 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 33 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 34 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 35 | ParallelLMHead, |
| 36 | VocabParallelEmbedding, |
| 37 | ) |
| 38 | from vllm.model_executor.models.interfaces import ( |
| 39 | EagleModelMixin, |
| 40 | MixtureOfExperts, |
| 41 | SupportsEagle3, |
| 42 | SupportsLoRA, |
| 43 | SupportsPP, |
| 44 | ) |
| 45 | from vllm.model_executor.models.llama import LlamaMLP as AfmoeMLP |
| 46 | from vllm.model_executor.models.utils import ( |
| 47 | AutoWeightsLoader, |
| 48 | PPMissingLayer, |
| 49 | WeightsMapper, |
| 50 | extract_layer_index, |
| 51 | make_layers, |
| 52 | maybe_prefix, |
| 53 | ) |
| 54 | from vllm.sequence import IntermediateTensors |
| 55 | from vllm.v1.attention.backend import AttentionType |
| 56 | |
| 57 | logger = init_logger(__name__) |
| 58 | |
| 59 | |
| 60 | class AfmoeMoE(nn.Module): |
| 61 | def __init__( |
| 62 | self, |
| 63 | config, # AfmoeConfig |
| 64 | quant_config: QuantizationConfig | None = None, |
| 65 | prefix: str = "", |
| 66 | enable_eplb: bool = False, |
| 67 | ): |
| 68 | super().__init__() |
| 69 | self.tp_size = get_tensor_model_parallel_world_size() |
| 70 | self.route_scale = config.route_scale |
| 71 | self.score_func = config.score_func |
| 72 | self.route_norm = config.route_norm |
| 73 | |
| 74 | self.ep_group = get_ep_group().device_group |
| 75 | self.ep_size = self.ep_group.size() |
| 76 | self.n_routed_experts: int = config.num_experts |
| 77 | self.n_shared_experts: int = config.num_shared_experts |
| 78 | |
| 79 | if config.hidden_act != "silu": |
| 80 | raise ValueError( |
| 81 | f"Unsupported activation: {config.hidden_act}. " |
| 82 | "Only silu is supported for now." |
| 83 | ) |
| 84 | |
| 85 | # Router gate |
| 86 | self.gate = GateLinear( |
| 87 | config.hidden_size, |
| 88 | config.num_experts, |
| 89 | out_dtype=torch.float32, |
| 90 | params_dtype=torch.float32, |
| 91 | prefix=f"{prefix}.gate", |
| 92 | ) |
| 93 | self.expert_bias = nn.Parameter( |
| 94 | torch.empty(config.num_experts, dtype=torch.float32) |
| 95 | ) |
| 96 | |
| 97 | # Load balancing settings |
| 98 | vllm_config = get_current_vllm_config() |
| 99 | eplb_config = vllm_config.parallel_config.eplb_config |
| 100 | self.enable_eplb = enable_eplb |
| 101 | |
| 102 | self.n_redundant_experts = eplb_config.num_redundant_experts |
| 103 | self.n_logical_experts = self.n_routed_experts |
| 104 | self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts |
| 105 | self.n_local_physical_experts = self.n_physical_experts // self.ep_size |
| 106 | |
| 107 | self.shared_experts = None |
| 108 | # Shared experts |
| 109 | if config.num_shared_experts > 0: |
| 110 | intermediate_size = config.moe_intermediate_size * config.num_shared_experts |
| 111 | self.shared_experts = AfmoeMLP( |
| 112 | hidden_size=config.hidden_size, |
| 113 | intermediate_size=intermediate_size, |
| 114 | hidden_act=config.hidden_act, |
| 115 | quant_config=quant_config, |
| 116 | reduce_results=False, |
| 117 | prefix=f"{prefix}.shared_experts", |
| 118 | ) |
| 119 | |
| 120 | # Routed experts using FusedMoEFactory |
| 121 | self.experts = FusedMoEFactory( |
| 122 | shared_experts=self.shared_experts, |
| 123 | num_experts=config.num_experts, |
| 124 | top_k=config.num_experts_per_tok, |
| 125 | hidden_size=config.hidden_size, |
| 126 | intermediate_size=config.moe_intermediate_size, |
| 127 | renormalize=self.route_norm if self.score_func == "sigmoid" else False, |
| 128 | quant_config=quant_config, |
| 129 | use_grouped_topk=True, |
| 130 | num_expert_group=config.n_group, |
| 131 | topk_group=config.topk_group, |
| 132 | prefix=f"{prefix}.experts", |
| 133 | scoring_func=self.score_func, |
| 134 | routed_scaling_factor=self.route_scale, |
| 135 | e_score_correction_bias=self.expert_bias, |
| 136 | enable_eplb=self.enable_eplb, |
| 137 | num_redundant_experts=self.n_redundant_experts, |
| 138 | router_logits_dtype=torch.float32, |
| 139 | ) |
| 140 | |
| 141 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 142 | num_tokens, hidden_dim = hidden_states.shape |
| 143 | hidden_states = hidden_states.view(-1, hidden_dim) |
| 144 | |
| 145 | router_logits, _ = self.gate(hidden_states) |
| 146 | |
| 147 | final_hidden_states = self.experts( |
| 148 | hidden_states=hidden_states, router_logits=router_logits |
| 149 | ) |
| 150 | |
| 151 | return final_hidden_states.view(num_tokens, hidden_dim) |
| 152 | |
| 153 | |
| 154 | class AfmoeAttention(nn.Module): |
| 155 | def __init__( |
| 156 | self, |
| 157 | config, # AfmoeConfig |
| 158 | layer_idx: int, |
| 159 | hidden_size: int, |
| 160 | num_heads: int, |
| 161 | num_kv_heads: int, |
| 162 | max_position_embeddings: int = 131072, |
| 163 | head_dim: int | None = None, |
| 164 | rms_norm_eps: float = 1e-05, |
| 165 | cache_config: CacheConfig | None = None, |
| 166 | quant_config: QuantizationConfig | None = None, |
| 167 | prefix: str = "", |
| 168 | attn_type: str = AttentionType.DECODER, |
| 169 | ) -> None: |
| 170 | super().__init__() |
| 171 | self.layer_idx = layer_idx |
| 172 | self.hidden_size = hidden_size |
| 173 | tp_size = get_tensor_model_parallel_world_size() |
| 174 | self.total_num_heads = num_heads |
| 175 | assert self.total_num_heads % tp_size == 0 |
| 176 | self.num_heads = self.total_num_heads // tp_size |
| 177 | self.total_num_kv_heads = num_kv_heads |
| 178 | if self.total_num_kv_heads >= tp_size: |
| 179 | # Number of KV heads is greater than TP size, so we partition |
| 180 | # the KV heads across multiple tensor parallel GPUs. |
| 181 | assert self.total_num_kv_heads % tp_size == 0 |
| 182 | else: |
| 183 | # Number of KV heads is less than TP size, so we replicate |
| 184 | # the KV heads across multiple tensor parallel GPUs. |
| 185 | assert tp_size % self.total_num_kv_heads == 0 |
| 186 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 187 | self.head_dim = head_dim or (hidden_size // self.total_num_heads) |
| 188 | self.q_size = self.num_heads * self.head_dim |
| 189 | self.kv_size = self.num_kv_heads * self.head_dim |
| 190 | self.scaling = self.head_dim**-0.5 |
| 191 | self.max_position_embeddings = max_position_embeddings |
| 192 | |
| 193 | # Check if this is a local attention layer |
| 194 | self.is_local_attention = config.layer_types[layer_idx] == "sliding_attention" |
| 195 | self.sliding_window = config.sliding_window if self.is_local_attention else None |
| 196 | |
| 197 | self.qkv_proj = QKVParallelLinear( |
| 198 | self.hidden_size, |
| 199 | self.head_dim, |
| 200 | self.total_num_heads, |
| 201 | self.total_num_kv_heads, |
| 202 | bias=False, |
| 203 | quant_config=quant_config, |
| 204 | prefix=f"{prefix}.qkv_proj", |
| 205 | ) |
| 206 | |
| 207 | self.o_proj = RowParallelLinear( |
| 208 | self.total_num_heads * self.head_dim, |
| 209 | self.hidden_size, |
| 210 | bias=False, |
| 211 | quant_config=quant_config, |
| 212 | prefix=f"{prefix}.o_proj", |
| 213 | ) |
| 214 | |
| 215 | # Gating projection |
| 216 | self.gate_proj = ColumnParallelLinear( |
| 217 | hidden_size, |
| 218 | self.total_num_heads * self.head_dim, |
| 219 | bias=False, |
| 220 | quant_config=quant_config, |
| 221 | prefix=f"{prefix}.gate_proj", |
| 222 | ) |
| 223 | |
| 224 | # Q/K normalization |
| 225 | self.q_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps) |
| 226 | self.k_norm = RMSNorm(self.head_dim, eps=config.rms_norm_eps) |
| 227 | |
| 228 | # Only create rotary embeddings for local attention |
| 229 | if self.is_local_attention: |
| 230 | self.rotary_emb: nn.Module | None = get_rope( |
| 231 | self.head_dim, |
| 232 | max_position=max_position_embeddings, |
| 233 | rope_parameters=config.rope_parameters, |
| 234 | is_neox_style=True, |
| 235 | ) |
| 236 | else: |
| 237 | self.rotary_emb = None |
| 238 | |
| 239 | self.attn = Attention( |
| 240 | self.num_heads, |
| 241 | self.head_dim, |
| 242 | self.scaling, |
| 243 | num_kv_heads=self.num_kv_heads, |
| 244 | cache_config=cache_config, |
| 245 | quant_config=quant_config, |
| 246 | per_layer_sliding_window=self.sliding_window, |
| 247 | prefix=f"{prefix}.attn", |
| 248 | attn_type=attn_type, |
| 249 | ) |
| 250 | |
| 251 | def forward( |
| 252 | self, |
| 253 | positions: torch.Tensor, |
| 254 | hidden_states: torch.Tensor, |
| 255 | ) -> torch.Tensor: |
| 256 | qkv, _ = self.qkv_proj(hidden_states) |
| 257 | gate, _ = self.gate_proj(hidden_states) |
| 258 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 259 | |
| 260 | # Apply Q/K normalization |
| 261 | q = self.q_norm(q.reshape(-1, self.num_heads, self.head_dim)).reshape(q.shape) |
| 262 | k = self.k_norm(k.reshape(-1, self.num_kv_heads, self.head_dim)).reshape( |
| 263 | k.shape |
| 264 | ) |
| 265 | |
| 266 | # Apply rotary embeddings only for local attention |
| 267 | if self.is_local_attention and self.rotary_emb is not None: |
| 268 | q, k = self.rotary_emb(positions, q, k) |
| 269 | |
| 270 | attn_output = self.attn(q, k, v) |
| 271 | |
| 272 | # Apply gating |
| 273 | attn_output = attn_output * torch.sigmoid(gate) |
| 274 | output, _ = self.o_proj(attn_output) |
| 275 | return output |
| 276 | |
| 277 | |
| 278 | class AfmoeDecoderLayer(nn.Module): |
| 279 | def __init__( |
| 280 | self, |
| 281 | config, # AfmoeConfig |
| 282 | cache_config: CacheConfig | None = None, |
| 283 | quant_config: QuantizationConfig | None = None, |
| 284 | prefix: str = "", |
| 285 | enable_eplb: bool = False, |
| 286 | ) -> None: |
| 287 | super().__init__() |
| 288 | self.hidden_size = config.hidden_size |
| 289 | max_position_embeddings = getattr(config, "max_position_embeddings", 131072) |
| 290 | |
| 291 | # DecoderLayers are created with `make_layers` which passes the prefix |
| 292 | # with the layer's index. |
| 293 | self.layer_idx = extract_layer_index(prefix) |
| 294 | |
| 295 | self.self_attn = AfmoeAttention( |
| 296 | config=config, |
| 297 | layer_idx=self.layer_idx, |
| 298 | hidden_size=self.hidden_size, |
| 299 | num_heads=config.num_attention_heads, |
| 300 | num_kv_heads=config.num_key_value_heads, |
| 301 | max_position_embeddings=max_position_embeddings, |
| 302 | head_dim=config.head_dim, |
| 303 | rms_norm_eps=config.rms_norm_eps, |
| 304 | cache_config=cache_config, |
| 305 | quant_config=quant_config, |
| 306 | prefix=f"{prefix}.self_attn", |
| 307 | ) |
| 308 | |
| 309 | # MoE or dense FFN |
| 310 | self.moe_enabled = self.layer_idx >= config.num_dense_layers |
| 311 | if self.moe_enabled: |
| 312 | self.mlp = AfmoeMoE( |
| 313 | config=config, |
| 314 | quant_config=quant_config, |
| 315 | prefix=f"{prefix}.mlp", |
| 316 | enable_eplb=enable_eplb, |
| 317 | ) |
| 318 | else: |
| 319 | self.mlp = AfmoeMLP( |
| 320 | hidden_size=config.hidden_size, |
| 321 | intermediate_size=config.intermediate_size, |
| 322 | hidden_act=config.hidden_act, |
| 323 | quant_config=quant_config, |
| 324 | prefix=f"{prefix}.mlp", |
| 325 | ) |
| 326 | |
| 327 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 328 | self.post_attention_layernorm = RMSNorm( |
| 329 | config.hidden_size, eps=config.rms_norm_eps |
| 330 | ) |
| 331 | self.pre_mlp_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 332 | self.post_mlp_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 333 | |
| 334 | def forward( |
| 335 | self, |
| 336 | positions: torch.Tensor, |
| 337 | hidden_states: torch.Tensor, |
| 338 | residual: torch.Tensor | None, |
| 339 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 340 | if residual is None: |
| 341 | residual = hidden_states |
| 342 | hidden_states = self.input_layernorm(hidden_states) |
| 343 | else: |
| 344 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 345 | |
| 346 | hidden_states = self.self_attn( |
| 347 | positions=positions, |
| 348 | hidden_states=hidden_states, |
| 349 | ) |
| 350 | hidden_states = self.post_attention_layernorm(hidden_states) # attn norm b |
| 351 | |
| 352 | # Fully Connected |
| 353 | hidden_states, residual = self.pre_mlp_layernorm( # ffn norm a |
| 354 | hidden_states, residual |
| 355 | ) |
| 356 | hidden_states = self.mlp(hidden_states) |
| 357 | hidden_states = self.post_mlp_layernorm(hidden_states) # ffn norm b |
| 358 | |
| 359 | return hidden_states, residual |
| 360 | |
| 361 | |
| 362 | @support_torch_compile( |
| 363 | dynamic_arg_dims={ |
| 364 | "input_ids": 0, |
| 365 | "positions": -1, |
| 366 | "intermediate_tensors": 0, |
| 367 | "inputs_embeds": 0, |
| 368 | } |
| 369 | ) |
| 370 | class AfmoeModel(nn.Module, EagleModelMixin): |
| 371 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 372 | super().__init__() |
| 373 | |
| 374 | config = vllm_config.model_config.hf_config |
| 375 | cache_config = vllm_config.cache_config |
| 376 | quant_config = vllm_config.quant_config |
| 377 | enable_eplb = vllm_config.parallel_config.enable_eplb |
| 378 | self.config = config |
| 379 | |
| 380 | self.vocab_size = config.vocab_size |
| 381 | self.mup_enabled = config.mup_enabled |
| 382 | |
| 383 | if get_pp_group().is_first_rank: |
| 384 | self.embed_tokens = VocabParallelEmbedding( |
| 385 | config.vocab_size, config.hidden_size, prefix=f"{prefix}.embed_tokens" |
| 386 | ) |
| 387 | else: |
| 388 | self.embed_tokens = PPMissingLayer() |
| 389 | |
| 390 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 391 | config.num_hidden_layers, |
| 392 | lambda prefix: AfmoeDecoderLayer( |
| 393 | config=config, |
| 394 | cache_config=cache_config, |
| 395 | quant_config=quant_config, |
| 396 | prefix=prefix, |
| 397 | enable_eplb=enable_eplb, |
| 398 | ), |
| 399 | prefix=f"{prefix}.layers", |
| 400 | ) |
| 401 | |
| 402 | if get_pp_group().is_last_rank: |
| 403 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 404 | else: |
| 405 | self.norm = PPMissingLayer() |
| 406 | |
| 407 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 408 | return self.embed_tokens(input_ids) |
| 409 | |
| 410 | def forward( |
| 411 | self, |
| 412 | input_ids: torch.Tensor | None, |
| 413 | positions: torch.Tensor, |
| 414 | intermediate_tensors: IntermediateTensors | None = None, |
| 415 | inputs_embeds: torch.Tensor | None = None, |
| 416 | ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]: |
| 417 | if get_pp_group().is_first_rank: |
| 418 | if inputs_embeds is not None: |
| 419 | hidden_states = inputs_embeds |
| 420 | else: |
| 421 | hidden_states = self.embed_input_ids(input_ids) |
| 422 | |
| 423 | # Apply muP input scaling if enabled |
| 424 | if self.mup_enabled: |
| 425 | hidden_states = hidden_states * (self.config.hidden_size**0.5) |
| 426 | |
| 427 | residual = None |
| 428 | else: |
| 429 | assert intermediate_tensors is not None |
| 430 | hidden_states = intermediate_tensors["hidden_states"] |
| 431 | residual = intermediate_tensors["residual"] |
| 432 | |
| 433 | aux_hidden_states = self._maybe_add_hidden_state([], 0, hidden_states, residual) |
| 434 | for idx, layer in enumerate( |
| 435 | islice(self.layers, self.start_layer, self.end_layer) |
| 436 | ): |
| 437 | hidden_states, residual = layer(positions, hidden_states, residual) |
| 438 | self._maybe_add_hidden_state( |
| 439 | aux_hidden_states, idx + 1, hidden_states, residual |
| 440 | ) |
| 441 | |
| 442 | if not get_pp_group().is_last_rank: |
| 443 | return IntermediateTensors( |
| 444 | {"hidden_states": hidden_states, "residual": residual} |
| 445 | ) |
| 446 | |
| 447 | hidden_states, _ = self.norm(hidden_states, residual) |
| 448 | |
| 449 | if len(aux_hidden_states) > 0: |
| 450 | return hidden_states, aux_hidden_states |
| 451 | |
| 452 | return hidden_states |
| 453 | |
| 454 | def make_empty_intermediate_tensors( |
| 455 | self, batch_size: int, dtype: torch.dtype, device: torch.device |
| 456 | ) -> IntermediateTensors: |
| 457 | return IntermediateTensors( |
| 458 | { |
| 459 | "hidden_states": torch.zeros( |
| 460 | (batch_size, self.config.hidden_size), dtype=dtype, device=device |
| 461 | ), |
| 462 | "residual": torch.zeros( |
| 463 | (batch_size, self.config.hidden_size), dtype=dtype, device=device |
| 464 | ), |
| 465 | } |
| 466 | ) |
| 467 | |
| 468 | |
| 469 | class AfmoeForCausalLM( |
| 470 | nn.Module, SupportsPP, SupportsEagle3, SupportsLoRA, MixtureOfExperts |
| 471 | ): |
| 472 | packed_modules_mapping = { |
| 473 | "qkv_proj": [ |
| 474 | "q_proj", |
| 475 | "k_proj", |
| 476 | "v_proj", |
| 477 | ], |
| 478 | "gate_up_proj": [ |
| 479 | "gate_proj", |
| 480 | "up_proj", |
| 481 | ], |
| 482 | } |
| 483 | |
| 484 | hf_to_vllm_mapper = WeightsMapper( |
| 485 | orig_to_new_suffix={ |
| 486 | ".router.gate.weight": ".gate.weight", |
| 487 | }, |
| 488 | orig_to_new_stacked={ |
| 489 | # weight_name: (param_name, shard_id) |
| 490 | ".q_proj": (".qkv_proj", "q"), |
| 491 | ".k_proj": (".qkv_proj", "k"), |
| 492 | ".v_proj": (".qkv_proj", "v"), |
| 493 | # `.self_attn.gate_proj` is a gated-attention projection (not fused). |
| 494 | ".mlp.gate_proj": (".mlp.gate_up_proj", 0), |
| 495 | ".mlp.up_proj": (".mlp.gate_up_proj", 1), |
| 496 | ".shared_experts.gate_proj": (".shared_experts.gate_up_proj", 0), |
| 497 | ".shared_experts.up_proj": (".shared_experts.gate_up_proj", 1), |
| 498 | }, |
| 499 | ) |
| 500 | |
| 501 | fall_back_to_pt_during_load = False |
| 502 | |
| 503 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 504 | super().__init__() |
| 505 | config = vllm_config.model_config.hf_config |
| 506 | quant_config = vllm_config.quant_config |
| 507 | self.config = config |
| 508 | self.quant_config = quant_config |
| 509 | self.model = AfmoeModel( |
| 510 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 511 | ) |
| 512 | if get_pp_group().is_last_rank: |
| 513 | self.lm_head = ParallelLMHead( |
| 514 | config.vocab_size, config.hidden_size, quant_config=quant_config |
| 515 | ) |
| 516 | else: |
| 517 | self.lm_head = PPMissingLayer() |
| 518 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 519 | self.make_empty_intermediate_tensors = ( |
| 520 | self.model.make_empty_intermediate_tensors |
| 521 | ) |
| 522 | # Set MoE hyperparameters |
| 523 | self.num_moe_layers = config.num_hidden_layers - config.num_dense_layers |
| 524 | self.num_expert_groups = config.n_group |
| 525 | |
| 526 | self.moe_layers: list[MoERunner] = [] |
| 527 | example_moe = None |
| 528 | for layer in self.model.layers: |
| 529 | if isinstance(layer, PPMissingLayer): |
| 530 | continue |
| 531 | |
| 532 | assert isinstance(layer, AfmoeDecoderLayer) |
| 533 | if layer.moe_enabled: |
| 534 | example_moe = layer.mlp |
| 535 | self.moe_layers.append(layer.mlp.experts) |
| 536 | |
| 537 | if example_moe is None and self.num_moe_layers > 0: |
| 538 | raise RuntimeError("No AfmoeMoE layer found in model.layers.") |
| 539 | |
| 540 | if example_moe is not None: |
| 541 | self.num_logical_experts = example_moe.n_logical_experts |
| 542 | self.num_physical_experts = example_moe.n_physical_experts |
| 543 | self.num_local_physical_experts = example_moe.n_local_physical_experts |
| 544 | self.num_routed_experts = example_moe.n_routed_experts |
| 545 | self.num_shared_experts = example_moe.n_shared_experts |
| 546 | self.num_redundant_experts = example_moe.n_redundant_experts |
| 547 | |
| 548 | def update_physical_experts_metadata( |
| 549 | self, |
| 550 | num_physical_experts: int, |
| 551 | num_local_physical_experts: int, |
| 552 | ) -> None: |
| 553 | assert self.num_local_physical_experts == num_local_physical_experts |
| 554 | self.num_physical_experts = num_physical_experts |
| 555 | self.num_local_physical_experts = num_local_physical_experts |
| 556 | self.num_redundant_experts = num_physical_experts - self.num_logical_experts |
| 557 | for layer in self.model.layers: |
| 558 | if isinstance(layer, PPMissingLayer): |
| 559 | continue |
| 560 | if layer.moe_enabled: |
| 561 | moe = layer.mlp |
| 562 | moe.n_local_physical_experts = num_local_physical_experts |
| 563 | moe.n_physical_experts = num_physical_experts |
| 564 | moe.n_redundant_experts = self.num_redundant_experts |
| 565 | moe.experts.update_expert_map() |
| 566 | |
| 567 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 568 | return self.model.embed_input_ids(input_ids) |
| 569 | |
| 570 | def forward( |
| 571 | self, |
| 572 | input_ids: torch.Tensor | None, |
| 573 | positions: torch.Tensor, |
| 574 | intermediate_tensors: IntermediateTensors | None = None, |
| 575 | inputs_embeds: torch.Tensor | None = None, |
| 576 | ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]: |
| 577 | hidden_states = self.model( |
| 578 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 579 | ) |
| 580 | return hidden_states |
| 581 | |
| 582 | def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor | None: |
| 583 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 584 | return logits |
| 585 | |
| 586 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 587 | loader = AutoWeightsLoader(self) |
| 588 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |