| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Inference-only Intern-S2-Mobius 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 VllmConfig, replace, set_current_vllm_config |
| 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.fused_moe import ( |
| 21 | FusedMoEFactory, |
| 22 | GateLinear, |
| 23 | ) |
| 24 | from vllm.model_executor.layers.fusion.mm_input_norm import build_mm_input_norm |
| 25 | from vllm.model_executor.layers.linear import ReplicatedLinear |
| 26 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 27 | from vllm.model_executor.layers.mamba.gdn.qwen_gdn_linear_attn import ( |
| 28 | QwenGatedDeltaNetAttention, |
| 29 | ) |
| 30 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 31 | ParallelLMHead, |
| 32 | VocabParallelEmbedding, |
| 33 | ) |
| 34 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 35 | from vllm.sequence import IntermediateTensors |
| 36 | from vllm.tokenizers.registry import cached_tokenizer_from_config |
| 37 | |
| 38 | from .interfaces import EagleModelMixin |
| 39 | from .interns2_preview import ( |
| 40 | InternS2PreviewForConditionalGeneration, |
| 41 | InternS2PreviewProcessingInfo, |
| 42 | ) |
| 43 | from .qwen2_moe import Qwen2MoeMLP |
| 44 | from .qwen3_5 import ( |
| 45 | Qwen3_5ForCausalLMBase, |
| 46 | Qwen3_5Model, |
| 47 | Qwen3_5RMSNorm, |
| 48 | ) |
| 49 | from .qwen3_5_mtp import Qwen3_5MoeMTP |
| 50 | from .qwen3_next import Qwen3NextAttention |
| 51 | from .qwen3_vl import ( |
| 52 | Qwen3_VisionTransformer, |
| 53 | Qwen3VLDummyInputsBuilder, |
| 54 | Qwen3VLMultiModalProcessor, |
| 55 | ) |
| 56 | from .utils import ( |
| 57 | AutoWeightsLoader, |
| 58 | PPMissingLayer, |
| 59 | extract_layer_index, |
| 60 | make_empty_intermediate_tensors_factory, |
| 61 | make_layers, |
| 62 | maybe_prefix, |
| 63 | sequence_parallel_chunk, |
| 64 | ) |
| 65 | |
| 66 | |
| 67 | class InternS2MobiusProcessingInfo(InternS2PreviewProcessingInfo): |
| 68 | pass |
| 69 | |
| 70 | |
| 71 | class InternS2MobiusMetaMoeBlock(nn.Module): |
| 72 | """A routed MoE bank shared by multiple decoder layers.""" |
| 73 | |
| 74 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 75 | super().__init__() |
| 76 | |
| 77 | config = vllm_config.model_config.hf_text_config |
| 78 | parallel_config = vllm_config.parallel_config |
| 79 | |
| 80 | self.tp_size = get_tensor_model_parallel_world_size() |
| 81 | self.ep_group = get_ep_group().device_group |
| 82 | self.ep_rank = get_ep_group().rank_in_group |
| 83 | self.ep_size = self.ep_group.size() |
| 84 | self.n_routed_experts = config.num_experts |
| 85 | self.is_sequence_parallel = parallel_config.use_sequence_parallel_moe |
| 86 | |
| 87 | if self.tp_size > self.n_routed_experts: |
| 88 | raise ValueError( |
| 89 | f"Tensor parallel size {self.tp_size} is greater than " |
| 90 | f"the number of experts {self.n_routed_experts}." |
| 91 | ) |
| 92 | |
| 93 | eplb_config = parallel_config.eplb_config |
| 94 | self.enable_eplb = parallel_config.enable_eplb |
| 95 | self.n_logical_experts = self.n_routed_experts |
| 96 | self.n_redundant_experts = eplb_config.num_redundant_experts |
| 97 | self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts |
| 98 | self.n_local_physical_experts = self.n_physical_experts // self.ep_size |
| 99 | |
| 100 | self.gate = GateLinear( |
| 101 | config.hidden_size, |
| 102 | config.num_experts, |
| 103 | prefix=f"{prefix}.gate", |
| 104 | ) |
| 105 | self.experts = FusedMoEFactory( |
| 106 | gate=self.gate, |
| 107 | num_experts=config.num_experts, |
| 108 | top_k=config.num_experts_per_tok, |
| 109 | hidden_size=config.hidden_size, |
| 110 | intermediate_size=config.moe_intermediate_size, |
| 111 | renormalize=getattr(config, "norm_topk_prob", True), |
| 112 | quant_config=vllm_config.quant_config, |
| 113 | prefix=f"{prefix}.experts", |
| 114 | enable_eplb=self.enable_eplb, |
| 115 | num_redundant_experts=self.n_redundant_experts, |
| 116 | is_sequence_parallel=self.is_sequence_parallel, |
| 117 | activation=config.hidden_act, |
| 118 | ) |
| 119 | |
| 120 | def forward( |
| 121 | self, |
| 122 | hidden_states: torch.Tensor, |
| 123 | already_sequence_parallel: bool = False, |
| 124 | ) -> torch.Tensor: |
| 125 | orig_shape = hidden_states.shape |
| 126 | num_tokens, hidden_dim = hidden_states.shape |
| 127 | hidden_states = hidden_states.view(-1, hidden_dim) |
| 128 | |
| 129 | if self.is_sequence_parallel and not already_sequence_parallel: |
| 130 | hidden_states = sequence_parallel_chunk(hidden_states) |
| 131 | |
| 132 | final_hidden_states = self.experts( |
| 133 | hidden_states=hidden_states, |
| 134 | router_logits=hidden_states, |
| 135 | ) |
| 136 | |
| 137 | if self.is_sequence_parallel and not already_sequence_parallel: |
| 138 | final_hidden_states = tensor_model_parallel_all_gather( |
| 139 | final_hidden_states, 0 |
| 140 | ) |
| 141 | final_hidden_states = final_hidden_states[:num_tokens] |
| 142 | |
| 143 | return final_hidden_states.view(orig_shape) |
| 144 | |
| 145 | |
| 146 | class InternS2MobiusSharedExpertBlock(nn.Module): |
| 147 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 148 | super().__init__() |
| 149 | |
| 150 | config = vllm_config.model_config.hf_text_config |
| 151 | is_sequence_parallel = vllm_config.parallel_config.use_sequence_parallel_moe |
| 152 | self.shared_expert_gate = ReplicatedLinear( |
| 153 | config.hidden_size, |
| 154 | 1, |
| 155 | bias=False, |
| 156 | quant_config=None, |
| 157 | prefix=f"{prefix}.shared_expert_gate", |
| 158 | ) |
| 159 | self.shared_expert = Qwen2MoeMLP( |
| 160 | hidden_size=config.hidden_size, |
| 161 | intermediate_size=config.shared_expert_intermediate_size, |
| 162 | hidden_act=config.hidden_act, |
| 163 | quant_config=vllm_config.quant_config, |
| 164 | expert_gate=self.shared_expert_gate, |
| 165 | is_sequence_parallel=is_sequence_parallel, |
| 166 | prefix=f"{prefix}.shared_expert", |
| 167 | ) |
| 168 | |
| 169 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 170 | return self.shared_expert(hidden_states) |
| 171 | |
| 172 | |
| 173 | class InternS2MobiusDecoderLayer(nn.Module): |
| 174 | def __init__( |
| 175 | self, |
| 176 | vllm_config: VllmConfig, |
| 177 | layer_type: str, |
| 178 | prefix: str = "", |
| 179 | ) -> None: |
| 180 | super().__init__() |
| 181 | |
| 182 | config = vllm_config.model_config.hf_text_config |
| 183 | model_config = vllm_config.model_config |
| 184 | parallel_config = vllm_config.parallel_config |
| 185 | |
| 186 | self.layer_type = layer_type |
| 187 | self.layer_idx = extract_layer_index(prefix) |
| 188 | self.num_blocks = config.num_blocks |
| 189 | self.use_attn_reduce_scatter_for_moe = ( |
| 190 | parallel_config.use_sequence_parallel_moe |
| 191 | and parallel_config.pipeline_parallel_size == 1 |
| 192 | ) |
| 193 | |
| 194 | if self.layer_type == "linear_attention": |
| 195 | self.linear_attn = QwenGatedDeltaNetAttention( |
| 196 | config=config, |
| 197 | vllm_config=vllm_config, |
| 198 | prefix=f"{prefix}.linear_attn", |
| 199 | gqa_interleaved_layout=False, |
| 200 | reduce_results=not self.use_attn_reduce_scatter_for_moe, |
| 201 | ) |
| 202 | elif self.layer_type == "full_attention": |
| 203 | self.self_attn = Qwen3NextAttention( |
| 204 | config, |
| 205 | model_config=model_config, |
| 206 | cache_config=vllm_config.cache_config, |
| 207 | quant_config=vllm_config.quant_config, |
| 208 | prefix=f"{prefix}.self_attn", |
| 209 | reduce_results=not self.use_attn_reduce_scatter_for_moe, |
| 210 | ) |
| 211 | else: |
| 212 | raise ValueError(f"Invalid layer_type {self.layer_type}") |
| 213 | |
| 214 | self.mlp = InternS2MobiusSharedExpertBlock( |
| 215 | vllm_config=vllm_config, |
| 216 | prefix=f"{prefix}.mlp", |
| 217 | ) |
| 218 | self.input_layernorm = Qwen3_5RMSNorm( |
| 219 | config.hidden_size, eps=config.rms_norm_eps |
| 220 | ) |
| 221 | self.post_attention_layernorm = Qwen3_5RMSNorm( |
| 222 | config.hidden_size, eps=config.rms_norm_eps |
| 223 | ) |
| 224 | |
| 225 | self.layer_scale = getattr(config, "layer_scale", False) |
| 226 | if self.layer_scale: |
| 227 | self.attn_layer_scale = nn.Parameter(torch.zeros(1, 1, config.hidden_size)) |
| 228 | self.ffn_layer_scale = nn.Parameter(torch.zeros(1, 1, config.hidden_size)) |
| 229 | |
| 230 | def forward( |
| 231 | self, |
| 232 | hidden_states: torch.Tensor, |
| 233 | residual: torch.Tensor | None, |
| 234 | positions: torch.Tensor, |
| 235 | meta_mlp: nn.ModuleList, |
| 236 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 237 | full_num_tokens = positions.shape[-1] |
| 238 | |
| 239 | if residual is None: |
| 240 | residual = hidden_states |
| 241 | hidden_states = self.input_layernorm(hidden_states) |
| 242 | else: |
| 243 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 244 | |
| 245 | if self.use_attn_reduce_scatter_for_moe: |
| 246 | hidden_states = tensor_model_parallel_all_gather(hidden_states, 0) |
| 247 | hidden_states = hidden_states[:full_num_tokens] |
| 248 | |
| 249 | if self.layer_type == "linear_attention": |
| 250 | hidden_states = self.linear_attn(hidden_states=hidden_states) |
| 251 | else: |
| 252 | hidden_states = self.self_attn( |
| 253 | hidden_states=hidden_states, |
| 254 | positions=positions, |
| 255 | ) |
| 256 | |
| 257 | if self.layer_scale: |
| 258 | scale = self.attn_layer_scale.to(hidden_states.dtype) |
| 259 | hidden_states = hidden_states * (scale[0] + 1) |
| 260 | |
| 261 | if self.use_attn_reduce_scatter_for_moe: |
| 262 | tp_world_size = get_tensor_model_parallel_world_size() |
| 263 | sp_pad = (-hidden_states.shape[0]) % tp_world_size |
| 264 | hidden_states = torch.nn.functional.pad(hidden_states, (0, 0, 0, sp_pad)) |
| 265 | hidden_states = tensor_model_parallel_reduce_scatter(hidden_states, 0) |
| 266 | |
| 267 | hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) |
| 268 | block_idx = self.layer_idx % self.num_blocks |
| 269 | routed_output = meta_mlp[block_idx]( |
| 270 | hidden_states, |
| 271 | already_sequence_parallel=self.use_attn_reduce_scatter_for_moe, |
| 272 | ) |
| 273 | hidden_states = routed_output + self.mlp(hidden_states) |
| 274 | |
| 275 | if self.layer_scale: |
| 276 | scale = self.ffn_layer_scale.to(hidden_states.dtype) |
| 277 | hidden_states = hidden_states * (scale[0] + 1) |
| 278 | |
| 279 | return hidden_states, residual |
| 280 | |
| 281 | |
| 282 | @support_torch_compile( |
| 283 | dynamic_arg_dims={ |
| 284 | "input_ids": 0, |
| 285 | "positions": -1, |
| 286 | "intermediate_tensors": 0, |
| 287 | "inputs_embeds": 0, |
| 288 | } |
| 289 | ) |
| 290 | class InternS2MobiusModel(nn.Module, EagleModelMixin): |
| 291 | hf_to_vllm_mapper = Qwen3_5Model.hf_to_vllm_mapper |
| 292 | |
| 293 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 294 | nn.Module.__init__(self) |
| 295 | |
| 296 | config = vllm_config.model_config.hf_text_config |
| 297 | self.config = config |
| 298 | self.quant_config = vllm_config.quant_config |
| 299 | self.num_redundant_experts = ( |
| 300 | vllm_config.parallel_config.eplb_config.num_redundant_experts |
| 301 | ) |
| 302 | self.vocab_size = config.vocab_size |
| 303 | self.embed_tokens = VocabParallelEmbedding( |
| 304 | self.vocab_size, |
| 305 | config.hidden_size, |
| 306 | ) |
| 307 | self.meta_mlp = nn.ModuleList( |
| 308 | InternS2MobiusMetaMoeBlock( |
| 309 | vllm_config=vllm_config, |
| 310 | prefix=f"{prefix}.meta_mlp.{idx}", |
| 311 | ) |
| 312 | for idx in range(config.num_blocks) |
| 313 | ) |
| 314 | |
| 315 | def get_layer(prefix: str) -> InternS2MobiusDecoderLayer: |
| 316 | return InternS2MobiusDecoderLayer( |
| 317 | vllm_config, |
| 318 | layer_type=config.layer_types[extract_layer_index(prefix)], |
| 319 | prefix=prefix, |
| 320 | ) |
| 321 | |
| 322 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 323 | config.num_hidden_layers, |
| 324 | get_layer, |
| 325 | prefix=f"{prefix}.layers", |
| 326 | ) |
| 327 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 328 | ["hidden_states", "residual"], config.hidden_size |
| 329 | ) |
| 330 | |
| 331 | if get_pp_group().is_last_rank: |
| 332 | self.norm = Qwen3_5RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 333 | else: |
| 334 | self.norm = PPMissingLayer() |
| 335 | |
| 336 | self.aux_hidden_state_layers: tuple[int, ...] = () |
| 337 | |
| 338 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 339 | return self.embed_tokens(input_ids) |
| 340 | |
| 341 | @property |
| 342 | def use_sequence_parallel(self) -> bool: |
| 343 | return self.layers[self.start_layer].use_attn_reduce_scatter_for_moe |
| 344 | |
| 345 | def forward( |
| 346 | self, |
| 347 | input_ids: torch.Tensor | None, |
| 348 | positions: torch.Tensor, |
| 349 | intermediate_tensors: IntermediateTensors | None = None, |
| 350 | inputs_embeds: torch.Tensor | None = None, |
| 351 | ) -> torch.Tensor | IntermediateTensors | tuple[torch.Tensor, list[torch.Tensor]]: |
| 352 | if get_pp_group().is_first_rank: |
| 353 | hidden_states = ( |
| 354 | inputs_embeds |
| 355 | if inputs_embeds is not None |
| 356 | else self.embed_input_ids(input_ids) |
| 357 | ) |
| 358 | residual = None |
| 359 | else: |
| 360 | assert intermediate_tensors is not None |
| 361 | hidden_states = intermediate_tensors["hidden_states"] |
| 362 | residual = intermediate_tensors["residual"] |
| 363 | |
| 364 | full_num_tokens = positions.shape[-1] |
| 365 | if self.use_sequence_parallel: |
| 366 | hidden_states = sequence_parallel_chunk(hidden_states) |
| 367 | assert residual is None |
| 368 | |
| 369 | aux_hidden_states = self._maybe_add_hidden_state([], 0, hidden_states, residual) |
| 370 | for layer_idx, layer in enumerate( |
| 371 | islice(self.layers, self.start_layer, self.end_layer), |
| 372 | start=self.start_layer, |
| 373 | ): |
| 374 | hidden_states, residual = layer( |
| 375 | positions=positions, |
| 376 | hidden_states=hidden_states, |
| 377 | residual=residual, |
| 378 | meta_mlp=self.meta_mlp, |
| 379 | ) |
| 380 | self._maybe_add_hidden_state( |
| 381 | aux_hidden_states, |
| 382 | layer_idx + 1, |
| 383 | hidden_states, |
| 384 | residual, |
| 385 | ) |
| 386 | |
| 387 | if not get_pp_group().is_last_rank: |
| 388 | return IntermediateTensors( |
| 389 | {"hidden_states": hidden_states, "residual": residual} |
| 390 | ) |
| 391 | hidden_states, _ = self.norm(hidden_states, residual) |
| 392 | if self.use_sequence_parallel: |
| 393 | if aux_hidden_states: |
| 394 | hidden_size = hidden_states.shape[-1] |
| 395 | hidden_states = torch.cat([hidden_states, *aux_hidden_states], dim=-1) |
| 396 | hidden_states = tensor_model_parallel_all_gather(hidden_states, 0) |
| 397 | hidden_states = hidden_states[:full_num_tokens] |
| 398 | hidden_states, *aux_hidden_states = hidden_states.split( |
| 399 | hidden_size, dim=-1 |
| 400 | ) |
| 401 | else: |
| 402 | hidden_states = tensor_model_parallel_all_gather(hidden_states, 0) |
| 403 | hidden_states = hidden_states[:full_num_tokens] |
| 404 | if aux_hidden_states: |
| 405 | return hidden_states, aux_hidden_states |
| 406 | return hidden_states |
| 407 | |
| 408 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 409 | loader = AutoWeightsLoader(self) |
| 410 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 411 | |
| 412 | |
| 413 | class InternS2MobiusForCausalLM(Qwen3_5ForCausalLMBase): |
| 414 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 415 | config = vllm_config.model_config.hf_text_config |
| 416 | self.vllm_config = vllm_config |
| 417 | self.model_config = vllm_config.model_config |
| 418 | self.quant_config = vllm_config.quant_config |
| 419 | self.scheduler_config = vllm_config.scheduler_config |
| 420 | |
| 421 | nn.Module.__init__(self) |
| 422 | self.config = config |
| 423 | self.model = InternS2MobiusModel( |
| 424 | vllm_config=vllm_config, |
| 425 | prefix=maybe_prefix(prefix, "model"), |
| 426 | ) |
| 427 | |
| 428 | if get_pp_group().is_last_rank: |
| 429 | self.lm_head = ParallelLMHead( |
| 430 | config.vocab_size, |
| 431 | config.hidden_size, |
| 432 | quant_config=self.quant_config, |
| 433 | prefix=maybe_prefix(prefix, "lm_head"), |
| 434 | ) |
| 435 | if config.tie_word_embeddings: |
| 436 | self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) |
| 437 | else: |
| 438 | self.lm_head = PPMissingLayer() |
| 439 | |
| 440 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 441 | self.make_empty_intermediate_tensors = ( |
| 442 | self.model.make_empty_intermediate_tensors |
| 443 | ) |
| 444 | |
| 445 | |
| 446 | class InternS2MobiusMTP(Qwen3_5MoeMTP): |
| 447 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 448 | speculative_config = vllm_config.speculative_config |
| 449 | if speculative_config is None or speculative_config.draft_model_config is None: |
| 450 | raise ValueError( |
| 451 | "Intern-S2-Mobius MTP requires a draft model configuration." |
| 452 | ) |
| 453 | |
| 454 | draft_vllm_config = replace( |
| 455 | vllm_config, |
| 456 | model_config=speculative_config.draft_model_config, |
| 457 | ) |
| 458 | with set_current_vllm_config(draft_vllm_config, prefix=prefix): |
| 459 | super().__init__(vllm_config=draft_vllm_config, prefix=prefix) |
| 460 | |
| 461 | |
| 462 | @MULTIMODAL_REGISTRY.register_processor( |
| 463 | Qwen3VLMultiModalProcessor, |
| 464 | info=InternS2MobiusProcessingInfo, |
| 465 | dummy_inputs=Qwen3VLDummyInputsBuilder, |
| 466 | ) |
| 467 | class InternS2MobiusForConditionalGeneration(InternS2PreviewForConditionalGeneration): |
| 468 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "model") -> None: |
| 469 | nn.Module.__init__(self) |
| 470 | |
| 471 | config = vllm_config.model_config.hf_config |
| 472 | if getattr(config, "image_token_index", None) is None: |
| 473 | config.image_token_index = config.image_token_id |
| 474 | quant_config = vllm_config.quant_config |
| 475 | multimodal_config = vllm_config.model_config.get_multimodal_config() |
| 476 | |
| 477 | self.config = config |
| 478 | self.model_config = vllm_config.model_config |
| 479 | self.multimodal_config = multimodal_config |
| 480 | self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data" |
| 481 | self.is_multimodal_pruning_enabled = ( |
| 482 | multimodal_config.is_multimodal_pruning_enabled() |
| 483 | ) |
| 484 | self.video_pruning_rate = multimodal_config.video_pruning_rate |
| 485 | self._tokenizer = cached_tokenizer_from_config(vllm_config.model_config) |
| 486 | |
| 487 | self.use_deepstack = False |
| 488 | |
| 489 | with self._mark_tower_model(vllm_config, {"image", "video"}): |
| 490 | self.visual = Qwen3_VisionTransformer( |
| 491 | config.vision_config, |
| 492 | norm_eps=getattr(config, "rms_norm_eps", 1e-6), |
| 493 | quant_config=quant_config, |
| 494 | input_norm=build_mm_input_norm(self.model_config), |
| 495 | prefix=maybe_prefix(prefix, "visual"), |
| 496 | ) |
| 497 | |
| 498 | with self._mark_language_model(vllm_config): |
| 499 | self.language_model = InternS2MobiusForCausalLM( |
| 500 | vllm_config=vllm_config, |
| 501 | prefix=maybe_prefix(prefix, "language_model"), |
| 502 | ) |
| 503 | |
| 504 | self.make_empty_intermediate_tensors = ( |
| 505 | self.language_model.make_empty_intermediate_tensors |
| 506 | ) |
| 507 | self.set_moe_parameters() |
| 508 | |
| 509 | def set_moe_parameters(self) -> None: |
| 510 | meta_mlp = self.language_model.model.meta_mlp |
| 511 | if not meta_mlp: |
| 512 | raise RuntimeError("Intern-S2-Mobius requires at least one MoE bank.") |
| 513 | |
| 514 | self.moe_layers = [moe.experts for moe in meta_mlp] |
| 515 | example_moe = meta_mlp[0] |
| 516 | self.num_moe_layers = len(self.moe_layers) |
| 517 | self.num_expert_groups = 1 |
| 518 | self.num_shared_experts = 0 |
| 519 | self.num_logical_experts = example_moe.n_logical_experts |
| 520 | self.num_physical_experts = example_moe.n_physical_experts |
| 521 | self.num_local_physical_experts = example_moe.n_local_physical_experts |
| 522 | self.num_routed_experts = example_moe.n_routed_experts |
| 523 | self.num_redundant_experts = example_moe.n_redundant_experts |
| 524 | |
| 525 | def update_physical_experts_metadata( |
| 526 | self, |
| 527 | num_physical_experts: int, |
| 528 | num_local_physical_experts: int, |
| 529 | ) -> None: |
| 530 | assert self.num_local_physical_experts == num_local_physical_experts |
| 531 | self.num_physical_experts = num_physical_experts |
| 532 | self.num_local_physical_experts = num_local_physical_experts |
| 533 | self.num_redundant_experts = num_physical_experts - self.num_logical_experts |
| 534 | for moe in self.language_model.model.meta_mlp: |
| 535 | moe.n_local_physical_experts = num_local_physical_experts |
| 536 | moe.n_physical_experts = num_physical_experts |
| 537 | moe.n_redundant_experts = self.num_redundant_experts |
| 538 | moe.experts.update_expert_map() |