| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Inkling model implementation for AMD GPUs.""" |
| 4 | |
| 5 | from __future__ import annotations |
| 6 | |
| 7 | from collections.abc import Iterable |
| 8 | from typing import Any |
| 9 | |
| 10 | import regex as re |
| 11 | import torch |
| 12 | from torch import nn |
| 13 | |
| 14 | from vllm.config import VllmConfig |
| 15 | from vllm.distributed import ( |
| 16 | get_pp_group, |
| 17 | get_tensor_model_parallel_rank, |
| 18 | get_tensor_model_parallel_world_size, |
| 19 | tensor_model_parallel_all_gather, |
| 20 | tensor_model_parallel_reduce_scatter, |
| 21 | ) |
| 22 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 23 | from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead |
| 24 | from vllm.model_executor.models.interfaces import ( |
| 25 | MultiModalEmbeddings, |
| 26 | SupportsLoRA, |
| 27 | SupportsMultiModal, |
| 28 | SupportsPP, |
| 29 | ) |
| 30 | from vllm.model_executor.models.utils import ( |
| 31 | AutoWeightsLoader, |
| 32 | WeightsMapper, |
| 33 | make_empty_intermediate_tensors_factory, |
| 34 | make_layers, |
| 35 | maybe_prefix, |
| 36 | ) |
| 37 | from vllm.models.inkling.common.mm_preprocess import ( |
| 38 | InklingDummyInputsBuilder, |
| 39 | InklingMultiModalProcessor, |
| 40 | InklingProcessingInfo, |
| 41 | inkling_audio_enabled, |
| 42 | inkling_vision_enabled, |
| 43 | ) |
| 44 | from vllm.models.inkling.common.towers import InklingAudio, InklingVision |
| 45 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 46 | from vllm.sequence import IntermediateTensors |
| 47 | |
| 48 | from ..configs import InklingMMConfig, InklingModelConfig |
| 49 | from .attention import InklingAttention, compute_log_scaling_tau |
| 50 | from .layernorm import InklingRMSNorm |
| 51 | from .logits_processor import InklingLogitsProcessor |
| 52 | from .mlp import InklingDenseMLP |
| 53 | from .moe import InklingMoE |
| 54 | from .ops.norm import add_rmsnorm, embed_rmsnorm |
| 55 | from .sconv_swa_attn import _ATTN, _MLP, InklingConvState |
| 56 | from .short_conv import InklingShortConv |
| 57 | |
| 58 | |
| 59 | def _layer_id(name: str) -> int | None: |
| 60 | m = re.search(r"\.layers\.(\d+)\.", name) |
| 61 | return int(m.group(1)) if m else None |
| 62 | |
| 63 | |
| 64 | def _sconv_add_norm( |
| 65 | delta: torch.Tensor, |
| 66 | hidden: torch.Tensor, |
| 67 | sconv: InklingShortConv, |
| 68 | norm: InklingRMSNorm | None, |
| 69 | positions: torch.Tensor, |
| 70 | ) -> tuple[torch.Tensor | None, torch.Tensor]: |
| 71 | """``h = hidden + sconv(TP-sum(delta)); y = rmsnorm(h)``. |
| 72 | |
| 73 | ROCm uses the portable collective path. The Lamport P2P implementation in |
| 74 | the NVIDIA model relies on CUDA GDC and CUDA-specific symmetric-memory |
| 75 | publication semantics which cannot be linked into a gfx950 HSACO.""" |
| 76 | norm_w = norm.weight if norm is not None else None |
| 77 | eps = norm.variance_epsilon if norm is not None else 0.0 |
| 78 | |
| 79 | # RCCL RS -> shard sconv -> AG -> fused add(+rmsnorm). |
| 80 | shard = tensor_model_parallel_reduce_scatter(delta, dim=-1) |
| 81 | shard = sconv(shard.contiguous(), positions) |
| 82 | full = tensor_model_parallel_all_gather(shard, dim=-1) |
| 83 | if norm is None: |
| 84 | return None, hidden + full |
| 85 | return add_rmsnorm(hidden, full, norm_w, eps) |
| 86 | |
| 87 | |
| 88 | class InklingDecoderLayer(nn.Module): |
| 89 | def __init__( |
| 90 | self, |
| 91 | config: InklingModelConfig, |
| 92 | layer_id: int, |
| 93 | is_local: bool, |
| 94 | quant_config: QuantizationConfig | None, |
| 95 | prefix: str, |
| 96 | force_dense_mlp: bool = False, |
| 97 | ) -> None: |
| 98 | super().__init__() |
| 99 | # Per-layer owner of the conv state as a paged SWA cache. The 4 sconv |
| 100 | # streams (K/V/attn/mlp) are packed head-major into one block and share |
| 101 | # it. Built first so the attention layer can wire its K/V sconv to it. |
| 102 | self.conv_state = InklingConvState( |
| 103 | num_kv_heads=( |
| 104 | config.swa_num_key_value_heads |
| 105 | if is_local |
| 106 | else config.num_key_value_heads |
| 107 | ), |
| 108 | head_dim=config.swa_head_dim if is_local else config.head_dim, |
| 109 | hidden_size=config.hidden_size, |
| 110 | kernel_size=config.sconv_kernel_size, |
| 111 | prefix=f"{prefix}.conv_state", |
| 112 | ) |
| 113 | self.attn_norm = InklingRMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 114 | self.attn = InklingAttention( |
| 115 | config, |
| 116 | num_heads=( |
| 117 | config.swa_num_attention_heads |
| 118 | if is_local |
| 119 | else config.num_attention_heads |
| 120 | ), |
| 121 | num_kv_heads=( |
| 122 | config.swa_num_key_value_heads |
| 123 | if is_local |
| 124 | else config.num_key_value_heads |
| 125 | ), |
| 126 | head_dim=config.swa_head_dim if is_local else config.head_dim, |
| 127 | rel_extent=config.rel_extent, |
| 128 | local_extent=config.sliding_window_size, |
| 129 | is_local=is_local, |
| 130 | prefix=f"{prefix}.attn", |
| 131 | quant_config=quant_config, |
| 132 | conv_owner=self.conv_state, |
| 133 | ) |
| 134 | self.mlp_norm = InklingRMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 135 | if force_dense_mlp or layer_id < config.dense_mlp_idx: |
| 136 | self.mlp: nn.Module = InklingDenseMLP( |
| 137 | hidden_size=config.hidden_size, |
| 138 | intermediate_size=config.dense_intermediate_size, |
| 139 | use_global_scale=config.use_global_scale, |
| 140 | quant_config=quant_config, |
| 141 | prefix=f"{prefix}.mlp", |
| 142 | ) |
| 143 | else: |
| 144 | # InklingMoE decides per layer (from the checkpoint exclude list) |
| 145 | # whether the routed experts are NVFP4 or bf16; the shared sink |
| 146 | # experts are always bf16. |
| 147 | self.mlp = InklingMoE( |
| 148 | config, |
| 149 | prefix=f"{prefix}.mlp", |
| 150 | quant_config=quant_config, |
| 151 | ) |
| 152 | |
| 153 | # Short convolution on the attention-output and MLP-output residual |
| 154 | # streams, hidden-sharded: the sublayer outputs are reduce-scattered |
| 155 | # to [T, H/tp], the sconv runs on the shard, and an all-gather |
| 156 | # restores the full residual — all fused with the residual add + next |
| 157 | # rmsnorm via the Lamport P2P kernels for decode-sized batches. |
| 158 | tp_size = get_tensor_model_parallel_world_size() |
| 159 | sconv_dim = config.hidden_size // tp_size |
| 160 | self.attn_sconv = InklingShortConv( |
| 161 | sconv_dim, config.sconv_kernel_size, owner=self.conv_state, stream_idx=_ATTN |
| 162 | ) |
| 163 | self.mlp_sconv = InklingShortConv( |
| 164 | sconv_dim, config.sconv_kernel_size, owner=self.conv_state, stream_idx=_MLP |
| 165 | ) |
| 166 | |
| 167 | def forward( |
| 168 | self, |
| 169 | positions: torch.Tensor, |
| 170 | hidden_states: torch.Tensor, |
| 171 | pending: tuple[torch.Tensor | None, InklingShortConv] | None = None, |
| 172 | defer_mlp_add: bool = False, |
| 173 | attn_in: torch.Tensor | None = None, |
| 174 | log_scaling: torch.Tensor | None = None, |
| 175 | ) -> ( |
| 176 | torch.Tensor | tuple[torch.Tensor, tuple[torch.Tensor | None, InklingShortConv]] |
| 177 | ): |
| 178 | # The previous sublayer's (pre-reduce, pre-sconv) delta is folded in |
| 179 | # fused with its RS/sconv/AG and this layer's pre-attention rmsnorm. |
| 180 | # A None delta means the partials sit in the NVLS symm buffer. |
| 181 | if pending is None: |
| 182 | if attn_in is None: |
| 183 | # First layer; on the text path attn_norm comes fused with |
| 184 | # the embedding gather (chain_weight in embed_rmsnorm). |
| 185 | attn_in = self.attn_norm(hidden_states) |
| 186 | else: |
| 187 | attn_in, hidden_states = _sconv_add_norm( |
| 188 | pending[0], hidden_states, pending[1], self.attn_norm, positions |
| 189 | ) |
| 190 | attn_output = self.attn(positions, attn_in, log_scaling) |
| 191 | mlp_in, hidden_states = _sconv_add_norm( |
| 192 | attn_output, hidden_states, self.attn_sconv, self.mlp_norm, positions |
| 193 | ) |
| 194 | mlp_output = self.mlp(mlp_in) |
| 195 | if defer_mlp_add: |
| 196 | # Caller folds mlp_output (pre-reduce, pre-sconv) into the next |
| 197 | # fused sconv+add+rmsnorm. |
| 198 | return hidden_states, (mlp_output, self.mlp_sconv) |
| 199 | # Standalone (MTP) tail: finish the sublayer without a norm. |
| 200 | return _sconv_add_norm( |
| 201 | mlp_output, hidden_states, self.mlp_sconv, None, positions |
| 202 | )[1] |
| 203 | |
| 204 | |
| 205 | class InklingReplicatedEmbedding(nn.Module): |
| 206 | """Full-vocab embedding table replicated on every TP rank. |
| 207 | |
| 208 | Trades the full table per rank (~2.3 GiB at V=201k / H=6144 bf16, vs a |
| 209 | 1/tp shard) for no masked lookup and no per-lookup TP all-reduce — one |
| 210 | all-reduce per MTP draft step plus one per verify pass — and keeps the |
| 211 | full table on-rank for the fused gather+norm kernels (``embed_rmsnorm``, |
| 212 | ``embed_dual_rmsnorm_cat``). Bit-exact vs vocab-parallel: the all-reduce |
| 213 | there only ever summed one real row against exact zeros. The LM head |
| 214 | stays vocab-sharded. |
| 215 | """ |
| 216 | |
| 217 | def __init__(self, num_embeddings: int, embedding_dim: int) -> None: |
| 218 | super().__init__() |
| 219 | self.weight = nn.Parameter( |
| 220 | torch.empty(num_embeddings, embedding_dim, dtype=torch.get_default_dtype()), |
| 221 | requires_grad=False, |
| 222 | ) |
| 223 | |
| 224 | def forward(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 225 | return embed_rmsnorm(input_ids, self.weight, None, 0.0) |
| 226 | |
| 227 | |
| 228 | class InklingModel(nn.Module): |
| 229 | def __init__( |
| 230 | self, |
| 231 | *, |
| 232 | config: InklingModelConfig, |
| 233 | quant_config: QuantizationConfig | None, |
| 234 | prefix: str, |
| 235 | ) -> None: |
| 236 | super().__init__() |
| 237 | self.config = config |
| 238 | self.embed_tokens = InklingReplicatedEmbedding( |
| 239 | config.padded_vocab_size, config.hidden_size |
| 240 | ) |
| 241 | self.embed_norm = ( |
| 242 | InklingRMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 243 | if config.use_embed_norm |
| 244 | else None |
| 245 | ) |
| 246 | local_ids = set(config.local_layer_ids) |
| 247 | |
| 248 | def get_layer(prefix: str) -> InklingDecoderLayer: |
| 249 | idx = _layer_id(prefix + ".") or int(prefix.split(".")[-1]) |
| 250 | return InklingDecoderLayer( |
| 251 | config, idx, idx in local_ids, quant_config, prefix |
| 252 | ) |
| 253 | |
| 254 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 255 | config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers" |
| 256 | ) |
| 257 | self.norm = InklingRMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 258 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 259 | ["hidden_states"], config.hidden_size |
| 260 | ) |
| 261 | |
| 262 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 263 | # Row gather + embed_norm in one launch. |
| 264 | norm = self.embed_norm |
| 265 | return embed_rmsnorm( |
| 266 | input_ids, |
| 267 | self.embed_tokens.weight, |
| 268 | norm.weight if norm is not None else None, |
| 269 | norm.variance_epsilon if norm is not None else 0.0, |
| 270 | ) |
| 271 | |
| 272 | def forward( |
| 273 | self, |
| 274 | input_ids: torch.Tensor | None, |
| 275 | positions: torch.Tensor, |
| 276 | intermediate_tensors: IntermediateTensors | None, |
| 277 | inputs_embeds: torch.Tensor | None = None, |
| 278 | ) -> torch.Tensor | IntermediateTensors: |
| 279 | attn_in0: torch.Tensor | None = None |
| 280 | if get_pp_group().is_first_rank: |
| 281 | if inputs_embeds is not None: |
| 282 | # embed_norm was already applied when producing inputs_embeds. |
| 283 | hidden_states = inputs_embeds |
| 284 | else: |
| 285 | # Gather + embed_norm + the first layer's attn_norm, one launch. |
| 286 | norm = self.embed_norm |
| 287 | hidden_states, attn_in0 = embed_rmsnorm( |
| 288 | input_ids, |
| 289 | self.embed_tokens.weight, |
| 290 | norm.weight if norm is not None else None, |
| 291 | self.config.rms_norm_eps, |
| 292 | chain_weight=self.layers[self.start_layer].attn_norm.weight, |
| 293 | ) |
| 294 | else: |
| 295 | assert intermediate_tensors is not None |
| 296 | hidden_states = intermediate_tensors["hidden_states"] |
| 297 | hidden_states = hidden_states.view(-1, hidden_states.shape[-1]) |
| 298 | log_scaling = None |
| 299 | if self.config.log_scaling_n_floor is not None: |
| 300 | log_scaling = compute_log_scaling_tau( |
| 301 | positions, |
| 302 | self.config.log_scaling_n_floor, |
| 303 | self.config.log_scaling_alpha, |
| 304 | ) |
| 305 | |
| 306 | pending: tuple[torch.Tensor | None, InklingShortConv] | None = None |
| 307 | for layer in self.layers[self.start_layer : self.end_layer]: |
| 308 | hidden_states, pending = layer( |
| 309 | positions, |
| 310 | hidden_states, |
| 311 | pending=pending, |
| 312 | defer_mlp_add=True, |
| 313 | attn_in=attn_in0, |
| 314 | log_scaling=log_scaling, |
| 315 | ) |
| 316 | attn_in0 = None |
| 317 | |
| 318 | if not get_pp_group().is_last_rank: |
| 319 | if pending is not None: |
| 320 | hidden_states = _sconv_add_norm( |
| 321 | pending[0], hidden_states, pending[1], None, positions |
| 322 | )[1] |
| 323 | return IntermediateTensors({"hidden_states": hidden_states}) |
| 324 | if pending is not None: |
| 325 | # Final RS/sconv/AG + residual add fused with the final rmsnorm. |
| 326 | norm_out = _sconv_add_norm( |
| 327 | pending[0], hidden_states, pending[1], self.norm, positions |
| 328 | )[0] |
| 329 | assert norm_out is not None |
| 330 | return norm_out |
| 331 | return self.norm(hidden_states) |
| 332 | |
| 333 | |
| 334 | class _TmlForCausalLMBase(nn.Module, SupportsPP, SupportsLoRA): |
| 335 | """Shared text-backbone causal-LM scaffolding for both entry classes.""" |
| 336 | |
| 337 | hf_to_vllm_mapper = WeightsMapper( |
| 338 | orig_to_new_substr={ |
| 339 | ".w13_dn": ".gate_up_proj", |
| 340 | ".w2_md": ".down_proj", |
| 341 | }, |
| 342 | orig_to_new_stacked={ |
| 343 | ".attn.wq_du.": (".attn.qkvr.", 0), |
| 344 | ".attn.wk_dv.": (".attn.qkvr.", 1), |
| 345 | ".attn.wv_dv.": (".attn.qkvr.", 2), |
| 346 | ".attn.wr_du.": (".attn.qkvr.", 3), |
| 347 | }, |
| 348 | orig_to_new_prefix={ |
| 349 | "model.llm.layers.": "model.layers.", |
| 350 | "model.llm.embed_norm": "model.embed_norm", |
| 351 | "model.llm.embed": "model.embed_tokens", |
| 352 | "model.llm.norm": "model.norm", |
| 353 | "model.llm.unembed": "lm_head", |
| 354 | "language_model.layers.": "model.layers.", |
| 355 | "language_model.lm_head.": "lm_head.", |
| 356 | }, |
| 357 | orig_to_new_suffix={ |
| 358 | # NVFP4 scale |
| 359 | ".w13_weight.scale": ".w13_weight_scale", |
| 360 | ".w13_weight.scale2": ".w13_weight_scale_2", |
| 361 | ".w2_weight.scale": ".w2_weight_scale", |
| 362 | ".w2_weight.scale2": ".w2_weight_scale_2", |
| 363 | }, |
| 364 | ) |
| 365 | # Quark uses this mapping when resolving quantization exclusions for the |
| 366 | # four checkpoint attention projections fused into qkvr. |
| 367 | packed_modules_mapping = { |
| 368 | "qkvr": ["wq_du", "wk_dv", "wv_dv", "wr_du"], |
| 369 | "w13": ["w1", "w3"], |
| 370 | } |
| 371 | embedding_modules = { |
| 372 | "lm_head": "output_embeddings", |
| 373 | } |
| 374 | |
| 375 | def _build( |
| 376 | self, |
| 377 | vllm_config: VllmConfig, |
| 378 | text_config: InklingModelConfig, |
| 379 | prefix: str, |
| 380 | ) -> None: |
| 381 | quant_config = vllm_config.quant_config |
| 382 | self.config = text_config |
| 383 | # ROCm checkpoints use Quark OCP MXFP4. The global Quark config is |
| 384 | # passed into each routed MoE and its exclusion list keeps the rest of |
| 385 | # Inkling in bf16. |
| 386 | # Read by the MRV2 runner to publish per-request short-conv metadata. |
| 387 | # Short convolution is intrinsic to Inkling, so this is always set. |
| 388 | self.uses_sconv = True |
| 389 | self.model = InklingModel( |
| 390 | config=text_config, |
| 391 | quant_config=quant_config, |
| 392 | prefix=maybe_prefix(prefix, "model"), |
| 393 | ) |
| 394 | self.lm_head = ParallelLMHead( |
| 395 | text_config.padded_vocab_size, |
| 396 | text_config.hidden_size, |
| 397 | org_num_embeddings=text_config.padded_vocab_size, |
| 398 | quant_config=quant_config, |
| 399 | prefix=maybe_prefix(prefix, "lm_head"), |
| 400 | ) |
| 401 | self.logits_processor = InklingLogitsProcessor( |
| 402 | text_config.padded_vocab_size, |
| 403 | org_vocab_size=text_config.vocab_size, |
| 404 | soft_cap=text_config.final_logit_softcapping, |
| 405 | logits_mup_width_multiplier=text_config.logits_mup_width_multiplier, |
| 406 | ) |
| 407 | self.make_empty_intermediate_tensors = ( # type: ignore[method-assign] |
| 408 | self.model.make_empty_intermediate_tensors |
| 409 | ) |
| 410 | |
| 411 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 412 | return self.model.embed_input_ids(input_ids) |
| 413 | |
| 414 | def forward( |
| 415 | self, |
| 416 | input_ids: torch.Tensor | None, |
| 417 | positions: torch.Tensor, |
| 418 | intermediate_tensors: IntermediateTensors | None = None, |
| 419 | inputs_embeds: torch.Tensor | None = None, |
| 420 | **kwargs: object, |
| 421 | ) -> torch.Tensor | IntermediateTensors: |
| 422 | return self.model( |
| 423 | input_ids, |
| 424 | positions, |
| 425 | intermediate_tensors, |
| 426 | inputs_embeds, |
| 427 | ) |
| 428 | |
| 429 | def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor | None: |
| 430 | return self.logits_processor(self.lm_head, hidden_states) |
| 431 | |
| 432 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 433 | return _load_inkling_weights(self, weights, self.config) |
| 434 | |
| 435 | |
| 436 | class InklingForCausalLM(_TmlForCausalLMBase): |
| 437 | """Text-only entry point (``inkling_model`` checkpoints).""" |
| 438 | |
| 439 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 440 | super().__init__() |
| 441 | self._build(vllm_config, vllm_config.model_config.hf_config, prefix) |
| 442 | |
| 443 | |
| 444 | @MULTIMODAL_REGISTRY.register_processor( |
| 445 | InklingMultiModalProcessor, |
| 446 | info=InklingProcessingInfo, |
| 447 | dummy_inputs=InklingDummyInputsBuilder, |
| 448 | ) |
| 449 | class InklingForConditionalGeneration(_TmlForCausalLMBase, SupportsMultiModal): |
| 450 | """Top-level (multimodal) entry point. |
| 451 | |
| 452 | Builds the vision + audio towers on top of the shared text backbone. Inkling has |
| 453 | NO cross-modal fusion (the vision tower emits one token per patch, the audio |
| 454 | tower one token per frame), so generation reuses the inherited backbone |
| 455 | ``forward`` / ``compute_logits`` (the latter already applies muP) and this |
| 456 | class only adds multimodal embedding + merge. |
| 457 | """ |
| 458 | |
| 459 | hf_to_vllm_mapper = _TmlForCausalLMBase.hf_to_vllm_mapper | WeightsMapper( |
| 460 | orig_to_new_prefix={ |
| 461 | "model.audio.": "audio.", |
| 462 | "model.visual.": "visual.vision_encoder.", |
| 463 | }, |
| 464 | ) |
| 465 | |
| 466 | @classmethod |
| 467 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 468 | if modality.startswith("image"): |
| 469 | return "<|content_image|>" |
| 470 | if modality.startswith("audio"): |
| 471 | return "<|content_audio_input|>" |
| 472 | raise ValueError("Only image or audio modality is supported") |
| 473 | |
| 474 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 475 | super().__init__() |
| 476 | config: InklingMMConfig = vllm_config.model_config.hf_config |
| 477 | |
| 478 | self.visual = ( |
| 479 | InklingVision(config.vision_config, prefix=maybe_prefix(prefix, "visual")) |
| 480 | if inkling_vision_enabled(config) |
| 481 | else None |
| 482 | ) |
| 483 | self.audio = ( |
| 484 | InklingAudio(config.audio_config, prefix=maybe_prefix(prefix, "audio")) |
| 485 | if inkling_audio_enabled(config) |
| 486 | else None |
| 487 | ) |
| 488 | |
| 489 | self._build(vllm_config, config.text_config, prefix) |
| 490 | |
| 491 | # -- multimodal embedding ------------------------------------------- |
| 492 | |
| 493 | def _process_image_input( |
| 494 | self, pixel_values: Any, num_patches: Any |
| 495 | ) -> tuple[torch.Tensor, ...]: |
| 496 | assert self.visual is not None |
| 497 | # pixel_values is a list (per item) of [P_i, 2, P, P, 3] tensors, |
| 498 | # or a single concatenated tensor. Normalize to a flat batch, run the |
| 499 | # tower once, then split back per item. |
| 500 | if isinstance(pixel_values, (list, tuple)): |
| 501 | if not pixel_values: |
| 502 | return () |
| 503 | sizes = [int(p.shape[0]) for p in pixel_values] |
| 504 | patches = torch.cat(list(pixel_values), dim=0) |
| 505 | else: |
| 506 | patches = pixel_values |
| 507 | sizes = self._sizes_from(num_patches, patches.shape[0]) |
| 508 | |
| 509 | patches = patches.to(device=self.visual.device, dtype=self.visual.dtype) |
| 510 | embeds = self.visual(patches) # [total_patches, D] |
| 511 | return tuple(embeds.split(sizes)) |
| 512 | |
| 513 | def _process_audio_input( |
| 514 | self, input_audio_features: Any, num_audio_tokens: Any |
| 515 | ) -> tuple[torch.Tensor, ...]: |
| 516 | assert self.audio is not None |
| 517 | if isinstance(input_audio_features, (list, tuple)): |
| 518 | if not input_audio_features: |
| 519 | return () |
| 520 | sizes = [int(d.shape[0]) for d in input_audio_features] |
| 521 | dmel = torch.cat(list(input_audio_features), dim=0) |
| 522 | else: |
| 523 | dmel = input_audio_features |
| 524 | sizes = self._sizes_from(num_audio_tokens, dmel.shape[0]) |
| 525 | |
| 526 | dmel = dmel.to(device=self.audio.device) |
| 527 | embeds = self.audio(dmel) # [total_frames, D] |
| 528 | return tuple(embeds.split(sizes)) |
| 529 | |
| 530 | @staticmethod |
| 531 | def _sizes_from(counts: Any, total: int) -> list[int]: |
| 532 | if counts is None: |
| 533 | return [total] |
| 534 | if isinstance(counts, torch.Tensor): |
| 535 | return [int(c) for c in counts.flatten().tolist()] |
| 536 | if isinstance(counts, (list, tuple)): |
| 537 | flat: list[int] = [] |
| 538 | for c in counts: |
| 539 | flat.append(int(c.item()) if isinstance(c, torch.Tensor) else int(c)) |
| 540 | return flat |
| 541 | return [int(counts)] |
| 542 | |
| 543 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 544 | # Iterate modalities in a stable order so the returned per-item tensors |
| 545 | # line up with their appearance order; the positional merge in |
| 546 | # embed_input_ids handles actual placement. |
| 547 | pixel_values = kwargs.get("pixel_values") |
| 548 | num_patches = kwargs.get("num_patches") |
| 549 | input_audio_features = kwargs.get("input_audio_features") |
| 550 | num_audio_tokens = kwargs.get("num_audio_tokens") |
| 551 | |
| 552 | embeddings: tuple[torch.Tensor, ...] = () |
| 553 | if pixel_values is not None and self.visual is not None: |
| 554 | embeddings += self._process_image_input(pixel_values, num_patches) |
| 555 | if input_audio_features is not None and self.audio is not None: |
| 556 | embeddings += self._process_audio_input( |
| 557 | input_audio_features, num_audio_tokens |
| 558 | ) |
| 559 | return embeddings |
| 560 | |
| 561 | def embed_input_ids( |
| 562 | self, |
| 563 | input_ids: torch.Tensor, |
| 564 | multimodal_embeddings: MultiModalEmbeddings | None = None, |
| 565 | *, |
| 566 | is_multimodal: torch.Tensor | None = None, |
| 567 | ) -> torch.Tensor: |
| 568 | # Override the base's 1-arg embed_input_ids: the runner calls this 3-arg |
| 569 | # signature for multimodal models. Text embeddings come from the shared |
| 570 | # backbone (which applies embed_norm); MM embeddings are scattered in. |
| 571 | from vllm.model_executor.models.utils import _merge_multimodal_embeddings |
| 572 | |
| 573 | # Placeholder ids use unused vocabulary slots and these positions are |
| 574 | # overwritten by MM embeds below. |
| 575 | inputs_embeds = self.model.embed_input_ids(input_ids) |
| 576 | if multimodal_embeddings is None or len(multimodal_embeddings) == 0: |
| 577 | return inputs_embeds |
| 578 | assert is_multimodal is not None |
| 579 | return _merge_multimodal_embeddings( |
| 580 | inputs_embeds=inputs_embeds, |
| 581 | multimodal_embeddings=multimodal_embeddings, |
| 582 | is_multimodal=is_multimodal, |
| 583 | ) |
| 584 | |
| 585 | def get_language_model(self) -> nn.Module: |
| 586 | # This class IS the causal LM (the towers are side branches), so the |
| 587 | # language model is self — callers expect a module exposing ``.model`` |
| 588 | # / ``.lm_head`` (e.g. the MTP/eagle loader shares embeddings via |
| 589 | # ``get_language_model().model.embed_tokens``). |
| 590 | return self |
| 591 | |
| 592 | |
| 593 | # =========================================================================== |
| 594 | # Weight loading |
| 595 | # =========================================================================== |
| 596 | |
| 597 | |
| 598 | _MOE_EXPERT_WEIGHT_RE = re.compile( |
| 599 | r"^(?P<mlp>.*\.mlp)\.(?P<rest>(?:shared_)?experts\..+)$" |
| 600 | ) |
| 601 | |
| 602 | |
| 603 | def _is_peft_adapter_weight(name: str) -> bool: |
| 604 | return ".lora_A." in name or ".lora_B." in name |
| 605 | |
| 606 | |
| 607 | def _load_inkling_weights( |
| 608 | module: nn.Module, |
| 609 | weights: Iterable[tuple[str, torch.Tensor]], |
| 610 | config: InklingModelConfig, |
| 611 | ) -> set[str]: |
| 612 | moe_modules = { |
| 613 | name: mod for name, mod in module.named_modules() if isinstance(mod, InklingMoE) |
| 614 | } |
| 615 | loaded: set[str] = set() |
| 616 | tp_size = get_tensor_model_parallel_world_size() |
| 617 | tp_rank = get_tensor_model_parallel_rank() |
| 618 | local_ids = set(config.local_layer_ids) |
| 619 | |
| 620 | def _iter_loadable_weights() -> Iterable[tuple[str, torch.Tensor]]: |
| 621 | for name, weight in module.hf_to_vllm_mapper.apply(weights): |
| 622 | # LightSeek's MXFP4 conversion bundles a quantized copy of the |
| 623 | # standalone PEFT adapter in the base safetensors index. These are |
| 624 | # not base-model parameters; adapters remain opt-in at serving. |
| 625 | if _is_peft_adapter_weight(name): |
| 626 | continue |
| 627 | shard_id = getattr(weight, "shard_id", None) |
| 628 | # Replicate K/V conv-free GQA heads when tp_size > num_kv_heads. |
| 629 | if ( |
| 630 | shard_id in (1, 2) |
| 631 | and name.endswith(".attn.qkvr.weight") |
| 632 | and weight.shape[0] > 0 |
| 633 | ): |
| 634 | lid = _layer_id(name) |
| 635 | if lid is not None: |
| 636 | is_local = lid in local_ids |
| 637 | n_kv = ( |
| 638 | config.swa_num_key_value_heads |
| 639 | if is_local |
| 640 | else config.num_key_value_heads |
| 641 | ) |
| 642 | head_dim = config.swa_head_dim if is_local else config.head_dim |
| 643 | if tp_size > n_kv and weight.shape[0] == n_kv * head_dim: |
| 644 | kv_idx = (tp_rank * n_kv) // tp_size |
| 645 | weight = weight.narrow(0, kv_idx * head_dim, head_dim) |
| 646 | weight.shard_id = shard_id |
| 647 | |
| 648 | # MoE expert tensors (fused stacked, routed + shared sink): translate |
| 649 | # the checkpoint layout to per-expert MoERunner loads. |
| 650 | moe_match = _MOE_EXPERT_WEIGHT_RE.match(name) |
| 651 | if moe_match is not None and moe_match.group("mlp") in moe_modules: |
| 652 | moe = moe_modules[moe_match.group("mlp")] |
| 653 | for rel in moe.load_expert_weight(moe_match.group("rest"), weight): |
| 654 | loaded.add(f"{moe_match.group('mlp')}.{rel}") |
| 655 | continue |
| 656 | |
| 657 | yield name, weight |
| 658 | |
| 659 | mapper = WeightsMapper(orig_to_new_prefix={"model.mtp.": None}) |
| 660 | loader = AutoWeightsLoader(module) |
| 661 | loaded |= loader.load_weights(_iter_loadable_weights(), mapper=mapper) |
| 662 | |
| 663 | # Post-load MoE fixups (default input scales, zeroed EP-padding experts). |
| 664 | for moe_name, moe in moe_modules.items(): |
| 665 | for rel in moe.finalize_load(): |
| 666 | loaded.add(f"{moe_name}.{rel}") |
| 667 | return loaded |
| 668 | |
| 669 | |
| 670 | EntryClass = [InklingForCausalLM, InklingForConditionalGeneration] |