| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Inference-only Kimi-K3 Multi-Token-Prediction (MTP) draft model.""" |
| 4 | |
| 5 | import copy |
| 6 | from collections.abc import Iterable |
| 7 | |
| 8 | import torch |
| 9 | import torch.nn as nn |
| 10 | |
| 11 | import vllm.envs as envs |
| 12 | from vllm.config import VllmConfig |
| 13 | from vllm.forward_context import get_forward_context, is_forward_context_available |
| 14 | from vllm.logger import init_logger |
| 15 | from vllm.model_executor.layers.fused_moe import ( |
| 16 | fused_moe_make_expert_params_mapping, |
| 17 | ) |
| 18 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 19 | from vllm.model_executor.layers.linear import ReplicatedLinear |
| 20 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 21 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 22 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 23 | ParallelLMHead, |
| 24 | VocabParallelEmbedding, |
| 25 | ) |
| 26 | from vllm.model_executor.model_loader.weight_utils import ( |
| 27 | default_weight_loader, |
| 28 | maybe_remap_kv_scale_name, |
| 29 | ) |
| 30 | from vllm.model_executor.models.utils import get_pp_missing_layer_names, maybe_prefix |
| 31 | from vllm.models.common.ops.sequence_parallel import ( |
| 32 | sp_all_gather, |
| 33 | sp_padding_mask, |
| 34 | sp_shard, |
| 35 | ) |
| 36 | from vllm.sequence import IntermediateTensors |
| 37 | from vllm.transformers_utils.configs.kimi_linear import KimiLinearConfig |
| 38 | |
| 39 | from ..common.mtp import fused_mtp_input |
| 40 | from .low_latency_gemm import enable_kimi_k3_low_latency_gemm |
| 41 | from .model import ( |
| 42 | KimiDecoderLayer, |
| 43 | KimiMoE, |
| 44 | get_spec_layer_idx_from_weight_name, |
| 45 | make_kimi_k3_mega_moe_expert_params_mapping, |
| 46 | ) |
| 47 | |
| 48 | logger = init_logger(__name__) |
| 49 | |
| 50 | |
| 51 | class SharedHead(nn.Module): |
| 52 | def __init__( |
| 53 | self, |
| 54 | config: KimiLinearConfig, |
| 55 | prefix: str, |
| 56 | quant_config: QuantizationConfig | None = None, |
| 57 | ) -> None: |
| 58 | super().__init__() |
| 59 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 60 | self.head = ParallelLMHead( |
| 61 | config.vocab_size, |
| 62 | config.hidden_size, |
| 63 | quant_config=quant_config, |
| 64 | prefix=maybe_prefix(prefix, "head"), |
| 65 | ) |
| 66 | |
| 67 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 68 | return self.norm(hidden_states) |
| 69 | |
| 70 | |
| 71 | class KimiK3MultiTokenPredictorLayer(nn.Module): |
| 72 | def __init__( |
| 73 | self, |
| 74 | config: KimiLinearConfig, |
| 75 | vllm_config: VllmConfig, |
| 76 | prefix: str, |
| 77 | ) -> None: |
| 78 | super().__init__() |
| 79 | self.config = config |
| 80 | quant_config = vllm_config.quant_config |
| 81 | |
| 82 | self.enorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 83 | self.hnorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 84 | self.eh_proj = ReplicatedLinear( |
| 85 | config.hidden_size * 2, |
| 86 | config.hidden_size, |
| 87 | bias=False, |
| 88 | quant_config=None, |
| 89 | prefix=maybe_prefix(prefix, "eh_proj"), |
| 90 | return_bias=False, |
| 91 | ) |
| 92 | |
| 93 | self.shared_head = SharedHead( |
| 94 | config=config, prefix=prefix, quant_config=quant_config |
| 95 | ) |
| 96 | # The MTP block starts without the base model's AttnRes state. |
| 97 | block_config = copy.copy(config) |
| 98 | block_config.attn_res_block_size = None |
| 99 | # NOTE: the prefix must end in the numeric spec-layer index so that |
| 100 | # KimiDecoderLayer can parse ``layer_idx`` and pick MLA (full attn). |
| 101 | # Own aux stream for the MLA g_proj output-gate overlap (DeepseekV4 |
| 102 | # convention: the MTP block creates its own stream). |
| 103 | aux_stream = torch.cuda.Stream() |
| 104 | self.mtp_block = KimiDecoderLayer( |
| 105 | block_config, vllm_config, prefix=prefix, aux_stream=aux_stream |
| 106 | ) |
| 107 | |
| 108 | def forward( |
| 109 | self, |
| 110 | input_ids: torch.Tensor, |
| 111 | positions: torch.Tensor, |
| 112 | previous_hidden_states: torch.Tensor, |
| 113 | inputs_embeds: torch.Tensor | None = None, |
| 114 | spec_step_index: int = 0, |
| 115 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 116 | assert inputs_embeds is not None |
| 117 | hidden_states = self.eh_proj( |
| 118 | fused_mtp_input( |
| 119 | positions, |
| 120 | inputs_embeds, |
| 121 | previous_hidden_states, |
| 122 | self.enorm.weight, |
| 123 | self.hnorm.weight, |
| 124 | self.enorm.variance_epsilon, |
| 125 | ) |
| 126 | ) |
| 127 | if self.mtp_block.use_sequence_parallel: |
| 128 | if envs.VLLM_MOE_SKIP_PADDING and is_forward_context_available(): |
| 129 | forward_context = get_forward_context() |
| 130 | forward_context.is_padding = sp_padding_mask( |
| 131 | forward_context.is_padding, hidden_states |
| 132 | ) |
| 133 | hidden_states = sp_shard(hidden_states) |
| 134 | |
| 135 | hidden_states, _, residual = self.mtp_block( |
| 136 | positions=positions, |
| 137 | hidden_states=hidden_states, |
| 138 | residual=None, |
| 139 | ) |
| 140 | if self.mtp_block.use_sequence_parallel: |
| 141 | assert residual is not None |
| 142 | hidden_states = hidden_states + residual |
| 143 | hidden_states = sp_all_gather(hidden_states)[: positions.shape[0]] |
| 144 | logits_hidden_states = self.shared_head.norm(hidden_states) |
| 145 | return logits_hidden_states, hidden_states |
| 146 | |
| 147 | # Produce the normalized logits input and the pre-norm recurrent state |
| 148 | # in one fused add-RMSNorm launch. |
| 149 | logits_hidden_states, hidden_states = self.shared_head.norm( |
| 150 | hidden_states, residual |
| 151 | ) |
| 152 | return logits_hidden_states, hidden_states |
| 153 | |
| 154 | |
| 155 | class KimiK3MultiTokenPredictor(nn.Module): |
| 156 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 157 | super().__init__() |
| 158 | config: KimiLinearConfig = vllm_config.model_config.hf_text_config |
| 159 | self.config = config |
| 160 | self.mtp_start_layer_idx = config.num_hidden_layers |
| 161 | self.num_mtp_layers = config.num_nextn_predict_layers |
| 162 | |
| 163 | self.layers = torch.nn.ModuleDict( |
| 164 | { |
| 165 | str(idx): KimiK3MultiTokenPredictorLayer( |
| 166 | config, vllm_config, f"{prefix}.layers.{idx}" |
| 167 | ) |
| 168 | for idx in range( |
| 169 | self.mtp_start_layer_idx, |
| 170 | self.mtp_start_layer_idx + self.num_mtp_layers, |
| 171 | ) |
| 172 | } |
| 173 | ) |
| 174 | self.embed_tokens = VocabParallelEmbedding( |
| 175 | config.vocab_size, |
| 176 | config.hidden_size, |
| 177 | prefix=maybe_prefix(prefix, "embed_tokens"), |
| 178 | ) |
| 179 | self.logits_processor = LogitsProcessor(config.vocab_size) |
| 180 | |
| 181 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 182 | return self.embed_tokens(input_ids) |
| 183 | |
| 184 | def forward( |
| 185 | self, |
| 186 | input_ids: torch.Tensor, |
| 187 | positions: torch.Tensor, |
| 188 | previous_hidden_states: torch.Tensor, |
| 189 | inputs_embeds: torch.Tensor | None = None, |
| 190 | spec_step_idx: int = 0, |
| 191 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 192 | if inputs_embeds is None: |
| 193 | inputs_embeds = self.embed_tokens(input_ids) |
| 194 | current_step_idx = spec_step_idx % self.num_mtp_layers |
| 195 | return self.layers[str(self.mtp_start_layer_idx + current_step_idx)]( |
| 196 | input_ids, |
| 197 | positions, |
| 198 | previous_hidden_states, |
| 199 | inputs_embeds, |
| 200 | current_step_idx, |
| 201 | ) |
| 202 | |
| 203 | def compute_logits( |
| 204 | self, |
| 205 | hidden_states: torch.Tensor, |
| 206 | spec_step_idx: int = 0, |
| 207 | ) -> torch.Tensor: |
| 208 | current_step_idx = spec_step_idx % self.num_mtp_layers |
| 209 | mtp_layer = self.layers[str(self.mtp_start_layer_idx + current_step_idx)] |
| 210 | logits = self.logits_processor(mtp_layer.shared_head.head, hidden_states) |
| 211 | return logits |
| 212 | |
| 213 | |
| 214 | class KimiK3MTP(nn.Module): |
| 215 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 216 | super().__init__() |
| 217 | self.config = vllm_config.model_config.hf_text_config |
| 218 | self.quant_config = vllm_config.quant_config |
| 219 | self.model = KimiK3MultiTokenPredictor( |
| 220 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 221 | ) |
| 222 | enable_kimi_k3_low_latency_gemm(self, vllm_config.model_config.dtype) |
| 223 | |
| 224 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 225 | return self.model.embed_input_ids(input_ids) |
| 226 | |
| 227 | def forward( |
| 228 | self, |
| 229 | input_ids: torch.Tensor | None, |
| 230 | positions: torch.Tensor, |
| 231 | hidden_states: torch.Tensor, |
| 232 | intermediate_tensors: IntermediateTensors | None = None, |
| 233 | inputs_embeds: torch.Tensor | None = None, |
| 234 | spec_step_idx: int = 0, |
| 235 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 236 | return self.model( |
| 237 | input_ids, |
| 238 | positions, |
| 239 | hidden_states, |
| 240 | inputs_embeds, |
| 241 | spec_step_idx, |
| 242 | ) |
| 243 | |
| 244 | def compute_logits( |
| 245 | self, |
| 246 | hidden_states: torch.Tensor, |
| 247 | spec_step_idx: int = 0, |
| 248 | ) -> torch.Tensor | None: |
| 249 | return self.model.compute_logits(hidden_states, spec_step_idx) |
| 250 | |
| 251 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 252 | # Mirror KimiLinearForCausalLM.load_weights naming: leading-dot shard |
| 253 | # names, q_lora-conditional fused QKV, and w1/w2/w3 expert weights. |
| 254 | kda_config = self.config.linear_attn_config |
| 255 | use_full_rank_gate = bool( |
| 256 | kda_config and kda_config.get("use_full_rank_gate", False) |
| 257 | ) |
| 258 | beta_shard_id = 5 if use_full_rank_gate else 3 |
| 259 | stacked_params_mapping = [ |
| 260 | # (param_name, shard_name, shard_id) |
| 261 | (".in_proj_qkvgfab", ".q_proj", 0), |
| 262 | (".in_proj_qkvgfab", ".k_proj", 1), |
| 263 | (".in_proj_qkvgfab", ".v_proj", 2), |
| 264 | (".in_proj_qkvgfab", ".b_proj", beta_shard_id), |
| 265 | (".in_proj_qkvgfab", ".f_a_proj", 4), |
| 266 | (".conv1d", ".q_conv1d", 0), |
| 267 | (".conv1d", ".k_conv1d", 1), |
| 268 | (".conv1d", ".v_conv1d", 2), |
| 269 | (".gate_up_proj", ".gate_proj", 0), |
| 270 | (".gate_up_proj", ".up_proj", 1), |
| 271 | ] |
| 272 | if use_full_rank_gate: |
| 273 | stacked_params_mapping.append((".in_proj_qkvgfab", ".g_proj", 3)) |
| 274 | if getattr(self.config, "q_lora_rank", None) is not None: |
| 275 | if self.config.mla_use_output_gate: |
| 276 | stacked_params_mapping += [ |
| 277 | (".fused_qkv_a_g_proj", ".q_a_proj", 0), |
| 278 | (".fused_qkv_a_g_proj", ".kv_a_proj_with_mqa", 1), |
| 279 | (".fused_qkv_a_g_proj", ".g_proj", 2), |
| 280 | ] |
| 281 | else: |
| 282 | stacked_params_mapping += [ |
| 283 | (".fused_qkv_a_proj", ".q_a_proj", 0), |
| 284 | (".fused_qkv_a_proj", ".kv_a_proj_with_mqa", 1), |
| 285 | ] |
| 286 | |
| 287 | use_mega_moe = any( |
| 288 | module.use_mega_moe |
| 289 | for module in self.modules() |
| 290 | if isinstance(module, KimiMoE) |
| 291 | ) |
| 292 | if self.config.is_moe and use_mega_moe: |
| 293 | expert_params_mapping = make_kimi_k3_mega_moe_expert_params_mapping( |
| 294 | self.config.num_experts |
| 295 | ) |
| 296 | elif self.config.is_moe: |
| 297 | expert_params_mapping = fused_moe_make_expert_params_mapping( |
| 298 | self, |
| 299 | ckpt_gate_proj_name="w1", |
| 300 | ckpt_down_proj_name="w2", |
| 301 | ckpt_up_proj_name="w3", |
| 302 | num_experts=self.config.num_experts, |
| 303 | ) |
| 304 | else: |
| 305 | expert_params_mapping = [] |
| 306 | |
| 307 | pp_missing_layer_names = get_pp_missing_layer_names(self) |
| 308 | params_dict = dict(self.named_parameters()) |
| 309 | # Under the MXFP4 quant interface the routed experts register unpacked |
| 310 | # params (``w13_weight``), while the compressed-tensors checkpoint names |
| 311 | # them ``.weight_packed``. Rebind so the expert mapping resolves; scales |
| 312 | # already share the ``.weight_scale`` suffix. |
| 313 | experts_unpacked = not use_mega_moe and not any( |
| 314 | n.endswith("w13_weight_packed") for n in params_dict |
| 315 | ) |
| 316 | loaded_params: set[str] = set() |
| 317 | for name, loaded_weight in weights: |
| 318 | if "rotary_emb.inv_freq" in name: |
| 319 | continue |
| 320 | # The multimodal checkpoint prefixes text weights with |
| 321 | # ``language_model.``; strip it so names match this draft model's |
| 322 | # parameter paths (``model.layers.{i}.``). Non-text weights |
| 323 | # (vision_tower, mm_projector, ...) never match a spec layer below. |
| 324 | if name.startswith("language_model."): |
| 325 | name = name[len("language_model.") :] |
| 326 | if experts_unpacked and name.endswith(".weight_packed"): |
| 327 | name = name.replace(".weight_packed", ".weight") |
| 328 | spec_layer = get_spec_layer_idx_from_weight_name(self.config, name) |
| 329 | if spec_layer is None: |
| 330 | continue |
| 331 | name = self._rewrite_spec_layer_name(spec_layer, name) |
| 332 | |
| 333 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 334 | if weight_name not in name: |
| 335 | continue |
| 336 | # Routed experts (``.experts.{i}.w1/w2/w3``) are handled by the |
| 337 | # expert mapping below; skip them here. Shared experts |
| 338 | # (``.shared_experts.``) use gate/up_proj and fall through. |
| 339 | if ".experts." in name: |
| 340 | continue |
| 341 | name_mapped = name.replace(weight_name, param_name) |
| 342 | # Only take this mapping if the fused destination actually |
| 343 | # exists (e.g. QKV fusion is only present when q_lora is used). |
| 344 | if name_mapped not in params_dict: |
| 345 | continue |
| 346 | if name_mapped in pp_missing_layer_names: |
| 347 | continue |
| 348 | name = name_mapped |
| 349 | param = params_dict[name] |
| 350 | weight_loader = param.weight_loader |
| 351 | weight_loader(param, loaded_weight, shard_id) |
| 352 | break |
| 353 | else: |
| 354 | for ( |
| 355 | expert_param_name, |
| 356 | expert_weight_name, |
| 357 | expert_id, |
| 358 | expert_shard_id, |
| 359 | ) in expert_params_mapping: |
| 360 | if expert_weight_name not in name: |
| 361 | continue |
| 362 | name_mapped = name.replace(expert_weight_name, expert_param_name) |
| 363 | if name_mapped in pp_missing_layer_names: |
| 364 | continue |
| 365 | param = params_dict[name_mapped] |
| 366 | weight_loader = param.weight_loader |
| 367 | weight_loader( |
| 368 | param, |
| 369 | loaded_weight, |
| 370 | name_mapped, |
| 371 | shard_id=expert_shard_id, |
| 372 | expert_id=expert_id, |
| 373 | ) |
| 374 | name = name_mapped |
| 375 | break |
| 376 | else: |
| 377 | if name.endswith(".bias") and name not in params_dict: |
| 378 | continue |
| 379 | remapped_name = maybe_remap_kv_scale_name(name, params_dict) |
| 380 | if remapped_name is None: |
| 381 | continue |
| 382 | name = remapped_name |
| 383 | |
| 384 | # The embedding is shared across MTP layers; only the first |
| 385 | # spec layer carries the hoisted (non-".layers") copy. |
| 386 | if spec_layer != self.model.mtp_start_layer_idx and ( |
| 387 | ".layers" not in name |
| 388 | ): |
| 389 | continue |
| 390 | if name in pp_missing_layer_names: |
| 391 | continue |
| 392 | # The base model uses an attn-residual scheme whose per-layer |
| 393 | # weights (self_attention_res_*, mlp_res_*) are not used by |
| 394 | # the draft block; such names have no matching parameter and |
| 395 | # are safely skipped. |
| 396 | if name not in params_dict: |
| 397 | continue |
| 398 | |
| 399 | param = params_dict[name] |
| 400 | weight_loader = getattr( |
| 401 | param, "weight_loader", default_weight_loader |
| 402 | ) |
| 403 | weight_loader(param, loaded_weight) |
| 404 | loaded_params.add(name) |
| 405 | |
| 406 | # Validate that weights were loaded for each expected MTP layer. |
| 407 | loaded_layers: set[int] = set() |
| 408 | for param_name in loaded_params: |
| 409 | spec_layer = get_spec_layer_idx_from_weight_name(self.config, param_name) |
| 410 | if spec_layer is not None: |
| 411 | loaded_layers.add(spec_layer) |
| 412 | for layer_idx in range( |
| 413 | self.model.mtp_start_layer_idx, |
| 414 | self.model.mtp_start_layer_idx + self.model.num_mtp_layers, |
| 415 | ): |
| 416 | if layer_idx not in loaded_layers: |
| 417 | raise ValueError( |
| 418 | f"MTP speculative decoding layer {layer_idx} weights " |
| 419 | f"missing from checkpoint. The checkpoint may not include " |
| 420 | f"the MTP layer weights. Use a checkpoint that includes " |
| 421 | f"MTP layer weights, or disable speculative decoding." |
| 422 | ) |
| 423 | |
| 424 | if use_mega_moe: |
| 425 | for module in self.modules(): |
| 426 | if isinstance(module, KimiMoE) and module.use_mega_moe: |
| 427 | module.experts.finalize_weights() |
| 428 | |
| 429 | return loaded_params |
| 430 | |
| 431 | def _rewrite_spec_layer_name(self, spec_layer: int, name: str) -> str: |
| 432 | """Rewrite a checkpoint weight name to this module's parameter path. |
| 433 | |
| 434 | Top-level MTP submodules (enorm/hnorm/eh_proj/shared_head) stay under |
| 435 | ``model.layers.{spec_layer}.*``; the shared ``embed_tokens`` is hoisted |
| 436 | to ``model.*``; everything else is a transformer-block weight and gets |
| 437 | ``.mtp_block`` inserted. |
| 438 | """ |
| 439 | spec_layer_weight_names = [ |
| 440 | "embed_tokens", |
| 441 | "enorm", |
| 442 | "hnorm", |
| 443 | "eh_proj", |
| 444 | "shared_head", |
| 445 | ] |
| 446 | shared_weight_names = ["embed_tokens"] |
| 447 | spec_layer_weight = False |
| 448 | shared_weight = False |
| 449 | for weight_name in spec_layer_weight_names: |
| 450 | if weight_name in name: |
| 451 | spec_layer_weight = True |
| 452 | if weight_name in shared_weight_names: |
| 453 | shared_weight = True |
| 454 | break |
| 455 | if not spec_layer_weight: |
| 456 | name = name.replace( |
| 457 | f"model.layers.{spec_layer}.", |
| 458 | f"model.layers.{spec_layer}.mtp_block.", |
| 459 | ) |
| 460 | elif shared_weight: |
| 461 | name = name.replace(f"model.layers.{spec_layer}.", "model.") |
| 462 | return name |