| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Qwen3 DSpark draft model for semi-autoregressive drafting. |
| 4 | |
| 5 | DSpark drafts a whole block in one parallel pass (DFlash-style: context-KV |
| 6 | precompute + a non-causal query-block forward) and then injects intra-block |
| 7 | dependency with a lightweight sequential Markov head. |
| 8 | |
| 9 | The parallel backbone is a standard Qwen3 decoder stack reused from the |
| 10 | DFlash Qwen3 draft (see qwen3_dflash.py). DSpark adds: |
| 11 | * ``markov_head``: low-rank V x r / r x V transition bias added to the base |
| 12 | logits, sampled left-to-right by the speculator (the sequential stage). |
| 13 | * ``confidence_head``: per-position acceptance-probability estimate. |
| 14 | |
| 15 | DSparkMarkovHead and DSparkConfidenceHead are shared with the DSV4-style DSpark model. |
| 16 | """ |
| 17 | |
| 18 | from collections.abc import Iterable |
| 19 | |
| 20 | import torch |
| 21 | import torch.nn as nn |
| 22 | |
| 23 | from vllm.config import VllmConfig |
| 24 | from vllm.logger import init_logger |
| 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.quantization import QuantizationConfig |
| 28 | from vllm.model_executor.layers.quantization.utils.nvfp4_emulation_utils import ( |
| 29 | dequantize_to_dtype, |
| 30 | nvfp4_gathered_bias, |
| 31 | ) |
| 32 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 33 | ParallelLMHead, |
| 34 | ) |
| 35 | from vllm.platforms import current_platform |
| 36 | |
| 37 | from .qwen3_dflash import DFlashQwen3ForCausalLM, DFlashQwen3Model |
| 38 | from .utils import ( |
| 39 | AutoWeightsLoader, |
| 40 | WeightsMapper, |
| 41 | maybe_prefix, |
| 42 | process_eagle_weight, |
| 43 | ) |
| 44 | |
| 45 | logger = init_logger(__name__) |
| 46 | |
| 47 | |
| 48 | class DSparkMarkovHead(nn.Module): |
| 49 | """Sequential transition-bias head (low-rank V x r, r x V). |
| 50 | |
| 51 | ``markov_w1[token]`` embeds the previously sampled token (target vocab, |
| 52 | ``vocab_size``); ``markov_w2`` projects it to a draft-vocab bias |
| 53 | (``draft_vocab_size``) added to the base draft logits. The two sizes |
| 54 | coincide for full-vocab drafts. |
| 55 | |
| 56 | Both weights are replicated because the head runs sequentially for every |
| 57 | draft position. Sharding them would add an all-reduce and a full-vocab |
| 58 | gather to each position. |
| 59 | """ |
| 60 | |
| 61 | def __init__( |
| 62 | self, |
| 63 | vocab_size: int, |
| 64 | draft_vocab_size: int, |
| 65 | markov_rank: int, |
| 66 | prefix: str, |
| 67 | quant_config: QuantizationConfig | None = None, |
| 68 | retain_weight_for_gather: bool = False, |
| 69 | ) -> None: |
| 70 | super().__init__() |
| 71 | self.markov_w1 = nn.Embedding(vocab_size, markov_rank) |
| 72 | self.markov_w2 = ParallelLMHead( |
| 73 | draft_vocab_size, |
| 74 | markov_rank, |
| 75 | bias=False, |
| 76 | quant_config=quant_config, |
| 77 | prefix=maybe_prefix(prefix, "markov_w2"), |
| 78 | disable_tp=True, |
| 79 | ) |
| 80 | self.markov_w2._retain_weight_for_gather = retain_weight_for_gather |
| 81 | self.markov_w2.is_w4a16_nvfp4 = False |
| 82 | |
| 83 | def embed(self, token_ids: torch.Tensor) -> torch.Tensor: |
| 84 | """r-dim Markov embedding of ``token_ids`` ([B] -> [B, r]).""" |
| 85 | return self.markov_w1(token_ids) |
| 86 | |
| 87 | def bias( |
| 88 | self, |
| 89 | markov_embed: torch.Tensor, |
| 90 | logits_processor: LogitsProcessor, |
| 91 | ) -> torch.Tensor: |
| 92 | """Vocab-size transition bias from a Markov embedding ([B, r] -> [B, V]).""" |
| 93 | return logits_processor(self.markov_w2, markov_embed) |
| 94 | |
| 95 | def apply_bias_gathered( |
| 96 | self, |
| 97 | markov_embed: torch.Tensor, |
| 98 | logits: torch.Tensor, |
| 99 | values: torch.Tensor, |
| 100 | index: torch.Tensor, |
| 101 | scale: float = 1.0, |
| 102 | ) -> torch.Tensor: |
| 103 | """Apply the Markov bias only to selected rows of ``logits``. |
| 104 | |
| 105 | The caller initializes ``logits`` to ``-inf`` once for all draft |
| 106 | positions. This method scatters the corrected candidate values into |
| 107 | that dense buffer so the normal sampler sees the truncated proposal. |
| 108 | """ |
| 109 | if self.markov_w2._retain_weight_for_gather and self.markov_w2.is_w4a16_nvfp4: |
| 110 | # Use the fast Triton kernel for W4A16 NVFP4 weights with BF16 |
| 111 | # activations on CUDA. FP16 activations, non-CUDA accelerators, and |
| 112 | # CPU use the slower dequantize/scatter fallback below. |
| 113 | if ( |
| 114 | current_platform.is_cuda() |
| 115 | and markov_embed.is_cuda |
| 116 | and markov_embed.dtype == torch.bfloat16 |
| 117 | and values.dtype == torch.bfloat16 |
| 118 | and logits.dtype == torch.bfloat16 |
| 119 | ): |
| 120 | nvfp4_gathered_bias( |
| 121 | markov_embed, |
| 122 | self.markov_w2._nvfp4_weight_for_gather, |
| 123 | self.markov_w2._nvfp4_weight_scale_for_gather, |
| 124 | self.markov_w2._nvfp4_weight_global_scale_for_gather, |
| 125 | values, |
| 126 | index, |
| 127 | logits, |
| 128 | scale, |
| 129 | ) |
| 130 | return logits |
| 131 | |
| 132 | flat_index = index.reshape(-1) |
| 133 | packed_weight = self.markov_w2._nvfp4_weight_for_gather.index_select( |
| 134 | 0, flat_index |
| 135 | ) |
| 136 | weight_scale = self.markov_w2._nvfp4_weight_scale_for_gather.index_select( |
| 137 | 0, flat_index |
| 138 | ) |
| 139 | weight = dequantize_to_dtype( |
| 140 | packed_weight, |
| 141 | weight_scale, |
| 142 | self.markov_w2._nvfp4_weight_global_scale_for_gather, |
| 143 | dtype=markov_embed.dtype, |
| 144 | block_size=self.markov_w2._nvfp4_group_size_for_gather, |
| 145 | swizzle=False, |
| 146 | ).view(*index.shape, -1) |
| 147 | else: |
| 148 | weight = self.markov_w2.weight[index] |
| 149 | if weight.dtype != markov_embed.dtype: |
| 150 | raise RuntimeError( |
| 151 | "Gathered Markov projection requires unquantized weights or " |
| 152 | "retained ModelOpt W4A16 NVFP4 weights." |
| 153 | ) |
| 154 | corrected = values.unsqueeze(-1).clone() |
| 155 | corrected.baddbmm_( |
| 156 | weight, |
| 157 | markov_embed.unsqueeze(-1), |
| 158 | beta=1.0, |
| 159 | alpha=scale, |
| 160 | ) |
| 161 | return logits.scatter_(1, index, corrected.squeeze(-1)) |
| 162 | |
| 163 | |
| 164 | class DSparkConfidenceHead(nn.Module): |
| 165 | """DSpark acceptance-confidence head.""" |
| 166 | |
| 167 | def __init__( |
| 168 | self, |
| 169 | input_dim: int, |
| 170 | prefix: str, |
| 171 | bias: bool = False, |
| 172 | with_markov: bool = True, |
| 173 | ) -> None: |
| 174 | super().__init__() |
| 175 | self.with_markov = with_markov |
| 176 | self.proj = ReplicatedLinear( |
| 177 | input_dim, |
| 178 | 1, |
| 179 | bias=bias, |
| 180 | return_bias=False, |
| 181 | params_dtype=torch.float32, |
| 182 | prefix=maybe_prefix(prefix, "proj"), |
| 183 | ) |
| 184 | |
| 185 | def forward(self, hidden: torch.Tensor, markov_embed: torch.Tensor) -> torch.Tensor: |
| 186 | x = ( |
| 187 | torch.cat([hidden, markov_embed], dim=-1) if self.with_markov else hidden |
| 188 | ).float() |
| 189 | return self.proj(x).squeeze(-1) |
| 190 | |
| 191 | |
| 192 | class Qwen3DSparkModel(DFlashQwen3Model): |
| 193 | """DFlash Qwen3 backbone + DSpark Markov / confidence heads.""" |
| 194 | |
| 195 | def __init__( |
| 196 | self, |
| 197 | *, |
| 198 | vllm_config: VllmConfig, |
| 199 | start_layer_id: int = 0, |
| 200 | prefix: str = "", |
| 201 | ) -> None: |
| 202 | super().__init__( |
| 203 | vllm_config=vllm_config, start_layer_id=start_layer_id, prefix=prefix |
| 204 | ) |
| 205 | config = self.config |
| 206 | draft_vocab_size = ( |
| 207 | getattr(config, "draft_vocab_size", None) or config.vocab_size |
| 208 | ) |
| 209 | self.markov_head = DSparkMarkovHead( |
| 210 | config.vocab_size, |
| 211 | draft_vocab_size, |
| 212 | config.markov_rank, |
| 213 | prefix=maybe_prefix(prefix, "markov_head"), |
| 214 | quant_config=self.quant_config, |
| 215 | retain_weight_for_gather=( |
| 216 | getattr(config, "dspark_draft_topk", None) is not None |
| 217 | ), |
| 218 | ) |
| 219 | self.confidence_head: DSparkConfidenceHead | None = None |
| 220 | if getattr(config, "enable_confidence_head", False): |
| 221 | with_markov = getattr(config, "confidence_head_with_markov", False) |
| 222 | input_dim = config.hidden_size |
| 223 | if with_markov: |
| 224 | input_dim += config.markov_rank |
| 225 | self.confidence_head = DSparkConfidenceHead( |
| 226 | input_dim, |
| 227 | prefix=maybe_prefix(prefix, "confidence_head"), |
| 228 | bias=True, |
| 229 | with_markov=with_markov, |
| 230 | ) |
| 231 | |
| 232 | |
| 233 | class Qwen3DSparkForCausalLM(DFlashQwen3ForCausalLM): |
| 234 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 235 | nn.Module.__init__(self) |
| 236 | speculative_config = vllm_config.speculative_config |
| 237 | assert speculative_config is not None |
| 238 | self.draft_model_config = speculative_config.draft_model_config |
| 239 | self.config = self.draft_model_config.hf_config |
| 240 | if getattr(self.config, "draft_vocab_size", None) is None: |
| 241 | self.config.draft_vocab_size = getattr(self.config, "vocab_size", None) |
| 242 | target_layer_num = vllm_config.model_config.get_total_num_hidden_layers() |
| 243 | self.model = Qwen3DSparkModel( |
| 244 | vllm_config=vllm_config, |
| 245 | prefix=maybe_prefix(prefix, "model"), |
| 246 | start_layer_id=target_layer_num, |
| 247 | ) |
| 248 | |
| 249 | logit_scale = getattr(self.config, "logit_scale", 1.0) |
| 250 | self.lm_head = ParallelLMHead( |
| 251 | self.config.draft_vocab_size, |
| 252 | self.config.hidden_size, |
| 253 | prefix=maybe_prefix(prefix, "lm_head"), |
| 254 | ) |
| 255 | self.logits_processor = LogitsProcessor( |
| 256 | self.config.draft_vocab_size, scale=logit_scale |
| 257 | ) |
| 258 | self.target_vocab_size = vllm_config.model_config.get_vocab_size() |
| 259 | # Only a pruned draft (subset of the target vocab) needs a d2t mapping to |
| 260 | # scatter ids up. A draft >= the target vocab (e.g. padded to the target's |
| 261 | # physical embedding size) maps identically, so no d2t is needed. |
| 262 | if self.config.draft_vocab_size < self.target_vocab_size: |
| 263 | self.draft_id_to_target_id = nn.Parameter( |
| 264 | torch.zeros(self.config.draft_vocab_size, dtype=torch.long), |
| 265 | requires_grad=False, |
| 266 | ) |
| 267 | else: |
| 268 | self.draft_id_to_target_id = None |
| 269 | |
| 270 | def get_draft_kv_cache_layer_names(self) -> list[str]: |
| 271 | return [layer.self_attn.attn.layer_name for layer in self.model.layers] |
| 272 | |
| 273 | def compute_draft_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 274 | # Draft-vocab logits without the d2t scatter: the speculator adds the |
| 275 | # Markov bias in draft space, then remaps via map_draft_to_target. |
| 276 | return self.logits_processor(self.lm_head, hidden_states) |
| 277 | |
| 278 | def map_draft_to_target(self, draft_ids: torch.Tensor) -> torch.Tensor: |
| 279 | # Map draft-vocab ids to target ids (identity for full-vocab drafts). |
| 280 | if self.draft_id_to_target_id is None: |
| 281 | return draft_ids |
| 282 | return draft_ids + self.draft_id_to_target_id[draft_ids] |
| 283 | |
| 284 | def markov_embed(self, token_ids: torch.Tensor) -> torch.Tensor: |
| 285 | return self.model.markov_head.embed(token_ids) |
| 286 | |
| 287 | def markov_bias(self, markov_embed: torch.Tensor) -> torch.Tensor: |
| 288 | return self.model.markov_head.bias(markov_embed, self.logits_processor) |
| 289 | |
| 290 | def apply_markov_bias_gathered( |
| 291 | self, |
| 292 | markov_embed: torch.Tensor, |
| 293 | logits: torch.Tensor, |
| 294 | values: torch.Tensor, |
| 295 | index: torch.Tensor, |
| 296 | ) -> torch.Tensor: |
| 297 | return self.model.markov_head.apply_bias_gathered( |
| 298 | markov_embed, |
| 299 | logits, |
| 300 | values, |
| 301 | index, |
| 302 | self.logits_processor.scale, |
| 303 | ) |
| 304 | |
| 305 | def compute_confidence( |
| 306 | self, head_hidden: torch.Tensor, markov_embed: torch.Tensor |
| 307 | ) -> torch.Tensor: |
| 308 | """Per-position acceptance probability for each drafted token.""" |
| 309 | assert self.model.confidence_head is not None |
| 310 | return torch.sigmoid(self.model.confidence_head(head_hidden, markov_embed)) |
| 311 | |
| 312 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): |
| 313 | model_weights = {} |
| 314 | includes_embed_tokens = False |
| 315 | includes_lm_head = False |
| 316 | includes_draft_id_mapping = False |
| 317 | includes_confidence_head = False |
| 318 | for name, loaded_weight in weights: |
| 319 | # t2d is training-only; the draft remaps via d2t at sampling time. |
| 320 | if "t2d" in name: |
| 321 | continue |
| 322 | if "d2t" in name: |
| 323 | name = name.replace("d2t", "draft_id_to_target_id") |
| 324 | includes_draft_id_mapping = True |
| 325 | elif "lm_head" not in name: |
| 326 | name = "model." + name |
| 327 | if "embed_tokens" in name: |
| 328 | includes_embed_tokens = True |
| 329 | if "lm_head" in name: |
| 330 | includes_lm_head = True |
| 331 | if "confidence_head" in name: |
| 332 | includes_confidence_head = True |
| 333 | model_weights[name] = loaded_weight |
| 334 | # Sets has_own_embed_tokens / has_own_lm_head so load_dspark_model |
| 335 | # knows whether to keep these or alias the target's. |
| 336 | process_eagle_weight(self, name) |
| 337 | |
| 338 | # mask_embedding is an unused placeholder param; DSpark masks via the vocab row. |
| 339 | # embed_tokens / lm_head are optional; when omitted they are shared from |
| 340 | # the target by load_dspark_model, so skip the unloaded params here. |
| 341 | uses_expanded_input_vocab = self.config.vocab_size > self.target_vocab_size |
| 342 | uses_reduced_vocab = self.config.draft_vocab_size < self.target_vocab_size |
| 343 | if uses_expanded_input_vocab and not includes_embed_tokens: |
| 344 | raise ValueError( |
| 345 | "Qwen3 DSpark checkpoints whose input vocab_size is larger than " |
| 346 | "the target vocabulary must include embed_tokens weights." |
| 347 | ) |
| 348 | if uses_reduced_vocab and not includes_lm_head: |
| 349 | raise ValueError( |
| 350 | "Reduced-vocabulary Qwen3 DSpark checkpoints must include " |
| 351 | "lm_head weights; the full target lm_head cannot be shared." |
| 352 | ) |
| 353 | if uses_reduced_vocab and not includes_draft_id_mapping: |
| 354 | raise ValueError( |
| 355 | "Reduced-vocabulary Qwen3 DSpark checkpoints must include a " |
| 356 | "d2t mapping so sampled draft ids can be converted to target ids." |
| 357 | ) |
| 358 | |
| 359 | orig_to_new_substr = {"mask_embedding": None} |
| 360 | if not includes_embed_tokens: |
| 361 | orig_to_new_substr["embed_tokens"] = None |
| 362 | if not includes_lm_head: |
| 363 | orig_to_new_substr["lm_head"] = None |
| 364 | if not includes_draft_id_mapping: |
| 365 | orig_to_new_substr["draft_id_to_target_id"] = None |
| 366 | if self.model.confidence_head is None or not includes_confidence_head: |
| 367 | self.model.confidence_head = None |
| 368 | orig_to_new_substr["confidence_head"] = None |
| 369 | mapper = WeightsMapper(orig_to_new_substr=orig_to_new_substr) |
| 370 | loader = AutoWeightsLoader(self) |
| 371 | loader.load_weights(model_weights.items(), mapper=mapper) |
| 372 | self.model._build_fused_kv_buffers() |