| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """DSpark draft model for DeepSeek-V4 on Intel XPU. |
| 4 | |
| 5 | Minimal XPU port of nvidia/dspark.py. Replaces tilelang MHC kernels with |
| 6 | the platform-agnostic custom ops (HCHeadOp, MHCPostOp) already used by the |
| 7 | XPU MTP path, and uses the XPU Triton-based qnorm_rope_kv_fp8_insert for |
| 8 | context KV precomputation. |
| 9 | """ |
| 10 | |
| 11 | from collections.abc import Iterable |
| 12 | |
| 13 | import regex as re |
| 14 | import torch |
| 15 | import torch.nn as nn |
| 16 | |
| 17 | from vllm.config import VllmConfig, get_current_vllm_config |
| 18 | from vllm.distributed import ( |
| 19 | get_tensor_model_parallel_rank, |
| 20 | get_tensor_model_parallel_world_size, |
| 21 | ) |
| 22 | from vllm.logger import init_logger |
| 23 | from vllm.model_executor.layers.fused_moe import ( |
| 24 | fused_moe_make_expert_params_mapping, |
| 25 | ) |
| 26 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 27 | from vllm.model_executor.layers.linear import ReplicatedLinear |
| 28 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 29 | from vllm.model_executor.layers.mhc import HCHeadOp, MHCPostOp |
| 30 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 31 | ParallelLMHead, |
| 32 | VocabParallelEmbedding, |
| 33 | ) |
| 34 | from vllm.model_executor.model_loader.weight_utils import default_weight_loader |
| 35 | from vllm.model_executor.models.qwen3_dspark import DSparkMarkovHead |
| 36 | from vllm.model_executor.models.utils import maybe_prefix |
| 37 | |
| 38 | from .model import ( |
| 39 | DeepseekV4DecoderLayer, |
| 40 | make_deepseek_v4_expert_params_mapping, |
| 41 | ) |
| 42 | |
| 43 | logger = init_logger(__name__) |
| 44 | |
| 45 | _EXPERT_SCALE_RE = re.compile(r"\.experts\.\d+\.w[123]\.scale$") |
| 46 | |
| 47 | |
| 48 | class DSparkDeepseekV4Model(nn.Module): |
| 49 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 50 | super().__init__() |
| 51 | assert vllm_config.speculative_config is not None |
| 52 | config = vllm_config.speculative_config.draft_model_config.hf_config |
| 53 | self.config = config |
| 54 | self.hidden_size = config.hidden_size |
| 55 | self.hc_mult = config.hc_mult |
| 56 | self.hc_eps = config.hc_eps |
| 57 | self.rms_norm_eps = config.rms_norm_eps |
| 58 | self.num_hidden_layers = config.num_hidden_layers |
| 59 | self.target_layer_ids = tuple(config.dspark_target_layer_ids) |
| 60 | |
| 61 | self.num_dspark_layers = getattr(config, "n_mtp_layers", None) or 3 |
| 62 | |
| 63 | # Shared with target (aliased by speculator loading utility). |
| 64 | self.embed_tokens = VocabParallelEmbedding( |
| 65 | config.vocab_size, |
| 66 | config.hidden_size, |
| 67 | prefix=maybe_prefix(prefix, "embed_tokens"), |
| 68 | ) |
| 69 | |
| 70 | self.main_proj = ReplicatedLinear( |
| 71 | config.hidden_size * len(self.target_layer_ids), |
| 72 | config.hidden_size, |
| 73 | bias=False, |
| 74 | return_bias=False, |
| 75 | quant_config=vllm_config.quant_config, |
| 76 | prefix=maybe_prefix(prefix, "main_proj"), |
| 77 | ) |
| 78 | self.main_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 79 | |
| 80 | current_vllm_config = get_current_vllm_config() |
| 81 | self.layers = nn.ModuleList( |
| 82 | [ |
| 83 | DeepseekV4DecoderLayer( |
| 84 | current_vllm_config, |
| 85 | prefix=maybe_prefix(prefix, f"layers.{self.num_hidden_layers + i}"), |
| 86 | ) |
| 87 | for i in range(self.num_dspark_layers) |
| 88 | ] |
| 89 | ) |
| 90 | |
| 91 | # Heads: final norm + hc_head, and the Markov head |
| 92 | self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) |
| 93 | hc_dim = self.hc_mult * config.hidden_size |
| 94 | self.hc_head_fn = nn.Parameter( |
| 95 | torch.empty(self.hc_mult, hc_dim, dtype=torch.float32), |
| 96 | requires_grad=False, |
| 97 | ) |
| 98 | self.hc_head_base = nn.Parameter( |
| 99 | torch.empty(self.hc_mult, dtype=torch.float32), requires_grad=False |
| 100 | ) |
| 101 | self.hc_head_scale = nn.Parameter( |
| 102 | torch.empty(1, dtype=torch.float32), requires_grad=False |
| 103 | ) |
| 104 | draft_vocab_size = ( |
| 105 | getattr(config, "draft_vocab_size", None) or config.vocab_size |
| 106 | ) |
| 107 | self.markov_head = DSparkMarkovHead( |
| 108 | config.vocab_size, |
| 109 | draft_vocab_size, |
| 110 | config.dspark_markov_rank, |
| 111 | prefix=maybe_prefix(prefix, "markov_head"), |
| 112 | ) |
| 113 | |
| 114 | # XPU MHC ops (replaces tilelang) |
| 115 | self.mhc_post_op = MHCPostOp() |
| 116 | self.hc_head_op = HCHeadOp() |
| 117 | |
| 118 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 119 | return self.embed_tokens(input_ids) |
| 120 | |
| 121 | def combine_hidden_states(self, aux_hidden_states: torch.Tensor) -> torch.Tensor: |
| 122 | """main_x = main_norm(main_proj(concat of target aux hidden states)).""" |
| 123 | return self.main_norm(self.main_proj(aux_hidden_states)) |
| 124 | |
| 125 | @torch.inference_mode() |
| 126 | def precompute_and_store_context_kv( |
| 127 | self, |
| 128 | main_x: torch.Tensor, |
| 129 | context_positions: torch.Tensor, |
| 130 | context_slot_mappings: list[torch.Tensor | None] | None = None, |
| 131 | ) -> None: |
| 132 | """Insert the sliding-window context KV for every draft layer. |
| 133 | |
| 134 | Each layer derives its context KV from the SAME projected target hidden |
| 135 | ``main_x``, via that layer's own wkv + kv_norm + RoPE + quant, then |
| 136 | writes it at the layer's context slots. |
| 137 | """ |
| 138 | for i, layer in enumerate(self.layers): |
| 139 | slot_mapping = ( |
| 140 | None if context_slot_mappings is None else context_slot_mappings[i] |
| 141 | ) |
| 142 | attn = layer.attn |
| 143 | # wkv part of the fused wq_a|wkv projection (q_lora part discarded) |
| 144 | qr_kv, _ = attn.fused_wqa_wkv(main_x) |
| 145 | kv = qr_kv[..., attn.q_lora_rank :] |
| 146 | kv = attn.kv_norm(kv) |
| 147 | if slot_mapping is None: |
| 148 | continue |
| 149 | _insert_context_kv(attn, kv, context_positions, slot_mapping) |
| 150 | |
| 151 | def forward( |
| 152 | self, |
| 153 | input_ids: torch.Tensor, |
| 154 | positions: torch.Tensor, |
| 155 | inputs_embeds: torch.Tensor | None = None, |
| 156 | ) -> torch.Tensor: |
| 157 | if inputs_embeds is None: |
| 158 | inputs_embeds = self.embed_input_ids(input_ids) |
| 159 | # Expand to hc_mult copies for hyper-connections ([T, H] -> [T, hc, H]). |
| 160 | hidden_states = inputs_embeds.unsqueeze(-2).repeat(1, self.hc_mult, 1) |
| 161 | |
| 162 | residual = post_mix = res_mix = None |
| 163 | for layer in self.layers: |
| 164 | hidden_states, residual, post_mix, res_mix = layer( |
| 165 | hidden_states, |
| 166 | positions, |
| 167 | input_ids, |
| 168 | post_mix, |
| 169 | res_mix, |
| 170 | residual, |
| 171 | ) |
| 172 | # mhc_post: merge hyper-connection copies |
| 173 | hidden_states = self.mhc_post_op(hidden_states, residual, post_mix, res_mix) |
| 174 | # hc_head: reduces hc copies; return pre-norm head hidden |
| 175 | hidden_states = self.hc_head_op( |
| 176 | hidden_states, |
| 177 | self.hc_head_fn, |
| 178 | self.hc_head_scale, |
| 179 | self.hc_head_base, |
| 180 | self.rms_norm_eps, |
| 181 | self.hc_eps, |
| 182 | ) |
| 183 | return hidden_states |
| 184 | |
| 185 | |
| 186 | def _insert_context_kv( |
| 187 | attn: nn.Module, |
| 188 | kv: torch.Tensor, |
| 189 | positions: torch.Tensor, |
| 190 | slot_mapping: torch.Tensor, |
| 191 | ) -> None: |
| 192 | """RoPE + quant + paged-cache insert of (already kv_norm'd) context KV. |
| 193 | |
| 194 | On XPU, we reuse the same xpu_qnorm_rope_kv_fp8_insert kernel used in |
| 195 | the forward path, passing a dummy q (result discarded). |
| 196 | """ |
| 197 | from vllm.models.deepseek_v4.xpu.xpu_qnorm_rope_kv_fp8_insert import ( |
| 198 | xpu_qnorm_rope_kv_fp8_insert, |
| 199 | ) |
| 200 | |
| 201 | swa_cache = attn.swa_cache_layer.kv_cache |
| 202 | block_size = attn.swa_cache_layer.block_size |
| 203 | cos_sin_cache = attn.rotary_emb.cos_sin_cache |
| 204 | n_ctx = kv.shape[0] |
| 205 | |
| 206 | # Dummy q — we only care about the KV insert side effect. |
| 207 | dummy_q = torch.empty( |
| 208 | (n_ctx, attn.n_local_heads, attn.head_dim), |
| 209 | dtype=kv.dtype, |
| 210 | device=kv.device, |
| 211 | ) |
| 212 | xpu_qnorm_rope_kv_fp8_insert( |
| 213 | dummy_q, |
| 214 | kv, |
| 215 | swa_cache, |
| 216 | slot_mapping, |
| 217 | positions, |
| 218 | cos_sin_cache, |
| 219 | attn.eps, |
| 220 | block_size, |
| 221 | ) |
| 222 | |
| 223 | |
| 224 | class DSparkDeepseekV4ForCausalLM(nn.Module): |
| 225 | """XPU DSpark draft model entry point for DeepSeek-V4.""" |
| 226 | |
| 227 | has_own_embed_tokens = False |
| 228 | has_own_lm_head = False |
| 229 | |
| 230 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 231 | super().__init__() |
| 232 | assert vllm_config.speculative_config is not None |
| 233 | self.draft_model_config = vllm_config.speculative_config.draft_model_config |
| 234 | self.config = self.draft_model_config.hf_config |
| 235 | self.model = DSparkDeepseekV4Model( |
| 236 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 237 | ) |
| 238 | # Shared with the target (aliased by the speculator's load utility). |
| 239 | self.lm_head = ParallelLMHead( |
| 240 | self.config.vocab_size, |
| 241 | self.config.hidden_size, |
| 242 | prefix=maybe_prefix(prefix, "lm_head"), |
| 243 | ) |
| 244 | self.logits_processor = LogitsProcessor(self.config.vocab_size) |
| 245 | |
| 246 | # --- Hooks used by the speculator --- |
| 247 | |
| 248 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 249 | return self.model.embed_input_ids(input_ids) |
| 250 | |
| 251 | def combine_hidden_states(self, aux_hidden_states: torch.Tensor) -> torch.Tensor: |
| 252 | return self.model.combine_hidden_states(aux_hidden_states) |
| 253 | |
| 254 | def get_draft_kv_cache_layer_names(self) -> list[str]: |
| 255 | return [layer.attn.swa_cache_layer.prefix for layer in self.model.layers] |
| 256 | |
| 257 | def precompute_and_store_context_kv( |
| 258 | self, |
| 259 | context_states: torch.Tensor, |
| 260 | context_positions: torch.Tensor, |
| 261 | context_slot_mappings: list[torch.Tensor | None] | None = None, |
| 262 | ) -> None: |
| 263 | self.model.precompute_and_store_context_kv( |
| 264 | context_states, context_positions, context_slot_mappings |
| 265 | ) |
| 266 | |
| 267 | def forward( |
| 268 | self, |
| 269 | input_ids: torch.Tensor, |
| 270 | positions: torch.Tensor, |
| 271 | inputs_embeds: torch.Tensor | None = None, |
| 272 | ) -> torch.Tensor: |
| 273 | return self.model(input_ids, positions, inputs_embeds) |
| 274 | |
| 275 | def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 276 | """Base logits U_k = lm_head(norm(head_hidden)).""" |
| 277 | return self.logits_processor(self.lm_head, self.model.norm(hidden_states)) |
| 278 | |
| 279 | def compute_draft_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 280 | return self.compute_logits(hidden_states) |
| 281 | |
| 282 | def map_draft_to_target(self, draft_ids: torch.Tensor) -> torch.Tensor: |
| 283 | return draft_ids # full-vocab: draft ids are target ids |
| 284 | |
| 285 | def markov_embed(self, token_ids: torch.Tensor) -> torch.Tensor: |
| 286 | return self.model.markov_head.embed(token_ids) |
| 287 | |
| 288 | def markov_bias(self, markov_embed: torch.Tensor) -> torch.Tensor: |
| 289 | return self.model.markov_head.bias(markov_embed, self.logits_processor) |
| 290 | |
| 291 | # --- Weight loading --- |
| 292 | |
| 293 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 294 | """Load ``mtp.{0,1,2}.*`` draft weights from the target checkpoint.""" |
| 295 | first_layer = self.model.layers[0] |
| 296 | use_mega_moe = first_layer.ffn.use_mega_moe |
| 297 | if use_mega_moe: |
| 298 | expert_mapping = make_deepseek_v4_expert_params_mapping( |
| 299 | self.config.n_routed_experts |
| 300 | ) |
| 301 | else: |
| 302 | expert_mapping = fused_moe_make_expert_params_mapping( |
| 303 | self, |
| 304 | ckpt_gate_proj_name="w1", |
| 305 | ckpt_down_proj_name="w2", |
| 306 | ckpt_up_proj_name="w3", |
| 307 | num_experts=self.config.n_routed_experts, |
| 308 | ) |
| 309 | expert_scale_suffix = ( |
| 310 | ".weight_scale" |
| 311 | if getattr(self.config, "expert_dtype", "fp4") == "fp4" |
| 312 | else ".weight_scale_inv" |
| 313 | ) |
| 314 | |
| 315 | stacked_params_mapping = [ |
| 316 | ("gate_up_proj", "w1", 0), |
| 317 | ("gate_up_proj", "w3", 1), |
| 318 | ("attn.fused_wqa_wkv", "attn.wq_a", 0), |
| 319 | ("attn.fused_wqa_wkv", "attn.wkv", 1), |
| 320 | ] |
| 321 | |
| 322 | params_dict = dict(self.named_parameters()) |
| 323 | loaded_params: set[str] = set() |
| 324 | |
| 325 | tp_size = get_tensor_model_parallel_world_size() |
| 326 | tp_rank = get_tensor_model_parallel_rank() |
| 327 | n_local_head = self.config.num_attention_heads // tp_size |
| 328 | head_start = n_local_head * tp_rank |
| 329 | head_end = n_local_head * (tp_rank + 1) |
| 330 | |
| 331 | for name, loaded_weight in weights: |
| 332 | mapped = self._remap_dspark_name(name) |
| 333 | if mapped is None: |
| 334 | continue |
| 335 | name = mapped |
| 336 | |
| 337 | # .scale -> per-method scale suffix |
| 338 | if name.endswith(".scale"): |
| 339 | suffix = ( |
| 340 | expert_scale_suffix |
| 341 | if _EXPERT_SCALE_RE.search(name) |
| 342 | else ".weight_scale_inv" |
| 343 | ) |
| 344 | name = name.removesuffix(".scale") + suffix |
| 345 | |
| 346 | # Expert weights |
| 347 | if ".experts." in name: |
| 348 | if ( |
| 349 | "weight_scale" in name |
| 350 | and loaded_weight.dtype == torch.float8_e8m0fnu |
| 351 | ): |
| 352 | loaded_weight = loaded_weight.view(torch.uint8) |
| 353 | for param_name, weight_name, expert_id, shard_id in expert_mapping: |
| 354 | if weight_name not in name: |
| 355 | continue |
| 356 | name_mapped = name.replace(weight_name, param_name) |
| 357 | if name_mapped not in params_dict: |
| 358 | continue |
| 359 | param = params_dict[name_mapped] |
| 360 | success = param.weight_loader( |
| 361 | param, |
| 362 | loaded_weight, |
| 363 | name_mapped, |
| 364 | shard_id=shard_id, |
| 365 | expert_id=expert_id, |
| 366 | return_success=True, |
| 367 | ) |
| 368 | if success: |
| 369 | loaded_params.add(name_mapped) |
| 370 | break |
| 371 | continue |
| 372 | |
| 373 | # Stacked params (decoder-layer only) |
| 374 | is_layer_param = name.startswith("model.layers.") |
| 375 | for param_name, weight_name, stacked_shard_id in stacked_params_mapping: |
| 376 | if not is_layer_param or weight_name not in name: |
| 377 | continue |
| 378 | name = name.replace(weight_name, param_name) |
| 379 | if name not in params_dict: |
| 380 | break |
| 381 | param = params_dict[name] |
| 382 | param.weight_loader(param, loaded_weight, stacked_shard_id) |
| 383 | loaded_params.add(name) |
| 384 | break |
| 385 | else: |
| 386 | if "attn_sink" in name: |
| 387 | if name not in params_dict: |
| 388 | continue |
| 389 | narrow = loaded_weight[head_start:head_end] |
| 390 | params_dict[name][: narrow.shape[0]].copy_(narrow) |
| 391 | loaded_params.add(name) |
| 392 | continue |
| 393 | if ".shared_experts.w2" in name: |
| 394 | name = name.replace( |
| 395 | ".shared_experts.w2", ".shared_experts.down_proj" |
| 396 | ) |
| 397 | if name.endswith(".ffn.gate.bias"): |
| 398 | name = name.replace( |
| 399 | ".ffn.gate.bias", ".ffn.gate.e_score_correction_bias" |
| 400 | ) |
| 401 | if name not in params_dict: |
| 402 | continue |
| 403 | param = params_dict[name] |
| 404 | weight_loader = getattr(param, "weight_loader", default_weight_loader) |
| 405 | weight_loader(param, loaded_weight) |
| 406 | loaded_params.add(name) |
| 407 | |
| 408 | self._finalize_moe() |
| 409 | logger.info_once("DSpark XPU draft model loaded: %d params", len(loaded_params)) |
| 410 | return loaded_params |
| 411 | |
| 412 | def _finalize_moe(self) -> None: |
| 413 | for layer in self.model.layers: |
| 414 | layer.ffn.finalize_mega_moe_weights() |
| 415 | |
| 416 | def _remap_dspark_name(self, name: str) -> str | None: |
| 417 | """Map checkpoint ``mtp.{i}.*`` name to this model's parameter path.""" |
| 418 | m = re.match(r"mtp\.(\d+)\.(.*)", name) |
| 419 | if m is None: |
| 420 | return None |
| 421 | stage = int(m.group(1)) |
| 422 | rest = m.group(2) |
| 423 | if rest.startswith("confidence_head."): |
| 424 | return None |
| 425 | head_prefixes = ( |
| 426 | "norm.", |
| 427 | "hc_head_fn", |
| 428 | "hc_head_base", |
| 429 | "hc_head_scale", |
| 430 | "markov_head.", |
| 431 | ) |
| 432 | if rest.startswith(("main_proj.", "main_norm.")) or rest.startswith( |
| 433 | head_prefixes |
| 434 | ): |
| 435 | return f"model.{rest}" |
| 436 | return f"model.layers.{stage}.{rest}" |