| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import copy |
| 4 | from typing import Any |
| 5 | |
| 6 | import torch |
| 7 | import torch.nn as nn |
| 8 | |
| 9 | from vllm.config import VllmConfig, replace |
| 10 | from vllm.config.compilation import CUDAGraphMode |
| 11 | from vllm.forward_context import BatchDescriptor, set_forward_context |
| 12 | from vllm.logger import init_logger |
| 13 | from vllm.triton_utils import tl, triton |
| 14 | from vllm.v1.attention.backend import AttentionCGSupport |
| 15 | from vllm.v1.attention.backends.utils import PAD_SLOT_ID |
| 16 | from vllm.v1.kv_cache_interface import ( |
| 17 | KVCacheConfig, |
| 18 | get_kv_cache_spec_sliding_window, |
| 19 | ) |
| 20 | from vllm.v1.worker.gpu.attn_utils import build_slot_mappings_by_layer |
| 21 | from vllm.v1.worker.gpu.block_table import BlockTables |
| 22 | from vllm.v1.worker.gpu.cp_utils import cp_local_slot |
| 23 | from vllm.v1.worker.gpu.dp_utils import DPSyncState, dispatch_cg_and_sync_dp |
| 24 | from vllm.v1.worker.gpu.input_batch import InputBatch, InputBuffers |
| 25 | from vllm.v1.worker.gpu.model_states.interface import ModelState |
| 26 | from vllm.v1.worker.gpu.spec_decode.dflash.cudagraph import DFlashCudaGraphManager |
| 27 | from vllm.v1.worker.gpu.spec_decode.dflash.utils import load_dflash_model |
| 28 | from vllm.v1.worker.gpu.spec_decode.speculator import DraftModelSpeculator |
| 29 | from vllm.v1.worker.gpu.spec_decode.utils import get_parallel_drafting_token_id |
| 30 | from vllm.v1.worker.utils import AttentionGroup |
| 31 | |
| 32 | logger = init_logger(__name__) |
| 33 | |
| 34 | |
| 35 | class DFlashSpeculator(DraftModelSpeculator): |
| 36 | _speculator_name = "DFlash" # For logging, so we can share methods with subclasses |
| 37 | |
| 38 | def __init__(self, vllm_config: VllmConfig, device: torch.device): |
| 39 | parallel_config = vllm_config.parallel_config |
| 40 | speculative_config = vllm_config.speculative_config |
| 41 | assert speculative_config is not None |
| 42 | vllm_config = copy.copy(vllm_config) |
| 43 | vllm_config.parallel_config = replace( |
| 44 | parallel_config, |
| 45 | prefill_context_parallel_size=1, |
| 46 | decode_context_parallel_size=( |
| 47 | parallel_config.decode_context_parallel_size |
| 48 | if speculative_config.draft_model_config.use_mla |
| 49 | else 1 |
| 50 | ), |
| 51 | ) |
| 52 | super().__init__(vllm_config, device) |
| 53 | |
| 54 | self.hidden_states = torch.zeros( |
| 55 | self.max_num_tokens, self.hidden_size, dtype=self.dtype, device=device |
| 56 | ) |
| 57 | |
| 58 | # Multimodal inputs not currently supported. |
| 59 | self.supports_mm_inputs = False |
| 60 | |
| 61 | # Each request emits exactly (bonus + N mask) query tokens per step. |
| 62 | self.num_query_per_req = 1 + self.num_speculative_steps |
| 63 | |
| 64 | self.parallel_drafting_token_id = get_parallel_drafting_token_id( |
| 65 | self.draft_model_config.hf_config |
| 66 | ) |
| 67 | |
| 68 | from vllm.model_executor.models.qwen3_dflash import dflash_has_any_non_causal |
| 69 | |
| 70 | self.requires_non_causal = dflash_has_any_non_causal( |
| 71 | self.draft_model_config.hf_config |
| 72 | ) |
| 73 | |
| 74 | # Whether the anchor query position is itself a prediction. DFlash default uses |
| 75 | # the anchor as the bonus token (only mask tokens predict); DSpark samples from |
| 76 | # the anchor and the N-1 mask token positions. See _prepare_dflash_inputs_kernel |
| 77 | dflash_config = ( |
| 78 | getattr(self.draft_model_config.hf_config, "dflash_config", None) or {} |
| 79 | ) |
| 80 | if dflash_config.get("sample_from_anchor", False): |
| 81 | raise ValueError( |
| 82 | "sample_from_anchor=True is not supported for DFlash. " |
| 83 | "DFlash uses a fixed 1+N query layout where the anchor " |
| 84 | "is the bonus token." |
| 85 | ) |
| 86 | self.sample_from_anchor = False |
| 87 | |
| 88 | # Context positions for the K/V precompute, populated by prepare_dflash_inputs. |
| 89 | self.context_positions = torch.zeros( |
| 90 | self.max_num_tokens, dtype=torch.int64, device=device |
| 91 | ) |
| 92 | |
| 93 | # Per-mask-token sampling buffers. Flattened from (num_reqs, num_spec_tokens). |
| 94 | max_num_sampled_tokens = self.max_num_reqs * self.num_speculative_steps |
| 95 | self.sample_indices = torch.zeros( |
| 96 | max_num_sampled_tokens, dtype=torch.int64, device=device |
| 97 | ) |
| 98 | self.sample_pos = torch.zeros( |
| 99 | max_num_sampled_tokens, dtype=torch.int64, device=device |
| 100 | ) |
| 101 | # -1 marks an inert sampling row. CUDA graph capture can execute the |
| 102 | # full buffer before a real batch has populated it, so zero would make |
| 103 | # every padding row scatter into request slot 0. |
| 104 | self.sample_idx_mapping = torch.full( |
| 105 | (max_num_sampled_tokens,), -1, dtype=torch.int32, device=device |
| 106 | ) |
| 107 | # [0, 1, ..., N-1, 0, 1, ..., N-1, ...] -> the per-token column index into |
| 108 | # draft_logits[req, step, :]. |
| 109 | self.sample_col = torch.arange( |
| 110 | self.num_speculative_steps, dtype=torch.int32, device=device |
| 111 | ).repeat(self.max_num_reqs) |
| 112 | |
| 113 | self.query_cudagraph_manager: DFlashCudaGraphManager | None = None |
| 114 | self.draft_kv_cache_group_id: int = -1 |
| 115 | |
| 116 | @property |
| 117 | def attn_vllm_config(self) -> VllmConfig: |
| 118 | # The draft's attention differs from the target's in causality. |
| 119 | config = copy.copy(super().attn_vllm_config) |
| 120 | config.attention_config = replace( |
| 121 | self.vllm_config.attention_config, |
| 122 | use_non_causal=self.requires_non_causal, |
| 123 | ) |
| 124 | return config |
| 125 | |
| 126 | def init_cudagraph_manager(self, cudagraph_mode: CUDAGraphMode) -> None: |
| 127 | wants_full = cudagraph_mode.decode_mode() == CUDAGraphMode.FULL |
| 128 | supports_full = ( |
| 129 | self.attn_cg_support.min_cg_support.value |
| 130 | >= AttentionCGSupport.UNIFORM_BATCH.value |
| 131 | ) |
| 132 | if wants_full and not supports_full: |
| 133 | logger.warning( |
| 134 | "%s draft attention (%s) does not support full CUDA graphs; " |
| 135 | "running the draft eagerly.", |
| 136 | self._speculator_name, |
| 137 | self.attn_cg_support.min_cg_attn_backend, |
| 138 | ) |
| 139 | # PIECEWISE cudagraphs are not supported for dflash. |
| 140 | if wants_full and supports_full: |
| 141 | cudagraph_mode = CUDAGraphMode.FULL_DECODE_ONLY |
| 142 | else: |
| 143 | cudagraph_mode = CUDAGraphMode.NONE |
| 144 | |
| 145 | self.query_cudagraph_manager = DFlashCudaGraphManager( |
| 146 | self.vllm_config, |
| 147 | self.device, |
| 148 | cudagraph_mode, |
| 149 | decode_query_len=self.num_query_per_req, |
| 150 | ) |
| 151 | |
| 152 | def capture(self) -> None: |
| 153 | logger.info("Capturing model for %s speculator...", self._speculator_name) |
| 154 | # Padded sample rows must not scatter into a live request during capture. |
| 155 | self.sample_indices.zero_() |
| 156 | self.sample_pos.zero_() |
| 157 | self.sample_idx_mapping.fill_(-1) |
| 158 | # Capture must not write context K/V. |
| 159 | self._context_slot_mappings.fill_(PAD_SLOT_ID) |
| 160 | assert self.query_cudagraph_manager is not None |
| 161 | self.query_cudagraph_manager.capture( |
| 162 | self._generate_draft, |
| 163 | self.input_buffers, |
| 164 | self.block_tables, |
| 165 | self.attn_groups, |
| 166 | self.kv_cache_config, |
| 167 | self.max_model_len, |
| 168 | causal=self._group_causal, |
| 169 | precompute_context_kv=lambda num_reqs: self._precompute_context_kv( |
| 170 | 0, self._num_graph_context_tokens(num_reqs) |
| 171 | ), |
| 172 | progress_bar_desc=f"Capturing {self._speculator_name.lower()} CUDA graphs", |
| 173 | ) |
| 174 | |
| 175 | def load_draft_model( |
| 176 | self, |
| 177 | target_model: nn.Module, |
| 178 | target_attn_layer_names: set[str], |
| 179 | ) -> nn.Module: |
| 180 | return load_dflash_model(target_model, self.vllm_config) |
| 181 | |
| 182 | def set_attn( |
| 183 | self, |
| 184 | model_state: ModelState, |
| 185 | kv_cache_config: KVCacheConfig, |
| 186 | block_tables: BlockTables, |
| 187 | target_input_buffers: InputBuffers, |
| 188 | target_attn_groups: list[list[AttentionGroup]], |
| 189 | ) -> None: |
| 190 | super().set_attn( |
| 191 | model_state, |
| 192 | kv_cache_config, |
| 193 | block_tables, |
| 194 | target_input_buffers, |
| 195 | target_attn_groups, |
| 196 | ) |
| 197 | |
| 198 | # FlashAttention's AOT split schedule is wrong for a windowed drafter, |
| 199 | # and `_get_sliding_window_configs` leaves it on or off depending on |
| 200 | # whether the target also runs FlashAttention. Decide it here instead. |
| 201 | for groups in self.attn_groups: |
| 202 | for group in groups: |
| 203 | builder = group.get_metadata_builder() |
| 204 | if getattr( |
| 205 | builder, "aot_schedule", False |
| 206 | ) and get_kv_cache_spec_sliding_window(builder.kv_cache_spec): |
| 207 | # `aot_schedule` belongs to FlashAttention's builder, not |
| 208 | # to the base class this loop is typed against. |
| 209 | builder.aot_schedule = False # type: ignore[attr-defined] |
| 210 | |
| 211 | self.draft_kv_cache_group_ids = [ |
| 212 | gid for gid, g in enumerate(self.attn_groups) if g |
| 213 | ] |
| 214 | assert self.draft_kv_cache_group_ids, "No draft attention groups found." |
| 215 | self.draft_kv_cache_group_id = self.draft_kv_cache_group_ids[0] |
| 216 | |
| 217 | # Per-group context slot buffers for the precompute (one row per group). |
| 218 | self._context_slot_mappings = torch.zeros( |
| 219 | len(self.draft_kv_cache_group_ids), |
| 220 | self.max_num_tokens, |
| 221 | dtype=torch.int64, |
| 222 | device=self.device, |
| 223 | ) |
| 224 | |
| 225 | # Map each draft decoder layer to the index (within draft_kv_cache_group_ids) |
| 226 | # of the kv-cache group its cache belongs to. Models that share a single group |
| 227 | # leave this as None and share one context slot mapping. |
| 228 | self._layer_group_idx: list[int] | None = None |
| 229 | # Per-KV-group causal, falling back to whether the drafter is all-causal. |
| 230 | self._group_causal: dict[int, bool] | bool = not self.requires_non_causal |
| 231 | if hasattr(self.model, "get_draft_kv_cache_layer_names"): |
| 232 | layer_names = self.model.get_draft_kv_cache_layer_names() |
| 233 | name_to_gid = { |
| 234 | ln: gid |
| 235 | for gid, group in enumerate(kv_cache_config.kv_cache_groups) |
| 236 | for ln in group.layer_names |
| 237 | } |
| 238 | gid_to_idx = {gid: i for i, gid in enumerate(self.draft_kv_cache_group_ids)} |
| 239 | self._layer_group_idx = [ |
| 240 | gid_to_idx[name_to_gid[name]] for name in layer_names |
| 241 | ] |
| 242 | if hasattr(self.model, "get_draft_attn_causal"): |
| 243 | self._group_causal = { |
| 244 | name_to_gid[name]: layer_causal |
| 245 | for name, layer_causal in zip( |
| 246 | layer_names, self.model.get_draft_attn_causal() |
| 247 | ) |
| 248 | } |
| 249 | |
| 250 | @torch.inference_mode() |
| 251 | def _run_model( |
| 252 | self, |
| 253 | num_tokens: int, |
| 254 | attn_metadata: dict[str, Any] | None, |
| 255 | slot_mappings: dict[str, torch.Tensor] | None, |
| 256 | num_tokens_across_dp: torch.Tensor | None, |
| 257 | cudagraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE, |
| 258 | ) -> torch.Tensor: |
| 259 | batch_descriptor = BatchDescriptor(num_tokens=num_tokens) |
| 260 | with set_forward_context( |
| 261 | attn_metadata, |
| 262 | self.vllm_config, |
| 263 | num_tokens=num_tokens, |
| 264 | cudagraph_runtime_mode=cudagraph_runtime_mode, |
| 265 | num_tokens_across_dp=num_tokens_across_dp, |
| 266 | slot_mapping=slot_mappings, |
| 267 | batch_descriptor=batch_descriptor, |
| 268 | ): |
| 269 | last_hidden_states = self.model( |
| 270 | input_ids=self.input_buffers.input_ids[:num_tokens], |
| 271 | positions=self.input_buffers.positions[:num_tokens], |
| 272 | inputs_embeds=None, |
| 273 | ) |
| 274 | return last_hidden_states |
| 275 | |
| 276 | def _generate_draft( |
| 277 | self, |
| 278 | num_reqs: int, |
| 279 | num_tokens_padded: int, |
| 280 | attn_metadata: dict[str, Any] | None, |
| 281 | slot_mappings: dict[str, torch.Tensor] | None, |
| 282 | num_tokens_across_dp: torch.Tensor | None, |
| 283 | cudagraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE, |
| 284 | ) -> None: |
| 285 | last_hidden_states = self._run_model( |
| 286 | num_tokens_padded, |
| 287 | attn_metadata, |
| 288 | slot_mappings, |
| 289 | num_tokens_across_dp, |
| 290 | cudagraph_runtime_mode, |
| 291 | ) |
| 292 | num_sample = num_reqs * self.num_speculative_steps |
| 293 | sample_hidden_states = last_hidden_states[self.sample_indices[:num_sample]] |
| 294 | # sample_pos is the predicted token's position P. Sampling keys a draw |
| 295 | # by the position before the sampled token, P-1. |
| 296 | draft_tokens = self.sample_draft( |
| 297 | sample_hidden_states, |
| 298 | self.sample_pos[:num_sample] - 1, |
| 299 | self.sample_idx_mapping[:num_sample], |
| 300 | self.temperature, |
| 301 | self.seeds, |
| 302 | self.sample_col[:num_sample], |
| 303 | self.draft_logits, |
| 304 | ) |
| 305 | self.draft_tokens[:num_reqs] = draft_tokens.view( |
| 306 | num_reqs, self.num_speculative_steps |
| 307 | ) |
| 308 | |
| 309 | def _num_graph_context_tokens(self, num_reqs: int) -> int: |
| 310 | # Context rows a captured draft step stores: one full verify per request. |
| 311 | return min(num_reqs * (self.num_speculative_steps + 1), self.max_num_tokens) |
| 312 | |
| 313 | def _precompute_context_kv( |
| 314 | self, start: int, end: int, dummy_run: bool = False |
| 315 | ) -> None: |
| 316 | if dummy_run: |
| 317 | context_slots: torch.Tensor | list[torch.Tensor | None] | None = None |
| 318 | elif self._layer_group_idx is not None: |
| 319 | context_slots = [ |
| 320 | self._context_slot_mappings[gidx][start:end] |
| 321 | for gidx in self._layer_group_idx |
| 322 | ] |
| 323 | else: |
| 324 | context_slots = self._context_slot_mappings[0][start:end] |
| 325 | self.model.precompute_and_store_context_kv( |
| 326 | self.hidden_states[start:end], |
| 327 | self.context_positions[start:end], |
| 328 | context_slots, |
| 329 | ) |
| 330 | |
| 331 | def prepare_context_anchor( |
| 332 | self, input_batch: InputBatch, num_rejected: torch.Tensor |
| 333 | ) -> None: |
| 334 | """Publish context features required by a draft's candidate head.""" |
| 335 | |
| 336 | @torch.inference_mode() |
| 337 | def propose( |
| 338 | self, |
| 339 | input_batch: InputBatch, |
| 340 | attn_metadata: dict[str, Any], |
| 341 | slot_mappings: dict[str, torch.Tensor], |
| 342 | # [num_tokens, hidden_size] |
| 343 | last_hidden_states: torch.Tensor, |
| 344 | # num_layers x [num_tokens, hidden_size] |
| 345 | aux_hidden_states: list[torch.Tensor] | None, |
| 346 | # [num_reqs] |
| 347 | num_sampled: torch.Tensor, |
| 348 | # [num_reqs] |
| 349 | num_rejected: torch.Tensor, |
| 350 | # [max_num_reqs] |
| 351 | last_sampled: torch.Tensor, |
| 352 | # [max_num_reqs] |
| 353 | next_prefill_tokens: torch.Tensor, |
| 354 | # [max_num_reqs] |
| 355 | temperature: torch.Tensor, |
| 356 | # [max_num_reqs] |
| 357 | seeds: torch.Tensor, |
| 358 | dp_sync: DPSyncState | None = None, |
| 359 | dummy_run: bool = False, |
| 360 | skip_attn_for_dummy_run: bool = False, |
| 361 | mm_inputs: tuple[list[torch.Tensor], torch.Tensor] | None = None, |
| 362 | is_profile: bool = False, |
| 363 | ) -> torch.Tensor: |
| 364 | num_reqs = input_batch.num_reqs |
| 365 | num_target_tokens = input_batch.num_tokens |
| 366 | num_query_tokens = num_reqs * self.num_query_per_req |
| 367 | max_seq_len = input_batch.seq_lens_cpu_upper_bound[:num_reqs].max().item() |
| 368 | self.draft_max_seq_len = min( |
| 369 | max_seq_len + self.num_query_per_req, self.max_model_len |
| 370 | ) |
| 371 | |
| 372 | # NOTE: To avoid CPU-GPU synchronization without CPU knowing the |
| 373 | # number of rejected tokens, we maintain the size of input_ids and |
| 374 | # hidden_states the same as the target model's. This means, we pad each |
| 375 | # request's query length to include any rejected positions. |
| 376 | if aux_hidden_states: |
| 377 | hidden_states = self.model.combine_hidden_states( |
| 378 | torch.cat(aux_hidden_states, dim=-1) |
| 379 | ) |
| 380 | else: |
| 381 | hidden_states = last_hidden_states |
| 382 | self.hidden_states[:num_target_tokens].copy_(hidden_states[:num_target_tokens]) |
| 383 | self.prepare_context_anchor(input_batch, num_rejected) |
| 384 | |
| 385 | if dummy_run and skip_attn_for_dummy_run: |
| 386 | # Memory profiling path: block_tables / kv_cache_config are not initialized. |
| 387 | # Since DFlash needs to build its own attention metadata, we must skip the |
| 388 | # preparation in this path and run a minimal forward pass. |
| 389 | self.model.precompute_and_store_context_kv( |
| 390 | self.hidden_states[:num_target_tokens], |
| 391 | self.context_positions[:num_target_tokens], |
| 392 | ) |
| 393 | # DFlash processes all speculative tokens in one forward pass, |
| 394 | # so the real token count is num_query_tokens. |
| 395 | self._prepare_eplb_forward(num_query_tokens) |
| 396 | self._generate_draft( |
| 397 | num_reqs, |
| 398 | num_query_tokens, |
| 399 | attn_metadata=None, |
| 400 | slot_mappings=None, |
| 401 | num_tokens_across_dp=None, |
| 402 | cudagraph_runtime_mode=CUDAGraphMode.NONE, |
| 403 | ) |
| 404 | return self.draft_tokens[:num_reqs] |
| 405 | |
| 406 | if self.pcp_manager is not None and not dummy_run: |
| 407 | self.block_tables.gather_block_tables( |
| 408 | input_batch.idx_mapping, num_reqs_padded=num_reqs |
| 409 | ) |
| 410 | |
| 411 | # The query slot mapping is written into the shared BlockTables slot_mappings. |
| 412 | # That buffer's address is what the captured CUDA graph reads from at replay. |
| 413 | assert self.draft_kv_cache_group_id >= 0 |
| 414 | # Support multiple draft KV cache groups by preparing inputs once for each |
| 415 | for i, gid in enumerate(self.draft_kv_cache_group_ids): |
| 416 | prepare_dflash_inputs( |
| 417 | self.input_buffers, |
| 418 | self.block_tables.slot_mappings[gid], |
| 419 | self.context_positions, |
| 420 | self._context_slot_mappings[i], |
| 421 | self.sample_indices, |
| 422 | self.sample_pos, |
| 423 | self.sample_idx_mapping, |
| 424 | self.temperature, |
| 425 | self.seeds, |
| 426 | input_batch, |
| 427 | num_sampled, |
| 428 | num_rejected, |
| 429 | last_sampled, |
| 430 | next_prefill_tokens, |
| 431 | temperature, |
| 432 | seeds, |
| 433 | self.block_tables.input_block_tables[gid], |
| 434 | self.block_tables.kernel_block_sizes[gid], |
| 435 | self.block_tables.cp_rank, |
| 436 | self.dcp_size, |
| 437 | self.block_tables.cp_interleave, |
| 438 | self.parallel_drafting_token_id, |
| 439 | self.num_query_per_req, |
| 440 | self.num_speculative_steps, |
| 441 | self.max_num_reqs, |
| 442 | self.max_num_tokens, |
| 443 | self.max_model_len, |
| 444 | self.sample_from_anchor, |
| 445 | ) |
| 446 | |
| 447 | batch_sync, num_batch_tokens = ( |
| 448 | self._build_uniform_batch_dp_sync(dp_sync, num_reqs, self.num_query_per_req) |
| 449 | if dp_sync is not None |
| 450 | else (None, num_query_tokens) |
| 451 | ) |
| 452 | # Every DFlash step has exactly num_query_per_req tokens, so we can use FULL CGs |
| 453 | batch_desc, batch_sync = dispatch_cg_and_sync_dp( |
| 454 | self.query_cudagraph_manager, |
| 455 | num_reqs, |
| 456 | num_batch_tokens, |
| 457 | uniform_token_count=self.num_query_per_req, |
| 458 | dp_size=self.dp_size, |
| 459 | dp_rank=self.dp_rank, |
| 460 | need_eager=is_profile, |
| 461 | dp_sync=batch_sync, |
| 462 | ) |
| 463 | num_tokens_padded = batch_desc.num_tokens |
| 464 | num_tokens_across_dp = ( |
| 465 | batch_sync.num_tokens_across_dp if batch_sync is not None else None |
| 466 | ) |
| 467 | |
| 468 | if batch_desc.cg_mode == CUDAGraphMode.FULL: |
| 469 | # The graph stores the first num_context context rows. |
| 470 | assert batch_desc.num_reqs is not None |
| 471 | num_context = self._num_graph_context_tokens(batch_desc.num_reqs) |
| 472 | if dummy_run: |
| 473 | # Dummy block tables are placeholders: write no context K/V. |
| 474 | self._context_slot_mappings[:, :num_context].fill_(PAD_SLOT_ID) |
| 475 | elif num_target_tokens <= num_context: |
| 476 | # Rows past the batch keep stale positions but write no K/V. |
| 477 | self._context_slot_mappings[:, num_target_tokens:num_context].fill_( |
| 478 | PAD_SLOT_ID |
| 479 | ) |
| 480 | else: |
| 481 | # Prefill context beyond the graph's rows is stored before replay. |
| 482 | self._precompute_context_kv(num_context, num_target_tokens) |
| 483 | else: |
| 484 | self._precompute_context_kv(0, num_target_tokens, dummy_run) |
| 485 | |
| 486 | # Rebuild the draft attention metadata even when replaying the FULL |
| 487 | # graph so that any attention metadata builder state is updated. |
| 488 | draft_attn_metadata = self._build_uniform_attn_metadata( |
| 489 | num_reqs=num_reqs, |
| 490 | batch_desc=batch_desc, |
| 491 | num_query_per_req=self.num_query_per_req, |
| 492 | seq_lens_cpu_upper_bound=input_batch.seq_lens_cpu_upper_bound, |
| 493 | step=self.num_query_per_req, |
| 494 | causal=self._group_causal, |
| 495 | ) |
| 496 | draft_slot_mappings_by_layer = build_slot_mappings_by_layer( |
| 497 | self.block_tables.slot_mappings[:, :num_tokens_padded], |
| 498 | self.kv_cache_config, |
| 499 | ) |
| 500 | |
| 501 | # DFlash processes all speculative tokens in one forward pass, |
| 502 | # so the real token count is num_query_tokens. |
| 503 | self._prepare_eplb_forward(num_query_tokens) |
| 504 | |
| 505 | if batch_desc.cg_mode == CUDAGraphMode.FULL: |
| 506 | assert self.query_cudagraph_manager is not None |
| 507 | self.query_cudagraph_manager.run_fullgraph(batch_desc) |
| 508 | else: |
| 509 | self._generate_draft( |
| 510 | num_reqs, |
| 511 | num_tokens_padded, |
| 512 | draft_attn_metadata, |
| 513 | draft_slot_mappings_by_layer, |
| 514 | num_tokens_across_dp=num_tokens_across_dp, |
| 515 | cudagraph_runtime_mode=batch_desc.cg_mode, |
| 516 | ) |
| 517 | |
| 518 | return self.draft_tokens[:num_reqs] |
| 519 | |
| 520 | |
| 521 | @triton.jit |
| 522 | def _prepare_dflash_inputs_kernel( |
| 523 | # Outputs |
| 524 | out_input_ids_ptr, |
| 525 | out_query_positions_ptr, |
| 526 | out_query_start_loc_ptr, |
| 527 | out_seq_lens_ptr, |
| 528 | out_query_slot_mapping_ptr, |
| 529 | out_context_positions_ptr, |
| 530 | out_context_slot_mapping_ptr, |
| 531 | out_sample_indices_ptr, |
| 532 | out_sample_pos_ptr, |
| 533 | out_sample_idx_mapping_ptr, |
| 534 | out_temperature_ptr, |
| 535 | out_seeds_ptr, |
| 536 | # Inputs from target batch |
| 537 | target_positions_ptr, |
| 538 | target_query_start_loc_ptr, |
| 539 | idx_mapping_ptr, |
| 540 | last_sampled_ptr, |
| 541 | next_prefill_tokens_ptr, |
| 542 | num_sampled_ptr, |
| 543 | num_rejected_ptr, |
| 544 | # Sampling params |
| 545 | temperature_ptr, |
| 546 | seeds_ptr, |
| 547 | # Block table for slot mapping lookup. |
| 548 | block_table_ptr, |
| 549 | block_table_stride, |
| 550 | # Scalars |
| 551 | parallel_drafting_token_id, |
| 552 | block_size, |
| 553 | num_query_per_req, |
| 554 | num_speculative_steps, |
| 555 | max_num_reqs, |
| 556 | max_num_tokens, |
| 557 | max_model_len, |
| 558 | cp_rank, |
| 559 | SAMPLE_FROM_ANCHOR: tl.constexpr, |
| 560 | PAD_SLOT_ID: tl.constexpr, |
| 561 | CP_SIZE: tl.constexpr, |
| 562 | CP_INTERLEAVE: tl.constexpr, |
| 563 | BLOCK_SIZE: tl.constexpr, |
| 564 | ): |
| 565 | req_idx = tl.program_id(0) |
| 566 | block_idx = tl.program_id(1) |
| 567 | num_reqs = tl.num_programs(0) |
| 568 | req_state_idx = tl.load(idx_mapping_ptr + req_idx) |
| 569 | |
| 570 | ctx_start = tl.load(target_query_start_loc_ptr + req_idx) |
| 571 | ctx_end = tl.load(target_query_start_loc_ptr + req_idx + 1) |
| 572 | num_ctx = ctx_end - ctx_start |
| 573 | |
| 574 | num_rejected = tl.load(num_rejected_ptr + req_idx) |
| 575 | valid_ctx_end = ctx_end - num_rejected |
| 576 | num_valid_ctx = valid_ctx_end - ctx_start |
| 577 | |
| 578 | num_sampled = tl.load(num_sampled_ptr + req_idx) |
| 579 | if num_sampled > 0: |
| 580 | bonus_token = tl.load(last_sampled_ptr + req_state_idx).to(tl.int32) |
| 581 | else: |
| 582 | # Chunked prefilling: splice in the next prefill token. |
| 583 | bonus_token = tl.load(next_prefill_tokens_ptr + req_state_idx).to(tl.int32) |
| 584 | |
| 585 | last_valid_pos = tl.load(target_positions_ptr + valid_ctx_end - 1) |
| 586 | query_base = req_idx * num_query_per_req |
| 587 | |
| 588 | j = block_idx * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) |
| 589 | is_ctx = j < num_ctx |
| 590 | is_valid_ctx = j < num_valid_ctx |
| 591 | is_query = (j >= num_valid_ctx) & (j < num_valid_ctx + num_query_per_req) |
| 592 | query_off = j - num_valid_ctx |
| 593 | |
| 594 | # --- Context positions / slots --- |
| 595 | ctx_pos_idx = ctx_start + tl.where(is_ctx, j, 0) |
| 596 | ctx_pos = tl.load(target_positions_ptr + ctx_pos_idx, mask=is_valid_ctx, other=0) |
| 597 | ctx_block_num = ctx_pos // (block_size * CP_SIZE) |
| 598 | ctx_block_num = tl.minimum(ctx_block_num, block_table_stride - 1) |
| 599 | ctx_block_id = tl.load( |
| 600 | block_table_ptr + req_idx * block_table_stride + ctx_block_num, |
| 601 | mask=is_valid_ctx, |
| 602 | other=0, |
| 603 | ).to(tl.int64) |
| 604 | # Block 0 is the null block. Old sliding-window context positions can map |
| 605 | # to it after eviction; rejected suffix rows are invalid context as well. |
| 606 | # Neither kind of row may write draft KV into physical block 0. |
| 607 | ctx_resident = is_valid_ctx & (ctx_block_id != 0) |
| 608 | local_ctx_slot = cp_local_slot( |
| 609 | ctx_pos, ctx_block_id, block_size, cp_rank, CP_SIZE, CP_INTERLEAVE, PAD_SLOT_ID |
| 610 | ) |
| 611 | ctx_slot = tl.where( |
| 612 | ctx_resident, |
| 613 | local_ctx_slot, |
| 614 | PAD_SLOT_ID, |
| 615 | ) |
| 616 | # Stored over the full [0, num_ctx) span while the loads above are masked to |
| 617 | # [0, num_valid_ctx): the rejected suffix rows in between get position 0 and |
| 618 | # PAD_SLOT_ID. That is intentional — those rows write no KV and their |
| 619 | # positions are never consumed, but the span must stay fully initialized so |
| 620 | # a replayed graph cannot observe a stale value from an earlier batch. |
| 621 | tl.store(out_context_positions_ptr + ctx_start + j, ctx_pos, mask=is_ctx) |
| 622 | tl.store(out_context_slot_mapping_ptr + ctx_start + j, ctx_slot, mask=is_ctx) |
| 623 | |
| 624 | # --- Query positions / input_ids / slots --- |
| 625 | query_pos = last_valid_pos + 1 + query_off |
| 626 | query_idx = query_base + query_off |
| 627 | is_bonus = is_query & (query_off == 0) |
| 628 | input_id = tl.where(is_bonus, bonus_token, parallel_drafting_token_id) |
| 629 | |
| 630 | q_block_num = query_pos // (block_size * CP_SIZE) |
| 631 | q_block_num = tl.minimum(q_block_num, block_table_stride - 1) |
| 632 | q_block_id = tl.load( |
| 633 | block_table_ptr + req_idx * block_table_stride + q_block_num, |
| 634 | mask=is_query, |
| 635 | other=0, |
| 636 | ).to(tl.int64) |
| 637 | # A null block is never a writable cache slot. This can occur when a |
| 638 | # sliding-window block table contains evicted/global padding entries. |
| 639 | q_resident = is_query & (q_block_id != 0) |
| 640 | local_q_slot = cp_local_slot( |
| 641 | query_pos, |
| 642 | q_block_id, |
| 643 | block_size, |
| 644 | cp_rank, |
| 645 | CP_SIZE, |
| 646 | CP_INTERLEAVE, |
| 647 | PAD_SLOT_ID, |
| 648 | ) |
| 649 | q_slot = tl.where( |
| 650 | q_resident, |
| 651 | local_q_slot, |
| 652 | PAD_SLOT_ID, |
| 653 | ) |
| 654 | |
| 655 | tl.store(out_input_ids_ptr + query_idx, input_id, mask=is_query) |
| 656 | clamped_query_pos = tl.minimum(query_pos, max_model_len - 1) |
| 657 | tl.store(out_query_positions_ptr + query_idx, clamped_query_pos, mask=is_query) |
| 658 | tl.store(out_query_slot_mapping_ptr + query_idx, q_slot, mask=is_query) |
| 659 | |
| 660 | # --- Sample indices / positions / idx_mapping --- |
| 661 | # When SAMPLE_FROM_ANCHOR (DSpark), so we sample at EVERY query position |
| 662 | # and each position k predicts the NEXT token (sampled position = query_pos + 1). |
| 663 | # Otherwise (DFlash default) the anchor is the bonus token and only the mask tokens |
| 664 | # at offsets > 0 are sampled from, each AT its own position. |
| 665 | sample_off = 0 if SAMPLE_FROM_ANCHOR else 1 |
| 666 | is_sample = is_query & (query_off >= sample_off) |
| 667 | sample_idx = req_idx * num_speculative_steps + (query_off - sample_off) |
| 668 | sample_pos = query_pos + 1 if SAMPLE_FROM_ANCHOR else query_pos |
| 669 | tl.store(out_sample_indices_ptr + sample_idx, query_idx, mask=is_sample) |
| 670 | tl.store(out_sample_pos_ptr + sample_idx, sample_pos, mask=is_sample) |
| 671 | tl.store(out_sample_idx_mapping_ptr + sample_idx, req_state_idx, mask=is_sample) |
| 672 | |
| 673 | if block_idx == 0: |
| 674 | tl.store(out_query_start_loc_ptr + req_idx, query_base) |
| 675 | # seq_lens is the absolute sequence length the draft attention |
| 676 | # reads up to (context + query), not just the count of accepted |
| 677 | # tokens this step. |
| 678 | tl.store( |
| 679 | out_seq_lens_ptr + req_idx, |
| 680 | tl.minimum(last_valid_pos + 1 + num_query_per_req, max_model_len), |
| 681 | ) |
| 682 | # Copy sampling state. |
| 683 | tl.store( |
| 684 | out_temperature_ptr + req_state_idx, |
| 685 | tl.load(temperature_ptr + req_state_idx), |
| 686 | ) |
| 687 | tl.store(out_seeds_ptr + req_state_idx, tl.load(seeds_ptr + req_state_idx)) |
| 688 | if req_idx == num_reqs - 1: |
| 689 | # Pad per-request buffers to max_num_reqs for CUDA graph safety. |
| 690 | last_query_end = num_reqs * num_query_per_req |
| 691 | for i in range(num_reqs, max_num_reqs + 1, BLOCK_SIZE): |
| 692 | block = i + tl.arange(0, BLOCK_SIZE) |
| 693 | mask = block < max_num_reqs + 1 |
| 694 | tl.store(out_query_start_loc_ptr + block, last_query_end, mask=mask) |
| 695 | for i in range(num_reqs, max_num_reqs, BLOCK_SIZE): |
| 696 | block = i + tl.arange(0, BLOCK_SIZE) |
| 697 | mask = block < max_num_reqs |
| 698 | tl.store(out_seq_lens_ptr + block, 0, mask=mask) |
| 699 | # Padded sample slots point at query index 0 (a valid row in |
| 700 | # last_hidden_states) so CG replay never reads OOB. Padded |
| 701 | # sample idx mappings point to -1, which is ignored during |
| 702 | # sampling to prevent writing stale values to draft logits. |
| 703 | pad_start = num_reqs * num_speculative_steps |
| 704 | pad_end = max_num_reqs * num_speculative_steps |
| 705 | for i in range(pad_start, pad_end, BLOCK_SIZE): |
| 706 | block = i + tl.arange(0, BLOCK_SIZE) |
| 707 | mask = block < pad_end |
| 708 | tl.store(out_sample_indices_ptr + block, 0, mask=mask) |
| 709 | tl.store(out_sample_pos_ptr + block, 0, mask=mask) |
| 710 | tl.store(out_sample_idx_mapping_ptr + block, -1, mask=mask) |
| 711 | # Pad query slot mappings past num_query_tokens with PAD so the |
| 712 | # captured CG sees PAD slots (no K/V write) for replay sizes |
| 713 | # larger than the current request count. |
| 714 | q_pad_start = num_reqs * num_query_per_req |
| 715 | for i in range(q_pad_start, max_num_tokens, BLOCK_SIZE): |
| 716 | block = i + tl.arange(0, BLOCK_SIZE) |
| 717 | mask = block < max_num_tokens |
| 718 | tl.store(out_query_slot_mapping_ptr + block, PAD_SLOT_ID, mask=mask) |
| 719 | |
| 720 | |
| 721 | def prepare_dflash_inputs( |
| 722 | input_buffers: InputBuffers, |
| 723 | query_slot_mapping: torch.Tensor, |
| 724 | context_positions: torch.Tensor, |
| 725 | context_slot_mapping: torch.Tensor, |
| 726 | sample_indices: torch.Tensor, |
| 727 | sample_pos: torch.Tensor, |
| 728 | sample_idx_mapping: torch.Tensor, |
| 729 | temperature: torch.Tensor, |
| 730 | seeds: torch.Tensor, |
| 731 | input_batch: InputBatch, |
| 732 | # [num_reqs] |
| 733 | num_sampled: torch.Tensor, |
| 734 | # [num_reqs] |
| 735 | num_rejected: torch.Tensor, |
| 736 | # [max_num_reqs] |
| 737 | last_sampled: torch.Tensor, |
| 738 | # [max_num_reqs] |
| 739 | next_prefill_tokens: torch.Tensor, |
| 740 | # [max_num_reqs] |
| 741 | input_temperature: torch.Tensor, |
| 742 | # [max_num_reqs] |
| 743 | input_seeds: torch.Tensor, |
| 744 | # [max_num_reqs, max_num_blocks] |
| 745 | block_table: torch.Tensor, |
| 746 | block_size: int, |
| 747 | cp_rank: int, |
| 748 | cp_size: int, |
| 749 | cp_interleave: int, |
| 750 | parallel_drafting_token_id: int, |
| 751 | num_query_per_req: int, |
| 752 | num_speculative_steps: int, |
| 753 | max_num_reqs: int, |
| 754 | max_num_tokens: int, |
| 755 | max_model_len: int, |
| 756 | sample_from_anchor: bool = False, |
| 757 | ) -> None: |
| 758 | num_reqs = input_batch.num_reqs |
| 759 | assert num_reqs > 0 |
| 760 | # Cover the longest possible per-request span (ctx + query). Use the max |
| 761 | # per-request query length, not the total token count across the batch. |
| 762 | max_target_query_len = int(input_batch.num_scheduled_tokens.max()) |
| 763 | max_tokens_per_req = max_target_query_len + num_query_per_req |
| 764 | BLOCK_SIZE = min(256, triton.next_power_of_2(max(1, max_tokens_per_req))) |
| 765 | num_blocks = triton.cdiv(max_tokens_per_req, BLOCK_SIZE) |
| 766 | _prepare_dflash_inputs_kernel[(num_reqs, num_blocks)]( |
| 767 | input_buffers.input_ids, |
| 768 | input_buffers.positions, |
| 769 | input_buffers.query_start_loc, |
| 770 | input_buffers.seq_lens, |
| 771 | query_slot_mapping, |
| 772 | context_positions, |
| 773 | context_slot_mapping, |
| 774 | sample_indices, |
| 775 | sample_pos, |
| 776 | sample_idx_mapping, |
| 777 | temperature, |
| 778 | seeds, |
| 779 | input_batch.positions, |
| 780 | input_batch.query_start_loc, |
| 781 | input_batch.idx_mapping, |
| 782 | last_sampled, |
| 783 | next_prefill_tokens, |
| 784 | num_sampled, |
| 785 | num_rejected, |
| 786 | input_temperature, |
| 787 | input_seeds, |
| 788 | block_table, |
| 789 | block_table.stride(0), |
| 790 | parallel_drafting_token_id, |
| 791 | block_size, |
| 792 | num_query_per_req, |
| 793 | num_speculative_steps, |
| 794 | max_num_reqs, |
| 795 | max_num_tokens, |
| 796 | max_model_len, |
| 797 | cp_rank, |
| 798 | SAMPLE_FROM_ANCHOR=sample_from_anchor, |
| 799 | PAD_SLOT_ID=PAD_SLOT_ID, |
| 800 | CP_SIZE=cp_size, |
| 801 | CP_INTERLEAVE=cp_interleave, |
| 802 | BLOCK_SIZE=BLOCK_SIZE, |
| 803 | ) |