| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import dataclasses |
| 4 | from importlib.util import find_spec |
| 5 | from typing import TYPE_CHECKING, Any, cast |
| 6 | |
| 7 | import numpy as np |
| 8 | import torch |
| 9 | import torch.nn as nn |
| 10 | |
| 11 | from vllm.compilation.breakable_cudagraph import BreakableCUDAGraphWrapper |
| 12 | from vllm.config import ( |
| 13 | CUDAGraphMode, |
| 14 | VllmConfig, |
| 15 | get_layers_from_vllm_config, |
| 16 | replace, |
| 17 | ) |
| 18 | |
| 19 | if TYPE_CHECKING: |
| 20 | from vllm.v1.spec_decode.vocab_mapping import VocabMapping |
| 21 | |
| 22 | from vllm.distributed.eplb.eplb_state import EplbState |
| 23 | from vllm.distributed.parallel_state import get_pp_group |
| 24 | from vllm.forward_context import set_forward_context |
| 25 | from vllm.logger import init_logger |
| 26 | from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase |
| 27 | from vllm.model_executor.model_loader import get_model |
| 28 | from vllm.model_executor.model_loader.utils import get_draft_load_config |
| 29 | from vllm.model_executor.models import ( |
| 30 | supports_multimodal, |
| 31 | supports_multimodal_embeddings, |
| 32 | ) |
| 33 | from vllm.model_executor.models.deepseek_eagle3 import Eagle3DeepseekV2ForCausalLM |
| 34 | from vllm.model_executor.models.interfaces import SupportsMultiModal |
| 35 | from vllm.model_executor.models.laguna_dflash import DFlashLagunaForCausalLM |
| 36 | from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM |
| 37 | from vllm.model_executor.models.qwen3_dflash import DFlashQwen3ForCausalLM |
| 38 | from vllm.model_executor.models.qwen3_eagle3 import Eagle3Qwen3ForCausalLM |
| 39 | from vllm.platforms import current_platform |
| 40 | from vllm.utils.torch_utils import PIN_MEMORY, async_tensor_h2d |
| 41 | from vllm.v1.attention.backend import CommonAttentionMetadata |
| 42 | from vllm.v1.attention.backends.registry import AttentionBackendEnum |
| 43 | from vllm.v1.attention.backends.triton_attn import TritonAttentionMetadata |
| 44 | from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher |
| 45 | from vllm.v1.kv_cache_interface import KVCacheConfig, UniformTypeKVCacheSpecs |
| 46 | from vllm.v1.sample.metadata import SamplingMetadata |
| 47 | from vllm.v1.sample.ops.topk_topp_sampler import ( |
| 48 | empty_exponential_noise_like, |
| 49 | sample_with_exponential_noise, |
| 50 | ) |
| 51 | from vllm.v1.sample.sampler import _SAMPLING_EPS |
| 52 | from vllm.v1.spec_decode.metadata import SpecDecodeMetadata |
| 53 | from vllm.v1.spec_decode.utils import ( |
| 54 | PADDING_SLOT_ID, |
| 55 | _copy_and_expand_eagle_inputs, |
| 56 | _eagle_prepare_inputs_padded, |
| 57 | _eagle_prepare_next_token_padded, |
| 58 | _eagle_step_slot_mapping_metadata, |
| 59 | compute_new_slot_mapping, |
| 60 | eagle_step_update_slot_mapping_and_metadata, |
| 61 | extend_all_queries_by_N, |
| 62 | next_power_of_2, |
| 63 | ) |
| 64 | from vllm.v1.utils import CpuGpuBuffer |
| 65 | from vllm.v1.worker.dp_utils import coordinate_batch_across_dp |
| 66 | from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch |
| 67 | from vllm.v1.worker.utils import AttentionGroup |
| 68 | |
| 69 | logger = init_logger(__name__) |
| 70 | |
| 71 | |
| 72 | class SpecDecodeBaseProposer: |
| 73 | def __init__( |
| 74 | self, |
| 75 | vllm_config: VllmConfig, |
| 76 | device: torch.device, |
| 77 | pass_hidden_states_to_model: bool, |
| 78 | runner=None, |
| 79 | ): |
| 80 | self.vllm_config = vllm_config |
| 81 | assert vllm_config.speculative_config is not None |
| 82 | self.speculative_config = vllm_config.speculative_config |
| 83 | self.draft_model_config = self.speculative_config.draft_model_config |
| 84 | self.method = self.speculative_config.method |
| 85 | self.pass_hidden_states_to_model = pass_hidden_states_to_model |
| 86 | self._share_mtp_indices = False |
| 87 | |
| 88 | self.device = device |
| 89 | self.dtype = vllm_config.model_config.dtype |
| 90 | self.max_model_len = vllm_config.model_config.max_model_len |
| 91 | self.dp_rank = vllm_config.parallel_config.data_parallel_rank |
| 92 | self.eplb_state: EplbState | None = None |
| 93 | self.num_speculative_tokens = self.speculative_config.num_speculative_tokens |
| 94 | |
| 95 | # We need to get the hidden size from the draft model config because |
| 96 | # the draft model's hidden size can be different from the target model's |
| 97 | # hidden size (e.g., Llama 3.3 70B). |
| 98 | self.hidden_size = self.draft_model_config.get_hidden_size() |
| 99 | self.inputs_embeds_size = self.draft_model_config.get_inputs_embeds_size() |
| 100 | |
| 101 | # DeepSeek V4 MTP consumes the target's pre-hc_head residual stream, |
| 102 | # shape (T, hc_mult * hidden_size). Expand the hidden_states buffer |
| 103 | # so target_hidden_states fits; detect DeepseekV4 via draft hf_config. |
| 104 | draft_hf_config = self.draft_model_config.hf_config |
| 105 | if hasattr(draft_hf_config, "compress_ratios") and hasattr( |
| 106 | draft_hf_config, "hc_mult" |
| 107 | ): |
| 108 | self.hidden_size = self.hidden_size * draft_hf_config.hc_mult |
| 109 | |
| 110 | # Unifying eagle, draft model, and parallel drafting support. |
| 111 | # DFlash always uses parallel drafting (all tokens in one pass), |
| 112 | # but has an additional slot for the next_token_id (does not shift like EAGLE) |
| 113 | self.parallel_drafting: bool = self.speculative_config.parallel_drafting |
| 114 | self.extra_slots_per_request = ( |
| 115 | 1 if not self.parallel_drafting else self.num_speculative_tokens |
| 116 | ) |
| 117 | self.net_num_new_slots_per_request = self.extra_slots_per_request - ( |
| 118 | 1 if (self.pass_hidden_states_to_model and self.method != "dflash") else 0 |
| 119 | ) |
| 120 | self.needs_extra_input_slots = self.net_num_new_slots_per_request > 0 |
| 121 | |
| 122 | # When True, all draft steps reuse the same position as the |
| 123 | # first step instead of advancing by one each iteration. |
| 124 | # Used by draft models with Q-only attention that share KV |
| 125 | # with the target and always predict from the same position. |
| 126 | self.constant_draft_positions: bool = False |
| 127 | |
| 128 | self.parallel_drafting_token_id: int = 0 |
| 129 | self.parallel_drafting_hidden_state_tensor: torch.Tensor | None = None |
| 130 | if self.parallel_drafting: |
| 131 | self._init_parallel_drafting_params() |
| 132 | self.use_local_argmax_reduction: bool = ( |
| 133 | self.speculative_config.use_local_argmax_reduction |
| 134 | ) |
| 135 | self.use_fp64_gumbel = vllm_config.model_config.use_fp64_gumbel |
| 136 | |
| 137 | self.use_heterogeneous_vocab: bool = ( |
| 138 | self.speculative_config.use_heterogeneous_vocab |
| 139 | ) |
| 140 | self.vocab_mapping: VocabMapping | None = None |
| 141 | |
| 142 | self.max_batch_size = vllm_config.scheduler_config.max_num_seqs |
| 143 | self.max_num_tokens = vllm_config.scheduler_config.max_num_batched_tokens |
| 144 | _eagle_step_slot_mapping_metadata.register_warmup( |
| 145 | max_model_len=self.max_model_len, |
| 146 | max_batch_size=self.max_batch_size, |
| 147 | ) |
| 148 | _eagle_prepare_inputs_padded.register_warmup(max_batch_size=self.max_batch_size) |
| 149 | _eagle_prepare_next_token_padded.register_warmup( |
| 150 | vocab_size=vllm_config.model_config.get_vocab_size(), |
| 151 | num_sampled_tokens_per_req=self.num_speculative_tokens + 1, |
| 152 | max_batch_size=self.max_batch_size, |
| 153 | ) |
| 154 | if self.needs_extra_input_slots: |
| 155 | _copy_and_expand_eagle_inputs.register_warmup( |
| 156 | parallel_drafting_token_id=self.parallel_drafting_token_id, |
| 157 | num_padding_slots_per_request=self.extra_slots_per_request, |
| 158 | shift_input_ids=self.pass_hidden_states_to_model, |
| 159 | max_num_tokens=self.max_num_tokens, |
| 160 | ) |
| 161 | self.token_arange_np = np.arange(self.max_num_tokens, dtype=np.int32) |
| 162 | |
| 163 | # Can be specialized by methods like DFlash to reduce the limit |
| 164 | self.max_query_tokens = self.max_num_tokens |
| 165 | self.max_positions = self.max_num_tokens |
| 166 | |
| 167 | # Multi-modal data support |
| 168 | self.supports_mm_inputs = vllm_config.model_config.supports_multimodal_inputs |
| 169 | |
| 170 | self.draft_attn_groups: list[AttentionGroup] = [] |
| 171 | self.kv_cache_gid: int = -1 |
| 172 | self.eagle3_use_aux_hidden_state: bool = ( |
| 173 | self._get_eagle3_use_aux_hidden_state_from_config() |
| 174 | ) |
| 175 | |
| 176 | self.compilation_config = self.vllm_config.compilation_config |
| 177 | |
| 178 | # Cudagraph dispatcher for PIECEWISE-only dispatching in eagle. |
| 179 | # Keys are initialized later via initialize_cudagraph_keys() called from |
| 180 | # gpu_model_runner._check_and_update_cudagraph_mode after |
| 181 | # adjust_cudagraph_sizes_for_spec_decode is called. |
| 182 | self.cudagraph_dispatcher = CudagraphDispatcher(self.vllm_config) |
| 183 | |
| 184 | # persistent buffers for cuda graph |
| 185 | self.input_ids = torch.zeros( |
| 186 | self.max_num_tokens, dtype=torch.int32, device=device |
| 187 | ) |
| 188 | # Use draft model's M-RoPE setting, not target model's |
| 189 | # Draft models may be text-only even if target is multimodal |
| 190 | self.uses_mrope = self.draft_model_config.uses_mrope |
| 191 | if self.uses_mrope: |
| 192 | # NOTE: `mrope_positions` is implemented with one additional dummy |
| 193 | # position on purpose to make it non-contiguous so that it can work |
| 194 | # with torch compile. |
| 195 | # See detailed explanation in https://github.com/vllm-project/vllm/pull/12128#discussion_r1926431923 |
| 196 | |
| 197 | # NOTE: For text-only inputs, each dimension has identical position |
| 198 | # IDs, making M-RoPE functionally equivalent to 1D-RoPE. |
| 199 | # See page 5 of https://arxiv.org/abs/2409.12191 |
| 200 | self.mrope_positions = torch.zeros( |
| 201 | (self.draft_model_config.mrope_num_dims, self.max_positions + 1), |
| 202 | dtype=torch.int64, |
| 203 | device=device, |
| 204 | ) |
| 205 | else: |
| 206 | # RoPE need (max_num_tokens,) |
| 207 | self.positions = torch.zeros( |
| 208 | self.max_positions, |
| 209 | dtype=torch.int64, |
| 210 | device=device, |
| 211 | ) |
| 212 | self.hidden_states = torch.zeros( |
| 213 | (self.max_num_tokens, self.hidden_size), dtype=self.dtype, device=device |
| 214 | ) |
| 215 | |
| 216 | # Will be set when we initialize the attention backend |
| 217 | self.block_size: int = -1 |
| 218 | |
| 219 | # We need +1 here because the arange is used to set query_start_loc, |
| 220 | # which has one more element than batch_size. |
| 221 | max_num_slots_for_arange = max(self.max_batch_size + 1, self.max_num_tokens) |
| 222 | self.arange = torch.arange( |
| 223 | max_num_slots_for_arange, device=device, dtype=torch.int32 |
| 224 | ) |
| 225 | |
| 226 | if self.needs_extra_input_slots: |
| 227 | self._raise_if_padded_drafter_batch_disabled() |
| 228 | self._warn_if_multimodal() |
| 229 | self._raise_if_mrope() |
| 230 | |
| 231 | self.is_rejected_token_mask: torch.Tensor | None = None |
| 232 | self.is_masked_token_mask: torch.Tensor | None = None |
| 233 | if self.needs_extra_input_slots: |
| 234 | # For draft models and parallel drafting, we need to keep track of |
| 235 | # which tokens are rejected to update the slot mapping with padding slots. |
| 236 | self.is_rejected_token_mask = torch.zeros( |
| 237 | (self.max_num_tokens,), dtype=torch.bool, device=device |
| 238 | ) |
| 239 | # For parallel drafting, we also need to keep track of which tokens |
| 240 | # are parallel-padding tokens used to sample at later positions. |
| 241 | # We populate this tensor even when using draft models for simplicity. |
| 242 | self.is_masked_token_mask = torch.zeros( |
| 243 | (self.max_num_tokens,), dtype=torch.bool, device=device |
| 244 | ) |
| 245 | |
| 246 | self.inputs_embeds = torch.zeros( |
| 247 | (self.max_num_tokens, self.inputs_embeds_size), |
| 248 | dtype=self.dtype, |
| 249 | device=device, |
| 250 | ) |
| 251 | |
| 252 | self.backup_next_token_ids = CpuGpuBuffer( |
| 253 | self.max_batch_size, |
| 254 | dtype=torch.int32, |
| 255 | device=device, |
| 256 | with_numpy=True, |
| 257 | ) |
| 258 | self._enable_probabilistic_draft_probs = ( |
| 259 | self.speculative_config.rejection_sample_method == "standard" |
| 260 | and self.speculative_config.draft_sample_method == "probabilistic" |
| 261 | ) |
| 262 | self._last_draft_probs: torch.Tensor | None = None |
| 263 | |
| 264 | self._slot_mapping_buffer = torch.zeros( |
| 265 | self.max_positions, dtype=torch.int64, device=device |
| 266 | ) |
| 267 | |
| 268 | # Determine allowed attention backends once during initialization. |
| 269 | self.allowed_attn_types: tuple | None = None |
| 270 | if current_platform.is_rocm(): |
| 271 | from vllm.models.deepseek_v4.amd.rocm import ( |
| 272 | DeepseekV4ROCMAiterMLASparseMetadata, |
| 273 | DeepseekV4ROCMAiterSparseSWAMetadata, |
| 274 | ) |
| 275 | |
| 276 | # MiniMax-M3 sparse (lightning-indexer) attention. The multi-step |
| 277 | # drafting machinery is shared code at num_speculative_tokens>1. |
| 278 | # this just opts the metadata into the ROCm allowlist. |
| 279 | from vllm.models.minimax_m3.common.sparse_attention import ( |
| 280 | MiniMaxM3SparseMetadata, |
| 281 | ) |
| 282 | from vllm.v1.attention.backends.mla.indexer import ( |
| 283 | DeepseekV32IndexerMetadata, |
| 284 | ) |
| 285 | from vllm.v1.attention.backends.mla.rocm_aiter_mla_sparse import ( |
| 286 | ROCMAiterMLASparseMetadata, |
| 287 | ) |
| 288 | from vllm.v1.attention.backends.rocm_attn import RocmAttentionMetadata |
| 289 | |
| 290 | rocm_types = [ |
| 291 | TritonAttentionMetadata, |
| 292 | RocmAttentionMetadata, |
| 293 | ROCMAiterMLASparseMetadata, |
| 294 | DeepseekV4ROCMAiterMLASparseMetadata, |
| 295 | DeepseekV4ROCMAiterSparseSWAMetadata, |
| 296 | DeepseekV32IndexerMetadata, |
| 297 | MiniMaxM3SparseMetadata, |
| 298 | ] |
| 299 | # ROCM_AITER_FA is an optional backend |
| 300 | # We check is_enabled() here to avoid importing the backend module during |
| 301 | # auto-discovery when VLLM_ROCM_USE_AITER=0, which would trigger aiter |
| 302 | # import and JIT compilation warnings. Explicit backend selection via |
| 303 | # attention_config still works because the backend module is loaded |
| 304 | # directly when selected, not through this auto-discovery path. |
| 305 | # Check if backend module exists to allow explicit selection |
| 306 | if find_spec( |
| 307 | AttentionBackendEnum.ROCM_AITER_FA.get_path(include_classname=False) |
| 308 | ): |
| 309 | from vllm.v1.attention.backends.rocm_aiter_fa import ( |
| 310 | AiterFlashAttentionMetadata, |
| 311 | ) |
| 312 | |
| 313 | rocm_types.append(AiterFlashAttentionMetadata) |
| 314 | |
| 315 | # TRITON_MLA backend support for MLA models (e.g., DeepSeek) |
| 316 | from vllm.model_executor.layers.attention.mla_attention import ( |
| 317 | MLACommonMetadata, |
| 318 | ) |
| 319 | |
| 320 | rocm_types.append(MLACommonMetadata) |
| 321 | |
| 322 | # FlexAttention backend support |
| 323 | from vllm.v1.attention.backends.flex_attention import FlexAttentionMetadata |
| 324 | |
| 325 | rocm_types.append(FlexAttentionMetadata) |
| 326 | |
| 327 | self.allowed_attn_types = tuple(rocm_types) |
| 328 | |
| 329 | def _raise_if_padded_drafter_batch_disabled(self): |
| 330 | if self.speculative_config.disable_padded_drafter_batch: |
| 331 | raise NotImplementedError( |
| 332 | "Speculative Decoding with draft models or parallel drafting only " |
| 333 | "supports padded drafter batch. Please unset " |
| 334 | "disable_padded_drafter_batch in the speculative_config." |
| 335 | ) |
| 336 | |
| 337 | def _warn_if_multimodal(self): |
| 338 | if self.supports_mm_inputs: |
| 339 | logger.warning( |
| 340 | "Speculative Decoding with draft models or parallel drafting " |
| 341 | "does not fully support multimodal models yet. " |
| 342 | "Proceeding with text-only speculative decoding." |
| 343 | ) |
| 344 | |
| 345 | def _raise_if_mrope(self): |
| 346 | if self.draft_model_config.uses_mrope: |
| 347 | raise NotImplementedError( |
| 348 | "Speculative Decoding with draft models or parallel drafting " |
| 349 | "does not support M-RoPE yet" |
| 350 | ) |
| 351 | |
| 352 | def set_eplb_state(self, eplb_state: EplbState) -> None: |
| 353 | """Inject EPLB state after construction.""" |
| 354 | self.eplb_state = eplb_state |
| 355 | |
| 356 | def _init_parallel_drafting_params(self): |
| 357 | # For parallel drafting, we need the token ID to use for masked slots |
| 358 | # And for EAGLE + parallel drafting, we need the hidden state tensor to use |
| 359 | # for those masked slots. |
| 360 | |
| 361 | model_hf_config = self.draft_model_config.hf_config |
| 362 | # DFlash stores mask_token_id in dflash_config |
| 363 | dflash_config = getattr(model_hf_config, "dflash_config", None) |
| 364 | if dflash_config and "mask_token_id" in dflash_config: |
| 365 | self.parallel_drafting_token_id = dflash_config["mask_token_id"] |
| 366 | elif getattr(model_hf_config, "mask_token_id", None) is not None: |
| 367 | self.parallel_drafting_token_id = model_hf_config.mask_token_id |
| 368 | elif hasattr(model_hf_config, "dspark_noise_token_id"): |
| 369 | self.parallel_drafting_token_id = model_hf_config.dspark_noise_token_id |
| 370 | elif hasattr(model_hf_config, "pard_token"): |
| 371 | self.parallel_drafting_token_id = model_hf_config.pard_token |
| 372 | elif hasattr(model_hf_config, "ptd_token_id"): |
| 373 | self.parallel_drafting_token_id = model_hf_config.ptd_token_id |
| 374 | else: |
| 375 | raise ValueError( |
| 376 | "For parallel drafting, the draft model config must have " |
| 377 | "`dflash_config.mask_token_id`, `mask_token_id`, " |
| 378 | "`dspark_noise_token_id`, `pard_token`, or `ptd_token_id` " |
| 379 | "specified in its config.json." |
| 380 | ) |
| 381 | |
| 382 | if self.pass_hidden_states_to_model: |
| 383 | self.parallel_drafting_hidden_state_tensor = torch.empty( |
| 384 | self.hidden_size, dtype=self.dtype, device=self.device |
| 385 | ) |
| 386 | |
| 387 | def _get_positions(self, num_tokens: int): |
| 388 | if self.uses_mrope: |
| 389 | return self.mrope_positions[:, :num_tokens] |
| 390 | return self.positions[:num_tokens] |
| 391 | |
| 392 | def _set_positions(self, num_tokens: int, positions: torch.Tensor): |
| 393 | if self.uses_mrope: |
| 394 | self.mrope_positions[:, :num_tokens] = positions |
| 395 | else: |
| 396 | # Convert M-RoPE positions if target model uses M-RoPE |
| 397 | # but draft doesn't, For text inputs, all M-RoPE |
| 398 | # dimensions are identical |
| 399 | if self.vllm_config.model_config.uses_mrope: |
| 400 | positions = positions[0] |
| 401 | self.positions[:num_tokens] = positions |
| 402 | |
| 403 | def _get_slot_mapping( |
| 404 | self, |
| 405 | num_tokens: int, |
| 406 | slot_mapping: torch.Tensor | None = None, |
| 407 | ) -> dict[str, torch.Tensor]: |
| 408 | """Return slot_mapping dict for EAGLE layers. |
| 409 | |
| 410 | If slot_mapping is provided, copies it into the buffer first. |
| 411 | """ |
| 412 | if slot_mapping is not None: |
| 413 | num_actual = slot_mapping.shape[0] |
| 414 | self._slot_mapping_buffer[:num_actual].copy_(slot_mapping) |
| 415 | if num_tokens > num_actual: |
| 416 | self._slot_mapping_buffer[num_actual:num_tokens].fill_(PADDING_SLOT_ID) |
| 417 | |
| 418 | view = self._slot_mapping_buffer[:num_tokens] |
| 419 | return {name: view for name in self._draft_attn_layer_names} |
| 420 | |
| 421 | def initialize_cudagraph_keys(self, cudagraph_mode: CUDAGraphMode) -> None: |
| 422 | """Initialize cudagraph dispatcher keys for the drafter. |
| 423 | |
| 424 | Only supports PIECEWISE cudagraphs (via mixed_mode). |
| 425 | This should be called after adjust_cudagraph_sizes_for_spec_decode. |
| 426 | """ |
| 427 | if ( |
| 428 | not self.speculative_config.enforce_eager |
| 429 | and cudagraph_mode.mixed_mode() |
| 430 | in [CUDAGraphMode.PIECEWISE, CUDAGraphMode.FULL] |
| 431 | ): |
| 432 | eagle_cudagraph_mode = CUDAGraphMode.PIECEWISE |
| 433 | else: |
| 434 | eagle_cudagraph_mode = CUDAGraphMode.NONE |
| 435 | |
| 436 | self.cudagraph_dispatcher.initialize_cudagraph_keys(eagle_cudagraph_mode) |
| 437 | |
| 438 | def _greedy_sample(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 439 | """Greedy-sample draft tokens from hidden states.""" |
| 440 | if self.use_local_argmax_reduction: |
| 441 | return self.model.get_top_tokens(hidden_states) |
| 442 | if self.use_heterogeneous_vocab: |
| 443 | logits = self.model.compute_logits(hidden_states) |
| 444 | assert self.vocab_mapping is not None |
| 445 | logits = self.vocab_mapping.constrain_draft_logits(logits) |
| 446 | draft_token_ids = logits.argmax(dim=-1) |
| 447 | return self.vocab_mapping.map_draft_to_target_ids(draft_token_ids) |
| 448 | return self.model.compute_logits(hidden_states).argmax(dim=-1) |
| 449 | |
| 450 | def _sample_from_logits( |
| 451 | self, |
| 452 | logits: torch.Tensor, |
| 453 | sampling_metadata: SamplingMetadata, |
| 454 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 455 | if not self._enable_probabilistic_draft_probs: |
| 456 | return logits.argmax(dim=-1), None |
| 457 | if sampling_metadata.all_greedy: |
| 458 | return logits.argmax(dim=-1), None |
| 459 | |
| 460 | # Parallel drafting (e.g. DFlash) samples num_speculative_tokens rows |
| 461 | # per request in a single pass, so logits has batch_size * K rows while |
| 462 | # the sampling metadata is per-request. The rows are request-major |
| 463 | # (K consecutive slots per request), so repeat_interleave the |
| 464 | # per-request temperature to match before probabilistic sampling. |
| 465 | temperature = sampling_metadata.temperature |
| 466 | if temperature is not None and temperature.shape[0] != logits.shape[0]: |
| 467 | assert logits.shape[0] % temperature.shape[0] == 0 |
| 468 | factor = logits.shape[0] // temperature.shape[0] |
| 469 | sampling_metadata = dataclasses.replace( |
| 470 | sampling_metadata, |
| 471 | temperature=temperature.repeat_interleave(factor, dim=0), |
| 472 | ) |
| 473 | |
| 474 | return compute_probs_and_sample_next_token( |
| 475 | logits, sampling_metadata, self.use_fp64_gumbel |
| 476 | ) |
| 477 | |
| 478 | def _sample_draft_tokens( |
| 479 | self, |
| 480 | hidden_states: torch.Tensor, |
| 481 | sampling_metadata: SamplingMetadata, |
| 482 | ) -> tuple[torch.Tensor, torch.Tensor | None]: |
| 483 | if not self._enable_probabilistic_draft_probs or sampling_metadata.all_greedy: |
| 484 | return self._greedy_sample(hidden_states), None |
| 485 | logits = self.model.compute_logits(hidden_states) |
| 486 | if self.use_heterogeneous_vocab: |
| 487 | assert self.vocab_mapping is not None |
| 488 | logits = self.vocab_mapping.constrain_draft_logits(logits) |
| 489 | draft_token_ids, draft_probs = self._sample_from_logits( |
| 490 | logits, sampling_metadata |
| 491 | ) |
| 492 | if self.use_heterogeneous_vocab: |
| 493 | assert self.vocab_mapping is not None |
| 494 | draft_token_ids = self.vocab_mapping.map_draft_to_target_ids( |
| 495 | draft_token_ids |
| 496 | ) |
| 497 | # Config validation ensures draft_sample_method == "greedy" when |
| 498 | # use_heterogeneous_vocab is True, so this branch should never be |
| 499 | # reached. Kept as a safety fallback until probabilistic rejection |
| 500 | # sampling with heterogeneous vocabularies is implemented. |
| 501 | # TODO: remap draft_probs to target-vocab space for lossless |
| 502 | # probabilistic rejection sampling with heterogeneous vocabularies. |
| 503 | assert draft_probs is None, ( |
| 504 | "probabilistic draft sampling is not supported with " |
| 505 | "use_heterogeneous_vocab" |
| 506 | ) |
| 507 | return draft_token_ids, draft_probs |
| 508 | |
| 509 | def take_last_draft_probs(self) -> torch.Tensor | None: |
| 510 | return self._last_draft_probs |
| 511 | |
| 512 | def propose( |
| 513 | self, |
| 514 | num_speculative_tokens, |
| 515 | # [num_tokens] |
| 516 | target_token_ids: torch.Tensor, |
| 517 | # [num_tokens] or [3, num_tokens] when M-RoPE is enabled |
| 518 | target_positions: torch.Tensor, |
| 519 | # [num_tokens, hidden_size] |
| 520 | target_hidden_states: torch.Tensor, |
| 521 | # [batch_size] |
| 522 | next_token_ids: torch.Tensor, |
| 523 | token_indices_to_sample: torch.Tensor | None, |
| 524 | common_attn_metadata: CommonAttentionMetadata, |
| 525 | sampling_metadata: SamplingMetadata, |
| 526 | mm_embed_inputs: tuple[list[torch.Tensor], torch.Tensor] | None = None, |
| 527 | num_rejected_tokens_gpu: torch.Tensor | None = None, |
| 528 | slot_mappings: dict[str, torch.Tensor] |
| 529 | | list[dict[str, torch.Tensor]] |
| 530 | | None = None, |
| 531 | ) -> torch.Tensor: |
| 532 | self.num_speculative_tokens = num_speculative_tokens |
| 533 | self._last_draft_probs = None |
| 534 | batch_size = common_attn_metadata.batch_size() |
| 535 | |
| 536 | if self.method in ("eagle3", "dflash"): |
| 537 | model = self.model |
| 538 | if isinstance(model, BreakableCUDAGraphWrapper): |
| 539 | model = model.unwrap() |
| 540 | assert isinstance( |
| 541 | model, |
| 542 | ( |
| 543 | Eagle3LlamaForCausalLM, |
| 544 | Eagle3DeepseekV2ForCausalLM, |
| 545 | DFlashQwen3ForCausalLM, |
| 546 | Eagle3Qwen3ForCausalLM, |
| 547 | DFlashLagunaForCausalLM, |
| 548 | ), |
| 549 | ) |
| 550 | target_hidden_states = self.model.combine_hidden_states( |
| 551 | target_hidden_states |
| 552 | ) |
| 553 | assert target_hidden_states.shape[-1] == self.hidden_size |
| 554 | |
| 555 | num_tokens, token_indices_to_sample, common_attn_metadata = ( |
| 556 | self.set_inputs_first_pass( |
| 557 | target_token_ids=target_token_ids, |
| 558 | next_token_ids=next_token_ids, |
| 559 | target_positions=target_positions, |
| 560 | target_hidden_states=target_hidden_states, |
| 561 | token_indices_to_sample=token_indices_to_sample, |
| 562 | cad=common_attn_metadata, |
| 563 | num_rejected_tokens_gpu=num_rejected_tokens_gpu, |
| 564 | ) |
| 565 | ) |
| 566 | |
| 567 | per_group_attn_metadata, per_layer_attn_metadata = ( |
| 568 | self.build_per_group_and_layer_attn_metadata(common_attn_metadata) |
| 569 | ) |
| 570 | |
| 571 | cudagraph_runtime_mode, num_input_tokens, num_tokens_across_dp = ( |
| 572 | self._determine_batch_execution_and_padding(num_tokens) |
| 573 | ) |
| 574 | |
| 575 | model_kwargs, slot_mapping_size = self.build_model_inputs_first_pass( |
| 576 | num_tokens, num_input_tokens, mm_embed_inputs |
| 577 | ) |
| 578 | # Step 0 of index_share_for_mtp_iteration: let the MTP layer |
| 579 | # compute its own indices (skip_topk=False) so subsequent steps |
| 580 | # can reuse them. |
| 581 | if self._share_mtp_indices and hasattr(self.model.model, "set_skip_topk"): |
| 582 | self.model.model.set_skip_topk(False) |
| 583 | |
| 584 | if self.eplb_state is not None: |
| 585 | self.eplb_state.prepare_forward( |
| 586 | self.draft_model_config, |
| 587 | num_tokens, |
| 588 | ) |
| 589 | |
| 590 | with set_forward_context( |
| 591 | per_layer_attn_metadata, |
| 592 | self.vllm_config, |
| 593 | num_tokens=num_input_tokens, |
| 594 | num_tokens_across_dp=num_tokens_across_dp, |
| 595 | cudagraph_runtime_mode=cudagraph_runtime_mode, |
| 596 | slot_mapping=self._get_slot_mapping( |
| 597 | slot_mapping_size, common_attn_metadata.slot_mapping |
| 598 | ), |
| 599 | ): |
| 600 | ret_hidden_states = self.model(**model_kwargs) |
| 601 | if not self.model_returns_tuple(): |
| 602 | last_hidden_states = ret_hidden_states |
| 603 | hidden_states = last_hidden_states |
| 604 | else: |
| 605 | last_hidden_states, hidden_states = ret_hidden_states |
| 606 | |
| 607 | # After step 0: switch to reuse mode so steps 1+ skip the indexer |
| 608 | # and read the indices that step 0 just wrote into the shared buffer. |
| 609 | if self._share_mtp_indices and hasattr(self.model.model, "set_skip_topk"): |
| 610 | self.model.model.set_skip_topk(True) |
| 611 | # The topk indices were written for each query token in the multi-token |
| 612 | # batch. Compact the topk indices for each request's last token. |
| 613 | self.model.model.compact_topk_indices(token_indices_to_sample) |
| 614 | |
| 615 | sample_hidden_states = last_hidden_states[token_indices_to_sample] |
| 616 | |
| 617 | # No draft tokens requested (e.g. Dynamic SD decided K=0). |
| 618 | # The prefill forward pass above already ran to keep the drafter |
| 619 | # KV cache in sync, so just return an empty tensor. |
| 620 | if self.num_speculative_tokens == 0: |
| 621 | return torch.empty( |
| 622 | batch_size, |
| 623 | 0, |
| 624 | device=sample_hidden_states.device, |
| 625 | dtype=torch.int64, |
| 626 | ) |
| 627 | |
| 628 | # Early exit if there is only one draft token to be generated. |
| 629 | if self.num_speculative_tokens == 1 or self.parallel_drafting: |
| 630 | draft_token_ids, draft_probs = self._sample_draft_tokens( |
| 631 | sample_hidden_states, sampling_metadata |
| 632 | ) |
| 633 | if draft_probs is not None: |
| 634 | self._last_draft_probs = draft_probs.view( |
| 635 | -1, self.num_speculative_tokens, draft_probs.shape[-1] |
| 636 | ).contiguous() |
| 637 | return draft_token_ids.view(-1, self.num_speculative_tokens) |
| 638 | |
| 639 | if self.uses_mrope: |
| 640 | positions = self.mrope_positions[:, token_indices_to_sample] |
| 641 | else: |
| 642 | positions = self.positions[token_indices_to_sample] |
| 643 | hidden_states = hidden_states[token_indices_to_sample] |
| 644 | |
| 645 | if self.constant_draft_positions: |
| 646 | # Write the sampling positions into the front of the |
| 647 | # positions buffer so that subsequent loop iterations |
| 648 | # (which read via _get_positions) use the correct values. |
| 649 | self.positions[:batch_size] = positions |
| 650 | |
| 651 | draft_token_ids, draft_probs = self._sample_draft_tokens( |
| 652 | sample_hidden_states, sampling_metadata |
| 653 | ) |
| 654 | draft_probs_list = None if draft_probs is None else [draft_probs] |
| 655 | |
| 656 | if self.allowed_attn_types is not None: |
| 657 | for group_md in per_group_attn_metadata: |
| 658 | if not isinstance(group_md, self.allowed_attn_types): |
| 659 | raise ValueError( |
| 660 | f"Unsupported attention metadata type for speculative " |
| 661 | "decoding with num_speculative_tokens > 1: " |
| 662 | f"{type(group_md)}. Supported types are: " |
| 663 | f"{self.allowed_attn_types}" |
| 664 | ) |
| 665 | |
| 666 | # Generate the remaining draft tokens. |
| 667 | draft_token_ids_list = [draft_token_ids] |
| 668 | |
| 669 | cudagraph_runtime_mode, input_batch_size, batch_size_across_dp = ( |
| 670 | self._determine_batch_execution_and_padding(batch_size) |
| 671 | ) |
| 672 | |
| 673 | common_attn_metadata.num_actual_tokens = batch_size |
| 674 | common_attn_metadata.max_query_len = 1 |
| 675 | common_attn_metadata.query_start_loc = self.arange[: batch_size + 1] |
| 676 | common_attn_metadata.query_start_loc_cpu = torch.from_numpy( |
| 677 | self.token_arange_np[: batch_size + 1] |
| 678 | ).clone() |
| 679 | |
| 680 | # In padded drafter batch, we need to adjust the sequence lengths |
| 681 | # to remove the "padding" (i.e. rejected tokens). |
| 682 | # Only apply this adjustment when we have rejected tokens |
| 683 | # (i.e., not the first proposal). |
| 684 | if self.num_speculative_tokens > 1 and num_rejected_tokens_gpu is not None: |
| 685 | common_attn_metadata.seq_lens -= num_rejected_tokens_gpu |
| 686 | |
| 687 | block_size = self.block_size |
| 688 | assert block_size > 0, "block_size has not been initialized." |
| 689 | for token_index in range(self.num_speculative_tokens - 1): |
| 690 | # Update the inputs. |
| 691 | # cast to int32 is crucial when eagle model is compiled. |
| 692 | # tensor.argmax() returns int64 by default. |
| 693 | input_ids = draft_token_ids_list[-1].int() |
| 694 | |
| 695 | if self.use_heterogeneous_vocab: |
| 696 | # Map target token IDs to draft vocab space (TLI algorithm) |
| 697 | assert self.vocab_mapping is not None |
| 698 | input_ids = self.vocab_mapping.map_target_to_draft_ids(input_ids) |
| 699 | |
| 700 | if not self.constant_draft_positions: |
| 701 | positions = self._update_positions_dependent_metadata( |
| 702 | positions, |
| 703 | common_attn_metadata, |
| 704 | batch_size, |
| 705 | input_batch_size, |
| 706 | block_size, |
| 707 | ) |
| 708 | |
| 709 | # Rebuild attention metadata. When draft positions are constant |
| 710 | # (e.g. Gemma4 MTP), common_attn_metadata is invariant across |
| 711 | # loop iterations so we build once and reuse. |
| 712 | if not self.constant_draft_positions or token_index == 0: |
| 713 | _, per_layer_attn_metadata = ( |
| 714 | self.build_per_group_and_layer_attn_metadata( |
| 715 | common_attn_metadata, draft_index=token_index + 1 |
| 716 | ) |
| 717 | ) |
| 718 | |
| 719 | # copy inputs to buffer for cudagraph |
| 720 | self.input_ids[:batch_size] = input_ids |
| 721 | self.hidden_states[:batch_size] = hidden_states |
| 722 | if self.supports_mm_inputs: |
| 723 | self.inputs_embeds[:batch_size] = self.model.embed_input_ids(input_ids) |
| 724 | |
| 725 | input_ids = None |
| 726 | inputs_embeds = self.inputs_embeds[:input_batch_size] |
| 727 | else: |
| 728 | input_ids = self.input_ids[:input_batch_size] |
| 729 | inputs_embeds = None |
| 730 | |
| 731 | # Run the model. |
| 732 | model_kwargs = { |
| 733 | "input_ids": input_ids, |
| 734 | "positions": self._get_positions(input_batch_size), |
| 735 | "inputs_embeds": inputs_embeds, |
| 736 | } |
| 737 | if self.pass_hidden_states_to_model: |
| 738 | model_kwargs["hidden_states"] = self.hidden_states[:input_batch_size] |
| 739 | |
| 740 | if self.eplb_state is not None: |
| 741 | self.eplb_state.prepare_forward( |
| 742 | self.draft_model_config, |
| 743 | batch_size, |
| 744 | ) |
| 745 | |
| 746 | with set_forward_context( |
| 747 | per_layer_attn_metadata, |
| 748 | self.vllm_config, |
| 749 | num_tokens=input_batch_size, |
| 750 | num_tokens_across_dp=batch_size_across_dp, |
| 751 | cudagraph_runtime_mode=cudagraph_runtime_mode, |
| 752 | slot_mapping=self._get_slot_mapping(input_batch_size), |
| 753 | ): |
| 754 | ret_hidden_states = self.model(**model_kwargs) |
| 755 | if not self.model_returns_tuple(): |
| 756 | last_hidden_states = ret_hidden_states |
| 757 | hidden_states = ret_hidden_states |
| 758 | else: |
| 759 | last_hidden_states, hidden_states = ret_hidden_states |
| 760 | |
| 761 | hidden_states = hidden_states[:batch_size] |
| 762 | draft_token_ids, draft_probs = self._sample_draft_tokens( |
| 763 | last_hidden_states[:batch_size], sampling_metadata |
| 764 | ) |
| 765 | if draft_probs is not None: |
| 766 | assert draft_probs_list is not None |
| 767 | draft_probs_list.append(draft_probs) |
| 768 | draft_token_ids_list.append(draft_token_ids) |
| 769 | |
| 770 | # [batch_size, num_speculative_tokens] |
| 771 | draft_token_ids = torch.stack(draft_token_ids_list, dim=1) |
| 772 | if draft_probs_list is not None: |
| 773 | self._last_draft_probs = torch.stack(draft_probs_list, dim=1).contiguous() |
| 774 | return draft_token_ids |
| 775 | |
| 776 | def _update_positions_dependent_metadata( |
| 777 | self, |
| 778 | positions: torch.Tensor, |
| 779 | common_attn_metadata, |
| 780 | batch_size: int, |
| 781 | input_batch_size: int, |
| 782 | block_size: int, |
| 783 | ) -> torch.Tensor: |
| 784 | """Update positions, slot mappings, and sequence metadata for the |
| 785 | next draft step. Returns the updated positions tensor.""" |
| 786 | positions_1d = positions[0] if self.uses_mrope else positions |
| 787 | if self.uses_mrope: |
| 788 | out_pos = self.mrope_positions[0, :batch_size] |
| 789 | else: |
| 790 | out_pos = self.positions[:batch_size] |
| 791 | eagle_step_update_slot_mapping_and_metadata( |
| 792 | positions_1d=positions_1d, |
| 793 | block_table_tensor=common_attn_metadata.block_table_tensor, |
| 794 | seq_lens=common_attn_metadata.seq_lens, |
| 795 | block_size=block_size, |
| 796 | max_model_len=self.max_model_len, |
| 797 | out_clamped_positions=out_pos, |
| 798 | out_slot_mapping=self._slot_mapping_buffer[:input_batch_size], |
| 799 | input_batch_size=input_batch_size, |
| 800 | ) |
| 801 | common_attn_metadata.slot_mapping = self._slot_mapping_buffer[:batch_size] |
| 802 | if self.uses_mrope: |
| 803 | self.mrope_positions[1:, :batch_size] = self.mrope_positions[0, :batch_size] |
| 804 | positions = self.mrope_positions[:, :batch_size] |
| 805 | else: |
| 806 | positions = self.positions[:batch_size] |
| 807 | common_attn_metadata.max_seq_len = min( |
| 808 | common_attn_metadata.max_seq_len + 1, |
| 809 | self.max_model_len, |
| 810 | ) |
| 811 | |
| 812 | if common_attn_metadata.seq_lens_cpu_upper_bound is not None: |
| 813 | common_attn_metadata.seq_lens_cpu_upper_bound += 1 |
| 814 | |
| 815 | return positions |
| 816 | |
| 817 | def set_inputs_first_pass( |
| 818 | self, |
| 819 | target_token_ids: torch.Tensor, |
| 820 | next_token_ids: torch.Tensor, |
| 821 | target_positions: torch.Tensor, |
| 822 | target_hidden_states: torch.Tensor, |
| 823 | token_indices_to_sample: torch.Tensor | None, |
| 824 | cad: CommonAttentionMetadata, |
| 825 | num_rejected_tokens_gpu: torch.Tensor | None, |
| 826 | ) -> tuple[int, torch.Tensor, CommonAttentionMetadata]: |
| 827 | # Map target token IDs to draft vocab space (TLI algorithm) |
| 828 | if self.use_heterogeneous_vocab: |
| 829 | assert self.vocab_mapping is not None |
| 830 | target_token_ids = self.vocab_mapping.map_target_to_draft_ids( |
| 831 | target_token_ids |
| 832 | ) |
| 833 | next_token_ids = self.vocab_mapping.map_target_to_draft_ids(next_token_ids) |
| 834 | if not self.needs_extra_input_slots: |
| 835 | # Default EAGLE pathway: no reshaping of input tensors needed. |
| 836 | # Simply rotate the input ids and leave the positions unchanged, |
| 837 | # Inserting the next token ids at the last slot in each request. |
| 838 | if token_indices_to_sample is None: |
| 839 | token_indices_to_sample = cad.query_start_loc[1:] - 1 |
| 840 | |
| 841 | num_tokens = target_token_ids.shape[0] |
| 842 | # Shift the input ids by one token. |
| 843 | # E.g., [a1, b1, b2, c1, c2, c3] -> [b1, b2, c1, c2, c3, c3] |
| 844 | self.input_ids[: num_tokens - 1] = target_token_ids[1:] |
| 845 | # Replace the last token with the next token. |
| 846 | # E.g., [b1, b2, c1, c2, c3, c3] -> [a2, b2, b3, c2, c3, c4] |
| 847 | self.input_ids[token_indices_to_sample] = next_token_ids |
| 848 | |
| 849 | # copy inputs to buffer for cudagraph |
| 850 | self._set_positions(num_tokens, target_positions) |
| 851 | |
| 852 | self.hidden_states[:num_tokens] = target_hidden_states |
| 853 | |
| 854 | return num_tokens, token_indices_to_sample, cad |
| 855 | else: |
| 856 | assert self.is_rejected_token_mask is not None |
| 857 | assert self.is_masked_token_mask is not None |
| 858 | # 1. |
| 859 | # Call a custom triton kernel to copy input_ids and positions |
| 860 | # into the correct slots in the preallocated buffers self.input_ids, |
| 861 | # self.positions. |
| 862 | batch_size = cad.batch_size() |
| 863 | # Since we might have to copy a lot of data for prefills, we select the |
| 864 | # block size based on the max query length and limit to max 256 slots/block. |
| 865 | max_num_tokens_per_request = ( |
| 866 | cad.max_query_len + self.net_num_new_slots_per_request |
| 867 | ) |
| 868 | BLOCK_SIZE_TOKENS = min(256, next_power_of_2(max_num_tokens_per_request)) |
| 869 | num_blocks = ( |
| 870 | max_num_tokens_per_request + BLOCK_SIZE_TOKENS - 1 |
| 871 | ) // BLOCK_SIZE_TOKENS |
| 872 | total_num_input_tokens = target_token_ids.shape[0] |
| 873 | total_num_output_tokens = total_num_input_tokens + ( |
| 874 | self.net_num_new_slots_per_request * batch_size |
| 875 | ) |
| 876 | |
| 877 | token_indices_to_sample = torch.empty( |
| 878 | batch_size * self.extra_slots_per_request, |
| 879 | dtype=torch.int32, |
| 880 | device=self.device, |
| 881 | ) |
| 882 | |
| 883 | # Destination indices to write target_hidden_states into drafting buffer. |
| 884 | out_hidden_state_mapping = torch.empty( |
| 885 | total_num_input_tokens, dtype=torch.int32, device=self.device |
| 886 | ) |
| 887 | |
| 888 | # Kernel grid: one program per request (row) |
| 889 | query_start_loc = cad.query_start_loc |
| 890 | query_end_loc = cad.query_start_loc[1:] - 1 |
| 891 | if num_rejected_tokens_gpu is not None: |
| 892 | query_end_loc = query_end_loc - num_rejected_tokens_gpu |
| 893 | |
| 894 | _copy_and_expand_eagle_inputs( |
| 895 | # (Padded) Inputs from the target model |
| 896 | target_token_ids_ptr=target_token_ids, |
| 897 | target_positions_ptr=target_positions, |
| 898 | next_token_ids_ptr=next_token_ids, # sampled tokens, one per request |
| 899 | # Outputs to the drafting buffers |
| 900 | out_input_ids_ptr=self.input_ids, |
| 901 | out_positions_ptr=self.positions, # Doesn't support mrope for now |
| 902 | out_is_rejected_token_mask_ptr=self.is_rejected_token_mask, |
| 903 | out_is_masked_token_mask_ptr=self.is_masked_token_mask, |
| 904 | out_new_token_indices_ptr=token_indices_to_sample, |
| 905 | out_hidden_state_mapping_ptr=out_hidden_state_mapping, |
| 906 | # Input metadata |
| 907 | query_start_loc_ptr=query_start_loc, |
| 908 | query_end_loc_ptr=query_end_loc, |
| 909 | padding_token_id=0, |
| 910 | parallel_drafting_token_id=self.parallel_drafting_token_id, |
| 911 | # Sizing info |
| 912 | # Note that we can deduce batch_size for free from the grid size |
| 913 | total_input_tokens=total_num_input_tokens, |
| 914 | num_padding_slots_per_request=self.extra_slots_per_request, |
| 915 | shift_input_ids=self.pass_hidden_states_to_model, |
| 916 | BLOCK_SIZE_TOKENS=BLOCK_SIZE_TOKENS, |
| 917 | batch_size=batch_size, |
| 918 | num_blocks=num_blocks, |
| 919 | ) |
| 920 | if self.pass_hidden_states_to_model: |
| 921 | assert self.parallel_drafting_hidden_state_tensor is not None |
| 922 | self.hidden_states[out_hidden_state_mapping] = target_hidden_states |
| 923 | # Use torch.where to avoid DtoH sync from boolean indexing |
| 924 | mask = self.is_masked_token_mask[:total_num_output_tokens] |
| 925 | torch.where( |
| 926 | mask.unsqueeze(1), |
| 927 | self.parallel_drafting_hidden_state_tensor, |
| 928 | self.hidden_states[:total_num_output_tokens], |
| 929 | out=self.hidden_states[:total_num_output_tokens], |
| 930 | ) |
| 931 | |
| 932 | # 2. |
| 933 | # Recompute the slot mapping based on the new positions and |
| 934 | # rejection mask. |
| 935 | assert self.block_size > 0, "block_size has not been initialized." |
| 936 | new_slot_mapping = compute_new_slot_mapping( |
| 937 | cad=cad, |
| 938 | new_positions=self.positions[:total_num_output_tokens], |
| 939 | is_rejected_token_mask=self.is_rejected_token_mask[ |
| 940 | :total_num_output_tokens |
| 941 | ], |
| 942 | block_size=self.block_size, |
| 943 | num_new_tokens=self.net_num_new_slots_per_request, |
| 944 | max_model_len=self.max_model_len, |
| 945 | ) |
| 946 | |
| 947 | # 3. Update the common attention metadata with the new (meta)data |
| 948 | new_cad = extend_all_queries_by_N( |
| 949 | cad, |
| 950 | N=self.net_num_new_slots_per_request, |
| 951 | arange=self.arange, |
| 952 | new_slot_mapping=new_slot_mapping, |
| 953 | ) |
| 954 | |
| 955 | return total_num_output_tokens, token_indices_to_sample, new_cad |
| 956 | |
| 957 | def build_model_inputs_first_pass( |
| 958 | self, |
| 959 | num_tokens: int, |
| 960 | num_input_tokens: int, |
| 961 | mm_embed_inputs: tuple[list[torch.Tensor], torch.Tensor] | None, |
| 962 | ) -> tuple[dict[str, Any], int]: |
| 963 | if self.supports_mm_inputs: |
| 964 | mm_embeds, is_mm_embed = mm_embed_inputs or (None, None) |
| 965 | |
| 966 | self.inputs_embeds[:num_tokens] = self.model.embed_input_ids( |
| 967 | self.input_ids[:num_tokens], |
| 968 | multimodal_embeddings=mm_embeds, |
| 969 | is_multimodal=is_mm_embed, |
| 970 | ) |
| 971 | |
| 972 | input_ids = None |
| 973 | inputs_embeds = self.inputs_embeds[:num_input_tokens] |
| 974 | else: |
| 975 | input_ids = self.input_ids[:num_input_tokens] |
| 976 | inputs_embeds = None |
| 977 | |
| 978 | model_kwargs = { |
| 979 | "input_ids": input_ids, |
| 980 | "positions": self._get_positions(num_input_tokens), |
| 981 | "inputs_embeds": inputs_embeds, |
| 982 | } |
| 983 | if self.pass_hidden_states_to_model: |
| 984 | model_kwargs["hidden_states"] = self.hidden_states[:num_input_tokens] |
| 985 | |
| 986 | return model_kwargs, num_input_tokens |
| 987 | |
| 988 | def build_per_group_and_layer_attn_metadata( |
| 989 | self, common_attn_metadata: CommonAttentionMetadata, draft_index: int = 0 |
| 990 | ) -> tuple[list[object], dict[str, object]]: |
| 991 | per_group_attn_metadata: list[object] = [] |
| 992 | per_layer_attn_metadata: dict[str, object] = {} |
| 993 | for attn_group in self.draft_attn_groups: |
| 994 | attn_metadata = attn_group.get_metadata_builder().build_for_drafting( |
| 995 | common_attn_metadata=common_attn_metadata, draft_index=draft_index |
| 996 | ) |
| 997 | per_group_attn_metadata.append(attn_metadata) |
| 998 | for layer_name in attn_group.layer_names: |
| 999 | per_layer_attn_metadata[layer_name] = attn_metadata |
| 1000 | return per_group_attn_metadata, per_layer_attn_metadata |
| 1001 | |
| 1002 | def model_returns_tuple(self) -> bool: |
| 1003 | if self.method == "mtp": |
| 1004 | # These models return separate hidden states for logits and for |
| 1005 | # feedback into the next draft step. |
| 1006 | architectures = self.draft_model_config.hf_config.architectures or [] |
| 1007 | return bool( |
| 1008 | { |
| 1009 | "DeepSeekMTPModel", |
| 1010 | "DeepseekV32MTPModel", |
| 1011 | "Glm5NextMTPModel", |
| 1012 | "KimiK3MTPModel", |
| 1013 | }.intersection(architectures) |
| 1014 | ) |
| 1015 | return self.method not in ("mtp", "draft_model", "dflash") |
| 1016 | |
| 1017 | def prepare_next_token_ids_cpu( |
| 1018 | self, |
| 1019 | sampled_token_ids: list[list[int]], |
| 1020 | requests: dict[str, CachedRequestState], |
| 1021 | gpu_input_batch: InputBatch, |
| 1022 | num_scheduled_tokens: dict[str, int], |
| 1023 | ) -> torch.Tensor: |
| 1024 | """This function is used to prepare the inputs for speculative decoding. |
| 1025 | It calculates the next token ids for each request based on the sampled |
| 1026 | token ids from the CPU. If a request has no sampled token ids (e.g., |
| 1027 | during the initial decoding steps), it falls back to using the request |
| 1028 | state to get the next token id. |
| 1029 | """ |
| 1030 | req_ids = gpu_input_batch.req_ids |
| 1031 | next_token_ids: list[int] = [] |
| 1032 | for i, token_ids in enumerate(sampled_token_ids): |
| 1033 | if token_ids: |
| 1034 | # Common case. |
| 1035 | next_token_id = token_ids[-1] |
| 1036 | else: |
| 1037 | # Partial prefill (rare case). |
| 1038 | # Get the next token id from the request state. |
| 1039 | req_id = req_ids[i] |
| 1040 | req_state = requests[req_id] |
| 1041 | seq_len = req_state.num_computed_tokens + num_scheduled_tokens[req_id] |
| 1042 | next_token_id = req_state.get_token_id(seq_len) |
| 1043 | next_token_ids.append(next_token_id) |
| 1044 | next_token_ids = torch.tensor( |
| 1045 | next_token_ids, dtype=torch.int32, device=self.input_ids.device |
| 1046 | ) |
| 1047 | return next_token_ids |
| 1048 | |
| 1049 | def prepare_next_token_ids_padded( |
| 1050 | self, |
| 1051 | sampled_token_ids: torch.Tensor, |
| 1052 | requests: dict[str, CachedRequestState], |
| 1053 | gpu_input_batch: InputBatch, |
| 1054 | discard_request_mask: torch.Tensor, |
| 1055 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 1056 | """This function is used to prepare the inputs for speculative decoding. |
| 1057 | It calculates the next token ids and the number of valid sampled tokens |
| 1058 | for each request, considering the "discarded" requests whose next token |
| 1059 | is not sampled and comes from `request.get_token_id()` instead. This is denoted |
| 1060 | the "backup" token id. It also counts rejected tokens via `sampled_token_ids`. |
| 1061 | """ |
| 1062 | # Precompute backup token IDs for discarded requests. |
| 1063 | num_reqs = gpu_input_batch.num_reqs |
| 1064 | for i in range(num_reqs): |
| 1065 | self.backup_next_token_ids.np[i] = requests[ |
| 1066 | gpu_input_batch.req_ids[i] |
| 1067 | ].get_token_id(gpu_input_batch.num_tokens_no_spec[i] - 1) |
| 1068 | self.backup_next_token_ids.copy_to_gpu(num_reqs) |
| 1069 | backup_tokens_gpu = self.backup_next_token_ids.gpu |
| 1070 | |
| 1071 | batch_size, num_tokens = sampled_token_ids.shape |
| 1072 | device = sampled_token_ids.device |
| 1073 | |
| 1074 | assert discard_request_mask.dtype == torch.bool |
| 1075 | assert backup_tokens_gpu.dtype == torch.int32 |
| 1076 | |
| 1077 | next_token_ids = torch.empty(batch_size, dtype=torch.int32, device=device) |
| 1078 | valid_sampled_tokens_count = next_token_ids.new_empty(batch_size) |
| 1079 | |
| 1080 | # Kernel grid: one program per request (row) |
| 1081 | |
| 1082 | # Find the next power of 2 for block sizes |
| 1083 | BLOCK_SIZE_TOKENS = next_power_of_2(num_tokens) |
| 1084 | _eagle_prepare_next_token_padded( |
| 1085 | sampled_token_ids, |
| 1086 | discard_request_mask, |
| 1087 | backup_tokens_gpu, |
| 1088 | next_token_ids, |
| 1089 | valid_sampled_tokens_count, |
| 1090 | gpu_input_batch.vocab_size, |
| 1091 | num_tokens, |
| 1092 | batch_size, |
| 1093 | sampled_token_ids.stride(0), |
| 1094 | BLOCK_SIZE_TOKENS=BLOCK_SIZE_TOKENS, |
| 1095 | ) |
| 1096 | |
| 1097 | return next_token_ids, valid_sampled_tokens_count |
| 1098 | |
| 1099 | def prepare_inputs_padded( |
| 1100 | self, |
| 1101 | common_attn_metadata: CommonAttentionMetadata, |
| 1102 | spec_decode_metadata: SpecDecodeMetadata, |
| 1103 | valid_sampled_tokens_count: torch.Tensor, |
| 1104 | ) -> tuple[CommonAttentionMetadata, torch.Tensor, torch.Tensor]: |
| 1105 | """This function is used to prepare the inputs for speculative decoding |
| 1106 | It updates the common_attn_metadata for speculative decoding, |
| 1107 | but does not consider the rejected tokens. Instead, all tokens |
| 1108 | are included as inputs to the speculator, with the rejected tokens |
| 1109 | used as padding and filtered out later by `token_indices_to_sample`. |
| 1110 | No blocking CPU operations should be introduced in this function. |
| 1111 | """ |
| 1112 | num_reqs = common_attn_metadata.num_reqs |
| 1113 | device = valid_sampled_tokens_count.device |
| 1114 | |
| 1115 | token_indices_to_sample = torch.empty( |
| 1116 | (num_reqs,), dtype=torch.int32, device=device |
| 1117 | ) |
| 1118 | num_rejected_tokens_gpu = torch.empty( |
| 1119 | (num_reqs,), dtype=torch.int32, device=device |
| 1120 | ) |
| 1121 | |
| 1122 | _eagle_prepare_inputs_padded( |
| 1123 | spec_decode_metadata.cu_num_draft_tokens, |
| 1124 | valid_sampled_tokens_count, |
| 1125 | common_attn_metadata.query_start_loc, |
| 1126 | token_indices_to_sample, |
| 1127 | num_rejected_tokens_gpu, |
| 1128 | num_reqs, |
| 1129 | ) |
| 1130 | |
| 1131 | query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu |
| 1132 | new_query_len_per_req = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1] |
| 1133 | |
| 1134 | total_num_tokens = query_start_loc_cpu[-1].item() |
| 1135 | |
| 1136 | spec_common_attn_metadata = CommonAttentionMetadata( |
| 1137 | query_start_loc=common_attn_metadata.query_start_loc, |
| 1138 | seq_lens=common_attn_metadata.seq_lens, |
| 1139 | query_start_loc_cpu=query_start_loc_cpu, |
| 1140 | seq_lens_cpu_upper_bound=common_attn_metadata.seq_lens_cpu_upper_bound, |
| 1141 | num_reqs=common_attn_metadata.num_reqs, |
| 1142 | num_actual_tokens=total_num_tokens, |
| 1143 | max_query_len=new_query_len_per_req.max().item(), |
| 1144 | max_seq_len=common_attn_metadata.max_seq_len, |
| 1145 | block_table_tensor=common_attn_metadata.block_table_tensor, |
| 1146 | slot_mapping=common_attn_metadata.slot_mapping[:total_num_tokens], |
| 1147 | causal=True, |
| 1148 | dcp_local_seq_lens=common_attn_metadata.dcp_local_seq_lens, |
| 1149 | ) |
| 1150 | |
| 1151 | return ( |
| 1152 | spec_common_attn_metadata, |
| 1153 | token_indices_to_sample, |
| 1154 | num_rejected_tokens_gpu, |
| 1155 | ) |
| 1156 | |
| 1157 | def prepare_inputs( |
| 1158 | self, |
| 1159 | common_attn_metadata: CommonAttentionMetadata, |
| 1160 | sampled_token_ids: list[list[int]], |
| 1161 | num_draft_tokens: list[int], |
| 1162 | ) -> tuple[CommonAttentionMetadata, torch.Tensor]: |
| 1163 | """This function is used to prepare the inputs for speculative decoding. |
| 1164 | It updates to the common_attn_metadata to account for the rejected |
| 1165 | tokens (and newly sampled tokens). It also returns the token indices |
| 1166 | of the tokens that should be fed to the speculator. |
| 1167 | """ |
| 1168 | # E.g. |
| 1169 | # common_attn_metadata.query_start_loc{_cpu}: |
| 1170 | # [0, q1, q1 + q2, q1 + q2 + q3] |
| 1171 | # common_attn_metadata.seq_lens{_cpu}: [s1, s2, s3] |
| 1172 | # num_rejected_tokens: [n1, n2, n3] |
| 1173 | # This function computes the intermediate values: |
| 1174 | # num_tokens_per_req: [q1 - n1, q2 - n2, q3 - n3] |
| 1175 | # And returns: |
| 1176 | # common_attn_metadata.query_start_loc{_cpu}: |
| 1177 | # [0, q1 - n1, q1 + q2 - n1 - n2, q1 + q2 + q3 - n1 - n2 - n3] |
| 1178 | # common_attn_metadata.seq_lens{_cpu}: |
| 1179 | # [s1 - n1 + 1, s2 - n2 + 1, s3 - n3 + 1] |
| 1180 | # token_indices: [0, 1, ..., q1 - n1 - 1, |
| 1181 | # q1, q1 + 1, ..., q1 + q2 - n2 - 1, |
| 1182 | # q1 + q2, q1 + q2 + 1, ..., q1 + q2 + q3 - n3 - 1] |
| 1183 | |
| 1184 | num_rejected_tokens = [ |
| 1185 | n + 1 - len(sampled_token_ids[i]) if n > 0 else 0 |
| 1186 | for i, n in enumerate(num_draft_tokens) |
| 1187 | ] |
| 1188 | num_rejected_tokens = torch.tensor(num_rejected_tokens, dtype=torch.int32) |
| 1189 | |
| 1190 | device = common_attn_metadata.query_start_loc.device |
| 1191 | query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu |
| 1192 | # upper_bound - rejected = actual post-rejection seq_lens (no D2H sync). |
| 1193 | assert common_attn_metadata.seq_lens_cpu_upper_bound is not None |
| 1194 | new_seq_lens_cpu = ( |
| 1195 | common_attn_metadata.seq_lens_cpu_upper_bound - num_rejected_tokens |
| 1196 | ) |
| 1197 | |
| 1198 | # [0, q1, q1 + q2, q1 + q2 + q3] -> [q1, q2, q3] |
| 1199 | new_query_len_per_req = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1] |
| 1200 | # [q1, q2, q3] -> [q1 - n1, q2 - n2, q3 - n3] |
| 1201 | new_num_tokens_per_req = new_query_len_per_req - num_rejected_tokens |
| 1202 | new_num_tokens_per_req_np = new_num_tokens_per_req.numpy() |
| 1203 | |
| 1204 | # [q1 - n1, q2 - n2, q3 - n3] -> |
| 1205 | # [0, q1 - n1, q1 + q2 - n1 - n2, q1 + q2 + q3 - n1 - n2 - n3] |
| 1206 | new_query_start_loc_cpu = torch.zeros( |
| 1207 | query_start_loc_cpu.shape, |
| 1208 | dtype=torch.int32, |
| 1209 | pin_memory=PIN_MEMORY, |
| 1210 | ) |
| 1211 | new_query_start_loc_np = new_query_start_loc_cpu.numpy() |
| 1212 | np.cumsum(new_num_tokens_per_req_np, out=new_query_start_loc_np[1:]) |
| 1213 | |
| 1214 | total_num_tokens = new_query_start_loc_np[-1] |
| 1215 | # Example assuming num_tokens_per_req_np = [2, 4, 3] |
| 1216 | # this implies that `new_query_start_locs` is: |
| 1217 | # [0, 2, 6, 9] -> |
| 1218 | # [0, 0, 2, 2, 2, 2, 6, 6, 6] |
| 1219 | # _r1_ ____r2____ ___r3__ |
| 1220 | new_query_start_locs_expanded = np.repeat( |
| 1221 | new_query_start_loc_np[:-1], new_num_tokens_per_req_np |
| 1222 | ) |
| 1223 | # [0, 1, 2, 3, 4, 5, 6, 7, 8] -> |
| 1224 | # [0, 1, 0, 1, 2, 3, 0, 1, 2] |
| 1225 | # _r1_ ____r2____ ___r3__ |
| 1226 | token_offsets = ( |
| 1227 | self.token_arange_np[:total_num_tokens] - new_query_start_locs_expanded |
| 1228 | ) |
| 1229 | |
| 1230 | # Expand starting positions to match token pattern |
| 1231 | # [0, q1, q1 + q2] -> |
| 1232 | # [0, 0, q1, q1, q1, q1, q1 + q2, q1 + q2, q1 + q2] |
| 1233 | # _r1_ _____r2_______ ___________r3____________ |
| 1234 | old_query_start_locs_expanded = np.repeat( |
| 1235 | query_start_loc_cpu[:-1].numpy(), new_num_tokens_per_req_np |
| 1236 | ) |
| 1237 | # Final token indices are: |
| 1238 | # [0, 1, // req 1 |
| 1239 | # q1 + 0, q1 + 1, q1 + 2, q1 + 3, // req 2 |
| 1240 | # q1 + q2 + 0, q1 + q2 + 1, q1 + q2 + 2] // req 3 |
| 1241 | token_indices_np = token_offsets + old_query_start_locs_expanded |
| 1242 | token_indices = async_tensor_h2d(token_indices_np, device=device) |
| 1243 | |
| 1244 | spec_common_attn_metadata = CommonAttentionMetadata( |
| 1245 | query_start_loc=async_tensor_h2d(new_query_start_loc_cpu, device=device), |
| 1246 | seq_lens=async_tensor_h2d(new_seq_lens_cpu, device=device), |
| 1247 | query_start_loc_cpu=new_query_start_loc_cpu, |
| 1248 | seq_lens_cpu_upper_bound=new_seq_lens_cpu, |
| 1249 | num_reqs=common_attn_metadata.num_reqs, |
| 1250 | num_actual_tokens=total_num_tokens, |
| 1251 | max_query_len=new_query_len_per_req.max().item(), |
| 1252 | max_seq_len=new_seq_lens_cpu.max().item(), |
| 1253 | block_table_tensor=common_attn_metadata.block_table_tensor, |
| 1254 | slot_mapping=common_attn_metadata.slot_mapping[token_indices], |
| 1255 | causal=True, |
| 1256 | dcp_local_seq_lens=common_attn_metadata.dcp_local_seq_lens, |
| 1257 | ) |
| 1258 | |
| 1259 | return spec_common_attn_metadata, token_indices |
| 1260 | |
| 1261 | def get_model_name(self, model: nn.Module) -> str: |
| 1262 | if hasattr(model, "module"): # multi-GPU |
| 1263 | model = model.module |
| 1264 | return model.__class__.__name__ |
| 1265 | |
| 1266 | def _create_draft_vllm_config(self) -> VllmConfig: |
| 1267 | """Return a VllmConfig with kernel-level overrides for the proposer. |
| 1268 | Subclasses may override to apply additional config changes. |
| 1269 | """ |
| 1270 | spec_cfg = self.speculative_config |
| 1271 | base = self.vllm_config |
| 1272 | |
| 1273 | if spec_cfg.moe_backend is not None: |
| 1274 | base = replace( |
| 1275 | base, |
| 1276 | kernel_config=replace( |
| 1277 | base.kernel_config, |
| 1278 | moe_backend=spec_cfg.moe_backend, |
| 1279 | ), |
| 1280 | ) |
| 1281 | |
| 1282 | # Note (matt): Never inherit the attention backend from base, because there are |
| 1283 | # many opportunities for incompatibility, so we always independently autoselect |
| 1284 | # unless explicitly specified in the speculative config. |
| 1285 | base = replace( |
| 1286 | base, |
| 1287 | attention_config=replace( |
| 1288 | base.attention_config, |
| 1289 | backend=spec_cfg.attention_backend, |
| 1290 | ), |
| 1291 | ) |
| 1292 | |
| 1293 | if spec_cfg.kv_cache_dtype is not None: |
| 1294 | base = replace( |
| 1295 | base, |
| 1296 | cache_config=replace( |
| 1297 | base.cache_config, |
| 1298 | cache_dtype=spec_cfg.kv_cache_dtype, |
| 1299 | ), |
| 1300 | ) |
| 1301 | |
| 1302 | return base |
| 1303 | |
| 1304 | def _get_model(self) -> nn.Module: |
| 1305 | """Default method to call get_model(). Can be overridden by subclasses which |
| 1306 | need to customize model loading. |
| 1307 | """ |
| 1308 | from vllm.compilation.backends import set_model_tag |
| 1309 | |
| 1310 | draft_vllm_config = self._create_draft_vllm_config() |
| 1311 | with set_model_tag("eagle_head"): |
| 1312 | model = get_model( |
| 1313 | vllm_config=draft_vllm_config, |
| 1314 | model_config=self.speculative_config.draft_model_config, |
| 1315 | load_config=get_draft_load_config(draft_vllm_config), |
| 1316 | ) |
| 1317 | return model |
| 1318 | |
| 1319 | def load_model(self, target_model: nn.Module) -> None: |
| 1320 | target_attn_layer_names = set( |
| 1321 | get_layers_from_vllm_config( |
| 1322 | self.vllm_config, |
| 1323 | AttentionLayerBase, # type: ignore[type-abstract] |
| 1324 | ).keys() |
| 1325 | ) |
| 1326 | |
| 1327 | self.model = self._get_model() |
| 1328 | |
| 1329 | # Find draft layers (attention layers added by draft model) |
| 1330 | all_attn_layers = get_layers_from_vllm_config( |
| 1331 | self.vllm_config, |
| 1332 | AttentionLayerBase, # type: ignore[type-abstract] |
| 1333 | ) |
| 1334 | # Filter to only layers that have KV cache specs. |
| 1335 | self._draft_attn_layer_names = { |
| 1336 | name |
| 1337 | for name in (set(all_attn_layers.keys()) - target_attn_layer_names) |
| 1338 | if all_attn_layers[name].get_kv_cache_spec(self.vllm_config) is not None |
| 1339 | } |
| 1340 | |
| 1341 | # Even if the target model is multimodal, we can also use |
| 1342 | # text-only draft models |
| 1343 | if self.supports_mm_inputs and not supports_multimodal_embeddings(self.model): |
| 1344 | logger.warning_once( |
| 1345 | "Draft model %s does not support external multimodal embeddings. " |
| 1346 | "Embeddings from the target model will not be passed to the " |
| 1347 | "drafter; using text-only draft inputs instead.", |
| 1348 | type(self.model).__name__, |
| 1349 | ) |
| 1350 | self.supports_mm_inputs = False |
| 1351 | |
| 1352 | if supports_multimodal(target_model): |
| 1353 | # handle multimodality |
| 1354 | assert hasattr(target_model, "config") |
| 1355 | if self.get_model_name(target_model) in [ |
| 1356 | "BailingMoeV3VLForConditionalGeneration", |
| 1357 | "Cohere2VisionForConditionalGeneration", |
| 1358 | "Exaone4_5_ForConditionalGeneration", |
| 1359 | "GlmOcrForConditionalGeneration", |
| 1360 | "HunYuanVLForConditionalGeneration", |
| 1361 | "InternS2PreviewForConditionalGeneration", |
| 1362 | "MiMoV2OmniForCausalLM", |
| 1363 | "Qwen2_5_VLForConditionalGeneration", |
| 1364 | "Qwen3_5ForConditionalGeneration", |
| 1365 | "Qwen3_5MoeForConditionalGeneration", |
| 1366 | "Qwen3VLForConditionalGeneration", |
| 1367 | "Qwen3VLMoeForConditionalGeneration", |
| 1368 | "Gemma4ForConditionalGeneration", |
| 1369 | "Gemma4UnifiedForConditionalGeneration", |
| 1370 | "Step3p7ForConditionalGeneration", |
| 1371 | ]: |
| 1372 | self.model.config.image_token_index = target_model.config.image_token_id |
| 1373 | elif self.get_model_name(target_model) == "PixtralForConditionalGeneration": |
| 1374 | self.model.config.image_token_index = ( |
| 1375 | target_model.config.vision_config.image_token_id |
| 1376 | ) |
| 1377 | elif self.get_model_name(target_model) in ( |
| 1378 | "KimiK25ForConditionalGeneration", |
| 1379 | "KimiK3ForConditionalGeneration", |
| 1380 | ): |
| 1381 | self.model.config.image_token_index = ( |
| 1382 | target_model.config.media_placeholder_token_id |
| 1383 | ) |
| 1384 | elif self.get_model_name(target_model) == "NemotronH_Nano_VL_V2": |
| 1385 | self.model.config.image_token_index = ( |
| 1386 | target_model.config.img_context_token_id |
| 1387 | ) |
| 1388 | else: |
| 1389 | self.model.config.image_token_index = ( |
| 1390 | target_model.config.image_token_index |
| 1391 | ) |
| 1392 | target_language_model = cast( |
| 1393 | SupportsMultiModal, target_model |
| 1394 | ).get_language_model() |
| 1395 | else: |
| 1396 | target_language_model = target_model |
| 1397 | |
| 1398 | self._maybe_share_embeddings(target_language_model) |
| 1399 | self._maybe_share_lm_head(target_language_model) |
| 1400 | |
| 1401 | if ( |
| 1402 | self.parallel_drafting |
| 1403 | and self.pass_hidden_states_to_model |
| 1404 | and self.parallel_drafting_hidden_state_tensor is not None |
| 1405 | ): |
| 1406 | flat_mask = self.model.mask_hidden.view(-1) |
| 1407 | if self.eagle3_use_aux_hidden_state: |
| 1408 | # EAGLE3: mask_hidden stores all aux hidden states, |
| 1409 | # project through combine_hidden_states |
| 1410 | self.parallel_drafting_hidden_state_tensor.copy_( |
| 1411 | self.model.combine_hidden_states(flat_mask) |
| 1412 | ) |
| 1413 | else: |
| 1414 | self.parallel_drafting_hidden_state_tensor.copy_(flat_mask) |
| 1415 | |
| 1416 | def _maybe_share_embeddings(self, target_language_model: nn.Module) -> None: |
| 1417 | """Some draft models may not have their own embedding layers, and some may |
| 1418 | have a duplicate copy of the target model's embedding layers. In these cases, |
| 1419 | we share the target model's embedding layers with the draft model to save |
| 1420 | memory. |
| 1421 | """ |
| 1422 | if get_pp_group().world_size == 1: |
| 1423 | inner_model = getattr(target_language_model, "model", None) |
| 1424 | if inner_model is None: |
| 1425 | raise AttributeError("Target model does not have 'model' attribute") |
| 1426 | if hasattr(inner_model, "embed_tokens"): |
| 1427 | target_embed_tokens = inner_model.embed_tokens |
| 1428 | elif hasattr(inner_model, "embedding"): |
| 1429 | target_embed_tokens = inner_model.embedding |
| 1430 | else: |
| 1431 | raise AttributeError( |
| 1432 | "Target model does not have 'embed_tokens' or 'embedding' attribute" |
| 1433 | ) |
| 1434 | |
| 1435 | share_embeddings = False |
| 1436 | if hasattr(self.model, "has_own_embed_tokens"): |
| 1437 | # EAGLE model |
| 1438 | if not self.model.has_own_embed_tokens: |
| 1439 | share_embeddings = True |
| 1440 | logger.info( |
| 1441 | "Detected EAGLE model without its own embed_tokens in the" |
| 1442 | " checkpoint. Sharing target model embedding weights with the" |
| 1443 | " draft model." |
| 1444 | ) |
| 1445 | elif ( |
| 1446 | isinstance(target_embed_tokens.weight, torch.Tensor) |
| 1447 | and isinstance(self.model.model.embed_tokens.weight, torch.Tensor) |
| 1448 | # TODO: Offload to CPU for comparison to avoid extra GPU memory |
| 1449 | # usage in CI testing environments with limited GPU memory |
| 1450 | and torch.equal( |
| 1451 | target_embed_tokens.weight.cpu(), |
| 1452 | self.model.model.embed_tokens.weight.cpu(), |
| 1453 | ) |
| 1454 | ): |
| 1455 | share_embeddings = True |
| 1456 | logger.info( |
| 1457 | "Detected EAGLE model with embed_tokens identical to the target" |
| 1458 | " model. Sharing target model embedding weights with the draft" |
| 1459 | " model." |
| 1460 | ) |
| 1461 | else: |
| 1462 | logger.info( |
| 1463 | "Detected EAGLE model with distinct embed_tokens weights. " |
| 1464 | "Keeping separate embedding weights from the target model." |
| 1465 | ) |
| 1466 | else: |
| 1467 | # MTP model |
| 1468 | share_embeddings = True |
| 1469 | logger.info( |
| 1470 | "Detected MTP model. " |
| 1471 | "Sharing target model embedding weights with the draft model." |
| 1472 | ) |
| 1473 | |
| 1474 | if share_embeddings and hasattr(self.model, "has_own_embed_tokens"): |
| 1475 | # EAGLE drafts consume input embeddings at their own hidden |
| 1476 | # size, so only share when the widths match. MTP drafts |
| 1477 | # project target-width embeddings (e.g. Gemma4 MTP's |
| 1478 | # pre_projection takes 2 * backbone_hidden_size), so the |
| 1479 | # width check does not apply to them. |
| 1480 | draft_embed = self.model.model.embed_tokens |
| 1481 | # Guard with isinstance so non-Tensor weights (e.g. in tests) |
| 1482 | # are not affected — mirrors the weight-equality check above. |
| 1483 | if isinstance(target_embed_tokens.weight, torch.Tensor) and isinstance( |
| 1484 | draft_embed.weight, torch.Tensor |
| 1485 | ): |
| 1486 | target_dim = target_embed_tokens.weight.shape[-1] |
| 1487 | draft_dim = draft_embed.weight.shape[-1] |
| 1488 | if target_dim != draft_dim: |
| 1489 | share_embeddings = False |
| 1490 | logger.info( |
| 1491 | "Target embedding dim (%d) differs from draft " |
| 1492 | "embedding dim (%d). Keeping separate embedding " |
| 1493 | "weights.", |
| 1494 | target_dim, |
| 1495 | draft_dim, |
| 1496 | ) |
| 1497 | |
| 1498 | if share_embeddings: |
| 1499 | if hasattr(self.model.model, "embed_tokens"): |
| 1500 | del self.model.model.embed_tokens |
| 1501 | self.model.model.embed_tokens = target_embed_tokens |
| 1502 | else: |
| 1503 | logger.info( |
| 1504 | "The draft model's vocab embedding will be loaded separately" |
| 1505 | " from the target model." |
| 1506 | ) |
| 1507 | |
| 1508 | def _maybe_share_lm_head(self, target_language_model: nn.Module) -> None: |
| 1509 | """Some draft models may not have their own LM head, and some may have a |
| 1510 | duplicate copy of the target model's LM head. In these cases, we share |
| 1511 | the target model's LM head with the draft model to save memory. |
| 1512 | """ |
| 1513 | share_lm_head = False |
| 1514 | if hasattr(self.model, "has_own_lm_head"): |
| 1515 | # EAGLE model |
| 1516 | if not self.model.has_own_lm_head: |
| 1517 | share_lm_head = True |
| 1518 | logger.info( |
| 1519 | "Detected EAGLE model without its own lm_head in the checkpoint. " |
| 1520 | "Sharing target model lm_head weights with the draft model." |
| 1521 | ) |
| 1522 | elif ( |
| 1523 | hasattr(target_language_model, "lm_head") |
| 1524 | and hasattr(target_language_model.lm_head, "weight") |
| 1525 | and hasattr(self.model.lm_head, "weight") |
| 1526 | and isinstance(target_language_model.lm_head.weight, torch.Tensor) |
| 1527 | and isinstance(self.model.lm_head.weight, torch.Tensor) |
| 1528 | # TODO: Offload to CPU for comparison to avoid extra GPU memory |
| 1529 | # usage in CI testing environments with limited GPU memory |
| 1530 | and torch.equal( |
| 1531 | target_language_model.lm_head.weight.cpu(), |
| 1532 | self.model.lm_head.weight.cpu(), |
| 1533 | ) |
| 1534 | ): |
| 1535 | share_lm_head = True |
| 1536 | logger.info( |
| 1537 | "Detected EAGLE model with lm_head identical to the target model. " |
| 1538 | "Sharing target model lm_head weights with the draft model." |
| 1539 | ) |
| 1540 | else: |
| 1541 | logger.info( |
| 1542 | "Detected EAGLE model with distinct lm_head weights. " |
| 1543 | "Keeping separate lm_head weights from the target model." |
| 1544 | ) |
| 1545 | else: |
| 1546 | # MTP model |
| 1547 | share_lm_head = True |
| 1548 | logger.info( |
| 1549 | "Detected MTP model. " |
| 1550 | "Sharing target model lm_head weights with the draft model." |
| 1551 | ) |
| 1552 | |
| 1553 | if share_lm_head and hasattr(target_language_model, "lm_head"): |
| 1554 | if hasattr(self.model, "lm_head"): |
| 1555 | del self.model.lm_head |
| 1556 | self.model.lm_head = target_language_model.lm_head |
| 1557 | |
| 1558 | # MTP models call compute_logits via shared_head.head (a |
| 1559 | # ParallelLMHead inside each MTP layer), not self.model.lm_head. |
| 1560 | # If the checkpoint omits a copy of the lm_head weights at the |
| 1561 | # MTP layer path, shared_head.head stays uninitialised and |
| 1562 | # produces NaN logits. Always share it explicitly. |
| 1563 | inner = getattr(self.model, "model", None) |
| 1564 | layers = getattr(inner, "layers", None) if inner else None |
| 1565 | if layers is not None: |
| 1566 | items = layers.values() if isinstance(layers, nn.ModuleDict) else layers |
| 1567 | for layer in items: |
| 1568 | sh = getattr(layer, "shared_head", None) |
| 1569 | if sh is not None and hasattr(sh, "head"): |
| 1570 | del sh.head |
| 1571 | sh.head = target_language_model.lm_head |
| 1572 | logger.info( |
| 1573 | "Shared target model lm_head with MTP shared_head.head." |
| 1574 | ) |
| 1575 | |
| 1576 | if hasattr(target_language_model.model, "topk_indices_buffer"): |
| 1577 | target_buffer = target_language_model.model.topk_indices_buffer |
| 1578 | if hasattr(self.model.model, "topk_indices_buffer"): |
| 1579 | del self.model.model.topk_indices_buffer |
| 1580 | self.model.model.topk_indices_buffer = target_buffer |
| 1581 | # Also share at per-module level so that the indexer and |
| 1582 | # sparse-attention backends in each MTP layer read from |
| 1583 | # the target model's buffer. |
| 1584 | for _, module in self.model.model.named_modules(): |
| 1585 | if hasattr(module, "topk_indices_buffer"): |
| 1586 | module.topk_indices_buffer = target_buffer |
| 1587 | logger.info( |
| 1588 | "Detected MTP model with topk_indices_buffer. " |
| 1589 | "Sharing target model topk_indices_buffer with the draft model." |
| 1590 | ) |
| 1591 | |
| 1592 | # Detect index_share_for_mtp_iteration: when True, the proposer |
| 1593 | # toggles skip_topk so step 0 computes MTP's own indices and |
| 1594 | # steps 1+ reuse them. |
| 1595 | spec_config = self.vllm_config.speculative_config |
| 1596 | draft_hf_config = ( |
| 1597 | spec_config.draft_model_config.hf_config |
| 1598 | if spec_config is not None |
| 1599 | else None |
| 1600 | ) |
| 1601 | self._share_mtp_indices = getattr( |
| 1602 | draft_hf_config, "index_share_for_mtp_iteration", False |
| 1603 | ) |
| 1604 | |
| 1605 | if self.use_local_argmax_reduction: |
| 1606 | if not hasattr(self.model, "get_top_tokens"): |
| 1607 | raise ValueError( |
| 1608 | "use_local_argmax_reduction is enabled but draft model " |
| 1609 | f"{self.model.__class__.__name__} does not implement " |
| 1610 | "get_top_tokens()." |
| 1611 | ) |
| 1612 | logger.info( |
| 1613 | "Using local argmax reduction for draft token generation " |
| 1614 | "(communication: O(2*tp_size) vs O(vocab_size))." |
| 1615 | ) |
| 1616 | |
| 1617 | @torch.inference_mode() |
| 1618 | def dummy_run( |
| 1619 | self, |
| 1620 | num_tokens: int, |
| 1621 | use_cudagraphs: bool = True, |
| 1622 | is_graph_capturing: bool = False, |
| 1623 | slot_mappings: dict[str, torch.Tensor] | None = None, |
| 1624 | ) -> None: |
| 1625 | # FIXME: when using tree-based specdec, adjust number of forward-passes |
| 1626 | # according to the depth of the tree. |
| 1627 | only_one_forward_pass = is_graph_capturing or self.parallel_drafting |
| 1628 | for fwd_idx in range( |
| 1629 | 1 if only_one_forward_pass else self.num_speculative_tokens |
| 1630 | ): |
| 1631 | if fwd_idx <= 1: |
| 1632 | cudagraph_runtime_mode, num_input_tokens, num_tokens_across_dp = ( |
| 1633 | self._determine_batch_execution_and_padding( |
| 1634 | num_tokens, use_cudagraphs=use_cudagraphs |
| 1635 | ) |
| 1636 | ) |
| 1637 | |
| 1638 | # Make sure to use EAGLE's own buffer during cudagraph capture. |
| 1639 | if ( |
| 1640 | self._draft_attn_layer_names |
| 1641 | and slot_mappings is not None |
| 1642 | and next(iter(self._draft_attn_layer_names)) in slot_mappings |
| 1643 | ): |
| 1644 | slot_mapping_dict = self._get_slot_mapping(num_input_tokens) |
| 1645 | else: |
| 1646 | slot_mapping_dict = slot_mappings or {} |
| 1647 | |
| 1648 | with set_forward_context( |
| 1649 | None, |
| 1650 | self.vllm_config, |
| 1651 | num_tokens=num_input_tokens, |
| 1652 | num_tokens_across_dp=num_tokens_across_dp, |
| 1653 | cudagraph_runtime_mode=cudagraph_runtime_mode, |
| 1654 | slot_mapping=slot_mapping_dict, |
| 1655 | ): |
| 1656 | if self.supports_mm_inputs: |
| 1657 | input_ids = None |
| 1658 | inputs_embeds = self.inputs_embeds[:num_input_tokens] |
| 1659 | else: |
| 1660 | input_ids = self.input_ids[:num_input_tokens] |
| 1661 | inputs_embeds = None |
| 1662 | |
| 1663 | kwargs = dict( |
| 1664 | input_ids=input_ids, |
| 1665 | positions=self._get_positions(num_input_tokens), |
| 1666 | inputs_embeds=inputs_embeds, |
| 1667 | ) |
| 1668 | if self.pass_hidden_states_to_model: |
| 1669 | kwargs["hidden_states"] = self.hidden_states[:num_input_tokens] |
| 1670 | self.model(**kwargs) |
| 1671 | |
| 1672 | def _get_eagle3_use_aux_hidden_state_from_config(self) -> bool: |
| 1673 | """Some eagle3 heads (e.g., nvidia/gpt-oss-120b-Eagle3-v2) do not use auxiliary |
| 1674 | hidden states and directly uses the last layer output just like eagle1. |
| 1675 | They might indicate this by setting "use_aux_hidden_state" to False |
| 1676 | inside the "eagle_config" dict of their hf_config. |
| 1677 | """ |
| 1678 | if self.method != "eagle3": |
| 1679 | return False |
| 1680 | # Assume that eagle3 heads use aux hidden states by default |
| 1681 | use_aux_hidden_state = True |
| 1682 | eagle_config = getattr(self.draft_model_config.hf_config, "eagle_config", None) |
| 1683 | if eagle_config is not None: |
| 1684 | use_aux_hidden_state = eagle_config.get("use_aux_hidden_state", True) |
| 1685 | return use_aux_hidden_state |
| 1686 | |
| 1687 | def validate_same_kv_cache_group(self, kv_cache_config: KVCacheConfig) -> None: |
| 1688 | """Validate that all drafting layers belong to the same KVCacheGroup. |
| 1689 | Need this assumption to ensure all drafting layers can use the |
| 1690 | same AttentionMetadata. |
| 1691 | May extend to multiple AttentionMetadata in the future. |
| 1692 | """ |
| 1693 | kv_cache_groups: dict[str, int] = {} |
| 1694 | for id, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups): |
| 1695 | for layer_name in kv_cache_group.layer_names: |
| 1696 | kv_cache_groups[layer_name] = id |
| 1697 | assert ( |
| 1698 | len( |
| 1699 | set( |
| 1700 | [ |
| 1701 | kv_cache_groups[layer_name] |
| 1702 | for layer_name in self._draft_attn_layer_names |
| 1703 | ] |
| 1704 | ) |
| 1705 | ) |
| 1706 | == 1 |
| 1707 | ), "All drafting layers should belong to the same kv cache group" |
| 1708 | |
| 1709 | def initialize_attn_backend( |
| 1710 | self, |
| 1711 | kv_cache_config: KVCacheConfig, |
| 1712 | kernel_block_sizes: list[int] | None = None, |
| 1713 | ) -> None: |
| 1714 | """Initialize AttentionGroups for draft layers using kv_cache_config. |
| 1715 | Called from the model runner's initialize_metadata_builders. |
| 1716 | """ |
| 1717 | all_attn_layers = get_layers_from_vllm_config( |
| 1718 | self.vllm_config, |
| 1719 | AttentionLayerBase, # type: ignore[type-abstract] |
| 1720 | ) |
| 1721 | |
| 1722 | # Find which kv_cache_group the draft layers belong to |
| 1723 | self.validate_same_kv_cache_group(kv_cache_config) |
| 1724 | kv_cache_spec = None |
| 1725 | for gid, group in enumerate(kv_cache_config.kv_cache_groups): |
| 1726 | if self._draft_attn_layer_names & set(group.layer_names): |
| 1727 | self.kv_cache_gid = gid |
| 1728 | kv_cache_spec = group.kv_cache_spec |
| 1729 | break |
| 1730 | |
| 1731 | attention_groups: dict[tuple[str, str], AttentionGroup] = {} |
| 1732 | if kv_cache_spec is not None: |
| 1733 | # _draft_attn_layer_names is a set; iterate in sorted order so |
| 1734 | # that attention_groups (and anything derived from its first |
| 1735 | # element) is deterministic across processes. |
| 1736 | for layer_name in sorted(self._draft_attn_layer_names): |
| 1737 | attn_backend = all_attn_layers[layer_name].get_attn_backend() |
| 1738 | backend_key = attn_backend.full_cls_name() |
| 1739 | if backend_key not in attention_groups: |
| 1740 | layer_kv_cache_spec = kv_cache_spec |
| 1741 | if isinstance(layer_kv_cache_spec, UniformTypeKVCacheSpecs): |
| 1742 | layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[ |
| 1743 | layer_name |
| 1744 | ] |
| 1745 | |
| 1746 | kernel_block_size = ( |
| 1747 | kernel_block_sizes[self.kv_cache_gid] |
| 1748 | if kernel_block_sizes is not None |
| 1749 | and self.kv_cache_gid < len(kernel_block_sizes) |
| 1750 | else None |
| 1751 | ) |
| 1752 | attn_group = AttentionGroup( |
| 1753 | backend=attn_backend, |
| 1754 | layer_names=[layer_name], |
| 1755 | kv_cache_spec=layer_kv_cache_spec, |
| 1756 | kv_cache_group_id=self.kv_cache_gid, |
| 1757 | ) |
| 1758 | attn_group.create_metadata_builders( |
| 1759 | self.vllm_config, |
| 1760 | self.device, |
| 1761 | kernel_block_size=kernel_block_size, |
| 1762 | ) |
| 1763 | attention_groups[backend_key] = attn_group |
| 1764 | else: |
| 1765 | attention_groups[backend_key].layer_names.append(layer_name) |
| 1766 | |
| 1767 | self.draft_attn_groups = list(attention_groups.values()) |
| 1768 | if kernel_block_sizes is not None and 0 <= self.kv_cache_gid < len( |
| 1769 | kernel_block_sizes |
| 1770 | ): |
| 1771 | # Slot mappings are computed against the block table, which is |
| 1772 | # stored at kernel-block granularity. Use the kernel block size |
| 1773 | # rather than the KV cache manager's block size; the two differ |
| 1774 | # when manager blocks are split for the attention kernel. |
| 1775 | self.block_size = kernel_block_sizes[self.kv_cache_gid] |
| 1776 | else: |
| 1777 | self.block_size = ( |
| 1778 | self.draft_attn_groups[0] |
| 1779 | .get_metadata_builder() |
| 1780 | .kv_cache_spec.block_size |
| 1781 | ) |
| 1782 | logger.debug("Using block size %d for drafting layers", self.block_size) |
| 1783 | |
| 1784 | def _determine_batch_execution_and_padding( |
| 1785 | self, |
| 1786 | num_tokens: int, |
| 1787 | use_cudagraphs: bool = True, |
| 1788 | ) -> tuple[CUDAGraphMode, int, torch.Tensor | None]: |
| 1789 | cudagraph_mode, batch_desc = self.cudagraph_dispatcher.dispatch( |
| 1790 | num_tokens, |
| 1791 | valid_modes=({CUDAGraphMode.NONE} if not use_cudagraphs else None), |
| 1792 | ) |
| 1793 | num_tokens_padded = batch_desc.num_tokens |
| 1794 | |
| 1795 | # Extra coordination when running data-parallel since we need to |
| 1796 | # coordinate across ranks |
| 1797 | # TODO(Flechman): support DBO ubatching |
| 1798 | should_ubatch, num_tokens_across_dp = False, None |
| 1799 | if self.vllm_config.parallel_config.data_parallel_size > 1: |
| 1800 | should_ubatch, num_tokens_across_dp, synced_cudagraph_mode = ( |
| 1801 | coordinate_batch_across_dp( |
| 1802 | num_tokens_unpadded=num_tokens, |
| 1803 | parallel_config=self.vllm_config.parallel_config, |
| 1804 | allow_microbatching=False, |
| 1805 | num_tokens_padded=num_tokens_padded, |
| 1806 | cudagraph_mode=cudagraph_mode.value, |
| 1807 | ) |
| 1808 | ) |
| 1809 | assert not should_ubatch, "DBO ubatching not implemented for EAGLE" |
| 1810 | |
| 1811 | # Extract DP-synced values |
| 1812 | if num_tokens_across_dp is not None: |
| 1813 | dp_rank = self.dp_rank |
| 1814 | num_tokens_padded = int(num_tokens_across_dp[dp_rank].item()) |
| 1815 | # Re-dispatch with DP padding so we have the correct |
| 1816 | # batch_descriptor |
| 1817 | cudagraph_mode, batch_desc = self.cudagraph_dispatcher.dispatch( |
| 1818 | num_tokens_padded, |
| 1819 | valid_modes={CUDAGraphMode(synced_cudagraph_mode)}, |
| 1820 | ) |
| 1821 | # Assert to make sure the agreed upon token count is correct |
| 1822 | # otherwise num_tokens_across_dp will no-longer be valid |
| 1823 | assert batch_desc.num_tokens == num_tokens_padded |
| 1824 | num_tokens_across_dp[dp_rank] = num_tokens_padded |
| 1825 | |
| 1826 | return cudagraph_mode, num_tokens_padded, num_tokens_across_dp |
| 1827 | |
| 1828 | |
| 1829 | # NOTE(woosuk): Currently, the below code is not used and we always use argmax |
| 1830 | # to sample the draft tokens. We will use this after we find a way to manage |
| 1831 | # the draft prob tensor. |
| 1832 | # Refer to https://github.com/vllm-project/vllm/pull/16899 for the details. |
| 1833 | # FIXME(woosuk): The logic here is duplicated with the main sampling code. |
| 1834 | # We should refactor this to reuse the same sampling implementation. |
| 1835 | def compute_probs_and_sample_next_token( |
| 1836 | logits: torch.Tensor, |
| 1837 | sampling_metadata: SamplingMetadata, |
| 1838 | use_fp64_gumbel: bool = False, |
| 1839 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 1840 | if sampling_metadata.all_greedy: |
| 1841 | # For greedy requests, draft_probs is not used in rejection sampling. |
| 1842 | # Therefore, we can just return the logits. |
| 1843 | probs = logits |
| 1844 | next_token_ids = logits.argmax(dim=-1) |
| 1845 | return next_token_ids, probs |
| 1846 | |
| 1847 | assert sampling_metadata.temperature is not None |
| 1848 | |
| 1849 | # Use epsilon comparison to detect greedy sampling (temperature ~ 0.0) |
| 1850 | # consistent with sampler.py's _SAMPLING_EPS threshold |
| 1851 | temperature = sampling_metadata.temperature |
| 1852 | # Avoid division by zero if there are greedy requests. |
| 1853 | if not sampling_metadata.all_random: |
| 1854 | is_greedy = temperature < _SAMPLING_EPS |
| 1855 | temperature = torch.where(is_greedy, 1.0, temperature) |
| 1856 | logits.div_(temperature.view(-1, 1)) |
| 1857 | probs = logits.softmax(dim=-1, dtype=torch.float32) |
| 1858 | |
| 1859 | # NOTE(woosuk): Currently, we ignore most of the sampling parameters in |
| 1860 | # generating the draft tokens. We only use the temperature. While this |
| 1861 | # could degrade the acceptance rate, it does not affect the distribution |
| 1862 | # of the generated tokens after rejection sampling. |
| 1863 | |
| 1864 | # TODO(woosuk): Consider seeds. |
| 1865 | q = empty_exponential_noise_like(probs, use_fp64_gumbel) |
| 1866 | q.exponential_() |
| 1867 | # NOTE(woosuk): We shouldn't use `probs.div_(q)` because the draft_probs |
| 1868 | # will be used later for rejection sampling. |
| 1869 | next_token_ids = sample_with_exponential_noise(probs.clone(), q) |
| 1870 | if not sampling_metadata.all_random: |
| 1871 | greedy_token_ids = probs.argmax(dim=-1) |
| 1872 | next_token_ids = torch.where(is_greedy, greedy_token_ids, next_token_ids) |
| 1873 | return next_token_ids, probs |