| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from typing import Any |
| 4 | |
| 5 | import torch |
| 6 | import torch.nn as nn |
| 7 | |
| 8 | from vllm.compilation.backends import set_model_tag |
| 9 | from vllm.config import VllmConfig |
| 10 | from vllm.config.compilation import CUDAGraphMode |
| 11 | from vllm.forward_context import set_forward_context |
| 12 | from vllm.model_executor.model_loader import get_model |
| 13 | from vllm.v1.worker.gpu.dp_utils import DPSyncState |
| 14 | from vllm.v1.worker.gpu.input_batch import InputBatch |
| 15 | from vllm.v1.worker.gpu.spec_decode.speculator import DraftModelSpeculator |
| 16 | |
| 17 | |
| 18 | class ExtractHiddenStatesSpeculator(DraftModelSpeculator): |
| 19 | """Cache target hidden states while returning always-accepted draft tokens.""" |
| 20 | |
| 21 | def __init__(self, vllm_config: VllmConfig, device: torch.device): |
| 22 | assert vllm_config.speculative_config is not None |
| 23 | if vllm_config.speculative_config.draft_sample_method != "greedy": |
| 24 | raise ValueError( |
| 25 | "extract_hidden_states only supports draft_sample_method='greedy'" |
| 26 | ) |
| 27 | super().__init__(vllm_config, device) |
| 28 | |
| 29 | if self.num_speculative_steps != 1: |
| 30 | raise ValueError( |
| 31 | "extract_hidden_states requires num_speculative_tokens to be 1" |
| 32 | ) |
| 33 | if self.speculative_config.disable_padded_drafter_batch: |
| 34 | raise ValueError( |
| 35 | "disable_padded_drafter_batch is not supported with " |
| 36 | "extract_hidden_states method" |
| 37 | ) |
| 38 | |
| 39 | self.supports_mm_inputs = False |
| 40 | layer_ids = getattr( |
| 41 | self.draft_model_config.hf_config, |
| 42 | "eagle_aux_hidden_state_layer_ids", |
| 43 | None, |
| 44 | ) |
| 45 | if not layer_ids: |
| 46 | raise ValueError( |
| 47 | "eagle_aux_hidden_state_layer_ids must be set in the draft " |
| 48 | "model config for extract_hidden_states method" |
| 49 | ) |
| 50 | |
| 51 | self.num_hidden_states = len(layer_ids) |
| 52 | assert isinstance(self.dtype, torch.dtype) |
| 53 | self.hidden_states = torch.zeros( |
| 54 | self.max_num_tokens, |
| 55 | self.num_hidden_states, |
| 56 | self.vllm_config.model_config.get_hidden_size(), |
| 57 | dtype=self.dtype, |
| 58 | device=device, |
| 59 | ) |
| 60 | |
| 61 | def load_draft_model( |
| 62 | self, |
| 63 | target_model: nn.Module, |
| 64 | target_attn_layer_names: set[str], |
| 65 | ) -> nn.Module: |
| 66 | del target_model, target_attn_layer_names |
| 67 | with set_model_tag("extract_hidden_states"): |
| 68 | return get_model( |
| 69 | vllm_config=self.vllm_config, |
| 70 | model_config=self.draft_model_config, |
| 71 | ) |
| 72 | |
| 73 | def load_model(self, target_model: nn.Module) -> None: |
| 74 | super().load_model(target_model) |
| 75 | if len(self.draft_attn_layer_names) != 1: |
| 76 | raise ValueError( |
| 77 | "ExtractHiddenStatesModel should have exactly one attention " |
| 78 | f"layer, found {len(self.draft_attn_layer_names)}" |
| 79 | ) |
| 80 | |
| 81 | def init_cudagraph_manager(self, cudagraph_mode: CUDAGraphMode) -> None: |
| 82 | del cudagraph_mode |
| 83 | |
| 84 | def capture(self) -> None: |
| 85 | return None |
| 86 | |
| 87 | @torch.inference_mode() |
| 88 | def propose( |
| 89 | self, |
| 90 | input_batch: InputBatch, |
| 91 | attn_metadata: dict[str, Any], |
| 92 | slot_mappings: dict[str, torch.Tensor], |
| 93 | last_hidden_states: torch.Tensor, |
| 94 | aux_hidden_states: list[torch.Tensor] | None, |
| 95 | num_sampled: torch.Tensor, |
| 96 | num_rejected: torch.Tensor, |
| 97 | last_sampled: torch.Tensor, |
| 98 | next_prefill_tokens: torch.Tensor, |
| 99 | temperature: torch.Tensor, |
| 100 | seeds: torch.Tensor, |
| 101 | dp_sync: DPSyncState | None = None, |
| 102 | dummy_run: bool = False, |
| 103 | skip_attn_for_dummy_run: bool = False, |
| 104 | mm_inputs: tuple[list[torch.Tensor], torch.Tensor] | None = None, |
| 105 | is_profile: bool = False, |
| 106 | ) -> torch.Tensor: |
| 107 | del ( |
| 108 | last_hidden_states, |
| 109 | num_sampled, |
| 110 | num_rejected, |
| 111 | next_prefill_tokens, |
| 112 | temperature, |
| 113 | seeds, |
| 114 | dummy_run, |
| 115 | mm_inputs, |
| 116 | is_profile, |
| 117 | ) |
| 118 | |
| 119 | draft_tokens = last_sampled[input_batch.idx_mapping, :1] |
| 120 | if skip_attn_for_dummy_run: |
| 121 | return draft_tokens |
| 122 | if aux_hidden_states is None: |
| 123 | raise ValueError( |
| 124 | "aux_hidden_states are required when using extract_hidden_states" |
| 125 | ) |
| 126 | if len(aux_hidden_states) != self.num_hidden_states: |
| 127 | raise ValueError( |
| 128 | f"Expected {self.num_hidden_states} auxiliary hidden states, " |
| 129 | f"got {len(aux_hidden_states)}" |
| 130 | ) |
| 131 | |
| 132 | stacked_hidden_states = torch.stack(aux_hidden_states, dim=1) |
| 133 | num_tokens = stacked_hidden_states.shape[0] |
| 134 | self.hidden_states[:num_tokens].copy_(stacked_hidden_states) |
| 135 | |
| 136 | draft_attn_metadata = { |
| 137 | name: attn_metadata[name] for name in self.draft_attn_layer_names |
| 138 | } |
| 139 | draft_slot_mappings = { |
| 140 | name: slot_mappings[name][:num_tokens] |
| 141 | for name in self.draft_attn_layer_names |
| 142 | } |
| 143 | with set_forward_context( |
| 144 | draft_attn_metadata, |
| 145 | self.vllm_config, |
| 146 | num_tokens=num_tokens, |
| 147 | num_tokens_across_dp=( |
| 148 | dp_sync.num_tokens_across_dp if dp_sync is not None else None |
| 149 | ), |
| 150 | cudagraph_runtime_mode=CUDAGraphMode.NONE, |
| 151 | slot_mapping=draft_slot_mappings, |
| 152 | is_padding=input_batch.is_padding[:num_tokens], |
| 153 | ): |
| 154 | self.model(hidden_states=self.hidden_states[:num_tokens]) |
| 155 | |
| 156 | return draft_tokens |