| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import torch |
| 5 | import torch.nn as nn |
| 6 | from typing_extensions import override |
| 7 | |
| 8 | from vllm.config import VllmConfig |
| 9 | from vllm.config.utils import replace |
| 10 | from vllm.logger import init_logger |
| 11 | from vllm.model_executor.model_loader import get_model |
| 12 | from vllm.tokenizers.registry import get_tokenizer |
| 13 | from vllm.v1.spec_decode.llm_base_proposer import SpecDecodeBaseProposer |
| 14 | from vllm.v1.spec_decode.vocab_mapping import VocabMapping |
| 15 | |
| 16 | logger = init_logger(__name__) |
| 17 | |
| 18 | |
| 19 | class DraftModelProposer(SpecDecodeBaseProposer): |
| 20 | def __init__( |
| 21 | self, |
| 22 | vllm_config: VllmConfig, |
| 23 | device: torch.device, |
| 24 | runner=None, |
| 25 | ): |
| 26 | super().__init__( |
| 27 | vllm_config=vllm_config, |
| 28 | device=device, |
| 29 | pass_hidden_states_to_model=False, |
| 30 | runner=runner, |
| 31 | ) |
| 32 | self._raise_if_draft_tp_mismatch() |
| 33 | |
| 34 | self.use_heterogeneous_vocab = self.speculative_config.use_heterogeneous_vocab |
| 35 | |
| 36 | spec = self.speculative_config |
| 37 | if self.use_heterogeneous_vocab: |
| 38 | # Heterogeneous vocabularies: build a VocabMapping to translate |
| 39 | # token IDs between the two tokenizers and constrain draft logits |
| 40 | # to the intersection so rejection sampling stays lossless. |
| 41 | target_tokenizer = get_tokenizer( |
| 42 | spec.target_model_config.tokenizer, |
| 43 | trust_remote_code=spec.target_model_config.trust_remote_code, |
| 44 | ) |
| 45 | draft_tokenizer = get_tokenizer( |
| 46 | spec.draft_model_config.model, |
| 47 | trust_remote_code=spec.draft_model_config.trust_remote_code, |
| 48 | ) |
| 49 | self.vocab_mapping: VocabMapping | None = VocabMapping( |
| 50 | target_tokenizer=target_tokenizer, |
| 51 | draft_tokenizer=draft_tokenizer, |
| 52 | target_vocab_size=spec.target_model_config.get_vocab_size(), |
| 53 | draft_vocab_size=spec.draft_model_config.get_vocab_size(), |
| 54 | device=device, |
| 55 | ) |
| 56 | else: |
| 57 | self._raise_if_vocab_size_mismatch() |
| 58 | self.vocab_mapping = None |
| 59 | |
| 60 | def _raise_if_vocab_size_mismatch(self): |
| 61 | self.speculative_config.verify_equal_vocab_size_if_draft_model() |
| 62 | |
| 63 | def _raise_if_draft_tp_mismatch(self): |
| 64 | # Note(Tomas Ruiz) If we run the target model with TP > 1 and |
| 65 | # the draft model with TP = 1, then the different TP ranks collide. |
| 66 | # Specifically when all ranks compile the draft model on rank 0 |
| 67 | # (because TP=1), then the torch compile cache is overwritten and corrupted. |
| 68 | # We need a mechanism like this: https://github.com/vllm-project/vllm/pull/5414 |
| 69 | # To prevent this error, we assert that both TP sizes must be the same. |
| 70 | spec_cfg = self.speculative_config |
| 71 | tgt_tp = spec_cfg.target_parallel_config.tensor_parallel_size |
| 72 | draft_tp = spec_cfg.draft_parallel_config.tensor_parallel_size |
| 73 | if draft_tp != tgt_tp: |
| 74 | raise ValueError( |
| 75 | f"Currently, 'draft_tensor_parallel_size' and 'tensor_parallel_size' " |
| 76 | f"must be the same. Got {draft_tp} and {tgt_tp}. " |
| 77 | "Please pass 'draft_tensor_parallel_size' in the speculative_config." |
| 78 | ) |
| 79 | |
| 80 | @override |
| 81 | def _create_draft_vllm_config(self) -> VllmConfig: |
| 82 | base = super()._create_draft_vllm_config() |
| 83 | spec = self.speculative_config |
| 84 | target_parallel = base.parallel_config |
| 85 | |
| 86 | return replace( |
| 87 | base, |
| 88 | quant_config=None, |
| 89 | parallel_config=replace( |
| 90 | spec.draft_parallel_config, |
| 91 | rank=target_parallel.rank, |
| 92 | decode_context_parallel_size=( |
| 93 | target_parallel.decode_context_parallel_size |
| 94 | ), |
| 95 | cp_kv_cache_interleave_size=target_parallel.cp_kv_cache_interleave_size, |
| 96 | dcp_comm_backend=target_parallel.dcp_comm_backend, |
| 97 | dcp_q_replicate=target_parallel.dcp_q_replicate, |
| 98 | ), |
| 99 | model_config=spec.draft_model_config, |
| 100 | ) |
| 101 | |
| 102 | @override |
| 103 | def _get_model(self) -> nn.Module: |
| 104 | from vllm.compilation.backends import set_model_tag |
| 105 | |
| 106 | draft_vllm_config = self._create_draft_vllm_config() |
| 107 | with set_model_tag("draft_model"): |
| 108 | model = get_model( |
| 109 | vllm_config=draft_vllm_config, |
| 110 | prefix="draft_model", |
| 111 | ) |
| 112 | return model |
| 113 | |
| 114 | @override |
| 115 | def _maybe_share_embeddings(self, target_language_model: nn.Module) -> None: |
| 116 | # Draft models don't share embeddings with the target model |
| 117 | pass |
| 118 | |
| 119 | @override |
| 120 | def _maybe_share_lm_head(self, target_language_model: nn.Module) -> None: |
| 121 | # Draft models don't share lm_head with the target model |
| 122 | pass |