vllm.spec_decode.smaller_tp_proposer_worker
SmallerTpProposerWorker
¶
Bases: ProposerWorkerBase
Class which allows a speculative draft model to run with smaller tensor parallel degree than target model. This reduces the communication overhead of small draft models.
To implement this feature, this class differs behavior based on is_dummy flag, where dummy means worker that does not participate draft generation. Participating workers use a smaller tp group by patching vLLM's tensor parallel group temporarily during forward passes of draft models.
Source code in vllm/spec_decode/smaller_tp_proposer_worker.py
27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 |
|
__init__
¶
__init__(worker: MultiStepWorker, draft_ranks: List[int])
Create a SmallerTpProposerWorker.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
worker
|
~MultiStepWorker
|
an |
required |
draft_ranks
|
List[int]
|
if this value is given, only the GPU ranks |
required |
Source code in vllm/spec_decode/smaller_tp_proposer_worker.py
_patch_tensor_parallel_group
¶
Temporarily patch the global tp group state with its own tp group state.
determine_num_available_blocks
¶
Source code in vllm/spec_decode/smaller_tp_proposer_worker.py
execute_model
¶
execute_model(
execute_model_req: Optional[ExecuteModelRequest] = None,
) -> List[SamplerOutput]
Source code in vllm/spec_decode/smaller_tp_proposer_worker.py
get_spec_proposals
¶
get_spec_proposals(
execute_model_req: ExecuteModelRequest,
seq_ids_with_bonus_token_in_last_step: Set[int],
) -> SpeculativeProposals
Produce speculations given an input batch of sequences. The number of speculative tokens per sequence is determined by max_proposal_len.
Source code in vllm/spec_decode/smaller_tp_proposer_worker.py
init_device
¶
Source code in vllm/spec_decode/smaller_tp_proposer_worker.py
initialize_cache
¶
load_model
¶
maybe_load_lm_head_weight
¶
maybe_load_lm_head_weight(lm_head_weight: Tensor) -> None
Source code in vllm/spec_decode/smaller_tp_proposer_worker.py
maybe_wrap_worker
classmethod
¶
Wrap the worker in a SmallerTpProposerWorker if necessary.
Source code in vllm/spec_decode/smaller_tp_proposer_worker.py
sampler_output
¶
sampler_output(
execute_model_req: ExecuteModelRequest,
sample_len: int,
seq_ids_with_bonus_token_in_last_step: Set[int],
) -> Tuple[List[SamplerOutput], bool]