[Core] Support min_tokens with speculative decoding (#32642)

Signed-off-by: qianlihuang <yiliu.dong@qq.com>
Co-authored-by: qianlihuang <yiliu.dong@qq.com>
This commit is contained in:
Yiliu Dong
2026-02-26 12:31:28 -05:00
committed by GitHub
co-authored by qianlihuang
parent 99c7892c5b
commit d940607629
7 changed files with 75 additions and 11 deletions
+7
View File
@@ -10,6 +10,7 @@ import torch.nn as nn
from vllm.logger import init_logger
from vllm.triton_utils import tl, triton
from vllm.v1.outputs import LogprobsLists, LogprobsTensors, SamplerOutput
from vllm.v1.sample.logits_processor.builtin import MinTokensLogitsProcessor
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.sample.ops.bad_words import apply_bad_words_with_drafts
from vllm.v1.sample.ops.penalties import apply_all_penalties
@@ -292,6 +293,12 @@ class RejectionSampler(nn.Module):
logits, bad_words_token_ids, output_token_ids, metadata.num_draft_tokens
)
for processor in sampling_metadata.logitsprocs.non_argmax_invariant:
if isinstance(processor, MinTokensLogitsProcessor):
logits = processor.apply_with_spec_decode(
logits, metadata.num_draft_tokens
)
return logits
@staticmethod