biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix] Qwen-vl output is inconsistent in speculative decoding (#10350)

Browse Source
This commit is contained in:
Sky Lee
2024-11-15 13:40:10 +08:00
committed by GitHub
parent b40cf6402e
commit 2ec8827288
1 changed files with 2 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/spec_decode/batch_expansion.py
View File

@@ -353,6 +353,7 @@ class BatchExpansionTop1Scorer(SpeculativeScorer):
seq_data = seq_group_metadata.seq_data[seq_id]
prompt_token_ids = seq_data.prompt_token_ids_array
new_output_token_ids = [*seq_data.get_output_token_ids(), *token_ids]
mrope_position_delta = seq_data.mrope_position_delta
new_seq_data_dict = {
target_seq_id:
@@ -368,6 +369,7 @@ class BatchExpansionTop1Scorer(SpeculativeScorer):
# the kv cache is filled by a previous batch in the batch expansion.
for data in new_seq_data_dict.values():
data.update_num_computed_tokens(data.get_len() - 1)
data.mrope_position_delta = mrope_position_delta
return SequenceGroupMetadata(
request_id=seq_group_metadata.request_id,