A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 07:00 UTC
[Bugfix] Qwen-vl output is inconsistent in speculative decoding (#10350)
Sky Lee committed
Nov 15, 2024 at 13:40 UTC
2ec88272881a49d40d91ae0cd858b19d22996c70
1 file changed
+2
vllm/spec_decode/batch_expansion.py
+2
@@ -353,6 +353,7 @@ class BatchExpansionTop1Scorer(SpeculativeScorer):
353
seq_data = seq_group_metadata.seq_data[seq_id]
354
prompt_token_ids = seq_data.prompt_token_ids_array
355
new_output_token_ids = [*seq_data.get_output_token_ids(), *token_ids]
356
+ mrope_position_delta = seq_data.mrope_position_delta
357
358
new_seq_data_dict = {
359
target_seq_id:
@@ -368,6 +369,7 @@ class BatchExpansionTop1Scorer(SpeculativeScorer):
369
# the kv cache is filled by a previous batch in the batch expansion.
370
for data in new_seq_data_dict.values():
371
data.update_num_computed_tokens(data.get_len() - 1)
372
+ data.mrope_position_delta = mrope_position_delta
373
374
return SequenceGroupMetadata(
375
request_id=seq_group_metadata.request_id,