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, 09:30 UTC
Fix beam search eos (#9627)
Robert Shaw committed
Oct 28, 2024 at 02:59 UTC
feb92fbe4ab6803527df48658a87ebd00b99969f
1 file changed
+6
-1
vllm/engine/protocol.py
+6
-1
@@ -140,7 +140,12 @@ class EngineClient(ABC):
140
best_beams = sorted_completed[:beam_width]
141
142
for beam in best_beams:
143
- beam.text = tokenizer.decode(beam.tokens[tokenized_length:])
143
+ if (beam.tokens[-1] == tokenizer.eos_token_id and not ignore_eos):
144
+ # Skip the eos token in the text.
145
+ tokens = beam.tokens[tokenized_length:-1]
146
+ else:
147
+ tokens = beam.tokens[tokenized_length:]
148
+ beam.text = tokenizer.decode(tokens)
149
150
beam_search_output = RequestOutput(
151
request_id=request_id,