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 15, 13:30 UTC
[Bugfix] fix error due to an uninitialized tokenizer when using `skip_tokenizer_init` with `num_scheduler_steps` (#9276)
Signed-off-by: changjun.lee <pord7457@gmail.com>
changjun.lee committed
Apr 27, 2025 at 00:51 UTC
10fd1d7380a42950822447cb9326c3550773b358
1 file changed
+1
-1
vllm/engine/output_processor/multi_step.py
+1
-1
@@ -178,7 +178,7 @@ class MultiStepOutputProcessor(SequenceGroupOutputProcessor):
178
# generates a fixed number of tokens without evaluating stopping
179
# conditions within the block. This can cause an eos token to be
180
# unintentionally ignored.
181
- if not sampling_params.ignore_eos:
181
+ if not sampling_params.ignore_eos and self.detokenizer:
182
eos_token_id = self.get_tokenizer_for_seq(seq).eos_token_id
183
# Avoiding .index calls as exception throwing in the happy path
184
# is expensive.