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 8, 19:00 UTC
[Bugfix] Fix DeepSeek v3.2 tokenizer outputting None issue (#33832)
Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Wei Zhao committed
Feb 5, 2026 at 18:50 UTC
91a07ff6187e7308794b2a4863ab9e1f821ed464
1 file changed
+4
vllm/tokenizers/detokenizer_utils.py
+4
@@ -157,6 +157,10 @@ def detokenize_incrementally(
157
)
158
if isinstance(new_tokens, str):
159
new_tokens = [new_tokens]
160
+ else:
161
+ # This is required to guard against out-of-vocab prompt token ids
162
+ # (for example when using dummy weights)
163
+ _replace_none_with_empty(new_tokens) # type: ignore[arg-type]
164
else:
165
new_tokens = [""]
166
output_tokens = prev_tokens + new_tokens