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, 14:45 UTC
[Bugfix] catch AssertionError in MistralTokenizer as ValueError (#16344)
Signed-off-by: Guillaume Calmettes <gcalmettes@scaleway.com>
Guillaume Calmettes committed
Apr 9, 2025 at 19:33 UTC
c3b51891371e659c92936697c622df1162fb6eff
1 file changed
+12
-5
vllm/entrypoints/chat_utils.py
+12
-5
@@ -1193,8 +1193,15 @@ def apply_mistral_chat_template(
1193
**kwargs,
1194
)
1195
1196
- return tokenizer.apply_chat_template(
1197
- messages=messages,
1198
- tools=tools,
1199
- **kwargs,
1200
- )
1196
+ try:
1197
+ return tokenizer.apply_chat_template(
1198
+ messages=messages,
1199
+ tools=tools,
1200
+ **kwargs,
1201
+ )
1202
+ # mistral-common uses assert statements to stop processing of input
1203
+ # if input does not comply with the expected format.
1204
+ # We convert those assertion errors to ValueErrors so they can be
1205
+ # are properly caught in the preprocessing_input step
1206
+ except AssertionError as e:
1207
+ raise ValueError from e