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, 09:45 UTC
[Frontend] Fix chat template content format detection (#18190)
Signed-off-by: Sebastian Schönnenbeck <sebastian.schoennenbeck@comma-soft.com>
Sebastian Schoennenbeck committed
May 15, 2025 at 18:00 UTC
2aa5470ac586f1603eb9d9b30d6c41ba3b5b9cd4
1 file changed
+3
-4
vllm/entrypoints/chat_utils.py
+3
-4
@@ -387,7 +387,6 @@ def resolve_hf_chat_template(
387
def _resolve_chat_template_content_format(
388
chat_template: Optional[str],
389
tools: Optional[list[dict[str, Any]]],
390
- given_format: ChatTemplateContentFormatOption,
390
tokenizer: AnyTokenizer,
391
*,
392
model_config: ModelConfig,
@@ -408,7 +407,7 @@ def _resolve_chat_template_content_format(
407
detected_format = ("string" if jinja_text is None else
408
_detect_content_format(jinja_text, default="string"))
409
411
- return detected_format if given_format == "auto" else given_format
410
+ return detected_format
411
412
413
@lru_cache
@@ -451,7 +450,6 @@ def resolve_chat_template_content_format(
450
detected_format = _resolve_chat_template_content_format(
451
chat_template,
452
tools,
454
- given_format,
453
tokenizer,
454
model_config=model_config,
455
)
@@ -462,7 +460,8 @@ def resolve_chat_template_content_format(
460
detected_format=detected_format,
461
)
462
465
- return detected_format
463
+ return detected_format if given_format == "auto" else given_format
464
+
465
466
467
ModalityStr = Literal["image", "audio", "video", "image_embeds"]