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
[Misc] Add warning for multimodal data in LLM.beam_search (#16241)
Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>
Alex Brooks committed
Apr 8, 2025 at 05:05 UTC
69ecaa7c7916499730f50813c6435c8f11e65c52
1 file changed
+10
vllm/entrypoints/llm.py
+10
@@ -536,6 +536,16 @@ class LLM:
536
tokenizer.eos_token_id,
537
length_penalty)
538
539
+ # TODO - fix handling of multimodal data for beam search; we pass it
540
+ # through in the async version on the abstract EngineClient, but not
541
+ # here.
542
+ if any("multi_modal_data" in prompt
543
+ and prompt["multi_modal_data"] is not None
544
+ for prompt in prompts):
545
+ logger.warning(
546
+ "Multimodal data appears to have been provided, but is not"
547
+ " currently being passed through in LLM.beam_search()!")
548
+
549
tokenizer = self.get_tokenizer()
550
# generate 2 * beam_width candidates at each step
551
# following the huggingface transformers implementation