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] Revert max_prompt_len validation for decoder-only models. (#16741)
Signed-off-by: David Heineman <david@davidheineman.com>
David Heineman committed
Apr 16, 2025 at 21:33 UTC
607029e51513933fff7edabe4691c96a82cfd289
2 files changed
+2
-2
vllm/engine/llm_engine.py
+1
-1
@@ -2062,7 +2062,7 @@ class LLMEngine:
2062
raise ValueError(f"The {prompt_type} prompt cannot be empty")
2063
2064
max_prompt_len = self.model_config.max_model_len
2065
- if len(prompt_ids) >= max_prompt_len:
2065
+ if len(prompt_ids) > max_prompt_len:
2066
if prompt_type == "encoder" and model_config.is_multimodal_model:
2067
mm_registry = self.input_preprocessor.mm_registry
2068
mm_processor = mm_registry.create_processor(
vllm/v1/engine/processor.py
+1
-1
@@ -354,7 +354,7 @@ class Processor:
354
raise ValueError(f"Token id {max_input_id} is out of vocabulary")
355
356
max_prompt_len = self.model_config.max_model_len
357
- if len(prompt_ids) >= max_prompt_len:
357
+ if len(prompt_ids) > max_prompt_len:
358
if prompt_type == "encoder" and model_config.is_multimodal_model:
359
mm_registry = self.input_preprocessor.mm_registry
360
mm_processor = mm_registry.create_processor(