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, 23:30 UTC
[V1][Misc] Avoid unnecessary log output (#13289)
Jee Jee Li committed
Feb 17, 2025 at 11:26 UTC
2010f04c17e76c7d1f70f6e1c9d3857a93036114
1 file changed
+7
-5
vllm/v1/worker/gpu_model_runner.py
+7
-5
@@ -96,11 +96,13 @@ class GPUModelRunner(LoRAModelRunnerMixin):
96
self.mm_registry = MULTIMODAL_REGISTRY
97
self.uses_mrope = model_config.uses_mrope
98
99
- # NOTE: Initialized client is only used for processing dummy
100
- # multimodal data into multimodal kwargs for GPU memory profiling.
101
- # Only applicable to multimodal models with legacy input mapper.
102
- self.mm_input_mapper_profiling = MMInputCacheClient(self.model_config)
103
- self.mm_input_mapper_profiling.use_cache = False
99
+ if self.is_multimodal_model:
100
+ # NOTE: Initialized client is only used for processing dummy
101
+ # multimodal data into multimodal kwargs for GPU memory profiling.
102
+ # Only applicable to multimodal models with legacy input mapper.
103
+ self.mm_input_mapper_profiling = MMInputCacheClient(
104
+ self.model_config)
105
+ self.mm_input_mapper_profiling.use_cache = False
106
107
encoder_compute_budget, encoder_cache_size = compute_encoder_budget(
108
model_config=model_config,