fix(ci): Reduce the free gpu memory fraction (#2433)

components/backends/trtllm/engine_configs/agg.yaml

@@ -22,7 +22,7 @@ backend: pytorch
 enable_chunked_prefill: true
 
 kv_cache_config:
-  free_gpu_memory_fraction: 0.95
+  free_gpu_memory_fraction: 0.85
 
 # NOTE: pytorch_backend_config section flattened since: https://github.com/NVIDIA/TensorRT-LLM/pull/4603
 # NOTE: overlap_scheduler enabled by default since this commit and changed

components/backends/trtllm/engine_configs/decode.yaml

@@ -25,7 +25,7 @@ cuda_graph_config:
   max_batch_size: 16
 
 kv_cache_config:
-  free_gpu_memory_fraction: 0.95
+  free_gpu_memory_fraction: 0.85
 
 cache_transceiver_config:
   backend: default

components/backends/trtllm/engine_configs/prefill.yaml

@@ -24,7 +24,7 @@ disable_overlap_scheduler: true
 cuda_graph_config:
   max_batch_size: 16
 kv_cache_config:
-  free_gpu_memory_fraction: 0.95
+  free_gpu_memory_fraction: 0.85
 
 cache_transceiver_config:
   backend: default
\ No newline at end of file