[https://nvbugs/5512734][fix] Update kv cache config for maverick (NVIDIA#7710)

Signed-off-by: Ivy Zhang <[email protected]>
Signed-off-by: Wangshanshan <[email protected]>

Author: crazydemo

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -704,13 +704,15 @@ def test_fp8(self, cuda_graph, tp_size, pp_size, ep_size):
         if get_device_count() != tp_size * pp_size:
             pytest.skip("Device count mismatch with world size")
 
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75)
         with LLM(
                 f"{llm_models_root()}/llama4-models/nvidia/Llama-4-Maverick-17B-128E-Instruct-FP8",
                 tensor_parallel_size=tp_size,
                 # Keep this low to avoid warmup OOM in CI
                 max_seq_len=8192,
                 pipeline_parallel_size=pp_size,
                 moe_expert_parallel_size=ep_size,
+                kv_cache_config=kv_cache_config,
                 cuda_graph_config=CudaGraphConfig()
                 if cuda_graph else None) as llm:
             assert llm.args.quant_config.quant_algo == QuantAlgo.FP8