[TRTLLM-5630] restore free_gpu_memory_fraction=0.9 in tests (NVIDIA#4859)

Signed-off-by: ixlmar <[email protected]>

Author: ixlmar

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -460,8 +460,7 @@ def test_bfloat16(self, mtp_nextn, attention_dp, cuda_graph,
             pytest.skip("https://nvbugs/5252313")
         if torch_compile and attention_dp:
             pytest.skip("https://nvbugs/5252559")
-        # OOM on H100 with default free_gpu_memory_fraction=0.9
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.6)
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
         pytorch_config = dict(
             disable_overlap_scheduler=not overlap_scheduler,
             use_cuda_graph=cuda_graph,
@@ -507,8 +506,7 @@ def test_bfloat16_4gpus(self, tp_size, pp_size, ep_size, mtp_nextn,
             pytest.skip("https://nvbugs/5252559")
         if torch_compile and pp_size > 1:
             pytest.skip("PP with torch.compile is not supported yet.")
-        # OOM on H100 with default free_gpu_memory_fraction=0.9
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.7)
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
         pytorch_config = dict(
             disable_overlap_scheduler=not overlap_scheduler,
             use_cuda_graph=cuda_graph,
@@ -552,8 +550,7 @@ def test_fp8_block_scales(self, mtp_nextn, fp8kv, attention_dp, cuda_graph,
             pytest.skip("https://nvbugs/5252313")
         if torch_compile and attention_dp:
             pytest.skip("https://nvbugs/5252559")
-        # OOM on H100 with default free_gpu_memory_fraction=0.9
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.8)
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
         pytorch_config = dict(
             disable_overlap_scheduler=not overlap_scheduler,
             use_cuda_graph=cuda_graph,
@@ -592,8 +589,7 @@ def test_fp8_block_scales(self, mtp_nextn, fp8kv, attention_dp, cuda_graph,
 
     @pytest.mark.skip_device_not_contain(["H100"])
     def test_fp8_block_scales_cuda_graph_padding(self):
-        # OOM on H100 with default free_gpu_memory_fraction=0.9
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.8)
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
         pytorch_config = dict(
             disable_overlap_scheduler=False,
             use_cuda_graph=True,
@@ -638,8 +634,7 @@ def test_fp8_block_scales_4gpus(self, tp_size, pp_size, ep_size, mtp_nextn,
             pytest.skip("https://nvbugs/5252559")
         if torch_compile and pp_size > 1:
             pytest.skip("PP with torch.compile is not supported yet.")
-        # OOM on H100 with default free_gpu_memory_fraction=0.9
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.8)
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
         pytorch_config = dict(
             disable_overlap_scheduler=not overlap_scheduler,
             use_cuda_graph=cuda_graph,
@@ -682,8 +677,7 @@ def test_fp8_block_scales_4gpus(self, tp_size, pp_size, ep_size, mtp_nextn,
     @pytest.mark.skip_less_device(4)
     @pytest.mark.skip_device_not_contain(["H100", "H200"])
     def test_fp8_block_scales_4gpus_static_eplb(self):
-        # OOM on H100 with default free_gpu_memory_fraction=0.9
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.8)
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
 
         num_experts = 72
         num_slots = 80
@@ -840,7 +834,7 @@ def test_no_kv_cache_reuse(self, quant_dtype, mtp_nextn, fp8kv,
         elif quant_dtype == "nvfp4":
             model_path = f"{llm_models_root()}/DeepSeek-V3-Lite/nvfp4_moe_only"
 
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.6,
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9,
                                         enable_block_reuse=False)
         pytorch_config = dict(
             disable_overlap_scheduler=not overlap_scheduler,
@@ -907,7 +901,7 @@ def test_nvfp4_8gpus(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
                          attention_dp, cuda_graph, overlap_scheduler,
                          max_batch_size, moe_backend):
 
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.4)
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
         pytorch_config = dict(disable_overlap_scheduler=not overlap_scheduler,
                               use_cuda_graph=cuda_graph,
                               moe_backend=moe_backend)
@@ -956,7 +950,7 @@ def test_nvfp4_8gpus(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
     def test_fp8_blockscale(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
                             attention_dp, cuda_graph, overlap_scheduler,
                             max_batch_size):
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.4)
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
         pytorch_config = dict(
             disable_overlap_scheduler=not overlap_scheduler,
             use_cuda_graph=cuda_graph,
@@ -1010,7 +1004,7 @@ class TestNemotronNas(LlmapiAccuracyTestHarness):
 
     @pytest.mark.skip_less_device(8)
     def test_auto_dtype_tp8(self):
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.7)
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
         pytorch_config = dict()
 
         with LLM(self.MODEL_PATH,
@@ -1306,7 +1300,7 @@ def test_fp8(self, tp_size, pp_size, ep_size, attention_dp, cuda_graph,
         pytorch_config = dict(disable_overlap_scheduler=not overlap_scheduler,
                               use_cuda_graph=cuda_graph)
 
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.7)
+        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
         llm = LLM(
             f"{llm_models_root()}/Qwen3/saved_models_Qwen3-235B-A22B_fp8_hf",
             tensor_parallel_size=tp_size,