[Perf] Raise Blackwell CUDA graph capture default to 1024 (#49390)

This commit is contained in:
Lucas Wilkinson
2026-08-07 15:18:14 -04:00
committed by GitHub
parent 56a4b63d44
commit 021b7d985b
3 changed files with 48 additions and 8 deletions
+32
View File
@@ -492,6 +492,38 @@ def test_cudagraph_sizes_post_init(
)
@pytest.mark.skipif(
not current_platform.support_static_graph_mode(),
reason="Skip if static graph mode is not supported",
)
@pytest.mark.parametrize(
("is_blackwell", "expected_max_size"), [(False, 512), (True, 1024)]
)
def test_blackwell_cudagraph_default(is_blackwell, expected_max_size):
vllm_config = VllmConfig()
vllm_config.model_config = MagicMock(enforce_eager=False)
vllm_config.scheduler_config = SchedulerConfig(
max_num_seqs=512,
max_num_batched_tokens=2048,
max_model_len=2048,
is_encoder_decoder=False,
)
vllm_config.compilation_config = CompilationConfig(
cudagraph_mode=CUDAGraphMode.FULL_AND_PIECEWISE,
)
with patch.object(
current_platform,
"is_device_capability_family",
return_value=is_blackwell,
):
vllm_config._set_cudagraph_sizes()
assert (
vllm_config.compilation_config.max_cudagraph_capture_size == expected_max_size
)
@pytest.mark.skipif(
not current_platform.support_static_graph_mode(),
reason="Skip if not cudagraph mode supported",