[Bugfix] Re-enable use_cudagraph in vLLM v1 (#19299)

Signed-off-by: Richard Zou <zou3519@gmail.com>
This commit is contained in:
Richard Zou
2025-06-08 08:56:12 +08:00
committed by GitHub
parent d77f7fb871
commit eaa2e51088
6 changed files with 52 additions and 8 deletions
+1 -1
View File
@@ -95,7 +95,7 @@ def _test_simple_piecewise_compile(*, use_inductor):
num_piecewise_graphs_seen=5, # 2 * num_layers + 1
num_piecewise_capturable_graphs_seen=3, # 1 + num_layers
num_backend_compilations=3, # num_piecewise_capturable_graphs_seen
num_cudagraph_caputured=
num_cudagraph_captured=
6, # num_cudagraph_sizes * num_piecewise_capturable_graphs_seen
):
+3 -3
View File
@@ -327,7 +327,7 @@ def _test_toy_llama(*, use_inductor):
num_piecewise_graphs_seen=0,
num_piecewise_capturable_graphs_seen=0,
num_backend_compilations=0,
num_cudagraph_caputured=0,
num_cudagraph_captured=0,
):
outputs.append(
run_model(llama_config, use_inductor=False, use_compile=False))
@@ -343,7 +343,7 @@ def _test_toy_llama(*, use_inductor):
num_piecewise_graphs_seen=1,
num_piecewise_capturable_graphs_seen=1,
num_backend_compilations=1, # num_piecewise_capturable_graphs_seen
num_cudagraph_caputured=
num_cudagraph_captured=
2, # num_cudagraph_sizes * num_piecewise_capturable_graphs_seen
**kwargs,
):
@@ -361,7 +361,7 @@ def _test_toy_llama(*, use_inductor):
llama_config.num_layers, # 1 + num_layers
num_backend_compilations=1 +
llama_config.num_layers, # num_piecewise_capturable_graphs_seen
num_cudagraph_caputured=2 *
num_cudagraph_captured=2 *
(1 + llama_config.num_layers
), # num_cudagraph_sizes * num_piecewise_capturable_graphs_seen
):