[torch.compile] Make inductor partition rules respect splitting_ops #25691 (#25845)

Signed-off-by: baonudesifeizhai <baonudesifeizhai@gmail.com>
Signed-off-by: baonudesifeizhai <85092850+baonudesifeizhai@users.noreply.github.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
This commit is contained in:
baonudesifeizhai
2025-10-10 16:35:28 +00:00
committed by GitHub
co-authored by Luka Govedič
parent e519281920
commit cddce79fda
9 changed files with 267 additions and 112 deletions
@@ -198,7 +198,7 @@ def test_multi_graph_piecewise_compile_outputs_equal():
compilation_config=CompilationConfig(
level=CompilationLevel.PIECEWISE,
use_cudagraph=True,
splitting_ops=["silly.attention"],
splitting_ops=["silly::attention"],
cudagraph_capture_sizes=[1, 2],
)
)
@@ -267,7 +267,7 @@ def test_multi_graph_piecewise_compile_outputs_equal():
compilation_config=CompilationConfig(
level=CompilationLevel.PIECEWISE,
use_cudagraph=False,
splitting_ops=["silly.attention"],
splitting_ops=["silly::attention"],
)
)
cudagraph_runtime_mode = CUDAGraphMode.PIECEWISE
+2 -2
View File
@@ -127,7 +127,7 @@ def _run_simple_model(
@torch.inference_mode()
def test_simple_piecewise_compile(use_inductor):
_run_simple_model(
splitting_ops=["silly.attention"],
splitting_ops=["silly::attention"],
use_inductor_graph_partition=False,
use_inductor=use_inductor,
# 2 * num_layers + 1
@@ -142,7 +142,7 @@ def test_simple_piecewise_compile(use_inductor):
@torch.inference_mode()
@pytest.mark.parametrize("splitting_ops", [["silly.attention"], []])
@pytest.mark.parametrize("splitting_ops", [["silly::attention"], []])
def test_simple_inductor_graph_partition(splitting_ops, monkeypatch):
if not is_torch_equal_or_newer("2.9.0.dev"):
pytest.skip("inductor graph partition is only available in PyTorch 2.9+")
+2 -2
View File
@@ -268,7 +268,7 @@ def run_model(
cudagraph_capture_sizes=[1, 2],
)
if split_attn:
compilation_config.splitting_ops = ["silly.attention"]
compilation_config.splitting_ops = ["silly::attention"]
cudagraph_runtime_mode = CUDAGraphMode.PIECEWISE
else:
compilation_config = CompilationConfig(
@@ -438,7 +438,7 @@ def benchmark():
compilation_config = CompilationConfig(
level=CompilationLevel.PIECEWISE,
use_cudagraph=True,
splitting_ops=["silly.attention"],
splitting_ops=["silly::attention"],
cudagraph_capture_sizes=cudagraph_sizes,
)
else: