biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[BugFix] Work around graph partition x torch.compile cache issue (#26956)

Browse Source 
Signed-off-by: Richard Zou <zou3519@gmail.com>
This commit is contained in:
Richard Zou
2025-10-15 23:06:11 -04:00
committed by GitHub
parent e19b16dde6
commit 9b6504c307
2 changed files with 31 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
tests/compile/piecewise/test_toy_llama.py
View File

@@ -337,9 +337,8 @@ def run_model(llama_config, compile_config: CompilationConfig) -> torch.Tensor:
def test_toy_llama(
backend: str, use_inductor_graph_partition: bool, monkeypatch, tmp_path
):
# We disable the vLLM compile cache into a new tmp dir for 2 reasons:
# We disable the vLLM compile cache into a new tmp dir for 1 reason:
# 1. To make sure we can properly track the number of Inductor compilations.
# 2. Inductor partitioning does not play nicely with Autograd cache (below)
monkeypatch.setenv("VLLM_DISABLE_COMPILE_CACHE", "1")
if use_inductor_graph_partition and not is_torch_equal_or_newer("2.9.0.dev"):
@@ -369,15 +368,6 @@ def test_toy_llama(
cudagraph_capture_sizes=[1, 2],
)
# FIXME(luka/boyuan): the graph from the previous test case
# (no inductor partition) gets cached by AotAutograd so then the
# compilation with inductor partitioning incorrectly loads an unpartitioned
# graph and never partitions. I think this is a bug with custom inductor
# partitioning but does not affect vLLM more generally as vLLM uses its own
# cache (which takes inductor partitioning into account).
if use_inductor_graph_partition:
compile_config_no_split.inductor_compile_config["force_disable_caches"] = True
compile_config_split = deepcopy(compile_config_no_split)
compile_config_split.splitting_ops = ["silly::attention"]