Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[V1] Do not use inductor for piecewise CUDA graphs #10225

Merged
merged 2 commits into from
Nov 11, 2024
Merged

[V1] Do not use inductor for piecewise CUDA graphs #10225

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
195b833
[V1] Do not use inductor for piecewise CUDA graphs
WoosukKwon Nov 11, 2024
6c2750c
minor
WoosukKwon Nov 11, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 3 additions & 4 deletions vllm/v1/worker/gpu_model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -404,15 +404,14 @@ def execute_model(

def load_model(self) -> None:
if self.use_cuda_graph:
# FIXME(woosuk): Currently, the custom ops are not supported
# in the piecewise compilation mode. We rely on TorchInductor
# to optimize the model.
# FIXME(woosuk): Currently, we do not use inductor to reduce the
# compilation time and any potential issues with the inductor.
os.environ["VLLM_CUSTOM_OPS"] = "none"
set_compilation_config(
CompilationConfig(
use_cudagraph=True,
non_cudagraph_ops=["vllm.unified_v1_flash_attention"],
use_inductor=True,
use_inductor=False,
))

logger.info("Starting to load model %s...", self.model_config.model)
Expand Down