Skip to content

Fix Qwen3.5 AOTI kernel packaging on CUDA - #22758

Open
Gasoonjia wants to merge 1 commit into
mainfrom
codex/fix-qwen35-pytorch214-autotune
Open

Fix Qwen3.5 AOTI kernel packaging on CUDA#22758
Gasoonjia wants to merge 1 commit into
mainfrom
codex/fix-qwen35-pytorch214-autotune

Conversation

@Gasoonjia

@Gasoonjia Gasoonjia commented Sep 11, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

Summary:

  • enable compile-time Triton autotuning for Qwen3.5 CUDA export
  • embed selected kernel binaries in AOTI shared objects instead of retaining temporary file paths
  • avoid PyTorch 2.14 lazy-JIT illegal-memory-access and runtime file-not-found failures

Validation:

  • reproduced the lazy-autotune CUDA illegal-memory-access and missing .fatbin runtime failure
  • completed full Qwen3.5-35B-A3B-HQQ-INT4 export on a dual RTX 5090 server
  • export peak GPU memory: 18816.96 MiB (limit: 21504 MiB)
  • verified the .pte contains no temporary .fatbin path references
  • built and ran the CUDA C++ runner successfully; generated The capital of France is Paris.
  • black --check examples/models/qwen3_5_moe/export.py

@pytorch-bot

pytorch-bot Bot commented Sep 11, 2026 •
edited
Loading

Copy link
Copy Markdown

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22758

Note: Links to docs will display an error until the docs builds have been completed.

❌ You can merge normally! (4 Unrelated Failures), 1 Unclassified Failure

As of commit 1acdeb4 with merge base 1abaa87 (image):

UNCLASSIFIED FAILURE - DrCI could not classify the following job because the workflow did not run on the merge base. The failure may be pre-existing on trunk or introduced by this PR:

BROKEN TRUNK - The following jobs failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

This comment was automatically generated by Dr. CI and updates every 15 minutes.

@meta-cla meta-cla Bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Sep 11, 2026
@Gasoonjia
Gasoonjia force-pushed the codex/fix-qwen35-pytorch214-autotune branch from 6d4b1fb to 17845c4 Compare September 11, 2026 22:52
@Gasoonjia Gasoonjia changed the title Fix Qwen3.5 lazy AOTI autotuning on CUDA Fix Qwen3.5 AOTI kernel packaging on CUDA Sep 11, 2026
@Gasoonjia
Fix Qwen3.5 AOTI kernel packaging on CUDA
1acdeb4 
Signed-off-by: Songhao Jia <gasoonjia@meta.com>
@Gasoonjia
Gasoonjia force-pushed the codex/fix-qwen35-pytorch214-autotune branch from 17845c4 to 1acdeb4 Compare September 12, 2026 08:19
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. release notes: backends

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@Gasoonjia