Skip to content

Arm backend: Add Qwen3 VL language model MXFP8 test#20533

Merged
tom-arm merged 1 commit into
pytorch:mainfrom
tom-arm:add_qwen_mxfp_test
Jun 29, 2026
Merged

Arm backend: Add Qwen3 VL language model MXFP8 test#20533
tom-arm merged 1 commit into
pytorch:mainfrom
tom-arm:add_qwen_mxfp_test

Conversation

@tom-arm

@tom-arm tom-arm commented Jun 26, 2026

Copy link
Copy Markdown
Collaborator
  • Add an E2E TOSA test for the Qwen3 VL language model with linear layers converted to MXFP8.

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani

Copilot AI review requested due to automatic review settings June 26, 2026 11:48
@tom-arm tom-arm requested a review from digantdesai as a code owner June 26, 2026 11:48
@pytorch-bot

pytorch-bot Bot commented Jun 26, 2026

Copy link
Copy Markdown

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20533

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

✅ You can merge normally! (3 Unrelated Failures)

As of commit d5a486a with merge base a2438e4 (image):

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 Jun 26, 2026
@linux-foundation-easycla

linux-foundation-easycla Bot commented Jun 26, 2026

Copy link
Copy Markdown

CLA Signed
The committers listed above are authorized under a signed CLA.

  • ✅ login: bdemirb / name: Baris (d5a486a)
  • ✅ login: tom-arm / name: Tom Allsop (d5a486a)

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Adds an end-to-end Arm TOSA test to validate MXFP8 (FP8 e4m3) conversion for the Qwen3-VL TextModel path, ensuring BF16 + MXFP lowering and accuracy checks work together for this model family.

Changes:

  • Introduces an MXFP8 + BF16 TOSA pipeline test for TextModelWrapper using MXFPTosaPipelineFP.
  • Adds a Linear-module filter and a post-export op-count assertion to verify all nn.Linear layers are rewritten to torch.ops.tosa_mxfp.linear.default.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

@tom-arm tom-arm force-pushed the add_qwen_mxfp_test branch from 1038f14 to 0728a47 Compare June 26, 2026 11:59
@tom-arm

tom-arm commented Jun 26, 2026

Copy link
Copy Markdown
Collaborator Author

/easycla

@tom-arm tom-arm added the release notes: arm Changes to the ARM backend delegate label Jun 26, 2026
@tom-arm

tom-arm commented Jun 29, 2026

Copy link
Copy Markdown
Collaborator Author

/easycla

* Add an E2E TOSA test for the Qwen3 VL language model with
  linear layers converted to MXFP8.

Signed-off-by: Baris Demir <baris.demir@arm.com>
Signed-off-by: Tom Allsop <tom.allsop@arm.com>

Co-authored-by: Baris Demir <baris.demir@arm.com>
@tom-arm tom-arm force-pushed the add_qwen_mxfp_test branch from 0728a47 to d5a486a Compare June 29, 2026 08:56
@tom-arm

tom-arm commented Jun 29, 2026

Copy link
Copy Markdown
Collaborator Author

/easycla

1 similar comment
@tom-arm

tom-arm commented Jun 29, 2026

Copy link
Copy Markdown
Collaborator Author

/easycla

@tom-arm tom-arm merged commit a812429 into pytorch:main Jun 29, 2026
484 of 489 checks passed
@tom-arm tom-arm deleted the add_qwen_mxfp_test branch June 29, 2026 12:36
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

ciflow/trunk CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. module: arm Issues related to arm backend release notes: arm Changes to the ARM backend delegate

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants