Add Qwen 3.5 MoE to cuda-perf CI and add prefill throughput tracking by Gasoonjia · Pull Request #18903 · pytorch/executorch

Gasoonjia · 2026-04-15T08:02:52Z

Add PyTorchObserver stats output to qwen3_5_moe runner (enables cuda_benchmark.py parsing), --prompt_file flag, and GPU memory stats
Add prefill_throughput metric to cuda_benchmark.py (prefill tok/s alongside existing decode tok/s)
Add Qwen3.5-35B-A3B-HQQ-INT4 to cuda-perf.yml with >1000 token prompt and 512 output tokens, on linux.aws.a100
Creata a new ciflow tag for cuda-erf ci
Remove random model selection and schedule trigger; always run all models when triggered

pytorch-bot · 2026-04-15T08:02:57Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18903

📄 Preview Python docs built from this PR

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

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

Rolling out OSDC (ARC) runners on pull & trunk workflows in PyTorch main

❌ 3 New Failures, 2 Cancelled Jobs, 4 Unrelated Failures

As of commit 9e72e8b with merge base 656850a ():

NEW FAILURES - The following jobs have failed:

pull / android / run-emulator (gh)
The process '/usr/local/lib/android/sdk/platform-tools/adb' failed with exit code 224
pull / unittest-arm-backend-with-no-deps (test_pytest_models_tosa) / linux-job (gh)
RuntimeError: Command docker exec -t 6b98484c92b8738be4847fe50ef78c42deb2d07dc44dbd7e150e4d941eed345a /exec failed with exit code 1
trunk / test-arm-backend-ethos-u (test_pytest_ops_ethos_u85) / linux-job (gh)
RuntimeError: Command docker exec -t 24b066054793970da292850d2f9259eb8a86c25f2466c29506e58ad8a9ef5e5a /exec failed with exit code 1

CANCELLED JOBS - The following jobs were cancelled. Please retry:

cuda-perf / benchmark-cuda (SocialLocalMobile/Qwen3.5-35B-A3B-HQQ-INT4, quantized-int4-tile-packed, SocialLoc... / linux-job (gh)
trunk / test-models-macos-coreml (edsr) / macos-job (gh)
##[error]The operation was canceled.

FLAKY - The following job failed but was likely due to flakiness present on trunk:

trunk / test-huggingface-transformers-macos (llama3.2-1b|coreml_fp32_gpu|--quantize) / macos-job (gh) (detected as infra flaky with no log or failing log classifier)

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

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

pull / unittest / windows / windows-job (gh) (trunk failure)
##[error]The operation was canceled.
pull / unittest-editable / windows / windows-job (gh) (trunk failure)
##[error]The operation was canceled.
trunk / unittest-release / windows / windows-job (gh) (trunk failure)
##[error]The operation was canceled.

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

github-actions · 2026-04-15T08:03:38Z

This PR needs a `release notes:` label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

- Add PyTorchObserver stats output to qwen3_5_moe runner (enables cuda_benchmark.py parsing), --prompt_file flag, and GPU memory stats - Add prefill_throughput metric to cuda_benchmark.py (prefill tok/s alongside existing decode tok/s) - Add Qwen3.5-35B-A3B-HQQ-INT4 to cuda-perf.yml with >1000 token prompt and 512 output tokens, on linux.aws.a100 - Align cuda-perf.yml triggers with cuda.yml (push main/release, ciflow/cuda tags, PR on backends/cuda and backends/aoti paths) - Remove random model selection and schedule trigger; always run all models when triggered

digantdesai · 2026-04-21T03:57:30Z

+            TOKENIZER="model_artifacts/tokenizer.json"
+            # Generate a >1000 token prompt for benchmarking
+            python3 -c "
+        text = 'The quick brown fox jumps over the lazy dog. ' * 200


can we just check in a file, instead of creating one on the fly, with different large prompts and use those everywhere, such repetitions can lead to weird response sometimes. For a pure performance benchmarking this is OK however since this is a CI in case of a suspect diff, we will come look at the output and it better not be garbage because of this prompt. :)

have created a specific file for long prompt.

digantdesai · 2026-04-21T03:59:09Z

 DEFINE_string(data_path, "", "Data file (.ptd) for CUDA backend.");
 DEFINE_string(tokenizer_path, "", "HuggingFace tokenizer.json path.");
 DEFINE_string(prompt, "Hello", "Prompt text.");
+DEFINE_string(


digantdesai · 2026-04-21T03:59:36Z

+  stats.gpu_peak_usage_mb =
+      (stats.gpu_total_bytes - gpu_free_bytes) / 1024.0 / 1024.0;
+
+  llm::print_report(stats);


…18903) - Add PyTorchObserver stats output to qwen3_5_moe runner (enables cuda_benchmark.py parsing), --prompt_file flag, and GPU memory stats - Add prefill_throughput metric to cuda_benchmark.py (prefill tok/s alongside existing decode tok/s) - Add Qwen3.5-35B-A3B-HQQ-INT4 to cuda-perf.yml with >1000 token prompt and 512 output tokens, on linux.aws.a100 - Creata a new ciflow tag for cuda-erf ci - Remove random model selection and schedule trigger; always run all models when triggered --------- Co-authored-by: gasoonjia <gasoonjia@fb.com>

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 Apr 15, 2026

Gasoonjia temporarily deployed to upload-benchmark-results April 15, 2026 11:33 — with GitHub Actions Inactive

Gasoonjia force-pushed the cuda-perf-qwen35-moe branch from a23b5b1 to 49d0aa1 Compare April 15, 2026 18:46

Gasoonjia temporarily deployed to upload-benchmark-results April 15, 2026 21:00 — with GitHub Actions Inactive

fix 0 first token

36213f9

Gasoonjia force-pushed the cuda-perf-qwen35-moe branch from e50e3fa to 36213f9 Compare April 15, 2026 22:15

Merge branch 'main' into cuda-perf-qwen35-moe

7336416

Gasoonjia marked this pull request as ready for review April 15, 2026 22:15

Gasoonjia requested a review from lucylq as a code owner April 15, 2026 22:15

Gasoonjia requested a review from digantdesai April 15, 2026 22:17

lint

bd5b601

Gasoonjia temporarily deployed to upload-benchmark-results April 16, 2026 00:50 — with GitHub Actions Inactive

digantdesai reviewed Apr 21, 2026

View reviewed changes

digantdesai approved these changes Apr 21, 2026

View reviewed changes

solve comment; decouple cuda ci with perf ci by creating perf ci tag

69b5042

Gasoonjia force-pushed the cuda-perf-qwen35-moe branch from df2a4bf to 69b5042 Compare April 22, 2026 07:54

Merge branch 'main' into cuda-perf-qwen35-moe

feaa1a7

Gasoonjia added the ciflow/cuda label Apr 22, 2026

make prompt shorter without template

3635fdf

Gasoonjia temporarily deployed to upload-benchmark-results April 22, 2026 12:33 — with GitHub Actions Inactive

remove extra tensor move action

9e72e8b

Gasoonjia temporarily deployed to upload-benchmark-results April 22, 2026 22:15 — with GitHub Actions Inactive

Gasoonjia temporarily deployed to upload-benchmark-results April 23, 2026 00:18 — with GitHub Actions Inactive

Gasoonjia merged commit c48ea12 into main Apr 23, 2026
472 of 481 checks passed

Gasoonjia deleted the cuda-perf-qwen35-moe branch April 23, 2026 03:13

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add Qwen 3.5 MoE to cuda-perf CI and add prefill throughput tracking#18903

Add Qwen 3.5 MoE to cuda-perf CI and add prefill throughput tracking#18903
Gasoonjia merged 8 commits intomainfrom
cuda-perf-qwen35-moe

Gasoonjia commented Apr 15, 2026 •

edited

Loading

Uh oh!

pytorch-bot Bot commented Apr 15, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Apr 15, 2026

Uh oh!

digantdesai Apr 21, 2026 •

edited

Loading

Uh oh!

Gasoonjia Apr 22, 2026

Uh oh!

digantdesai Apr 21, 2026

Uh oh!

digantdesai Apr 21, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

Gasoonjia commented Apr 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-bot Bot commented Apr 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18903

❗ 1 Active SEVs

❌ 3 New Failures, 2 Cancelled Jobs, 4 Unrelated Failures

Uh oh!

github-actions Bot commented Apr 15, 2026

This PR needs a release notes: label

Uh oh!

digantdesai Apr 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Gasoonjia Apr 22, 2026

Choose a reason for hiding this comment

Uh oh!

digantdesai Apr 21, 2026

Choose a reason for hiding this comment

Uh oh!

digantdesai Apr 21, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Gasoonjia commented Apr 15, 2026 •

edited

Loading

pytorch-bot Bot commented Apr 15, 2026 •

edited

Loading

This PR needs a `release notes:` label

digantdesai Apr 21, 2026 •

edited

Loading