Qualcomm AI Engine Direct - Mimi Enablement Stage 1

[winskuo-quic]

Summary

This is the stage 1 of Mimi Enablement.
Stage 2 will consist of actual model enablement.

• Support OP:
  • exp
  • expm1
  • elu
  • transpose conv1d
  • bitwise_and
  • scalar_tensor
  • stack
  • unbind

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/9570

• 📄 Preview Python docs built from this PR

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

❌ 2 New Failures

As of commit f277d44 with merge base 90f0843 ():

NEW FAILURES - The following jobs have failed:

• pull / android / build-llm-demo / linux-job (gh)
  RuntimeError: Command docker exec -t 017c2ffa40c46acfc35ab9ee6017c55b76f8093f9ed7ad1fecc3a9876658887a /exec failed with exit code 1
• pull / test-models-linux (phi_4_mini, portable, linux.4xlarge.memory) / linux-job (gh)
  RuntimeError: Model phi_4_mini is not a valid name. Available models are ['mul', 'linear', 'add', 'add_mul', 'softmax', 'dl3', 'edsr', 'emformer_transcribe', 'emformer_predict', 'emformer_join', 'llama2', 'llama', 'llama3_2_vision_encoder', 'lstm', 'mobilebert', 'mv2', 'mv2_untrained', 'mv3', 'vit', 'w2l', 'ic3', 'ic4', 'resnet18', 'resnet50', 'llava', 'efficient_sam', 'qwen2_5', 'phi-4-mini'].

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

[winskuo-quic]

Hi @cccclai, @iseeyuan, @billmguo,
This PR is the stage 1 of Mimi Enablement, which is mainly focusing on supporting operations and adding some passes.

Please notice the change under examples/models/llama/export_llama_lib.py is to preserve the original behavior. The pass AnnotateDecomposed is now disabled by default. However, since the export_llama_lib.py actually needs this pass to be on, I have manually enabled the pass.

Please have a look.
Thanks

[facebook-github-bot]

@cccclai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[cccclai]

Hi @cccclai, @iseeyuan, @billmguo, This PR is the stage 1 of Mimi Enablement, which is mainly focusing on supporting operations and adding some passes.

Please notice the change under examples/models/llama/export_llama_lib.py is to preserve the original behavior. The pass AnnotateDecomposed is now disabled by default. However, since the export_llama_lib.py actually needs this pass to be on, I have manually enabled the pass.

Please have a look. Thanks

Mind adding this note as part of the code command, so people know the context in the future? Thanks.

[facebook-github-bot]

@cccclai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[winskuo-quic]

Hi @cccclai, @iseeyuan, @billmguo, This PR is the stage 1 of Mimi Enablement, which is mainly focusing on supporting operations and adding some passes.
Please notice the change under examples/models/llama/export_llama_lib.py is to preserve the original behavior. The pass AnnotateDecomposed is now disabled by default. However, since the export_llama_lib.py actually needs this pass to be on, I have manually enabled the pass.
Please have a look. Thanks

Mind adding this note as part of the code command, so people know the context in the future? Thanks.

Thanks for reviewing and the reminder.
I will add this during stage2 of Mimi Enablement.

[winskuo-quic]

@pytorchbot label "release notes: qualcomm"

[facebook-github-bot]

@cccclai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[cccclai]

Looks good to me, thank you!