Add fused QKV HQQ triton_mm test

[jeromeku]

Add fused QKV HQQ triton_mm test

Add test / example of how to fuse qkv projections when using the hqq triton_mm fused int4 matmul kernel.

Description

Common pattern in transformer models is to fuse query, key, and value weights such that qkv projection can be fused into single matmul.

In order to use torchao.prototype.hqq.triton_mixed_mm during int4 quantized model training / inference, query, key, and value weights are quantized and packed. This introduces additional complexities when additionally applying this fusion pattern.

Contribution

This PR adds a test that demonstrates how to properly fuse and pack query, key, and v projection weights such that the fused weight can be used with the triton_mixed_mm kernel.

Tests for equivalence of passing single fused qkv into kernel against individual q, k, v kernels for both transposed and non-transposed cases.

@msaroufim @mobicham @KeremTurgutlu

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/306

• 📄 Preview Python docs built from this PR

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 50a9ce7 with merge base 8dbf031 ():

FLAKY - The following jobs failed but were likely due to flakiness present on trunk:

• Run Regression Tests / test (CUDA 2.2.2, linux.g5.12xlarge.nvidia.gpu, torch==2.2.2, cuda, 12.1) / linux-job (gh) (matched linux rule in flaky-rules.json)
  The process '/usr/bin/git' failed with exit code 128
• Run Regression Tests / test (CUDA 2.3, linux.g5.12xlarge.nvidia.gpu, torch==2.3.0, cuda, 12.1) / linux-job (gh) (matched linux rule in flaky-rules.json)
  The process '/usr/bin/git' failed with exit code 128
• Run Regression Tests / test (CUDA Nightly, linux.g5.12xlarge.nvidia.gpu, --pre torch --index-url https://download.pytorc... / linux-job (gh) (matched linux rule in flaky-rules.json)
  The process '/usr/bin/git' failed with exit code 128

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