3/x mx cleanup: speed up local test suite

[vkuzo]

Summary:

• Before, worst case: ~4 minutes
• Before, best case: ~2 minutes
• After: ~82 sec

Changes:

1. test smaller shapes in triton kernel tests, sometimes the large
   shapes are expensive in local runs
2. remove testing of mxfp4 + compile, it's really slow and not useful
   yet

Test Plan:

pytest test/prototype/mx_formats/ -s

Reviewers:

Subscribers:

Tasks:

Tags:

[vkuzo]

Stack from ghstack (oldest at bottom):

• 6/x mx cleanup: make NVFP4Tensor use base implements #3387
• 5/x mx cleanup: rename mx inf config to MXDynamicActivationMXWeightConfig #3386
• 4/x: mx cleanup: use kernel_preference instead of gemm_kernel_choice #3385
• -> 3/x mx cleanup: speed up local test suite #3383
• 2/x mx cleanup: remove pack_fp6 #3382
• 1/x mx cleanup: mxtensor: rename _block_size to block_size #3381

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/3383

• 📄 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:

• Dr CI is temporarily not working due to API fairewall

❌ 2 New Failures

As of commit d4f3b17 with merge base 5f33595 ():

NEW FAILURES - The following jobs have failed:

• Run 1xL4 Tests / test (SM-89, linux.g6.4xlarge.experimental.nvidia.gpu, --pre torch --index-url https://download.p... / linux-job (gh)
  RuntimeError: Command docker exec -t 5e8a274bee56ef2cabc806744659ec3b5df71f294debec70616b716e7f8cb9f2 /exec failed with exit code 1
• Run Regression Tests / test-nightly (CUDA Nightly, linux.g5.12xlarge.nvidia.gpu, --pre torch --index-url https://downloa... / linux-job (gh)
  RuntimeError: Command docker exec -t 63d73a15abfa785621d20348732dcf26288b3251d5d625c7526d2c9771cea52e /exec failed with exit code 1

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

[danielvegamyhre]

is this removed because in practice we're only using floor and rceil?

[vkuzo]

yep!

CEIL is still tested for eager, this removes it from compile