support Per Group in Float8DynamicActivationFloat8WeightConfig (#4182)

[zlin888]

Summary:

as title

Differential Revision: D97987011

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/4182

• 📄 Preview Python docs built from this PR

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

❌ 1 New Failure, 1 Unrelated Failure

As of commit 1695a45 with merge base 6f56403 ():

NEW FAILURE - The following job has failed:

• PR Label Check / Check PR Labels (gh)
  Process completed with exit code 1.

BROKEN TRUNK - The following job failed but were present on the merge base:

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

• Run Regression Tests / test-nightly (CUDA Nightly, linux.g5.12xlarge.nvidia.gpu, --pre torch --index-url https://downloa... / linux-job (gh) (trunk failure)
  test/dtypes/test_nf4.py::TestComm::test_comm

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

[meta-codesync]

@zlin888 has exported this pull request. If you are a Meta employee, you can view the originating Diff in D97987011.

[zlin888]

@jerryzh168 @jcaip pls take a look. thanks!

[JiwaniZakir]

In inference.py, logging.getLogger() (line added near the top) creates a root logger rather than a module-scoped one. The standard convention is logging.getLogger(__name__), which avoids accidentally affecting the logging behavior of unrelated modules that share the root logger.

Additionally, in _normalize_granularity and _check_hardware_support, the new is_per_group path checks that both granularities are PerGroup instances, but never validates that their group_size values match. A mismatch between activation and weight group sizes would likely produce silent incorrect results or a cryptic error downstream — an explicit check like granularity[0].group_size == granularity[1].group_size with a descriptive ValueError would catch this early.

Finally, the elif is_per_group branch in _check_hardware_support unconditionally emits a warning on every call. If this function is invoked repeatedly (e.g., per-layer during model conversion), this will flood logs. Emitting it once via warnings.warn(..., stacklevel=2) with stacklevel pointing to the caller, or using logger.warning with a rate-limiting approach, would be more user-friendly.