[matthiasdiener]

Description

Fixes #790 (related PRs: #646, #615, #593).

Triton's autotuner picks a launch config per process by timing. For the
RMSNorm and LayerNorm kernels, num_warps (and the RMSNorm dgamma-reduce block
shape) sets the reduction order, so processes running identical
shapes can produce slightly different results.

Launch configs are now fixed by row width:

• RMSNorm fwd num_warps: 1 (<= 1024), 4 (<= 8192), else 16
• RMSNorm bwd num_warps: 1 (<= 1024), 4 (<= 2048), else 16
• RMSNorm dgamma reduce: BLOCK_SIZE_M=128, BLOCK_SIZE_N=64, num_warps=8
• LayerNorm fwd num_warps: 1 (<= 2048), 4 (<= 6144), else 8; num_stages=1

On gfx950 these are within ~1% geomean of the per-shape best. LayerNorm fwd
is 1.2-2.4x faster at widths <= 2048, which the old config space missed.
The autotune kwargs remain for API compatibility. The FP8 transpose stays
autotuned because its output does not depend on the config.

Also adds a test.

cc @jbelofamd

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[Micky774]

Since this is a TE-only API I think we can deprecate this parameter, e.g. set the default to "deprecated" and then warn on usage, and internally force-set it to false.

[matthiasdiener]

Good idea, thanks. Done (in spirit) in 15d8189; te_rmsnorm_fwd_triton still uses autotuning in part (for the FP8 transpose kernel).

[Micky774]

LGTM

[jbelofamd]

Re-checked at 15d8189. The new commit only changes how the autotune argument is handled (deprecation warnings) and the tests. The launch-config tables and the kernels are the same as at 0ec0e4a, which is the head we validated above, so that result still holds.

On CI: both red sGPU jobs (mi30x and mi35x) fail only in the examples group, in the JAX MNIST example (examples/jax/mnist/test_single_gpu_mnist.py). Its call to datasets.load_dataset dies while hashing the config: TypeError: Pickler._batch_setitems() takes 2 positional arguments but 3 were given, which comes from dill 0.3.8 running under Python 3.14. The torch, jax and core groups pass on both, so the failure is unrelated to this PR.

Thanks again, @matthiasdiener, and @Micky774 for the review.