Current scaling: two-stage Triton amax kernel

[matthiasdiener]

Description

The corresponding 2-stage HIP kernel amax implementation is in #369.

Partially addresses https://github.com/ROCm/frameworks-internal/issues/14303.

See https://github.com/ROCm/frameworks-internal/issues/14303#issuecomment-3554900809 for a performance analysis.

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

• Added a 2-stage Triton amax implementation, optional, but enabled by default
  • Disable by setting export NVTE_USE_ATOMIC_AMAX=1 (this will use the previous atomic implementation)

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

[wangye805]

The triton path is not enabled by default so I think you will need to test with both NVTE_USE_ATOMIC_AMAX=1 and NVTE_USE_ATOMIC_AMAX=0 when NVTE_USE_CAST_TRANSPOSE_TRITON is 1.

Also not sure about the runtime cost of adding two new pytests in level 3

[matthiasdiener]

The triton path is not enabled by default so I think you will need to test with both NVTE_USE_ATOMIC_AMAX=1 and NVTE_USE_ATOMIC_AMAX=0 when NVTE_USE_CAST_TRANSPOSE_TRITON is 1.

I added both cases in d7259d1.

Also not sure about the runtime cost of adding two new pytests in level 3

test_numerics.py takes about 5 min, test_fusible_ops.py takes about 1 min (on gfx942), times 2 since we run it with NVTE_USE_ATOMIC_AMAX=0 and =1. Perhaps adding just the test in 188b7ca is enough?

[wangye805]

5 mins sounds okay for level 3. @ipanfilo , what do you think?

[matthiasdiener]

After discussing with @wenchenvincent, we concluded that it is worth keeping the extra tests around.

[matthiasdiener]

The triton path is not enabled by default so I think you will need to test with both NVTE_USE_ATOMIC_AMAX=1 and NVTE_USE_ATOMIC_AMAX=0 when NVTE_USE_CAST_TRANSPOSE_TRITON is 1.

I added both cases in d7259d1.

Also not sure about the runtime cost of adding two new pytests in level 3

test_numerics.py takes about 5 min, test_fusible_ops.py takes about 1 min (on gfx942), times 2 since we run it with NVTE_USE_ATOMIC_AMAX=0 and =1. Perhaps adding just the test in 188b7ca is enough?

[matthiasdiener]

@wangye805 Do you remember why current scaling was disabled here (in #374)?

[wangye805]

I recall I moved this line to

TransformerEngine/transformer_engine/pytorch/triton_kernels/cast.py

Line 59 in 7f6682c

from ..tensor.float8_tensor import Float8CurrentScalingQuantizer

instead of disabling curent scaling quantizer entirely, in order to resolve a circular inclusion issue