[PyTorch] Avoid select op in PyTorch extensions

[timmoon10]

Description

While profiling CPU overheads in te.Linear with FP8 support enabled, I find that a non-trivial amount of CPU time is spent in cast-transpose functions:

Much of this runtime is spent indexing into FP8 scaling factors. It takes a ~5 us on CPU to create each PyTorch tensor, and this can become significant when GEMM kernels take ~100 us on GPU. (Actually, it seems much of this is profiler overhead. There's some overhead, but it's not as bad as 5 us.)

This PR makes some backward-compatible changes to some of the PyTorch extensions to avoid calling the select op, and thus avoiding some tensor creation. It adds optional arguments for tensor offsets so that we can do pointer arithmetic instead of relying on PyTorch. With these changes, I see a 1.05x speedup in the forward pass of very small te.Linears.

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)

Changes

• Add FP8 scaling factor offset args to cast-transpose PyTorch extensions
• Add FP8 scaling factor offset args to LayerNorm and RMSNorm PyTorch extensions

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

[timmoon10]

/te-ci pytorch

[timmoon10]

/te-ci pytorch

[timmoon10]

/te-ci pytorch

[timmoon10]

/te-ci pytorch

[cyanguwa]

LGTM