Description

#1127 changed the default behavior of Float8Tensor so that the FP8 transpose is no longer memoized, but rather treated as part of the data layout. This helps us avoid headaches where the cache is invalidated (since the FP8 data and transpose are updated together in Float8Tensor.quantize_) and reduces the FP8-specific logic in the TE modules (e.g. whether to update the transpose in the first microbatch). However, it also means that Float8Tensor needs to know whether to store the FP8 transpose at construction time.

This PR adds a heuristic kwarg to the fp8_model_init context. Right now the only supported value is "memory", which will cause TE modules to initialize FP8 params without FP8 transposes. In the future, I can imagine extending this logic to fp8_autocast so it can handle FP8 data tensors. We could also support other heuristics (e.g. "performance", "communication", etc) and add logic for other quantization schemes. However, I've purposely kept this PR small to avoid interfering with other changes in the quantization infrastructure (see #1251).

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 refractor

Changes

• Add heuristic option to fp8_model_init context

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