[HollowMan6]

Description

Set the default backward-gradient amax_epsilon in Float8BlockScaling to 1e-12. This gives zero-amax E5M2 gradient blocks a nonzero floor in the recipe, instead of requiring downstream projects to override the recipe instance. Forward input and weight quantization defaults remain unchanged.

Fixes https://github.com/NVIDIA/Megatron-LM/pull/7523/changes#r4211643974

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

• Use amax_epsilon=1e-12 for the default blockwise FP8 backward-gradient quantizer.
• Document the backward-only default.
• Test zero-gradient quantization and Linear backward, and compare ordinary nonzero-block quantization bitwise with the previous epsilon-zero setting.

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

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Changes default FP8 quantization epsilon for backward gradients.

The PR appears safe to merge, with the previous test gap addressed.

What we checked:

• Expected scales match the kernel: The kernel floors the block maximum to epsilon, divides the FP8 maximum by it, and rounds down to a power of two when requested. This matches the test. The 128-by-128 input leaves no padded scale entries.

Summary

Sets the default backward-gradient amax_epsilon in Float8BlockScaling to 1e-12. Forward defaults stay unchanged.

• Adds recipe-default and zero-gradient backward checks.
• Strengthens zero-block tests to check the expected inverse scales in both directions.
• No new actionable issues found.

_{Reviews (2) · Last reviewed commit: "address review" · Reviewed by Greptile}

[ptrendx]

Not fully against merging this PR, but this sounds more like the problem with the quantization method to be honest - having amax equal to 0 should be perfectly fine, we just should guard against creating the infinite scale from it (and instead have some predefined scale in that case, like 1 or something).

[HollowMan6]

Thank you @ptrendx for your suggestions, please find discussions at NVIDIA/Megatron-LM#7523 (comment), for that I will convert this into draft and let you know if the changes are really necessary here.

[HollowMan6]

Closing as it seems this defensive change is no longer needed.