Graph Safe Current Scaling Support for GroupedLinear Module/Ops

[vthumbe1503]

Description

Please include a brief summary of the changes, relevant motivation and context.

Fixes # (issue)

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

[vthumbe1503]

/te-ci pytorch

[greptile-apps]

Greptile Summary

This PR extends the CUDA-graph-safe grouped-GEMM path in both the module (_GroupedLinear) and ops (GroupedLinear) layers to support FP8 per-tensor current scaling, which is backed by tex.group_quantize and cuBLASLt grouped GEMM and runs on Hopper (CC 9.0) as well as Blackwell.

• Core logic change: Float8CurrentScalingQuantizer is now detected before the Blackwell-only CC check in both _is_grouped_tensor_path_supported (module) and _is_graph_safe_path_supported (ops), enabling an early return True on Hopper.
• Bug fix carried forward: The guard that frees grouped_x.rowwise_data/scale_inv before save_for_backward is now gated on grouped_x.columnwise_data is not None, preventing null-activation corruption for per-tensor FP8. The single_grouped_weight restriction for Float8CurrentScaling in make_grouped_weights is also removed, and _is_graph_safe_path_supported gains a single_grouped_weight parameter to properly restrict only NVFP4.
• Tests: New fp8_current_scaling parametrize cases are added to test_grouped_linear.py with correctly relaxed Hopper skip logic; test_grouped_mlp.py adds fp8_current_scaling/nvfp4_rht to the cuda-graph-safe test suite with appropriate per-recipe skip guards.

Confidence Score: 5/5

The changes are safe to merge — the core guard fixes are correct in both code paths and the new Float8CurrentScaling early-return is logically sound.

Both the module and ops layers correctly gate rowwise_data cleanup on columnwise_data is not None, preventing null-activation corruption for per-tensor FP8. The new dispatch is consistent between paths, and the single_grouped_weight restriction is properly scoped to NVFP4 only.

No files require special attention.

Important Files Changed

Filename	Overview
transformer_engine/pytorch/module/grouped_linear.py	Adds Float8CurrentScalingQuantizer early-return in _is_grouped_tensor_path_supported and fixes the backward-pass guard so rowwise_data/scale_inv are only cleared when columnwise_data is present.
transformer_engine/pytorch/ops/basic/grouped_linear.py	Imports Float8CurrentScalingQuantizer, threads single_grouped_weight into _is_graph_safe_path_supported, removes the erroneous float8_current_scaling restriction in make_grouped_weights, and applies the columnwise_data is not None guard before freeing rowwise_data.
tests/pytorch/test_grouped_linear.py	Adds fp8_current_scaling parametrize cases and correctly relaxes the Blackwell-only skip for that recipe.
tests/pytorch/test_grouped_mlp.py	Adds fp8_current_scaling and nvfp4_rht to the cuda-graph-safe test; stale skip reason string for the delayed-scaling guard.

_{Reviews (4): Last reviewed commit: "Merge branch 'nvfp4_and_fp8_current_scal..." | Re-trigger Greptile}