Graph Safe Current Scaling Support for GroupedLinear Module/Ops + Fix CUBLAS GGEMM heuristics for Wgrad

[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 graph-safe GroupedLinear path (both the module and ops implementations) to support FP8 per-tensor current scaling on Hopper (CC 9.0) and fixes incorrect cuBLAS grouped GEMM heuristics that were supplying wrong dimension hints for the wgrad GEMM.

• FP8 current-scaling graph-safe path: Float8CurrentScalingQuantizer is now recognised as a valid quantizer for the grouped-tensor GEMM path on Hopper and Blackwell; the Blackwell-only guard that previously rejected it has been moved after an early-return for current-scaling. The bug where rowwise_data/scale_inv were cleared unconditionally (destroying forward activations before save_for_backward when columnwise_data is None) is fixed by adding the grouped_x.columnwise_data is not None guard in both module and ops codepaths.
• cuBLAS wgrad heuristics fix: nvte_grouped_gemm_with_discrete_out was computing avg_m/avg_n/avg_k from the wrong operands for the NT-layout wgrad GEMM call; avg_k now correctly uses first_dim(A) (token count, the contracted dimension) instead of last_dim(A) (in_features), and avg_m/avg_n sources are swapped to match the row-major/column-major convention used internally.
• Test coverage: New fp8_current_scaling parametrised cases added to test_grouped_linear and test_grouped_mlp graph-safe tests, with Hopper-compatible skip logic replacing the previous blanket Blackwell-only skip.

Confidence Score: 5/5

The changes are well-scoped and additive: the correctness fix for the rowwise_data guard is straightforward, the heuristics change only affects algorithm selection (not numerical output), and the test additions provide direct coverage of the new path.

The core logic change — guarding rowwise_data/scale_inv clearing behind columnwise_data is not None — is clearly correct and resolves the previously flagged bug. The cuBLAS heuristic corrections are performance-only (wrong hints lead to suboptimal algorithm selection, not wrong results). No data-loss or correctness regressions are introduced by this PR.

tests/pytorch/test_grouped_mlp.py contains a stale skip message; all other files look correct.

Important Files Changed

Filename	Overview
transformer_engine/common/gemm/cublaslt_grouped_gemm.cu	Fixes wrong avg_m/avg_n/avg_k heuristic sources in nvte_grouped_gemm_with_discrete_out for the wgrad GEMM case: avg_k now correctly uses first_dim(A) (token count — the true contracted dimension) instead of last_dim(A) (in_features) for the NT-layout wgrad call.
transformer_engine/pytorch/module/grouped_linear.py	Correctly guards rowwise_data/scale_inv clearing with columnwise_data is not None, fixing the unconditional clear that destroyed FP8 per-tensor current-scaling activations before save_for_backward. Extends _is_grouped_tensor_path_supported to allow Float8CurrentScalingQuantizer on Hopper by returning True before the Blackwell-only check.
transformer_engine/pytorch/ops/basic/grouped_linear.py	Mirrors the module-level guard fix, adds single_grouped_weight parameter to _is_graph_safe_path_supported to restrict NVFP4 to discrete weights, removes the erroneous float8_current_scaling runtime error for single_grouped_weight, and imports Float8CurrentScalingQuantizer.
tests/pytorch/test_grouped_linear.py	Adds fp8_current_scaling parametrized cases to two graph-safe grouped linear tests and refines the Blackwell-only skip to allow current-scaling on Hopper (SM90).
tests/pytorch/test_grouped_mlp.py	Expands cuda_graph_safe test to cover fp8_current_scaling and nvfp4_rht, adds BF16-only and discrete-weight-only skips for NVFP4, but retains a stale skip message that still mentions current scaling after the condition was narrowed to delayed scaling only.

_{Reviews (6): Last reviewed commit: "Merge branch 'main' into nvfp4_and_fp8_c..." | Re-trigger Greptile}

[vthumbe1503]

/te-ci pytorch

[vthumbe1503]

/te-ci pytorch