[PyTorch] Change order of args in another permutation triton kernel

[tdophung]

Description

Change order of args in function that I missed from #2416

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

change order of arhgs in _unpermute_bwd_with_merging_probs_kernel

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

[tdophung]

/te_ci L0 pytorch

[greptile-apps]

Greptile Overview

Greptile Summary

Completes the argument reordering work from PR #2416 by updating _unpermute_bwd_with_merging_probs_kernel to follow the JAX-Triton compatible parameter pattern: input pointers, strides, output pointers, then metas.

• Moved fwd_input_grad_ptr and merging_probs_grad_ptr from input section to output section
• Moved num_experts and hidden_size from sizes section to metas section
• Updated the kernel call site to match the new parameter order
• All arguments correctly align between caller and kernel definition

Confidence Score: 5/5

• This PR is safe to merge with no risk
• This is a straightforward parameter reordering that completes the work from PR [PyTorch] Change arguments order in triton kernels to make jax-triton work #2416. The changes are mechanical, both the kernel definition and call site are updated consistently, and the argument alignment has been verified to be correct. No logic changes were made.
• No files require special attention

Important Files Changed

File Analysis

Filename	Score	Overview
transformer_engine/common/triton/permutation.py	5/5	Reordered kernel parameters to follow standard pattern: input pointers, strides, output pointers, then metas
transformer_engine/pytorch/triton/permutation.py	5/5	Updated kernel call to match reordered parameters, ensuring correct argument alignment

Sequence Diagram

sequenceDiagram
    participant Caller as PyTorch Layer<br/>(permutation.py)
    participant Wrapper as unpermute_with_mask_map_bwd_with_merging_probs<br/>(triton/permutation.py)
    participant Kernel as _unpermute_bwd_with_merging_probs_kernel<br/>(common/triton/permutation.py)

    Caller->>Wrapper: Call with tensors and parameters
    Note over Wrapper: Create output tensors:<br/>act_grad, merging_probs_grad
    Wrapper->>Kernel: Pass args in JAX-Triton compatible order:<br/>1. Input pointers (fwd_output_grad, fwd_input, etc.)<br/>2. Strides (8 stride parameters)<br/>3. Output pointers (act_grad, merging_probs_grad)<br/>4. Metas (num_experts, hidden_size, etc.)
    Note over Kernel: Process unpermute backward pass<br/>with merging probabilities
    Kernel-->>Wrapper: Writes to output tensors
    Wrapper-->>Caller: Return act_grad, merging_probs_grad

Loading

[greptile-apps]

_{2 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[ksivaman]

LGTM

[jberchtold-nvidia]

/te-ci L0 pytorch