Pass input_output_alias to TritonAutotunedKernelCall

[tdophung]

Description

https://nvbugspro.nvidia.com/bug/5810384
To remove the WAR that was put in place for this bug.

This should also serves as part 2 to WAR to the intermittent sort_chunks_by_index bug seen before in #2730

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

[greptile-apps]

Greptile Summary

This PR removes the workaround that was passing an empty input_output_aliases_with_sizes tuple to TritonAutotunedKernelCall, and replaces it with the correct alias computation. The workaround was masking a bug in jaxlib/gpu/triton_kernels.cc where the restore phase would unconditionally access input_copies[input_idx] even when the save phase had skipped saving (because XLA did not actually alias the buffers), leading to a CUDA_ERROR_INVALID_VALUE. The referenced internal bug (5810384) claims the C++ bug is now fixed.

Key changes:

• Computes (input_idx, buffer_output_idx, size_bytes) tuples for each alias entry in input_output_aliases, where buffer_output_idx = num_inputs + output_idx to match the C++ buffer-indexing convention (inputs first, then outputs).
• Passes the computed tuple to TritonAutotunedKernelCall instead of the previously hardcoded empty ().
• This also fixes a silent correctness issue: without proper save/restore, subsequent autotuning configs were running on already-overwritten input data, which could select the wrong config (the root cause of the intermittent sort_chunks_by_index bug from PR WAR sort_chunks_by_index intermittent failures in L0 JAX unitttest #2730).

Minor note: size_bytes is derived from the input aval's shape/dtype; an assertion that the aliased output has the same shape/dtype would make the assumption explicit and catch any unexpected future misuse.

Confidence Score: 5/5

• Safe to merge — the alias computation is logically correct and removes both the crash risk and the silent autotuning correctness bug.
• The only finding is a P2 style suggestion (add an assertion that aliased input/output shapes match). The core logic — computing buffer indices with num_inputs + output_idx and byte sizes with aval.size * itemsize — is correct. The WAR removal is well-motivated by the referenced C++ bugfix, and the previous workaround was already causing silent correctness issues in autotuning.
• No files require special attention.

Important Files Changed

Filename	Overview
transformer_engine/jax/triton_extensions/utils.py	Removes the empty-alias workaround for the jaxlib save/restore bug and properly computes input_output_aliases_with_sizes for TritonAutotunedKernelCall; correctness depends on the referenced C++ bug being fixed in all targeted jaxlib versions.

Sequence Diagram

sequenceDiagram
    participant Caller as triton_call_lowering
    participant Alias as alias computation (new)
    participant ATCK as TritonAutotunedKernelCall (C++)
    participant CUDA as CUDA runtime

    Caller->>Alias: iterate input_output_aliases.items()
    Alias-->>Caller: (input_idx, num_inputs+output_idx, size_bytes) tuples
    Caller->>ATCK: TritonAutotunedKernelCall(name, kernel_calls, aliases_with_sizes)

    loop for each autotuning config
        ATCK->>CUDA: cudaMemcpy — save aliased input buffer (save phase)
        ATCK->>CUDA: launch kernel config N
        ATCK->>CUDA: record timing
        ATCK->>CUDA: cudaMemcpy — restore original input buffer (restore phase)
    end

    ATCK-->>Caller: best config selected, correct input state preserved

Loading

_{Reviews (2): Last reviewed commit: "[pre-commit.ci] auto fixes from pre-comm..." | Re-trigger Greptile}