[JAX] Estimate post-RHT amax using regular amax

[jberchtold-nvidia]

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

[greptile-apps]

Greptile Overview

Greptile Summary

This PR modifies the post-RHT (Random Hadamard Transform) amax calculation in JAX quantization to use an approximation instead of the exact computation. The change replaces the call to calculate_post_rht_amax() with a simple multiplication: post_rht_amax = amax * 2.0.

• Performance optimization: Avoids the kernel call to RHTAmaxCalculationPrimitive for computing exact post-RHT amax
• Approximation method: Uses a 2x scaling factor on regular amax as a conservative estimate
• Dead code introduced: The hardcoded use_approx_amax = True flag leaves the original exact calculation path unreachable

Confidence Score: 4/5

• This PR is relatively safe - it's a performance optimization that uses a conservative 2x approximation factor
• The change is straightforward and the approximation approach (multiplying amax by 2.0) appears to be a reasonable conservative estimate. However, the hardcoded boolean creates unreachable dead code which could be cleaned up.
• No files require special attention beyond the style note about dead code

Important Files Changed

File Analysis

Filename	Score	Overview
transformer_engine/jax/cpp_extensions/quantization.py	4/5	Changed post-RHT amax calculation from exact computation to approximation using amax * 2.0. Introduces hardcoded use_approx_amax = True flag making the else branch dead code.

Sequence Diagram

sequenceDiagram
    participant Q as _quantize_dbias_impl
    participant A as calculate_amax
    participant K as TE Kernel
    
    Note over Q: When use_rht=True
    
    alt use_approx_amax=True (New Path)
        Q->>A: calculate_amax(x.data)
        A-->>Q: amax
        Q->>Q: post_rht_amax = amax * 2.0
    else use_approx_amax=False (Dead Code)
        Q->>K: calculate_post_rht_amax(x.data)
        K-->>Q: new_amax, post_rht_amax
    end
    
    Q->>K: PrimitiveClass.outer_primitive.bind(..., post_rht_amax)

Loading

[greptile-apps]

Additional Comments (1)

1. transformer_engine/jax/cpp_extensions/quantization.py, line 810 (link)

   style: Hardcoded use_approx_amax = True makes the else branch (lines 820-831) unreachable dead code. Consider either removing the dead code or making this configurable (e.g., via quantizer attribute or function parameter).

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}