Remove old validation code for dequantization of nvfp4

[protonu]

Stacked on top of #5591.

This removes old validation code in favor of new dequantization of nvfp4 that was added in above mentioned PR.
No tests needed.

[github-actions]

Review updated until commit c6cf36a

Description

• Remove old nvfp4 dequantization validation code (unpack_fp4_bytes, e2m1_to_fp32)

• Replace with new dequantization functions (fp4_to_fp32, unpack_fp4)

• Update test files to use new function signatures

• Simplify dequantize_to_dtype implementation using dequantize_fp4

Changes walkthrough

Relevant files

Enhancement

test_cutlass_nvfp4_gemm.py Update test imports and function calls                                      tests/python/direct/test_cutlass_nvfp4_gemm.py Update imports to use fp4_to_fp32 and unpack_fp4 instead of unpack_fp4_bytes Modify function calls to use new signature with view(torch.uint8) +4/-3      test_python_frontend.py Update test imports and function calls                                      tests/python/direct/test_python_frontend.py Update imports to use fp4_to_fp32 and unpack_fp4 instead of unpack_fp4_bytes Modify function calls to use new signature with view(torch.uint8) +3/-2      narrow_precision.py Remove old dequantization validation code                                tests/python/direct_utils/narrow_precision.py Remove old validation code: kE2M1ToFloatArray, e2m1_to_fp32, unpack_fp4_bytes Update dequantize_to_dtype to use new dequantize_fp4 function Simplify implementation by removing manual unpacking logic +4/-47

PR Reviewer Guide

Here are some key observations to aid the review process:

🧪 PR contains tests

⚡ Recommended focus areas for review

Missing dequantize_fp4 function The code references a dequantize_fp4 function that is not defined in this file. This function appears to be part of the new dequantization implementation that should be available. Need to verify this function exists and is properly imported. out = dequantize_fp4( tensor_fp4.view(torch.uint8), tensor_sf, (6.0 * 448.0) / global_scale ) Import consistency The imports are changed to use fp4_to_fp32 and unpack_fp4 from python.direct_utils, but these functions should be verified to exist and provide the same functionality as the removed unpack_fp4_bytes. fp4_to_fp32, unpack_fp4, Import path verification The imports are changed to use fp4_to_fp32 and unpack_fp4 from python.direct_utils.narrow_precision. Need to ensure these functions are properly exported from this module and provide equivalent functionality. fp4_to_fp32, unpack_fp4,

[protonu]

!test

[greptile-apps]

Greptile Overview

Greptile Summary

This PR removes deprecated FP4 (nvfp4) dequantization code in favor of the newer functions (fp4_to_fp32, unpack_fp4, dequantize_fp4) that were added in PR #5591.

• Removed kE2M1ToFloatArray, e2m1_to_fp32, and unpack_fp4_bytes functions from narrow_precision.py
• Refactored dequantize_to_dtype to delegate to dequantize_fp4 instead of manually implementing the dequantization logic
• Updated test files to use the new fp4_to_fp32(unpack_fp4(...)) pattern instead of the removed unpack_fp4_bytes
• The new implementation is mathematically equivalent to the old code, using a lookup table approach for FP4→FP32 conversion

Confidence Score: 5/5

• This PR is safe to merge - it's a straightforward code cleanup that removes redundant validation code in favor of newer, equivalent implementations.
• The changes remove duplicate code and consolidate FP4 dequantization into the newer functions. Mathematical equivalence between old and new implementations has been verified. The changes are limited to test utility files and don't affect production code paths.
• No files require special attention

Important Files Changed

File Analysis

Filename	Score	Overview
tests/python/direct_utils/narrow_precision.py	5/5	Removed deprecated kE2M1ToFloatArray, e2m1_to_fp32, and unpack_fp4_bytes functions. Refactored dequantize_to_dtype to use the newer dequantize_fp4 function. The new implementation is mathematically equivalent to the old code.
tests/python/direct/test_cutlass_nvfp4_gemm.py	5/5	Updated imports and usages to use new fp4_to_fp32 and unpack_fp4 functions instead of removed unpack_fp4_bytes. The new function chain produces equivalent results.
tests/python/direct/test_python_frontend.py	5/5	Updated imports and test reference calculation to use new fp4_to_fp32(unpack_fp4(...)) instead of removed unpack_fp4_bytes. Functionally equivalent change.

Sequence Diagram

sequenceDiagram
    participant Test as Test Code
    participant DP as narrow_precision.py
    participant U as unpack_fp4()
    participant F as fp4_to_fp32()
    participant D as dequantize_fp4()

    Note over Test,D: Old Flow (Removed)
    Test->>DP: unpack_fp4_bytes(tensor)
    DP->>DP: e2m1_to_fp32() per element
    DP-->>Test: float32 tensor

    Note over Test,D: New Flow
    Test->>U: unpack_fp4(tensor.view(uint8))
    U-->>Test: unpacked nibbles
    Test->>F: fp4_to_fp32(unpacked)
    F-->>Test: float32 tensor (via LUT)

    Note over Test,D: dequantize_to_dtype (Refactored)
    Test->>DP: dequantize_to_dtype(tensor, sf, scale)
    DP->>D: dequantize_fp4(tensor, sf, amax)
    D->>U: unpack_fp4()
    D->>F: fp4_to_fp32()
    D-->>DP: scaled float32 tensor
    DP-->>Test: reshaped result

Loading

[greptile-apps]

Additional Comments (1)

1. tests/python/direct_utils/narrow_precision.py, line 14-54 (link)

   logic: removing unpack_fp4_bytes, e2m1_to_fp32, and kE2M1ToFloatArray will break tests that still import and use these functions:

   • tests/python/direct/test_python_frontend.py:2700 uses unpack_fp4_bytes
   • tests/python/direct/test_cutlass_nvfp4_gemm.py:155-156 uses unpack_fp4_bytes

   these test files need to be updated to either use the new dequantize_fp4 function or have unpack_fp4_bytes remain available

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[protonu]

!test

[greptile-apps]

_{3 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[jjsjann123]

LGTM

[greptile-apps]

_{3 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[protonu]

!test

[greptile-apps]

_{3 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[protonu]

!test

[greptile-apps]

_{3 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}