Greptile Summary

This PR adds unswizzle support for MXFP8 and NVFP4 scaling factors, providing the inverse operation to the existing nvte_swizzle_scaling_factors API. It introduces nvte_unswizzle_scaling_factors and nvte_multi_tensor_unswizzle_scaling_factors, along with the corresponding GPU kernels (unswizzle_row_scaling_kernel_impl, unswizzle_col_scaling_kernel_impl) and byte-level inverse shuffle helpers (regs_unshuffle, regs_unshuffle_with_bit_shifts). Tests covering standalone unswizzle and swizzle→unswizzle round-trips are also added.

Key observations:

• The two new inverse shuffle functions (regs_unshuffle and regs_unshuffle_with_bit_shifts) are mathematically correct inverses of their forward counterparts for all supported LType variants (int, int2, int4).
• Bug in test helpers: In both performTestUnswizzle1D and performTestSwizzleUnswizzleRoundtrip, the variables SF_MODE_X and SF_MODE_Y are uninitialized when the !(rowwise || columnwise) branch of the skip guard is taken, causing undefined behaviour in the GTEST_SKIP message.
• The rowwise_swizzle/columnwise_swizzle variable names in multi_tensor_unswizzle_scaling_factors should be named rowwise_unswizzle/columnwise_unswizzle to avoid confusion about data-flow direction.
• Minor: both skip messages are missing a space before "is not implemented.", and the regs_unshuffle_with_bit_shifts function body is not followed by a blank line before the next template.

Confidence Score: 4/5

• Production kernel logic is sound; two UB issues in test helpers should be fixed before merging.
• The core GPU kernels and inverse shuffle helpers are mathematically correct and mirror the existing swizzle structure. The only defects found are confined to test code: two instances of reading uninitialized variables in GTEST_SKIP messages (undefined behaviour, though unlikely to cause silent data corruption in production). The remaining issues are cosmetic naming/style concerns. No functional regression risks exist in the shipped library code.
• tests/cpp/operator/test_swizzle.cu — fix uninitialized SF_MODE_X/SF_MODE_Y before merging.

Important Files Changed

Sequence Diagram

sequenceDiagram
    participant Caller
    participant nvte_unswizzle_scaling_factors
    participant unswizzle_scaling_factors
    participant unswizzle_scaling_kernel
    participant unswizzle_row_impl as unswizzle_row_scaling_kernel_impl
    participant unswizzle_col_impl as unswizzle_col_scaling_kernel_impl

    Caller->>nvte_unswizzle_scaling_factors: input (swizzled), output (compact), stream
    nvte_unswizzle_scaling_factors->>unswizzle_scaling_factors: convertNVTETensorCheck()
    unswizzle_scaling_factors->>unswizzle_scaling_factors: validate scaling_mode, dtype, shapes
    alt rowwise_unswizzle
        unswizzle_scaling_factors->>unswizzle_scaling_kernel: launch<<<grid,block,slm,stream>>>
        unswizzle_scaling_kernel->>unswizzle_row_impl: row_scaling=true
        unswizzle_row_impl->>unswizzle_row_impl: load tiles to SLM
        unswizzle_row_impl->>unswizzle_row_impl: regs_unshuffle()
        unswizzle_row_impl->>unswizzle_row_impl: write compact output
    else columnwise_unswizzle
        unswizzle_scaling_factors->>unswizzle_scaling_kernel: launch<<<grid,block,slm,stream>>>
        unswizzle_scaling_kernel->>unswizzle_col_impl: row_scaling=false
        unswizzle_col_impl->>unswizzle_col_impl: load tiles to SLM
        unswizzle_col_impl->>unswizzle_col_impl: regs_unshuffle_with_bit_shifts()
        unswizzle_col_impl->>unswizzle_col_impl: write compact output
    end
    unswizzle_scaling_factors-->>Caller: compact scale_inv in output

_{Last reviewed commit: 621bc16}