Skip to content

Model L: propagate a NaN through enc0's ReLU + FP8 step as the original does; converter check for non-finite weight codes #251

Open

Description

@phantoslayer21

Problem

Model L's first input linear ends in a ReLU and an FP8 conversion: x1 = q8(max(0, RN16(b + W q))). When the
accumulator is NaN (a NaN input code in q, or a NaN weight code), the original keeps the NaN through the fused
ReLU + FP8 conversion. Our kernels give 0: max(0, NaN) returns 0 on the hardware and in the HIP code.

This affects every implementation of that step alike:

  • main's kernels: clean/l/kernels/l_ops.hip (linear_32_32<RELU = true>) and its D3D12 ports in
    clean/dx12/kernels/l_ops_linear.hlsl (l_linear_32_32_relu, l_linear_32_32_fused, via
    clean/dx12/hlsl/l_lin_px.hlsli);
  • the encoder block with the linears as its prologue (clean/dx12/kernels/swin_block.hlsl SW_SRC == 3,
    clean/dx12/kernels/fused_swin.hlsl);
  • the hybrid twins of the same step (D3D12 and HIP).

The reference oracle (clean/l/reference/oracle.py, _out_round mode relu_q8) keeps the NaN, so the hybrid stage
check's random NaN-input cases differ from it (the second linear's chain and the pack12 coefficients downstream).
Finite inputs are unaffected.

Spec references: L P3 item 13; spec 05 edge cases. Spec 03 maps NaN inputs to 0, and real converted models contain no
FP8 NaN weight codes, so this is not visible on real data. It was kept out of 0.6.5.

Until it is fixed, the random-weight legs of clean/l/runner/l_hyb_stage_check.py draw finite codes only, and one
dedicated NaN-input leg is reported separately (listed, neither pass nor fail).

Proposed fix

  1. In every implementation listed above, make the ReLU + FP8 step keep a NaN as the original does (one form shared by
    main's kernels, the hybrid twins and HIP, so the byte checks between them stay identical).
  2. Fold the dedicated NaN-input leg back into the pass/fail random legs of l_hyb_stage_check.py once the kernels
    match the oracle.

Converter validation

Add a check to the model converter's validation: a real converted model must contain no non-finite weight codes (no FP8
NaN codes, no FP16 NaN or infinity in the converted blobs). A model that fails is refused with one line naming the stage
and blob.

Tests

  • l_hyb_stage_check.py: the NaN-input leg passes under the normal rule (then merged into the random legs).
  • The byte checks (HIP == D3D12) for the linears and the fused encoder block stay identical, NaN cases included.
  • The converter check: one synthetic model with a NaN weight code is refused; the real models pass.

Activity

  1. added this to the 0.6.6 milestone on Nov 16, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions