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

Aberta
#251 0 comentários 0 reações 0 responsáveis Ver no GitHub

Mantenedores costumam responder em até 1 dia

Ninguém assumiu esta issue ainda.

Avaliação

Dificuldade
5/5
Tempo estimado
Mais de uma semana
Facilidade para iniciantes
35/100
Tipo de issue
Bug
Clareza
Claramente especificada
Status de atividade
Ativa
Stack de tecnologia
cpp, python

Direção de pesquisa

Start with clean/l/reference/oracle.py and its _out_round relu_q8 mode, then compare the listed HIP and D3D12 kernels and the encoder block paths. Trace how the NaN-input case reaches the second linear and pack12 coefficients, and inspect clean/l/runner/l_hyb_stage_check.py for the separately reported leg. Done when the NaN cases match the oracle across implementations and the converter rejects non-finite weight codes with the stage and blob named.

Escrita pelo modelo de indexação a partir do texto da issue.

Descrição

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.
Linguagem predominante
C++
Estrelas
12
Forks
2
Merge médio
3h 10min
PRs com merge (30d)
171

Preparar o ambiente

Primeiros passos

  1. Leia a issue inteira e depois o guia de contribuição do projeto.
  2. Comente na issue dizendo que vai assumir — evita que duas pessoas façam o mesmo trabalho.
  3. Faça um fork do repositório e trabalhe em uma branch.
  4. Abra um pull request que referencie o número da issue.

Mais de phantoslayer21/wronggpu

Todas as issues de phantoslayer21/wronggpu

Issues semelhantes

Mais issues de C++

Receba novas issues na sua caixa de entrada

Um resumo curto de issues do GitHub para quem está começando.