Open
feat(qbmmv3): ScaleA+ScaleB dual NZ layout support #2
HuangRZzzzAdd on Aug 20
feat(qbmmv3): ScaleA+ScaleB dual NZ layout support #2
Open
HuangRZzzzAdd on Aug 20
HuangRZzzz
HuangRZzzzCollaborators
Aug 20

ScaleA NZ + ScaleB NZ 双 NZ 布局支持

修改内容

Wrapper (4文件):

  • qbmm_mx_tensor_api_blaze.h / without_batch / l0c_pingpong / streamk: 增加 ScaleALayout 模板参数

Kernel (quant_batch_matmul_v3.cpp):

  • QbmmV3ScaleALayout 类型别名 (FORMAT_SCALE_A=29 -> ZZLayoutPtn<C0=2>)
  • QBMMV3_SCALE_A_NZ_FORCE_BLAZE 宏: HIGH_LEVEL key 映射到 Blaze 实现

Host tiling:

  • tiling_base: isScaleANz/scaleAFormat 字段, SetFormat() 检测 pertoken NZ
  • AnalyzeInputs(): pertoken origin shape 包装 (isScaleANz 时)
  • tiling_cache: isScaleANz hash 位
  • checker: MxPertokenScaleShapeCheck/MxScaleShapeCheck/CheckBatchValidInMxPerGroupMode NZ skip
  • GetApiLevel: isScaleANz 返回 BLAZE_LEVEL

Golden:

  • format_transform hook: index 2 (ScaleB) + index 5 (ScaleA)
  • ScaleA unpack: x_before_kg=True
  • transA=true golden: 修复 x1_scale_mx reshape (M, Kg*2) + broadcast 后 swapaxes
  • customize_inputs: ScaleA NZ 跳过 sanitize

CSV:

  • ttk_kernel_qbmm_mxfp8_scaleAB_nz_st.csv: 29 cases, scale + pertoken 双 NZ

测试结果

ScaleAB NZ 29/29 PASS, 100% 精度, Blaze tensor API.
GM->L1: copy_gm_to_cbuf_align_v2, 无 nd2nz 二次转排.

依赖

  • ops-tensor: scaleAB-nz-layout 分支
  • ops-test-kit: format_transform hook
likedislike
Merge blocked