| Files | Last commit | Last update |
|---|---|---|
feat: weight_quant_preprocess 新增 A8W4(F8F4) T-CG 数据流 MM_A8F4_TCG Co-authored-by: maqijun<maqijun@h-partners.com> # message auto-generated for no-merge-commit merge: !6068 merge feat/weight-quant-preprocess-f8f4-tcg into master feat: weight_quant_preprocess 新增 A8W4(F8F4) T-CG 数据流 MM_A8F4_TCG Created-by: maqijun Commit-by: maqijun Merged-by: cann-robot Description: ## 描述 为 weight_quant_preprocess(伪量化 Matmul 的前置参数预处理算子)新增 A8W4(F8F4) T-CG 数据流 MM_A8F4_TCG:x 为 FLOAT8_E4M3FN、weight 为 FLOAT4_E2M1(紧凑 4-bit)、weightScale 为 FP16/BF16 per-group({ceildiv(K,32),N},G>1),配套 QuantBatchMatmulV4 的 T-CG 伪量化场景,共 2 个子场景(weight 转置 → ND 直拷透传;weight 非转置 → ND→FRACTAL_NZ_C0_32 分形转换)。check/process 全部复用既有组件(含与 master 重名的存量符号统一加 9201 后缀),TCG 零新增 kernel。 ## 关联的Issue #3839 https://gitcode.com/cann/ops-math/issues/3839 ## 测试 二级冒烟(新增 MM_A8F4_TCG UT 用例全过,存量 91 例回归正常) ## 文档更新 docs/aclnnWeightQuantPreprocess.md ## 类型标签 - [ ] Bug修复 - [x] 新特性 - [ ] 性能优化 - [ ] 文档更新 - [ ] 其他,请描述: See merge request: cann/ops-math!6068 | 20 days ago |