| feat(swiglu_quant): Ascend950 ACLNN SwiGluQuantV2 with Ops-aligned scale output ## Summary Migrate SwigluQuant on Ascend 950 to CANN aclnnSwiGluQuantV2 while keeping Ascend 910B on the existing OpsRunner (AscendC) path. Unify the second output (scale) semantics with 910B via aclnnInplaceReciprocal after the ACLNN kernel. Consolidate CSV coverage into a single swiglu_quant.csv with SOC-specific cases. ## Motivation - Reuse CANN-maintained SwiGluQuantV2 on 950 instead of maintaining a separate kernel path for that platform. - Preserve upstream API compatibility: SwigluQuantParam and tensor I/O unchanged. - Keep 910B behavior and golden stable on OpsRunner (UB limits, scale = dequant). - Expose ACLNN scale as dequant scale (y_max/127) on 950 to match 910B consumers. Co-authored-by: ivanshan05075<shanzidan@h-partners.com> # message auto-generated for no-merge-commit merge: !2097 merge br_personal/ivan/switch_swiglu_quant into master feat(swiglu_quant): switch swigluquant to ACLNN aclnnSwiGluQuantV2 on 950 Created-by: ivanshan05075 Commit-by: ivanshan05075 Merged-by: cann-robot Description: ## 描述 <!--在这里详细描述你的改动，包括改动的原因和所采取的方法。--> SwigluQuant 950 芯片切换 aclnnSwiGluQuantV2 + scale 输出 拼接 aclnnInplaceReciprocal保证输出一致 ## checklist <!-- [x] 表示选中 --> - [x] 是否通过本地IDE对代码进行静态检查 - [x] 是否通过本地IDE对代码进行格式化处理 - [x] 是否进行空指针校验 - [x] 是否进行返回值校验 - [x] 是否正确释放new/malloc申请的内存 - [x] 是否充分考虑接口的异常场景 - [x] 是否正确记录错误日志 See merge request: cann/ascend-transformer-boost!2097 | 1 个月前 |
