Closed
[Bug-Report|缺陷反馈]: ascend960PR quant_batch_matmul_v3 MX/HIF4场景 MTE1 M/N拆分路径计算结果错误 #8904
chen-shuaiAdd on 5 days agoClosed at 4 days ago
5 days ago Added label: bug-report
chen-shuai
5 days ago Comments:
5 days ago Comments:
/assign


5 days ago Assign chen-shuai as responsible person
5 days ago Associated pull request
yuning_chen
5 days ago Comments:
5 days ago Comments:
收到,quant_batch_matmul_v3 在 Ascend960PR 的 MX 量化路径问题需要确认两件事:①是 MX 量化参数(scale 形态/block 大小)与文档不一致,还是实现缺失;②失败在 tiling 阶段还是 kernel 阶段(决定归属)。麻烦贴一下报错码与最小复现 shape。


4 days ago Close issue
4 days ago changed custom state from 待办的 to 已完成
3 days ago Added label: resolved
Describe the current behavior / 问题描述 (Mandatory / 必填)
ascend960(DAV_9202, L0C 512KB)上,quant_batch_matmul_v3 走 per-group scale(MXFP8/MXFP4/HIF4)且 baseMN 较大触发 MTE1 M/N 拆分时,输出与参考结果不一致。已定位三类错误模式:
Environment Information / 环境信息 (Mandatory / 必填)
昇腾硬件型号:Ascend 960PR
Steps to reproduce the issue / 重现步骤 (Mandatory / 必填)
Describe the expected behavior / 预期结果 (Mandatory / 必填)
拆分路径与不拆分路径(CPU 参考/其它已验证架构)结果一致;切点落在 L1 分形边界,两 half 偏移与 MTE2 写入几何一致,所有拆分场景数值正确。
Related log / screenshot / 日志 / 截图 (Mandatory / 必填)
NA
Special notes for this issue/备注 (Optional / 选填)