Merged
fix/a6-kc-quant-core-ratio #14301
huangyuxiaaaaaAdd on 28 days ago
fix/a6-kc-quant-core-ratio #14301
Merged
Pull Request has been successfully merged by CANN-robot.
(Thank you for huangyuxiaaaaa's contribution.)28 days ago Created pull request, commit 27341893
28 days ago Pre merge succeeded (commit_id: cd2fc07932bfa4f5a11a0abdc92eb915ac7190f1)
28 days ago Pre merge succeeded (commit_id: b9d6e918bbb78a6d9f79fc6eda41cbbad211e321)
28 days ago Added label: cann-cla/yes
56 messages are collapsed here。 See more
26 days ago Added label: approved
huangshuai59
26 days ago Comments:
26 days ago Comments:
/lgtm


CANN-robot
26 days ago Comments:
26 days ago Comments:
26 days ago Added label: lgtm
26 days ago Merged pull request
描述
all2allmm和mmall2all算子在A6上跑kc量化场景会出现卡死,修改相关核配比后精度pass
关联的Issue
测试
A5 RDV已过,上板精度pass
文档更新
类型标签
MC2 专项代码检视意见
针对 PR cann/ops-transformer#14301 的 MC2 变更完成专项静态检视。
2d715c31d8cc54973a6450f3fb773f8ed118844cupstream/master,merge-base2734189379775d7d39dcf5f823eebff27c87b7ef__NPU_ARCH__ == 9201)C:V=1:1,同时保留其他架构 C:V=1:2检视结论
未发现阻塞合入的问题。
本次修改在同一编译期架构条件下统一处理了以下三组依赖关系:
CV_RATIO。A6 只等待/通知一个 AIV,其他架构维持两个 AIV,未发现等待与通知数量不匹配。GetBlockNum() * AIV_TO_AIC_RATIO,A6 使用 1,其他架构使用 2,与实际 C:V 配比一致。ASCENDC_TPL_MIX_AIC_1_1,其他架构继续选择ASCENDC_TPL_MIX_AIC_1_2;APACE 路径仍保持ASCENDC_TPL_AIC_ONLY。关键证据
mc2/3rd/quant_batch_matmul_v3/op_kernel/quant_batch_matmul_v3_base.h:52-57:CV_RATIO在 9201 上为 1,其余架构为 2。mc2/3rd/quant_batch_matmul_v3/op_kernel/arch35/qbmm_mix_online_dynamic.h:97-109:第二组 AIC/AIV 同步仅在CV_RATIO == 2时执行。mc2/3rd/quant_batch_matmul_v3/op_kernel/arch35/qbmm_mix_online_dynamic.h:160-164:AIV block 映射按CV_RATIO计算。mc2/3rd/quant_batch_matmul_v3/op_kernel/arch35/qbmm_mix_online_dynamic.h:348-359,394-409:dequant 和 GM→UB 搬运使用相同的 M 切分及偏移计算。mc2/common/op_kernel/mc2_templates/quantization/fp8/fp8_dynamic_quant_pertoken.h:51-55,177-178:动态量化参与核数按架构配比计算。mc2/matmul_allto_all/op_kernel/arch35/matmul_allto_all_tiling_key.h:39-43,64-86:A6 的三种原生模式统一选用 1:1 kernel type。本地检查
pre-commit:通过;trailing whitespace、EOF、merge conflict、private key、clang-format、codespell、duplicate header、OAT 等适用检查全部通过。git diff --check upstream/master...HEAD:通过。origin/fix/a6-kc-quant-core-ratio一致。建议的动态验证
本地 Windows 环境不具备 A6 CANN 编译与多卡运行条件,建议合入前在 960DT(9201)补充以下验证:
singleCoreM=1、奇数 M、M 小于 AIC 核数、M 大于 AIC 核数,检查结果精度及无卡死。