Closed
[Requirement|需求建议]: norm系4算子新增ascend350平台支持(batch_norm_grad/sync_bn_training_update/rms_norm_quant_v2/rms_norm_quant_v3) #7464
zhuzixian-lrAdd on  26 days agoClosed at  22 days ago
zhuzixian-lr
zhuzixian-lrCollaborators
26 days ago Create

Backgroud(背景信息)

350 平台适配需求清单核对后,norm 系第四波 4 个算子在 ascend350 上零支持(def/CMake/json/kernel 注册均无 ascend350),350 上算子不可用。4 算子在 950/960dt 均已有支持。

涉及算子:batch_norm_grad、sync_bn_training_update、rms_norm_quant_v2、rms_norm_quant_v3(均位于 ops-nn norm/ 目录)。

Origin(信息来源)

Ascend350 平台适配需求(81 项业务清单与前几波适配核对后的第四波主清单)。

Benefit / Necessity (价值/作用)

350 与 950 同属 DAV_3510(arch35),同源 kernel/tiling 可复用;补齐后 norm 系 BN 反传/rms 量化族在 350 可用,支撑 350 业务落地。

Design(设计方案)

四件套适配:

  1. def 增加 AddConfig("ascend350")(复用 950 同一 config 对象)
  2. config/ascend350/ 平行目录新增 binary.json(cp 自 950,语义一致)
  3. sync_bn_training_update/rms_norm_quant_v3 顶层 CMake SUPPORT_COMPUTE_UNIT/TILING_DIR 补 ascend350/arch35;rms_norm_quant_v3 op_kernel COMPUTE_UNITS 补 ascend350
  4. binary_config 登记 BatchNormGrad/RmsNormQuantV2/SyncBNTrainingUpdate 的 ascend350(编译选项严格对标 950)

验证:TTK kernel 12 条 + aclnn 1 条(aclnnRmsNormQuantV3)在 950 全 PASS(精度 100%)。

关联 PR:https://gitcode.com/cann/ops-nn/merge_requests/12408

likedislike
zhuzixian-lrzhuzixian-lrCollaborators
26 days ago Added label: requirement
zhuzixian-lr
zhuzixian-lrCollaborators
26 days ago Comments:
likedislike
CANN-robotCANN-robotCollaborators
26 days ago Assign zhuzixian-lr as responsible person
zhuzixian-lrzhuzixian-lrCollaborators
25 days ago Associated pull request
zhuzixian-lrzhuzixian-lrCollaborators
25 days ago Deleted associated pull request: feat: 4个算子新增ascend350平台支持(norm系第四波)
zhuzixian-lrzhuzixian-lrCollaborators
25 days ago Associated pull request
zhuzixian-lrzhuzixian-lrCollaborators
22 days ago changed custom state from 进行中 to 已完成
zhuzixian-lrzhuzixian-lrCollaborators
22 days ago Close issue