| Files | Last commit | Last update |
|---|---|---|
Moe目录日志规范整改 Co-authored-by: guoqiuhao<guoqiuhao1@huawei.com> # message auto-generated for no-merge-commit merge: !10386 merge log into master Moe目录日志规范整改 Created-by: guoqiuhao Commit-by: guoqiuhao Merged-by: cann-robot Description: ## 描述 Moe目录日志规范整改. 1.禁止中文日志 2.英文拼写错误 3.日志级别错误 4.语法错误 5.禁止自定义打印 6.错误描述不完整 7.冗余日志 8.未知缩写 9.内部敏感词 10.缺少度量单位 ## 关联的Issue [#4476](https://gitcode.com/cann/ops-transformer/issues/4476) ## 测试 不涉及 ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10386 | 2 months ago | |
增加moe蓝区ttk用例适配 Co-authored-by: cpy_123456<chenpengyi@huawei.com> # message auto-generated for no-merge-commit merge: !10344 merge ut into master 增加moe蓝区ttk用例适配 Created-by: cpy_123456 Commit-by: cpy_123456 Merged-by: cann-robot Description: ## 描述 # MoE 算子 ST 用例选取文档 ## 1. 选取原则 - **ACLNN 模式**:每个算子选取约 20 个用例(ACLNN 调用开销较大,约 2s/用例) - **Kernel 模式**:每个算子选取约 30 个用例(Kernel 编译+执行较快,约 0.4s/用例) - 覆盖所有 dtype(float32/float16/bfloat16/int8/int32/int64) - 覆盖所有关键属性组合(dropPadMode/quantMode/normType/renorm 等) - 覆盖可选输入的组合(有/无 optional inputs) - 优先选取小 shape 用例(减少运行时间) - 排除已知精度边界用例(量化精度差异、bf16 reduction 差异等) - 排除依赖算子代码修改的用例(如 moe_token_unpermute_grad active 场景) - 总运行时间控制在 10 分钟以内 ## 2. 选取结果 ### ACLNN 模式(12 个 CSV,150 个用例) | CSV 文件 | 算子目录 | 选取/全量 | 运行时间 | PASS | |---------|---------|----------|---------|------| | test_aclnn_moe_gating_top_k_arch35.csv | moe_gating_top_k | 20/38 | 42s | 20/20 | | test_aclnn_moe_gating_top_k_v2_arch35.csv | moe_gating_top_k | 20/32 | 46s | 20/20 | | test_aclnn_moe_gating_top_k_softmax_v2_arch35.csv | moe_gating_top_k_softmax_v2 | 20/232 | 47s | 20/20 | | test_aclnn_moe_init_routing_arch35.csv | moe_init_routing | 10/10 | 36s | 10/10 | | test_aclnn_moe_init_routing_v2_arch35.csv | moe_init_routing_v2 | 20/27 | 57s | 20/20 | | test_aclnn_moe_init_routing_v2_grad_arch35.csv | moe_init_routing_v2_grad | 20/116 | 51s | 20/20 | | test_aclnn_moe_init_routing_quant_v2_arch35.csv | moe_init_routing_quant_v2 | 15/34 | 48s | 15/15 | | test_aclnn_moe_finalize_routing_v2_grad_arch35.csv | moe_finalize_routing_v2_grad | 11/94 | 30s | 11/11 | | test_aclnn_moe_token_permute_arch35.csv | moe_token_permute | 4/4 | 19s | 4/4 | | test_aclnn_moe_token_permute_grad_arch35.csv | moe_token_permute_grad | 4/4 | 18s | 4/4 | | test_aclnn_moe_token_unpermute_arch35.csv | moe_token_unpermute | 2/2 | 15s | 2/2 | | test_aclnn_moe_token_unpermute_grad_arch35.csv | moe_token_unpermute_grad | 4/4 | 19s | 4/4 | ### Kernel 模式(10 个 CSV,300 个用例) | CSV 文件 | 算子目录 | 选取/全量 | 运行时间 | PASS | |---------|---------|----------|---------|------| | test_moe_gating_top_k_arch35.csv | moe_gating_top_k | 30/829 | 11s | 30/30 | | test_moe_gating_top_k_softmax_v2_arch35.csv | moe_gating_top_k_softmax_v2 | 30/832 | 13s | 30/30 | | test_moe_init_routing_arch35.csv | moe_init_routing | 30/203 | 13s | 30/30 | | test_moe_init_routing_v2_arch35.csv | moe_init_routing_v2 | 30/200 | 13s | 30/30 | | test_moe_init_routing_v2_grad_arch35.csv | moe_init_routing_v2_grad | 30/144 | 11s | 30/30 | | test_moe_init_routing_quant_v2_arch35.csv | moe_init_routing_quant_v2 | 30/190 | 13s | 30/30 | | test_moe_finalize_routing_v2_arch35.csv | moe_finalize_routing_v2 | 30/112 | 13s | 30/30 | | test_moe_finalize_routing_v2_grad_arch35.csv | moe_finalize_routing_v2_grad | 30/198 | 13s | 30/30 | | test_moe_compute_expert_tokens_arch35.csv | moe_compute_expert_tokens | 30/159 | 13s | 30/30 | | test_moe_re_routing_arch35.csv | moe_re_routing | 30/578 | 11s | 30/30 | ## 3. 验证汇总 | 指标 | 值 | |------|------| | 选取用例总数 | 450 | | 全量用例总数 | 3726 | | PASS | 450 (100%) | | FAIL | 0 | | ACLNN 运行时间 | 428s (7m08s) | | Kernel 运行时间 | 124s (2m04s) | | 总运行时间 | 552s (9m12s) | | 平均每用例 | 1.23s | ## 4. 排除用例说明 ## 5. 功能覆盖说明 ### moe_gating_top_k - **norm_type**:覆盖 0(softmax)/1(sigmoid)/2(sqrtsoftplus) - **group_count**:覆盖 1(无分组)/8(标准分组)/大值(多分组) - **group_select_mode**:覆盖 0(amax)/1(top2sum) - **out_flag**:覆盖 True/False - **renorm**:覆盖 0/1 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_gating_top_k_v2 - **normType**:覆盖 1(sigmoid) - **inputIdsOptional/tid2eidOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_gating_top_k_softmax_v2 - **renorm**:覆盖 0(softmax后取topk)/1(topk后renorm) - **finishedOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing - **active_num**:覆盖小值(1~5)/大值(100+) - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_v2 - **drop_pad_mode**:覆盖 0(dropless)/1(drop/pad) - **expert_tokens_count_or_cumsum_flag**:覆盖 0/1(cumsum)/2(count) - **expert_tokens_before_capacity_flag**:覆盖 True/False - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_v2_grad - **drop_pad_mode**:覆盖 0/1 - **active_num**:覆盖 0(full)/1(active)/大值 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_quant_v2 - **drop_pad_mode**:覆盖 0/1 - **quant_mode**:覆盖 0(静态量化)/1(动态量化) - **expert_tokens_count_or_cumsum_flag**:覆盖 0/1/2 - **expert_tokens_before_capacity_flag**:覆盖 True/False - **dtype**:覆盖 fp32/fp16/bf16 ### moe_finalize_routing_v2 - **drop_pad_mode**:覆盖 0/2/3 - **zero/copy/constant_expert_range**:覆盖各种 expert range 配置 - **optional inputs(x/a1/a2/v)**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_finalize_routing_v2_grad - **drop_pad_mode**:覆盖 0/1 - **active_num**:覆盖 0/1/大值 - **scales**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_compute_expert_tokens - **num_experts**:覆盖小值(8)/大值(800+) - **dtype**:int32 ### moe_re_routing - **idx_type**:覆盖 0/1 - **expert_token_num_type**:覆盖 0/1 - **per_token_scales**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16/int8/int32/int64 ### moe_token_permute - **numOutTokens**:覆盖 0(full)/10(active)/20(active) - **paddedMode**:False - **dtype**:覆盖 fp32/fp16 ### moe_token_permute_grad - **numTopk**:覆盖 8 - **paddedMode**:False - **dtype**:覆盖 fp32/fp16 ### moe_token_unpermute - **paddedMode**:False - **probsOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16 ### moe_token_unpermute_grad - **paddedMode**:False - **probsOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16 ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4472 ## 测试 <!--描述进行了哪些测试来验证你的改动。包括但不限于二级冒烟、算子泛化等。--> ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [x] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [x] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10344 | 2 months ago | |
Moe目录日志规范整改 Co-authored-by: guoqiuhao<guoqiuhao1@huawei.com> # message auto-generated for no-merge-commit merge: !10386 merge log into master Moe目录日志规范整改 Created-by: guoqiuhao Commit-by: guoqiuhao Merged-by: cann-robot Description: ## 描述 Moe目录日志规范整改. 1.禁止中文日志 2.英文拼写错误 3.日志级别错误 4.语法错误 5.禁止自定义打印 6.错误描述不完整 7.冗余日志 8.未知缩写 9.内部敏感词 10.缺少度量单位 ## 关联的Issue [#4476](https://gitcode.com/cann/ops-transformer/issues/4476) ## 测试 不涉及 ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10386 | 2 months ago | |
moe_init_routing和moe_finalize_routing支持自动反向 Co-authored-by: JimmyLam2<linjingmin2@huawei.com> # message auto-generated for no-merge-commit merge: !9881 merge master into master moe_init_routing和moe_finalize_routing支持自动反向 Created-by: JimmyLam2 Commit-by: JimmyLam2 Merged-by: cann-robot Description: ## 描述 moe_init_routing和moe_finalize_routing支持自动反向 ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4359 ## 测试 不涉及ST与二级冒烟。采用构造torch用例的方式进行验证。  ## 文档更新 新增 1.torch_extension/cann_ops_transformer/docs/zh/moe_finalize_routing_grad.md 2.torch_extension/cann_ops_transformer/docs/zh/moe_init_routing_grad.md 修改 1.torch_extension/cann_ops_transformer/docs/zh/moe_finalize_routing.md 2.torch_extension/cann_ops_transformer/docs/zh/moe_init_routing.md ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [x] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!9881 | 2 months ago | |
添加moe_init_routing_v4和moe_re_routing_v2算子的st&标杆 Co-authored-by: cpy_123456<chenpengyi@huawei.com> # message auto-generated for no-merge-commit merge: !10530 merge ut into master 添加moe_init_routing_v4和moe_re_routing_v2算子的st&标杆 Created-by: cpy_123456 Commit-by: cpy_123456 Merged-by: cann-robot Description: ## 描述 添加moe_init_routing_v4和moe_re_routing_v2算子的st&标杆 ## 一、moe_init_routing_v4(25 个用例,全部 PASS) ### 1.1 场景覆盖矩阵 | 维度 | 覆盖值 | 用例数 | |---|---|---| | **输入数据类型 (x)** | int8 | 2 | | | float16 | 7 | | | bfloat16 | 8 | | | float32 | 8 | | **quant_mode** | -1(非量化,DropLess) | 13 | | | 1(动态量化) | 12 | | **expert_tokens_num_type (etnt)** | 1(直方图统计) | 13 | | | 2(key-value 对) | 12 | | **drop_pad_mode** | 0(DropLess 模式) | 25 | | **expert_tokens_num_flag** | True | 25 | | **row_idx_type** | 0(gather) | 7 | | | 1(scatter) | 18 | | **scale 输入** | 有 | 10 | | | 无 | 15 | | **topk_weight 输入** | 有 | 25 | | **expert_num 范围** | 256 ~ 2930 | — | | **active_expert_range 跨度** | 6 ~ 580 | — | ## 二、moe_re_routing_v2(25 个用例,全部 PASS) ### 2.1 场景覆盖矩阵 | 维度 | 覆盖值 | 用例数 | |---|---|---| | **tokens 数据类型** | int8 | 8 | | | float16 | 8 | | | bfloat16 | 9 | | **idx_type** | 0(正向索引,permute_token_idx=src_offset) | 12 | | | 1(反向索引,permute_token_idx=dst_offset) | 13 | | **per_token_scales 输入** | 有 | 12 | | | 无 | 13 | | **expert_topk_weight 输入** | 有 | 25 | | **rank_num(专家并行数)** | 1 | 17 | | | 2 | 3 | | | 3 | 1 | | | 4~5 | 2 | | | 7~52 | 2 | | **expert_num(每 rank 专家数)** | 1 | 17 | | | 2 | 5 | | | 3 | 1 | | | 5~6 | 2 | | **tokens 第二维 (hidden_size)** | 9 ~ 15267 | — | ## 关联的Issue <!-- 如果这个PR是为了解决特定的Issue,请在这里提供Issue链接。例如:关联Issue #000--> <!-- 如果这个PR是为了解决特定的问题单,请在这里描述问题单单号。--> ## 测试 st&标杆通过 ## 文档更新 https://gitcode.com/cann/ops-transformer/issues/4543 ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [x] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10530 | 2 months ago | |
Moe目录日志规范整改 Co-authored-by: guoqiuhao<guoqiuhao1@huawei.com> # message auto-generated for no-merge-commit merge: !10386 merge log into master Moe目录日志规范整改 Created-by: guoqiuhao Commit-by: guoqiuhao Merged-by: cann-robot Description: ## 描述 Moe目录日志规范整改. 1.禁止中文日志 2.英文拼写错误 3.日志级别错误 4.语法错误 5.禁止自定义打印 6.错误描述不完整 7.冗余日志 8.未知缩写 9.内部敏感词 10.缺少度量单位 ## 关联的Issue [#4476](https://gitcode.com/cann/ops-transformer/issues/4476) ## 测试 不涉及 ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10386 | 2 months ago | |
安装重名头文件整改:common.h Co-authored-by: chenyifan<chenyifan66@h-partners.com> # message auto-generated for no-merge-commit merge: !10537 merge clear_repeat_h_comm into master 安装重名头文件整改:common.h Created-by: mutex_lock Commit-by: chenyifan Merged-by: cann-robot Description: ## 描述 按**安装重名**头文件规范,完成 ** common.h / moe_common.h / moe_v2_common.h 跨算子同名整改**,消除安装包内跨算子同名头冲突。 **原因**: **common.h(13 份)**、**moe_common.h(5 份)**、**moe_v2_common.h(7 份)** 属安装规则 **3.9**(仓内非对外、不同算子/组件目录拷贝同名头),且口径为「同名不同实现亦改名」。权威 common/include/op_kernel/common.h 与 mc2/common/op_kernel/moe_distribute_base.h **保留原名**。 **方法**:低风险只改次要副本——各算子/组件侧 common.h / moe_common.h / moe_v2_common.h / setup 侧 moe_distribute_base.h 加算子或组件前缀(如 agm_v2_、matmul_all_reduce_、moe_init_routing_、mega_moe_、setup_ 等);同步更新 #include、\file、include guard。无业务逻辑变更。禁止软链接。 ### 安装分类规则 | # | 场景 | 是否整改 | 处置 | |---|------|---------|------| | 3.9 | 仓内非对外同名拷贝 | 是 | 次要侧改名 | | 权威 common | common/include/op_kernel/common.h | 否 | 保留原名 | | 权威 mc2 base | mc2/common/op_kernel/moe_distribute_base.h | 否 | 保留原名(v2/combine 等仍 #include 此路径) | ### 范围与结果 | 项 | 说明 | |----|------| | 重命名组数 | **26** | | 涉及域 | attention(scatter_pa / sfag / sfmg)、examples fkle、mc2(agm_v2 / matmul_all_reduce / mega_moe / dispatch setup/teardown/v2、mhc tiling)、moe(gating / init_routing / token_permute 系列) | | 附带修复 | dispatch setup base/common 映射;teardown 跨组件引用;v2 arch35 a5_ccu 裸 include;v3/combine_v3 apt 依赖 v2/v2_combine 的 __has_include 多布局路径 | | 变更文件数 | **166** | | 行变更 | +253 / -238 | ### 26 组重命名(旧 → 新) | # | 旧路径 | 新路径 | |--:|--------|--------| | 1 | attention/scatter_pa_kv_cache/op_kernel/arch35/common.h | attention/scatter_pa_kv_cache/op_kernel/arch35/scatter_pa_kv_cache_arch35_common.h | | 2 | attention/sparse_flash_attention_grad/op_kernel/arch35/common.h | attention/sparse_flash_attention_grad/op_kernel/arch35/sparse_flash_attention_grad_arch35_common.h | | 3 | attention/sparse_flash_mla_grad/op_kernel/arch35/common.h | attention/sparse_flash_mla_grad/op_kernel/arch35/sparse_flash_mla_grad_arch35_common.h | | 4 | examples/fast_kernel_launch_example/csrc/moe_distribute_dispatch_v2/ascend910_93/op_kernel/common.h | examples/fast_kernel_launch_example/csrc/moe_distribute_dispatch_v2/ascend910_93/op_kernel/fkle_mdd_v2_common.h | | 5 | mc2/all_gather_matmul_v2/op_kernel/common.h | mc2/all_gather_matmul_v2/op_kernel/agm_v2_common.h | | 6 | mc2/matmul_all_reduce/op_kernel/common.h | mc2/matmul_all_reduce/op_kernel/matmul_all_reduce_common.h | | 7 | mc2/moe_distribute_dispatch_setup/op_kernel/common.h | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_common.h | | 8 | mc2/moe_distribute_dispatch_v2/op_kernel/common.h | mc2/moe_distribute_dispatch_v2/op_kernel/moe_distribute_dispatch_v2_common.h | | 9 | mhc/mhc_pre_sinkhorn_backward/op_host/op_tiling/arch22/common.h | mhc/mhc_pre_sinkhorn_backward/op_host/op_tiling/arch22/mhc_pre_sinkhorn_backward_common.h | | 10 | moe/moe_gating_top_k/op_kernel/arch35/common.h | moe/moe_gating_top_k/op_kernel/arch35/moe_gating_top_k_arch35_common.h | | 11 | moe/moe_gating_top_k/op_kernel/common.h | moe/moe_gating_top_k/op_kernel/moe_gating_top_k_common.h | | 12 | moe/moe_gating_top_k_backward/op_kernel/arch35/common.h | moe/moe_gating_top_k_backward/op_kernel/arch35/moe_gating_top_k_backward_arch35_common.h | | 13 | moe/moe_gating_top_k_backward/op_kernel/common.h | moe/moe_gating_top_k_backward/op_kernel/moe_gating_top_k_backward_common.h | | 14 | moe/moe_init_routing/op_kernel/moe_common.h | moe/moe_init_routing/op_kernel/moe_init_routing_common.h | | 15 | moe/moe_init_routing/op_kernel/arch35/moe_common.h | moe/moe_init_routing/op_kernel/arch35/moe_init_routing_arch35_common.h | | 16 | moe/moe_token_permute/op_kernel/moe_common.h | moe/moe_token_permute/op_kernel/moe_token_permute_common.h | | 17 | moe/moe_token_permute_with_ep/op_kernel/moe_common.h | moe/moe_token_permute_with_ep/op_kernel/moe_token_permute_with_ep_common.h | | 18 | moe/moe_token_permute_with_routing_map/op_kernel/moe_common.h | moe/moe_token_permute_with_routing_map/op_kernel/moe_token_permute_with_routing_map_common.h | | 19 | mc2/mega_moe/op_kernel/arch22/moe_init_routing_v2/moe_v2_common.h | mc2/mega_moe/op_kernel/arch22/moe_init_routing_v2/mega_moe_moe_v2_common.h | | 20 | mc2/mega_moe/op_kernel/arch22/moe_init_routing_quant_v2/moe_v2_common.h | mc2/mega_moe/op_kernel/arch22/moe_init_routing_quant_v2/mega_moe_quant_moe_v2_common.h | | 21 | moe/moe_init_routing_v2/op_kernel/moe_v2_common.h | moe/moe_init_routing_v2/op_kernel/moe_init_routing_v2_common.h | | 22 | moe/moe_init_routing_v2/op_kernel/arch35/moe_v2_common.h | moe/moe_init_routing_v2/op_kernel/arch35/moe_init_routing_v2_arch35_common.h | | 23 | moe/moe_init_routing_quant_v2/inner/moe_v2_common.h | moe/moe_init_routing_quant_v2/inner/moe_init_routing_quant_v2_pkg_inner_common.h | | 24 | moe/moe_init_routing_quant_v2/op_kernel/arch35/moe_v2_common.h | moe/moe_init_routing_quant_v2/op_kernel/arch35/moe_init_routing_quant_v2_arch35_common.h | | 25 | moe/moe_init_routing_quant_v2/op_kernel/inner/moe_v2_common.h | moe/moe_init_routing_quant_v2/op_kernel/inner/moe_init_routing_quant_v2_inner_common.h | | 26 | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_base.h | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_base.h | ### 命名约定 | 场景 | 约定 | |------|------| | all_gather_matmul_v2 | agm_v2_* | | matmul_all_reduce | matmul_all_reduce_common.h | | mega_moe 内 moe_v2 | mega_moe_* / mega_moe_quant_* | | moe 算子 | moe_<op>_common.h、arch35 加 _arch35_common | | dispatch setup | setup_base.h / setup_common.h(与 mc2/common 权威 base 区分) | | include guard | 与文件名一致(如 MOE_DISTRIBUTE_DISPATCH_V2_COMMON_H) | ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4356 ## 测试 ## 文档更新 ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [x] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10537 | 2 months ago | |
安装重名头文件整改:common.h Co-authored-by: chenyifan<chenyifan66@h-partners.com> # message auto-generated for no-merge-commit merge: !10537 merge clear_repeat_h_comm into master 安装重名头文件整改:common.h Created-by: mutex_lock Commit-by: chenyifan Merged-by: cann-robot Description: ## 描述 按**安装重名**头文件规范,完成 ** common.h / moe_common.h / moe_v2_common.h 跨算子同名整改**,消除安装包内跨算子同名头冲突。 **原因**: **common.h(13 份)**、**moe_common.h(5 份)**、**moe_v2_common.h(7 份)** 属安装规则 **3.9**(仓内非对外、不同算子/组件目录拷贝同名头),且口径为「同名不同实现亦改名」。权威 common/include/op_kernel/common.h 与 mc2/common/op_kernel/moe_distribute_base.h **保留原名**。 **方法**:低风险只改次要副本——各算子/组件侧 common.h / moe_common.h / moe_v2_common.h / setup 侧 moe_distribute_base.h 加算子或组件前缀(如 agm_v2_、matmul_all_reduce_、moe_init_routing_、mega_moe_、setup_ 等);同步更新 #include、\file、include guard。无业务逻辑变更。禁止软链接。 ### 安装分类规则 | # | 场景 | 是否整改 | 处置 | |---|------|---------|------| | 3.9 | 仓内非对外同名拷贝 | 是 | 次要侧改名 | | 权威 common | common/include/op_kernel/common.h | 否 | 保留原名 | | 权威 mc2 base | mc2/common/op_kernel/moe_distribute_base.h | 否 | 保留原名(v2/combine 等仍 #include 此路径) | ### 范围与结果 | 项 | 说明 | |----|------| | 重命名组数 | **26** | | 涉及域 | attention(scatter_pa / sfag / sfmg)、examples fkle、mc2(agm_v2 / matmul_all_reduce / mega_moe / dispatch setup/teardown/v2、mhc tiling)、moe(gating / init_routing / token_permute 系列) | | 附带修复 | dispatch setup base/common 映射;teardown 跨组件引用;v2 arch35 a5_ccu 裸 include;v3/combine_v3 apt 依赖 v2/v2_combine 的 __has_include 多布局路径 | | 变更文件数 | **166** | | 行变更 | +253 / -238 | ### 26 组重命名(旧 → 新) | # | 旧路径 | 新路径 | |--:|--------|--------| | 1 | attention/scatter_pa_kv_cache/op_kernel/arch35/common.h | attention/scatter_pa_kv_cache/op_kernel/arch35/scatter_pa_kv_cache_arch35_common.h | | 2 | attention/sparse_flash_attention_grad/op_kernel/arch35/common.h | attention/sparse_flash_attention_grad/op_kernel/arch35/sparse_flash_attention_grad_arch35_common.h | | 3 | attention/sparse_flash_mla_grad/op_kernel/arch35/common.h | attention/sparse_flash_mla_grad/op_kernel/arch35/sparse_flash_mla_grad_arch35_common.h | | 4 | examples/fast_kernel_launch_example/csrc/moe_distribute_dispatch_v2/ascend910_93/op_kernel/common.h | examples/fast_kernel_launch_example/csrc/moe_distribute_dispatch_v2/ascend910_93/op_kernel/fkle_mdd_v2_common.h | | 5 | mc2/all_gather_matmul_v2/op_kernel/common.h | mc2/all_gather_matmul_v2/op_kernel/agm_v2_common.h | | 6 | mc2/matmul_all_reduce/op_kernel/common.h | mc2/matmul_all_reduce/op_kernel/matmul_all_reduce_common.h | | 7 | mc2/moe_distribute_dispatch_setup/op_kernel/common.h | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_common.h | | 8 | mc2/moe_distribute_dispatch_v2/op_kernel/common.h | mc2/moe_distribute_dispatch_v2/op_kernel/moe_distribute_dispatch_v2_common.h | | 9 | mhc/mhc_pre_sinkhorn_backward/op_host/op_tiling/arch22/common.h | mhc/mhc_pre_sinkhorn_backward/op_host/op_tiling/arch22/mhc_pre_sinkhorn_backward_common.h | | 10 | moe/moe_gating_top_k/op_kernel/arch35/common.h | moe/moe_gating_top_k/op_kernel/arch35/moe_gating_top_k_arch35_common.h | | 11 | moe/moe_gating_top_k/op_kernel/common.h | moe/moe_gating_top_k/op_kernel/moe_gating_top_k_common.h | | 12 | moe/moe_gating_top_k_backward/op_kernel/arch35/common.h | moe/moe_gating_top_k_backward/op_kernel/arch35/moe_gating_top_k_backward_arch35_common.h | | 13 | moe/moe_gating_top_k_backward/op_kernel/common.h | moe/moe_gating_top_k_backward/op_kernel/moe_gating_top_k_backward_common.h | | 14 | moe/moe_init_routing/op_kernel/moe_common.h | moe/moe_init_routing/op_kernel/moe_init_routing_common.h | | 15 | moe/moe_init_routing/op_kernel/arch35/moe_common.h | moe/moe_init_routing/op_kernel/arch35/moe_init_routing_arch35_common.h | | 16 | moe/moe_token_permute/op_kernel/moe_common.h | moe/moe_token_permute/op_kernel/moe_token_permute_common.h | | 17 | moe/moe_token_permute_with_ep/op_kernel/moe_common.h | moe/moe_token_permute_with_ep/op_kernel/moe_token_permute_with_ep_common.h | | 18 | moe/moe_token_permute_with_routing_map/op_kernel/moe_common.h | moe/moe_token_permute_with_routing_map/op_kernel/moe_token_permute_with_routing_map_common.h | | 19 | mc2/mega_moe/op_kernel/arch22/moe_init_routing_v2/moe_v2_common.h | mc2/mega_moe/op_kernel/arch22/moe_init_routing_v2/mega_moe_moe_v2_common.h | | 20 | mc2/mega_moe/op_kernel/arch22/moe_init_routing_quant_v2/moe_v2_common.h | mc2/mega_moe/op_kernel/arch22/moe_init_routing_quant_v2/mega_moe_quant_moe_v2_common.h | | 21 | moe/moe_init_routing_v2/op_kernel/moe_v2_common.h | moe/moe_init_routing_v2/op_kernel/moe_init_routing_v2_common.h | | 22 | moe/moe_init_routing_v2/op_kernel/arch35/moe_v2_common.h | moe/moe_init_routing_v2/op_kernel/arch35/moe_init_routing_v2_arch35_common.h | | 23 | moe/moe_init_routing_quant_v2/inner/moe_v2_common.h | moe/moe_init_routing_quant_v2/inner/moe_init_routing_quant_v2_pkg_inner_common.h | | 24 | moe/moe_init_routing_quant_v2/op_kernel/arch35/moe_v2_common.h | moe/moe_init_routing_quant_v2/op_kernel/arch35/moe_init_routing_quant_v2_arch35_common.h | | 25 | moe/moe_init_routing_quant_v2/op_kernel/inner/moe_v2_common.h | moe/moe_init_routing_quant_v2/op_kernel/inner/moe_init_routing_quant_v2_inner_common.h | | 26 | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_base.h | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_base.h | ### 命名约定 | 场景 | 约定 | |------|------| | all_gather_matmul_v2 | agm_v2_* | | matmul_all_reduce | matmul_all_reduce_common.h | | mega_moe 内 moe_v2 | mega_moe_* / mega_moe_quant_* | | moe 算子 | moe_<op>_common.h、arch35 加 _arch35_common | | dispatch setup | setup_base.h / setup_common.h(与 mc2/common 权威 base 区分) | | include guard | 与文件名一致(如 MOE_DISTRIBUTE_DISPATCH_V2_COMMON_H) | ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4356 ## 测试 ## 文档更新 ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [x] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10537 | 2 months ago | |
moe_gating_top_k_softmax_st Co-authored-by: gcw_Lf43RugY<mingrui5@huawei.com> # message auto-generated for no-merge-commit merge: !10494 merge master into master moe_gating_top_k_softmax_st Created-by: gcw_Lf43RugY Commit-by: gcw_Lf43RugY Merged-by: cann-robot Description: ## 描述 ## MoE算子:moe_gating_top_k_softmax ST用例选取文档 ### 1.选取原则 - ACLNN 模式:每个算子选取约 20 个用例(ACLNN 调用开销较大,约 2s/用例) - Kernel 模式:每个算子选取约 30 个用例(Kernel 编译+执行较快,约 0.4s/用例) - 覆盖所有 dtype(float32/float16/bfloat16/int8/int32/int64) - 覆盖所有关键属性组合(dropPadMode/quantMode/normType/renorm 等) - 覆盖可选输入的组合(有/无 optional inputs) - 优先选取小 shape 用例(减少运行时间) - 排除已知精度边界用例(量化精度差异、bf16 reduction 差异等) - 排除依赖算子代码修改的用例(如 moe_token_unpermute_grad active 场景) - 总运行时间控制在 10 分钟以内 ### 2.选取结果 ### ACLNN模式 (1个csv,20个用例) | csv文件 | 算子目录 | 选取/全量 | 运行时间 | PASS | |--|--|--| --| --| | test_aclnn_moe_gating_top_k_softmax_arch35.csv | moe_gating_top_k_softmax | 20/20 | 110s | 20/20 | ### Kernel 模式 (1个csv,30个用例) | csv文件 | 算子目录 | 选取/全量 | 运行时间 | PASS | |--|--|--| --| --| | test_moe_gating_top_k_softmax_arch35.csv | moe_gating_top_k_softmax | 30/350 | 216s | 30/30 | ### 3 验证汇总 | 指标 | 值 | |--|--| | 选取用例总数 | 50 | | 全量用例总数 | 370 | | PASS | 50 | | FAIL | 0 | | ACLNN运行时间 | 110s(1min50s) | | Kernel运行时间 | 216s(3min36s) | | 总运行时间 | 317s(5min26s) | | 平均每用例 | 4s | ## 4. 功能覆盖说明 moe_gating_top_k_softmax - dtype:float32 / float16 / bfloat16 - finishedOptional: 覆盖 有/无 ## 关联的Issue [# MoE 算子 ST 用例选取文档](https://gitcode.com/cann/ops-transformer/issues/4546) ## 测试 <!--描述进行了哪些测试来验证你的改动。包括但不限于二级冒烟、算子泛化等。--> ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [x] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [x] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10494 | 2 months ago | |
增加moe蓝区ttk用例适配 Co-authored-by: cpy_123456<chenpengyi@huawei.com> # message auto-generated for no-merge-commit merge: !10344 merge ut into master 增加moe蓝区ttk用例适配 Created-by: cpy_123456 Commit-by: cpy_123456 Merged-by: cann-robot Description: ## 描述 # MoE 算子 ST 用例选取文档 ## 1. 选取原则 - **ACLNN 模式**:每个算子选取约 20 个用例(ACLNN 调用开销较大,约 2s/用例) - **Kernel 模式**:每个算子选取约 30 个用例(Kernel 编译+执行较快,约 0.4s/用例) - 覆盖所有 dtype(float32/float16/bfloat16/int8/int32/int64) - 覆盖所有关键属性组合(dropPadMode/quantMode/normType/renorm 等) - 覆盖可选输入的组合(有/无 optional inputs) - 优先选取小 shape 用例(减少运行时间) - 排除已知精度边界用例(量化精度差异、bf16 reduction 差异等) - 排除依赖算子代码修改的用例(如 moe_token_unpermute_grad active 场景) - 总运行时间控制在 10 分钟以内 ## 2. 选取结果 ### ACLNN 模式(12 个 CSV,150 个用例) | CSV 文件 | 算子目录 | 选取/全量 | 运行时间 | PASS | |---------|---------|----------|---------|------| | test_aclnn_moe_gating_top_k_arch35.csv | moe_gating_top_k | 20/38 | 42s | 20/20 | | test_aclnn_moe_gating_top_k_v2_arch35.csv | moe_gating_top_k | 20/32 | 46s | 20/20 | | test_aclnn_moe_gating_top_k_softmax_v2_arch35.csv | moe_gating_top_k_softmax_v2 | 20/232 | 47s | 20/20 | | test_aclnn_moe_init_routing_arch35.csv | moe_init_routing | 10/10 | 36s | 10/10 | | test_aclnn_moe_init_routing_v2_arch35.csv | moe_init_routing_v2 | 20/27 | 57s | 20/20 | | test_aclnn_moe_init_routing_v2_grad_arch35.csv | moe_init_routing_v2_grad | 20/116 | 51s | 20/20 | | test_aclnn_moe_init_routing_quant_v2_arch35.csv | moe_init_routing_quant_v2 | 15/34 | 48s | 15/15 | | test_aclnn_moe_finalize_routing_v2_grad_arch35.csv | moe_finalize_routing_v2_grad | 11/94 | 30s | 11/11 | | test_aclnn_moe_token_permute_arch35.csv | moe_token_permute | 4/4 | 19s | 4/4 | | test_aclnn_moe_token_permute_grad_arch35.csv | moe_token_permute_grad | 4/4 | 18s | 4/4 | | test_aclnn_moe_token_unpermute_arch35.csv | moe_token_unpermute | 2/2 | 15s | 2/2 | | test_aclnn_moe_token_unpermute_grad_arch35.csv | moe_token_unpermute_grad | 4/4 | 19s | 4/4 | ### Kernel 模式(10 个 CSV,300 个用例) | CSV 文件 | 算子目录 | 选取/全量 | 运行时间 | PASS | |---------|---------|----------|---------|------| | test_moe_gating_top_k_arch35.csv | moe_gating_top_k | 30/829 | 11s | 30/30 | | test_moe_gating_top_k_softmax_v2_arch35.csv | moe_gating_top_k_softmax_v2 | 30/832 | 13s | 30/30 | | test_moe_init_routing_arch35.csv | moe_init_routing | 30/203 | 13s | 30/30 | | test_moe_init_routing_v2_arch35.csv | moe_init_routing_v2 | 30/200 | 13s | 30/30 | | test_moe_init_routing_v2_grad_arch35.csv | moe_init_routing_v2_grad | 30/144 | 11s | 30/30 | | test_moe_init_routing_quant_v2_arch35.csv | moe_init_routing_quant_v2 | 30/190 | 13s | 30/30 | | test_moe_finalize_routing_v2_arch35.csv | moe_finalize_routing_v2 | 30/112 | 13s | 30/30 | | test_moe_finalize_routing_v2_grad_arch35.csv | moe_finalize_routing_v2_grad | 30/198 | 13s | 30/30 | | test_moe_compute_expert_tokens_arch35.csv | moe_compute_expert_tokens | 30/159 | 13s | 30/30 | | test_moe_re_routing_arch35.csv | moe_re_routing | 30/578 | 11s | 30/30 | ## 3. 验证汇总 | 指标 | 值 | |------|------| | 选取用例总数 | 450 | | 全量用例总数 | 3726 | | PASS | 450 (100%) | | FAIL | 0 | | ACLNN 运行时间 | 428s (7m08s) | | Kernel 运行时间 | 124s (2m04s) | | 总运行时间 | 552s (9m12s) | | 平均每用例 | 1.23s | ## 4. 排除用例说明 ## 5. 功能覆盖说明 ### moe_gating_top_k - **norm_type**:覆盖 0(softmax)/1(sigmoid)/2(sqrtsoftplus) - **group_count**:覆盖 1(无分组)/8(标准分组)/大值(多分组) - **group_select_mode**:覆盖 0(amax)/1(top2sum) - **out_flag**:覆盖 True/False - **renorm**:覆盖 0/1 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_gating_top_k_v2 - **normType**:覆盖 1(sigmoid) - **inputIdsOptional/tid2eidOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_gating_top_k_softmax_v2 - **renorm**:覆盖 0(softmax后取topk)/1(topk后renorm) - **finishedOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing - **active_num**:覆盖小值(1~5)/大值(100+) - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_v2 - **drop_pad_mode**:覆盖 0(dropless)/1(drop/pad) - **expert_tokens_count_or_cumsum_flag**:覆盖 0/1(cumsum)/2(count) - **expert_tokens_before_capacity_flag**:覆盖 True/False - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_v2_grad - **drop_pad_mode**:覆盖 0/1 - **active_num**:覆盖 0(full)/1(active)/大值 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_quant_v2 - **drop_pad_mode**:覆盖 0/1 - **quant_mode**:覆盖 0(静态量化)/1(动态量化) - **expert_tokens_count_or_cumsum_flag**:覆盖 0/1/2 - **expert_tokens_before_capacity_flag**:覆盖 True/False - **dtype**:覆盖 fp32/fp16/bf16 ### moe_finalize_routing_v2 - **drop_pad_mode**:覆盖 0/2/3 - **zero/copy/constant_expert_range**:覆盖各种 expert range 配置 - **optional inputs(x/a1/a2/v)**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_finalize_routing_v2_grad - **drop_pad_mode**:覆盖 0/1 - **active_num**:覆盖 0/1/大值 - **scales**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_compute_expert_tokens - **num_experts**:覆盖小值(8)/大值(800+) - **dtype**:int32 ### moe_re_routing - **idx_type**:覆盖 0/1 - **expert_token_num_type**:覆盖 0/1 - **per_token_scales**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16/int8/int32/int64 ### moe_token_permute - **numOutTokens**:覆盖 0(full)/10(active)/20(active) - **paddedMode**:False - **dtype**:覆盖 fp32/fp16 ### moe_token_permute_grad - **numTopk**:覆盖 8 - **paddedMode**:False - **dtype**:覆盖 fp32/fp16 ### moe_token_unpermute - **paddedMode**:False - **probsOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16 ### moe_token_unpermute_grad - **paddedMode**:False - **probsOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16 ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4472 ## 测试 <!--描述进行了哪些测试来验证你的改动。包括但不限于二级冒烟、算子泛化等。--> ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [x] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [x] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10344 | 2 months ago | |
安装重名头文件整改:common.h Co-authored-by: chenyifan<chenyifan66@h-partners.com> # message auto-generated for no-merge-commit merge: !10537 merge clear_repeat_h_comm into master 安装重名头文件整改:common.h Created-by: mutex_lock Commit-by: chenyifan Merged-by: cann-robot Description: ## 描述 按**安装重名**头文件规范,完成 ** common.h / moe_common.h / moe_v2_common.h 跨算子同名整改**,消除安装包内跨算子同名头冲突。 **原因**: **common.h(13 份)**、**moe_common.h(5 份)**、**moe_v2_common.h(7 份)** 属安装规则 **3.9**(仓内非对外、不同算子/组件目录拷贝同名头),且口径为「同名不同实现亦改名」。权威 common/include/op_kernel/common.h 与 mc2/common/op_kernel/moe_distribute_base.h **保留原名**。 **方法**:低风险只改次要副本——各算子/组件侧 common.h / moe_common.h / moe_v2_common.h / setup 侧 moe_distribute_base.h 加算子或组件前缀(如 agm_v2_、matmul_all_reduce_、moe_init_routing_、mega_moe_、setup_ 等);同步更新 #include、\file、include guard。无业务逻辑变更。禁止软链接。 ### 安装分类规则 | # | 场景 | 是否整改 | 处置 | |---|------|---------|------| | 3.9 | 仓内非对外同名拷贝 | 是 | 次要侧改名 | | 权威 common | common/include/op_kernel/common.h | 否 | 保留原名 | | 权威 mc2 base | mc2/common/op_kernel/moe_distribute_base.h | 否 | 保留原名(v2/combine 等仍 #include 此路径) | ### 范围与结果 | 项 | 说明 | |----|------| | 重命名组数 | **26** | | 涉及域 | attention(scatter_pa / sfag / sfmg)、examples fkle、mc2(agm_v2 / matmul_all_reduce / mega_moe / dispatch setup/teardown/v2、mhc tiling)、moe(gating / init_routing / token_permute 系列) | | 附带修复 | dispatch setup base/common 映射;teardown 跨组件引用;v2 arch35 a5_ccu 裸 include;v3/combine_v3 apt 依赖 v2/v2_combine 的 __has_include 多布局路径 | | 变更文件数 | **166** | | 行变更 | +253 / -238 | ### 26 组重命名(旧 → 新) | # | 旧路径 | 新路径 | |--:|--------|--------| | 1 | attention/scatter_pa_kv_cache/op_kernel/arch35/common.h | attention/scatter_pa_kv_cache/op_kernel/arch35/scatter_pa_kv_cache_arch35_common.h | | 2 | attention/sparse_flash_attention_grad/op_kernel/arch35/common.h | attention/sparse_flash_attention_grad/op_kernel/arch35/sparse_flash_attention_grad_arch35_common.h | | 3 | attention/sparse_flash_mla_grad/op_kernel/arch35/common.h | attention/sparse_flash_mla_grad/op_kernel/arch35/sparse_flash_mla_grad_arch35_common.h | | 4 | examples/fast_kernel_launch_example/csrc/moe_distribute_dispatch_v2/ascend910_93/op_kernel/common.h | examples/fast_kernel_launch_example/csrc/moe_distribute_dispatch_v2/ascend910_93/op_kernel/fkle_mdd_v2_common.h | | 5 | mc2/all_gather_matmul_v2/op_kernel/common.h | mc2/all_gather_matmul_v2/op_kernel/agm_v2_common.h | | 6 | mc2/matmul_all_reduce/op_kernel/common.h | mc2/matmul_all_reduce/op_kernel/matmul_all_reduce_common.h | | 7 | mc2/moe_distribute_dispatch_setup/op_kernel/common.h | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_common.h | | 8 | mc2/moe_distribute_dispatch_v2/op_kernel/common.h | mc2/moe_distribute_dispatch_v2/op_kernel/moe_distribute_dispatch_v2_common.h | | 9 | mhc/mhc_pre_sinkhorn_backward/op_host/op_tiling/arch22/common.h | mhc/mhc_pre_sinkhorn_backward/op_host/op_tiling/arch22/mhc_pre_sinkhorn_backward_common.h | | 10 | moe/moe_gating_top_k/op_kernel/arch35/common.h | moe/moe_gating_top_k/op_kernel/arch35/moe_gating_top_k_arch35_common.h | | 11 | moe/moe_gating_top_k/op_kernel/common.h | moe/moe_gating_top_k/op_kernel/moe_gating_top_k_common.h | | 12 | moe/moe_gating_top_k_backward/op_kernel/arch35/common.h | moe/moe_gating_top_k_backward/op_kernel/arch35/moe_gating_top_k_backward_arch35_common.h | | 13 | moe/moe_gating_top_k_backward/op_kernel/common.h | moe/moe_gating_top_k_backward/op_kernel/moe_gating_top_k_backward_common.h | | 14 | moe/moe_init_routing/op_kernel/moe_common.h | moe/moe_init_routing/op_kernel/moe_init_routing_common.h | | 15 | moe/moe_init_routing/op_kernel/arch35/moe_common.h | moe/moe_init_routing/op_kernel/arch35/moe_init_routing_arch35_common.h | | 16 | moe/moe_token_permute/op_kernel/moe_common.h | moe/moe_token_permute/op_kernel/moe_token_permute_common.h | | 17 | moe/moe_token_permute_with_ep/op_kernel/moe_common.h | moe/moe_token_permute_with_ep/op_kernel/moe_token_permute_with_ep_common.h | | 18 | moe/moe_token_permute_with_routing_map/op_kernel/moe_common.h | moe/moe_token_permute_with_routing_map/op_kernel/moe_token_permute_with_routing_map_common.h | | 19 | mc2/mega_moe/op_kernel/arch22/moe_init_routing_v2/moe_v2_common.h | mc2/mega_moe/op_kernel/arch22/moe_init_routing_v2/mega_moe_moe_v2_common.h | | 20 | mc2/mega_moe/op_kernel/arch22/moe_init_routing_quant_v2/moe_v2_common.h | mc2/mega_moe/op_kernel/arch22/moe_init_routing_quant_v2/mega_moe_quant_moe_v2_common.h | | 21 | moe/moe_init_routing_v2/op_kernel/moe_v2_common.h | moe/moe_init_routing_v2/op_kernel/moe_init_routing_v2_common.h | | 22 | moe/moe_init_routing_v2/op_kernel/arch35/moe_v2_common.h | moe/moe_init_routing_v2/op_kernel/arch35/moe_init_routing_v2_arch35_common.h | | 23 | moe/moe_init_routing_quant_v2/inner/moe_v2_common.h | moe/moe_init_routing_quant_v2/inner/moe_init_routing_quant_v2_pkg_inner_common.h | | 24 | moe/moe_init_routing_quant_v2/op_kernel/arch35/moe_v2_common.h | moe/moe_init_routing_quant_v2/op_kernel/arch35/moe_init_routing_quant_v2_arch35_common.h | | 25 | moe/moe_init_routing_quant_v2/op_kernel/inner/moe_v2_common.h | moe/moe_init_routing_quant_v2/op_kernel/inner/moe_init_routing_quant_v2_inner_common.h | | 26 | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_base.h | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_base.h | ### 命名约定 | 场景 | 约定 | |------|------| | all_gather_matmul_v2 | agm_v2_* | | matmul_all_reduce | matmul_all_reduce_common.h | | mega_moe 内 moe_v2 | mega_moe_* / mega_moe_quant_* | | moe 算子 | moe_<op>_common.h、arch35 加 _arch35_common | | dispatch setup | setup_base.h / setup_common.h(与 mc2/common 权威 base 区分) | | include guard | 与文件名一致(如 MOE_DISTRIBUTE_DISPATCH_V2_COMMON_H) | ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4356 ## 测试 ## 文档更新 ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [x] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10537 | 2 months ago | |
fix moe and posembling Co-authored-by: caiwenwen<caiwenwen6@h-partners.com> # message auto-generated for no-merge-commit merge: !9719 merge master into master fix moe and posembling Created-by: caiwenwen Commit-by: caiwenwen Merged-by: cann-robot Description: ## 描述 修改moe and posembling两个文件夹标签 ## 关联的Issue 关联Issue [#4040](https://gitcode.com/cann/ops-transformer/issues/4040) <!-- 如果这个PR是为了解决特定的问题单,请在这里描述问题单单号。--> ## 测试 <!--描述进行了哪些测试来验证你的改动。包括但不限于二级冒烟、算子泛化等。--> ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [x] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!9719 | 3 months ago | |
添加moe_init_routing_v4和moe_re_routing_v2算子的st&标杆 Co-authored-by: cpy_123456<chenpengyi@huawei.com> # message auto-generated for no-merge-commit merge: !10530 merge ut into master 添加moe_init_routing_v4和moe_re_routing_v2算子的st&标杆 Created-by: cpy_123456 Commit-by: cpy_123456 Merged-by: cann-robot Description: ## 描述 添加moe_init_routing_v4和moe_re_routing_v2算子的st&标杆 ## 一、moe_init_routing_v4(25 个用例,全部 PASS) ### 1.1 场景覆盖矩阵 | 维度 | 覆盖值 | 用例数 | |---|---|---| | **输入数据类型 (x)** | int8 | 2 | | | float16 | 7 | | | bfloat16 | 8 | | | float32 | 8 | | **quant_mode** | -1(非量化,DropLess) | 13 | | | 1(动态量化) | 12 | | **expert_tokens_num_type (etnt)** | 1(直方图统计) | 13 | | | 2(key-value 对) | 12 | | **drop_pad_mode** | 0(DropLess 模式) | 25 | | **expert_tokens_num_flag** | True | 25 | | **row_idx_type** | 0(gather) | 7 | | | 1(scatter) | 18 | | **scale 输入** | 有 | 10 | | | 无 | 15 | | **topk_weight 输入** | 有 | 25 | | **expert_num 范围** | 256 ~ 2930 | — | | **active_expert_range 跨度** | 6 ~ 580 | — | ## 二、moe_re_routing_v2(25 个用例,全部 PASS) ### 2.1 场景覆盖矩阵 | 维度 | 覆盖值 | 用例数 | |---|---|---| | **tokens 数据类型** | int8 | 8 | | | float16 | 8 | | | bfloat16 | 9 | | **idx_type** | 0(正向索引,permute_token_idx=src_offset) | 12 | | | 1(反向索引,permute_token_idx=dst_offset) | 13 | | **per_token_scales 输入** | 有 | 12 | | | 无 | 13 | | **expert_topk_weight 输入** | 有 | 25 | | **rank_num(专家并行数)** | 1 | 17 | | | 2 | 3 | | | 3 | 1 | | | 4~5 | 2 | | | 7~52 | 2 | | **expert_num(每 rank 专家数)** | 1 | 17 | | | 2 | 5 | | | 3 | 1 | | | 5~6 | 2 | | **tokens 第二维 (hidden_size)** | 9 ~ 15267 | — | ## 关联的Issue <!-- 如果这个PR是为了解决特定的Issue,请在这里提供Issue链接。例如:关联Issue #000--> <!-- 如果这个PR是为了解决特定的问题单,请在这里描述问题单单号。--> ## 测试 st&标杆通过 ## 文档更新 https://gitcode.com/cann/ops-transformer/issues/4543 ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [x] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10530 | 2 months ago | |
安装重名头文件整改:common.h Co-authored-by: chenyifan<chenyifan66@h-partners.com> # message auto-generated for no-merge-commit merge: !10537 merge clear_repeat_h_comm into master 安装重名头文件整改:common.h Created-by: mutex_lock Commit-by: chenyifan Merged-by: cann-robot Description: ## 描述 按**安装重名**头文件规范,完成 ** common.h / moe_common.h / moe_v2_common.h 跨算子同名整改**,消除安装包内跨算子同名头冲突。 **原因**: **common.h(13 份)**、**moe_common.h(5 份)**、**moe_v2_common.h(7 份)** 属安装规则 **3.9**(仓内非对外、不同算子/组件目录拷贝同名头),且口径为「同名不同实现亦改名」。权威 common/include/op_kernel/common.h 与 mc2/common/op_kernel/moe_distribute_base.h **保留原名**。 **方法**:低风险只改次要副本——各算子/组件侧 common.h / moe_common.h / moe_v2_common.h / setup 侧 moe_distribute_base.h 加算子或组件前缀(如 agm_v2_、matmul_all_reduce_、moe_init_routing_、mega_moe_、setup_ 等);同步更新 #include、\file、include guard。无业务逻辑变更。禁止软链接。 ### 安装分类规则 | # | 场景 | 是否整改 | 处置 | |---|------|---------|------| | 3.9 | 仓内非对外同名拷贝 | 是 | 次要侧改名 | | 权威 common | common/include/op_kernel/common.h | 否 | 保留原名 | | 权威 mc2 base | mc2/common/op_kernel/moe_distribute_base.h | 否 | 保留原名(v2/combine 等仍 #include 此路径) | ### 范围与结果 | 项 | 说明 | |----|------| | 重命名组数 | **26** | | 涉及域 | attention(scatter_pa / sfag / sfmg)、examples fkle、mc2(agm_v2 / matmul_all_reduce / mega_moe / dispatch setup/teardown/v2、mhc tiling)、moe(gating / init_routing / token_permute 系列) | | 附带修复 | dispatch setup base/common 映射;teardown 跨组件引用;v2 arch35 a5_ccu 裸 include;v3/combine_v3 apt 依赖 v2/v2_combine 的 __has_include 多布局路径 | | 变更文件数 | **166** | | 行变更 | +253 / -238 | ### 26 组重命名(旧 → 新) | # | 旧路径 | 新路径 | |--:|--------|--------| | 1 | attention/scatter_pa_kv_cache/op_kernel/arch35/common.h | attention/scatter_pa_kv_cache/op_kernel/arch35/scatter_pa_kv_cache_arch35_common.h | | 2 | attention/sparse_flash_attention_grad/op_kernel/arch35/common.h | attention/sparse_flash_attention_grad/op_kernel/arch35/sparse_flash_attention_grad_arch35_common.h | | 3 | attention/sparse_flash_mla_grad/op_kernel/arch35/common.h | attention/sparse_flash_mla_grad/op_kernel/arch35/sparse_flash_mla_grad_arch35_common.h | | 4 | examples/fast_kernel_launch_example/csrc/moe_distribute_dispatch_v2/ascend910_93/op_kernel/common.h | examples/fast_kernel_launch_example/csrc/moe_distribute_dispatch_v2/ascend910_93/op_kernel/fkle_mdd_v2_common.h | | 5 | mc2/all_gather_matmul_v2/op_kernel/common.h | mc2/all_gather_matmul_v2/op_kernel/agm_v2_common.h | | 6 | mc2/matmul_all_reduce/op_kernel/common.h | mc2/matmul_all_reduce/op_kernel/matmul_all_reduce_common.h | | 7 | mc2/moe_distribute_dispatch_setup/op_kernel/common.h | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_common.h | | 8 | mc2/moe_distribute_dispatch_v2/op_kernel/common.h | mc2/moe_distribute_dispatch_v2/op_kernel/moe_distribute_dispatch_v2_common.h | | 9 | mhc/mhc_pre_sinkhorn_backward/op_host/op_tiling/arch22/common.h | mhc/mhc_pre_sinkhorn_backward/op_host/op_tiling/arch22/mhc_pre_sinkhorn_backward_common.h | | 10 | moe/moe_gating_top_k/op_kernel/arch35/common.h | moe/moe_gating_top_k/op_kernel/arch35/moe_gating_top_k_arch35_common.h | | 11 | moe/moe_gating_top_k/op_kernel/common.h | moe/moe_gating_top_k/op_kernel/moe_gating_top_k_common.h | | 12 | moe/moe_gating_top_k_backward/op_kernel/arch35/common.h | moe/moe_gating_top_k_backward/op_kernel/arch35/moe_gating_top_k_backward_arch35_common.h | | 13 | moe/moe_gating_top_k_backward/op_kernel/common.h | moe/moe_gating_top_k_backward/op_kernel/moe_gating_top_k_backward_common.h | | 14 | moe/moe_init_routing/op_kernel/moe_common.h | moe/moe_init_routing/op_kernel/moe_init_routing_common.h | | 15 | moe/moe_init_routing/op_kernel/arch35/moe_common.h | moe/moe_init_routing/op_kernel/arch35/moe_init_routing_arch35_common.h | | 16 | moe/moe_token_permute/op_kernel/moe_common.h | moe/moe_token_permute/op_kernel/moe_token_permute_common.h | | 17 | moe/moe_token_permute_with_ep/op_kernel/moe_common.h | moe/moe_token_permute_with_ep/op_kernel/moe_token_permute_with_ep_common.h | | 18 | moe/moe_token_permute_with_routing_map/op_kernel/moe_common.h | moe/moe_token_permute_with_routing_map/op_kernel/moe_token_permute_with_routing_map_common.h | | 19 | mc2/mega_moe/op_kernel/arch22/moe_init_routing_v2/moe_v2_common.h | mc2/mega_moe/op_kernel/arch22/moe_init_routing_v2/mega_moe_moe_v2_common.h | | 20 | mc2/mega_moe/op_kernel/arch22/moe_init_routing_quant_v2/moe_v2_common.h | mc2/mega_moe/op_kernel/arch22/moe_init_routing_quant_v2/mega_moe_quant_moe_v2_common.h | | 21 | moe/moe_init_routing_v2/op_kernel/moe_v2_common.h | moe/moe_init_routing_v2/op_kernel/moe_init_routing_v2_common.h | | 22 | moe/moe_init_routing_v2/op_kernel/arch35/moe_v2_common.h | moe/moe_init_routing_v2/op_kernel/arch35/moe_init_routing_v2_arch35_common.h | | 23 | moe/moe_init_routing_quant_v2/inner/moe_v2_common.h | moe/moe_init_routing_quant_v2/inner/moe_init_routing_quant_v2_pkg_inner_common.h | | 24 | moe/moe_init_routing_quant_v2/op_kernel/arch35/moe_v2_common.h | moe/moe_init_routing_quant_v2/op_kernel/arch35/moe_init_routing_quant_v2_arch35_common.h | | 25 | moe/moe_init_routing_quant_v2/op_kernel/inner/moe_v2_common.h | moe/moe_init_routing_quant_v2/op_kernel/inner/moe_init_routing_quant_v2_inner_common.h | | 26 | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_base.h | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_base.h | ### 命名约定 | 场景 | 约定 | |------|------| | all_gather_matmul_v2 | agm_v2_* | | matmul_all_reduce | matmul_all_reduce_common.h | | mega_moe 内 moe_v2 | mega_moe_* / mega_moe_quant_* | | moe 算子 | moe_<op>_common.h、arch35 加 _arch35_common | | dispatch setup | setup_base.h / setup_common.h(与 mc2/common 权威 base 区分) | | include guard | 与文件名一致(如 MOE_DISTRIBUTE_DISPATCH_V2_COMMON_H) | ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4356 ## 测试 ## 文档更新 ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [x] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10537 | 2 months ago | |
moe_init_routing和moe_finalize_routing支持自动反向 Co-authored-by: JimmyLam2<linjingmin2@huawei.com> # message auto-generated for no-merge-commit merge: !9881 merge master into master moe_init_routing和moe_finalize_routing支持自动反向 Created-by: JimmyLam2 Commit-by: JimmyLam2 Merged-by: cann-robot Description: ## 描述 moe_init_routing和moe_finalize_routing支持自动反向 ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4359 ## 测试 不涉及ST与二级冒烟。采用构造torch用例的方式进行验证。  ## 文档更新 新增 1.torch_extension/cann_ops_transformer/docs/zh/moe_finalize_routing_grad.md 2.torch_extension/cann_ops_transformer/docs/zh/moe_init_routing_grad.md 修改 1.torch_extension/cann_ops_transformer/docs/zh/moe_finalize_routing.md 2.torch_extension/cann_ops_transformer/docs/zh/moe_init_routing.md ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [x] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!9881 | 2 months ago | |
统一moeIintRoutingV3 mxfp8对于边界值的处理方式(quant=2\3\4\5\11\12\13\14\15\16\17) Co-authored-by: cpy_123456<chenpengyi@huawei.com> # message auto-generated for no-merge-commit merge: !10116 merge cpy_1 into master 统一moeIintRoutingV3 mxfp8对于边界值的处理方式(quant=2\3\4\5\11\12\13\14\15\16\17) Created-by: cpy_123456 Commit-by: cpy_123456 Merged-by: cann-robot Description: ## 描述 moeinitroutingv3 mxfp8量化对于nan、INF、-Inf处理逻辑不一致,输出expandedx,expandedscale应统一 ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4358 ## 测试 david冒烟、st通过 ## 文档更新 不涉及 ## 类型标签 <!-- [x] 表示选中 --> - [x] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10116 | 2 months ago | |
添加moe_init_routing_v4和moe_re_routing_v2算子的st&标杆 Co-authored-by: cpy_123456<chenpengyi@huawei.com> # message auto-generated for no-merge-commit merge: !10530 merge ut into master 添加moe_init_routing_v4和moe_re_routing_v2算子的st&标杆 Created-by: cpy_123456 Commit-by: cpy_123456 Merged-by: cann-robot Description: ## 描述 添加moe_init_routing_v4和moe_re_routing_v2算子的st&标杆 ## 一、moe_init_routing_v4(25 个用例,全部 PASS) ### 1.1 场景覆盖矩阵 | 维度 | 覆盖值 | 用例数 | |---|---|---| | **输入数据类型 (x)** | int8 | 2 | | | float16 | 7 | | | bfloat16 | 8 | | | float32 | 8 | | **quant_mode** | -1(非量化,DropLess) | 13 | | | 1(动态量化) | 12 | | **expert_tokens_num_type (etnt)** | 1(直方图统计) | 13 | | | 2(key-value 对) | 12 | | **drop_pad_mode** | 0(DropLess 模式) | 25 | | **expert_tokens_num_flag** | True | 25 | | **row_idx_type** | 0(gather) | 7 | | | 1(scatter) | 18 | | **scale 输入** | 有 | 10 | | | 无 | 15 | | **topk_weight 输入** | 有 | 25 | | **expert_num 范围** | 256 ~ 2930 | — | | **active_expert_range 跨度** | 6 ~ 580 | — | ## 二、moe_re_routing_v2(25 个用例,全部 PASS) ### 2.1 场景覆盖矩阵 | 维度 | 覆盖值 | 用例数 | |---|---|---| | **tokens 数据类型** | int8 | 8 | | | float16 | 8 | | | bfloat16 | 9 | | **idx_type** | 0(正向索引,permute_token_idx=src_offset) | 12 | | | 1(反向索引,permute_token_idx=dst_offset) | 13 | | **per_token_scales 输入** | 有 | 12 | | | 无 | 13 | | **expert_topk_weight 输入** | 有 | 25 | | **rank_num(专家并行数)** | 1 | 17 | | | 2 | 3 | | | 3 | 1 | | | 4~5 | 2 | | | 7~52 | 2 | | **expert_num(每 rank 专家数)** | 1 | 17 | | | 2 | 5 | | | 3 | 1 | | | 5~6 | 2 | | **tokens 第二维 (hidden_size)** | 9 ~ 15267 | — | ## 关联的Issue <!-- 如果这个PR是为了解决特定的Issue,请在这里提供Issue链接。例如:关联Issue #000--> <!-- 如果这个PR是为了解决特定的问题单,请在这里描述问题单单号。--> ## 测试 st&标杆通过 ## 文档更新 https://gitcode.com/cann/ops-transformer/issues/4543 ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [x] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10530 | 2 months ago | |
增加moe蓝区ttk用例适配 Co-authored-by: cpy_123456<chenpengyi@huawei.com> # message auto-generated for no-merge-commit merge: !10344 merge ut into master 增加moe蓝区ttk用例适配 Created-by: cpy_123456 Commit-by: cpy_123456 Merged-by: cann-robot Description: ## 描述 # MoE 算子 ST 用例选取文档 ## 1. 选取原则 - **ACLNN 模式**:每个算子选取约 20 个用例(ACLNN 调用开销较大,约 2s/用例) - **Kernel 模式**:每个算子选取约 30 个用例(Kernel 编译+执行较快,约 0.4s/用例) - 覆盖所有 dtype(float32/float16/bfloat16/int8/int32/int64) - 覆盖所有关键属性组合(dropPadMode/quantMode/normType/renorm 等) - 覆盖可选输入的组合(有/无 optional inputs) - 优先选取小 shape 用例(减少运行时间) - 排除已知精度边界用例(量化精度差异、bf16 reduction 差异等) - 排除依赖算子代码修改的用例(如 moe_token_unpermute_grad active 场景) - 总运行时间控制在 10 分钟以内 ## 2. 选取结果 ### ACLNN 模式(12 个 CSV,150 个用例) | CSV 文件 | 算子目录 | 选取/全量 | 运行时间 | PASS | |---------|---------|----------|---------|------| | test_aclnn_moe_gating_top_k_arch35.csv | moe_gating_top_k | 20/38 | 42s | 20/20 | | test_aclnn_moe_gating_top_k_v2_arch35.csv | moe_gating_top_k | 20/32 | 46s | 20/20 | | test_aclnn_moe_gating_top_k_softmax_v2_arch35.csv | moe_gating_top_k_softmax_v2 | 20/232 | 47s | 20/20 | | test_aclnn_moe_init_routing_arch35.csv | moe_init_routing | 10/10 | 36s | 10/10 | | test_aclnn_moe_init_routing_v2_arch35.csv | moe_init_routing_v2 | 20/27 | 57s | 20/20 | | test_aclnn_moe_init_routing_v2_grad_arch35.csv | moe_init_routing_v2_grad | 20/116 | 51s | 20/20 | | test_aclnn_moe_init_routing_quant_v2_arch35.csv | moe_init_routing_quant_v2 | 15/34 | 48s | 15/15 | | test_aclnn_moe_finalize_routing_v2_grad_arch35.csv | moe_finalize_routing_v2_grad | 11/94 | 30s | 11/11 | | test_aclnn_moe_token_permute_arch35.csv | moe_token_permute | 4/4 | 19s | 4/4 | | test_aclnn_moe_token_permute_grad_arch35.csv | moe_token_permute_grad | 4/4 | 18s | 4/4 | | test_aclnn_moe_token_unpermute_arch35.csv | moe_token_unpermute | 2/2 | 15s | 2/2 | | test_aclnn_moe_token_unpermute_grad_arch35.csv | moe_token_unpermute_grad | 4/4 | 19s | 4/4 | ### Kernel 模式(10 个 CSV,300 个用例) | CSV 文件 | 算子目录 | 选取/全量 | 运行时间 | PASS | |---------|---------|----------|---------|------| | test_moe_gating_top_k_arch35.csv | moe_gating_top_k | 30/829 | 11s | 30/30 | | test_moe_gating_top_k_softmax_v2_arch35.csv | moe_gating_top_k_softmax_v2 | 30/832 | 13s | 30/30 | | test_moe_init_routing_arch35.csv | moe_init_routing | 30/203 | 13s | 30/30 | | test_moe_init_routing_v2_arch35.csv | moe_init_routing_v2 | 30/200 | 13s | 30/30 | | test_moe_init_routing_v2_grad_arch35.csv | moe_init_routing_v2_grad | 30/144 | 11s | 30/30 | | test_moe_init_routing_quant_v2_arch35.csv | moe_init_routing_quant_v2 | 30/190 | 13s | 30/30 | | test_moe_finalize_routing_v2_arch35.csv | moe_finalize_routing_v2 | 30/112 | 13s | 30/30 | | test_moe_finalize_routing_v2_grad_arch35.csv | moe_finalize_routing_v2_grad | 30/198 | 13s | 30/30 | | test_moe_compute_expert_tokens_arch35.csv | moe_compute_expert_tokens | 30/159 | 13s | 30/30 | | test_moe_re_routing_arch35.csv | moe_re_routing | 30/578 | 11s | 30/30 | ## 3. 验证汇总 | 指标 | 值 | |------|------| | 选取用例总数 | 450 | | 全量用例总数 | 3726 | | PASS | 450 (100%) | | FAIL | 0 | | ACLNN 运行时间 | 428s (7m08s) | | Kernel 运行时间 | 124s (2m04s) | | 总运行时间 | 552s (9m12s) | | 平均每用例 | 1.23s | ## 4. 排除用例说明 ## 5. 功能覆盖说明 ### moe_gating_top_k - **norm_type**:覆盖 0(softmax)/1(sigmoid)/2(sqrtsoftplus) - **group_count**:覆盖 1(无分组)/8(标准分组)/大值(多分组) - **group_select_mode**:覆盖 0(amax)/1(top2sum) - **out_flag**:覆盖 True/False - **renorm**:覆盖 0/1 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_gating_top_k_v2 - **normType**:覆盖 1(sigmoid) - **inputIdsOptional/tid2eidOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_gating_top_k_softmax_v2 - **renorm**:覆盖 0(softmax后取topk)/1(topk后renorm) - **finishedOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing - **active_num**:覆盖小值(1~5)/大值(100+) - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_v2 - **drop_pad_mode**:覆盖 0(dropless)/1(drop/pad) - **expert_tokens_count_or_cumsum_flag**:覆盖 0/1(cumsum)/2(count) - **expert_tokens_before_capacity_flag**:覆盖 True/False - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_v2_grad - **drop_pad_mode**:覆盖 0/1 - **active_num**:覆盖 0(full)/1(active)/大值 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_quant_v2 - **drop_pad_mode**:覆盖 0/1 - **quant_mode**:覆盖 0(静态量化)/1(动态量化) - **expert_tokens_count_or_cumsum_flag**:覆盖 0/1/2 - **expert_tokens_before_capacity_flag**:覆盖 True/False - **dtype**:覆盖 fp32/fp16/bf16 ### moe_finalize_routing_v2 - **drop_pad_mode**:覆盖 0/2/3 - **zero/copy/constant_expert_range**:覆盖各种 expert range 配置 - **optional inputs(x/a1/a2/v)**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_finalize_routing_v2_grad - **drop_pad_mode**:覆盖 0/1 - **active_num**:覆盖 0/1/大值 - **scales**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_compute_expert_tokens - **num_experts**:覆盖小值(8)/大值(800+) - **dtype**:int32 ### moe_re_routing - **idx_type**:覆盖 0/1 - **expert_token_num_type**:覆盖 0/1 - **per_token_scales**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16/int8/int32/int64 ### moe_token_permute - **numOutTokens**:覆盖 0(full)/10(active)/20(active) - **paddedMode**:False - **dtype**:覆盖 fp32/fp16 ### moe_token_permute_grad - **numTopk**:覆盖 8 - **paddedMode**:False - **dtype**:覆盖 fp32/fp16 ### moe_token_unpermute - **paddedMode**:False - **probsOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16 ### moe_token_unpermute_grad - **paddedMode**:False - **probsOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16 ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4472 ## 测试 <!--描述进行了哪些测试来验证你的改动。包括但不限于二级冒烟、算子泛化等。--> ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [x] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [x] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10344 | 2 months ago | |
添加moe_init_routing_v4和moe_re_routing_v2算子的st&标杆 Co-authored-by: cpy_123456<chenpengyi@huawei.com> # message auto-generated for no-merge-commit merge: !10530 merge ut into master 添加moe_init_routing_v4和moe_re_routing_v2算子的st&标杆 Created-by: cpy_123456 Commit-by: cpy_123456 Merged-by: cann-robot Description: ## 描述 添加moe_init_routing_v4和moe_re_routing_v2算子的st&标杆 ## 一、moe_init_routing_v4(25 个用例,全部 PASS) ### 1.1 场景覆盖矩阵 | 维度 | 覆盖值 | 用例数 | |---|---|---| | **输入数据类型 (x)** | int8 | 2 | | | float16 | 7 | | | bfloat16 | 8 | | | float32 | 8 | | **quant_mode** | -1(非量化,DropLess) | 13 | | | 1(动态量化) | 12 | | **expert_tokens_num_type (etnt)** | 1(直方图统计) | 13 | | | 2(key-value 对) | 12 | | **drop_pad_mode** | 0(DropLess 模式) | 25 | | **expert_tokens_num_flag** | True | 25 | | **row_idx_type** | 0(gather) | 7 | | | 1(scatter) | 18 | | **scale 输入** | 有 | 10 | | | 无 | 15 | | **topk_weight 输入** | 有 | 25 | | **expert_num 范围** | 256 ~ 2930 | — | | **active_expert_range 跨度** | 6 ~ 580 | — | ## 二、moe_re_routing_v2(25 个用例,全部 PASS) ### 2.1 场景覆盖矩阵 | 维度 | 覆盖值 | 用例数 | |---|---|---| | **tokens 数据类型** | int8 | 8 | | | float16 | 8 | | | bfloat16 | 9 | | **idx_type** | 0(正向索引,permute_token_idx=src_offset) | 12 | | | 1(反向索引,permute_token_idx=dst_offset) | 13 | | **per_token_scales 输入** | 有 | 12 | | | 无 | 13 | | **expert_topk_weight 输入** | 有 | 25 | | **rank_num(专家并行数)** | 1 | 17 | | | 2 | 3 | | | 3 | 1 | | | 4~5 | 2 | | | 7~52 | 2 | | **expert_num(每 rank 专家数)** | 1 | 17 | | | 2 | 5 | | | 3 | 1 | | | 5~6 | 2 | | **tokens 第二维 (hidden_size)** | 9 ~ 15267 | — | ## 关联的Issue <!-- 如果这个PR是为了解决特定的Issue,请在这里提供Issue链接。例如:关联Issue #000--> <!-- 如果这个PR是为了解决特定的问题单,请在这里描述问题单单号。--> ## 测试 st&标杆通过 ## 文档更新 https://gitcode.com/cann/ops-transformer/issues/4543 ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [x] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10530 | 2 months ago | |
安装重名头文件整改:common.h Co-authored-by: chenyifan<chenyifan66@h-partners.com> # message auto-generated for no-merge-commit merge: !10537 merge clear_repeat_h_comm into master 安装重名头文件整改:common.h Created-by: mutex_lock Commit-by: chenyifan Merged-by: cann-robot Description: ## 描述 按**安装重名**头文件规范,完成 ** common.h / moe_common.h / moe_v2_common.h 跨算子同名整改**,消除安装包内跨算子同名头冲突。 **原因**: **common.h(13 份)**、**moe_common.h(5 份)**、**moe_v2_common.h(7 份)** 属安装规则 **3.9**(仓内非对外、不同算子/组件目录拷贝同名头),且口径为「同名不同实现亦改名」。权威 common/include/op_kernel/common.h 与 mc2/common/op_kernel/moe_distribute_base.h **保留原名**。 **方法**:低风险只改次要副本——各算子/组件侧 common.h / moe_common.h / moe_v2_common.h / setup 侧 moe_distribute_base.h 加算子或组件前缀(如 agm_v2_、matmul_all_reduce_、moe_init_routing_、mega_moe_、setup_ 等);同步更新 #include、\file、include guard。无业务逻辑变更。禁止软链接。 ### 安装分类规则 | # | 场景 | 是否整改 | 处置 | |---|------|---------|------| | 3.9 | 仓内非对外同名拷贝 | 是 | 次要侧改名 | | 权威 common | common/include/op_kernel/common.h | 否 | 保留原名 | | 权威 mc2 base | mc2/common/op_kernel/moe_distribute_base.h | 否 | 保留原名(v2/combine 等仍 #include 此路径) | ### 范围与结果 | 项 | 说明 | |----|------| | 重命名组数 | **26** | | 涉及域 | attention(scatter_pa / sfag / sfmg)、examples fkle、mc2(agm_v2 / matmul_all_reduce / mega_moe / dispatch setup/teardown/v2、mhc tiling)、moe(gating / init_routing / token_permute 系列) | | 附带修复 | dispatch setup base/common 映射;teardown 跨组件引用;v2 arch35 a5_ccu 裸 include;v3/combine_v3 apt 依赖 v2/v2_combine 的 __has_include 多布局路径 | | 变更文件数 | **166** | | 行变更 | +253 / -238 | ### 26 组重命名(旧 → 新) | # | 旧路径 | 新路径 | |--:|--------|--------| | 1 | attention/scatter_pa_kv_cache/op_kernel/arch35/common.h | attention/scatter_pa_kv_cache/op_kernel/arch35/scatter_pa_kv_cache_arch35_common.h | | 2 | attention/sparse_flash_attention_grad/op_kernel/arch35/common.h | attention/sparse_flash_attention_grad/op_kernel/arch35/sparse_flash_attention_grad_arch35_common.h | | 3 | attention/sparse_flash_mla_grad/op_kernel/arch35/common.h | attention/sparse_flash_mla_grad/op_kernel/arch35/sparse_flash_mla_grad_arch35_common.h | | 4 | examples/fast_kernel_launch_example/csrc/moe_distribute_dispatch_v2/ascend910_93/op_kernel/common.h | examples/fast_kernel_launch_example/csrc/moe_distribute_dispatch_v2/ascend910_93/op_kernel/fkle_mdd_v2_common.h | | 5 | mc2/all_gather_matmul_v2/op_kernel/common.h | mc2/all_gather_matmul_v2/op_kernel/agm_v2_common.h | | 6 | mc2/matmul_all_reduce/op_kernel/common.h | mc2/matmul_all_reduce/op_kernel/matmul_all_reduce_common.h | | 7 | mc2/moe_distribute_dispatch_setup/op_kernel/common.h | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_common.h | | 8 | mc2/moe_distribute_dispatch_v2/op_kernel/common.h | mc2/moe_distribute_dispatch_v2/op_kernel/moe_distribute_dispatch_v2_common.h | | 9 | mhc/mhc_pre_sinkhorn_backward/op_host/op_tiling/arch22/common.h | mhc/mhc_pre_sinkhorn_backward/op_host/op_tiling/arch22/mhc_pre_sinkhorn_backward_common.h | | 10 | moe/moe_gating_top_k/op_kernel/arch35/common.h | moe/moe_gating_top_k/op_kernel/arch35/moe_gating_top_k_arch35_common.h | | 11 | moe/moe_gating_top_k/op_kernel/common.h | moe/moe_gating_top_k/op_kernel/moe_gating_top_k_common.h | | 12 | moe/moe_gating_top_k_backward/op_kernel/arch35/common.h | moe/moe_gating_top_k_backward/op_kernel/arch35/moe_gating_top_k_backward_arch35_common.h | | 13 | moe/moe_gating_top_k_backward/op_kernel/common.h | moe/moe_gating_top_k_backward/op_kernel/moe_gating_top_k_backward_common.h | | 14 | moe/moe_init_routing/op_kernel/moe_common.h | moe/moe_init_routing/op_kernel/moe_init_routing_common.h | | 15 | moe/moe_init_routing/op_kernel/arch35/moe_common.h | moe/moe_init_routing/op_kernel/arch35/moe_init_routing_arch35_common.h | | 16 | moe/moe_token_permute/op_kernel/moe_common.h | moe/moe_token_permute/op_kernel/moe_token_permute_common.h | | 17 | moe/moe_token_permute_with_ep/op_kernel/moe_common.h | moe/moe_token_permute_with_ep/op_kernel/moe_token_permute_with_ep_common.h | | 18 | moe/moe_token_permute_with_routing_map/op_kernel/moe_common.h | moe/moe_token_permute_with_routing_map/op_kernel/moe_token_permute_with_routing_map_common.h | | 19 | mc2/mega_moe/op_kernel/arch22/moe_init_routing_v2/moe_v2_common.h | mc2/mega_moe/op_kernel/arch22/moe_init_routing_v2/mega_moe_moe_v2_common.h | | 20 | mc2/mega_moe/op_kernel/arch22/moe_init_routing_quant_v2/moe_v2_common.h | mc2/mega_moe/op_kernel/arch22/moe_init_routing_quant_v2/mega_moe_quant_moe_v2_common.h | | 21 | moe/moe_init_routing_v2/op_kernel/moe_v2_common.h | moe/moe_init_routing_v2/op_kernel/moe_init_routing_v2_common.h | | 22 | moe/moe_init_routing_v2/op_kernel/arch35/moe_v2_common.h | moe/moe_init_routing_v2/op_kernel/arch35/moe_init_routing_v2_arch35_common.h | | 23 | moe/moe_init_routing_quant_v2/inner/moe_v2_common.h | moe/moe_init_routing_quant_v2/inner/moe_init_routing_quant_v2_pkg_inner_common.h | | 24 | moe/moe_init_routing_quant_v2/op_kernel/arch35/moe_v2_common.h | moe/moe_init_routing_quant_v2/op_kernel/arch35/moe_init_routing_quant_v2_arch35_common.h | | 25 | moe/moe_init_routing_quant_v2/op_kernel/inner/moe_v2_common.h | moe/moe_init_routing_quant_v2/op_kernel/inner/moe_init_routing_quant_v2_inner_common.h | | 26 | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_base.h | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_base.h | ### 命名约定 | 场景 | 约定 | |------|------| | all_gather_matmul_v2 | agm_v2_* | | matmul_all_reduce | matmul_all_reduce_common.h | | mega_moe 内 moe_v2 | mega_moe_* / mega_moe_quant_* | | moe 算子 | moe_<op>_common.h、arch35 加 _arch35_common | | dispatch setup | setup_base.h / setup_common.h(与 mc2/common 权威 base 区分) | | include guard | 与文件名一致(如 MOE_DISTRIBUTE_DISPATCH_V2_COMMON_H) | ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4356 ## 测试 ## 文档更新 ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [x] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10537 | 2 months ago | |
增加moe蓝区ttk用例适配 Co-authored-by: cpy_123456<chenpengyi@huawei.com> # message auto-generated for no-merge-commit merge: !10344 merge ut into master 增加moe蓝区ttk用例适配 Created-by: cpy_123456 Commit-by: cpy_123456 Merged-by: cann-robot Description: ## 描述 # MoE 算子 ST 用例选取文档 ## 1. 选取原则 - **ACLNN 模式**:每个算子选取约 20 个用例(ACLNN 调用开销较大,约 2s/用例) - **Kernel 模式**:每个算子选取约 30 个用例(Kernel 编译+执行较快,约 0.4s/用例) - 覆盖所有 dtype(float32/float16/bfloat16/int8/int32/int64) - 覆盖所有关键属性组合(dropPadMode/quantMode/normType/renorm 等) - 覆盖可选输入的组合(有/无 optional inputs) - 优先选取小 shape 用例(减少运行时间) - 排除已知精度边界用例(量化精度差异、bf16 reduction 差异等) - 排除依赖算子代码修改的用例(如 moe_token_unpermute_grad active 场景) - 总运行时间控制在 10 分钟以内 ## 2. 选取结果 ### ACLNN 模式(12 个 CSV,150 个用例) | CSV 文件 | 算子目录 | 选取/全量 | 运行时间 | PASS | |---------|---------|----------|---------|------| | test_aclnn_moe_gating_top_k_arch35.csv | moe_gating_top_k | 20/38 | 42s | 20/20 | | test_aclnn_moe_gating_top_k_v2_arch35.csv | moe_gating_top_k | 20/32 | 46s | 20/20 | | test_aclnn_moe_gating_top_k_softmax_v2_arch35.csv | moe_gating_top_k_softmax_v2 | 20/232 | 47s | 20/20 | | test_aclnn_moe_init_routing_arch35.csv | moe_init_routing | 10/10 | 36s | 10/10 | | test_aclnn_moe_init_routing_v2_arch35.csv | moe_init_routing_v2 | 20/27 | 57s | 20/20 | | test_aclnn_moe_init_routing_v2_grad_arch35.csv | moe_init_routing_v2_grad | 20/116 | 51s | 20/20 | | test_aclnn_moe_init_routing_quant_v2_arch35.csv | moe_init_routing_quant_v2 | 15/34 | 48s | 15/15 | | test_aclnn_moe_finalize_routing_v2_grad_arch35.csv | moe_finalize_routing_v2_grad | 11/94 | 30s | 11/11 | | test_aclnn_moe_token_permute_arch35.csv | moe_token_permute | 4/4 | 19s | 4/4 | | test_aclnn_moe_token_permute_grad_arch35.csv | moe_token_permute_grad | 4/4 | 18s | 4/4 | | test_aclnn_moe_token_unpermute_arch35.csv | moe_token_unpermute | 2/2 | 15s | 2/2 | | test_aclnn_moe_token_unpermute_grad_arch35.csv | moe_token_unpermute_grad | 4/4 | 19s | 4/4 | ### Kernel 模式(10 个 CSV,300 个用例) | CSV 文件 | 算子目录 | 选取/全量 | 运行时间 | PASS | |---------|---------|----------|---------|------| | test_moe_gating_top_k_arch35.csv | moe_gating_top_k | 30/829 | 11s | 30/30 | | test_moe_gating_top_k_softmax_v2_arch35.csv | moe_gating_top_k_softmax_v2 | 30/832 | 13s | 30/30 | | test_moe_init_routing_arch35.csv | moe_init_routing | 30/203 | 13s | 30/30 | | test_moe_init_routing_v2_arch35.csv | moe_init_routing_v2 | 30/200 | 13s | 30/30 | | test_moe_init_routing_v2_grad_arch35.csv | moe_init_routing_v2_grad | 30/144 | 11s | 30/30 | | test_moe_init_routing_quant_v2_arch35.csv | moe_init_routing_quant_v2 | 30/190 | 13s | 30/30 | | test_moe_finalize_routing_v2_arch35.csv | moe_finalize_routing_v2 | 30/112 | 13s | 30/30 | | test_moe_finalize_routing_v2_grad_arch35.csv | moe_finalize_routing_v2_grad | 30/198 | 13s | 30/30 | | test_moe_compute_expert_tokens_arch35.csv | moe_compute_expert_tokens | 30/159 | 13s | 30/30 | | test_moe_re_routing_arch35.csv | moe_re_routing | 30/578 | 11s | 30/30 | ## 3. 验证汇总 | 指标 | 值 | |------|------| | 选取用例总数 | 450 | | 全量用例总数 | 3726 | | PASS | 450 (100%) | | FAIL | 0 | | ACLNN 运行时间 | 428s (7m08s) | | Kernel 运行时间 | 124s (2m04s) | | 总运行时间 | 552s (9m12s) | | 平均每用例 | 1.23s | ## 4. 排除用例说明 ## 5. 功能覆盖说明 ### moe_gating_top_k - **norm_type**:覆盖 0(softmax)/1(sigmoid)/2(sqrtsoftplus) - **group_count**:覆盖 1(无分组)/8(标准分组)/大值(多分组) - **group_select_mode**:覆盖 0(amax)/1(top2sum) - **out_flag**:覆盖 True/False - **renorm**:覆盖 0/1 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_gating_top_k_v2 - **normType**:覆盖 1(sigmoid) - **inputIdsOptional/tid2eidOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_gating_top_k_softmax_v2 - **renorm**:覆盖 0(softmax后取topk)/1(topk后renorm) - **finishedOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing - **active_num**:覆盖小值(1~5)/大值(100+) - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_v2 - **drop_pad_mode**:覆盖 0(dropless)/1(drop/pad) - **expert_tokens_count_or_cumsum_flag**:覆盖 0/1(cumsum)/2(count) - **expert_tokens_before_capacity_flag**:覆盖 True/False - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_v2_grad - **drop_pad_mode**:覆盖 0/1 - **active_num**:覆盖 0(full)/1(active)/大值 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_quant_v2 - **drop_pad_mode**:覆盖 0/1 - **quant_mode**:覆盖 0(静态量化)/1(动态量化) - **expert_tokens_count_or_cumsum_flag**:覆盖 0/1/2 - **expert_tokens_before_capacity_flag**:覆盖 True/False - **dtype**:覆盖 fp32/fp16/bf16 ### moe_finalize_routing_v2 - **drop_pad_mode**:覆盖 0/2/3 - **zero/copy/constant_expert_range**:覆盖各种 expert range 配置 - **optional inputs(x/a1/a2/v)**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_finalize_routing_v2_grad - **drop_pad_mode**:覆盖 0/1 - **active_num**:覆盖 0/1/大值 - **scales**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_compute_expert_tokens - **num_experts**:覆盖小值(8)/大值(800+) - **dtype**:int32 ### moe_re_routing - **idx_type**:覆盖 0/1 - **expert_token_num_type**:覆盖 0/1 - **per_token_scales**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16/int8/int32/int64 ### moe_token_permute - **numOutTokens**:覆盖 0(full)/10(active)/20(active) - **paddedMode**:False - **dtype**:覆盖 fp32/fp16 ### moe_token_permute_grad - **numTopk**:覆盖 8 - **paddedMode**:False - **dtype**:覆盖 fp32/fp16 ### moe_token_unpermute - **paddedMode**:False - **probsOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16 ### moe_token_unpermute_grad - **paddedMode**:False - **probsOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16 ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4472 ## 测试 <!--描述进行了哪些测试来验证你的改动。包括但不限于二级冒烟、算子泛化等。--> ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [x] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [x] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10344 | 2 months ago | |
安装重名头文件整改:common.h Co-authored-by: chenyifan<chenyifan66@h-partners.com> # message auto-generated for no-merge-commit merge: !10537 merge clear_repeat_h_comm into master 安装重名头文件整改:common.h Created-by: mutex_lock Commit-by: chenyifan Merged-by: cann-robot Description: ## 描述 按**安装重名**头文件规范,完成 ** common.h / moe_common.h / moe_v2_common.h 跨算子同名整改**,消除安装包内跨算子同名头冲突。 **原因**: **common.h(13 份)**、**moe_common.h(5 份)**、**moe_v2_common.h(7 份)** 属安装规则 **3.9**(仓内非对外、不同算子/组件目录拷贝同名头),且口径为「同名不同实现亦改名」。权威 common/include/op_kernel/common.h 与 mc2/common/op_kernel/moe_distribute_base.h **保留原名**。 **方法**:低风险只改次要副本——各算子/组件侧 common.h / moe_common.h / moe_v2_common.h / setup 侧 moe_distribute_base.h 加算子或组件前缀(如 agm_v2_、matmul_all_reduce_、moe_init_routing_、mega_moe_、setup_ 等);同步更新 #include、\file、include guard。无业务逻辑变更。禁止软链接。 ### 安装分类规则 | # | 场景 | 是否整改 | 处置 | |---|------|---------|------| | 3.9 | 仓内非对外同名拷贝 | 是 | 次要侧改名 | | 权威 common | common/include/op_kernel/common.h | 否 | 保留原名 | | 权威 mc2 base | mc2/common/op_kernel/moe_distribute_base.h | 否 | 保留原名(v2/combine 等仍 #include 此路径) | ### 范围与结果 | 项 | 说明 | |----|------| | 重命名组数 | **26** | | 涉及域 | attention(scatter_pa / sfag / sfmg)、examples fkle、mc2(agm_v2 / matmul_all_reduce / mega_moe / dispatch setup/teardown/v2、mhc tiling)、moe(gating / init_routing / token_permute 系列) | | 附带修复 | dispatch setup base/common 映射;teardown 跨组件引用;v2 arch35 a5_ccu 裸 include;v3/combine_v3 apt 依赖 v2/v2_combine 的 __has_include 多布局路径 | | 变更文件数 | **166** | | 行变更 | +253 / -238 | ### 26 组重命名(旧 → 新) | # | 旧路径 | 新路径 | |--:|--------|--------| | 1 | attention/scatter_pa_kv_cache/op_kernel/arch35/common.h | attention/scatter_pa_kv_cache/op_kernel/arch35/scatter_pa_kv_cache_arch35_common.h | | 2 | attention/sparse_flash_attention_grad/op_kernel/arch35/common.h | attention/sparse_flash_attention_grad/op_kernel/arch35/sparse_flash_attention_grad_arch35_common.h | | 3 | attention/sparse_flash_mla_grad/op_kernel/arch35/common.h | attention/sparse_flash_mla_grad/op_kernel/arch35/sparse_flash_mla_grad_arch35_common.h | | 4 | examples/fast_kernel_launch_example/csrc/moe_distribute_dispatch_v2/ascend910_93/op_kernel/common.h | examples/fast_kernel_launch_example/csrc/moe_distribute_dispatch_v2/ascend910_93/op_kernel/fkle_mdd_v2_common.h | | 5 | mc2/all_gather_matmul_v2/op_kernel/common.h | mc2/all_gather_matmul_v2/op_kernel/agm_v2_common.h | | 6 | mc2/matmul_all_reduce/op_kernel/common.h | mc2/matmul_all_reduce/op_kernel/matmul_all_reduce_common.h | | 7 | mc2/moe_distribute_dispatch_setup/op_kernel/common.h | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_common.h | | 8 | mc2/moe_distribute_dispatch_v2/op_kernel/common.h | mc2/moe_distribute_dispatch_v2/op_kernel/moe_distribute_dispatch_v2_common.h | | 9 | mhc/mhc_pre_sinkhorn_backward/op_host/op_tiling/arch22/common.h | mhc/mhc_pre_sinkhorn_backward/op_host/op_tiling/arch22/mhc_pre_sinkhorn_backward_common.h | | 10 | moe/moe_gating_top_k/op_kernel/arch35/common.h | moe/moe_gating_top_k/op_kernel/arch35/moe_gating_top_k_arch35_common.h | | 11 | moe/moe_gating_top_k/op_kernel/common.h | moe/moe_gating_top_k/op_kernel/moe_gating_top_k_common.h | | 12 | moe/moe_gating_top_k_backward/op_kernel/arch35/common.h | moe/moe_gating_top_k_backward/op_kernel/arch35/moe_gating_top_k_backward_arch35_common.h | | 13 | moe/moe_gating_top_k_backward/op_kernel/common.h | moe/moe_gating_top_k_backward/op_kernel/moe_gating_top_k_backward_common.h | | 14 | moe/moe_init_routing/op_kernel/moe_common.h | moe/moe_init_routing/op_kernel/moe_init_routing_common.h | | 15 | moe/moe_init_routing/op_kernel/arch35/moe_common.h | moe/moe_init_routing/op_kernel/arch35/moe_init_routing_arch35_common.h | | 16 | moe/moe_token_permute/op_kernel/moe_common.h | moe/moe_token_permute/op_kernel/moe_token_permute_common.h | | 17 | moe/moe_token_permute_with_ep/op_kernel/moe_common.h | moe/moe_token_permute_with_ep/op_kernel/moe_token_permute_with_ep_common.h | | 18 | moe/moe_token_permute_with_routing_map/op_kernel/moe_common.h | moe/moe_token_permute_with_routing_map/op_kernel/moe_token_permute_with_routing_map_common.h | | 19 | mc2/mega_moe/op_kernel/arch22/moe_init_routing_v2/moe_v2_common.h | mc2/mega_moe/op_kernel/arch22/moe_init_routing_v2/mega_moe_moe_v2_common.h | | 20 | mc2/mega_moe/op_kernel/arch22/moe_init_routing_quant_v2/moe_v2_common.h | mc2/mega_moe/op_kernel/arch22/moe_init_routing_quant_v2/mega_moe_quant_moe_v2_common.h | | 21 | moe/moe_init_routing_v2/op_kernel/moe_v2_common.h | moe/moe_init_routing_v2/op_kernel/moe_init_routing_v2_common.h | | 22 | moe/moe_init_routing_v2/op_kernel/arch35/moe_v2_common.h | moe/moe_init_routing_v2/op_kernel/arch35/moe_init_routing_v2_arch35_common.h | | 23 | moe/moe_init_routing_quant_v2/inner/moe_v2_common.h | moe/moe_init_routing_quant_v2/inner/moe_init_routing_quant_v2_pkg_inner_common.h | | 24 | moe/moe_init_routing_quant_v2/op_kernel/arch35/moe_v2_common.h | moe/moe_init_routing_quant_v2/op_kernel/arch35/moe_init_routing_quant_v2_arch35_common.h | | 25 | moe/moe_init_routing_quant_v2/op_kernel/inner/moe_v2_common.h | moe/moe_init_routing_quant_v2/op_kernel/inner/moe_init_routing_quant_v2_inner_common.h | | 26 | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_base.h | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_base.h | ### 命名约定 | 场景 | 约定 | |------|------| | all_gather_matmul_v2 | agm_v2_* | | matmul_all_reduce | matmul_all_reduce_common.h | | mega_moe 内 moe_v2 | mega_moe_* / mega_moe_quant_* | | moe 算子 | moe_<op>_common.h、arch35 加 _arch35_common | | dispatch setup | setup_base.h / setup_common.h(与 mc2/common 权威 base 区分) | | include guard | 与文件名一致(如 MOE_DISTRIBUTE_DISPATCH_V2_COMMON_H) | ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4356 ## 测试 ## 文档更新 ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [x] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10537 | 2 months ago | |
Moe目录日志规范整改 Co-authored-by: guoqiuhao<guoqiuhao1@huawei.com> # message auto-generated for no-merge-commit merge: !10386 merge log into master Moe目录日志规范整改 Created-by: guoqiuhao Commit-by: guoqiuhao Merged-by: cann-robot Description: ## 描述 Moe目录日志规范整改. 1.禁止中文日志 2.英文拼写错误 3.日志级别错误 4.语法错误 5.禁止自定义打印 6.错误描述不完整 7.冗余日志 8.未知缩写 9.内部敏感词 10.缺少度量单位 ## 关联的Issue [#4476](https://gitcode.com/cann/ops-transformer/issues/4476) ## 测试 不涉及 ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10386 | 2 months ago | |
安装重名头文件整改:common.h Co-authored-by: chenyifan<chenyifan66@h-partners.com> # message auto-generated for no-merge-commit merge: !10537 merge clear_repeat_h_comm into master 安装重名头文件整改:common.h Created-by: mutex_lock Commit-by: chenyifan Merged-by: cann-robot Description: ## 描述 按**安装重名**头文件规范,完成 ** common.h / moe_common.h / moe_v2_common.h 跨算子同名整改**,消除安装包内跨算子同名头冲突。 **原因**: **common.h(13 份)**、**moe_common.h(5 份)**、**moe_v2_common.h(7 份)** 属安装规则 **3.9**(仓内非对外、不同算子/组件目录拷贝同名头),且口径为「同名不同实现亦改名」。权威 common/include/op_kernel/common.h 与 mc2/common/op_kernel/moe_distribute_base.h **保留原名**。 **方法**:低风险只改次要副本——各算子/组件侧 common.h / moe_common.h / moe_v2_common.h / setup 侧 moe_distribute_base.h 加算子或组件前缀(如 agm_v2_、matmul_all_reduce_、moe_init_routing_、mega_moe_、setup_ 等);同步更新 #include、\file、include guard。无业务逻辑变更。禁止软链接。 ### 安装分类规则 | # | 场景 | 是否整改 | 处置 | |---|------|---------|------| | 3.9 | 仓内非对外同名拷贝 | 是 | 次要侧改名 | | 权威 common | common/include/op_kernel/common.h | 否 | 保留原名 | | 权威 mc2 base | mc2/common/op_kernel/moe_distribute_base.h | 否 | 保留原名(v2/combine 等仍 #include 此路径) | ### 范围与结果 | 项 | 说明 | |----|------| | 重命名组数 | **26** | | 涉及域 | attention(scatter_pa / sfag / sfmg)、examples fkle、mc2(agm_v2 / matmul_all_reduce / mega_moe / dispatch setup/teardown/v2、mhc tiling)、moe(gating / init_routing / token_permute 系列) | | 附带修复 | dispatch setup base/common 映射;teardown 跨组件引用;v2 arch35 a5_ccu 裸 include;v3/combine_v3 apt 依赖 v2/v2_combine 的 __has_include 多布局路径 | | 变更文件数 | **166** | | 行变更 | +253 / -238 | ### 26 组重命名(旧 → 新) | # | 旧路径 | 新路径 | |--:|--------|--------| | 1 | attention/scatter_pa_kv_cache/op_kernel/arch35/common.h | attention/scatter_pa_kv_cache/op_kernel/arch35/scatter_pa_kv_cache_arch35_common.h | | 2 | attention/sparse_flash_attention_grad/op_kernel/arch35/common.h | attention/sparse_flash_attention_grad/op_kernel/arch35/sparse_flash_attention_grad_arch35_common.h | | 3 | attention/sparse_flash_mla_grad/op_kernel/arch35/common.h | attention/sparse_flash_mla_grad/op_kernel/arch35/sparse_flash_mla_grad_arch35_common.h | | 4 | examples/fast_kernel_launch_example/csrc/moe_distribute_dispatch_v2/ascend910_93/op_kernel/common.h | examples/fast_kernel_launch_example/csrc/moe_distribute_dispatch_v2/ascend910_93/op_kernel/fkle_mdd_v2_common.h | | 5 | mc2/all_gather_matmul_v2/op_kernel/common.h | mc2/all_gather_matmul_v2/op_kernel/agm_v2_common.h | | 6 | mc2/matmul_all_reduce/op_kernel/common.h | mc2/matmul_all_reduce/op_kernel/matmul_all_reduce_common.h | | 7 | mc2/moe_distribute_dispatch_setup/op_kernel/common.h | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_common.h | | 8 | mc2/moe_distribute_dispatch_v2/op_kernel/common.h | mc2/moe_distribute_dispatch_v2/op_kernel/moe_distribute_dispatch_v2_common.h | | 9 | mhc/mhc_pre_sinkhorn_backward/op_host/op_tiling/arch22/common.h | mhc/mhc_pre_sinkhorn_backward/op_host/op_tiling/arch22/mhc_pre_sinkhorn_backward_common.h | | 10 | moe/moe_gating_top_k/op_kernel/arch35/common.h | moe/moe_gating_top_k/op_kernel/arch35/moe_gating_top_k_arch35_common.h | | 11 | moe/moe_gating_top_k/op_kernel/common.h | moe/moe_gating_top_k/op_kernel/moe_gating_top_k_common.h | | 12 | moe/moe_gating_top_k_backward/op_kernel/arch35/common.h | moe/moe_gating_top_k_backward/op_kernel/arch35/moe_gating_top_k_backward_arch35_common.h | | 13 | moe/moe_gating_top_k_backward/op_kernel/common.h | moe/moe_gating_top_k_backward/op_kernel/moe_gating_top_k_backward_common.h | | 14 | moe/moe_init_routing/op_kernel/moe_common.h | moe/moe_init_routing/op_kernel/moe_init_routing_common.h | | 15 | moe/moe_init_routing/op_kernel/arch35/moe_common.h | moe/moe_init_routing/op_kernel/arch35/moe_init_routing_arch35_common.h | | 16 | moe/moe_token_permute/op_kernel/moe_common.h | moe/moe_token_permute/op_kernel/moe_token_permute_common.h | | 17 | moe/moe_token_permute_with_ep/op_kernel/moe_common.h | moe/moe_token_permute_with_ep/op_kernel/moe_token_permute_with_ep_common.h | | 18 | moe/moe_token_permute_with_routing_map/op_kernel/moe_common.h | moe/moe_token_permute_with_routing_map/op_kernel/moe_token_permute_with_routing_map_common.h | | 19 | mc2/mega_moe/op_kernel/arch22/moe_init_routing_v2/moe_v2_common.h | mc2/mega_moe/op_kernel/arch22/moe_init_routing_v2/mega_moe_moe_v2_common.h | | 20 | mc2/mega_moe/op_kernel/arch22/moe_init_routing_quant_v2/moe_v2_common.h | mc2/mega_moe/op_kernel/arch22/moe_init_routing_quant_v2/mega_moe_quant_moe_v2_common.h | | 21 | moe/moe_init_routing_v2/op_kernel/moe_v2_common.h | moe/moe_init_routing_v2/op_kernel/moe_init_routing_v2_common.h | | 22 | moe/moe_init_routing_v2/op_kernel/arch35/moe_v2_common.h | moe/moe_init_routing_v2/op_kernel/arch35/moe_init_routing_v2_arch35_common.h | | 23 | moe/moe_init_routing_quant_v2/inner/moe_v2_common.h | moe/moe_init_routing_quant_v2/inner/moe_init_routing_quant_v2_pkg_inner_common.h | | 24 | moe/moe_init_routing_quant_v2/op_kernel/arch35/moe_v2_common.h | moe/moe_init_routing_quant_v2/op_kernel/arch35/moe_init_routing_quant_v2_arch35_common.h | | 25 | moe/moe_init_routing_quant_v2/op_kernel/inner/moe_v2_common.h | moe/moe_init_routing_quant_v2/op_kernel/inner/moe_init_routing_quant_v2_inner_common.h | | 26 | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_base.h | mc2/moe_distribute_dispatch_setup/op_kernel/moe_distribute_dispatch_setup_base.h | ### 命名约定 | 场景 | 约定 | |------|------| | all_gather_matmul_v2 | agm_v2_* | | matmul_all_reduce | matmul_all_reduce_common.h | | mega_moe 内 moe_v2 | mega_moe_* / mega_moe_quant_* | | moe 算子 | moe_<op>_common.h、arch35 加 _arch35_common | | dispatch setup | setup_base.h / setup_common.h(与 mc2/common 权威 base 区分) | | include guard | 与文件名一致(如 MOE_DISTRIBUTE_DISPATCH_V2_COMMON_H) | ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4356 ## 测试 ## 文档更新 ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [x] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10537 | 2 months ago | |
Moe目录日志规范整改 Co-authored-by: guoqiuhao<guoqiuhao1@huawei.com> # message auto-generated for no-merge-commit merge: !10386 merge log into master Moe目录日志规范整改 Created-by: guoqiuhao Commit-by: guoqiuhao Merged-by: cann-robot Description: ## 描述 Moe目录日志规范整改. 1.禁止中文日志 2.英文拼写错误 3.日志级别错误 4.语法错误 5.禁止自定义打印 6.错误描述不完整 7.冗余日志 8.未知缩写 9.内部敏感词 10.缺少度量单位 ## 关联的Issue [#4476](https://gitcode.com/cann/ops-transformer/issues/4476) ## 测试 不涉及 ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10386 | 2 months ago | |
增加moe蓝区ttk用例适配 Co-authored-by: cpy_123456<chenpengyi@huawei.com> # message auto-generated for no-merge-commit merge: !10344 merge ut into master 增加moe蓝区ttk用例适配 Created-by: cpy_123456 Commit-by: cpy_123456 Merged-by: cann-robot Description: ## 描述 # MoE 算子 ST 用例选取文档 ## 1. 选取原则 - **ACLNN 模式**:每个算子选取约 20 个用例(ACLNN 调用开销较大,约 2s/用例) - **Kernel 模式**:每个算子选取约 30 个用例(Kernel 编译+执行较快,约 0.4s/用例) - 覆盖所有 dtype(float32/float16/bfloat16/int8/int32/int64) - 覆盖所有关键属性组合(dropPadMode/quantMode/normType/renorm 等) - 覆盖可选输入的组合(有/无 optional inputs) - 优先选取小 shape 用例(减少运行时间) - 排除已知精度边界用例(量化精度差异、bf16 reduction 差异等) - 排除依赖算子代码修改的用例(如 moe_token_unpermute_grad active 场景) - 总运行时间控制在 10 分钟以内 ## 2. 选取结果 ### ACLNN 模式(12 个 CSV,150 个用例) | CSV 文件 | 算子目录 | 选取/全量 | 运行时间 | PASS | |---------|---------|----------|---------|------| | test_aclnn_moe_gating_top_k_arch35.csv | moe_gating_top_k | 20/38 | 42s | 20/20 | | test_aclnn_moe_gating_top_k_v2_arch35.csv | moe_gating_top_k | 20/32 | 46s | 20/20 | | test_aclnn_moe_gating_top_k_softmax_v2_arch35.csv | moe_gating_top_k_softmax_v2 | 20/232 | 47s | 20/20 | | test_aclnn_moe_init_routing_arch35.csv | moe_init_routing | 10/10 | 36s | 10/10 | | test_aclnn_moe_init_routing_v2_arch35.csv | moe_init_routing_v2 | 20/27 | 57s | 20/20 | | test_aclnn_moe_init_routing_v2_grad_arch35.csv | moe_init_routing_v2_grad | 20/116 | 51s | 20/20 | | test_aclnn_moe_init_routing_quant_v2_arch35.csv | moe_init_routing_quant_v2 | 15/34 | 48s | 15/15 | | test_aclnn_moe_finalize_routing_v2_grad_arch35.csv | moe_finalize_routing_v2_grad | 11/94 | 30s | 11/11 | | test_aclnn_moe_token_permute_arch35.csv | moe_token_permute | 4/4 | 19s | 4/4 | | test_aclnn_moe_token_permute_grad_arch35.csv | moe_token_permute_grad | 4/4 | 18s | 4/4 | | test_aclnn_moe_token_unpermute_arch35.csv | moe_token_unpermute | 2/2 | 15s | 2/2 | | test_aclnn_moe_token_unpermute_grad_arch35.csv | moe_token_unpermute_grad | 4/4 | 19s | 4/4 | ### Kernel 模式(10 个 CSV,300 个用例) | CSV 文件 | 算子目录 | 选取/全量 | 运行时间 | PASS | |---------|---------|----------|---------|------| | test_moe_gating_top_k_arch35.csv | moe_gating_top_k | 30/829 | 11s | 30/30 | | test_moe_gating_top_k_softmax_v2_arch35.csv | moe_gating_top_k_softmax_v2 | 30/832 | 13s | 30/30 | | test_moe_init_routing_arch35.csv | moe_init_routing | 30/203 | 13s | 30/30 | | test_moe_init_routing_v2_arch35.csv | moe_init_routing_v2 | 30/200 | 13s | 30/30 | | test_moe_init_routing_v2_grad_arch35.csv | moe_init_routing_v2_grad | 30/144 | 11s | 30/30 | | test_moe_init_routing_quant_v2_arch35.csv | moe_init_routing_quant_v2 | 30/190 | 13s | 30/30 | | test_moe_finalize_routing_v2_arch35.csv | moe_finalize_routing_v2 | 30/112 | 13s | 30/30 | | test_moe_finalize_routing_v2_grad_arch35.csv | moe_finalize_routing_v2_grad | 30/198 | 13s | 30/30 | | test_moe_compute_expert_tokens_arch35.csv | moe_compute_expert_tokens | 30/159 | 13s | 30/30 | | test_moe_re_routing_arch35.csv | moe_re_routing | 30/578 | 11s | 30/30 | ## 3. 验证汇总 | 指标 | 值 | |------|------| | 选取用例总数 | 450 | | 全量用例总数 | 3726 | | PASS | 450 (100%) | | FAIL | 0 | | ACLNN 运行时间 | 428s (7m08s) | | Kernel 运行时间 | 124s (2m04s) | | 总运行时间 | 552s (9m12s) | | 平均每用例 | 1.23s | ## 4. 排除用例说明 ## 5. 功能覆盖说明 ### moe_gating_top_k - **norm_type**:覆盖 0(softmax)/1(sigmoid)/2(sqrtsoftplus) - **group_count**:覆盖 1(无分组)/8(标准分组)/大值(多分组) - **group_select_mode**:覆盖 0(amax)/1(top2sum) - **out_flag**:覆盖 True/False - **renorm**:覆盖 0/1 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_gating_top_k_v2 - **normType**:覆盖 1(sigmoid) - **inputIdsOptional/tid2eidOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_gating_top_k_softmax_v2 - **renorm**:覆盖 0(softmax后取topk)/1(topk后renorm) - **finishedOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing - **active_num**:覆盖小值(1~5)/大值(100+) - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_v2 - **drop_pad_mode**:覆盖 0(dropless)/1(drop/pad) - **expert_tokens_count_or_cumsum_flag**:覆盖 0/1(cumsum)/2(count) - **expert_tokens_before_capacity_flag**:覆盖 True/False - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_v2_grad - **drop_pad_mode**:覆盖 0/1 - **active_num**:覆盖 0(full)/1(active)/大值 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_quant_v2 - **drop_pad_mode**:覆盖 0/1 - **quant_mode**:覆盖 0(静态量化)/1(动态量化) - **expert_tokens_count_or_cumsum_flag**:覆盖 0/1/2 - **expert_tokens_before_capacity_flag**:覆盖 True/False - **dtype**:覆盖 fp32/fp16/bf16 ### moe_finalize_routing_v2 - **drop_pad_mode**:覆盖 0/2/3 - **zero/copy/constant_expert_range**:覆盖各种 expert range 配置 - **optional inputs(x/a1/a2/v)**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_finalize_routing_v2_grad - **drop_pad_mode**:覆盖 0/1 - **active_num**:覆盖 0/1/大值 - **scales**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_compute_expert_tokens - **num_experts**:覆盖小值(8)/大值(800+) - **dtype**:int32 ### moe_re_routing - **idx_type**:覆盖 0/1 - **expert_token_num_type**:覆盖 0/1 - **per_token_scales**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16/int8/int32/int64 ### moe_token_permute - **numOutTokens**:覆盖 0(full)/10(active)/20(active) - **paddedMode**:False - **dtype**:覆盖 fp32/fp16 ### moe_token_permute_grad - **numTopk**:覆盖 8 - **paddedMode**:False - **dtype**:覆盖 fp32/fp16 ### moe_token_unpermute - **paddedMode**:False - **probsOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16 ### moe_token_unpermute_grad - **paddedMode**:False - **probsOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16 ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4472 ## 测试 <!--描述进行了哪些测试来验证你的改动。包括但不限于二级冒烟、算子泛化等。--> ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [x] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [x] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10344 | 2 months ago | |
增加moe蓝区ttk用例适配 Co-authored-by: cpy_123456<chenpengyi@huawei.com> # message auto-generated for no-merge-commit merge: !10344 merge ut into master 增加moe蓝区ttk用例适配 Created-by: cpy_123456 Commit-by: cpy_123456 Merged-by: cann-robot Description: ## 描述 # MoE 算子 ST 用例选取文档 ## 1. 选取原则 - **ACLNN 模式**:每个算子选取约 20 个用例(ACLNN 调用开销较大,约 2s/用例) - **Kernel 模式**:每个算子选取约 30 个用例(Kernel 编译+执行较快,约 0.4s/用例) - 覆盖所有 dtype(float32/float16/bfloat16/int8/int32/int64) - 覆盖所有关键属性组合(dropPadMode/quantMode/normType/renorm 等) - 覆盖可选输入的组合(有/无 optional inputs) - 优先选取小 shape 用例(减少运行时间) - 排除已知精度边界用例(量化精度差异、bf16 reduction 差异等) - 排除依赖算子代码修改的用例(如 moe_token_unpermute_grad active 场景) - 总运行时间控制在 10 分钟以内 ## 2. 选取结果 ### ACLNN 模式(12 个 CSV,150 个用例) | CSV 文件 | 算子目录 | 选取/全量 | 运行时间 | PASS | |---------|---------|----------|---------|------| | test_aclnn_moe_gating_top_k_arch35.csv | moe_gating_top_k | 20/38 | 42s | 20/20 | | test_aclnn_moe_gating_top_k_v2_arch35.csv | moe_gating_top_k | 20/32 | 46s | 20/20 | | test_aclnn_moe_gating_top_k_softmax_v2_arch35.csv | moe_gating_top_k_softmax_v2 | 20/232 | 47s | 20/20 | | test_aclnn_moe_init_routing_arch35.csv | moe_init_routing | 10/10 | 36s | 10/10 | | test_aclnn_moe_init_routing_v2_arch35.csv | moe_init_routing_v2 | 20/27 | 57s | 20/20 | | test_aclnn_moe_init_routing_v2_grad_arch35.csv | moe_init_routing_v2_grad | 20/116 | 51s | 20/20 | | test_aclnn_moe_init_routing_quant_v2_arch35.csv | moe_init_routing_quant_v2 | 15/34 | 48s | 15/15 | | test_aclnn_moe_finalize_routing_v2_grad_arch35.csv | moe_finalize_routing_v2_grad | 11/94 | 30s | 11/11 | | test_aclnn_moe_token_permute_arch35.csv | moe_token_permute | 4/4 | 19s | 4/4 | | test_aclnn_moe_token_permute_grad_arch35.csv | moe_token_permute_grad | 4/4 | 18s | 4/4 | | test_aclnn_moe_token_unpermute_arch35.csv | moe_token_unpermute | 2/2 | 15s | 2/2 | | test_aclnn_moe_token_unpermute_grad_arch35.csv | moe_token_unpermute_grad | 4/4 | 19s | 4/4 | ### Kernel 模式(10 个 CSV,300 个用例) | CSV 文件 | 算子目录 | 选取/全量 | 运行时间 | PASS | |---------|---------|----------|---------|------| | test_moe_gating_top_k_arch35.csv | moe_gating_top_k | 30/829 | 11s | 30/30 | | test_moe_gating_top_k_softmax_v2_arch35.csv | moe_gating_top_k_softmax_v2 | 30/832 | 13s | 30/30 | | test_moe_init_routing_arch35.csv | moe_init_routing | 30/203 | 13s | 30/30 | | test_moe_init_routing_v2_arch35.csv | moe_init_routing_v2 | 30/200 | 13s | 30/30 | | test_moe_init_routing_v2_grad_arch35.csv | moe_init_routing_v2_grad | 30/144 | 11s | 30/30 | | test_moe_init_routing_quant_v2_arch35.csv | moe_init_routing_quant_v2 | 30/190 | 13s | 30/30 | | test_moe_finalize_routing_v2_arch35.csv | moe_finalize_routing_v2 | 30/112 | 13s | 30/30 | | test_moe_finalize_routing_v2_grad_arch35.csv | moe_finalize_routing_v2_grad | 30/198 | 13s | 30/30 | | test_moe_compute_expert_tokens_arch35.csv | moe_compute_expert_tokens | 30/159 | 13s | 30/30 | | test_moe_re_routing_arch35.csv | moe_re_routing | 30/578 | 11s | 30/30 | ## 3. 验证汇总 | 指标 | 值 | |------|------| | 选取用例总数 | 450 | | 全量用例总数 | 3726 | | PASS | 450 (100%) | | FAIL | 0 | | ACLNN 运行时间 | 428s (7m08s) | | Kernel 运行时间 | 124s (2m04s) | | 总运行时间 | 552s (9m12s) | | 平均每用例 | 1.23s | ## 4. 排除用例说明 ## 5. 功能覆盖说明 ### moe_gating_top_k - **norm_type**:覆盖 0(softmax)/1(sigmoid)/2(sqrtsoftplus) - **group_count**:覆盖 1(无分组)/8(标准分组)/大值(多分组) - **group_select_mode**:覆盖 0(amax)/1(top2sum) - **out_flag**:覆盖 True/False - **renorm**:覆盖 0/1 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_gating_top_k_v2 - **normType**:覆盖 1(sigmoid) - **inputIdsOptional/tid2eidOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_gating_top_k_softmax_v2 - **renorm**:覆盖 0(softmax后取topk)/1(topk后renorm) - **finishedOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing - **active_num**:覆盖小值(1~5)/大值(100+) - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_v2 - **drop_pad_mode**:覆盖 0(dropless)/1(drop/pad) - **expert_tokens_count_or_cumsum_flag**:覆盖 0/1(cumsum)/2(count) - **expert_tokens_before_capacity_flag**:覆盖 True/False - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_v2_grad - **drop_pad_mode**:覆盖 0/1 - **active_num**:覆盖 0(full)/1(active)/大值 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_init_routing_quant_v2 - **drop_pad_mode**:覆盖 0/1 - **quant_mode**:覆盖 0(静态量化)/1(动态量化) - **expert_tokens_count_or_cumsum_flag**:覆盖 0/1/2 - **expert_tokens_before_capacity_flag**:覆盖 True/False - **dtype**:覆盖 fp32/fp16/bf16 ### moe_finalize_routing_v2 - **drop_pad_mode**:覆盖 0/2/3 - **zero/copy/constant_expert_range**:覆盖各种 expert range 配置 - **optional inputs(x/a1/a2/v)**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_finalize_routing_v2_grad - **drop_pad_mode**:覆盖 0/1 - **active_num**:覆盖 0/1/大值 - **scales**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16 ### moe_compute_expert_tokens - **num_experts**:覆盖小值(8)/大值(800+) - **dtype**:int32 ### moe_re_routing - **idx_type**:覆盖 0/1 - **expert_token_num_type**:覆盖 0/1 - **per_token_scales**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16/bf16/int8/int32/int64 ### moe_token_permute - **numOutTokens**:覆盖 0(full)/10(active)/20(active) - **paddedMode**:False - **dtype**:覆盖 fp32/fp16 ### moe_token_permute_grad - **numTopk**:覆盖 8 - **paddedMode**:False - **dtype**:覆盖 fp32/fp16 ### moe_token_unpermute - **paddedMode**:False - **probsOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16 ### moe_token_unpermute_grad - **paddedMode**:False - **probsOptional**:覆盖 有/无 - **dtype**:覆盖 fp32/fp16 ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4472 ## 测试 <!--描述进行了哪些测试来验证你的改动。包括但不限于二级冒烟、算子泛化等。--> ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [x] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [x] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10344 | 2 months ago | |
Moe目录日志规范整改 Co-authored-by: guoqiuhao<guoqiuhao1@huawei.com> # message auto-generated for no-merge-commit merge: !10386 merge log into master Moe目录日志规范整改 Created-by: guoqiuhao Commit-by: guoqiuhao Merged-by: cann-robot Description: ## 描述 Moe目录日志规范整改. 1.禁止中文日志 2.英文拼写错误 3.日志级别错误 4.语法错误 5.禁止自定义打印 6.错误描述不完整 7.冗余日志 8.未知缩写 9.内部敏感词 10.缺少度量单位 ## 关联的Issue [#4476](https://gitcode.com/cann/ops-transformer/issues/4476) ## 测试 不涉及 ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10386 | 2 months ago | |
完成(moe / attention 外围)重名头文件整改 Co-authored-by: chenyifan<chenyifan66@h-partners.com> # message auto-generated for no-merge-commit merge: !10202 merge clear_pr3_pr4 into master 完成(moe / attention 外围)重名头文件整改 Created-by: mutex_lock Commit-by: chenyifan Merged-by: cann-robot Description: ## 描述 按仓内重名头文件规范,完成 **PR3 + PR4**( moe 纯组件 / attention 外围)整改,消除该范围内单仓同名头冲突。 **原因**:清单内 moe 与 attention 外围仍有大量同名头,违反「单仓内原则上不允许重名头文件」;易导致 include 误解析、维护成本上升。 **方法**:优先只改内部/次要副本(保留对外公开名与主路径原名),并同步更新 #include、\file 及必要 include guard。无业务逻辑与格式化变更。 ### 仓内分类规则 对每个重名头文件,先判断 **内容是否相同**、**是否对外/包间头文件**,再套用: | # | 场景 | 是否整改 | 处置 | |---|------|---------|------| | 1.1 | 内容不同,均为内部头文件 | 是 | 建议加模块前缀 / arch 后缀改名 | | 1.2 | 内容相同,均为内部头文件 | 是 | 只保留一份;**改 include 路径,禁止软链接**(本批对 arch 等同内容组改为 *_archXX.h 消歧,不删文件,避免相对 include 误解析) | | 1.3 | 部分内部、部分对外/包间 | 是 | **只改内部** 那份 | | 1.4 | 内容相同,均为对外/包间 | 否(兼容期) | 可用软链接保留多路径 | | 1.5 | 内容不同,均为对外/包间 | 兼容期内可暂缓 | 兼容期结束后改名 | | 1.6 | 其他 | 原则上不允许重名 | 特例需评审 | **原则**:除 1.4 外,单仓内原则上不允许重名头文件。 本分支实际用到:**1.1**、**1.2(同内容改 arch / 算子后缀,不删文件)**、**1.3**。 ### 范围与结果 | PR | 范围 | 组数(计划) | 本分支状态 | |---:|------|-------------:|------------| | 3 | moe 纯组件 | 48 | 完成 | | 4 | attention 外围(indexer / NSA 杂项 / mla-chunk / compressor 等) | 64 | 完成 | ### 命名约定 | 场景 | 约定 | |------|------| | arch 分档 | *_arch22.h / *_arch35.h / *_arch38.h | | 1.3 对外+包内 | 保留 op_api;包内 → *_kernel.h 或 *_archXX.h | | 跨算子同名 | 保留主算子原名;次要副本加算子/场景后缀 | | quant 与 routing 消歧(moe) | quant_v2/inner → *_quant_pkg_inner.h;quant arch35 → *_quant_arch35.h | ### 本分支变更统计 | 项 | 数量 | |----|-----:| | rename | 189 | | 修改(含 include / \file / guard 同步) | 72 | | **合计** | **261** | 主要目录(按变更文件数): | 目录 | 约计 | |------|-----:| | moe/moe_init_routing_quant_v2 | 31 | | attention/mla_prolog、attention/compressor | 各 25 | | moe/moe_init_routing_v3 | 16 | | attention/lightning_indexer_v2、quant_lightning_indexer_v2 | 各 11 | | 其余 moe init/permute、indexer、NSA、chunk、gating 等 | 余量 | ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4356 ## 测试 ## 文档更新 ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [x] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10202 | 2 months ago | |
docs(moe): fix routing-map token op A5 support and index dtype Co-authored-by: weixin_44156099<heyuhang20@h-partners.com> # message auto-generated for no-merge-commit merge: !10534 merge fix/moe-routing-map-docs into master docs(moe): fix routing-map token op A5 support and index dtype Created-by: weixin_44156099 Commit-by: weixin_44156099 Merged-by: cann-robot Description: ## 描述 修正 MoeTokenUnpermuteWithRoutingMap README 中输出参数 outIndex、permuteTokenId 的数据类型描述(FLOAT → INT32)。 ## 关联的Issue 关联 Issue #4544 ## 测试 文档更新,未涉及算子代码变更,无需编译/精度测试。 ## 文档更新 - 更新了上述四个算子目录下的 README.md 与 aclnn 接口文档。 ## 类型标签 - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [ ] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [x] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [ ] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10534 | 2 months ago | |
完成(moe / attention 外围)重名头文件整改 Co-authored-by: chenyifan<chenyifan66@h-partners.com> # message auto-generated for no-merge-commit merge: !10202 merge clear_pr3_pr4 into master 完成(moe / attention 外围)重名头文件整改 Created-by: mutex_lock Commit-by: chenyifan Merged-by: cann-robot Description: ## 描述 按仓内重名头文件规范,完成 **PR3 + PR4**( moe 纯组件 / attention 外围)整改,消除该范围内单仓同名头冲突。 **原因**:清单内 moe 与 attention 外围仍有大量同名头,违反「单仓内原则上不允许重名头文件」;易导致 include 误解析、维护成本上升。 **方法**:优先只改内部/次要副本(保留对外公开名与主路径原名),并同步更新 #include、\file 及必要 include guard。无业务逻辑与格式化变更。 ### 仓内分类规则 对每个重名头文件,先判断 **内容是否相同**、**是否对外/包间头文件**,再套用: | # | 场景 | 是否整改 | 处置 | |---|------|---------|------| | 1.1 | 内容不同,均为内部头文件 | 是 | 建议加模块前缀 / arch 后缀改名 | | 1.2 | 内容相同,均为内部头文件 | 是 | 只保留一份;**改 include 路径,禁止软链接**(本批对 arch 等同内容组改为 *_archXX.h 消歧,不删文件,避免相对 include 误解析) | | 1.3 | 部分内部、部分对外/包间 | 是 | **只改内部** 那份 | | 1.4 | 内容相同,均为对外/包间 | 否(兼容期) | 可用软链接保留多路径 | | 1.5 | 内容不同,均为对外/包间 | 兼容期内可暂缓 | 兼容期结束后改名 | | 1.6 | 其他 | 原则上不允许重名 | 特例需评审 | **原则**:除 1.4 外,单仓内原则上不允许重名头文件。 本分支实际用到:**1.1**、**1.2(同内容改 arch / 算子后缀,不删文件)**、**1.3**。 ### 范围与结果 | PR | 范围 | 组数(计划) | 本分支状态 | |---:|------|-------------:|------------| | 3 | moe 纯组件 | 48 | 完成 | | 4 | attention 外围(indexer / NSA 杂项 / mla-chunk / compressor 等) | 64 | 完成 | ### 命名约定 | 场景 | 约定 | |------|------| | arch 分档 | *_arch22.h / *_arch35.h / *_arch38.h | | 1.3 对外+包内 | 保留 op_api;包内 → *_kernel.h 或 *_archXX.h | | 跨算子同名 | 保留主算子原名;次要副本加算子/场景后缀 | | quant 与 routing 消歧(moe) | quant_v2/inner → *_quant_pkg_inner.h;quant arch35 → *_quant_arch35.h | ### 本分支变更统计 | 项 | 数量 | |----|-----:| | rename | 189 | | 修改(含 include / \file / guard 同步) | 72 | | **合计** | **261** | 主要目录(按变更文件数): | 目录 | 约计 | |------|-----:| | moe/moe_init_routing_quant_v2 | 31 | | attention/mla_prolog、attention/compressor | 各 25 | | moe/moe_init_routing_v3 | 16 | | attention/lightning_indexer_v2、quant_lightning_indexer_v2 | 各 11 | | 其余 moe init/permute、indexer、NSA、chunk、gating 等 | 余量 | ## 关联的Issue https://gitcode.com/cann/ops-transformer/issues/4356 ## 测试 ## 文档更新 ## 类型标签 <!-- [x] 表示选中 --> - [ ] 🐛 Bug修复 - [ ] ✨ 新特性 - [ ] ⚡ 性能优化 - [x] ♻️ 重构 - [ ] 🧪 测试 - [ ] 📦 构建/CI - [ ] 🔧 配置变更 - [ ] 📝 文档更新 - [ ] ⬆️ 依赖升级 - [ ] 🔒 安全修复 - [x] 🧹 代码清理 - [ ] ❓ 其他,请描述: See merge request: cann/ops-transformer!10202 | 2 months ago | |
更新license Co-authored-by: PerrySkywalker<wangmingkang1@huawei.com> # message auto-generated for no-merge-commit merge: !568 merge lic into master 更新license Created-by: PerrySkywalker Commit-by: PerrySkywalker Merged-by: cann-robot Description: ## 描述 更新license <!--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。--> ## 关联的Issue <!-- 如果这个PR是为了解决特定的Issue,请在这里提供Issue链接。例如:关联Issue #000--> <!-- 如果这个PR是为了解决特定的问题单,请在这里描述问题单单号。--> ## 测试 <!--描述进行了哪些测试来验证你的改动。包括但不限于二级冒烟、算子泛化等。--> ## 文档更新 <!--如果这个PR包含文档的更新,请在这里指出。例如:更新了README.md文件。--> ## 类型标签 <!-- [x] 表示选中 --> - [ ] Bug修复 - [ ] 新特性 - [ ] 性能优化 - [ ] 文档更新 - [ ] 其他,请描述: See merge request: cann/ops-transformer!568 | 10 months ago |
| Files | Last commit | Last update |
|---|---|---|
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 3 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 2 months ago | ||
| 10 months ago |