From 6de71ded06a7967ba1f0176c31e6e822037bc13b Mon Sep 17 00:00:00 2001 From: zeroRains Date: Sat, 27 Jul 2024 13:46:38 +0800 Subject: [PATCH 1/8] add weekly report --- .../[WeeklyReports]2024.05.11~2024.05.24.md | 111 ++++++++++++++++++ 1 file changed, 111 insertions(+) create mode 100644 WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.05.11~2024.05.24.md diff --git a/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.05.11~2024.05.24.md b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.05.11~2024.05.24.md new file mode 100644 index 00000000..66c65e98 --- /dev/null +++ b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.05.11~2024.05.24.md @@ -0,0 +1,111 @@ +### 姓名 + +卢林军 + +### 实习项目 + +组合机制建设和机制推全 + +### 本周工作 + +本项目的主要工作是对尚未支持组合机制的算子添加组合机制并完善机制,本周主要工作如下: + +1. 为sum和mean op的反向拆解添加动态shape支持 + + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/64789 + +2. 完善Reduce_as 算子计算方式 + +之前reduce_as op的实现是在计算好reduce_dim之后将reduce_dim作为输入传给reduce_sum kernel。当reduce_dim为空数组时,reduce_sum会默认执行reduce_all的计算,但是在reduce_as op中,当reduce_dim为空时,期望是不对输入做任何操作。 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/65002 + +3. 为add, subtract, multiply和 divide op的反向拆解添加动态shape支持 + +同时修复了之前写法导致算子性能下降的BUG,修复multiply_grad的反向拆解在一个输入动态shape另一个输入是静态shape的场景下出现的bug,同时为同类算子新增对应的单测 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/65003 +- https://github.com/PaddlePaddle/Paddle/pull/65005 +- https://github.com/PaddlePaddle/Paddle/pull/65007 +- https://github.com/PaddlePaddle/Paddle/pull/65006 +- https://github.com/PaddlePaddle/Paddle/pull/65357 +- https://github.com/PaddlePaddle/Paddle/pull/65643 + +5. 为concat_grad添加动态shape支持,添加split_grad动态shape的单测 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/65148 + +6. 为relu_grad的反向拆解过程支持动态shape,添加relu_grad和sigmoid_grad的动态shape单测 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/65482 +- https://github.com/PaddlePaddle/Paddle/pull/65832 + +7. 完善get_reduce_dims_from_out函数功能 + +目前发现get_reduce_dims_from_out在某些动态shape场景下也能正常工作,但是函数设计的本身目的是用于处理静态shape的场景,因此在该函数中添加了动态shap的检测 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/65666 + +8. 为elementwise_pow_grad添加动态shape支持,添加pow_grad动态shape的单测 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/65692 + +9. 为softmax_grad添加动态shape支持 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/65961 + +10. 前向拆解lerp op并添加动态shape支持 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/65967 + +11. 前向拆解log_loss op并添加动态shape支持 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/65968 + +12. 为动态shape场景,添加GetOutputDimsForDynamicShape函数 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/66172 + +13. 修复stack op在旧IR组合机制下的BUG + +当Tensor不开启梯度计算时,在旧IR下梯度初始化似乎不是初始化为空值,而是初始化成一个维度为[]的Tensor,这旧导致了其会进入到反向计算prim拆解的流程中,而其中涉及到了一个对梯度(维度为[])reshape操作,这就导致了这个测试失败。因此需要在初始化变量时开启梯度计算 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/66457 + +14. 前向拆解kldiv_loss op,并支持动态shape + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/66510 + +### 下周工作 + +1. 收集待拆解算子并为其支持组合机制 + +### 导师点评 + From 1402b21b07740bdb5bd96914b34a75e48c9fe6a8 Mon Sep 17 00:00:00 2001 From: zeroRains Date: Sat, 27 Jul 2024 13:51:31 +0800 Subject: [PATCH 2/8] modify the file name --- ...5.11~2024.05.24.md => [WeeklyReports]2024.07.15~2024.07.28.md} | 0 1 file changed, 0 insertions(+), 0 deletions(-) rename WeeklyReports/Hackathon_7th/04_zeroRains/{[WeeklyReports]2024.05.11~2024.05.24.md => [WeeklyReports]2024.07.15~2024.07.28.md} (100%) diff --git a/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.05.11~2024.05.24.md b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.15~2024.07.28.md similarity index 100% rename from WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.05.11~2024.05.24.md rename to WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.15~2024.07.28.md From a3c3536e66c87b6fe0aa483e009d52ee5f2cad77 Mon Sep 17 00:00:00 2001 From: cyber-pioneer <116002591+cyber-pioneer@users.noreply.github.com> Date: Mon, 29 Jul 2024 10:34:45 +0800 Subject: [PATCH 3/8] Update WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.15~2024.07.28.md --- .../04_zeroRains/[WeeklyReports]2024.07.15~2024.07.28.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.15~2024.07.28.md b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.15~2024.07.28.md index 66c65e98..2d4fe747 100644 --- a/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.15~2024.07.28.md +++ b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.15~2024.07.28.md @@ -108,4 +108,4 @@ 1. 收集待拆解算子并为其支持组合机制 ### 导师点评 - +靠谱,高效 From 03c15999f8c80d59ea7b1ca556d2e15a0e66ef16 Mon Sep 17 00:00:00 2001 From: cyber-pioneer <116002591+cyber-pioneer@users.noreply.github.com> Date: Mon, 12 Aug 2024 21:37:56 +0800 Subject: [PATCH 4/8] Update WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.29~2024.08.11.md --- .../04_zeroRains/[WeeklyReports]2024.07.29~2024.08.11.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.29~2024.08.11.md b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.29~2024.08.11.md index 8cab5c80..56334870 100644 --- a/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.29~2024.08.11.md +++ b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.29~2024.08.11.md @@ -41,4 +41,4 @@ ### 导师点评 - +Nice From 82a34c2c0d6eb9e25a67309fd993ffddc6b26df4 Mon Sep 17 00:00:00 2001 From: zeroRains Date: Mon, 26 Aug 2024 16:33:34 +0800 Subject: [PATCH 5/8] add weeklyreports --- .../[WeeklyReports]2024.07.29~2024.08.11.md | 4 -- .../[WeeklyReports]2024.08.12~2024.08.25.md | 44 +++++++++++++++++++ 2 files changed, 44 insertions(+), 4 deletions(-) create mode 100644 WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.12~2024.08.25.md diff --git a/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.29~2024.08.11.md b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.29~2024.08.11.md index cf0876a3..56334870 100644 --- a/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.29~2024.08.11.md +++ b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.07.29~2024.08.11.md @@ -41,8 +41,4 @@ ### 导师点评 -<<<<<<< HEAD - -======= Nice ->>>>>>> 5f903d04cee7dd6955f0ed5b1f968da18d2c5e6c diff --git a/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.12~2024.08.25.md b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.12~2024.08.25.md new file mode 100644 index 00000000..13a85a20 --- /dev/null +++ b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.12~2024.08.25.md @@ -0,0 +1,44 @@ +### 姓名 + +卢林军 + +### 实习项目 + +组合机制建设和机制推全 + +### 本周工作 + +本项目的主要工作是对尚未支持组合机制的算子添加组合机制并完善机制,本周主要工作如下: + +1. 为expand_grad op 添加动态shape支持 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/67481 + +2. 为stack_grad op 添加动态shape支持 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/67502 + +3. 为pad_grad 添加动态shape支持 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/67606 + +4. 为scale_grad, square_grad, transpose_grad, swiglu_grad添加动态shape单测,并按照字母序重新整理单测 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/67590 + + +### 下周工作 + +1. batch_norm,prod,dropout反向适配动态shape + +### 导师点评 + + From 7e785c05645c8a358b32bffbcd72abec89a7b51b Mon Sep 17 00:00:00 2001 From: Zero Rains Date: Wed, 28 Aug 2024 17:08:41 +0800 Subject: [PATCH 6/8] Update [WeeklyReports]2024.08.12~2024.08.25.md --- .../04_zeroRains/[WeeklyReports]2024.08.12~2024.08.25.md | 1 + 1 file changed, 1 insertion(+) diff --git a/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.12~2024.08.25.md b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.12~2024.08.25.md index 13a85a20..86644e40 100644 --- a/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.12~2024.08.25.md +++ b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.12~2024.08.25.md @@ -41,4 +41,5 @@ ### 导师点评 +工作完成度高,独立思考并解决问题,后续可尝试独立规划每周任务 From b142a3b1382eea0560006d7483337e793e7e4dc7 Mon Sep 17 00:00:00 2001 From: zerorains Date: Sun, 15 Sep 2024 14:38:35 +0000 Subject: [PATCH 7/8] add weeklyreport --- .../[WeeklyReports]2024.08.26~2024.9.8.md | 62 +++++++++++++++++++ 1 file changed, 62 insertions(+) create mode 100644 WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.26~2024.9.8.md diff --git a/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.26~2024.9.8.md b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.26~2024.9.8.md new file mode 100644 index 00000000..d7c70604 --- /dev/null +++ b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.26~2024.9.8.md @@ -0,0 +1,62 @@ +### 姓名 + +卢林军 + +### 实习项目 + +组合机制建设和机制推全 + +### 本周工作 + +本项目的主要工作是对尚未支持组合机制的算子添加组合机制并完善机制,本周主要工作如下: + +1. 为dropout_grad op 添加动态shape支持 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/67757 + +2. 为prod_grad op 添加动态shape支持 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/67775 + +3. 为cumprod_grad 支持动态shape + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/67965 + +4. 为gather_grad, gather_nd_grad 支持动态shape + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/67968 + +5. 为scatter_grad 添加动态shape支持,添加scatter_nd_add_grad的动态shape单测,为所有反向动态shape测试添加拆解检测 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/68168 + +6. 为topk_grad添加动态shape支持 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/68228 + +7. 为unsqueeze_grad, squeeze_grad 添加动态shape单测 + +相关 PR: + +- https://github.com/PaddlePaddle/Paddle/pull/68076 + +### 下周工作 + +1. group_norm_grad, layer_norm_grad, masked_select_grad, roll_grad反向适配动态shape +2. one_hot, batch_norm_, batch_norm, bmm 前向适配动态shape + +### 导师点评 + + From 012da25c7cca45962472fd83f103dd1642c5d046 Mon Sep 17 00:00:00 2001 From: Zero Rains Date: Mon, 30 Sep 2024 15:21:46 +0800 Subject: [PATCH 8/8] Update WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.26~2024.9.8.md Co-authored-by: xiaoguoguo626807 <100397923+xiaoguoguo626807@users.noreply.github.com> --- .../04_zeroRains/[WeeklyReports]2024.08.26~2024.9.8.md | 1 + 1 file changed, 1 insertion(+) diff --git a/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.26~2024.9.8.md b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.26~2024.9.8.md index d7c70604..6dbba437 100644 --- a/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.26~2024.9.8.md +++ b/WeeklyReports/Hackathon_7th/04_zeroRains/[WeeklyReports]2024.08.26~2024.9.8.md @@ -58,5 +58,6 @@ 2. one_hot, batch_norm_, batch_norm, bmm 前向适配动态shape ### 导师点评 +按时完成工作任务,效率很高