From dbdf5b0577d7c27b0712b2197444ef7ae263ac0c Mon Sep 17 00:00:00 2001 From: WHoutstanding Date: Sun, 18 Jan 2026 09:54:42 +0800 Subject: [PATCH 1/5] WHoutstanding add 2026.1.5~2024.1.18 WeeklyReport --- .../[WeeklyReport]2026.1.5~2026.1.18.md | 21 +++++++++++++++++++ 1 file changed, 21 insertions(+) create mode 100644 WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.1.5~2026.1.18.md diff --git a/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.1.5~2026.1.18.md b/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.1.5~2026.1.18.md new file mode 100644 index 00000000..819a1c55 --- /dev/null +++ b/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.1.5~2026.1.18.md @@ -0,0 +1,21 @@ +### 姓名 +王豪 + +### 实习项目 +GraphNet 计算子图构建与推广 + +### 本周工作 + +1. 修复dtype_generalizer.py的bug：在判断计算图need_rewrite之前， parse_immutable_model_path_into_sole_graph_module()返回的trace_model不包含tensor_meta导致dtype_pass.need_rewrite(traced_model)返回总是false，在need_rewrite之前为trace_model加入tensor_meta，因此调用ShapeProp(traced_model).propagate(*inputs) + +2. 将dtype_gen_test.sh改为标准脚本格式并对samll100_samples.txt的进行dtype generalizaion pass测试，并用 graph_net.torch.validate 验证生成的 samples of dtype generalization pass + +3. 把dtype_generalizer.py做成了SamplePass的子类并放到torch/sample_pass目录下。 + + +### 下周工作 + +1. 调试分析未通过的sample的原因 +2. 开一个关于数据类型泛化的讨论区，并讨论如何集成进加工流程。 + +### 导师点评 From 08959b7560b7883ee9c17727d3ee0cd9bd77a054 Mon Sep 17 00:00:00 2001 From: WHoutstanding Date: Sun, 25 Jan 2026 13:44:19 +0800 Subject: [PATCH 2/5] fix WeekyReport --- .../WHoutstanding/[WeeklyReport]2026.1.5~2026.1.18.md | 1 + 1 file changed, 1 insertion(+) diff --git a/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.1.5~2026.1.18.md b/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.1.5~2026.1.18.md index 819a1c55..c473ebfe 100644 --- a/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.1.5~2026.1.18.md +++ b/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.1.5~2026.1.18.md @@ -19,3 +19,4 @@ GraphNet 计算子图构建与推广 2. 开一个关于数据类型泛化的讨论区，并讨论如何集成进加工流程。 ### 导师点评 +true From d10625ecde1edc3cd6ee65c093b9a04e7bae7a7a Mon Sep 17 00:00:00 2001 From: WHoutstanding Date: Sun, 1 Feb 2026 23:53:22 +0800 Subject: [PATCH 3/5] [WeeklyReport] WHoutstanding 2026.01.19~2026.02.01 --- .../[WeeklyReport]2026.01.19~2026.02.01.md | 17 +++++++++++++++++ 1 file changed, 17 insertions(+) create mode 100644 WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.01.19~2026.02.01.md diff --git a/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.01.19~2026.02.01.md b/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.01.19~2026.02.01.md new file mode 100644 index 00000000..a826ffd1 --- /dev/null +++ b/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.01.19~2026.02.01.md @@ -0,0 +1,17 @@ +### 姓名 +王豪 + +### 实习项目 +GraphNet 计算子图构建与推广 + +### 本周工作 + +1. 处理dtype泛化失败的计算图，修改dtype转化的代码逻辑是计算图dtype成功转化率提高至95% + + +### 下周工作 + +集成类型泛化的流程, 集成在子图和整图上。 + +### 导师点评 +true From 353ed6b672fd52848721664da0de7cc8f0be0a0a Mon Sep 17 00:00:00 2001 From: WHoutstanding Date: Sun, 1 Mar 2026 15:12:16 +0800 Subject: [PATCH 4/5] WHoutstanding add 2026.02.02~2026.02.15 WeeklyReport --- .../[WeeklyReport]2026.2.2~2026.2.15.md | 18 ++++++++++++++++++ 1 file changed, 18 insertions(+) create mode 100644 WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.2.2~2026.2.15.md diff --git a/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.2.2~2026.2.15.md b/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.2.2~2026.2.15.md new file mode 100644 index 00000000..b3488c07 --- /dev/null +++ b/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.2.2~2026.2.15.md @@ -0,0 +1,18 @@ +### 姓名 +王豪 + +### 实习项目 +GraphNet 计算子图构建与推广 + +### 本周工作 + +1. GraphNet的samples的计算图存在unstable api，使用Paconvert获取未转换的api，并通过unstableTostable backend将计算图的unstable api转换为stabel api。 +2. 用paconvert转换100个torch sample，成功率100的Sample有66个。 + + +### 下周工作 + +1. 使用LLM解决badcase + +### 导师点评 + \ No newline at end of file From 4ddac2f855b05c36d894d239fbe64f614d29513d Mon Sep 17 00:00:00 2001 From: WHoutstanding Date: Sun, 8 Mar 2026 21:32:14 +0800 Subject: [PATCH 5/5] WHoutstanding add 2026.02.23~2026.03.08 WeeklyReport --- .../[WeeklyReport]2026.2.23~2026.3.8.md | 17 +++++++++++++++++ 1 file changed, 17 insertions(+) create mode 100644 WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.2.23~2026.3.8.md diff --git a/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.2.23~2026.3.8.md b/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.2.23~2026.3.8.md new file mode 100644 index 00000000..f8aeaaef --- /dev/null +++ b/WeeklyReports/hackathon_10th/WHoutstanding/[WeeklyReport]2026.2.23~2026.3.8.md @@ -0,0 +1,17 @@ +### 姓名 +王豪 + +### 实习项目 +GraphNet 计算子图构建与推广 + +### 本周工作 + +1. 完成准备 GraphNet中 PaConvert 的bad case, 探索使用各家 LLM 来修复 PaConvert的 bug。 +2. 分析各家 LLM 对嫌疑代码的通透理解程度，并作为隐变量，估算 LLMs 是否有能力解决该 BUG。 + +### 下周工作 + +探索编译机床：直接驱动 LLMs 来解 目标软件包的 bug； + +### 导师点评 + \ No newline at end of file