Merge branch 'develop' of github.com:PaddlePaddle/GraphNet into develop

Author: lixinqi

.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,6 @@
+### PR Category
+<!-- One of [ New Sample | Feature Enhancement | Bug Fix | Other ] -->
+
+
+### Description
+<!-- Describe what you’ve done -->

CONTRIBUTE_TUTORIAL.md

@@ -48,7 +48,7 @@ from graph_net.torch.extractor import extract
 
 def run_model(name: str, device_str: str) -> None:
     """
-    Run computation graph extraction for the specified model.
+    Run computational graph extraction for the specified model.
 
     Args:
         name (str): Model name (e.g., 'resnet50', 'vit_b_16', 'bert-base-uncased').
@@ -222,14 +222,14 @@ Re-run `extract(name)(model)(input_data)` to complete extraction.
   * Insert hooks at specific layers by passing them into `wrapped = extract(...)(model, hooks=...)`.
 
 
-### 3. Extracting the Computation Graph
+### 3. Extracting the Computational Graph
 
 1. **Run extract**
 
 Execute scripts with `@graph_net.torch.extract` or `@graph_net.paddle.extract`. For example:
 
 ```bash
-# Extract the ResNet‑18 computation graph
+# Extract the ResNet‑18 computational graph
 python -m graph_net.test.vision_model_test
 ```
 
@@ -241,7 +241,7 @@ Expected output will be saved under `$GRAPH_NET_EXTRACT_WORKSPACE`.
 python -m graph_net.torch.validate --model-path $GRAPH_NET_EXTRACT_WORKSPACE/model_name
 ```
 
-`validate` checks if the extracted graph meets the Dataset Construction Constraints. If success, you’re ready to submit.
+`validate` checks if the extracted graph meets the Dataset Construction Constraints. If success, you’re ready to continue.
 
 
 ### 4. Submitting the Extracted Graph
@@ -263,16 +263,17 @@ python -m graph_net.pack --output /path/to/output.zip --clear-after-pack True
 
 This API:
 
-a. Packages all files under `$GRAPH_NET_EXTRACT_WORKSPACE` into `/path/to/output.zip`
+a. Packages all files under `$GRAPH_NET_EXTRACT_WORKSPACE` into `/path/to/output.zip` (You can set it to `GraphNet/samples`)
 
 b. Clears the workspace if `--clear-after-pack` is `True`
 
-Note: If third-party ops are used, contributors must include them manually in the graph directory. As long as `validate` passes, no specific folder structure is required.
+Note: If third-party ops are used, contributors must include them manually in the package. As long as `validate` passes, no specific folder structure is required.
 
 3. **Commit the changes**
 
+Move the packaged computational graph in the previous step to **samples** directory and commit.
 ```bash
-git add <new files>
+git add <the packaged computational graph>
 git commit -m "Description"
 ```
 

CONTRIBUTE_TUTORIAL_cn.md

@@ -234,7 +234,7 @@ python -m graph_net.test.vision_model_test
 ```bash
 python -m graph_net.torch.validate --model-path $GRAPH_NET_EXTRACT_WORKSPACE/model_name
 ```
-`validate` 验证您刚刚抽取的计算图符合Dataset Construction Constraints，如果结果为Success，则可以提交。
+`validate` 验证您刚刚抽取的计算图符合Dataset Construction Constraints，如果结果为Success，则可以继续。
 
 
 
@@ -254,16 +254,17 @@ python -m graph_net.pack --output /path/to/output.zip --clear-after-pack True
 ```
 该API的功能为：
 
- a. 打包`$GRAPH_NET_EXTRACT_WORKSPACE`下的所有文件到`/path/to/output.zip`
+ a. 打包`$GRAPH_NET_EXTRACT_WORKSPACE`下的所有文件到`/path/to/output.zip` （可以设置到`GraphNet/samples`）
 
  b. 若`--clear-after-pack`为`True`，则打包后清空`$GRAPH_NET_EXTRACT_WORKSPACE`
 
-请注意，如果有第三方算子，需要贡献者自行打包到计算图目录内。目前没有特别规定存放的目录结构，但只要通过了validate环节，就可以达到验收标准。
+请注意，如果有第三方算子，需要贡献者自行打包到计算图压缩包内。目前没有特别规定存放的目录结构，但只要通过了validate环节，就可以达到验收标准。
 
 3. **提交修改**
 
+移动上一步打包完成的计算图压缩包到**samples**目录，然后提交。
 ```bash
-git add <新增的文件>
+git add <计算图压缩包>
 git commit -m "描述"
 ```
 4. **推送分支到远程**（你的 Fork 仓库）

README.md

@@ -8,7 +8,7 @@ With GraphNet, users can:
 3. Train AI‑for‑Systems models to automatically generate compiler optimization passes.
 
 
-Dataset Construction Constraints：
+### Dataset Construction Constraints：
 1. Dynamic graphs must execute correctly.
 2. Each computation graph should include a standardized method for measuring performance.
 3. Graphs and their corresponding Python code must support serialization and deserialization.
@@ -91,10 +91,10 @@ Once you have packaged these extracted computation graphs, submit them to the Gr
 <table>
 <tr>
 <td align="center">
-    <img width="190" height="220" src="https://github.com/user-attachments/assets/31b4f0ba-417e-48b6-a860-124d74bd6643" />
+    <img width="200" src="https://github.com/user-attachments/assets/30f034dd-f7d9-49f5-bae8-30ba2ac9c6b4" />
 </td>
 <td align="center">
-    <img width="190" height="220" src="https://github.com/user-attachments/assets/140fa03e-36ef-44bf-8d9a-ca65c83b0139" />
+    <img width="200" src="https://github.com/user-attachments/assets/140fa03e-36ef-44bf-8d9a-ca65c83b0139" />
 </td>
 </tr>
 </table>

README_cn.md

@@ -8,7 +8,7 @@ GraphNet —— 一个面向编译器开发的大规模数据集，旨在为研
 2. 方便已有编译器做回归测试
 3. 训练AI-for-system模型以自动生成编译器优化Pass
 
-数据集构建约束：
+### 数据集构建约束：
 
 1. 动态图能正常运行
 2. 每份计算图有通用方法测定性能指标
@@ -90,7 +90,7 @@ python -m graph_net.config --global\
 <table>
 <tr>
 <td align="center">
-    <img width="190" height="220" src="https://github.com/user-attachments/assets/31b4f0ba-417e-48b6-a860-124d74bd6643" />
+    <img width="190" height="220" src="https://github.com/user-attachments/assets/1a42cceb-f026-44a6-acbe-dee810410893" />
 </td>
 <td align="center">
     <img width="190" height="220" src="https://github.com/user-attachments/assets/140fa03e-36ef-44bf-8d9a-ca65c83b0139" />
@@ -100,4 +100,6 @@ python -m graph_net.config --global\
 </div>
 
 ## 开源协议
-[MIT License](LICENSE)
+[MIT License](LICENSE)
+
+

graph_net/test/bert_model_test.py

@@ -24,7 +24,7 @@ def create_model():
     inputs = {k: v.to(device) for k, v in inputs.items()}
 
     model = create_model()
-    model = graph_net.torch.extract(name=get_model_name())(model)
+    model = graph_net.torch.extract(name=get_model_name(), dynamic=True)(model)
 
     print("Running inference...")
     output = model(**inputs)

graph_net/test/vision_model_test.py

@@ -28,7 +28,7 @@
     model.to(device)
     normalized_input = normalized_input.to(device)
 
-    model = graph_net.torch.extract(name="resnet18")(model)
+    model = graph_net.torch.extract(name="resnet18", dynamic=True)(model)
 
     print("Running inference...")
     print("Input shape:", normalized_input.shape)

graph_net/torch/utils.py

@@ -8,17 +8,23 @@
 import argparse
 import importlib
 import inspect
+import math
 
 
 def apply_templates(forward_code: str) -> str:
     tab = "    "
     forward_code = f"\n{tab}".join(forward_code.split("\n"))
-    return f"import torch\n\nclass GraphModule(torch.nn.Module):\n{tab}{forward_code}"
+    imports = "import torch"
+    if "device" in forward_code:
+        imports += "\n\nfrom torch import device"
+    return f"{imports}\n\nclass GraphModule(torch.nn.Module):\n{tab}{forward_code}"
 
 
 def get_limited_precision_float_str(value):
     if not isinstance(value, float):
         return value
+    if not math.isfinite(value):
+        return f'float("{value}")'
     return f"{value:.3f}"
 
 

samples/timm/aimv2_1b_patch14_224.apple_pt/graph_hash.txt

@@ -0,0 +1 @@
+ff10075e46eeae7615c783a3970e04ae79c1ea3f5b6020b167bec1e19489372f

samples/timm/aimv2_1b_patch14_224.apple_pt/graph_net.json

@@ -0,0 +1,5 @@
+{
+    "framework": "torch",
+    "num_devices_required": 1,
+    "num_nodes_required": 1
+}