Merge branch 'develop' of https://github.com/HeyDavid633/GraphNet into add_STFPM

Author: HeyDavid633

graph_net/torch/extractor.py

@@ -1,6 +1,7 @@
 import os
 import torch
 import json
+import shutil
 from typing import Union, Callable
 from . import utils
 
@@ -80,76 +81,105 @@ def forward(self, s0 : torch.SymInt, L_x_ : torch.Tensor):
     def wrapper(model: torch.nn.Module):
         assert isinstance(model, torch.nn.Module), f"{type(model)=}"
 
-        def extractor(gm: torch.fx.GraphModule, sample_inputs):
-            # 1. Get workspace path
-            workspace_path = os.environ.get("GRAPH_NET_EXTRACT_WORKSPACE")
-            if not workspace_path:
-                raise EnvironmentError(
-                    "Environment variable 'GRAPH_NET_EXTRACT_WORKSPACE' is not set."
+        class GraphExtractor:
+            def __init__(self):
+                self.subgraph_counter = 0
+
+            def move_files(self, source_dir, target_dir):
+                os.makedirs(target_dir, exist_ok=True)
+                for item in os.listdir(source_dir):
+                    source_path = os.path.join(source_dir, item)
+                    if os.path.isfile(source_path):
+                        target_path = os.path.join(target_dir, item)
+                        shutil.move(source_path, target_path)
+
+            def __call__(self, gm: torch.fx.GraphModule, sample_inputs):
+                # 1. Get workspace path
+                workspace_path = os.environ.get("GRAPH_NET_EXTRACT_WORKSPACE")
+                if not workspace_path:
+                    raise EnvironmentError(
+                        "Environment variable 'GRAPH_NET_EXTRACT_WORKSPACE' is not set."
+                    )
+                model_path = os.path.join(workspace_path, name)
+                os.makedirs(model_path, exist_ok=True)
+
+                if self.subgraph_counter == 0:
+                    subgraph_path = model_path
+                else:
+                    if self.subgraph_counter == 1:
+                        subgraph_0_path = os.path.join(model_path, f"subgraph_0")
+                        self.move_files(model_path, subgraph_0_path)
+
+                    subgraph_path = os.path.join(
+                        model_path, f"subgraph_{self.subgraph_counter}"
+                    )
+                    os.makedirs(subgraph_path, exist_ok=True)
+
+                self.subgraph_counter += 1
+
+                # 2. Get full params
+                params = {}
+                input_idx = 0
+                unique_id = 0
+
+                def try_rename_placeholder(node):
+                    assert node.op == "placeholder"
+                    if not placeholder_auto_rename:
+                        return
+                    nonlocal unique_id
+                    node.target = f"v{unique_id}"
+                    unique_id += 1
+                    node.name = f"v{unique_id}"
+                    unique_id += 1
+
+                for node in gm.graph.nodes:
+                    if node.op == "placeholder":
+                        try_rename_placeholder(node)
+                        input = sample_inputs[input_idx]
+                        if isinstance(input, torch.SymInt):
+                            input = torch.tensor(4)
+                        params[node.target] = input
+                        input_idx += 1
+                assert input_idx == len(sample_inputs)
+                if mut_graph_codes is not None:
+                    assert isinstance(mut_graph_codes, list)
+                    mut_graph_codes.append(gm.code)
+                # 3. Generate and save model code
+                base_code = gm.code
+                # gm.graph.print_tabular()
+                write_code = utils.apply_templates(base_code)
+                with open(os.path.join(subgraph_path, "model.py"), "w") as fp:
+                    fp.write(write_code)
+
+                # 4. Save metadata
+                metadata = {
+                    "framework": "torch",
+                    "num_devices_required": 1,
+                    "num_nodes_required": 1,
+                    "dynamic": bool(dynamic),
+                    "model_name": name,
+                }
+                with open(os.path.join(subgraph_path, "graph_net.json"), "w") as f:
+                    json.dump(metadata, f, indent=4)
+
+                # 5. Save tensor metadata
+                # Adapt to different input structures (e.g., single tensor vs. dict/tuple of tensors)
+                converted = utils.convert_state_and_inputs(params, [])
+                utils.save_converted_to_text(converted, file_path=subgraph_path)
+                utils.save_constraints_text(
+                    converted,
+                    file_path=os.path.join(
+                        subgraph_path, "input_tensor_constraints.py"
+                    ),
                 )
-            model_path = os.path.join(workspace_path, name)
-            os.makedirs(model_path, exist_ok=True)
-
-            # 2. Get full params
-            params = {}
-            input_idx = 0
-            unique_id = 0
-
-            def try_rename_placeholder(node):
-                assert node.op == "placeholder"
-                if not placeholder_auto_rename:
-                    return
-                nonlocal unique_id
-                node.target = f"v{unique_id}"
-                unique_id += 1
-                node.name = f"v{unique_id}"
-                unique_id += 1
-
-            for node in gm.graph.nodes:
-                if node.op == "placeholder":
-                    try_rename_placeholder(node)
-                    input = sample_inputs[input_idx]
-                    if isinstance(input, torch.SymInt):
-                        input = torch.tensor(4)
-                    params[node.target] = input
-                    input_idx += 1
-            assert input_idx == len(sample_inputs)
-            if mut_graph_codes is not None:
-                assert isinstance(mut_graph_codes, list)
-                mut_graph_codes.append(gm.code)
-            # 3. Generate and save model code
-            base_code = gm.code
-            # gm.graph.print_tabular()
-            write_code = utils.apply_templates(base_code)
-            with open(os.path.join(model_path, "model.py"), "w") as fp:
-                fp.write(write_code)
-
-            # 4. Save metadata
-            metadata = {
-                "framework": "torch",
-                "num_devices_required": 1,
-                "num_nodes_required": 1,
-                "dynamic": bool(dynamic),
-                "model_name": name,
-            }
-            with open(os.path.join(model_path, "graph_net.json"), "w") as f:
-                json.dump(metadata, f, indent=4)
-
-            # 5. Save tensor metadata
-            # Adapt to different input structures (e.g., single tensor vs. dict/tuple of tensors)
-            converted = utils.convert_state_and_inputs(params, [])
-            utils.save_converted_to_text(converted, file_path=model_path)
-            utils.save_constraints_text(
-                converted,
-                file_path=os.path.join(model_path, "input_tensor_constraints.py"),
-            )
 
-            print(
-                f"Graph and tensors for '{name}' extracted successfully to: {model_path}"
-            )
+                print(
+                    f"Graph and tensors for '{name}' extracted successfully to: {model_path}"
+                )
 
-            return gm.forward
+                return gm.forward
 
+        extractor = GraphExtractor()
         # return torch.compile(backend=extractor, dynamic=dynamic)
         compiled_model = torch.compile(model, backend=extractor, dynamic=dynamic)
 

graph_net/torch/validate.py

@@ -15,8 +15,7 @@ def temp_workspace():
         os.environ["GRAPH_NET_EXTRACT_WORKSPACE"] = old
 
 
-def main(args):
-    model_path = args.model_path
+def validate(args, model_path):
     with temp_workspace() as tmp_dir_name:
         print("Check extractability ...")
         cmd = f"{sys.executable} -m graph_net.torch.single_device_runner --model-path {model_path}"
@@ -36,16 +35,46 @@ def main(args):
                 if args.graph_net_samples_path is None
                 else args.graph_net_samples_path
             )
-            cmd = f"{sys.executable} -m graph_net.torch.check_redundant_incrementally --model-path {args.model_path} --graph-net-samples-path {graph_net_samples_path}"
+            cmd = f"{sys.executable} -m graph_net.torch.check_redundant_incrementally --model-path {model_path} --graph-net-samples-path {graph_net_samples_path}"
             cmd_ret = os.system(cmd)
-            rm_cmd = f"{sys.executable} -m graph_net.torch.remove_redundant_incrementally --model-path {args.model_path} --graph-net-samples-path {graph_net_samples_path}"
+            rm_cmd = f"{sys.executable} -m graph_net.torch.remove_redundant_incrementally --model-path {model_path} --graph-net-samples-path {graph_net_samples_path}"
             assert (
                 cmd_ret == 0
             ), f"\nPlease use the following command to remove redundant model directories:\n\n{rm_cmd}\n"
 
         print(f"Validation success, {model_path=}")
 
 
+def get_recursively_model_path(root_dir):
+    for sub_dir in get_immediate_subdirectory_paths(root_dir):
+        if is_single_model_dir(sub_dir):
+            yield sub_dir
+        else:
+            yield from get_recursively_model_path(sub_dir)
+
+
+def get_immediate_subdirectory_paths(parent_dir):
+    return [
+        sub_dir
+        for name in os.listdir(parent_dir)
+        for sub_dir in [os.path.join(parent_dir, name)]
+        if os.path.isdir(sub_dir)
+    ]
+
+
+def is_single_model_dir(model_dir):
+    return os.path.isfile(f"{model_dir}/graph_net.json")
+
+
+def main(args):
+    model_path = args.model_path
+    if is_single_model_dir(args.model_path):
+        validate(args, model_path)
+    else:
+        for model_path in get_recursively_model_path(args.model_path):
+            validate(args, model_path)
+
+
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(
         description="Validate a computation graph sample. return 0 if success"

paddle_samples/PaddleX/CLIP_vit_base_patch16_448_ML/subgraph_0/graph_hash.txt

@@ -0,0 +1 @@
+94f68a32a189a13a1025d0e8a1d6e6269b722b9018d7342b47c28793031842cc

paddle_samples/PaddleX/CLIP_vit_base_patch16_448_ML/subgraph_0/graph_net.json

@@ -0,0 +1,6 @@
+{
+    "framework": "paddle",
+    "model_name": "CLIP_vit_base_patch16_448_ML",
+    "num_devices_required": 1,
+    "num_nodes_required": 1
+}

paddle_samples/PaddleX/CLIP_vit_base_patch16_448_ML/subgraph_0/input_meta.py

@@ -0,0 +1,49 @@
+class Program_weight_tensor_data_0:
+    name = "data_0"
+    shape = [1, 1, 768]
+    dtype = "float32"
+    min_val = float("-1.57892")
+    max_val = float("0.974002")
+    mean = float("0.0033194")
+    std = float("0.219507")
+    data = None
+
+
+class Program_weight_tensor_data_1:
+    name = "data_1"
+    shape = [1, 197, 768]
+    dtype = "float32"
+    min_val = float("-0.226016")
+    max_val = float("0.248632")
+    mean = float("-0.00788177")
+    std = float("0.0294678")
+    data = None
+
+
+class Program_weight_tensor_data_2:
+    name = "data_2"
+    shape = [33, 768]
+    dtype = "float32"
+    min_val = float("-0.0865287")
+    max_val = float("0.0865482")
+    mean = float("-0.000143484")
+    std = float("0.0501765")
+    data = None
+
+
+class Program_weight_tensor_data_3:
+    name = "data_3"
+    shape = []
+    dtype = "int64"
+    data = [7]
+
+
+class Program_weight_tensor_data_4:
+    name = "data_4"
+    shape = [7, 3, 448, 448]
+    dtype = "float32"
+    min_val = float("-2.1179")
+    max_val = float("2.64")
+    mean = float("-0.154463")
+    std = float("1.24419")
+    data = None