add unstable_to_stable backend and check_unstable_api

Author: RbRe145

graph_net/torch/backend/unstable_to_stable_backend.py

@@ -0,0 +1,65 @@
+import os
+import torch
+from .graph_compiler_backend import GraphCompilerBackend
+
+
+class UnstableToStableBackend(GraphCompilerBackend):
+    def __call__(self, model):
+        # Perform unstable API check before running the model
+        self.model = model
+        self.unstable_to_stable()
+        self.check_unstable_api()
+        return self.model
+
+    """
+    TODO: 实现将 self.model 中的不稳定（unstable）API 转换为稳定（stable）API 的逻辑。
+    该 API 负责遍历 self.model，并将其中调用的实验性或不稳定接口替换为对应的稳定版本。
+    注意：此逻辑属于模型编译安全机制的重要组成部分，请勿随意修改或删除。
+    
+    api命名规范：
+    <unstable>_to_<stable>
+    
+    stable api链接： 
+    """
+
+    def unstable_to_stable(self):
+        return
+
+    def check_unstable_api(self):
+        """
+        Check whether gm contains the API specified in the environment
+        variable DISALLOWED_UNSTABLE_API. If it does, raise an exception and stop
+        execution immediately.
+
+        IMPORTANT:
+        This logic is part of the GraphNet compiler safety mechanism.
+        Do NOT modify, remove, or bypass this check under any circumstances.
+        """
+        unstable_api = os.getenv("DISALLOWED_UNSTABLE_API", "").strip()
+        if not unstable_api:
+            return  # Skip check if no environment variable is set
+
+        from torch.fx import symbolic_trace
+
+        try:
+            # Convert the model into a static computation graph (FX IR)
+            traced = symbolic_trace(self.model)
+            graph_text = str(traced.graph)
+        except Exception as e:
+            # In case tracing fails, fallback to textual model dump
+            graph_text = str(self.model)
+
+        # Search for the unstable API substring
+        if unstable_api in graph_text:
+            count = graph_text.count(unstable_api)
+            raise RuntimeError(
+                f"❌ Detected unstable API '{unstable_api}' '{count}' times in model graph.\n"
+                f"Please replace it with a stable API before proceeding.\n"
+            )
+        else:
+            print(f"✅ Model passed: no occurrence of '{unstable_api}' found.")
+
+    def synchronize(self):
+        # Synchronize CUDA operations if available
+        if torch.cuda.is_available():
+            torch.cuda.synchronize()

graph_net/torch/test_compiler.py

@@ -21,15 +21,18 @@
 from graph_net.torch.backend.tensorrt_backend import TensorRTBackend
 from graph_net.torch.backend.blade_disc_backend import BladeDISCBackend
 from graph_net.torch.backend.nope_backend import NopeBackend
+from graph_net.torch.backend.unstable_to_stable_backend import UnstableToStableBackend
 from graph_net.test_compiler_util import generate_allclose_configs
 
+
 registry_backend = {
     "tvm": TvmBackend(),
     "xla": XlaBackend(),
     "inductor": InductorBackend(),
     "tensorrt": TensorRTBackend(),
     "bladedisc": BladeDISCBackend(),
     "nope": NopeBackend(),
+    "unstable_to_stable": UnstableToStableBackend(),
 }
 
 
@@ -215,7 +218,7 @@ def test_single_model(args):
     )
 
     version_str = "unknown"
-    if args.compiler == "inductor":
+    if args.compiler in ["inductor", "unstable_to_stable"]:
         version_str = torch.__version__
     elif args.compiler in ["tvm", "xla", "tensorrt", "bladedisc"]:
         # Assuming compiler object has a version attribute