[SOT] Use original tuple as cache key to avoid hash collision (#69408)

SigureMo · web-flow · commit 96aaf0e71525 · 2024-11-15T02:38:49.000+08:00
diff --git a/python/paddle/jit/sot/infer_meta.py b/python/paddle/jit/sot/infer_meta.py
@@ -446,18 +446,12 @@ class InferMetaCache(Cache, metaclass=Singleton):
     def key_fn(
         self, func, *args, **kwargs
     ):  # args & kwargs have transformed to MetaInfo
-        try:
-            retval = hash(
-                (
-                    func,
-                    tuple(flatten(args)),
-                    tuple(kwargs.keys()),
-                    tuple(flatten(kwargs)),
-                )
-            )
-        except Exception as e:
-            return None
-        return retval
+        return (
+            func,
+            tuple(flatten(args)),
+            tuple(kwargs.keys()),
+            tuple(flatten(kwargs)),
+        )
 
     def value_fn(self, func, *args, **kwargs):
         return infer_meta(func, *args, **kwargs)
@@ -469,19 +463,13 @@ def key_fn(self, layer, *args, **kwargs):
             MetaInfo.from_value(x)
             for x in layer.parameters(include_sublayers=True)
         ]
-        try:
-            retval = hash(
-                (
-                    layer,
-                    tuple(params),
-                    tuple(flatten(args)),
-                    tuple(kwargs.keys()),
-                    tuple(flatten(kwargs)),
-                )
-            )
-        except Exception as e:
-            return None
-        return retval
+        return (
+            layer,
+            tuple(params),
+            tuple(flatten(args)),
+            tuple(kwargs.keys()),
+            tuple(flatten(kwargs)),
+        )
 
     def value_fn(self, layer, *args, **kwargs):
         return infer_meta_for_layer(layer, *args, **kwargs)
diff --git a/python/paddle/jit/sot/utils/utils.py b/python/paddle/jit/sot/utils/utils.py
@@ -278,7 +278,7 @@ def __init__(self, weak=False):
 
     def __call__(self, *args, **kwargs):
         cache_key = self.key_fn(*args, **kwargs)
-        if cache_key is None:
+        if not hashable(cache_key):
             return self.value_fn(*args, **kwargs)
         if cache_key in self.cache:
             log(5, "cache hit: ", cache_key, "\n")
diff --git a/test/sot/test_sot_cache.py b/test/sot/test_sot_cache.py
@@ -0,0 +1,51 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+from paddle.jit.sot.utils import Cache
+
+
+class CustomCache(Cache):
+    def key_fn(self, x):
+        return x
+
+    def value_fn(self, x):
+        return x
+
+
+class TestCache(unittest.TestCase):
+    def test_basic(self):
+        cache = CustomCache()
+        self.assertEqual(cache(1), 1)
+        self.assertEqual(cache(1), 1)
+
+    def test_int_collision(self):
+        cache = CustomCache()
+        # -1 and -2 hash to the same value
+        self.assertEqual(cache(-1), -1)
+        self.assertEqual(cache(-2), -2)
+
+    def test_bool_and_int_collision(self):
+        cache = CustomCache()
+        # 1 and True hash to the same value
+        self.assertEqual(cache(True), True)
+        self.assertEqual(cache(1), 1)
+        # 0 and False hash to the same value
+        self.assertEqual(cache(False), False)
+        self.assertEqual(cache(0), 0)
+
+
+if __name__ == "__main__":
+    unittest.main()
