零散问题修复

paconvert/api_alias_mapping.json

@@ -48,8 +48,8 @@
   "torch.bilinear": "torch.nn.functional.bilinear",
   "torch.celu_": "torch.nn.functional.celu_",
   "torch.channel_shuffle": "torch.nn.functional.channel_shuffle",
-  "torch.concatenate": "torch.cat",
   "torch.clip": "torch.clamp",
+  "torch.concatenate": "torch.cat",
   "torch.conv1d": "torch.nn.functional.conv1d",
   "torch.conv2d": "torch.nn.functional.conv2d",
   "torch.conv3d": "torch.nn.functional.conv3d",

paconvert/api_mapping.json

@@ -58,6 +58,7 @@
       "num_embeddings",
       "embedding_dim",
       "padding_idx",
+      "max_norm",
       "norm_type",
       "scale_grad_by_freq",
       "sparse",
@@ -72,7 +73,8 @@
       "padding_idx",
       "norm_type",
       "scale_grad_by_freq",
-      "sparse"
+      "sparse",
+      "max_norm"
     ]
   },
   "fairscale.nn.model_parallel.layers.RowParallelLinear": {
@@ -125,14 +127,14 @@
     "unsupport_args": [
       "window_size",
       "alibi_slopes",
-      "deterministic",
-      "return_attn_probs"
+      "deterministic"
     ],
     "kwargs_change": {
       "q": "query",
       "k": "key",
       "v": "value",
-      "dropout_p": "dropout"
+      "dropout_p": "dropout",
+      "return_attn_probs": "return_softmax"
     }
   },
   "flash_attn.flash_attn_interface.flash_attn_unpadded_func": {
@@ -157,15 +159,15 @@
     "unsupport_args": [
       "window_size",
       "alibi_slopes",
-      "deterministic",
-      "return_attn_probs"
+      "deterministic"
     ],
     "kwargs_change": {
       "q": "query",
       "k": "key",
       "v": "value",
       "dropout_p": "dropout",
-      "softmax_scale": "scale"
+      "softmax_scale": "scale",
+      "return_attn_probs": "return_softmax"
     }
   },
   "flash_attn.layers.rotary.apply_rotary_emb_func": {
@@ -189,7 +191,10 @@
       "x",
       "weight",
       "epsilon"
-    ]
+    ],
+    "kwargs_change": {
+      "weight": "norm_weight"
+    }
   },
   "os.environ.get": {
     "Matcher": "OsEnvironGetMatcher",
@@ -2172,6 +2177,7 @@
   },
   "torch.Tensor.is_inference": {
     "Matcher": "Is_InferenceMatcher",
+    "paddle_api": "paddle.Tensor.stop_gradient",
     "min_input_args": 0
   },
   "torch.Tensor.is_pinned": {
@@ -3192,7 +3198,7 @@
     ]
   },
   "torch.Tensor.positive": {
-	  "Matcher": "PositiveMatcher"
+    "Matcher": "PositiveMatcher"
   },
   "torch.Tensor.pow": {
     "Matcher": "GenericMatcher",
@@ -3571,7 +3577,7 @@
     }
   },
   "torch.Tensor.scatter_reduce": {
-	  "Matcher": "ScatterReduceMatcher",
+    "Matcher": "ScatterReduceMatcher",
     "paddle_api": "paddle.Tensor.put_along_axis",
     "min_input_args": 3,
     "args_list": [
@@ -4024,7 +4030,8 @@
     ]
   },
   "torch.Tensor.to_sparse_coo": {
-    "Matcher": "TensorToSparseCooMatcher"
+    "Matcher": "TensorToSparseCooMatcher",
+    "paddle_api": "paddle.Tensor.to_sparse_coo"
   },
   "torch.Tensor.tolist": {
     "Matcher": "UnchangeMatcher",
@@ -7836,6 +7843,16 @@
       "axis": 0
     }
   },
+  "torch.float_power": {
+    "Matcher": "FloatPowerMatcher",
+    "min_input_args": 2,
+    "args_list": [
+      "input",
+      "exponent",
+      "*",
+      "out"
+    ]
+  },
   "torch.floor": {
     "Matcher": "GenericMatcher",
     "paddle_api": "paddle.floor",
@@ -7849,16 +7866,6 @@
       "input": "x"
     }
   },
-  "torch.float_power": {
-    "Matcher": "FloatPowerMatcher",
-    "min_input_args": 2,
-    "args_list": [
-      "input",
-      "exponent",
-      "*",
-      "out"
-    ]
-  },
   "torch.floor_divide": {
     "Matcher": "Num2TensorBinaryMatcher",
     "paddle_api": "paddle.floor_divide",
@@ -8518,6 +8525,7 @@
   },
   "torch.is_inference": {
     "Matcher": "Is_InferenceMatcher",
+    "paddle_api": "paddle.Tensor.stop_gradient",
     "min_input_args": 1,
     "args_list": [
       "input"
@@ -14867,11 +14875,6 @@
       "dims": "perm"
     }
   },
-  "torch.pi": {
-    "Matcher": "GenericMatcher",
-    "paddle_api": "numpy.pi",
-    "min_input_args": 0
-  },
   "torch.pinverse": {
     "Matcher": "GenericMatcher",
     "paddle_api": "paddle.linalg.pinv",

paconvert/api_matcher.py

@@ -3238,19 +3238,26 @@ def generate_code(self, kwargs):
 
 class AvgPoolMatcher(BaseMatcher):
     def generate_code(self, kwargs):
-        if "input" in kwargs:
-            kwargs["x"] = kwargs.pop("input")
+        new_kwargs = {}
+        for k in list(kwargs.keys()):
+            if k == "input":
+                new_kwargs["x"] = kwargs.pop(k)
+            else:
+                new_kwargs[k] = kwargs.pop(k)
 
-        if "count_include_pad" in kwargs:
-            kwargs["exclusive"] = "not " + kwargs.pop("count_include_pad")
+        if "count_include_pad" in new_kwargs:
+            new_kwargs["exclusive"] = "not " + new_kwargs.pop("count_include_pad")
         else:
-            kwargs["exclusive"] = "False"
+            new_kwargs["exclusive"] = "False"
+
         API_TEMPLATE = textwrap.dedent(
             """
             {}({})
             """
         )
-        code = API_TEMPLATE.format(self.get_paddle_api(), self.kwargs_to_str(kwargs))
+        code = API_TEMPLATE.format(
+            self.get_paddle_api(), self.kwargs_to_str(new_kwargs)
+        )
 
         return code
 
@@ -4358,13 +4365,18 @@ def generate_code(self, kwargs):
 class ConstantLRMatcher(LRSchedulerMatcher):
     def generate_code(self, kwargs):
         optim = kwargs["optimizer"]
-        factor = 0.3333333333333333
         total_iters = 5
         if "factor" in kwargs:
             factor = kwargs.pop("factor")
+            kwargs["values"] = "[{}*{}.get_lr(), {}.get_lr()]".format(
+                factor, optim, optim
+            )
+        else:
+            kwargs["values"] = "[{}.get_lr()/3, {}.get_lr()]".format(optim, optim)
+
         if "total_iters" in kwargs:
             total_iters = kwargs.pop("total_iters")
-        kwargs["values"] = "[{}*{}.get_lr(), {}.get_lr()]".format(factor, optim, optim)
+
         kwargs["boundaries"] = "[{}]".format(total_iters)
         return super().generate_code(kwargs)
 
@@ -4847,7 +4859,7 @@ def generate_code(self, kwargs):
 class PositiveMatcher(BaseMatcher):
     def generate_aux_code(self):
         CODE_TEMPLATE = textwrap.dedent(
-        """
+            """
         def positive(x):
             if x.dtype != paddle.bool:
                 return x
@@ -5204,7 +5216,7 @@ def get_scalable_var(self):
         if not (arg_name.startswith("*") and len(arg_name) > 1):
             return None
         return arg_name[1:]
-    
+
     def get_paddle_nodes(self, args, kwargs):
         var_arg_name = self.get_scalable_var()
         dest_var_arg_name = self.api_mapping.get("kwargs_change", {}).get(
@@ -5222,7 +5234,7 @@ def get_paddle_nodes(self, args, kwargs):
         return ast.parse(code).body
 
 
-class ScalableVarMatcher(BaseMatcher): 
+class ScalableVarMatcher(BaseMatcher):
     def get_scalable_var(self):
         args_list = self.api_mapping.get("args_list", [])
         if len(args_list) != 1:

paconvert/attribute_mapping.json

@@ -190,6 +190,10 @@
   "torch.optim.Optimizer.state_dict": {
     "Matcher": "UnchangeMatcher"
   },
+  "torch.pi": {
+    "Matcher": "GenericMatcher",
+    "paddle_api": "numpy.pi"
+  },
   "torch.short": {
     "Matcher": "GenericMatcher",
     "paddle_api": "'int16'"

tests/test_Tensor_scatter_reduce.py

@@ -111,6 +111,7 @@ def test_case_7():
     )
     obj.run(pytorch_code, ["result"])
 
+
 def test_case_8():
     pytorch_code = textwrap.dedent(
         """

tests/test_block_diag.py

@@ -41,8 +41,8 @@ def test_case_2():
         A = torch.tensor([[4], [3], [2]])
         B = torch.tensor([7, 6, 5])
         C = torch.tensor(1)
-        result = torch.block_diag(torch.tensor([[4], [3], [2]]), 
-                                torch.tensor([7, 6, 5]), 
+        result = torch.block_diag(torch.tensor([[4], [3], [2]]),
+                                torch.tensor([7, 6, 5]),
                                 torch.tensor(1))
         """
     )