[Ascend] Wx/fix dtype cast bug of adamw op (#1260)

* fix dtype cast bug of adamw op

Author: POI-WX

adaptor/codegen/gen.py

@@ -65,6 +65,7 @@
     "IndexPut": ["out"],
     "Adadelta": ["input", "grad", "square_avg", "acc_delta"],
     "IndexBackward": ["zeros_like_input"],
+    "AdamW": ["param", "exp_avg", "exp_avg_sq", "max_exp_avg_sq"],
 }
 
 

impl/ascend/convert_config.yaml

@@ -1,6 +1,10 @@
 - common_config:
     layout: NCHW
 
+- diopiAdamW:
+    dtype: (float64)->float32
+    layout: ND
+
 - diopiSoftmax:
     layout: ND
 
@@ -447,8 +451,8 @@
 
 - diopiMaxPool2dWithIndices:
     tensor_dtype:
-        indices: (int64)->int32 
+        indices: (int64)->int32
 
 - diopiMaxPool2dBackward:
     tensor_dtype:
-        indices: (int64)->int32 
+        indices: (int64)->int32

impl/ascend/device_configs.py

@@ -1435,20 +1435,6 @@
         skip_all = True
     ),
 
-    'adam': dict(
-        name=['adamw'],
-        tensor_para=dict(
-            args=[
-                {
-                    "ins": ['param'],
-                    # float64 not supported yet on ascend
-                    # temporarily skip all test cases due to software stack version
-                    "dtype": [Skip(np.float16), Skip(np.float32), Skip(np.float64)],
-                },
-            ]
-        ),
-    ),
-
     # temporarily skip all test cases for flash_attention_varlen due to the version of software stack on ascend
     'flash_attention_varlen': dict(
         name=['flash_attention_varlen'],