!8685 [GraphKernel] For fp16 value, declare fp32 firstly and than cast to fp16 in expander

From: @tronzhang
Reviewed-by: 
Signed-off-by:

akg

@@ -1 +1 @@
-Subproject commit ef1a6b06781035540023819afead4bdfbd49af81
+Subproject commit 6d01f5e364224da0d58b8b20761b9af67587950b

mindspore/_extends/graph_kernel/expanders/gelu.py

@@ -36,9 +36,6 @@ def expand_gelu(expand_info):
         # create tensor input.
         input_x = graph_builder.tensor(input_desc['shape'], input_desc['data_type'], input_desc['format'])
         graph_scope.set_input(input_x)
-        dtype = input_x.dtype
-        if dtype == 'float16':
-            input_x = graph_builder.emit('Cast', [input_x], attrs={'dst_type': 'float32'})
 
         # cal y
         mul_0 = graph_builder.emit('Mul', [input_x, input_x])
@@ -58,8 +55,6 @@ def expand_gelu(expand_info):
         mul_x = graph_builder.emit('Mul', [input_x, tanh_y_add_one])
         result = graph_builder.emit('Mul', [const_half, mul_x])
 
-        if dtype == 'float16':
-            result = graph_builder.emit('Cast', [result], attrs={'dst_type': 'float16'})
         # set graph output.
         graph_scope.set_output(result)
 

mindspore/_extends/graph_kernel/expanders/gelu_grad.py

@@ -43,9 +43,6 @@ def expand_gelugrad(expand_info):
         input_x = graph_builder.tensor(input_desc_1['shape'], input_desc_1['data_type'], input_desc_1['format'])
         input_y = graph_builder.tensor(input_desc_2['shape'], input_desc_2['data_type'], input_desc_2['format'])
         graph_scope.set_input(input_dy, input_x, input_y)
-        dtype = input_dy.dtype
-        if dtype == 'float16':
-            input_dy = graph_builder.emit('Cast', [input_dy], attrs={'dst_type': 'float32'})
 
         # create some const var
         const_csvalue = graph_builder.value(input_dy.dtype, CSVALUE, input_desc_0['format'])
@@ -83,8 +80,6 @@ def expand_gelugrad(expand_info):
         result_tmp = graph_builder.emit('TensorAdd', [half_mul_tanh_res_add_one, mul_final])
         result = graph_builder.emit('Mul', [input_dy, result_tmp])
 
-        if dtype == 'float16':
-            result = graph_builder.emit('Cast', [result], attrs={'dst_type': 'float16'})
         # set graph output.
         graph_scope.set_output(result)
 

mindspore/_extends/graph_kernel/model/model_builder.py

@@ -151,7 +151,15 @@ class GraphBuilder:
         """Create a new Value"""
         if name in (None, ''):
             name = self._alloc_tensor_name()
-        return Value(name, dtype, value, data_format)
+
+        if dtype == "float16":
+            # For float16 value, it will be changed to float32 wrongly. And there is no good solution for now.
+            # So instead just declare float32 value and then cast it to float16.
+            v_fp32 = Value(name, "float32", value, data_format)
+            v = self.emit("Cast", [v_fp32], attrs={"dst_type": "float16"})
+        else:
+            v = Value(name, dtype, value, data_format)
+        return v
 
     def op(self, prim, output, inputs, attrs=None):
         """Insert an operator into graph"""
@@ -168,9 +176,9 @@ class GraphBuilder:
         """Emit a new operation"""
         if attrs is None:
             attrs = {}
-        if isinstance(inputs, Tensor):
+        if isinstance(inputs, (Tensor, Value)):
             inputs = [inputs]
-        tensor_inputs = [t for t in inputs if isinstance(t, Tensor)]
+        tensor_inputs = [t for t in inputs if isinstance(t, (Tensor, Value))]
         out_shape, out_dtype, out_format = OpInfer.infer(prim, tensor_inputs, attrs)
         output = self.tensor(out_shape, out_dtype, out_format, name)
         self.op(prim, output, inputs, attrs)