optimize op structure (#55988)

paddle/phi/api/yaml/fused_ops.yaml

@@ -83,7 +83,7 @@
   optional : bias, x_max
 
 - op : fused_bias_act
-  args : (Tensor x, Tensor bias, Tensor dequant_scales, Tensor shift, Tensor smooth, str act_method = "gelu", str compute_dtype = "default", int rows = -1, int cols = -1, float quant_scale = -1, int quant_round_type = 1, float quant_max_bound = 127.0, float quant_min_bound = -127.0)
+  args : (Tensor x, Tensor bias, Tensor dequant_scales, Tensor shift, Tensor smooth, str act_method = "gelu", str compute_dtype = "default", float quant_scale = -1, int quant_round_type = 1, float quant_max_bound = 127.0, float quant_min_bound = -127.0)
   output : Tensor(out)
   infer_meta :
     func: FusedBiasActInferMeta

paddle/phi/infermeta/multiary.cc

@@ -1342,8 +1342,6 @@ void FusedBiasActInferMeta(const MetaTensor& x,
                            const MetaTensor& smooth,
                            const std::string& act_method,
                            const std::string& compute_dtype,
-                           int rows,
-                           int cols,
                            float quant_scale,
                            int quant_round_type,
                            float quant_max_bound,
@@ -1358,10 +1356,14 @@ void FusedBiasActInferMeta(const MetaTensor& x,
   auto dim = x_dims[1];
 
   PADDLE_ENFORCE_GT(
-      rows, 0, phi::errors::InvalidArgument("The size of Attr(rows) must > 0"));
+      x_dims[0],
+      0,
+      phi::errors::InvalidArgument("The size of Attr(rows) must > 0"));
 
   PADDLE_ENFORCE_GT(
-      cols, 0, phi::errors::InvalidArgument("The size of Attr(cols) must > 0"));
+      x_dims[1],
+      0,
+      phi::errors::InvalidArgument("The size of Attr(cols) must > 0"));
 
   if (act_method == "geglu" || act_method == "swiglu") {
     PADDLE_ENFORCE_EQ(

paddle/phi/infermeta/multiary.h

@@ -286,8 +286,6 @@ void FusedBiasActInferMeta(const MetaTensor& x,
                            const MetaTensor& smooth,
                            const std::string& act_method,
                            const std::string& compute_dtype,
-                           int rows,
-                           int cols,
                            float quant_scale,
                            int quant_round_type,
                            float quant_max_bound,

paddle/phi/kernels/fusion/gpu/fused_bias_act_kernel.cu

@@ -438,14 +438,14 @@ void FusedBiasActKernel(const Context &dev_ctx,
                         const paddle::optional<DenseTensor> &smooth,
                         const std::string &act_method,
                         const std::string &compute_dtype,
-                        int rows,
-                        int cols,
                         float quant_scale,
                         int quant_round_type,
                         float quant_max_bound,
                         float quant_min_bound,
                         DenseTensor *out) {
 #ifndef PADDLE_WITH_HIP
+  int rows = x.dims()[0];
+  int cols = x.dims()[1];
   if (x.dtype() == phi::DataType::INT32) {
     if (compute_dtype == "bf16") {
       DispatchWithDtype<phi::dtype::bfloat16, Context>(

test/legacy_test/test_fused_bias_act_op.py

@@ -73,8 +73,6 @@ def fused_act_bias_wrapper(
     smooth=None,
     act_method='gelu',
     compute_dtype='default',
-    rows=0,
-    cols=0,
     quant_scale=-1,
     quant_round_type=0,
     quant_max_bound=0,
@@ -88,8 +86,6 @@ def fused_act_bias_wrapper(
         smooth,
         act_method,
         compute_dtype,
-        rows,
-        cols,
         quant_scale,
         quant_round_type,
         quant_max_bound,
@@ -140,8 +136,6 @@ class TestFusedBiasActOp(unittest.TestCase):
         return fused_act_bias_wrapper(
             x=x,
             bias=bias,
-            rows=self.rows,
-            cols=self.cols,
             act_method=self.act_method,
             compute_dtype=self.compute_dtype,
         )
@@ -197,8 +191,6 @@ class TestFastGeluFP16(TestFusedBiasActOp):
         out = fused_act_bias_wrapper(
             x=x,
             bias=bias,
-            rows=self.rows,
-            cols=self.cols,
             act_method=self.act_method,
         )
         self.use_fast_math(False)
@@ -284,8 +276,6 @@ class TestQuantFP32(TestFusedBiasActOp):
             smooth=smooth,
             act_method=self.act_method,
             compute_dtype=self.compute_dtype,
-            rows=self.rows,
-            cols=self.cols,
             quant_scale=self.quant_scale,
             quant_round_type=self.quant_round_type,
             quant_max_bound=self.quant_max_bound,
@@ -332,8 +322,6 @@ class TestDequantFP32(TestQuantFP32):
             dequant_scales=dequant_scales,
             act_method=self.act_method,
             compute_dtype=self.compute_dtype,
-            rows=self.rows,
-            cols=self.cols,
         )
         return out
 
@@ -441,8 +429,6 @@ class TestFusedBiasActOpBF16(unittest.TestCase):
             bias=bias,
             act_method=self.act_method,
             compute_dtype=self.compute_dtype,
-            rows=self.rows,
-            cols=self.cols,
         )
         return out
 
@@ -565,8 +551,6 @@ class TestQuantBF16(TestFusedBiasActOpBF16):
             smooth=smooth,
             act_method=self.act_method,
             compute_dtype=self.compute_dtype,
-            rows=self.rows,
-            cols=self.cols,
             quant_scale=self.quant_scale,
             quant_round_type=self.quant_round_type,
             quant_max_bound=self.quant_max_bound,
@@ -678,8 +662,6 @@ class TestAssert(unittest.TestCase):
             out = fused_act_bias_wrapper(
                 x=paddle.to_tensor(x),
                 bias=paddle.to_tensor(bias),
-                rows=self.rows,
-                cols=self.cols,
             )
         except ValueError as e:
             pass
@@ -696,8 +678,6 @@ class TestAssert(unittest.TestCase):
             out = fused_act_bias_wrapper(
                 x=paddle.to_tensor(x),
                 bias=paddle.to_tensor(bias),
-                rows=self.rows,
-                cols=self.cols,
                 compute_dtype='fp16',
             )
         except ValueError as e:
@@ -715,8 +695,6 @@ class TestAssert(unittest.TestCase):
             out = fused_act_bias_wrapper(
                 x=paddle.to_tensor(x),
                 bias=paddle.to_tensor(bias),
-                rows=self.rows,
-                cols=self.cols,
                 compute_dtype='fp16',
                 act_method=act_method,
             )
@@ -765,8 +743,6 @@ class TestWithoutBias(unittest.TestCase):
         return fused_act_bias_wrapper(
             x=x,
             bias=None,
-            rows=self.rows,
-            cols=self.cols,
             act_method=self.act_method,
         )