[0-size Tensor No.82] Add 0-size Tensor support for paddle.incubate.nn.functional.fused_bias_act [fluid_ops]

paddle/phi/infermeta/multiary.cc

@@ -2476,7 +2476,7 @@ void FusedBiasActInferMeta(const MetaTensor& x,
     x_shapes.push_back(x_dims[i]);
   }
 
-  if (config.is_runtime) {
+  if (config.is_runtime && x.numel() != 0) {
     PADDLE_ENFORCE_GT(
         x.numel() / dim,
         0,

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

@@ -556,6 +556,20 @@ void FusedBiasActKernel(const Context &dev_ctx,
                         float quant_max_bound,
                         float quant_min_bound,
                         DenseTensor *out) {
+  if (out && out->numel() == 0) {
+    if (quant_scale > 0) {
+      dev_ctx.template Alloc<int8_t>(out);
+    } else if (compute_dtype == "fp16") {
+      dev_ctx.template Alloc<phi::dtype::float16>(out);
+    } else if (compute_dtype == "bf16") {
+      dev_ctx.template Alloc<phi::dtype::bfloat16>(out);
+    } else if (compute_dtype == "fp32") {
+      dev_ctx.template Alloc<float>(out);
+    } else {
+      dev_ctx.template Alloc<T>(out);
+    }
+    return;
+  }
   int64_t cols = x.dims()[x.dims().size() - 1];
   int64_t rows = x.numel() / cols;
   if (x.dtype() == phi::DataType::INT32) {

paddle/phi/kernels/fusion/xpu/fused_bias_act_kernel.cc

@@ -112,6 +112,7 @@ void FusedBiasActKernel(const Context &dev_ctx,
                         DenseTensor *out) {
   auto xpu_ctx = static_cast<const phi::XPUContext *>(&dev_ctx);
   dev_ctx.template Alloc<T>(out);
+  if (out->numel() == 0) return;
 
   if (dequant_scales && dequant_scales.get().numel() > 0) {
     return DispatchComputeImpl<T>(xpu_ctx,

test/legacy_test/test_fused_bias_act_op.py

@@ -784,5 +784,31 @@ def test_check_output(self):
         )
 
 
+@unittest.skipIf(
+    not core.is_compiled_with_cuda() and not core.is_compiled_with_rocm(),
+    "core is not compiled with CUDA or ROCm",
+)
+class TestFusedBiasActOp_ZeroSize(TestWithoutBias):
+    def setUp(self):
+        paddle.seed(2017)
+        np.random.seed(2017)
+
+        self.op_type = "fused_bias_act"
+        self.rtol = 1e-5
+        self.atol = 1e-3
+
+        self.batch_size = 2
+        self.seq_len = 0
+        self.cols = 512
+
+        self.dtype = 'float32'
+        self.act_method = 'gelu'
+
+        self.use_glu = False
+
+        self.init_test_case()
+        self.generate_inputs()
+
+
 if __name__ == '__main__':
     unittest.main()