[cherry-pick] Add FP16 support for uniform in dygraph mode on Nvidia GPU (#46641)

Add FP16 support for uniform in dygraph mode on Nvidia GPU
Dev PR link PR46212

paddle/phi/kernels/gpu/uniform_random_kernel.cu

@@ -74,8 +74,12 @@ void UniformRandomRawKernel(const Context& dev_ctx,
     funcs::distribution_and_transform<T>(dev_ctx, out, dist, trans);
   } else {
     // Use OP seed
-    auto func = UniformGenerator<T>(
-        min.to<float>(), max.to<float>(), seed, diag_num, diag_step, diag_val);
+    auto func = UniformGenerator<T>(static_cast<T>(min.to<float>()),
+                                    static_cast<T>(max.to<float>()),
+                                    seed,
+                                    diag_num,
+                                    diag_step,
+                                    static_cast<T>(diag_val));
     IndexKernel<T, UniformGenerator<T>>(dev_ctx, out, func);
   }
 }
@@ -87,4 +91,5 @@ PD_REGISTER_KERNEL(uniform_random_raw,
                    ALL_LAYOUT,
                    phi::UniformRandomRawKernel,
                    float,
-                   double) {}
+                   double,
+                   phi::dtype::float16) {}

paddle/phi/kernels/uniform_random_kernel.cc

@@ -51,8 +51,13 @@ PD_REGISTER_KERNEL(uniform_random,
                    phi::dtype::bfloat16) {}
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
-PD_REGISTER_KERNEL(
-    uniform_random, GPU, ALL_LAYOUT, phi::UniformRandomKernel, float, double) {}
+PD_REGISTER_KERNEL(uniform_random,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::UniformRandomKernel,
+                   float,
+                   double,
+                   phi::dtype::float16) {}
 #endif
 
 #ifdef PADDLE_WITH_XPU

python/paddle/fluid/tests/unittests/test_uniform_random_op.py

@@ -587,8 +587,17 @@ class TestUniformDtype(unittest.TestCase):
             out = paddle.tensor.random.uniform([2, 3])
             self.assertEqual(out.dtype, fluid.core.VarDesc.VarType.FP64)
 
+        def test_dygraph_fp16():
+            if not paddle.is_compiled_with_cuda():
+                paddle.enable_static()
+                return
+            paddle.set_device('gpu')
+            out = paddle.uniform([2, 3], dtype=paddle.float16)
+            self.assertEqual(out.dtype, fluid.core.VarDesc.VarType.FP16)
+
         test_default_fp64()
         test_default_fp32()
+        test_dygraph_fp16()
 
         paddle.enable_static()