[XPU][PHI Kernels] refine bf16 test for fused_rope (#60439)

lj970926 · web-flow · commit e4b39bb56a4e · 2023-12-29T15:05:25.000+08:00
* refine fuesd_rope bf16 test

* format code
diff --git a/test/xpu/test_fused_rotary_position_embedding_op_xpu.py b/test/xpu/test_fused_rotary_position_embedding_op_xpu.py
@@ -196,16 +196,17 @@ def get_forward_backward(
         fw.append(out_q)
         fw.append(out_k)
         fw.append(out_v)
+        paddle.seed(seed + 1)
         out_gq = paddle.randn(out_q.shape, self.dtype)
         out_gk = paddle.randn(out_q.shape, self.dtype)
         out_gv = paddle.randn(out_q.shape, self.dtype)
 
         paddle.autograd.backward(
             [out_q, out_k, out_v], [out_gq, out_gk, out_gv], True
         )
-        bw.append(tensor_q)
-        bw.append(tensor_k)
-        bw.append(tensor_v)
+        bw.append(tensor_q.grad)
+        bw.append(tensor_k.grad)
+        bw.append(tensor_v.grad)
 
         return fw, bw
 
@@ -368,42 +369,89 @@ def setUp(self):
         self.shape = [2, 8, 2, 16]
 
     def test_api(self):
-        q_fp32 = paddle.rand(self.shape, dtype="float32")
-        k_fp32 = paddle.rand(self.shape, dtype="float32")
-        v_fp32 = paddle.rand(self.shape, dtype="float32")
-        sin_fp32 = paddle.rand(
-            [1, self.shape[1], 1, self.shape[3]], dtype="float32"
+        paddle.disable_static()
+        q_bf16 = paddle.randn(self.shape, dtype="bfloat16")
+        k_bf16 = paddle.randn(self.shape, dtype="bfloat16")
+        v_bf16 = paddle.randn(self.shape, dtype="bfloat16")
+        sin_bf16 = paddle.randn(
+            [1, self.shape[1], 1, self.shape[3]], dtype="bfloat16"
         )
-        cos_fp32 = paddle.rand(
-            [1, self.shape[1], 1, self.shape[3]], dtype="float32"
+        cos_bf16 = paddle.randn(
+            [1, self.shape[1], 1, self.shape[3]], dtype="bfloat16"
         )
-        q_bf16 = paddle.to_tensor(q_fp32, dtype="bfloat16")
-        k_bf16 = paddle.to_tensor(k_fp32, dtype="bfloat16")
-        v_bf16 = paddle.to_tensor(v_fp32, dtype="bfloat16")
-        sin_bf16 = paddle.to_tensor(sin_fp32, dtype="bfloat16")
-        cos_bf16 = paddle.to_tensor(cos_fp32, dtype="bfloat16")
-
-        out_fp32 = fused_rotary_position_embedding(
-            q_fp32,
-            k_fp32,
-            v_fp32,
-            sin_fp32,
-            cos_fp32,
-            use_neox_rotary_style=False,
+        q_bf16.stop_gradient = False
+        k_bf16.stop_gradient = False
+        v_bf16.stop_gradient = False
+        q_fp32 = paddle.to_tensor(q_bf16, dtype="float32", stop_gradient=False)
+        k_fp32 = paddle.to_tensor(k_bf16, dtype="float32", stop_gradient=False)
+        v_fp32 = paddle.to_tensor(v_bf16, dtype="float32", stop_gradient=False)
+        sin_fp32 = paddle.to_tensor(sin_bf16, dtype="float32")
+        cos_fp32 = paddle.to_tensor(cos_bf16, dtype="float32")
+
+        position_ids = paddle.arange(0, self.shape[1], dtype="int64")
+        position_ids = paddle.stack(
+            [position_ids for _ in range(self.shape[0])], axis=0
         )
         out_bf16 = fused_rotary_position_embedding(
             q_bf16,
             k_bf16,
             v_bf16,
             sin_bf16,
             cos_bf16,
+            position_ids=position_ids,
+            use_neox_rotary_style=False,
+        )
+
+        grad_out_q_bf16 = paddle.randn(self.shape, dtype="bfloat16")
+        grad_out_k_bf16 = paddle.randn(self.shape, dtype="bfloat16")
+        grad_out_v_bf16 = paddle.randn(self.shape, dtype="bfloat16")
+
+        paddle.autograd.backward(
+            out_bf16, [grad_out_q_bf16, grad_out_k_bf16, grad_out_v_bf16], True
+        )
+        grad_bf16 = [q_bf16.grad, k_bf16.grad, v_bf16.grad]
+
+        out_fp32 = paddle_fused_rotary_position_embedding(
+            q_fp32,
+            k_fp32,
+            v_fp32,
+            sin_fp32,
+            cos_fp32,
+            position_ids=position_ids,
             use_neox_rotary_style=False,
         )
+
+        grad_out_q_fp32 = paddle.to_tensor(grad_out_q_bf16, dtype="float32")
+        grad_out_k_fp32 = paddle.to_tensor(grad_out_k_bf16, dtype="float32")
+        grad_out_v_fp32 = paddle.to_tensor(grad_out_v_bf16, dtype="float32")
+        paddle.autograd.backward(
+            out_fp32, [grad_out_q_fp32, grad_out_k_fp32, grad_out_v_fp32], True
+        )
+        grad_fp32 = [q_fp32.grad, k_fp32.grad, v_fp32.grad]
+
         for fp32_val, bf16_val in zip(out_fp32, out_bf16):
             bf16_val = convert_uint16_to_float(bf16_val.numpy())
             np.testing.assert_allclose(
                 fp32_val.numpy(), bf16_val, rtol=1e-2, atol=1e-2
             )
+        for grad_fp32_val, grad_bf16_val in zip(grad_fp32, grad_bf16):
+            grad_bf16_val = convert_uint16_to_float(grad_bf16_val.numpy())
+            np.testing.assert_allclose(
+                grad_fp32_val.numpy(), grad_bf16_val, rtol=1e-2, atol=1e-2
+            )
+
+
+class XPUTestFusedRotaryPositionEmbeddingBf16_2(
+    XPUTestFusedRotaryPositionEmbeddingBf16_1
+):
+    def setUp(self):
+        self.shape = [2, 2048, 16, 128]
+
+
+# too long for CI
+# class XPUTestFusedRotaryPositionEmbeddingBf16_3(XPUTestFusedRotaryPositionEmbeddingBf16_1):
+#     def setUp(self):
+#         self.shape = [2, 8192, 8, 128]
 
 
 if __name__ == '__main__':
