Recover atol of fused_attention.

PaddlePaddle · Nov 9, 2021 · 68bc308 · 68bc308
1 parent 28e539e
commit 68bc308
Show file tree

Hide file tree

Showing 2 changed files with 8 additions and 8 deletions.
diff --git a/python/paddle/fluid/tests/unittests/test_fused_attention_op.py b/python/paddle/fluid/tests/unittests/test_fused_attention_op.py
@@ -219,9 +219,9 @@ def test_fused_attention_op(self):
         final_out_ref, x_grad_ref = self.GetBaselineOut()
         final_out, x_grad = self.GetFusedAttentionOut()
         np.testing.assert_allclose(
-            final_out_ref, final_out.numpy(), rtol=1e-5, atol=1e-3)
+            final_out_ref, final_out.numpy(), rtol=1e-5, atol=1e-5)
         np.testing.assert_allclose(
-            x_grad_ref, x_grad.numpy(), rtol=1e-5, atol=1e-3)
+            x_grad_ref, x_grad.numpy(), rtol=1e-5, atol=1e-5)
 
 
 class TestFusedAttentionOpPreLn(TestFusedAttentionOp):
@@ -249,9 +249,9 @@ def test_fused_attention_op(self):
         final_out_ref, x_grad_ref = self.GetBaselineOut()
         final_out, x_grad = self.GetFusedAttentionOut()
         np.testing.assert_allclose(
-            final_out_ref, final_out.numpy(), rtol=1e-5, atol=1e-1)
+            final_out_ref, final_out.numpy(), rtol=1e-5, atol=1e-5)
         np.testing.assert_allclose(
-            x_grad_ref, x_grad.numpy(), rtol=1e-5, atol=1e-1)
+            x_grad_ref, x_grad.numpy(), rtol=1e-5, atol=1e-5)
 
 
 class TestFusedAttentionOpNoneAttnMask(TestFusedAttentionOp):
@@ -279,9 +279,9 @@ def test_fused_attention_op(self):
         final_out_ref, x_grad_ref = self.GetBaselineOut()
         final_out, x_grad = self.GetFusedAttentionOut()
         np.testing.assert_allclose(
-            final_out_ref, final_out.numpy(), rtol=1e-5, atol=1e-1)
+            final_out_ref, final_out.numpy(), rtol=1e-5, atol=1e-5)
         np.testing.assert_allclose(
-            x_grad_ref, x_grad.numpy(), rtol=1e-5, atol=1e-1)
+            x_grad_ref, x_grad.numpy(), rtol=1e-5, atol=1e-5)
 
 
 class TestFusedAttentionOpFp16(TestFusedAttentionOp):

diff --git a/python/paddle/fluid/tests/unittests/test_fused_attention_op_api.py b/python/paddle/fluid/tests/unittests/test_fused_attention_op_api.py
@@ -224,7 +224,7 @@ def run_imperative(self):
                                     fused_attn.qkv_bias.numpy(),
                                     fused_attn.linear_weight.numpy(),
                                     fused_attn.linear_bias.numpy())
-        np.testing.assert_allclose(ref_out, out.numpy(), rtol=1e-5, atol=1e-3)
+        np.testing.assert_allclose(ref_out, out.numpy(), rtol=1e-5, atol=1e-5)
 
     def run_static(self):
         fused_attn = FusedMultiHeadAttention(
@@ -312,7 +312,7 @@ def test_static_api(self):
                                     self.attn_mask, ln_scale, ln_bias,
                                     ln_2_scale, ln_2_bias, qkv_weight, qkv_bias,
                                     linear_weight, linear_bias)
-        np.testing.assert_allclose(ref_out, out, rtol=1e-5, atol=1e-3)
+        np.testing.assert_allclose(ref_out, out, rtol=1e-5, atol=1e-5)
 
     def test_dynamic_api(self):
         paddle.disable_static(place=paddle.CUDAPlace(0))