fix ut

Author: fxyfxy777

python/test/unit/hopper/test_gemm_fusion.py

@@ -80,7 +80,7 @@ def test_gemm_fusion():
         BLOCK_M, BLOCK_N, BLOCK_K,  #
         num_warps=num_warps)
 
-    torch.testing.assert_close(ref_out, E, atol=3e-1, rtol=0)
+    torch.testing.assert_close(ref_out, E, atol=1e-2, rtol=0)
 
 
 @triton.jit

python/test/unit/hopper/test_persistent_warp_specialized_gemm.py

@@ -171,7 +171,7 @@ def test_user_defined_persistent_non_warp_specialized_gemm(M, N, K, BLOCK_M, BLO
                                               num_ctas=NUM_CTAS)
 
     th_c = torch.matmul(a, b)
-    torch.testing.assert_close(th_c, c, atol=5e-1, rtol=0, check_dtype=False)
+    torch.testing.assert_close(th_c, c, atol=1e-2, rtol=0, check_dtype=False)
 
 
 @triton.jit
@@ -317,7 +317,7 @@ def test_non_persistent_warp_specialized_gemm(M, N, K, BLOCK_M, BLOCK_N, BLOCK_K
             num_ctas=NUM_CTAS)
 
     th_c = torch.matmul(a, b)
-    torch.testing.assert_close(th_c, c, atol=5e-2, rtol=0, check_dtype=False)
+    torch.testing.assert_close(th_c, c, atol=1e-2, rtol=0, check_dtype=False)
 
 
 @triton.jit
@@ -467,7 +467,7 @@ def test_user_defined_persistent_warp_specialized_gemm(M, N, K, BLOCK_M, BLOCK_N
             num_warps=4, num_ctas=NUM_CTAS)
 
     th_c = torch.matmul(a, b)
-    torch.testing.assert_close(th_c, c, atol=5e-1, rtol=0, check_dtype=False)
+    torch.testing.assert_close(th_c, c, atol=1e-2, rtol=0, check_dtype=False)
 
 
 @triton.jit