[AsyncTP] Use assertEqual instead of allClose for bf16 tests (pytorch#162041)

fegin · pytorchmergebot · commit f044fa29021a · 2025-09-08T16:12:52.000Z
The async tp result and regular MM result are very close. If we adjust the allclose threshold, the test succeeds. This seems to indicate that the error is from numerical error of low precision. Pull Request resolved: pytorch#162041 Approved by: https://github.com/danielvegamyhre, https://github.com/ngimel ghstack dependencies: pytorch#162040
diff --git a/test/distributed/test_symmetric_memory.py b/test/distributed/test_symmetric_memory.py
@@ -558,7 +558,7 @@ def test_fused_scaled_matmul_reduce_scatter(
                 )
 
         assert outputs[0].stride() == outputs[1].stride()
-        assert torch.allclose(outputs[0], outputs[1]), (outputs[0], outputs[1])
+        self.assertEqual(outputs[0], outputs[1])
 
     @runOnRocmArch(MI300_ARCH)
     @parametrize("dim", [0, 1, 2])

Original file line number	Diff line number	Diff line change
`@@ -558,7 +558,7 @@ def test_fused_scaled_matmul_reduce_scatter(`
`558`	`558`	`)`
`559`	`559`
`560`	`560`	`assert outputs[0].stride() == outputs[1].stride()`
`561`		`- assert torch.allclose(outputs[0], outputs[1]), (outputs[0], outputs[1])`
	`561`	`+ self.assertEqual(outputs[0], outputs[1])`
`562`	`562`
`563`	`563`	`@runOnRocmArch(MI300_ARCH)`
`564`	`564`	`@parametrize("dim", [0, 1, 2])`