DISABLED test_cast_round_trip_soak_cuda_float8_e4m3fn (__main__.TestFloat8DtypeCUDA)

[huydhn]

Platforms: linux, win

This test was disabled because it is failing on main branch (recent examples).

This test is failing on periodic debug build with a Tensor-likes are not equal error, probably because of this change #107586

cc @galexite @malfet

[pytorch-bot]

Hello there! From the DISABLED prefix in this issue title, it looks like you are attempting to disable a test in PyTorch CI. The information I have parsed is below:

• Test name: test_cast_round_trip_soak_cuda_float8_e4m3fn (__main__.TestFloat8DtypeCUDA)
• Platforms for which to skip the test: linux, win
• Disabled by huydhn

Within ~15 minutes, test_cast_round_trip_soak_cuda_float8_e4m3fn (__main__.TestFloat8DtypeCUDA) will be disabled in PyTorch CI for these platforms: linux, win. Please verify that your test name looks correct, e.g., test_cuda_assert_async (__main__.TestCuda).

To modify the platforms list, please include a line in the issue body, like below. The default action will disable the test for all platforms if no platforms list is specified.

Platforms: case-insensitive, list, of, platforms

We currently support the following platforms: asan, dynamo, inductor, linux, mac, macos, rocm, slow, win, windows.

[galexite]

I don't think it's likely to be #107586, I think it's likely to be the same failure as before (from #113361) but the test has been renamed or moved since (from test_float8.py to test_quantization.py).

[malfet]

OK, I should have been more careful with code reviews: one should never compare floats with atol=0 as order of operations/compilation flags can have effect on the accuracy

[malfet]

I.e. here is a fix:

diff --git a/test/quantization/core/experimental/test_float8.py b/test/quantization/core/experimental/test_float8.py
index 54d45f905b1..fb4a23b1d3c 100644
--- a/test/quantization/core/experimental/test_float8.py
+++ b/test/quantization/core/experimental/test_float8.py
@@ -219,7 +219,7 @@ class TestFloat8Dtype(TestCase):
         x = torch.cat((x, -x))
         x8 = x.to(dtype)
         x8_simulated = simulate_fp8_precision(x, dtype)
-        self.assertEqual(x8_simulated, x8.float(), atol=0, rtol=0)
+        self.assertEqual(x8_simulated, x8.float())
 
     @dtypes(*FLOAT8_DTYPES)
     @dtypesIfCUDA(*CUDA_FLOAT8_DTYPES)