A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 07:00 UTC
[Misc] Bump up test_fused_moe tolerance (#10364)
Signed-off-by: ElizaWszola <eliza@neuralmagic.com>
ElizaWszola committed
Nov 15, 2024 at 17:31 UTC
79ee45b42822d750ead6121c8c741c8a947bfeaf
1 file changed
+1
-1
tests/kernels/test_moe.py
+1
-1
@@ -45,7 +45,7 @@ def test_fused_moe(
45
score = torch.randn((m, e), device="cuda", dtype=dtype)
46
triton_output = fused_moe(a, w1, w2, score, topk, renormalize=False)
47
torch_output = torch_moe(a, w1, w2, score, topk)
48
- torch.testing.assert_close(triton_output, torch_output, atol=1e-2, rtol=0)
48
+ torch.testing.assert_close(triton_output, torch_output, atol=2e-2, rtol=0)
49
50
51
@pytest.mark.parametrize("dtype",