tighten threshold

yibinl-nvidia · yibinl-nvidia · commit cf69b40b0b09 · 2025-11-21T20:21:38.000Z
Signed-off-by: Yibin Li &lt;109242046+yibinl-nvidia@users.noreply.github.com&gt;
diff --git a/tests/unittest/_torch/modeling/test_modeling_starcoder2.py b/tests/unittest/_torch/modeling/test_modeling_starcoder2.py
@@ -237,7 +237,7 @@ def test_starcoder2_allclose_to_hf(scenario: Scenario) -> None:
             position_ids=position_ids,
             use_cache=True,
         )
-        torch.testing.assert_close(logits, ref.logits[:, -1].float(), atol=0.4, rtol=0.4)
+        torch.testing.assert_close(logits, ref.logits[:, -1].float(), atol=0.1, rtol=0.1)
 
     # Generation phase (optionally with CUDA graphs)
     gen_input_ids = torch.tensor([900], dtype=torch.int, device=device)

Original file line number	Diff line number	Diff line change
`@@ -237,7 +237,7 @@ def test_starcoder2_allclose_to_hf(scenario: Scenario) -> None:`
`237`	`237`	`position_ids=position_ids,`
`238`	`238`	`use_cache=True,`
`239`	`239`	`)`
`240`		`- torch.testing.assert_close(logits, ref.logits[:, -1].float(), atol=0.4, rtol=0.4)`
	`240`	`+ torch.testing.assert_close(logits, ref.logits[:, -1].float(), atol=0.1, rtol=0.1)`
`241`	`241`
`242`	`242`	`# Generation phase (optionally with CUDA graphs)`
`243`	`243`	`gen_input_ids = torch.tensor([900], dtype=torch.int, device=device)`