change GPT2ForSequenceClassification inference accuracy tolerance (#136749)

Fixes https://github.com/pytorch/pytorch/issues/123503.

https://github.com/pytorch/pytorch/pull/121866 makes GPT2ForSequenceClassification hit the SDPA pattern 18 and then encounter the accuracy issue. The issue only happens with BF16 inference single thread. This PR tends to increase the model tolerance from 4e-3 to 5e-3 and make the check pass. Note that the issue is due to some small implementation diff. For example, the sdpa math backend scales q, k before matmul for stability; the flash attention backend has more diffs as a new algorithm.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/136749
Approved by: https://github.com/jgong5, https://github.com/jansel

benchmarks/dynamo/huggingface.py

@@ -501,12 +501,12 @@ class HuggingfaceRunner(BenchmarkRunner):
             else:
                 return 1e-2, cosine
         else:
-            if name in self._config["tolerance"]["higher_inference"]:
-                return 4e-3, cosine
             if (
                 current_device == "cpu"
                 and name in self._config["tolerance"]["higher_inference_cpu"]
             ):
+                return 5e-3, cosine
+            if name in self._config["tolerance"]["higher_inference"]:
                 return 4e-3, cosine
         return 1e-3, cosine
 

benchmarks/dynamo/huggingface.yaml

@@ -89,6 +89,7 @@ tolerance:
 
   higher_inference_cpu:
     - LayoutLMForSequenceClassification
+    - GPT2ForSequenceClassification
 
   cosine: []