fix summary naming

Author: titaiwangms

_unittests/ut_tasks/test_tasks_text_generation.py

@@ -24,11 +24,11 @@ def test_text_generation(self):
             patch=True,
         )
         self.assertIsInstance(summary, dict)
-        # token generation
+        # multi-turn conversation
         self.assertLess(summary["disc_onnx_ort_run_abs"], 3e-2)
         # prompt processing
         self.assertLess(summary["disc_onnx_ort_run2_abs"], 3e-2)
-        # multi-turn conversation
+        # token generation
         self.assertLess(summary["disc_onnx_ort_run3_abs"], 3e-2)
         self.assertIsInstance(data, dict)
         onnx_filename = data["onnx_filename"]

_unittests/ut_torch_models/test_validate_models.py

@@ -41,7 +41,7 @@ def test_validate_tiny_llms_bfloat16(self):
     @requires_transformers("4.53")
     @requires_torch("2.7.99")
     @requires_experimental()
-    @hide_stdout()
+    # @hide_stdout()
     def test_validate_microsoft_phi4_reasoning(self):
         # python -m onnx_diagnostic validate -m microsoft/Phi-4-mini-reasoning
         #       --run -v 1 --export custom  -o dump_test --no-quiet --device cuda --patch

_unittests/ut_torch_models/test_validate_whole_models.py

@@ -229,7 +229,7 @@ def test_m_validate_model_vit_model(self):
         self.assertIsInstance(summary, dict)
         self.assertIsInstance(data, dict)
         self.assertLess(summary["disc_onnx_ort_run_abs"], 1e-3)
-        self.assertLess(summary["disc_onnx_ort_run22_abs"], 1e-3)
+        self.assertLess(summary["disc_onnx_ort_run2_abs"], 1e-3)
         self.assertEqual("dict(pixel_values:A1s2x3x30x30)", summary["run_feeds_inputs"])
         self.assertEqual("dict(pixel_values:A1s3x3x31x31)", summary["run_feeds_inputs2"])
         self.assertEqual("#1[A1s2x2]", summary["run_output_inputs"])

onnx_diagnostic/tasks/text_generation.py

@@ -281,7 +281,7 @@ def get_inputs(
         )["inputs"]
         # Token generation (decode) testing
         # NOTE: We have to export model in decode mode to preserve the cache
-        res["token_generation"] = get_inputs(
+        res["inputs3"] = get_inputs(
             model=model,
             config=config,
             dummy_max_token_id=dummy_max_token_id,

onnx_diagnostic/torch_export_patches/patches/patch_transformers.py

@@ -1950,7 +1950,6 @@ def patched_sdpa_attention_forward(
         torch._check(
             attention_mask.shape[3] == key.shape[2] if attention_mask is not None else True
         )
-
         attn_output = torch.nn.functional.scaled_dot_product_attention(
             query,
             key,

onnx_diagnostic/torch_models/validate.py

@@ -683,8 +683,8 @@ def validate_model(
                     data,
                     summary,
                     k,
-                    f"run2{k[6:]}",
-                    f"run_expected2{k[6:]}",
+                    f"run{k[6:]}",
+                    f"run_expected{k[6:]}",
                     verbose,
                     1,
                     0,
@@ -1431,7 +1431,7 @@ def _mk(key, flavour=flavour):
 
     keys = [("inputs", "run_expected", "")]
     if second_input_keys:
-        keys.extend([(k, f"run_expected2{k[6:]}", f"2{k[6:]}") for k in second_input_keys])
+        keys.extend([(k, f"run_expected{k[6:]}", f"{k[6:]}") for k in second_input_keys])
     for k_input, k_expected, suffix in keys:
         # make_feeds
         assert k_input in data, f"Unable to find {k_input!r} in {sorted(data)}"