use 4.57.6 in CI

.github/workflows/ci.yml

@@ -17,7 +17,7 @@ jobs:
       matrix:
         os: [ubuntu-latest]
         python: ['3.10', '3.11', '3.12', '3.13']
-        transformers: ['4.48.3', '4.51.3', '4.52.4', '4.55.4', '4.56.2', '4.57.3', 'main']
+        transformers: ['4.48.3', '4.51.3', '4.52.4', '4.55.4', '4.56.2', '4.57.6', 'main']
         torch: ['2.9', 'main']
         exclude:
           - python: '3.10'  # 3.10
@@ -33,7 +33,7 @@ jobs:
           - python: '3.10'
             transformers: '4.56.2'
           - python: '3.10'
-            transformers: '4.57.3'
+            transformers: '4.57.6'
           - python: '3.11'  # 3.11
             torch: 'main'
           - python: '3.11'
@@ -43,7 +43,7 @@ jobs:
           - python: '3.11'
             transformers: '4.56.2'
           - python: '3.11'
-            transformers: '4.57.3'
+            transformers: '4.57.6'
           - python: '3.13'  # 3.11
             torch: '2.9'
           - python: '3.13'

.github/workflows/models.yml

@@ -17,7 +17,7 @@ jobs:
       matrix:
         os: [ubuntu-latest]
         python: ['3.12']
-        transformers: ['4.57.3']
+        transformers: ['4.57.6']
         torch: ['main']
     steps:
       - uses: actions/checkout@v3

onnx_diagnostic/torch_export_patches/patches/patch_torch.py

@@ -524,13 +524,16 @@ def compute_concrete_val() -> sympy.Basic:
 
             transmute_into_runtime_assert = False
 
+            backed_var_to_val = getattr(
+                self, "backed_var_to_val", getattr(self, "var_to_val", {})
+            )
             concrete_val = None
-            if not (expr.free_symbols <= self.var_to_val.keys()):
+            if not (expr.free_symbols <= backed_var_to_val.keys()):
                 # TODO: dedupe this with _maybe_evaluate_static
                 # Attempt to eliminate the unbacked SymInt
                 new_expr = self._maybe_evaluate_static(expr, unbacked_only=True)
                 assert new_expr is not None
-                if not (new_expr.free_symbols <= self.var_to_val.keys()):
+                if not (new_expr.free_symbols <= backed_var_to_val.keys()):
                     ok = False
 
                     # fallback_value is set when guard_or_true or guard_or_false are used.
@@ -541,17 +544,15 @@ def compute_concrete_val() -> sympy.Basic:
                     # oblivious_var_to_val will be defined iff we have sizes
                     # with DimDynamic.OBLIVIOUS_SIZE type.
                     # See https://github.com/pytorch/pytorch/issues/137100#issuecomment-2495778113
-                    var_to_val = getattr(
-                        self,
-                        "unbacked_var_to_val",
-                        getattr(self, "oblivious_var_to_val", False),
-                    )
                     if (
-                        var_to_val
-                        and not (correct_hint := orig_expr.xreplace(var_to_val)).free_symbols
+                        backed_var_to_val
+                        and getattr(self, "real_tensor_prop_unbacked_vals", True)
+                        and not (
+                            correct_hint := orig_expr.xreplace(backed_var_to_val)
+                        ).free_symbols
                         and not (
                             counterfactual_hint := orig_expr.xreplace(
-                                {k: max(2, v) for k, v in var_to_val.items()}
+                                {k: max(2, v) for k, v in backed_var_to_val.items()}
                             )
                         ).free_symbols
                         and correct_hint == counterfactual_hint
@@ -574,10 +575,10 @@ def compute_concrete_val() -> sympy.Basic:
                     # and if they pass we add a runtime assertions and continue.
                     if (
                         not ok
-                        and var_to_val
+                        and backed_var_to_val
                         and not (
-                            unsound_result := orig_expr.xreplace(var_to_val).xreplace(
-                                var_to_val
+                            unsound_result := orig_expr.xreplace(backed_var_to_val).xreplace(
+                                backed_var_to_val
                             )
                         ).free_symbols
                     ):

onnx_diagnostic/torch_models/hghub/hub_data_cached_configs.py

@@ -143,6 +143,7 @@ def _ccached_microsoft_phi2():
             "transformers_version": "4.51.0.dev0",
             "use_cache": true,
             "vocab_size": 51200,
+            "pad_token_id": 0,
         }
     )
 

onnx_diagnostic/torch_models/untrained/llm_phi2.py

@@ -53,6 +53,7 @@ def get_phi2(
         "transformers_version": "4.37.0",
         "use_cache": True,
         "vocab_size": 51200,
+        "pad_token_id": 0,
     }
     config.update(**kwargs)
     conf = transformers.PhiConfig(**config)