Fixes native inference input size mistmatch issue (#447)

Adds extra input during construct_input phase to avoid input vs
session_input size mismatch issue during native inference.

---------

Co-authored-by: Vivek Agrawal <[email protected]>

alt_e2eshark/onnx_tests/models/hf_models.py

@@ -45,7 +45,7 @@
 # model task. If a huge number of models that can be grouped
 # in a common category fall under this list, a new meta_constructor
 # should be created for them.
-update_tokenizer_input_names = [
+models_with_input_names_2 = {
     "hf_paraphrase-multilingual-MiniLM-L12-v2",
     "hf_all-MiniLM-L6-v2",
     "hf_jina-embeddings-v2-small-en",
@@ -138,7 +138,27 @@
     "hf_distilbert-base-nli-mean-tokens",
     "hf_distilbert-base-multilingual-cased",
     "hf_distilbert-base-cased",
-]
+}
+
+models_with_input_names_3 = {
+    "hf_bart-base",
+    "hf_gpt2-small-spanish",
+    "hf_opt-125m",
+    "hf_Qwen1.5-0.5B-Chat",
+    "hf_Qwen2-0.5B",
+    "hf_Qwen2.5-0.5B-Instruct",
+    "hf_really-tiny-falcon-testing",
+    "hf_tiny-dummy-qwen2",
+    "hf_tiny-Qwen2ForCausalLM-2.5",
+    "hf_tiny-random-GemmaForCausalLM",
+    "hf_tiny-random-LlamaForCausalLM",
+    "hf_tiny-random-mt5",
+    "hf_tiny-random-Phi3ForCausalLM",
+}
+
+models_with_input_names_4 = {
+    "hf_ivila-row-layoutlm-finetuned-s2vl-v2",
+}
 
 # Add a basic_opt list to apply O1 to the models.
 basic_opt = []
@@ -224,14 +244,49 @@ def construct_inputs(self):
         prompt = ["Deeds will not be less valiant because they are unpraised."]
 
         tokenizer = get_tokenizer_from_model_path(self.model_repo_path, self.cache_dir)
-        if self.name in update_tokenizer_input_names:
-            tokenizer.model_input_names = ["input_ids", "attention_mask"]
 
         tokens = tokenizer(prompt, return_tensors="pt", padding=True, truncation=True)
-        inputs = (*list(tokens.values()),)
-
         self.input_name_to_shape_map = {k: v.shape for (k, v) in tokens.items()}
 
+        if self.name in models_with_input_names_2:
+            # Handles 2 inputs
+            tokenizer.model_input_names = ["input_ids", "attention_mask"]
+            inputs = (*list(tokens.values()), )
+        else:
+            self.input_name_to_shape_map["position_ids"] = self.input_name_to_shape_map["input_ids"]
+            zeros = torch.zeros(*(self.input_name_to_shape_map["input_ids"]), dtype=int)
+            if self.name in models_with_input_names_3:
+                # Handles 3 inputs
+                tokenizer.model_input_names = ["input_ids", "attention_mask", "position_ids"]
+            elif self.name in models_with_input_names_4:
+                tokenizer.model_input_names = ["input_ids", "bbox", "attention_mask", "position_ids"]
+
+                # Handles 4 inputs
+                # Tokenizer is returning tokens dict with key token_type_ids" instead of "bbox".
+                # For now, "token_type_ids" will be reused as bbox in this case
+                # bbox is a bounding box with size [?, ?, 4]
+                #   where each 4 numbers represent x_min, y_min, x_max, y_max
+                tokens["token_type_ids"] = tokens["token_type_ids"].unsqueeze(-1).repeat(1, 1, 4)
+            else:
+                raise RuntimeError(f"Model: {self.name} not found in any of the registry lists.")
+
+            inputs = (*list(tokens.values()), zeros)
+
+        test_tensors = TestTensors(inputs)
+        return test_tensors
+
+
+class HfModelWithRandomInput(HfDownloadableModel):
+    def export_model(self, optim_level: str | None = None):
+        # We won't need optim_level.
+        del optim_level
+        super().export_model("O1" if self.name in basic_opt else None)
+
+    def construct_inputs(self):
+        inputs = torch.randn(1, 4, 16000)
+
+        self.input_name_to_shape_map = {'input_ids': torch.Size([16000, 4]), 'attention_mask': torch.Size([16000, 4])}
+
         test_tensors = TestTensors(inputs)
         return test_tensors
 
@@ -269,7 +324,7 @@ def construct_inputs(self):
         if (
             "deberta" in self.name
             or "roberta" in self.name
-            or self.name in update_tokenizer_input_names
+            or self.name in models_with_input_names_2
         ):
             tokenizer.model_input_names = ["input_ids", "attention_mask"]