Create create new method for example kwarg inputs instead

Author: jackzhxng

examples/models/llama2/export_llama_lib.py

@@ -814,7 +814,7 @@ def _load_llama_model(
             generate_full_logits=generate_full_logits,
             fairseq2=weight_type == WeightType.FAIRSEQ2,
             max_seq_len=max_seq_len,
-            enable_dynamic_shape=enable_dynamic_shape,
+            enable_dynamic_shape=enable_dynamic_shape ,
             output_prune_map_path=output_prune_map_path,
             args=args,
         )

examples/models/llama2/model.py

@@ -222,35 +222,32 @@ def get_eager_model(self) -> torch.nn.Module:
             # switch all to FP32
             return self.model_.to(torch.float32)
 
-    def get_example_inputs(self) -> Tuple[Tuple, Dict]:
+    def get_example_inputs(self):
         if self.use_kv_cache:
             return self.get_example_inputs_kvcache_sdpa()
         else:
-            positional_inputs = (
+            return (
                 torch.tensor(
                     [[1, 2, 3]], dtype=torch.long
                 ),  # tokens, with kv cache our input token length is always just 1 token.
             )
-            return (positional_inputs, {})
 
     # assumption is the custom op doesnt support dynamic shape right now. It might but its untested so lets first get static shape working
-    def get_example_inputs_kvcache_sdpa(self) -> Tuple[Tuple, Dict]:
+    def get_example_inputs_kvcache_sdpa(self):
         if self.enable_dynamic_shape:
-            positional_inputs = (
+            return (
                 torch.tensor([[2, 3, 4]], dtype=torch.long),
                 torch.tensor([0], dtype=torch.long),
             )
-            return (positional_inputs, {})
         else:
-            positional_inputs = (
+            return (
                 torch.tensor(
                     [[1]], dtype=torch.long
                 ),  # tokens, with kv cache our input token length is always just 1 token.
                 torch.tensor(
                     [0], dtype=torch.long
                 ),  # start_pos, what token of output are we on.
             )
-            return (positional_inputs, {})
 
     def _transform_for_pre_quantization(self, checkpoint):
         assert hasattr(self.args, "preq_mode"), "preq_mode must be specified"

examples/models/model_base.py

@@ -5,7 +5,6 @@
 # LICENSE file in the root directory of this source tree.
 
 from abc import ABC, abstractmethod
-from typing import Dict, Tuple
 
 import torch
 
@@ -38,11 +37,11 @@ def get_eager_model(self) -> torch.nn.Module:
         raise NotImplementedError("get_eager_model")
 
     @abstractmethod
-    def get_example_inputs(self) -> Tuple[Tuple, Dict]:
+    def get_example_inputs(self):
         """
         Abstract method to provide example inputs for the model.
 
         Returns:
-            Tuple[Tuple, Dict]: The positional inputs (Tuple) and the kwarg inputs (Dict).
+            Any: Example inputs that can be used for testing and tracing.
         """
         raise NotImplementedError("get_example_inputs")