Merge branch 'jz/native-runner-tt' into jz/tt-llama-kv-cache

Author: jackzhxng

.ci/scripts/gather_test_models.py

@@ -25,6 +25,7 @@
         "resnet50": "linux.12xlarge",
         "llava": "linux.12xlarge",
         "llama3_2_vision_encoder": "linux.12xlarge",
+        "llama3_2_text_decoder": "linux.12xlarge",
         # This one causes timeout on smaller runner, the root cause is unclear (T161064121)
         "dl3": "linux.12xlarge",
         "emformer_join": "linux.12xlarge",

examples/models/__init__.py

@@ -19,6 +19,7 @@
     "llama2": ("llama", "Llama2Model"),
     "llama": ("llama", "Llama2Model"),
     "llama3_2_vision_encoder": ("llama3_2_vision", "FlamingoVisionEncoderModel"),
+    "llama3_2_text_decoder": ("llama3_2_vision", "Llama3_2Decoder"),
     "lstm": ("lstm", "LSTMModel"),
     "mobilebert": ("mobilebert", "MobileBertModelExample"),
     "mv2": ("mobilenet_v2", "MV2Model"),

examples/models/llama/export_llama_lib.py

@@ -888,6 +888,8 @@ def _load_llama_model(
         if modelname == "llama3_2_vision":
             module_name = "llama3_2_vision"
             model_class_name = "Llama3_2Decoder"
+        else:
+            raise ValueError(f"{modelname} is not a valid Llama model.")
     else:
         raise ValueError(f"{modelname} is not a valid Llama model.")