Qwen and Phi-4-Mini targets (#13449)

Summary: Targets for Qwen 3, Phi-4-Mini

Reviewed By: cccclai

Differential Revision: D80015307

Author: rohansjoshi

examples/models/llama/TARGETS

@@ -84,6 +84,7 @@ runtime.python_binary(
     ],
     deps = [
         ":export_library",
+        ":eval_library",
         "//caffe2:torch",
         "//executorch/extension/pybindings:aten_lib",
         "//executorch/extension/llm/export:export_llm_lib",

examples/models/llama/export_llama_lib.py

@@ -605,17 +605,11 @@ def export_llama(
     if not llm_config.base.checkpoint and model_name in HUGGING_FACE_REPO_IDS:
         repo_id = HUGGING_FACE_REPO_IDS[model_name]
         if model_name == "qwen2_5":
-            from executorch.examples.models.qwen2_5 import (  # pyre-ignore[21]
-                convert_weights,
-            )
+            from executorch.examples.models.qwen2_5 import convert_weights
         elif model_name.startswith("qwen3"):
-            from executorch.examples.models.qwen3 import (  # pyre-ignore[21]
-                convert_weights,
-            )
+            from executorch.examples.models.qwen3 import convert_weights
         elif model_name == "phi_4_mini":
-            from executorch.examples.models.phi_4_mini import (  # pyre-ignore[21]
-                convert_weights,
-            )
+            from executorch.examples.models.phi_4_mini import convert_weights
         elif model_name == "smollm2":
             from executorch.examples.models.smollm2 import (  # pyre-ignore[21]
                 convert_weights,

examples/qualcomm/oss_scripts/llama/TARGETS

@@ -35,7 +35,7 @@ python_library(
 
 python_library(
     name = "llama_lib",
-    srcs = ["llama.py"],
+    srcs = ["__init__.py", "llama.py"],
     deps = [
         ":decoder_constants",
         ":decoder_utils",
@@ -47,6 +47,7 @@ python_library(
         "//executorch/devtools:lib",
         "//executorch/examples/models:models",
         "//executorch/examples/models/llama:hf_download",
+        "//executorch/examples/qualcomm/oss_scripts/llama:range_setting_pt2e",
         "//executorch/examples/qualcomm/oss_scripts/llama:static_llama",
         "//executorch/examples/qualcomm:utils",
         "//executorch/extension/export_util:export_util",
@@ -73,7 +74,6 @@ python_binary(
     ],
     deps = [
         ":llama_lib",
-        "//executorch/examples/qualcomm/oss_scripts/llama:range_setting_pt2e",
     ],
 )
 

examples/qualcomm/oss_scripts/llama/__init__.py

@@ -32,7 +32,7 @@ class HFModel(ABC):
     convert_weights: Callable
 
 
-SUPPORTED_HF_MODELS: Dict[str, Type[HFModel]] = {}
+SUPPORTED_HF_MODELS: Dict[str, HFModel] = {}
 
 
 def register_hf_model(name: str):

examples/qualcomm/oss_scripts/llama/llama.py

@@ -1040,7 +1040,7 @@ def _build_parser():
     parser.add_argument(
         "--model_mode",
         help="Export and inference kv mode, hybrid mode, or lookahead decoding mode",
-        default="kv",
+        default="hybrid",
         choices=["kv", "hybrid", "lookahead"],
         type=str,
     )