fix: accept Hugging Face configs without a model_type in HfModel

olive/common/hf/utils.py

@@ -9,14 +9,22 @@
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, Optional, Union
 
-from transformers import AutoConfig, AutoModel, AutoModelForSeq2SeqLM, AutoProcessor, AutoTokenizer, GenerationConfig
+from transformers import (
+    AutoConfig,
+    AutoModel,
+    AutoModelForSeq2SeqLM,
+    AutoProcessor,
+    AutoTokenizer,
+    GenerationConfig,
+    PretrainedConfig,
+)
 
 from olive.common.hf.mappings import TASK_TO_PEFT_TASK_TYPE
 from olive.common.hf.mlflow import get_pretrained_name_or_path
 from olive.common.utils import hardlink_copy_file
 
 if TYPE_CHECKING:
-    from transformers import PretrainedConfig, PreTrainedModel, PreTrainedTokenizer, PreTrainedTokenizerFast
+    from transformers import PreTrainedModel, PreTrainedTokenizer, PreTrainedTokenizerFast
 
 logger = logging.getLogger(__name__)
 TEST_MODEL_MARKER_FILE = "olive_test_model.json"
@@ -407,7 +415,17 @@ def get_model_config(
     model_name_or_path: str, test_model_config: Optional[dict[str, Any]] = None, **kwargs
 ) -> "PretrainedConfig":
     """Get HF Config for the given model_name_or_path."""
-    model_config = from_pretrained(AutoConfig, model_name_or_path, "config", **kwargs)
+    try:
+        model_config = from_pretrained(AutoConfig, model_name_or_path, "config", **kwargs)
+    except ValueError:
+        # A config without a model_type (e.g. LiquidAI/LFM2.5-Audio-1.5B) is not a transformers model, but passes
+        # like ModelBuilder only need the config. Remote-code configs keep their error asking for trust_remote_code.
+        config_dict, unused_kwargs = PretrainedConfig.get_config_dict(
+            get_pretrained_name_or_path(model_name_or_path, "config"), **kwargs
+        )
+        if "model_type" in config_dict or "auto_map" in config_dict:
+            raise
+        model_config = PretrainedConfig.from_dict(config_dict, **unused_kwargs)
 
     # add quantization config
     quantization_config = kwargs.get("quantization_config")

test/common/test_hf.py

@@ -15,6 +15,7 @@
     TEST_MODEL_MARKER_FILE,
     _apply_test_model_config,
     _load_test_model,
+    get_model_config,
     load_model_from_task,
 )
 
@@ -43,6 +44,32 @@ def test_apply_test_model_config_reduces_nested_vision_depth():
     assert reduced.vision_config.depth == 2
 
 
+def test_get_model_config_returns_plain_config_when_model_type_missing(tmp_path):
+    config = {"architectures": ["Lfm2AudioForConditionalGeneration"], "codebooks": 8, "lfm": {"hidden_size": 2048}}
+    (tmp_path / "config.json").write_text(json.dumps(config))
+
+    model_config = get_model_config(str(tmp_path), dtype="bfloat16")
+
+    assert model_config.architectures == ["Lfm2AudioForConditionalGeneration"]
+    assert model_config.codebooks == 8
+    assert model_config.lfm == {"hidden_size": 2048}
+    assert model_config.dtype == "bfloat16"
+
+
+def test_get_model_config_raises_when_model_type_unknown(tmp_path):
+    (tmp_path / "config.json").write_text(json.dumps({"model_type": "not_a_real_model_type"}))
+
+    with pytest.raises(ValueError, match="not_a_real_model_type"):
+        get_model_config(str(tmp_path))
+
+
+def test_get_model_config_raises_when_remote_code_not_trusted(tmp_path):
+    (tmp_path / "config.json").write_text(json.dumps({"auto_map": {"AutoConfig": "configuration_x.XConfig"}}))
+
+    with pytest.raises(ValueError, match="custom code"):
+        get_model_config(str(tmp_path), trust_remote_code=False)
+
+
 def test_load_model_from_task():
     # The model name and task type is gotten from
     # https://huggingface.co/docs/transformers/v4.28.1/en/main_classes/pipelines#transformers.pipeline