Converters refactor [base config/model]: Updates to the base config and model for better loading support and standardization

pyproject.toml

@@ -156,7 +156,7 @@ ignore = [
     "COM812",
     "ISC001",
     "TC002",
-    "TC003",  # allow imports outside of type checking blocks
+    "TC003", # allow imports for typing outside of type checking blocks
     "S311", # allow random number generators
     "PLW1514", # allow Path.open without encoding
     "RET505", # allow `else` blocks

src/speculators/models/eagle.py

@@ -309,6 +309,7 @@ def attach_verifier(
         self,
         verifier: Union[str, os.PathLike, PreTrainedModel],
         mode: Optional[Literal["full", "train_only"]] = None,
+        add_to_config: bool = True,
     ) -> PreTrainedModel:
         """
         Attach a verifier model to the EagleSpeculator for speculative decoding.
@@ -344,15 +345,19 @@ def attach_verifier(
             model directory, a Hugging Face model identifier, or an instance of
             PreTrainedModel. If a path or identifier is provided, the model will be
             loaded automatically. If an instance is provided, it will be used directly.
-        :param mode: The mode for attaching the verifier. Can be "full" or "train_only".
+        :param mode: The mode for attaching the verifier.
+            Can be "full" or "train_only".
             If None, defaults to "full". In "train_only" mode, only the layers
             required for a forward pass are attached, and the speculator cannot
             perform generation until a full verifier is attached.
+        :param add_to_config: Whether to update the speculator's configuration
+            with details from the attached verifier model.
         :return: The PreTrainedModel instance for the verifier that was attached.
         """
         verifier = super().attach_verifier(
             verifier=verifier,
             mode=mode,
+            add_to_config=add_to_config,
         )
 
         # Extract layers from the verifier model

tests/integration/test_config.py

@@ -22,7 +22,7 @@ def test_verifier_config_from_verifier_config():
             cache_dir=tmp_dir,
         )
 
-    config = VerifierConfig.from_config(
+    config = VerifierConfig.from_pretrained(
         pretrained_config, name_or_path="RedHatAI/Llama-3.1-8B-Instruct"
     )
     assert config.name_or_path == "RedHatAI/Llama-3.1-8B-Instruct"

tests/unit/models/test_eagle_model.py

@@ -94,6 +94,8 @@ def forward(self, input_ids, **kwargs):
 @pytest.fixture
 def sample_llama_config():
     return LlamaConfig(
+        name_or_path="test/verifier",
+        architectures=["LlamaForCausalLM"],
         attention_bias=False,
         attention_dropout=0.0,
         bos_token_id=128000,