[WIP] fill: from_eagle_format w/o weight loading

src/speculators/__init__.py

@@ -6,6 +6,8 @@
     TokenProposal,
     TokenProposalConfig,
     VerifierConfig,
+    DraftModelType,
+    TokenProposalType,
 )
 from .logging import configure_logger, logger
 from .settings import LoggingSettings, Settings, print_config, reload_settings, settings
@@ -25,4 +27,6 @@
     "print_config",
     "reload_settings",
     "settings",
+    "DraftModelType",
+    "TokenProposalType",
 ]

src/speculators/base/__init__.py

@@ -4,6 +4,9 @@
     TokenProposalConfig,
     VerifierConfig,
     speculators_config_version,
+    DraftModelType,
+    TokenProposalType,
+    AlgorithmType,
 )
 from .objects import Drafter, SpeculatorModel, TokenProposal
 
@@ -16,4 +19,7 @@
     "TokenProposalConfig",
     "VerifierConfig",
     "speculators_config_version",
+    "DraftModelType",
+    "TokenProposalType",
+    "AlgorithmType",
 ]

src/speculators/base/config.py

@@ -1,3 +1,4 @@
+from enum import Enum
 import json
 from pathlib import Path
 from typing import Any, Optional, Union
@@ -12,10 +13,22 @@
     "TokenProposalConfig",
     "VerifierConfig",
     "speculators_config_version",
+    "DraftModelType",
+    "TokenProposalType",
+    "AlgorithmType",
 ]
 
 speculators_config_version = "0.1.0"
 
+class DraftModelType(str, Enum):
+    LLAMA_EAGLE = "LlamaEagle"
+
+class TokenProposalType(str, Enum):
+    GREEDY = "GREEDY"
+
+class AlgorithmType(str, Enum):
+    EAGLE = "EAGLE"
+
 
 class DraftModelConfig(BaseModel):
     """
@@ -26,7 +39,7 @@ class DraftModelConfig(BaseModel):
     self-drafting (extensions of the verifier model).
     """
 
-    type_: str = Field(
+    type_: DraftModelType = Field(
         description=(
             "The type of the speculator architecture. "
             "This must be a valid architecture name from the speculators library."
@@ -55,7 +68,7 @@ class TokenProposalConfig(BaseModel):
     greedy, nucleus, token tree sampling, etc.
     """
 
-    type_: str = Field(
+    type_: TokenProposalType = Field(
         description=(
             "The type of the token proposal algorithm. "
             "This must be a valid proposer from the speculators library."
@@ -84,7 +97,7 @@ class VerifierConfig(BaseModel):
             "transformers library. Ex: `LlamaPreTrainedModel`"
         )
     )
-    model: str = Field(
+    model: Optional[str] = Field(
         description=(
             "The name of the verifier model. This must be a valid HF id/name or "
             "a path to a local model directory. Ex: `meta-llama/Llama-3.3-70B-Instruct`"

src/speculators/base/objects.py

@@ -243,6 +243,7 @@ def __init__(
         self.drafter = drafter
         self.verifier = verifier  # need to make optional and attach later
         self.proposals = proposals
+
 
     @abstractmethod
     @property
@@ -254,7 +255,7 @@ def config(self) -> SpeculatorConfig:
 
         :return: The config object for the speculator model.
         """
-        ...
+        return self._config
 
     def attach_verifier(self, verifier: Module):
         """

src/speculators/utils/convert.py

@@ -3,10 +3,18 @@
 
 from torch.nn import Module
 
+from transformers import AutoConfig
 from speculators.base import (
-    SpeculatorConfig,
     SpeculatorModel,
-)  # will need to fix circular import
+    SpeculatorConfig,
+    DraftModelConfig,
+    DraftModelType,
+    TokenProposalConfig,
+    TokenProposalType,
+    AlgorithmType,
+    VerifierConfig,
+)
+import logging
 
 __all__ = [
     "SpecDecodeLibraryFormats",
@@ -20,7 +28,7 @@
 
 
 SpecDecodeLibraryFormats = Literal["speculators", "eagle", "eagle2", "eagle3", "hass"]
-
+logger = logging.getLogger(__name__)
 
 def detect_model_format(
     source: Union[str, Path, Module],  # noqa: ARG001
@@ -96,8 +104,41 @@ def from_eagle_format(
         function signature rather than keeping them in kwargs as needed.
     :return: The converted speculator model.
     """
-    raise NotImplementedError("Eagle format conversion is not implemented yet.")
-
+    # Load config if it's not already a dict
+    if not isinstance(config, dict):
+        config = AutoConfig.from_pretrained(source if config is None else config)
+
+    # Build draft model config
+    draft_model_config = DraftModelConfig(
+        type_=DraftModelType.LLAMA_EAGLE,
+        inputs="model.layers[-1].*",
+        config=config,
+    )
+
+    # Set proposal config
+    proposal_config = TokenProposalConfig(type_=TokenProposalType.GREEDY)
+
+    # Construct verifier config
+    verifier_config = VerifierConfig(
+        architecture=config.get("architectures"),
+        model=getattr(verifier, "source", None) if verifier else None,
+    )
+
+    # Build full speculator config
+    speculator_config = SpeculatorConfig(
+        speculators_algorithm=AlgorithmType.EAGLE,
+        draft_model=draft_model_config,
+        proposal_methods=proposal_config,
+        default_proposal_method=proposal_config.type_,
+        verifier=verifier_config,
+    )
+
+    speculator_model = SpeculatorModel.from_config(speculator_config)
+    logger.info(
+        "Returning speculator model from config. "
+        "Note: weights have not been loaded from disk yet."
+        )
+    return speculator_model
 
 def from_eagle2_format(
     source: Union[str, Path, Module],  # noqa: ARG001