refactor: simplify BaseFactory and separate ModelFactory from AutoModel

- Extract _resolve_base_type and _validate_component as module-level helpers
- Replace ForwardRef._evaluate private API with eval in module namespace
- Remove broad except Exception in __init_subclass__, _component_base always set
- Replace direct _entries mutation in strategy.py with register() call form
- Remove dead TOKENIZER_CLASSES registry from AutoTokenizer
- Extract ModelFactory(BaseFactory[nn.Module]) as pure factory
- AutoModel now inherits only nn.Module, no factory state
- Move @AutoModel.register to @ModelFactory.register in transformer.py and encoder.py

astrai/factory.py

@@ -13,41 +13,63 @@ from typing import (
     Type,
     TypeVar,
     Union,
+    get_args,
+    get_origin,
 )
-from typing import get_args as _get_args
-from typing import get_origin as _get_origin
 
 T = TypeVar("T")
 
 
-def _resolve_type(
+def _resolve_base_type(
     arg: Union[Type, str, ForwardRef], factory_cls: type
 ) -> Optional[Type]:
-    """Resolve a generic type-arg (str forward-ref, ForwardRef, or class)."""
+    """Resolve the generic type-arg T to a concrete class.
-    if not isinstance(arg, (str, ForwardRef)):
+
+    - Concrete class (``BaseFactory[MyBase]``): returned directly.
+    - Forward reference (``BaseFactory["MyBase"]``): ``Base["X"]``
+      produces a ``ForwardRef("X")`` at class-creation time.  We
+      extract the name and evaluate it in the factory module's
+      global namespace — the same mechanism ``typing.get_type_hints``
+      uses internally.
+    """
+    if isinstance(arg, type):
         return arg
 
-    name = arg if isinstance(arg, str) else arg.__forward_arg__
+    if isinstance(arg, str):
-    if name == factory_cls.__name__:
+        name = arg
-        return factory_cls
+    elif isinstance(arg, ForwardRef):
+        name = arg.__forward_arg__
+    else:
+        return None
 
     mod = sys.modules.get(factory_cls.__module__)
     if mod is None:
         return None
-    ns = vars(mod)
+    try:
+        return eval(name, vars(mod))  # noqa: S307
+    except NameError:
+        return None
 
-    if isinstance(arg, ForwardRef):
-        return arg._evaluate(ns, None, recursive_guard=frozenset())
 
-    return ns.get(name)
+def _validate_component(component_cls: Type, base: Optional[Type]) -> None:
+    """Validate that *component_cls* inherits from *base*.
+
+    No-op when *base* is ``None`` (e.g. forward-ref resolution failed).
+    """
+    if base is not None and not issubclass(component_cls, base):
+        raise TypeError(f"{component_cls.__name__} must inherit from {base.__name__}")
 
 
 class BaseFactory(ABC, Generic[T]):
-    """Generic factory with decorator-based component registration.
+    """Generic factory with decorator-based registration.
+
+    Create a factory by subclassing with the desired base type::
 
         class MyFactory(BaseFactory[MyBase]):
             pass
 
+    Register components with the ``register`` decorator::
+
         @MyFactory.register("custom")
         class CustomComponent(MyBase):
             ...
@@ -64,13 +86,10 @@ class BaseFactory(ABC, Generic[T]):
     def __init_subclass__(cls, **kwargs):
         super().__init_subclass__(**kwargs)
         for orig_base in getattr(cls, "__orig_bases__", ()):
-            if _get_origin(orig_base) is BaseFactory:
+            if get_origin(orig_base) is BaseFactory:
-                (arg,) = _get_args(orig_base)
+                (arg,) = get_args(orig_base)
                 cls._entries = {}
-                try:
+                cls._component_base = _resolve_base_type(arg, cls)
-                    cls._component_base = _resolve_type(arg, cls)
-                except Exception:
-                    cls._component_base = None
                 return
 
     @classmethod
@@ -82,7 +101,7 @@ class BaseFactory(ABC, Generic[T]):
         """
 
         def decorator(component_cls: Type[T]) -> Type[T]:
-            cls._validate_component(component_cls)
+            _validate_component(component_cls, cls._component_base)
             if name in cls._entries:
                 raise ValueError(f"Component '{name}' is already registered")
             cls._entries[name] = component_cls
@@ -95,12 +114,11 @@ class BaseFactory(ABC, Generic[T]):
         """Create a component instance by name, filtering kwargs to match
         the component's ``__init__`` signature.
         """
-        entry = cls._entries.get(name)
+        component_cls = cls._entries.get(name)
-        if entry is None:
+        if component_cls is None:
             raise ValueError(
                 f"Unknown component: '{name}'. Supported types: {sorted(cls._entries)}"
             )
-        component_cls = entry
         sig = inspect.signature(component_cls.__init__)
         has_var_kwargs = any(
             p.kind == inspect.Parameter.VAR_KEYWORD for p in sig.parameters.values()
@@ -114,18 +132,6 @@ class BaseFactory(ABC, Generic[T]):
             kwargs = {k: v for k, v in kwargs.items() if k in valid}
         return component_cls(*args, **kwargs)
 
-    @classmethod
-    def _validate_component(cls, component_cls: Type[T]):
-        """Validate the decorated class inherits from the factory's base type.
-
-        Override for custom validation beyond ``issubclass``.
-        """
-        base = cls._component_base
-        if base is not None and not issubclass(component_cls, base):
-            raise TypeError(
-                f"{component_cls.__name__} must inherit from {base.__name__}"
-            )
-
     @classmethod
     def get_component_class(cls, name: str) -> Type[T]:
         """Get the registered component class without instantiating it."""

astrai/model/automodel.py

@@ -40,11 +40,12 @@ def _disable_random_init(enable: bool = True):
             setattr(nn.init, n, fn)
 
 
-class AutoModel(BaseFactory["AutoModel"], nn.Module):
+class ModelFactory(BaseFactory[nn.Module]):
-    """
+    """Pure factory for model dispatch, separated from nn.Module state."""
-    Autoregressive language model base class.
+
-    Provides model loading/saving, registration, and generation.
+
-    """
+class AutoModel(nn.Module):
+    """Model base class with loading/saving and generation."""
 
     def __init__(self, config: BaseModelConfig):
         super().__init__()
@@ -68,7 +69,7 @@ class AutoModel(BaseFactory["AutoModel"], nn.Module):
         config = ConfigFactory.load(raw)
         model_type = config.model_type or "autoregressive_lm"
 
-        actual_cls = AutoModel.get_component_class(model_type)
+        actual_cls = ModelFactory.get_component_class(model_type)
 
         with _disable_random_init(enable=disable_random_init):
             model = actual_cls(config)

astrai/model/encoder.py

@@ -5,7 +5,7 @@ import torch.nn as nn
 from torch import Tensor
 
 from astrai.config.model_config import EncoderConfig
-from astrai.model.automodel import AutoModel
+from astrai.model.automodel import AutoModel, ModelFactory
 from astrai.model.components.decoder_block import DecoderBlock
 from astrai.model.components.embedding import Embedding
 from astrai.model.components.norm import RMSNorm
@@ -13,7 +13,7 @@ from astrai.model.components.rope import RotaryEmbedding
 from astrai.model.transformer import process_attention_mask
 
 
-@AutoModel.register("embedding")
+@ModelFactory.register("embedding")
 class EmbeddingEncoder(AutoModel):
     def __init__(self, config: EncoderConfig):
         super().__init__(config)

astrai/model/transformer.py

@@ -6,7 +6,7 @@ from torch import Tensor
 
 from astrai.config.model_config import AutoRegressiveLMConfig
 from astrai.inference.core.cache import CacheView
-from astrai.model.automodel import AutoModel
+from astrai.model.automodel import AutoModel, ModelFactory
 from astrai.model.components.decoder_block import DecoderBlock
 from astrai.model.components.embedding import Embedding
 from astrai.model.components.linear import Linear
@@ -26,7 +26,7 @@ def process_attention_mask(
     return input_mask
 
 
-@AutoModel.register("autoregressive_lm")
+@ModelFactory.register("autoregressive_lm")
 class AutoRegressiveLM(AutoModel):
     """Autoregressive language model with paged KV cache."""
 

astrai/tokenize/tokenizer.py

@@ -20,8 +20,6 @@ Messages = List[Message]
 class AutoTokenizer:
     """Base tokenizer class with automatic loading support"""
 
-    TOKENIZER_CLASSES = {}  # Registry for auto-loading
-
     def __init__(
         self,
         path: Optional[Union[str, Path]] = None,
@@ -108,17 +106,6 @@ class AutoTokenizer:
         with open(save_path / "tokenizer_config.json", "w", encoding="utf-8") as f:
             json.dump(config, f, ensure_ascii=False, indent=2)
 
-    @classmethod
-    def register_tokenizer(cls, name: str, tokenizer_class: type):
-        """
-        Register a new tokenizer class.
-
-        Args:
-            name: Name to register the tokenizer class under
-            tokenizer_class: The tokenizer class to register
-        """
-        cls.TOKENIZER_CLASSES[name] = tokenizer_class
-
     def encode(
         self,
         tokens: Union[str, List[str]],

astrai/trainer/strategy.py

@@ -501,5 +501,5 @@ class GRPOStrategy(BaseStrategy):
 # Factory aliases: online variants use the same strategy class; the
 # ``RolloutRunner`` is injected by ``TrainContextBuilder`` to enable
 # online mode, so no separate subclass is needed.
-StrategyFactory._entries["online_grpo"] = GRPOStrategy
+StrategyFactory.register("online_grpo")(GRPOStrategy)
-StrategyFactory._entries["online_dpo"] = DPOStrategy
+StrategyFactory.register("online_dpo")(DPOStrategy)

tests/module/test_encoder.py

@@ -7,7 +7,7 @@ import safetensors.torch as st
 import torch
 
 from astrai.config.model_config import EncoderConfig
-from astrai.model.automodel import AutoModel
+from astrai.model.automodel import ModelFactory
 from astrai.model.encoder import EmbeddingEncoder
 from tests.helpers import TINY_CONFIG, assert_state_dicts_equal
 
@@ -69,8 +69,8 @@ def test_encoder_normalize(device):
 
 
 def test_encoder_register():
-    assert AutoModel.is_registered("embedding")
+    assert ModelFactory.is_registered("embedding")
-    cls = AutoModel.get_component_class("embedding")
+    cls = ModelFactory.get_component_class("embedding")
     assert cls is EmbeddingEncoder
 
 

tests/trainer/test_online_strategy.py

@@ -92,8 +92,8 @@ def _make_dpo(device, executor=None):
 def test_factory_registers_online_aliases():
     assert StrategyFactory.is_registered("online_grpo")
     assert StrategyFactory.is_registered("online_dpo")
-    assert StrategyFactory._entries["online_grpo"] is GRPOStrategy
+    assert StrategyFactory.get_component_class("online_grpo") is GRPOStrategy
-    assert StrategyFactory._entries["online_dpo"] is DPOStrategy
+    assert StrategyFactory.get_component_class("online_dpo") is DPOStrategy
 
 
 def test_grpo_supports_online(device):