packageification

example_trainer/__init__.py

@@ -8,13 +8,20 @@ Supports three training modes:
 - Legacy: Checkpoint-based training with vLLM restarts
 - Shared vLLM: Single-copy mode with CUDA IPC (no model duplication!)
 - LoRA: Adapter-only training with hot-swap capability
+- LoRA restart: Adapter training with periodic fast vLLM restarts
 
 Usage:
     # As CLI
     python -m example_trainer.grpo --model-name Qwen/Qwen2.5-3B-Instruct --training-steps 100
 
     # As library
-    from example_trainer import TrainingConfig, train_legacy, train_shared_vllm, train_lora
+    from example_trainer import (
+        TrainingConfig,
+        train_legacy,
+        train_shared_vllm,
+        train_lora,
+        train_lora_restart,
+    )
 
     config = TrainingConfig(model_name="Qwen/Qwen2.5-3B-Instruct", training_steps=100)
     train_legacy(config)
@@ -22,13 +29,14 @@ Usage:
 
 from .cli import config_from_args, parse_args
 from .config import TrainingConfig
-from .trainers import train_legacy, train_lora, train_shared_vllm
+from .trainers import train_legacy, train_lora, train_lora_restart, train_shared_vllm
 
 __all__ = [
     "TrainingConfig",
     "train_legacy",
     "train_shared_vllm",
     "train_lora",
+    "train_lora_restart",
     "parse_args",
     "config_from_args",
 ]