Fix client import error by vendoring transformers constants

This fixes issue #230 where importing art.Model on the client side
failed due to missing transformers dependencies. The issue was
introduced in PR #194 which moved backend dependencies to an optional
group but didn't account for art.dev importing transformers at the
module level.

The fix vendors all transformers enum constants into art/dev/model.py
with source comments, removing the hard dependency on transformers
for client imports while maintaining full compatibility.

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <[email protected]>

Author: corbt

dev/profile.ipynb

@@ -16,7 +16,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "from art.dev.model import get_model_config\n",
+    "from art.dev.get_model_config import get_model_config\n",
     "from art.local.state import ModelState\n",
     "\n",
     "config = get_model_config(\n",
@@ -188,4 +188,4 @@
  },
  "nbformat": 4,
  "nbformat_minor": 2
-}
+}

src/art/cli.py

@@ -1,5 +1,7 @@
 from fastapi import FastAPI, Body
 from fastapi.responses import StreamingResponse
+from fastapi import Request
+from fastapi.responses import JSONResponse
 import json
 import pydantic
 import socket
@@ -13,6 +15,7 @@
 from .trajectories import TrajectoryGroup
 from .types import TrainConfig
 from .utils.deploy_model import LoRADeploymentProvider
+from .errors import ARTError
 
 app = typer.Typer()
 
@@ -44,6 +47,12 @@ def __init__(self, *args: Any, **kwargs: Any) -> None:
 
     backend = LocalBackend()
     app = FastAPI()
+
+    # Add exception handler for ARTError
+    @app.exception_handler(ARTError)
+    async def art_error_handler(request: Request, exc: ARTError):
+        return JSONResponse(status_code=exc.status_code, content={"detail": exc.detail})
+
     app.get("/healthcheck")(lambda: {"status": "ok"})
     app.post("/close")(backend.close)
     app.post("/register")(backend.register)

src/art/dev/__init__.py

@@ -1,6 +1,5 @@
 from .engine import EngineArgs
 from .model import (
-    get_model_config,
     InternalModelConfig,
     InitArgs,
     PeftArgs,
@@ -12,7 +11,6 @@
 
 __all__ = [
     "EngineArgs",
-    "get_model_config",
     "InternalModelConfig",
     "InitArgs",
     "PeftArgs",

src/art/dev/get_model_config.py

@@ -0,0 +1,111 @@
+import torch
+
+from .engine import EngineArgs
+from .model import InitArgs, PeftArgs, TrainerArgs
+from .torchtune import TorchtuneArgs
+from .model import InternalModelConfig
+
+
+def get_model_config(
+    base_model: str,
+    output_dir: str,
+    config: "InternalModelConfig | None",
+) -> "InternalModelConfig":
+    from ..local.checkpoints import get_last_checkpoint_dir
+
+    if config is None:
+        config = InternalModelConfig()
+    enable_sleep_mode = config.get("engine_args", {}).get("enable_sleep_mode", True)
+    init_args = InitArgs(
+        model_name=base_model,
+        max_seq_length=32768,
+        load_in_4bit=True,  # False for LoRA 16bit
+        fast_inference=True,  # Enable vLLM fast inference
+        # vLLM args
+        disable_log_stats=False,
+        enable_prefix_caching=True,
+        gpu_memory_utilization=(
+            0.79 if enable_sleep_mode else 0.55
+        ),  # Reduce if out of memory
+        max_lora_rank=8,
+        use_async=True,
+    )
+    if config.get("_decouple_vllm_and_unsloth", False):
+        init_args["fast_inference"] = False
+        init_args.pop("disable_log_stats")
+        init_args.pop("enable_prefix_caching")
+        init_args.pop("gpu_memory_utilization")
+        init_args.pop("max_lora_rank")
+        init_args.pop("use_async")
+    engine_args = EngineArgs(
+        disable_log_requests=True,
+        # Multi-step processing is not supported for the Xformers attention backend
+        # which is the fallback for devices with compute capability < 8.0
+        num_scheduler_steps=(
+            16
+            if config.get("torchtune_args") is None
+            and not config.get("_decouple_vllm_and_unsloth", False)
+            and torch.cuda.get_device_capability()[0] >= 8
+            else 1
+        ),
+        enable_sleep_mode=enable_sleep_mode,
+        generation_config="vllm",
+    )
+    engine_args.update(config.get("engine_args", {}))
+    init_args.update(config.get("init_args", {}))
+    if last_checkpoint_dir := get_last_checkpoint_dir(output_dir):
+        init_args["model_name"] = last_checkpoint_dir
+        if config.get("torchtune_args") is not None:
+            engine_args["model"] = last_checkpoint_dir
+    elif config.get("torchtune_args") is not None:
+        engine_args["model"] = base_model
+    if config.get("_decouple_vllm_and_unsloth", False):
+        engine_args["model"] = base_model
+    peft_args = PeftArgs(
+        r=8,  # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
+        target_modules=[
+            "q_proj",
+            "k_proj",
+            "v_proj",
+            "o_proj",
+            "gate_proj",
+            "up_proj",
+            "down_proj",
+        ],  # Remove QKVO if out of memory
+        lora_alpha=16,
+        # Enable long context finetuning
+        use_gradient_checkpointing="unsloth",  # type: ignore
+        random_state=3407,
+    )
+    peft_args.update(config.get("peft_args", {}))
+    trainer_args = TrainerArgs(
+        learning_rate=5e-6,
+        adam_beta1=0.9,
+        adam_beta2=0.99,
+        weight_decay=0.1,
+        lr_scheduler_type="constant",
+        optim="paged_adamw_8bit",
+        logging_steps=1,
+        per_device_train_batch_size=2,
+        gradient_accumulation_steps=1,  # Increase to 4 for smoother training
+        num_generations=2,  # Decrease if out of memory
+        max_grad_norm=0.1,
+        save_strategy="no",
+        output_dir=output_dir,
+        disable_tqdm=True,
+        report_to="none",
+    )
+    trainer_args.update(config.get("trainer_args", {}))
+    if config.get("torchtune_args") is not None:
+        torchtune_args = TorchtuneArgs(model="qwen3_32b", model_type="QWEN3")
+        torchtune_args.update(config.get("torchtune_args", {}) or {})
+    else:
+        torchtune_args = None
+    return InternalModelConfig(
+        init_args=init_args,
+        engine_args=engine_args,
+        peft_args=peft_args,
+        trainer_args=trainer_args,
+        torchtune_args=torchtune_args,
+        _decouple_vllm_and_unsloth=config.get("_decouple_vllm_and_unsloth", False),
+    )