Fix compatibility with vLLM 0.10.x

MultiStepModelRunner was removed in vLLM 0.10.x and merged into ModelRunner,
which now has built-in LoRA support. This commit updates the code to handle
both old (<0.10) and new (>=0.10) versions of vLLM:

- Make MultiStepModelRunner import optional with try/except
- Update all isinstance checks to handle None case
- Convert patch_multi_step_model_runner to a no-op for vLLM 0.10.x
- Add deprecation notes explaining the changes

This fixes ModuleNotFoundError when using vLLM 0.10.2.

Author: benediktstroebl

src/art/unsloth/state.py

@@ -17,9 +17,14 @@
 from trl import GRPOConfig, GRPOTrainer
 from vllm.engine.arg_utils import AsyncEngineArgs
 from vllm.engine.async_llm_engine import AsyncLLMEngine
-from vllm.worker.multi_step_model_runner import MultiStepModelRunner
 from vllm.worker.worker_base import WorkerWrapperBase
 
+try:
+    from vllm.worker.multi_step_model_runner import MultiStepModelRunner
+except ImportError:
+    # MultiStepModelRunner was removed in vLLM 0.10.x
+    MultiStepModelRunner = None  # type: ignore
+
 from ..dev.model import InternalModelConfig
 from .train import gc_and_empty_cuda_cache
 
@@ -41,8 +46,8 @@ class ModelState:
     def __init__(self, config: InternalModelConfig) -> None:
         from vllm.engine import async_llm_engine
 
-        # Patch MultiStepModelRunner for Unsloth compatibility
-        if not hasattr(MultiStepModelRunner, "model"):
+        # Patch MultiStepModelRunner for Unsloth compatibility (vLLM < 0.10.x)
+        if MultiStepModelRunner is not None and not hasattr(MultiStepModelRunner, "model"):
             MultiStepModelRunner.model = property(  # type: ignore
                 lambda self: self._base_model_runner.model
             )
@@ -143,7 +148,10 @@ def __init__(self, async_engine: AsyncLLMEngine, enable_sleep_mode: bool) -> Non
             "WorkerWrapperBase",
             getattr(self.async_engine.engine.model_executor, "driver_worker"),
         )
-        if isinstance(self.driver_worker.model_runner, MultiStepModelRunner):
+        # Patch MultiStepModelRunner if it exists (vLLM < 0.10.x)
+        if MultiStepModelRunner is not None and isinstance(
+            self.driver_worker.model_runner, MultiStepModelRunner
+        ):
             patch_multi_step_model_runner(self.driver_worker.model_runner)
 
     @asynccontextmanager

src/art/vllm/patches.py

@@ -1,10 +1,12 @@
 """Monkey patches and modifications for vLLM."""
 
 import ctypes
-from typing import Any
+from typing import Any, TYPE_CHECKING
 
 import torch
-from vllm.worker.multi_step_model_runner import MultiStepModelRunner
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import ModelRunner
 
 
 def patch_allocator() -> None:
@@ -185,13 +187,13 @@ def patch(
     ToolParserManager.get_tool_parser = patched_get_tool_parser
 
 
-def patch_multi_step_model_runner(runner: MultiStepModelRunner) -> None:
+def patch_multi_step_model_runner(runner: "ModelRunner") -> None:
     """
     Patches a MultiStepModelRunner to support LoRA adapters.
+
+    Note: This function is deprecated as of vLLM 0.10.x. MultiStepModelRunner
+    was merged into ModelRunner, which now has built-in LoRA support.
+    This function is kept for backward compatibility but does nothing.
     """
-    base_runner = runner._base_model_runner
-    runner.set_active_loras = base_runner.set_active_loras
-    runner.add_lora = base_runner.add_lora
-    runner.remove_lora = base_runner.remove_lora
-    runner.pin_lora = base_runner.pin_lora
-    runner.list_loras = base_runner.list_loras
+    # No-op: ModelRunner in vLLM 0.10.x already has LoRA methods built-in
+    pass