📌 Pin vLLM version

docs/source/speeding_up_training.md

@@ -14,13 +14,7 @@ To speed up generation, you can use [vLLM](https://github.com/vllm-project/vllm)
 To use [vLLM](https://github.com/vllm-project/vllm), first install it using:
 
 ```bash
-pip install vllm
-```
-
-or 
-
-```bash
-pip install "trl[vllm]"
+pip install trl[vllm]
 ```
 
 <hfoptions id="vllm examples">

docs/source/vllm_integration.md

@@ -2,6 +2,12 @@
 
 This document will guide you through the process of using vLLM with TRL for faster generation in online methods like GRPO and Online DPO. We first summarize a tl;dr on how to use vLLM with TRL, and then we will go into the details of how it works under the hood. Let's go! 🔥
 
+<Tip warning={true}>
+
+TRL currently only supports vLLM versions `0.10.0`, `0.10.1`, and `0.10.2`. Please ensure you have one of these versions installed to avoid compatibility issues.
+
+</Tip>
+
 ## 🚀 How can I use vLLM with TRL to speed up training?
 
 💡 **Note**: Resources required for this specific example: a single node with 8 GPUs.

examples/scripts/evals/judge_tldr.py

@@ -14,8 +14,7 @@
 
 # /// script
 # dependencies = [
-#     "trl",
-#     "vllm",
+#     "trl[vllm]",
 # ]
 # ///
 

examples/scripts/rloo.py

@@ -14,12 +14,11 @@
 
 # /// script
 # dependencies = [
-#     "trl",
+#     "trl[vllm]",
 #     "peft",
 #     "math-verify",
 #     "latex2sympy2_extended",
 #     "trackio",
-#     "vllm",
 #     "kernels",
 # ]
 # ///

setup.cfg

@@ -64,7 +64,7 @@ test =
 vllm =
     # vLLM package does not yet support Python 3.13. These constraints can be lifted once support is added:
     # see https://github.com/vllm-project/vllm/pull/13164
-    vllm>=0.10.0; python_version < "3.13"
+    vllm>=0.10.0,<=0.10.2; python_version < "3.13"
     fastapi; python_version < "3.13"
     pydantic; python_version < "3.13"
     requests; python_version < "3.13"

trl/extras/vllm_client.py

@@ -114,7 +114,7 @@ def __init__(
         if not is_requests_available():
             raise ImportError("requests is not installed. Please install it with `pip install requests`.")
         if not is_vllm_available():
-            raise ImportError("vLLM is not installed. Please install it with `pip install vllm`.")
+            raise ImportError("vLLM is not installed. Please install it with `pip install trl[vllm]`.")
 
         self.session = requests.Session()
 

trl/import_utils.py

@@ -14,6 +14,7 @@
 
 import importlib
 import os
+import warnings
 from itertools import chain
 from types import ModuleType
 from typing import Any
@@ -35,7 +36,7 @@
 _requests_available = _is_package_available("requests")
 _unsloth_available = _is_package_available("unsloth")
 _uvicorn_available = _is_package_available("uvicorn")
-_vllm_available = _is_package_available("vllm")
+_vllm_available, _vllm_version = _is_package_available("vllm", return_version=True)
 _vllm_ascend_available = _is_package_available("vllm_ascend")
 _weave_available = _is_package_available("weave")
 
@@ -81,6 +82,15 @@ def is_uvicorn_available() -> bool:
 
 
 def is_vllm_available() -> bool:
+    if _vllm_available and (
+        version.parse(_vllm_version) < version.parse("0.10.0")
+        or version.parse(_vllm_version) > version.parse("0.10.2")
+    ):
+        warnings.warn(
+            "TRL currently only supports vLLM versions `0.10.0`, `0.10.1`, and `0.10.2`. You have version "
+            f"{_vllm_version} installed. We recommend to install one of these versions to avoid compatibility issues.",
+            UserWarning,
+        )
     return _vllm_available
 
 

trl/trainer/grpo_trainer.py

@@ -479,7 +479,7 @@ def __init__(
             if not is_vllm_available():
                 raise ImportError(
                     "vLLM is not available and `use_vllm` is set to True. Please install vLLM with "
-                    "`pip install vllm` to use it."
+                    "`pip install [vllm]` to use it."
                 )
 
             if self.vllm_mode == "server":

trl/trainer/online_dpo_config.py

@@ -293,7 +293,7 @@ class may differ from those in [`~transformers.TrainingArguments`].
         default=False,
         metadata={
             "help": "Whether to use vLLM for generating completions. Requires vLLM to be installed "
-            "(`pip install vllm`)."
+            "(`pip install trl[vllm]`)."
         },
     )
     vllm_model_impl: str = field(

trl/trainer/online_dpo_trainer.py

@@ -483,7 +483,7 @@ def __init__(
             if not is_vllm_available():
                 raise ImportError(
                     "vLLM is not available and `use_vllm` is set to True. Please install vLLM with "
-                    "`pip install vllm` to use it."
+                    "`pip install trl[vllm]` to use it."
                 )
 
             if self.vllm_mode == "server":

trl/trainer/rloo_trainer.py

@@ -551,7 +551,7 @@ def decode(example, tokenizer):
             if not is_vllm_available():
                 raise ImportError(
                     "vLLM is not available and `use_vllm` is set to True. Please install vLLM with "
-                    "`pip install vllm` to use it."
+                    "`pip install trl[vllm]` to use it."
                 )
 
             if self.vllm_mode == "server":