From c9ca2096b779b80ca65f0e30ace1899d582029b2 Mon Sep 17 00:00:00 2001
From: feng lui <3090641@qq.com>
Date: Mon, 12 May 2025 20:33:42 +0800
Subject: [PATCH] vLLM Windows CUDA support [tested] (#2158)

* Update loader.py

change vllm installed check by transformers utils function

* Update llama.py

change vllm installed check by transformers utils function

* add sample notebook

* fix Indentation

* add global is_vLLM_available function

* Pythonic style

* Delete nb/Qwen2.5_(3B)-GRPO-windows.ipynb

Would be great to move it to https://github.com/unslothai/notebooks - appreciate it!

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---
 unsloth/models/__init__.py | 4 ++--
 unsloth/models/_utils.py   | 4 ++++
 unsloth/models/llama.py    | 5 ++---
 unsloth/models/loader.py   | 6 +++---
 4 files changed, 11 insertions(+), 8 deletions(-)

diff --git a/unsloth/models/__init__.py b/unsloth/models/__init__.py
index 99db55c086..fb3fc5c455 100644
--- a/unsloth/models/__init__.py
+++ b/unsloth/models/__init__.py
@@ -20,5 +20,5 @@ from .qwen3     import FastQwen3Model
 from .qwen3_moe import FastQwen3MoeModel
 from .granite   import FastGraniteModel
 from .dpo       import PatchDPOTrainer, PatchKTOTrainer
-from ._utils    import is_bfloat16_supported, __version__
-from .rl        import PatchFastRL, vLLMSamplingParams
+from ._utils import is_bfloat16_supported, is_vLLM_available, __version__
+from .rl        import PatchFastRL, vLLMSamplingParams
\ No newline at end of file
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 82a1d02fce..4b9e2cce06 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -17,6 +17,7 @@ __version__ = "2025.5.1"
 __all__ = [
     "SUPPORTS_BFLOAT16",
     "is_bfloat16_supported",
+    "is_vLLM_available",
 
     "prepare_model_for_kbit_training",
     "xformers",
@@ -800,6 +801,9 @@ def is_bfloat16_supported():
     return SUPPORTS_BFLOAT16
 pass
 
+def is_vLLM_available():
+    return _is_package_available("vllm")
+pass
 
 # Patches models to add RoPE Scaling
 def patch_linear_scaling(
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 2145571ae1..207540abf5 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -1661,9 +1661,8 @@ class FastLlamaModel:
             )
         pass
         if fast_inference:
-            import platform
-            if platform.system().lower() == 'windows':
-                print("Unsloth: vLLM does not work in Windows! Will use Unsloth inference!")
+            if not is_vLLM_available():
+                print("Unsloth: vLLM is not installed! Will use Unsloth inference!")
                 fast_inference = False
             major_version, minor_version = torch.cuda.get_device_capability()
             if major_version < 7:
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 7e904471bc..c93c5d2742 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -14,6 +14,7 @@
 
 from ._utils import (
     is_bfloat16_supported,
+    is_vLLM_available,
     HAS_FLASH_ATTENTION,
     HAS_FLASH_ATTENTION_SOFTCAPPING,
     USE_MODELSCOPE,
@@ -351,9 +352,8 @@ class FastLanguageModel(FastLlamaModel):
         pass
 
         if fast_inference:
-            import platform
-            if platform.system().lower() == 'windows':
-                print("Unsloth: vLLM does not work in Windows! Will use Unsloth inference!")
+            if not is_vLLM_available():
+                print("Unsloth: vLLM is not installed! Will use Unsloth inference!")
                 fast_inference = False
             pass
             from unsloth_zoo.vllm_utils import (