feat: Add vLLM provider.

.github/workflows/tests-integration.yaml

@@ -27,7 +27,7 @@ jobs:
 
       - name: Set expected providers for local integration tests
         id: set_local_providers
-        run: echo "expected_providers=ollama,llamacpp,llamafile" >> $GITHUB_OUTPUT
+        run: echo "expected_providers=ollama,llamacpp,llamafile,vllm" >> $GITHUB_OUTPUT
 
   determine-jobs-to-run:
     needs: expected-providers
@@ -144,6 +144,14 @@ jobs:
           restore-keys: |
             ${{ runner.os }}-llamafile-
 
+      - uses: actions/cache@v4
+        if: github.event.inputs.filter == '' || contains(github.event.inputs.filter, 'vllm')
+        with:
+          path: ~/.cache/huggingface
+          key: ${{ runner.os }}-vllm-models-${{ hashFiles('tests/conftest.py') }}
+          restore-keys: |
+            ${{ runner.os }}-vllm-models-
+
       - name: Setup Ollama
         if: github.event.inputs.filter == '' || contains(github.event.inputs.filter, 'ollama')
         uses: ai-action/setup-ollama@v1
@@ -192,6 +200,19 @@ jobs:
         run: |
           timeout 60 bash -c 'until curl -s http://localhost:8090/health >/dev/null; do sleep 1; done'
 
+      - name: Install and Run vLLM (CPU)
+        if: github.event.inputs.filter == '' || contains(github.event.inputs.filter, 'vllm')
+        run: |
+          uv pip install vllm --extra-index-url https://wheels.vllm.ai/nightly/cpu --prerelease=allow --index-strategy unsafe-best-match --torch-backend cpu
+          export VLLM_CPU_KVCACHE_SPACE=4
+          vllm serve Qwen/Qwen2.5-0.5B-Instruct --dtype bfloat16 --max-model-len 2048 --port 8080 > vllm.log 2>&1 &
+          echo $! > vllm.pid
+
+      - name: Wait for vLLM to be ready
+        if: github.event.inputs.filter == '' || contains(github.event.inputs.filter, 'vllm')
+        run: |
+          timeout 600 bash -c 'until curl -s http://localhost:8080/v1/models >/dev/null; do echo "Waiting for vLLM..."; sleep 5; done' || (echo "=== vLLM logs ===" && cat vllm.log && exit 1)
+
       - name: Run Local Provider Integration tests
         env:
           INCLUDE_LOCAL_PROVIDERS: "true"
@@ -211,6 +232,14 @@ jobs:
             rm llamafile.pid
           fi
 
+      - name: Cleanup vLLM process
+        if: always()
+        run: |
+          if [ -f vllm.pid ]; then
+            kill $(cat vllm.pid) || true
+            rm vllm.pid
+          fi
+
       - name: Upload coverage reports to Codecov
         if: always()
         uses: codecov/codecov-action@v5

pyproject.toml

@@ -18,7 +18,7 @@ dependencies = [
 [project.optional-dependencies]
 
 all = [
-  "any-llm-sdk[mistral,anthropic,huggingface,gemini,vertexai,cohere,cerebras,fireworks,groq,bedrock,azure,azureopenai,watsonx,together,sambanova,ollama,moonshot,nebius,xai,databricks,deepseek,inception,openai,openrouter,portkey,lmstudio,llama,voyage,perplexity,platform,llamafile,llamacpp,sagemaker,gateway,zai,minimax]"
+  "any-llm-sdk[mistral,anthropic,huggingface,gemini,vertexai,cohere,cerebras,fireworks,groq,bedrock,azure,azureopenai,watsonx,together,sambanova,ollama,moonshot,nebius,xai,databricks,deepseek,inception,openai,openrouter,portkey,lmstudio,llama,voyage,perplexity,platform,llamafile,llamacpp,sagemaker,gateway,zai,minimax,vllm]"
 ]
 
 platform = [
@@ -109,6 +109,7 @@ openrouter = []
 portkey = []
 sambanova = []
 minimax = []
+vllm = []
 gateway = [
   "fastapi>=0.115.0",
   "uvicorn[standard]>=0.30.0",

src/any_llm/constants.py

@@ -45,6 +45,7 @@ class LLMProvider(StrEnum):
     SAGEMAKER = "sagemaker"
     TOGETHER = "together"
     VERTEXAI = "vertexai"
+    VLLM = "vllm"
     VOYAGE = "voyage"
     WATSONX = "watsonx"
     XAI = "xai"

src/any_llm/providers/vllm/__init__.py

@@ -0,0 +1,3 @@
+from .vllm import VllmProvider
+
+__all__ = ["VllmProvider"]

src/any_llm/providers/vllm/vllm.py

@@ -0,0 +1,18 @@
+from any_llm.providers.openai.base import BaseOpenAIProvider
+
+
+class VllmProvider(BaseOpenAIProvider):
+    API_BASE = "http://localhost:8000/v1"
+    ENV_API_KEY_NAME = "VLLM_API_KEY"
+    PROVIDER_NAME = "vllm"
+    PROVIDER_DOCUMENTATION_URL = "https://docs.vllm.ai/"
+
+    SUPPORTS_EMBEDDING = True
+    SUPPORTS_COMPLETION_REASONING = True
+    SUPPORTS_COMPLETION_STREAMING = True
+    SUPPORTS_COMPLETION_PDF = False
+
+    def _verify_and_set_api_key(self, api_key: str | None = None) -> str | None:
+        # vLLM server by default doesn't require an API key
+        # but can be configured to use one via --api-key flag
+        return api_key or ""

tests/conftest.py

@@ -23,6 +23,7 @@ def provider_reasoning_model_map() -> dict[LLMProvider, str]:
         LLMProvider.OPENROUTER: "google/gemini-2.5-flash-lite",
         LLMProvider.LLAMAFILE: "N/A",
         LLMProvider.LLAMACPP: "N/A",
+        LLMProvider.VLLM: "N/A",
         LLMProvider.LMSTUDIO: "openai/gpt-oss-20b",  # You must have LM Studio running and the server enabled
         LLMProvider.AZUREOPENAI: "azure/<your_deployment_name>",
         LLMProvider.CEREBRAS: "gpt-oss-120b",
@@ -60,6 +61,7 @@ def provider_model_map() -> dict[LLMProvider, str]:
         LLMProvider.OLLAMA: "llama3.2:1b",
         LLMProvider.LLAMAFILE: "N/A",
         LLMProvider.LMSTUDIO: "google/gemma-3n-e4b",  # You must have LM Studio running and the server enabled
+        LLMProvider.VLLM: "Qwen/Qwen2.5-0.5B-Instruct",
         LLMProvider.COHERE: "command-a-03-2025",
         LLMProvider.CEREBRAS: "llama-3.3-70b",
         LLMProvider.HUGGINGFACE: "huggingface/tgi",  # This is the syntax used in `litellm` when using HF Inference Endpoints (https://docs.litellm.ai/docs/providers/huggingface#dedicated-inference-endpoints)
@@ -131,6 +133,7 @@ def provider_client_config() -> dict[LLMProvider, dict[str, Any]]:
         LLMProvider.OPENAI: {"timeout": 100},
         LLMProvider.HUGGINGFACE: {"api_base": "https://oze7k8n86bjfzgjk.us-east-1.aws.endpoints.huggingface.cloud/v1"},
         LLMProvider.LLAMACPP: {"api_base": "http://127.0.0.1:8090/v1"},
+        LLMProvider.VLLM: {"api_base": "http://127.0.0.1:8080/v1"},
         LLMProvider.MISTRAL: {"timeout_ms": 100000},
         LLMProvider.NEBIUS: {"api_base": "https://api.studio.nebius.com/v1/"},
         LLMProvider.OPENAI: {"timeout": 10},

tests/constants.py

@@ -8,6 +8,7 @@
     LLMProvider.LMSTUDIO,
     LLMProvider.LLAMAFILE,
     LLMProvider.GATEWAY,
+    LLMProvider.VLLM,
 ]
 
 EXPECTED_PROVIDERS = os.environ.get("EXPECTED_PROVIDERS", "").split(",")

tests/unit/providers/test_vllm_provider.py

@@ -0,0 +1,13 @@
+from any_llm.providers.vllm.vllm import VllmProvider
+
+
+def test_provider_without_api_key() -> None:
+    provider = VllmProvider()
+    assert provider.PROVIDER_NAME == "vllm"
+    assert provider.API_BASE == "http://localhost:8000/v1"
+    assert provider.ENV_API_KEY_NAME == "VLLM_API_KEY"
+
+
+def test_provider_with_api_key() -> None:
+    provider = VllmProvider(api_key="test-api-key")
+    assert provider.PROVIDER_NAME == "vllm"

tests/unit/test_provider.py

@@ -147,6 +147,7 @@ def test_providers_raise_MissingApiKeyError(provider: LLMProvider) -> None:
         LLMProvider.OLLAMA,
         LLMProvider.SAGEMAKER,
         LLMProvider.VERTEXAI,
+        LLMProvider.VLLM,
         LLMProvider.GATEWAY,
     ):
         pytest.skip("This provider handles `api_key` differently.")