llama_cpp server: slight refactor to init_llama function

Define an init_llama function that starts llama with supplied settings instead of just doing it in the global context of app.py

This allows the test to be less brittle by not needing to mess with os.environ, then importing the app

llama_cpp/server/__main__.py

@@ -24,9 +24,10 @@ Then visit http://localhost:8000/docs to see the interactive API docs.
 import os
 import uvicorn
 
-from llama_cpp.server.app import app
+from llama_cpp.server.app import app, init_llama
 
 if __name__ == "__main__":
+    init_llama()
 
     uvicorn.run(
         app, host=os.getenv("HOST", "localhost"), port=int(os.getenv("PORT", 8000))

llama_cpp/server/app.py

@@ -13,7 +13,7 @@ from sse_starlette.sse import EventSourceResponse
 
 
 class Settings(BaseSettings):
-    model: str = os.environ["MODEL"]
+    model: str = os.environ.get("MODEL", "null")
     n_ctx: int = 2048
     n_batch: int = 512
     n_threads: int = max((os.cpu_count() or 2) // 2, 1)
@@ -38,31 +38,34 @@ app.add_middleware(
     allow_methods=["*"],
     allow_headers=["*"],
 )
-settings = Settings()
+
-llama = llama_cpp.Llama(
+llama: llama_cpp.Llama = None
-    settings.model,
+def init_llama(settings: Settings = None):
-    f16_kv=settings.f16_kv,
+    if settings is None:
-    use_mlock=settings.use_mlock,
+        settings = Settings()
-    use_mmap=settings.use_mmap,
+    global llama
-    embedding=settings.embedding,
+    llama = llama_cpp.Llama(
-    logits_all=settings.logits_all,
+        settings.model,
-    n_threads=settings.n_threads,
+        f16_kv=settings.f16_kv,
-    n_batch=settings.n_batch,
+        use_mlock=settings.use_mlock,
-    n_ctx=settings.n_ctx,
+        use_mmap=settings.use_mmap,
-    last_n_tokens_size=settings.last_n_tokens_size,
+        embedding=settings.embedding,
-    vocab_only=settings.vocab_only,
+        logits_all=settings.logits_all,
-)
+        n_threads=settings.n_threads,
-if settings.cache:
+        n_batch=settings.n_batch,
-    cache = llama_cpp.LlamaCache()
+        n_ctx=settings.n_ctx,
-    llama.set_cache(cache)
+        last_n_tokens_size=settings.last_n_tokens_size,
+        vocab_only=settings.vocab_only,
+    )
+    if settings.cache:
+        cache = llama_cpp.LlamaCache()
+        llama.set_cache(cache)
+
 llama_lock = Lock()
-
-
 def get_llama():
     with llama_lock:
         yield llama
 
-
 class CreateCompletionRequest(BaseModel):
     prompt: Union[str, List[str]]
     suffix: Optional[str] = Field(None)

tests/test_llama.py

@@ -132,10 +132,11 @@ def test_utf8(monkeypatch):
 
 def test_llama_server():
     from fastapi.testclient import TestClient
-    import os
+    from llama_cpp.server.app import app, init_llama, Settings
-    os.environ["MODEL"] = MODEL
+    s = Settings()
-    os.environ["VOCAB_ONLY"] = "true"
+    s.model = MODEL
-    from llama_cpp.server.app import app
+    s.vocab_only = True
+    init_llama(s)
     client = TestClient(app)
     response = client.get("/v1/models")
     assert response.json() == {