feat: add llama-swap as a backend

test/test_choose_endpoint.py

@@ -12,10 +12,11 @@ EP3 = "http://ep3:11434"
 LLAMA_EP = "http://llama:8080/v1"
 
 
-def _make_cfg(endpoints, llama_eps=None, max_conn=2, endpoint_config=None, priority_routing=False):
+def _make_cfg(endpoints, llama_eps=None, swap_eps=None, max_conn=2, endpoint_config=None, priority_routing=False):
     cfg = MagicMock()
     cfg.endpoints = endpoints
     cfg.llama_server_endpoints = llama_eps or []
+    cfg.llama_swap_endpoints = swap_eps or []
     cfg.api_keys = {}
     cfg.max_concurrent_connections = max_conn
     cfg.endpoint_config = endpoint_config or {}
@@ -46,6 +47,27 @@ class TestChooseEndpointBasic:
         assert ep == EP1
         assert tracking == "llama3.2:latest"
 
+    async def test_llama_swap_endpoint_is_a_candidate(self):
+        swap_ep = "http://swap:8080/v1"
+        cfg = _make_cfg([EP1], swap_eps=[swap_ep])
+
+        async def available(ep, *_):
+            # Only the llama-swap backend advertises this model
+            return {"org/model:Q4_K_M"} if ep == swap_ep else set()
+
+        async def loaded(ep):
+            return {"org/model:Q4_K_M"} if ep == swap_ep else set()
+
+        with (
+            patch.object(router, "config", cfg),
+            patch.object(router.fetch, "available_models", side_effect=available),
+            patch.object(router.fetch, "loaded_models", side_effect=loaded),
+        ):
+            ep, tracking = await router.choose_endpoint("org/model:Q4_K_M")
+        assert ep == swap_ep
+        # llama-swap models are tracked under their normalized name
+        assert tracking == "model"
+
     async def test_raises_when_no_endpoint_has_model(self):
         cfg = _make_cfg([EP1, EP2])
         with (