feat: add llama-swap as a backend

backends/probe.py

@@ -46,7 +46,7 @@ from backends.health import (
     _format_connection_issue,
     _is_llama_model_loaded,
 )
-from backends.normalize import is_ext_openai_endpoint, is_openai_compatible
+from backends.normalize import is_ext_openai_endpoint, is_openai_compatible, is_llama_server, is_llama_swap
 
 
 class fetch:
@@ -61,10 +61,10 @@ class fetch:
             headers["Authorization"] = "Bearer " + api_key
 
         ep_base = endpoint.rstrip("/")
-        if endpoint in cfg.llama_server_endpoints and "/v1" not in endpoint:
+        if is_llama_server(endpoint) and "/v1" not in endpoint:
             endpoint_url = f"{ep_base}/v1/models"
             key = "data"
-        elif "/v1" in endpoint or endpoint in cfg.llama_server_endpoints:
+        elif "/v1" in endpoint or is_llama_server(endpoint):
             endpoint_url = f"{ep_base}/models"
             key = "data"
         else:
@@ -194,6 +194,38 @@ class fetch:
         client: aiohttp.ClientSession = get_probe_session(endpoint)
         cfg = get_config()
 
+        # llama-swap: loaded/running workers are reported at /running (state == "ready"),
+        # NOT via a status field on /v1/models (which it omits). /running is a root route,
+        # so strip any /v1 suffix from the configured endpoint.
+        if is_llama_swap(endpoint):
+            base_url = endpoint.rstrip("/").removesuffix("/v1")
+            headers = {"Referer": default_headers.get("HTTP-Referer", "https://nomyo.ai")}
+            api_key = cfg.api_keys.get(endpoint)
+            if api_key is not None:
+                headers["Authorization"] = "Bearer " + api_key
+            try:
+                async with client.get(f"{base_url}/running", headers=headers) as resp:
+                    await _ensure_success(resp)
+                    data = await resp.json()
+
+                models = {
+                    item.get("model")
+                    for item in data.get("running", [])
+                    if item.get("model") and item.get("state") == "ready"
+                }
+
+                async with _loaded_models_cache_lock:
+                    _loaded_models_cache[endpoint] = (models, time.time())
+                async with _loaded_error_cache_lock:
+                    _loaded_error_cache.pop(endpoint, None)
+                return models
+            except Exception as e:
+                message = _format_connection_issue(f"{base_url}/running", e)
+                print(f"[fetch.loaded_models] {message}")
+                async with _loaded_error_cache_lock:
+                    _loaded_error_cache[endpoint] = time.time()
+                return set()
+
         # Check if this is a llama-server endpoint
         if endpoint in cfg.llama_server_endpoints:
             # Query /v1/models for llama-server. Send the configured key as a