feat: sglang backend support

1 files changed, 57 insertions, 2 deletions

diff --git a/app/backends/probe.py b/app/backends/probe.py
index 818aadb..d54a9f3 100644
--- a/app/backends/probe.py
+++ b/app/backends/probe.py
@@ -26,7 +26,7 @@ it stays cheap to import alongside ``backends.common``.
 import json
 import urllib.parse
 import urllib.request
-from typing import Optional
+from typing import List, Optional
 
 from backends import common
 
@@ -35,6 +35,7 @@ IDENTITY_FASTER = "faster"
 IDENTITY_QWEN_CUSTOM = "qwen-custom"
 IDENTITY_QWEN_CLONE = "qwen-clone"
 IDENTITY_QWEN_DESIGN = "qwen-design"
+IDENTITY_SGLOMNI = "sglomni"
 
 # Endpoint names the converter resolves for each qwen demo server (see
 # converter.clients QwenTTSClient). Mirror them here so identification matches
@@ -77,7 +78,7 @@ def _get_json(url: str, timeout: float) -> Optional[dict]:
 
 
 def _identify_health(base: str, timeout: float) -> Optional[str]:
-    """Identify audio.cpp / faster from their ``/health`` responses."""
+    """Identify audio.cpp / faster / sglang-omni from their /health responses."""
     payload = _get_json(f"{base}/health", timeout)
     if payload is None:
         return None
@@ -93,6 +94,15 @@ def _identify_health(base: str, timeout: float) -> Optional[str]:
         if isinstance(entries, list) and entries \
                 and any(isinstance(e, dict) and e.get("id") for e in entries):
             return IDENTITY_AUDIOCPP
+    # sglang-omni's /health reports {"status": "healthy", "stages": [...]}
+    # (200 when serving, 503 with "unhealthy" while booting). A 503 body
+    # still parses as JSON here, so require the healthy word explicitly —
+    # an "unhealthy" sgl-omni must not count as usable. The pipeline
+    # "stages" list is confirmed as a secondary mark (present on every
+    # sgl-omni 0.1.x server) before trusting the generic-sounding status.
+    if payload.get("status") == "healthy" \
+            and isinstance(payload.get("stages"), list):
+        return IDENTITY_SGLOMNI
     return None
 
 
@@ -184,3 +194,48 @@ def faster_model_loaded(url: str, timeout: float = DEFAULT_TIMEOUT) -> bool:
     """
     payload = health_payload(url, timeout)
     return bool(payload and payload.get("model_loaded"))
+
+
+def sglomni_served_model(url: str,
+                         timeout: float = DEFAULT_TIMEOUT) -> Optional[str]:
+    """The HuggingFace repo id a sglang-omni server at URL hosts, or None.
+
+    ``GET /v1/models`` answers ``{"data": [{"id": <served repo>}, ...]}``
+    with exactly one entry (one model per server process) — the same
+    model-identity role qwen's probe plays for its three demos. Used to
+    name the running model in statuses and to decide when a managed
+    server must be restarted to host the model a run selected.
+    """
+    if not url:
+        return None
+    models = _get_json(f"{url.rstrip('/')}/v1/models", timeout)
+    if models is None:
+        return None
+    entries = models.get("data")
+    if isinstance(entries, list) and entries \
+            and isinstance(entries[0], dict):
+        model_id = entries[0].get("id")
+        if isinstance(model_id, str) and model_id:
+            return model_id
+    return None
+
+
+def sglomni_voice_names(url: str,
+                        timeout: float = DEFAULT_TIMEOUT) -> Optional[List[str]]:
+    """The uploaded voice names registered on a sglang-omni server, or None.
+
+    ``GET /v1/audio/voices?names_only=true`` answers
+    ``{"uploaded_voice_names": [...]}`` — the server-side voices a remote
+    Convert form can offer in its voice picker (uploaded clips persist
+    across server restarts). None when the URL does not answer.
+    """
+    if not url:
+        return None
+    payload = _get_json(f"{url.rstrip('/')}/v1/audio/voices?names_only=true",
+                        timeout)
+    if payload is None:
+        return None
+    names = payload.get("uploaded_voice_names")
+    if isinstance(names, list):
+        return [str(name) for name in names if isinstance(name, str) and name]
+    return None