fix: model-specific instructions support and set cfg-scale 4 for breeze automaticallyHEADmain

2 files changed, 25 insertions, 2 deletions

diff --git a/app/backends/probe.py b/app/backends/probe.py
index d54a9f3..c84c88d 100644
--- a/app/backends/probe.py
+++ b/app/backends/probe.py
@@ -107,8 +107,15 @@ def _identify_health(base: str, timeout: float) -> Optional[str]:
 
 
 def _identify_gradio(base: str, timeout: float) -> Optional[str]:
-    """Identify a qwen-tts Gradio demo from its ``/info`` named endpoints."""
-    payload = _get_json(f"{base}/info", timeout)
+    """Identify a qwen-tts Gradio demo from its ``/info`` named endpoints.
+
+    Modern Gradio (>= 4.x / 5.x) routes its API under ``/gradio_api`` —
+    its ``/info`` lives at ``/gradio_api/info`` with the legacy ``/info``
+    path gone or deprecated — so both prefixes are probed.
+    """
+    payload = _get_json(f"{base}/gradio_api/info", timeout)
+    if payload is None:
+        payload = _get_json(f"{base}/info", timeout)
     if payload is None:
         return None
     endpoints = payload.get("named_endpoints")
diff --git a/app/backends/sglomni/catalog.py b/app/backends/sglomni/catalog.py
index 8bfb1fa..2452d82 100644
--- a/app/backends/sglomni/catalog.py
+++ b/app/backends/sglomni/catalog.py
@@ -52,6 +52,16 @@ class ModelEntry:
     system_hint: Optional[str] = None  # remediation when the binary is absent
     speakers: Optional[Tuple[str, ...]] = None  # preset voices (speaker)
     supports_seed: bool = False  # request-scoped seed accepted (Qwen3-TTS Base)
+    # Whether the model's serving pipeline consumes a separate
+    # "instructions" field alongside its normal voice conditioning
+    # (verified against the installed sglang_omni code, not assumed from
+    # the HTTP schema): Qwen3-TTS Base (clone + instruction conditioning
+    # in request_builders.py), Qwen3-TTS CustomVoice and VoiceDesign, and
+    # the MOSS v1.5 pair (reference + instruction in the user message).
+    # False would mean an instructions field is silently ignored (Higgs,
+    # Voxtral, fish, dots, ZONOS2 take none; fish's inline event tags
+    # belong in the text, not this field).
+    supports_instructions: bool = False
     # NOTE(unverified upstream): only the two Base entries are known to
     # accept a request-scoped seed (Voxtral rejects one outright); qwen's
     # demo client does send seeds to the CustomVoice/VoiceDesign models,
@@ -161,6 +171,7 @@ ENTRIES: Tuple[ModelEntry, ...] = (
         requires_reference=False,
         extras=_QWEN_EXTRAS, system_dep="sox", system_hint=_SOX_HINT,
         speakers=QWEN_CUSTOMVOICE_SPEAKERS,
+        supports_instructions=True,
         notes="built-in speakers, lightest model",
     ),
     ModelEntry(
@@ -172,6 +183,7 @@ ENTRIES: Tuple[ModelEntry, ...] = (
         requires_reference=True,
         extras=_QWEN_EXTRAS, system_dep="sox", system_hint=_SOX_HINT,
         supports_seed=True,
+        supports_instructions=True,
         notes="voice cloning from a reference clip",
     ),
     ModelEntry(
@@ -183,6 +195,7 @@ ENTRIES: Tuple[ModelEntry, ...] = (
         requires_reference=True,
         extras=_QWEN_EXTRAS, system_dep="sox", system_hint=_SOX_HINT,
         supports_seed=True,
+        supports_instructions=True,
         notes="voice cloning, higher quality",
     ),
     ModelEntry(
@@ -193,6 +206,7 @@ ENTRIES: Tuple[ModelEntry, ...] = (
         capability=CAPABILITY_DESIGN,
         requires_reference=False,
         extras=_QWEN_EXTRAS, system_dep="sox", system_hint=_SOX_HINT,
+        supports_instructions=True,
         notes="voice described by instructions",
     ),
     ModelEntry(
@@ -229,6 +243,7 @@ ENTRIES: Tuple[ModelEntry, ...] = (
         config="moss_tts.yaml",
         capability=CAPABILITY_CLONE,
         requires_reference=True,
+        supports_instructions=True,
         notes="voice cloning from a reference clip",
     ),
     ModelEntry(
@@ -238,6 +253,7 @@ ENTRIES: Tuple[ModelEntry, ...] = (
         config="moss_tts_local.yaml",
         capability=CAPABILITY_CLONE,
         requires_reference=False,
+        supports_instructions=True,
         notes="48 kHz, narration without a reference or cloning",
     ),
     ModelEntry(