feat: sglang backend support

1 files changed, 112 insertions, 23 deletions

diff --git a/audiobook.py b/audiobook.py
index b881b6e..88e8dca 100755
--- a/audiobook.py
+++ b/audiobook.py
@@ -18,6 +18,7 @@ externally-run server and never touches server state.
 """
 
 import argparse
+import logging
 import sys
 import traceback
 from pathlib import Path
@@ -53,6 +54,7 @@ from converter.clients import (
     BACKEND_AUDIOCPP,
     BACKEND_FASTER,
     BACKEND_QWEN,
+    BACKEND_SGLOMNI,
     VOICE_MODE_CLONE,
     VOICE_MODE_CUSTOM,
     VOICE_MODE_DESIGN,
@@ -363,6 +365,23 @@ def convert(backend: str, voice: str = None, clone: str = None,
         voice_mode = VOICE_MODE_CLONE
     elif backend == BACKEND_AUDIOCPP:
         voice_mode = VOICE_MODE_CLONE if voice else VOICE_MODE_CUSTOM
+    elif backend == BACKEND_SGLOMNI:
+        # Resolve the model first (the capability decides the voice mode):
+        # managed runs need the weights on disk, remote runs accept any
+        # catalog key. The same resolution runs again in the converter.
+        from backends.sglomni import models as sg_models
+        from backends.sglomni.catalog import entry_by_key
+        if api_url is None:
+            sg_entry = sg_models.resolve_model(model_id)
+        else:
+            sg_entry = entry_by_key((model_id or "").strip())
+            if sg_entry is None:
+                raise RuntimeError(
+                    f"Unknown SGLang-Omni model {model_id!r} — pick a "
+                    "catalog key for --model (see the backend docs).")
+        model_id = sg_entry.key
+        voice_mode = voice_mode_for(backend, voice, clone, instructions,
+                                    model=model_id)
     else:
         # qwen: instructions design the voice (VoiceDesign model), a
         # reference .wav clones one (Base), otherwise a built-in speaker.
@@ -396,8 +415,17 @@ def convert(backend: str, voice: str = None, clone: str = None,
     try:
         if manage_server and api_url is None:
             from backends import managed
-            server = managed.ensure_running(backend, voice_mode)
+            server = managed.ensure_running(backend, voice_mode,
+                                            model=model_id)
             if server is not None and not server.ok:
+                # The boot failed (crashed, refused, or timed out): the
+                # detail went to the server's own log and the console, so
+                # record it here too — the dated run log the failure
+                # pointers name must not stay empty.
+                from backends import servers as _servers
+                logging.error("the %s server failed to start; its output "
+                              "is in %s", server.spec.name,
+                              _servers.server_log_path(server.spec.name))
                 return 1
         converter = AudiobookConverter(
             voice_mode=voice_mode, voice_clone_ref_audio=clone,
@@ -501,6 +529,18 @@ Examples:
   python audiobook.py --backend qwen \\
       --instructions "A warm adult female narrator with a British accent"
 
+  # Use the SGLang-Omni server with a catalog model (one model per server
+  # process; the CLI starts/stops the managed instance around the run)
+  python audiobook.py --backend sglomni --model higgs_audio_v3_tts
+
+  # SGLang-Omni voice cloning from a reference clip
+  python audiobook.py --backend sglomni --model moss_tts_local \\
+      --clone path/to/reference.wav
+
+  # SGLang-Omni voice design (Qwen3-TTS VoiceDesign model)
+  python audiobook.py --backend sglomni --model qwen3_tts_1_7b_voicedesign \\
+      --instructions "A warm adult female narrator with a British accent"
+
   # Use the faster-qwen3-tts server (voice cloning, configured server-side)
   python audiobook.py --backend faster [--voice NAME]
 
@@ -582,13 +622,16 @@ Examples:
               "Ignored for m4b, which is always a single file.")
     )
     parser.add_argument(
-        "--backend", choices=[BACKEND_AUDIOCPP, BACKEND_QWEN, BACKEND_FASTER],
+        "--backend", choices=[BACKEND_AUDIOCPP, BACKEND_QWEN, BACKEND_FASTER,
+                              BACKEND_SGLOMNI],
         required=True,
         help=("TTS server to talk to: the qwen-tts demo server (qwen), the "
-              "faster-qwen3-tts OpenAI-compatible server (faster), or an "
+              "faster-qwen3-tts OpenAI-compatible server (faster), an "
               "audio.cpp audiocpp_server (audiocpp) hosting any of its TTS "
               "model families — Qwen3-TTS, Higgs Audio, VoxCPM2, IndexTTS2, "
-              "and more.")
+              "and more — or an SGLang-Omni server (sglomni) hosting one "
+              "of its catalog models (Qwen3-TTS, Higgs Audio v3, MOSS-TTS, "
+              "Voxtral TTS, Fish Speech, and more).")
     )
     parser.add_argument(
         "--voice", type=str, default=None, metavar="NAME",
@@ -599,7 +642,10 @@ Examples:
               "there), for every other family a voice_preset or voice_dir "
               "entry (cloning). qwen: a built-in CustomVoice speaker name "
               "(required for built-in-speaker runs; use --clone or "
-              "--instructions for the other voice modes).")
+              "--instructions for the other voice modes). sglomni: a "
+              "preset voice name on speaker-capable models (Qwen3-TTS "
+              "CustomVoice, Voxtral); other models use their default "
+              "voice, or --clone / --instructions.")
     )
     parser.add_argument(
         "--debug", action="store_true",
@@ -610,22 +656,25 @@ Examples:
     )
     parser.add_argument(
         "--model", type=str, default=None, metavar="ID",
-        help=("audio.cpp server model entry id to use for this run "
-              "(--backend audiocpp only). Required when the server hosts "
-              "several lazily-loaded models: generate one server.json with "
-              "backends.audiocpp, then pick the model per run with --model. "
-              "Leave unset to auto-select when the server hosts exactly one "
-              "entry.")
+        help=("audio.cpp server model entry id (--backend audiocpp) or "
+              "SGLang-Omni catalog model key (--backend sglomni, e.g. "
+              "higgs_audio_v3_tts) to use for this run. audiocpp: "
+              "required when the server hosts several lazily-loaded "
+              "models. sglomni: required when several models are "
+              "installed (one model per server process); auto-selected "
+              "when exactly one is installed.")
     )
     parser.add_argument(
         "--instructions", type=str, default=None, metavar="TEXT",
         help=("Voice design or style instruction sent with every request "
-              "(--backend audiocpp or qwen). With audiocpp it is required "
-              "for voice design models (server entries with task 'vdes', "
-              "e.g. Qwen3-TTS VoiceDesign) and optional style/delivery "
-              "control elsewhere; with qwen it selects the VoiceDesign "
-              "model and describes the voice to synthesize with, e.g. "
-              "'A warm adult female narrator with a British accent'.")
+              "(--backend audiocpp, qwen or sglomni). With audiocpp it is "
+              "required for voice design models (server entries with task "
+              "'vdes', e.g. Qwen3-TTS VoiceDesign) and optional "
+              "style/delivery control elsewhere; with qwen it selects the "
+              "VoiceDesign model and describes the voice to synthesize "
+              "with; with sglomni it is required for the VoiceDesign "
+              "model, e.g. 'A warm adult female narrator with a British "
+              "accent'.")
     )
     parser.add_argument(
         "--option", action="append", type=str, default=None, metavar="KEY=VALUE",
@@ -753,6 +802,43 @@ Examples:
                 args.language = normalize_language(args.language)
             except ValueError as exc:
                 parser.error(str(exc))
+    elif args.backend == BACKEND_SGLOMNI:
+        # sglomni: --clone hands a reference clip to the server per
+        # request (transcribed locally when no --transcription is given);
+        # --instructions design the voice on the VoiceDesign model;
+        # --voice names a preset on speaker-capable models. Language is
+        # sent through (the server accepts a display-name hint).
+        if args.language is not None:
+            try:
+                args.language = normalize_language(args.language)
+            except ValueError as exc:
+                parser.error(str(exc))
+        if not args.clone and (args.transcription or args.no_transcription):
+            print("[WARNING] --transcription/--no-transcription "
+                  "are ignored without --clone")
+        try:
+            from backends.sglomni import models as sg_models
+            installed = sg_models.installed_keys()
+        except Exception:
+            installed = None
+        if installed is not None:
+            if not installed:
+                parser.error("--backend sglomni: no models are downloaded "
+                             "— install one via the TUI's Configure "
+                             "Backends → SGLang-Omni first")
+            if args.model is None:
+                if len(installed) > 1:
+                    parser.error(
+                        "--backend sglomni requires --model when several "
+                        "models are installed (installed: "
+                        f"{', '.join(installed)})")
+                args.model = installed[0]
+            elif sg_models.entry_by_key(args.model) is None:
+                parser.error(f"--model {args.model!r} is not an "
+                             "SGLang-Omni catalog key (installed: "
+                             f"{', '.join(installed)})")
+        if args.clone and not Path(args.clone).is_file():
+            parser.error(f"--clone: no such reference audio file: {args.clone}")
     else:
         # qwen: --voice names a built-in CustomVoice speaker (the only
         # voice mode that needs one; --clone and --instructions pick the
@@ -779,14 +865,17 @@ Examples:
                          "reference .wav, or --instructions \"...\" to "
                          "design a voice")
 
-    if args.model is not None and args.backend != BACKEND_AUDIOCPP:
-        parser.error("--model requires --backend audiocpp; it selects an "
-                     "audio.cpp server model entry id")
+    if args.model is not None and args.backend not in \
+            (BACKEND_AUDIOCPP, BACKEND_SGLOMNI):
+        parser.error("--model requires --backend audiocpp or sglomni; it "
+                     "selects an audio.cpp server model entry id or an "
+                     "SGLang-Omni catalog model key")
 
     if args.instructions is not None and args.backend not in \
-            (BACKEND_AUDIOCPP, BACKEND_QWEN):
-        parser.error("--instructions requires --backend audiocpp or qwen; "
-                     "it is sent as the request's voice design/style field")
+            (BACKEND_AUDIOCPP, BACKEND_QWEN, BACKEND_SGLOMNI):
+        parser.error("--instructions requires --backend audiocpp, qwen or "
+                     "sglomni; it is sent as the request's voice "
+                     "design/style field")
 
     request_options = {}
     if args.option: