feat: sglang backend support

1 files changed, 277 insertions, 0 deletions

diff --git a/app/backends/sglomni/catalog.py b/app/backends/sglomni/catalog.py
new file mode 100644
index 0000000..c89e865
--- /dev/null
+++ b/app/backends/sglomni/catalog.py
@@ -0,0 +1,277 @@
+"""The sglang-omni model catalog: what can be installed, hosted, and how.
+
+One server process hosts one model (`sgl-omni serve --model-path <hf-repo>
+--config <yaml>`), so each entry is one launchable unit — unlike audio.cpp's
+multi-entry server.json. Every entry states how its voice is supplied
+(`capability`), whether a reference clip is mandatory for narration
+(`requires_reference`), and which model-companion packages (`extras`) must
+be pip-installed into the backend venv before its server will start.
+
+The catalog is static knowledge about sglang-omni's supported TTS models
+(v0.1.4-era), not a live query: the server's /v1/models answers only which
+model is currently hosted. Entries carry the HuggingFace repo id verbatim —
+it is both the `--model-path` value and the model name /v1/models reports,
+so a running server is matched back to its entry by that id.
+
+Voice capabilities (mirroring the audio.cpp client's vocabulary):
+  speaker  the voice is a named preset shipped with the model (Qwen3-TTS
+           CustomVoice speaker table; Voxtral's preset voices)
+  clone    the voice comes from a reference clip (per-request ref_audio +
+           ref_text, transcribed with Whisper when not provided)
+  design   the voice is described by instructions (Qwen3-TTS VoiceDesign)
+"""
+
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Tuple
+
+# Voice capabilities (same words the audio.cpp client uses).
+CAPABILITY_SPEAKER = "speaker"
+CAPABILITY_CLONE = "clone"
+CAPABILITY_DESIGN = "design"
+
+# A model-companion pip package: (requirement spec, no_deps). The no-deps
+# flag mirrors upstream install instructions that must not drag in their
+# own (conflicting) transformer pins — the Qwen3-TTS companion pins
+# Transformers 4 against sglang-omni's Transformers 5 stack.
+Extra = Tuple[str, bool]
+
+
+@dataclass(frozen=True)
+class ModelEntry:
+    """One installable/hostable sglang-omni TTS model."""
+    key: str                  # catalog id used by --model and the forms
+    label: str                # human-readable name (menus, status lines)
+    repo: str                 # HuggingFace repo id == --model-path value
+    config: Optional[str]     # vendored config yaml file name, None = none
+    capability: str           # speaker / clone / design
+    requires_reference: bool  # clone models: narration needs ref audio
+    extras: Tuple[Extra, ...] = field(default=())
+    system_dep: Optional[str] = None   # system binary the extras need
+    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)
+    notes: str = ""           # one-line description (documentation)
+    # The model's DEFAULT pipeline dynamically quantizes its MoE experts to
+    # FP8 at load time (sglang-omni's zonos2 config hardcodes it) — a Triton
+    # fp8e4nv kernel that only compiles on compute capability 8.9+ (RTX
+    # 4090/5090, Hopper). On older GPUs the server dies mid-boot; BF16_CONFIG
+    # is the vendored config that turns FP8 off so the model runs in bf16
+    # (~2x the MoE VRAM) there instead.
+    fp8_moe: bool = False
+    fp8_min_compute_capability: Optional[Tuple[int, int]] = None
+    bf16_config: Optional[str] = None
+
+
+# The Qwen3-TTS CustomVoice speaker table — the same built-in speakers the
+# qwen-tts demo and audio.cpp's CustomVoice entry expose (kept as a local
+# copy so this catalog stays importable without the converter package; the
+# test suite asserts it matches converter.clients.speakers).
+QWEN_CUSTOMVOICE_SPEAKERS: Tuple[str, ...] = (
+    "Vivian", "Serena", "Uncle_Fu", "Dylan", "Eric",
+    "Ryan", "Aiden", "Ono_Anna", "Sohee")
+
+# The Qwen3-TTS companions follow the upstream TTS guide exactly: the
+# qwen-tts demo package installs WITHOUT dependencies (its Transformers 4
+# pin would replace sglang-omni's pinned 5.12 stack; sglang-omni shims the
+# API differences), as do sox/einops (a normal sox resolve pulls numpy past
+# the ceiling numba imposes). The sox *binary* is a system package the
+# wizard detects and guides (it is required at synthesis time).
+_QWEN_EXTRAS: Tuple[Extra, ...] = (
+    ("sox", True), ("einops", True), ("qwen-tts==0.1.1", True))
+_SOX_HINT = ("install the sox system package (e.g. sudo pacman -S sox, "
+             "sudo apt install sox, brew install sox)")
+# The Fish Audio and ZONOS2 pipelines use the Descript DAC codec, which
+# upstream installs WITH dependencies (nothing conflicts).
+_DAC_EXTRAS: Tuple[Extra, ...] = (
+    ("descript-audiotools==0.7.2", False),
+    ("descript-audio-codec==1.0.0", False))
+
+ENTRIES: Tuple[ModelEntry, ...] = (
+    ModelEntry(
+        key="qwen3_tts_0_6b_customvoice",
+        label="Qwen3-TTS 0.6B CustomVoice",
+        repo="Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice",
+        config="qwen3_tts_0_6b_customvoice.yaml",
+        capability=CAPABILITY_SPEAKER,
+        requires_reference=False,
+        extras=_QWEN_EXTRAS, system_dep="sox", system_hint=_SOX_HINT,
+        speakers=QWEN_CUSTOMVOICE_SPEAKERS,
+        notes="built-in speakers, lightest model",
+    ),
+    ModelEntry(
+        key="qwen3_tts_0_6b_base",
+        label="Qwen3-TTS 0.6B Base",
+        repo="Qwen/Qwen3-TTS-12Hz-0.6B-Base",
+        config="qwen3_tts_0_6b.yaml",
+        capability=CAPABILITY_CLONE,
+        requires_reference=True,
+        extras=_QWEN_EXTRAS, system_dep="sox", system_hint=_SOX_HINT,
+        supports_seed=True,
+        notes="voice cloning from a reference clip",
+    ),
+    ModelEntry(
+        key="qwen3_tts_1_7b_base",
+        label="Qwen3-TTS 1.7B Base",
+        repo="Qwen/Qwen3-TTS-12Hz-1.7B-Base",
+        config="qwen3_tts_1_7b.yaml",
+        capability=CAPABILITY_CLONE,
+        requires_reference=True,
+        extras=_QWEN_EXTRAS, system_dep="sox", system_hint=_SOX_HINT,
+        supports_seed=True,
+        notes="voice cloning, higher quality",
+    ),
+    ModelEntry(
+        key="qwen3_tts_1_7b_voicedesign",
+        label="Qwen3-TTS 1.7B VoiceDesign",
+        repo="Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign",
+        config="qwen3_tts_1_7b_voicedesign.yaml",
+        capability=CAPABILITY_DESIGN,
+        requires_reference=False,
+        extras=_QWEN_EXTRAS, system_dep="sox", system_hint=_SOX_HINT,
+        notes="voice described by instructions",
+    ),
+    ModelEntry(
+        key="higgs_audio_v3_tts",
+        label="Higgs Audio v3 TTS",
+        repo="bosonai/higgs-audio-v3-tts-4b",
+        config=None,
+        capability=CAPABILITY_CLONE,
+        requires_reference=False,
+        notes="zero-shot narration, cloning from a reference clip",
+    ),
+    ModelEntry(
+        key="moss_tts",
+        label="MOSS-TTS v1.5",
+        repo="OpenMOSS-Team/MOSS-TTS-v1.5",
+        config="moss_tts.yaml",
+        capability=CAPABILITY_CLONE,
+        requires_reference=True,
+        notes="voice cloning from a reference clip",
+    ),
+    ModelEntry(
+        key="moss_tts_local",
+        label="MOSS-TTS Local v1.5",
+        repo="OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5",
+        config="moss_tts_local.yaml",
+        capability=CAPABILITY_CLONE,
+        requires_reference=False,
+        notes="48 kHz, narration without a reference or cloning",
+    ),
+    ModelEntry(
+        key="voxtral_tts",
+        label="Voxtral TTS 4B",
+        repo="mistralai/Voxtral-4B-TTS-2603",
+        config="voxtral_tts.yaml",
+        capability=CAPABILITY_SPEAKER,
+        requires_reference=False,
+        notes="preset voices, text-only requests",
+    ),
+    ModelEntry(
+        key="dots_tts_mf",
+        label="dots.tts (MeanFlow)",
+        repo="dots-studio/dots.tts-mf",
+        config="dots_tts.yaml",
+        capability=CAPABILITY_CLONE,
+        requires_reference=True,
+        notes="voice cloning, requires reference audio + transcript",
+    ),
+    ModelEntry(
+        key="fish_s2_pro",
+        label="Fish Speech S2-Pro",
+        repo="fishaudio/s2-pro",
+        config="s2pro_tts.yaml",
+        capability=CAPABILITY_CLONE,
+        requires_reference=False,
+        extras=_DAC_EXTRAS,
+        notes="zero-shot narration or cloning from a reference clip",
+    ),
+    ModelEntry(
+        key="zonos2",
+        label="ZONOS2",
+        repo="Zyphra/zonos2",
+        config=None,
+        capability=CAPABILITY_CLONE,
+        requires_reference=True,
+        extras=_DAC_EXTRAS,
+        notes="voice cloning, 44.1 kHz DAC vocoder",
+        fp8_moe=True,
+        fp8_min_compute_capability=(8, 9),
+        bf16_config="zonos2_bf16.yaml",
+    ),
+)
+
+_BY_KEY: Dict[str, ModelEntry] = {entry.key: entry for entry in ENTRIES}
+_BY_REPO: Dict[str, ModelEntry] = {entry.repo: entry for entry in ENTRIES}
+
+
+def entry_by_key(key: str) -> Optional[ModelEntry]:
+    """The catalog entry for a catalog KEY, or None."""
+    return _BY_KEY.get(key)
+
+
+def entry_by_repo(repo: str) -> Optional[ModelEntry]:
+    """The catalog entry hosting REPO (a /v1/models id), or None."""
+    return _BY_REPO.get(repo)
+
+
+def entries_by_keys(keys) -> List[ModelEntry]:
+    """The ENTRIES for KEYS, in catalog order (unknown keys dropped)."""
+    wanted = set(keys)
+    return [entry for entry in ENTRIES if entry.key in wanted]
+
+
+def config_path(entry: ModelEntry):
+    """The vendored config yaml path for ENTRY, or None when it runs on
+    --model-path alone (Higgs, ZONOS2). A declared-but-missing file means
+    a broken install — callers treat that like a missing entry."""
+    return _config_file(entry.config)
+
+
+def fallback_config_path(entry: ModelEntry):
+    """The vendored bf16 config yaml for ENTRY (None when it has none).
+
+    The config a GPU below the entry's fp8_min_compute_capability launches
+    with instead of the model's default FP8-quantized pipeline."""
+    return _config_file(entry.bf16_config)
+
+
+def _config_file(name: Optional[str]):
+    """CONFIGS_DIR/NAME, or None when NAME is None (no config for ENTRY)."""
+    if name is None:
+        return None
+    from .constants import CONFIGS_DIR
+    return CONFIGS_DIR / name
+
+
+def install_tree_families(available: List[ModelEntry]) -> List[dict]:
+    """The checkbox-tree shape for the model picker, grouped by upstream.
+
+    One family per model origin with one option per model, so the picker
+    reads like the audio.cpp one (a collapsed tree of related packages).
+    No ``detail`` is set: the tree's status line under the buttons would
+    only repeat the catalog keys under the cursor. AVAILABLE filters what
+    is shown (already-installed models stay listed so they can be
+    re-checked or repaired).
+    """
+    groups: List[Tuple[str, List[ModelEntry]]] = [
+        ("Qwen3-TTS (Qwen)", [e for e in available if e.repo.startswith("Qwen/")]),
+        ("Boson AI", [e for e in available if e.repo.startswith("bosonai/")]),
+        ("OpenMOSS", [e for e in available if e.repo.startswith("OpenMOSS-Team/")]),
+        ("Mistral AI", [e for e in available if e.repo.startswith("mistralai/")]),
+        ("dots.studio", [e for e in available if e.repo.startswith("dots-studio/")]),
+        ("Fish Audio", [e for e in available if e.repo.startswith("fishaudio/")]),
+        ("Zyphra", [e for e in available if e.repo.startswith("Zyphra/")]),
+    ]
+    families = []
+    for label, entries in groups:
+        if not entries:
+            continue
+        families.append({
+            "label": label,
+            "options": [
+                {"key": entry.key, "label": entry.label,
+                 "recommended": False}
+                for entry in entries
+            ],
+        })
+    return families