diff options
Diffstat (limited to 'app/backends/sglomni/catalog.py')
| -rw-r--r-- | app/backends/sglomni/catalog.py | 13 |
1 files changed, 13 insertions, 0 deletions
diff --git a/app/backends/sglomni/catalog.py b/app/backends/sglomni/catalog.py index f20509f..8bfb1fa 100644 --- a/app/backends/sglomni/catalog.py +++ b/app/backends/sglomni/catalog.py @@ -52,6 +52,11 @@ 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) + # 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, + # so those pipelines may accept one too — verify before flipping the + # flag (CONSTANT_SEED currently no-ops for every other entry). 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 @@ -103,6 +108,12 @@ _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 decode their codec assets through +# the ffmpeg binary (and the client concatenates multi-part chunks with +# it); the weights download fine without it, the server just fails to +# synthesize — so the install flow warns, like it does for sox. +_FFMPEG_HINT = ("install the ffmpeg system package (e.g. sudo pacman -S " + "ffmpeg, sudo apt install ffmpeg, brew install ffmpeg)") # The Fish Audio and ZONOS2 pipelines use the Descript DAC codec, which # upstream installs WITH dependencies — but descript-audiotools carries a # vestigial 2021-era pin, protobuf<3.20 (its code never imports protobuf), @@ -255,6 +266,7 @@ ENTRIES: Tuple[ModelEntry, ...] = ( capability=CAPABILITY_CLONE, requires_reference=False, extras=_DAC_EXTRAS, + system_dep="ffmpeg", system_hint=_FFMPEG_HINT, notes="zero-shot narration or cloning from a reference clip", ), ModelEntry( @@ -265,6 +277,7 @@ ENTRIES: Tuple[ModelEntry, ...] = ( capability=CAPABILITY_CLONE, requires_reference=True, extras=_DAC_EXTRAS, + system_dep="ffmpeg", system_hint=_FFMPEG_HINT, notes="voice cloning, 44.1 kHz DAC vocoder", fp8_moe=True, fp8_min_compute_capability=(8, 9), |