InferMatrix/Recipes

fishaudio/drama-3

Text to speech with drama-3-single-b300-engine on NVIDIA B300.

View on Hugging Face
Text to speechFP8 / BF16drama-3-single-b300-engine1 × B300unverified
INSTALLdrama-3-single-b300-engine · CUDAShell
Run in the prepared environment
git clone https://github.com/InferMatrixLab/drama-3-single-b300-engine.git
cd drama-3-single-b300-engine
git checkout 58a1882a4d941f8d21bfce6be37d3d5e7c9fde80
git clone https://github.com/fishaudio/fish-speech.git ../fish-speech
git -C ../fish-speech checkout 541566932abbaebcb0a40e562a1f846738383f38
python3 scripts/install_runtime.py --venv .venv --fish-source ../fish-speech
.venv/bin/hf download fishaudio/drama-3 \
  --revision e0a9218aa48888270f06ef1ed9ea67622c0e17d2 --local-dir models/drama-3
Unverified on NVIDIA B300
set -a
. configs/benchmark-native-b300.env
set +a
CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES-0}" OMP_NUM_THREADS=8 .venv/bin/python -m drama3.server \
  --checkpoint models/drama-3 --fish-source ../fish-speech --slots 64 --port 18082
HARDWARE
NVIDIA
Huawei Ascend
Apple Silicon
VARIANT
fp8 experts / bf16 state and codec
fishaudio/drama-3
ENGINE
BACKEND
Linux
TASK
Single device · recorded configuration
VERIFICATION
Deployment unverifiedAcceptance INCONCLUSIVE

Not revalidated for this catalog. See the original evidence for the exact scope.

ADVANCED Source & reproducibility
Deployment ref
58a1882a4d941f8d21bfce6be37d3d5e7c9fde80
Recipe source
InferMatrixLab/drama-3-single-b300-engine/recipes/drama3-b300-fp8.yaml
Content commit
c560a520913f51586687943ef0adb1ec764ef510
Updated
2026-10-19
Maintainers
@david6666666
Edit the source recipe

Environment & prerequisites

Linux

  • One allocated NVIDIA B300, Python 3.12, uv, Git and curl.
  • NVIDIA driver compatible with the locked CUDA 13 runtime; requirements-runtime.lock is installed by scripts/install_runtime.py.
  • Fish Speech source is pinned to 541566932abbaebcb0a40e562a1f846738383f38; model revision is e0a9218aa48888270f06ef1ed9ea67622c0e17d2.
  • Initial package installation and checkpoint download require network access.

Limitations & verification scope

  • This deployment ref has not been re-run on B300 during the catalog work; the Linux distribution is not frozen in the quickstart.
  • Mixed precision is intentional: native FP8 experts, BF16 GDN state and BF16 codec. No all-FP8 claim is made.
  • Output is 44.1 kHz mono signed 16-bit PCM. Allow model load, graph capture and codec warmup to finish before sending requests.
  • Historical matched results use standalone 33f6d7237196520be9f457d86df161264cbb087e and include incomplete baseline repetitions; they do not establish the all-concurrency 1.5x acceptance target.
  • Historical configuration, quality and raw evidence: https://github.com/InferMatrixLab/drama-3-single-b300-engine/blob/58a1882a4d941f8d21bfce6be37d3d5e7c9fde80/docs/current-results.md
  • Built with Fish Audio. Fish Audio Community License and dependency licenses apply; retain NOTICE and third_party_licenses.

Evidence & attribution

Source recipe Engine deployment revision

Imported reports retain their authorship and verification scope. Catalog checks do not establish inference quality or performance.