Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice

Text to speech with qwen3-tts-engine on Huawei Ascend 910B1.

Text to speechBF16qwen3-tts-engine1 × Ascend 910B1unverified

INSTALLqwen3-tts-engine · ASCENDShell

Run in the prepared environment

git clone https://github.com/InferMatrixLab/qwen3-tts-engine.git
cd qwen3-tts-engine
git checkout b4451682abac651f26046f8d27467de3f007e1e7
python -m pip install --no-deps -e .
npu-smi info
python -c "import torch, torch_npu, triton, qwen_tts; assert torch.npu.is_available()"
export QWEN_TTS_MODEL="$PWD/models/Qwen3-TTS-12Hz-1.7B-CustomVoice"
HF_HUB_OFFLINE=0 python - <<'PY'
import os
from huggingface_hub import snapshot_download
snapshot_download(repo_id="Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",
                  revision="0c0e3051f131929182e2c023b9537f8b1c68adfe",
                  local_dir=os.environ["QWEN_TTS_MODEL"])
PY

export ASCEND_RT_VISIBLE_DEVICES="${ASCEND_RT_VISIBLE_DEVICES-0}"
export OMP_NUM_THREADS=4
export HF_HUB_OFFLINE=1
export LD_PRELOAD=/usr/lib/aarch64-linux-gnu/libjemalloc.so.2
python - <<'PY'
import json
import os
import subprocess
import sys
from pathlib import Path
command = json.loads(Path("evidence/campaign-20261017/plan/continuous-direct-prompt-command.json").read_text())
command[0] = sys.executable
command[command.index("--model") + 1] = os.environ["QWEN_TTS_MODEL"]
subprocess.run(command, check=True)
PY

HARDWARE

NVIDIA

Huawei Ascend

Apple Silicon

VARIANT

bf16

Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice

ENGINE

BACKEND

Linux (prepared Ascend environment; distribution not recorded)

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: b4451682abac651f26046f8d27467de3f007e1e7
Recipe source: InferMatrixLab/qwen3-tts-engine/recipes/qwen3-tts-ascend-910b1.yaml
Content commit: 63620f6d906de103cc574f859124145998cc91a5
Updated: 2026-10-19
Maintainers: @amy-why-3459

Edit the source recipe

Environment & prerequisites

Linux (prepared Ascend environment; distribution not recorded)

One allocated Ascend 910B1; Python 3.12 with CANN and Triton Ascend already installed.
Prepared environment: torch 2.10.0, torch_npu 2.10.0.post4, qwen-tts 0.1.1, transformers 4.57.3.
Preserve the CANN PYTHONPATH; the repository does not install the NPU driver or toolchain.
Recorded ARM64 allocator workaround requires /usr/lib/aarch64-linux-gnu/libjemalloc.so.2.
GitHub, Python package index and Hugging Face access are required for initial installation and model download.

Limitations & verification scope

This catalog entry has not been executed on hardware during the 2026-10-19 documentation work.
The historical PASS covers measured engine e783a64b0fb4f9ade2ffb3980b3e1f74a00c9a6a and its frozen contract; it does not qualify this deployment ref automatically.
Historical evidence: https://github.com/InferMatrixLab/qwen3-tts-engine/blob/b4451682abac651f26046f8d27467de3f007e1e7/evidence/campaign-20261017/acceptance.md
Only this CustomVoice checkpoint, one 910B1 and the recorded BF16 configuration are represented. The separate 910B3 comparison is not validation of this recipe.
Keep the install and serve blocks in the same activated shell. Send the request from a second terminal after health succeeds.
The endpoint streams base64 PCM over SSE; it does not return a WAV directly. The OS distribution and CANN version were not fully frozen in the source documentation.

Evidence & attribution

Source recipe Engine deployment revision

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