Skip to content

v1.0.0

Latest
Latest

Choose a tag to compare

View all tags
@dangkhoa2016 dangkhoa2016 released this 27 Sep 14:51
v1.0.0
5ebc6a1

OpenBMB-VoxCPM2-Inference v1.0.0

Initial public release of an independent engineering project for portable OpenBMB VoxCPM2 inference.

This is not an official OpenBMB release. Model weights are not included in this repository.

Highlights

  • CPU inference on the measured Linux/Kaggle environment.
  • Single-GPU inference qualified on NVIDIA Tesla T4 16 GiB.
  • T4 x2 replica mode with one isolated model worker per GPU.
  • Deterministic execution profiles: cpu, cuda-single, cuda-replica, and auto.
  • Standard 48 kHz mono PCM16 WAV synthesis.
  • Voice design, controllable voice cloning, and audio continuation.
  • Backend-native incremental streaming.
  • Authenticated FastAPI runtime with health/readiness endpoints.
  • Bounded queueing, active-concurrency limits, cancellation, and timeout controls.
  • Reproducible Kaggle demo with English/Vietnamese playable audio.
  • CPU Docker portability contract on generic GitHub-hosted Linux.
  • MIT license for this repository's original code and documentation.
  • Bilingual community governance, issue/PR templates, and support/security guidance.
  • CI uses actions/checkout v7 and actions/setup-python v7.
  • Public README badges for CI, license, Python, PyTorch/VoxCPM2, T4x2, FastAPI, Kaggle, Docker, and release status.

Release authority

Git SHA:  5ebc6a151d8c3fcba9be2c159317f9b318375c97
Version:  1.0.0
CI run:   36382895414
Result:   SUCCESS

The public tree passed:

516 passed
1 skipped
Ruff: PASS
git diff --check: PASS
editable package build/install: PASS
community/license/link audits: PASS

The single skipped test is the root/filesystem-permission case where root bypasses ordinary read-permission semantics.

CI passed on Python 3.10, 3.11, and 3.12, including Docker portability.

License and third-party boundaries

This repository's original code and documentation are licensed under the MIT License:

Copyright (c) 2026 Đăng Khoa <i.am@dangkhoa.dev>

The VoxCPM2 model weights and their upstream OpenBMB runtime dependencies remain subject to their original license terms. This repository does not relicense third-party code or model weights.

Model identity

Model:           openbmb/VoxCPM2
Model revision:  32279effe8c19989596f05d353d1447f51d9e915
Upstream source: f772e498a45fbb5fb8e13fbf9b9c48be9fe33e69

Kaggle model mirror used by the demo flow:

dangkhoa2016/openbmb-voxcpm2

Kaggle demo

Canonical public notebook:

https://www.kaggle.com/code/dangkhoa2016/openbmb-voxcpm2-inference-kaggle-t4x2-demo

Community files

The repository includes:

  • CODEOWNERS;
  • bilingual Code of Conduct, Contributing, Security, and Support guides;
  • bilingual bug, feature, documentation, and question issue templates;
  • bilingual pull request templates;
  • weekly GitHub Actions Dependabot configuration.

Qualification

Public measured evidence and claim boundaries are summarized in:

provenance/QUALIFICATION.md

Important boundaries

The release does not claim universal CPU/RAM or GPU compatibility, tensor parallelism/model sharding, automatic failed-worker respawn, HA/SLA/autoscaling, real-model inference inside Docker, CUDA Docker qualification, semantic speaker-similarity guarantees, or universal real-time latency guarantees.

Assets 4
Loading