Repository navigation
nemo-speech-python v0.3.0rc1
Pre-release
Unofficial Python bindings for NeMo-Speech.cpp, v0.3.0rc1.
Changes
What's Changed
Fixes
- Allocate the ggml scheduler's split buffers on demand instead of 1.3 GB per session by @PNL-toshiyaishihara in #65
- Bound the RNNT greedy decoder's joint window with a NeMo-Speech.cpp patch by @PNL-toshiyaishihara in #67
Documentation
- Allow small, result-preserving patches to NeMo-Speech.cpp sources by @PNL-toshiyaishihara in #66
Full Changelog: v0.2.0...v0.3.0rc1
Vendored sources
| Source | Revision | Commit |
|---|---|---|
| NeMo-Speech.cpp | v0.2.0 | 6a3ca3693707 |
| llama.cpp | bd4f514 | bd4f514db14d |
| SentencePiece | v0.2.0 | 17d7580d6407 |
NeMo-Speech.cpp is unchanged since v0.2.0.
Wheels
Python: 3.10, 3.11, 3.12, 3.13, 3.14 (Requires-Python >=3.10). Every wheel is py3-none, so one file serves all of them.
The x86-64 wheels (Linux x86_64 and Windows x64, every variant) carry ggml's CPU backend once per CPU generation and load the best one for the CPU at run time, from plain x86-64 to AVX-VNNI and AVX-512.
| Variant | Platform | Requirements | File |
|---|---|---|---|
| CPU | Linux aarch64 (glibc 2.28+) | nemo_speech-0.3.0rc1-py3-none-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl |
|
| CPU | Linux x86_64 (glibc 2.28+) | nemo_speech-0.3.0rc1-py3-none-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl |
|
| CPU | Windows x64 | nemo_speech-0.3.0rc1-py3-none-win_amd64.whl |
|
| CPU + Metal | macOS arm64 (13.0+) | nemo_speech-0.3.0rc1-py3-none-macosx_13_0_arm64.whl |
|
| Vulkan | Windows x64 | GPU driver with Vulkan | nemo_speech-0.3.0rc1+vulkan-py3-none-win_amd64.whl |
| CUDA 12.8 | Linux x86_64 (glibc 2.28+) | NVIDIA driver R570+; compute capability 7.5, 8.0, 8.6, 8.9, 9.0, 10.0, 12.0; newer GPUs through PTX (9.0+) | nemo_speech-0.3.0rc1+cu128-py3-none-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl |
| CUDA 12.8 | Windows x64 | NVIDIA driver R570+; compute capability 7.5, 8.0, 8.6, 8.9, 9.0, 10.0, 12.0; newer GPUs through PTX (9.0+) | nemo_speech-0.3.0rc1+cu128-py3-none-win_amd64.whl |
| CUDA 13.0 | Linux aarch64 (glibc 2.28+) | NVIDIA driver R580+; compute capability 7.5, 8.0, 8.6, 8.9, 9.0, 10.0, 11.0, 12.0, 12.1; newer GPUs through PTX (9.0+) | nemo_speech-0.3.0rc1+cu130-py3-none-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl |
| CUDA 13.0 | Linux x86_64 (glibc 2.28+) | NVIDIA driver R580+; compute capability 7.5, 8.0, 8.6, 8.9, 9.0, 10.0, 12.0; newer GPUs through PTX (9.0+) | nemo_speech-0.3.0rc1+cu130-py3-none-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl |
| CUDA 13.0 | Windows x64 | NVIDIA driver R580+; compute capability 7.5, 8.0, 8.6, 8.9, 9.0, 10.0, 12.0; newer GPUs through PTX (9.0+) | nemo_speech-0.3.0rc1+cu130-py3-none-win_amd64.whl |
Install
All variants share the distribution name nemo-speech, so install the file for your platform and backend directly (a version specifier with -f would let pip pick a GPU variant). For example:
pip install https://github.com/PNL-toshiyaishihara/nemo-speech-python/releases/download/v0.3.0rc1/nemo_speech-0.3.0rc1+cu128-py3-none-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whlGPU wheels bundle their runtime libraries (CUDA runtime, cuBLAS) and need only the GPU driver. nemo_speech.build_info() reports how an installation was built. SHA256SUMS lists the checksums of all files.
Known issues
Upstream issues worked around in this release: https://github.com/PNL-toshiyaishihara/nemo-speech-python/issues?q=label%3Aupstream
