habemus-papadum-nvenc (import pdum.nvenc)
GPU NV12 → H.264/HEVC Annex B via NVIDIA's Video Codec SDK encoder, with no
PyAV and no host copy. The companion GPU encoder for
pdum.rfb (PyPI: habemus-papadum-rfb); a uv workspace member of
this repo. Full assessment: docs/nvenc_sdk_evaluation.md.
Why it exists:
- Builds + runs on CPython 3.14. Upstream
PyNvVideoCodecpins pybind11 2.10.0 (no 3.14) and ships no cp314 wheel/sdist; this package builds against pybind11 v3.0.4. - GPU-resident, PyAV-free. Encodes any
__cuda_array_interface__tensor (CuPy / PyTorch / Numba) directly, sidestepping the PyAV-18 requirement entirely.
What's ours vs NVIDIA's
src/cpp/nvenc_ext.cpp OURS — the only hand-written C++; thin pybind11 binding +
all NVTX ranges. Wraps NvEncoderCuda; no NVIDIA edits.
src/pdum/nvenc/__init__.py OURS — Python surface + the ABI loader (picks 12.1/13.0).
CMakeLists.txt OURS — pybind11 3.0.4 (the 3.14 fix), dual ABI, optional NVTX.
build-wheel.sh OURS — self-contained wheel build (auditwheel).
third_party/ VERBATIM, UNMODIFIED NVIDIA SDK (MIT). See PROVENANCE.md.
The NVIDIA source under third_party/ is copied byte-for-byte from PyNvVideoCodec 2.1.0 with its MIT headers intact; we made zero edits to it.
Dual NVENC ABI
The wheel ships two extensions built from the same source — _nvenc_121 (NVENC SDK
12.1) and _nvenc_130 (13.0) — and pdum/nvenc/__init__.py loads whichever the host
driver supports (newest first, via a cheap NvEncodeAPIGetMaxSupportedVersion probe),
so one wheel works across old and new drivers.
Build & test (local, CMake)
cmake -S . -B build -G Ninja -DUSE_NVTX=OFF # or -DUSE_NVTX=ON for profiling
cmake --build build -j
Build wheels (maintainer)
./build-wheel.sh # cp314 -> dist/habemus_papadum_nvenc-*.whl
./build-wheel.sh --nvtx # profiling wheel (NVTX ranges on)
PYTHON_VERSIONS="3.12 3.13 3.14" ./build-wheel.sh
The wheel bundles its C/C++ runtime deps but not libcuda / libnvidia-encode
— those come from the host NVIDIA driver, as they must. Publishing to PyPI is done
by scripts/publish.sh (which calls this), not from CI.
Usage
import cupy as cp
from pdum.nvenc import NvencEncoder
enc = NvencEncoder(1920, 1080, codec="h264", preset="p3", tuning="ll", fps=30, gop=30)
nv12 = cp.empty((1080 * 3 // 2, 1920), dtype=cp.uint8) # contiguous NV12
# ... render into nv12 ...
annexb = enc.encode(nv12, force_idr=True) # bytes; H.264 Annex B
annexb += enc.flush()
enc.close()
NvencEncoder(cuda_context=0) retains the device primary context — the same one
CuPy/PyTorch use — so device pointers are valid to NVENC with no cross-context copy.
NvencEncoder(extra_output_delay=0) (the default) is zero-latency: each frame's
access unit is returned by its own encode() call (synchronous 1-in-1-out), which is
what a low-latency stream wants. Raise it (NVIDIA's helper defaults to 3) to overlap
encode with rendering for throughput, at a matching cost in frames of latency.
NVTX profiling
Built with --nvtx / -DUSE_NVTX=ON, the binding emits ranges at the Python
boundary (pdum.encode, pdum.read_cai, pdum.copy_to_nvenc, pdum.submit,
pdum.collect_output) and activates NVIDIA's internal ranges (EncodeFrame,
DoEncode, MapResources, CopyToDeviceFrame_*). Profile with Nsight Systems:
nsys profile -t nvtx,cuda python your_script.py.
Scope / caveats
- Fixed-resolution NV12 in, Annex B out, one encoder per instance. No reconfigure /
SEI /
EncoderBackendwiring yet — that's thepdum.rfbintegration. - Input uses
GetNextInputFrame+CopyToDeviceFrame(one intra-GPU copy, no host round-trip). True zero-copy viaNvEncoder::RegisterResourceis a follow-up.
Metadata
Release files for habemus-papadum-nvenc 0.3.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Built distributions (wheels)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| habemus_papadum_nvenc-0.3.0-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl | CPython 3.14 | CPython 3.14 | Linux glibc 2.28+ x86-64, Linux glibc 2.27+ x86-64 | Details for habemus_papadum_nvenc-0.3.0-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl |
| habemus_papadum_nvenc-0.3.0-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl | CPython 3.13 | CPython 3.13 | Linux glibc 2.28+ x86-64, Linux glibc 2.27+ x86-64 | Details for habemus_papadum_nvenc-0.3.0-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl |
| habemus_papadum_nvenc-0.3.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl | CPython 3.12 | CPython 3.12 | Linux glibc 2.28+ x86-64, Linux glibc 2.27+ x86-64 | Details for habemus_papadum_nvenc-0.3.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl |
Total release size: 854.7 kB
Release files / habemus_papadum_nvenc-0.3.0-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl
| Download URL | habemus_papadum_nvenc-0.3.0-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl |
|---|---|
| Size | 285.0 kB |
| Tags | CPython 3.14 Linux glibc 2.27+ x86-64 Linux glibc 2.28+ x86-64 |
|
SHA-256 checksum How to use checksums |
6174860082ec6aa0c6c934c43ee695ad277d88ad29106277e359b7545e190293
|
|
BLAKE2b-256 checksum How to use checksums |
ffd86015c4c2b03880b46ede0638cfa8f1226a790bb350d33e1151ad137a26ff
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.1.0 CPython/3.13.12
|
Release files / habemus_papadum_nvenc-0.3.0-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl
| Download URL | habemus_papadum_nvenc-0.3.0-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl |
|---|---|
| Size | 285.0 kB |
| Tags | CPython 3.13 Linux glibc 2.27+ x86-64 Linux glibc 2.28+ x86-64 |
|
SHA-256 checksum How to use checksums |
bd11ccb6c01deca35bf80943b2bac0129624a845f0335b6c936bbd78bafd5464
|
|
BLAKE2b-256 checksum How to use checksums |
b5686fe6f6c605058ad51d8dda072726981d6d8cc6e14a4bfceb267ed9e4feaf
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.1.0 CPython/3.13.12
|
Release files / habemus_papadum_nvenc-0.3.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl
| Download URL | habemus_papadum_nvenc-0.3.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl |
|---|---|
| Size | 284.7 kB |
| Tags | CPython 3.12 Linux glibc 2.27+ x86-64 Linux glibc 2.28+ x86-64 |
|
SHA-256 checksum How to use checksums |
331d8cf584b7427eb96869578edaca7a4209314441d2923c1d9de3cb6d459937
|
|
BLAKE2b-256 checksum How to use checksums |
04c435de5be5e3661a79fc453fa8fdc51c101ebb755083d2db6357bb976b72a2
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.1.0 CPython/3.13.12
|
Security for habemus-papadum-nvenc 0.3.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.