Metadata-Version: 2.5
Name: jammi-server-cu12
Version: 0.49.1
Summary: CUDA (cu12) build of the Jammi engine server — GPU-accelerated `jammi-server` binary, pip-installable
Project-URL: Homepage, https://github.com/f-inverse/jammi-ai
Project-URL: Documentation, https://f-inverse.github.io/jammi-ai/
Project-URL: Repository, https://github.com/f-inverse/jammi-ai
Project-URL: Bug Tracker, https://github.com/f-inverse/jammi-ai/issues
Author: F-Inverse
License: Apache-2.0
Keywords: cuda,embeddings,fine-tune,gpu,inference,server,vector-search
Classifier: Development Status :: 4 - Beta
Classifier: Environment :: GPU :: NVIDIA CUDA :: 12
Classifier: Intended Audience :: Developers
Classifier: Intended Audience :: Science/Research
Classifier: License :: OSI Approved :: Apache Software License
Classifier: Operating System :: POSIX :: Linux
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Rust
Classifier: Topic :: Database
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Requires-Python: >=3.9
Requires-Dist: nvidia-cublas-cu12==12.6.*
Requires-Dist: nvidia-cuda-nvrtc-cu12==12.6.*
Requires-Dist: nvidia-cuda-runtime-cu12==12.6.*
Requires-Dist: nvidia-curand-cu12==10.3.7.*
Description-Content-Type: text/markdown

# jammi-server-cu12

GPU build of the Jammi engine server, packaged as a pip-installable wheel that
ships the CUDA-enabled `jammi-server` binary behind a `jammi-server` console
script.

```console
pip install jammi-server-cu12
jammi-server --help
```

The CUDA runtime libraries (`libcudart`, `libcublas`, `libcublasLt`,
`libcurand`, `libnvrtc`) are pulled in as `nvidia-*-cu12` wheel dependencies; the
console script prepends their `lib/` directories to `LD_LIBRARY_PATH` before
exec'ing the binary, so no system CUDA install is required (only an NVIDIA driver
on the host). This package and `jammi-server` (CPU) both provide the
`jammi-server` command — install exactly one.

## Requirements

- **NVIDIA GPU** of compute capability **8.0 or newer** (Ampere A100, A10/A6000,
  Ada L4/L40S, Hopper H100). The kernels are built at `compute_80` and the driver
  JIT-forwards them to 8.6 / 8.9 / 9.0. Turing (7.5, e.g. Tesla T4) is **not
  supported** by this build.
- **NVIDIA driver r560 or newer** (Linux: ≥ `560.28.03`). The kernels ship as PTX
  compiled with the CUDA 12.6 toolkit, which the deployment driver JIT-compiles at
  model load; a driver below the CUDA 12.6 line (for example `550.x`, which tops
  out at CUDA 12.4) cannot compile that PTX and the server exits at startup with a
  clear driver-too-old error rather than a raw `CUDA_ERROR_UNSUPPORTED_PTX_VERSION`
  at first inference. `nvidia-smi` shows the installed driver and its max CUDA
  version.
