Skip to content

cmp-b1 — CUDA 12.9 / SM70

Pre-release
Pre-release

Choose a tag to compare

View all tags
@Josephur Josephur released this 08 Sep 13:15
cmp-b1
16f6894

First SM70 binary release: all eight fork patches, including the Windows research-selector DLL export fix.

Website:

Linux:

Windows:

Requirements

  • NVIDIA CMP 100-210 / CMP 100HX-210 or a compatible SM70 GPU. Native SM70 cubins only; no PTX fallback.
  • An x64 CPU with AVX2, FMA, F16C and BMI2.
  • CUDA 12.9 runtime libraries (cuBLAS, cuBLASLt and cudart) and a compatible NVIDIA driver. CUDA runtime and driver libraries are not bundled. CUDA 13 alone does not provide the required CUDA 12 libraries.
  • Linux: Ubuntu 24.04-compatible glibc, libstdc++, libgomp and OpenSSL 3.
  • Windows: Microsoft Visual C++ x64 runtime, including the OpenMP runtime.

Included

llama-server, llama-cli, llama-bench, llama-quantize, their shared libraries, licenses and release.json source/build provenance. Extract the entire archive and keep the bin directory together.

The embedded web UI is disabled; llama-server provides the HTTP API.

Build and validation

Both platform builds passed with CUDA 12.9.2. Each independently verified that all eight patches reconstruct the public source and that the CUDA library contains only native SM70 cubins. Download checksums, packaged licenses, external dependencies and all six Windows selector exports were reviewed.

Linux help/startup was checked in an isolated CUDA 12.9 environment with a driver stub and no GPU access. Windows runtime execution and GPU inference have not been tested for these release archives; this initial release is marked prerelease.

Source: 16f6894 · Complete patch series

Assets 6
Loading