v0.32.2-rc0: CUDA v12 の Linux 向けに CC 10.0 サポートを追加v0.32.2-rc0: cuda: add CC 10.0 for linux in CUDA v12 (#17025)

Ollama has published v0.32.2-rc0, a release candidate that adds Compute Capability 10.0 (CC 10.0) support for CUDA v12 on Linux. The change matters for anyone running local large language models on the newest NVIDIA hardware, because without a matching compiled kernel, those GPUs may fall back to CPU execution or fail to load models efficiently.

Compute Capability is NVIDIA's versioning scheme that identifies the feature set and instruction support of a given GPU architecture. Each generation carries a number—for example, CC 8.x maps to Ampere and CC 9.0 to Hopper—and software that targets CUDA typically ships precompiled kernels for a specific list of these values. CC 10.0 corresponds to the latest generation of NVIDIA GPUs based on the Blackwell data-center architecture. By adding this target, Ollama can build and dispatch optimized code paths for those chips rather than relying on forward-compatible fallbacks, which are not always available or performant.

Ollama is an open-source tool that packages and runs LLMs locally, wrapping the widely used llama.cpp inference engine behind a simple command-line interface and HTTP API. It handles model downloading, quantization formats such as GGUF, and hardware acceleration across CUDA, ROCm, Metal, and CPU backends. Because the project distributes prebuilt binaries, the set of Compute Capabilities baked into its CUDA builds directly determines which GPUs receive native acceleration out of the box. Extending that list to CC 10.0 is therefore a prerequisite for smooth support of Blackwell-class accelerators on Linux.

The "-rc0" suffix indicates this is the first release candidate for v0.32.2, not a final stable build. Release candidates are intended for testing and validation before a general release; users who lack the relevant hardware or who need production stability may prefer to wait for the tagged stable version. The scoping of the change to CUDA v12 reflects the fact that newer architectures generally require a recent CUDA toolkit; older CUDA 11 builds do not include the compiler support needed to target CC 10.0.

For context, GPU acceleration is central to practical local inference. Model weights and the key-value cache consume large amounts of memory bandwidth, and modern data-center GPUs offer very high memory capacity and throughput that can run larger models or serve more concurrent requests than consumer cards. That said, the majority of Ollama users run consumer GeForce and workstation RTX cards; the consumer Blackwell line, sold as the RTX 50 series, reportedly uses a different Compute Capability and is handled by separate build targets. This particular change appears aimed primarily at the data-center Blackwell parts.

The update also fits a broader industry pattern. Runtime and framework maintainers, including PyTorch, TensorRT-LLM, vLLM, and llama.cpp itself, have been progressively adding Blackwell support as the hardware reaches customers, and each layer of the stack must be updated in turn. Ollama sits near the top of that stack, so it benefits from upstream CUDA and llama.cpp work while still needing its own build configuration to expose new architectures. Keeping the compiled kernel list current is routine but necessary maintenance as NVIDIA's lineup evolves.

Practically, the addition means Linux users with CC 10.0 GPUs should be able to install this release candidate and see native CUDA acceleration without manual recompilation. Those on Windows, or on other GPU vendors, are unaffected by this specific note. As with any pre-release, behavior may change before the stable v0.32.2 ships, and anyone deploying on new silicon is advised to verify driver versions, confirm the installed CUDA runtime, and benchmark against their expected workloads. Users can typically report issues through the project's GitHub repository, which is also where release notes and binaries are published.

Overall, v0.32.2-rc0 is an incremental but meaningful step that broadens Ollama's hardware coverage. It does not introduce new user-facing features so much as ensure that the latest NVIDIA accelerators are recognized and used efficiently, keeping the tool aligned with the current generation of GPU hardware.