offgrid-sys 0.2.2

# offgrid-sys

Raw bindgen bindings to [llama.cpp](https://github.com/ggml-org/llama.cpp) and mtmd, built
from the pinned `llama.cpp` submodule (always in CMake `Release`). The safe wrapper is
the `offgrid` crate.

On Windows laptops with an AMD iGPU plus another GPU, AMD's switchable-graphics Vulkan layer
hangs `vkCreateInstance`. `offgrid` sets `DISABLE_LAYER_AMD_SWITCHABLE_GRAPHICS_1=1` before
initializing; code calling `llama_backend_init` directly should do the same.

## Features

| Feature | Default | Effect |
|---|---|---|
| `vulkan` | yes | Enables the GGML Vulkan backend (`GGML_VULKAN=ON`) and links `vulkan`. |
| `cuda` | no | Enables the GGML CUDA backend (`GGML_CUDA=ON`). Requires a CUDA toolkit. |
| `bindgen` | no | Regenerate the bindings with bindgen (needs libclang) instead of using the committed `src/bindings/{linux,windows}.rs`; required on other targets. |
| `dynamic-backends` | no | Shared libs with runtime-loaded backends and all CPU variants (used by the packages). |

CPU-only builds: `cargo build -p offgrid-sys --no-default-features`.

## Environment variables

| Variable | Effect |
|---|---|
| `OFFGRID_NATIVE=1` | Sets `GGML_NATIVE=ON` (tune the CPU backend for the host CPU). Off by default so artifacts are portable. |
| `OFFGRID_TEST_MODEL` | GGUF model for the `tokenize_with_model` smoke test (skipped when unset). |

## Tests

```bash
# Device enumeration only
cargo test -p offgrid-sys -- --nocapture

# With model tokenization
OFFGRID_TEST_MODEL=/path/to/model.gguf cargo test -p offgrid-sys -- --nocapture

# CPU-only build (no Vulkan)
cargo build -p offgrid-sys --no-default-features
```

## Bumping the pinned llama.cpp commit

```bash
cd crates/offgrid-sys/llama.cpp
git fetch --depth 1 origin <new-commit-sha>
git checkout <new-commit-sha>
cd ../../..
scripts/gen-bindings.sh            # regenerate src/bindings/{linux,windows}.rs
git add crates/offgrid-sys/llama.cpp crates/offgrid-sys/src/bindings
cargo clean -p offgrid-sys && cargo build -p offgrid-sys
```