Build requirements and tests

In one line

provider-local needs a C compiler, the GPU features add more; the live suites run one real model per invocation.

You need

The provider-local feature and a C compiler.

Read this if

Your build fails on a native dependency, or you are running the local live tests.

Build requirements

provider-local owns these dependencies: candle-core and candle-nn 0.11, half, minijinja, minijinja-contrib and unicode-general-category. It reuses reqwest for the download, sha2 for the checksum and tokio. It does not use tokio-stream: the stream is a plain channel receiver. A GPU feature adds burn 0.22 with std, fusion and autotune.

• provider-local needs a C compiler. candle-core 0.11 depends on tokenizers[onig], so the dependency tree compiles the C crate onig_sys. cuca never calls that crate: the tokenizers, the chat templates and the engine code are Rust.
• Every GPU feature needs a C compiler as well, because CubeCL's runtime pulls an embedded store (turso) that builds C. On Windows a dependency's build script runs the SDK resource compiler, rc.exe (windres on the GNU target).
• provider-local-cuda and provider-local-rocm download and link a prebuilt LLVM on the first build, which needs the network once. ROCm builds on Linux only.
• The crate has no default provider, so every cargo command that builds the engine names its feature. candle selects its quantized SIMD kernels when it compiles: a generic build, without RUSTFLAGS="-C target-cpu=native", is several times slower.

Tests

tests/provider_local.rs runs one real model per invocation and needs no server. A scenario applies to a model only when LocalModel::capabilities() says the model can do it, and every request is greedy. A scenario that does not apply prints MATRIX|<scenario>|SKIP|<reason>, and a documented limit of a model prints MATRIX|<scenario>|KNOWN-LIMIT with its evidence. With no model configured, each test prints SKIP: local model not available: ... and passes. The CI test job compiles the suite under ALL_NO_LLVM, so CI exercises no model.

The scenarios:

• greedy_text_and_usage: a factual answer with exact usage. Prompt counts equal the prefix-reuse counts and a max_tokens cap is met to the token.
• thinking_precedes_answer and thinking_off_has_no_reasoning: Thinking blocks precede the answer when thinking is on, and are absent when it is off.
• tool_call_round_trip and tool_result_answer_without_tools: a get_weather call, then an answer from the result, with the tool declared on both requests and with no tool declared on the follow-up.
• long_context_needle: a prompt of 1,000 tokens or more with a code word to retrieve.
• prefix_reuse_hits_and_matches_cold: the warm prefix is read from the cache, an unrelated request shrinks it, and the warm and cold greedy outputs are identical.

tests/provider_local_mtp.rs is the drafter check described under Verified. Which models and devices these suites covered is on Local models and formats.