rakka-inference-core 0.2.6

Foundation crate for rakka-inference — typed errors, deployment value object, ModelRunner trait, batch primitives.
Documentation

inference-core

Foundation types for the rakka-inference workspace. Zero actor / GPU / HTTP dependencies — pure types that every other layer plugs into.

What it gives you

Type Purpose
ModelRunner The trait every backend implements (local GPU + remote network).
Deployment The shared declarative surface — what runs, where, with what limits.
ExecuteBatch / RunHandle Request input + streaming output handle.
TokenChunk / Tokens Per-chunk and aggregate output, with usage + cost.
InferenceError Typed error surface (rate-limited, circuit-open, content-filtered, context-length, …).
RuntimeKind / TransportKind / ProviderKind Backend identity used by placement, observability, dispatcher choice.
RateLimits / RetryPolicy / Timeouts / CircuitBreakerConfig Per-deployment policy primitives.
Secret<T> / SecretString Typed credentials — won't Debug, won't Display.
infer_runtime(model) Default backend selection from a model name (gpt-* → OpenAI, etc).

Why depend on this directly

If you're building a third-party runtime, this is the only crate you need. Implement ModelRunner and you slot into the rest of the workspace — gateway, supervision, distributed rate limiting, circuit breaking, streaming — for free.

use inference_core::{
    ExecuteBatch, InferenceResult, ModelRunner, RunHandle, RuntimeKind,
    SessionRebuildCause, TransportKind,
};

pub struct MyCustomRunner { /* ... */ }

#[async_trait::async_trait]
impl ModelRunner for MyCustomRunner {
    async fn execute(&mut self, batch: ExecuteBatch) -> InferenceResult<RunHandle> { /* ... */ }
    async fn rebuild_session(&mut self, _cause: SessionRebuildCause) -> InferenceResult<()> { Ok(()) }
    fn runtime_kind(&self) -> RuntimeKind { RuntimeKind::Custom("my-backend".into()) }
    fn transport_kind(&self) -> TransportKind { TransportKind::LocalGpu }
}

Dependency budget

inference-core depends only on:

  • serde / serde_json / thiserror / bytes — wire types
  • secrecy — typed secrets
  • async-trait — for the ModelRunner trait (documented exception)
  • futures — BoxStream for RunHandle
  • chrono (no-default-features, clock only) — for Tokens timestamps
  • url — for endpoint validation

No tokio, no rakka, no rakka-accel, no pyo3, no reqwest. This is what makes cargo build -p inference --features remote-only produce a binary with zero GPU deps — the foundation layer simply doesn't carry any.