This repository was archived by the owner on Oct 8, 2026. It is now read-only.
Repository navigation
v0.0.2 has landed #123
elibol
announced in
Announcements
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
v0.0.2 is now available on crates.io:
This release is focused on making cuTile Rust easier to install, easier to integrate, and more consistent across the host API, device DSL, and Tile IR compiler path.
Highlights
Low-friction crates.io install. Downstream projects can now depend on
cutile = "0.0.2"directly without requiring LLVM/MLIR installed on the system.Substantially improved host-side execution APIs. The
DeviceOpredesign adds a cleaner composition model, CUDA graph capture support, borrowed tensor inputs, mutable tensor partitions, and a unified kernel launch API. Docs: Device OperationsCleaner host and device APIs. The host API is more consistent around tensor ownership, borrowing, kernel launches, and third-party binding interop. On the device side, the low-level Tile IR operations in the DSL more closely track the Tile IR operation API. Docs: Host API | DSL API
Better JIT compiler lowering. The JIT compiler now uses proper lowering passes through a pass manager, including static dispatch lowering and type inference. Docs: Compiler README
Optional cuda-tile-rs tooling.
cuda-tile-rsis now an optional workspace crate with minimal builder tools and translation support for bidirectional MLIR and Tile IR bytecode translation. Docs: cuda-tile-rs READMERoadmap
Current issue priorities are tracked with
P0,P1, andP2labels.P0:
P1:
P2:
Please continue to provide feedback, especially on the v0.0.2 API and the roadmap items above.
All reactions