Repository navigation
- Faster speculative decoding with multi-token GDN/causal-convolution updates and optimized vocabulary-wide cumsums and gathers.
- Improved multi-device execution by reusing the active mesh and keeping more copies on-device.
- Faster host transfers through device-side untilization, stream-ordered reads, and larger zero-copy uploads.
- Fixed sharding, aliased custom-call, paged-cache, and unit-dimension gather issues.
- Added chunked-attention support for page tables of any width.
Full changelog: v0.1.4...v0.1.5