memra-kv 0.72.0

KV-cache format policy (q8_0/q5_1/q4_0/fp8 block layouts) for the memra inference engine

Coverage
43.14%
22 out of 51 items documented0 out of 24 items with examples
Size
Source code size: 25.22 kB This is the summed size of all the files inside the crates.io package for this release.
Documentation size: 454.51 kB This is the summed size of all files generated by rustdoc for all configured targets
Ø build duration
this release: 4s Average build duration of successful builds.
all releases: 5s Average build duration of successful builds in releases after 2024-10-23.
Links
avifenesh/memra
304 35 0
crates.io
Dependencies
- cudarc ^0.19 normal
- memra-gguf =0.72.0 normal
Versions
Owners

memra-kv — the dual KV/recurrent cache, extracted (Phase D, ARCHITECTURE-H100.md §5).

Moved VERBATIM from memra-engine/src/cache.rs behind the KvDev seam: the cache only ever needed 7 device ops (alloc/copy/set), so the trait is that surface and nothing more. The append/dequant KERNELS stay in the engine fatbins — this crate owns the structure, sizing math, and the KV format policy (env-selected, shared by the engine's fatbin router and every cache consumer). memra-engine re-exports this as cache so call sites are unchanged.