mirror of
https://github.com/hanzoai/engine.git
synced 2026-11-20 14:17:52 +00:00
- Demand-sized unified-memory KV cache: size to max_seq_len*max_batch_size (context*concurrency) bounded by the post-model/OS-reserve ceiling, instead of grabbing all free RAM. Fixes the ROCm/APU thrash where an ~88 GB grab on a 128 GB unified box never returned; 30B-A3B now 88 GB -> 24 GB, elastic with concurrency. (hanzo-engine/src/paged_attention/mod.rs) - Vulkan paged-attention wired (is_vulkan branch, opt-in via --paged-attn on): runs and eliminates the naive-Sdpa layout-copy churn, but the v1 scalar kernel regresses decode on this APU, so default OFF until v2 + MoE-glue fusion land. i64/i32 paged-attn index metadata now binds directly (backend coerces to u32). - hanzo-quant: Vulkan/Wgpu storage arms in the distributed pipeline. - Drop the temporary [patch.crates-io] hanzo-ml redirect; pin hanzo-ml 0.11.55. - Manifest hygiene: version on path deps (enso, ffi, server-core, zen3-serving).