Skip to content

ruvllm_sparse_attention v0.1.1 — FastGRNN-gated near-linear attention + no_std/ESP32-S3

Choose a tag to compare

View all tags
@ruvnet ruvnet released this 07 May 15:14
ruvllm_sparse_attention-v0.1.1
42021ba

What's new

ruvllm_sparse_attention v0.1.1 is the first release that runs on bare-metal ESP32-class hardware and drops attention's per-token cost from O(N log N) to near-linear O(N) via a FastGRNN salience gate.

📦 crates.io · 📚 docs.rs · 🧪 tutorial gist · 📋 PR #429

Highlights

  • FastGrnnGate + forward_gated (new public API) — recurrent O(N · D_h²) salience pass + gated attention forward. Combined cost O(N · (D_h² + W + G + K_keep + dim)) is linear in seq when K_keep is constant. Critical invariant: forward_gated with all-true mask is bit-identical to forward.
  • no_std + alloc support (ADR-192) — new std feature default-on, so existing consumers see zero behavioural change. Bare-metal targets opt out via default-features = false. Verified on attached ESP32-S3 (Xtensa LX7 @ 240 MHz, 16 MB flash) — release build 1.02 s, 376 KB rlib.
  • ADR-191 (proposed) — Pi Zero 2W production hardening: decode-deadline API, warm-up hook, pi_zero_2w() config preset.
  • Plain-language README with FAQ, scaling table, and SEO keywords. New end-to-end tutorial.
  • libm 0.2 added as always-on dep (~60 KB pure Rust). Restores f32::exp/sqrt/tanh/powi method syntax in no_std mode via an F32Ext trait — all 46 math call sites unchanged.

Verification matrix

Build Result
cargo test --lib 38/38 pass
cargo build --no-default-features clean
cargo build --no-default-features --features fp16 clean
cargo +esp build --release --target xtensa-esp32s3-none-elf -Z build-std=core,alloc clean (1.02 s, 376 KB rlib)
Native run of examples/esp32s3_smoke esp32s3_smoke: all checks passed

Migration

Zero action for std consumers — std is the default feature. New no_std consumers (ESP32 / Cortex-M / RISC-V MCU):

ruvllm_sparse_attention = { version = "0.1.1", default-features = false, features = ["fp16"] }

Bring your own allocator + panic handler + #[entry] (embedded-alloc, linked_list_allocator, esp-hal, xtensa-lx-rt).

What's next

  • ADR-191 implementation (decode-deadline API, warm-up hook, Pi Zero 2W preset)
  • Flash-able example app crate (embedded-alloc + esp-hal scaffolding around examples/esp32s3_smoke.rs)
Assets 2
Loading