[michalharakal]

Engine 0.61.0 and the Laya consumer of SKEEP-007 (quantize-at-load, the browser lane).

Engine bump

skainet = "0.61.0": SKEEP-007 Tiers 0–2 (packed-weight dispatch order, the skainet-backend-wasm-simd sidecar, quantize-at-load for the SafeTensors and GGUF loaders) and the scheduled removal of the SKEEP-006 compatibility aliases — nothing in this repository referenced them.

Laya on the packed Q8_0 lane

• LayaModelLoader.load is common code (commonMain): the four side files as text (LayaCheckpointFiles) plus a RandomAccessSource for model.safetensors, suspend, with weightForm / weightFormFor passed through to the engine's SafeTensors loader. The JVM load(directory: Path, ctx, exactGelu, weightFormFor) is now an extension over it — source-compatible for Kotlin callers (same call), a different owner class for Java callers (API dump updated). The Android replica of the six steps in the examples repository can collapse onto this function.
• LayaWeightForms.q8_0For: Q8_0 for the encoder and head projections (Wqkv, Wo, Wi, mlp.Wo, in_proj_weight, out_proj.weight, linear1/2.weight) and the 256k × 768 token table; norms, biases, the type embedding and the scorer stay dense. About a quarter of the FP32 bytes, served by the packed kernels on every tier (FFM on the JVM, JNI on Android, the SIMD128 sidecar in a browser after WasmSimd.install()).
• LayaWeightPreparation splits the fused Wqkv / Wi / in_proj matrices in their packed form (in row-major block storage a row block is a contiguous byte range, so no FP32 transient) and wraps the token table as PackedRowDequantTensorData, which the embedding layer reads row by row.
• Decision-quality gate: LayaQuantizedParityTest (model-gated like the other fixture tests) loads the FP32 and the Q8_0 agent, runs the predict.json single cases through both, and requires every choice to agree and no probability, score, noul value or confidence to shift by more than 0.05 (LAYA_Q8_TOL; measured 0.035). FP32 parity (1.5e-4) stays the contract of the reference lane. The fixture question parser moved to LayaFixtures.questionOf so both tests share it.
• Docs: typed-decisions.adoc gains a Quantized lane section and its Current limits no longer lists "JVM only"; CHANGELOG Unreleased.

Verification (local, 12-core laptop, engine 0.61.0 from Maven Central)

• compileKotlinJvm and compileTestKotlinJvm across the repository against 0.61.0: clean (nothing referenced the removed engine aliases).
• :llm-inference:laya:jvmTest with LAYA_MULTILINGUAL_PATH: 33 tests, the FP32 fixtures and batching tests unchanged.
• LayaQuantizedParityTest: 5/5 choices agree, 45 values compared, max shift 0.035 against the 0.05 gate; packed weights 325 MB; FP32 load 11.4 s, Q8_0 load 12.1 s. On the eager JVM path the Q8_0 lane runs the six fixture cases in 171 s against 69 s for FP32 — the row-major packed kernels take one activation row at a time — so on the JVM this lane is a memory measure; the speed is the browser sidecar's.
• apiDump for the laya module regenerated (the JVM load(Path, …) is now in LayaModelLoaderKt).