release: 0.33.0 — GRU, upsample2d Bilinear export, autodiff coverage fix

[michalharakal]

Release 0.33.0 — bundles everything merged to develop since 0.32.4.

Highlights (user value)

• GRU (Add GRU layer (composed, trainable, StableHLO-exportable) #772 / GRU #217) — the first recurrent layer; composed/unrolled, trainable, StableHLO-exportable, plus a gru(…) network-DSL builder.
• upsample2d Bilinear + StableHLO export (Add upsample2d Bilinear (eager+autodiff) and traceable StableHLO lowering (Nearest + Bilinear) #771) — Nearest & Bilinear lower to fixed reshape/broadcast/dot_general (no custom_call); unblocks resize/FPN export.
• Autodiff correctness fix (Autodiff coverage: close silent dispatch gaps + KSP-generated coverage guard #774) — elu/leakyRelu/permute had backward rules that were never wired into the trace dispatch, so their gradients were silently dropped to null. Now wired + guarded by a KSP-generated coverage test. Also makes cos/sin/tril/gather/indexSelect/unfold/convTranspose1d differentiable, and surfaces isDifferentiable in operators.json.
• Norms compile on stock IREE (fix(compile-hlo): lower layerNorm/rmsNorm/batchNorm to real stablehlo.reduce (compile on stock IREE) #769) — layerNorm/rmsNorm/batchNorm lower to real stablehlo.reduce instead of export-only custom_calls.

⚠️ Breaking

TensorOps.sin / cos / convTranspose1d are now abstract (were default-throwing) so they trace/export and become differentiable. Anyone implementing TensorOps directly must override them — both bundled backends (DefaultCpuOpsBase, VoidTensorOps) already do. Hence the minor bump.

Validation

Full downstream conformance suite — 12/12 models + 33/33 ops — exports, compiles, runs, and validates green end-to-end on IREE (llvm-cpu) against this exact tree (driven via a local-maven 0.32.5-localdev1 build of this branch). Includes the new GRU probe (max_abs_err 5.96e-08 vs the SKaiNET CPU oracle) and upsample2d (nearest 0.0, bilinear 1.19e-07 vs numpy).

Contents (gradle.properties 0.32.4→0.33.0, CHANGELOG.md, README "What's New")

This branch only bumps the version + docs. The code changes are already on develop via #769, #771, #772, #774.

Reviewer notes

• develop also carries perf(native-cpu): Q6_K NEON matmul kernel #768 (Q6_K NEON matmul kernel) from a separate workstream — not authored here; please confirm its changelog wording if you want it called out.
• GitHub reports Dependabot alerts on the default branch (mostly transitive/JS toolchain) — worth a glance but not release-blocking.
• Not tagged / not published — left for the release owner's go.

🤖 Generated with Claude Code