moe_combine_{f16,bf16,f32} replaces the broadcast_mul+sum expert-combine (fast_sum 487->8ms);
f32-direct activation quantize kills a 178ms cast. Qwen3-30B-A3B-Q4_K_M prefill 939->1167 T/s
(+19.5% isolated A/B, 1.11x of llama.cpp-HIP on the same gfx1151), decode flat. nbad=0
(moe_combine_numeric f16/bf16/f32 + existing MoE oracles). Stacks on the 0.11.7 fused GEMM.
Run the tests with:
RUST_LOG=wasm_bindgen_test_runner wasm-pack test --chrome --headless
Or:
wasm-pack test --chrome
If you get an "invalid session id" failure in headless mode, check that logs and it may well be that your ChromeDriver is not at the same version as your browser.