Integrate fused SwiGLU decode into Llama 3.2 1B#80

Draft

jgmelber wants to merge 11 commits into

develfrom

decode-fusion-llama

Commits on Mar 7, 2026

Fuse SiLU + elementwise multiply into single AIE kernel and operator

jgmelber
and
claude
committed
Fuse dual-GEMV + SiLU + Mul into single NPU design for SwiGLU decode

jgmelber
and
claude
committed
Add decode dataflow operators: fused SwiGLU, QKV proj, FlowKV attention, INT4 dequant-GEMV

jgmelber
and
claude
committed
Fix C++ formatting for CI clang-format compliance

jgmelber
and
claude
committed
Fix flowkv.cc clang-format: one param per line, join short expressions

jgmelber
and
claude
committed
On-chip reduction for swiglu_fused_decode, fused RoPE in flowkv_decode

jgmelber
and
claude
committed
Update README benchmarks with on-chip reduction results

jgmelber
and
claude
committed
Revert swiglu_fused_decode to DDR-partials reduction (faster)

jgmelber
and
claude
committed
Integrate swiglu_fused_decode into Llama 3.2 1B application

jgmelber
and
claude
committed
Skip redundant decode FFN operators when swiglu_fused is active

jgmelber
and
claude
committed
Enable fused SwiGLU decode path in Llama config

jgmelber
and
claude
committed