nki-moe-megakernel

General-purpose analysis of AWS Neuron profiler NTFF output. Given NTFF (and optionally NEFF) file paths, extracts and analyzes any aspect of device execution — instruction mix, DMA bandwidth, collective timing, engine utilization, per-op timing, or custom queries against the raw trace.

Generates, reviews, and optimizes AWS Neuron NKI kernels for Trainium2 (trn2) / NeuronCore-v3. Trigger ONLY when the user explicitly targets trn3, trn2, trn1, or inf2 hardware.

Generates, reviews, and optimizes AWS Neuron NKI kernels for Trainium3 (trn3) / NeuronCore-v4. Specializes in MXFP8/MXFP4 quantization (nc_matmul_mx, quantize_mx) and NeuronCore-v4 features. DEFAULT skill for all NKI kernel work in this repo (hardware is trn3). Trigger when the user asks to write, fix, optimize, benchmark, or profile any NKI kernel, mentions Trainium/SBUF/PSUM/HBM/nki.lang/nki.isa/neuron-profile, OR mentions trn3, NeuronCore-v4, MXFP8, MXFP4, microscaling, nc_matmul_mx, or quantize_mx. Prefer this over nki-kernel-optimizer unless the user explicitly targets trn2.

Port models defined in PyTorch to run on AWS Trainium accelerators.