Profile and tune Llaminar V2 CUDA kernels using LLAMINAR_PROFILING (per-kernel + per-stage timing), Nsight Systems (nsys), and Nsight Compute (ncu), then validate with the GEMM perf-test harness, the benchmark subcommand, and parity tests. Use when asked to find the slowest CUDA kernel, diagnose occupancy / register pressure / memory-bound stalls, A/B two kernel variants, or close a prefill/decode throughput gap while keeping PyTorch parity.
Tune, debug, or extend Llaminar V2 MTP/speculative decoding, prefix-cache interaction, vLLM-style accepted-state publication, MTP depth control, grouped verifier kernels, dense/MoE MTP parity, and MTP benchmark dashboards. Use when Codex is asked to improve MTP speed, fix MTP correctness, compare CUDA/ROCm/CPU MTP lanes, work on decode-equivalent grouped verifier rows, remove host/device coherence issues, update the MTP project plan/dashboard, or touch MTP-related tests and perf gates.
Profile and tune Llaminar V2 ROCm/HIP INT8 VNNI GEMM and GEMV kernels on AMD Instinct GPUs (gfx906 MI50/MI60) using rocprof per-dispatch timing and LLVM ISA analysis (llvm-objcopy/readelf/objdump), benchmarked against AMD Composable Kernel (CK). Use when asked to find a slow HIP kernel, diagnose occupancy / VGPR pressure / register spills / waitcnt stalls, A/B kernel variants (V1-V7), close a GEMM gap vs CK, or distinguish a real GPU-kernel gap from PCIe/wallclock noise while keeping parity.