What problem does it solve? It provides a single methodology for developing and using the mindie_bench benchmark toolchain to measure and compare FA/BSA/GMM/MM operator implementations on NPU, so model tuning decisions (dtype, quantization, sparsity, head shapes) are based on real measured MFU/latency data instead of guesswork. ## Core Features & Use Cases - Operator implementation selection: Run mindie_bench with --op and --config to compare real kernel dispatch across dtype/quant_algo (e.g., bf16 vs mxfp4 FA), sparsity levels, and head configurations, then merge runs into a single HTML report for MFU/latency comparison. - Toolchain development: Extend the benchmark with new operators via a five-step flow (op_defs accounting, NPU vendor kernel, schema tables, CLI registration, unit tests) following test-first development and single-source metrics conventions. - Data anomaly troubleshooting: Diagnose constant latency, all-zero outputs, MFU clamping, and incremental merge pitfalls using the operator-UT-first triage method and data validity guards. - Use Case: During model optimization (S1/S4 selection), compare bf16 vs mxfp4 FA implementations at the same seqlen with a shared peak_flops input, producing operator-level evidence for quantization and sparsity decisions. ## Quick Start Use the benchmark-dev skill to compare FA implementations across bf16 and mxfp4 dtypes at seqlen 8192 with mindie_bench and merge the runs into one HTML report.
