Benchmark each new/modified fastkernels operator against its SOTA-reference counterpart for correctness (same init/inputs -> same output) and performance (>= reference); fix and iterate. Invoke with /fastkernels-bench-ops <model/arch> [reference lib].
Adversarially check a newly added/modified fastkernels model against its SOTA reference library, hunting for interface/implementation mismatches and shortcuts; fix and iterate to zero differences. Invoke with /fastkernels-check-parity <model/arch> [reference…
Validate a fastkernels model end-to-end against its SOTA reference with tests/bench_vllm.py — token alignment >= 100 and speedup >= 1.0x, eager and compiled, small then full workload, TP if multi-GPU. Invoke with /fastkernels-validate-e2e <HF model id>.
Implement a new model architecture in fastkernels from a SOTA reference library (vLLM or SGLang), matching its interface, kernels, libraries, and dtypes exactly — no shortcuts. Invoke with /fastkernels-add-model <HF model id or arch> [reference lib].