Episode 05 · Planned
Attention kernels and precision
Compare verified attention backends, then separately compare weight and KV precision.
Compare two supported attention backends with verified dispatch, then run a separate weight/KV precision comparison. Unsupported combinations are excluded rather than forced. Record the actual attention backend and version; “Flash on/off” is insufficient provenance.
Evidence: compatibility record, profiler confirmation where available, memory use, latency, useful throughput and results on the fixed quality set against its acceptance threshold.