# E21f: the longest context a 4B model holds fully on one RTX 3080 10 GB, plus a 9B at 262K

Credit: the memory saving that makes these fit comes from upstream llama.cpp PR 29827, "CUDA: cap FA convert buffer for quantized KV" (unmerged at time of test). It is not our work.

- `e21f_chain.ps1`: starts `llama-server` (PR 29827 build) with q4_0 K and V cache, flash attention, ubatch 64. After the model loads it checks the process's shared GPU memory; above 300 MiB the run is logged as SKIPPED-WONTFIT instead of running with spill. Then one needle prompt per run, greedy, 24-token answer limit.
- `results.jsonl`: one row per run (raw server reply dropped). The 9B 10% row is regraded: the script's pass flag matched the whole server reply rather than the answer text, and the answer had no passcode.

Rerun (Windows, PowerShell): put the PR 29827 build and the GGUFs under `C:\rig`, then `powershell -ExecutionPolicy Bypass -File e21f_chain.ps1`.