# E21d: needle retrieval at 256K and 512K on one RTX 3080 10 GB

Credit: the memory saving that makes this 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. We built it and ran a needle test with it.

- `e21d_chain.ps1`: starts `llama-server` (PR 29827 build) with Qwen3.5-4B Q4_K_M, q4_0 K and V cache, flash attention, ubatch 64, then sends one prompt per config: filler text with the needle "The secret passcode for the vault is 7319-4826." at 50% depth, greedy (temperature 0). Peak VRAM from nvidia-smi, peak per-process shared GPU memory from the Windows counter.
- `results.jsonl`: one row per config (the raw server reply field was dropped; it only repeated the answer and local paths).

Configs: 256K native; 512K with YaRN x2 and context length overridden to 524,288; and the same 512K server with a 126K-token prompt.

Rerun (Windows, PowerShell, RTX 30-series): put the PR 29827 build and the GGUF where the script expects under `C:\rig`, then `powershell -ExecutionPolicy Bypass -File e21d_chain.ps1`.