# E21e: needle retrieval at three depths, 2B at about 1M tokens and 4B at 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.

- `e21e_chain.ps1`: starts `llama-server` (PR 29827 build) with q4_0 K and V cache, flash attention, ubatch 64, then sends one prompt per run: repeated filler text with the needle "The secret passcode for the vault is 7319-4826." at 10%, 50% or 90% depth, greedy (temperature 0). Peak VRAM from nvidia-smi, peak per-process shared GPU memory from the Windows counter.
- `results.jsonl`: one row per run (the raw server reply field was dropped; it only repeated the answer and local paths).

Runs: Qwen3.5-2B Q4_K_M at 1M context (YaRN x4, 996,043-token prompt) at depths 50%, 10%, 90%; Qwen3.5-4B Q4_K_M at 512K (YaRN x2, 504,043-token prompt) at depths 10%, 90%, and a repeat at 50% (the first 50% run was E21d).

Rerun (Windows, PowerShell, RTX 30-series): put the PR 29827 build and the GGUFs where the script expects under `C:\rig`, then `powershell -ExecutionPolicy Bypass -File e21e_chain.ps1`.