# E21 code and measurements

- e21_longctx.py: the 2B/4B sweep driver (llama-server, needle prompts, VRAM and shared-memory capture).
- The 9B run (e21_9b.py) and the ubatch/1M budget script (e21_ub_1m.py) were run on the test rig and are not in this snapshot yet. They use the same llama-server launch as e21_longctx.py:
  llama-server -m Qwen3.5-<size>-Q4_K_M.gguf -ngl 99 -fa on -c <ctx> -np 1 -ctk <q8_0|q4_0> -ctv q4_0 -t 8 --no-mmap
  plus --rope-scaling yarn --rope-scale <ctx/262144> --yarn-orig-ctx 262144 above 262144, and -ub 512/256/128 with -v for the budget runs (--no-kv-offload for the 1M compute measurement).
- Budget results: e21_4b_512k_vram_budget.md, e21_4b_ub_and_1m_budget.md.