[gspark-etri]

Background

ATOM was verified end to end, but RNGD had only storage numbers (#54), and no engine had served a model from an onggi volume (#55).

Changes

• docs/evaluation.md:
  • a summary matrix of the three levels (store less, node delta, NPU delta) and real-model serving, for ATOM, RNGD and the A100;
  • an RNGD section: four furiosa-llm artifacts, onggi diff --ranges, CSI cold and delta apply;
  • a section on serving real models from onggi volumes on ATOM and RNGD: tokens and pod-to-first-token times.

Verification

Cluster runs on npu-node-1 in a test namespace, deleted afterwards. Each engine Job held one card and ran to completion.

• RNGD storage: 32.44 GB → 10.79 GB. pfbatch → preset: 99.77 % at the same offset, 4.3 MB new. CSI preset by delta apply: Running in 4 s, 16.15 GB in place; every file identical.
• ATOM real model: Llama-3.1-8B from an onggi volume gives 48 greedy tokens identical to the original directory, cold, warm and after delta apply. First token 28.4 s via delta apply, 22.5 s from the original directory.
• RNGD real model: furiosa-llm 2026.3.0 gives tokens identical in every run. First token 21.8 s via delta apply (0 B from the store), 18.9 s from the original directory.
• Plugin fixes: the two problems the run found are fixed and merged (CSI: an exiting plugin removes the live plugin's socket #61 socket race, CSI: files of mode 0600 in an artifact are unreadable for non-root pods #62 file modes).

Compatibility and limits

• Docs only.
• A cold ATOM pull (11.58 GB from the store) puts the first token at 191 s.
• One RNGD cold run waited 248.8 s for a card that another host job held. That wait is reported separately.

Related Issue

Closes #54
Closes #55

🤖 Generated with Claude Code