[gspark-etri]

Background

#55 covered ATOM through rebel.RBLNCompiledModel and RNGD through furiosa-llm. The remaining engines, vLLM with vllm-rbln on ATOM and rngd-pd's own engine on RNGD, had not loaded a model from an onggi volume.

Changes

• docs/evaluation.md: two tables, for vLLM + vllm-rbln on ATOM and rngd-pd on RNGD, with cold and warm runs against the original directory.

Verification

Lab cluster run at onggi daf9884, one card per Job, test namespace deleted afterwards:

• vLLM 0.22.0 + vllm-rbln 0.11.0 (optimum-rbln 0.11 build of Llama-3.1-8B):
  • 48 greedy tokens identical in every run;
  • first token 54.6 s original, 55.2 s warm, 194.0 s cold (17.2 GB materialized).
• rngd-pd loads its compiled graphs from the volume through RNGD_ARTIFACT_CACHE:
  • tokens identical;
  • volume ready in 0.36 s; first token 101.8–118.7 s against 118.6 s original.

Compatibility and limits

• Docs only.
• The vLLM OpenAI server on the node fails every HTTP request because of a host package (prometheus_fastapi_instrumentator). The runs therefore used vLLM's offline API, which is the same engine and plugin.

Related Issue

Refs #55

🤖 Generated with Claude Code