Evaluating 1,350 Agent Sessions on One GPU for $5 via Shared LoRA Adapters
yb2698 · x · 2026-11-06
The author shares GPU-efficiency tricks for agent RL training:
- One vLLM per model, one trainer per harness, each with its own LoRA adapter on shared weights
- For LFM3-1.7B, all three trainers share a single GPU next to the vLLM GPU — agent rollouts spend most time in the sandbox, so trainers mostly wait and take turns instead of idling
- At eval time, one vLLM serves every checkpoint's adapter at once: 9 checkpoints, 1,350 agent sessions, one GPU, under an hour, $5 — versus $90 if each checkpoint were merged into full weights
More from coding & agent
- LLM agents run DNA experiments, rewriting protocols to grab idle lab instruments — bravo_abad · 2026-11-06
- Sentry's next MCP evolution: one CLI login powers both MCP and full CLI access — zeeg · 2026-11-06
- Eon Yukon: serverless engine feeds agents live Iceberg tables, 4x less compute than Spark — dr_cintas · 2026-11-06
- Perplexity launches Teammates: proactive AI coworkers that message you on Slack — inductionheads · 2026-11-06
- Box launches Box Content Retrieval with agentic platform and MCP support at BoxWorks — shashib · 2026-11-06
- Give your agent an interpreter, not tools: 19 min to 2 min on the same model — Itz_hany · 2026-11-06