Spark Memory & Thermal Ops
Plan unified-memory headroom and monitor temperature and power during long DGX Spark training runs
- Repository
- wshobson/agents
- Stars
- 42,095
- Licence
- MIT
- Files
- 3 files
What it does
Once installed, the agent plans whether a machine learning training run will fit in the DGX Spark's shared 128GB memory pool, using free memory readings rather than the GPU tool's numbers. When a job runs out of memory, it walks through a fixed order of fixes (clear the cache, shrink batch or packing length, then downgrade the training method) instead of jumping straight to smaller settings. It also helps interpret power and temperature during multi-hour jobs and can start a background sampler that logs readings to a CSV file.
Great for
- Engineers fine-tuning large models on an NVIDIA DGX Spark (GB10) who keep hitting out-of-memory errors
- Anyone running multi-hour training jobs who wants to tell thermal slowdowns apart from configuration bugs
- Users planning to run a training job and an inference server (such as vLLM or Ollama) on the same machine
Skip it if
- Machines other than DGX Spark/GB10 with discrete GPU memory, where the memory model and power guidance do not apply
- Launch-time failures such as library version mismatches or flash-attn problems, which the skill says belong to the separate spark-training-gotchas skill
- Anyone who cannot safely run commands as root, since one suggested step needs root and clearing caches is a between-run action
Try asking
- “Will a 70B QLoRA fine-tune with packing length 4096 fit in my Spark's memory before I launch it?”
- “My job ran out of memory on unified memory mid-step. What should I try first, second and third?”
- “Start a thermal sampler logging temperature and power every 30 seconds to thermal.log while my training runs.”
- “My training slowed down two hours in. Is this thermal throttling or a config problem?”
Install in 3 steps
- 1
Download the skill
The ZIP contains the 3 original files, the MIT licence and an attribution note.
spark-memory-thermal-ops.zip - 2
Add it to your agent
Claude Code (all projects)
mkdir -p ~/.claude/skills && unzip spark-memory-thermal-ops.zip -d ~/.claude/skills/Claude Code / Agent SDK (this project only)
mkdir -p .claude/skills && unzip spark-memory-thermal-ops.zip -d .claude/skills/Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.
- 3
Ask for the task
No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.
Get better results
- Budget memory using the free -g reading minus a few GB for overhead, not the 128GB spec or nvidia-smi output, and compare against the listed anchors.
- Start the thermal sampler before launching the job so timestamps line up with training logs, and review the script's output file afterwards.
- Before running the cache flush command, confirm you are on a between-run reset and that you have root rights, as the skill recommends.
What's inside
- assets/thermal-sample.sh
- references/uma-accounting.md
- SKILL.md
Redistributed unmodified from wshobson/agents under the MIT licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.