القائمة
Use when GPU services (Ollama, Whisper, ComfyUI/Flux, OCR) contend for VRAM on the RTX 3090 and hit CUDA OOM.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
Anthropic API rate limit handling - retry logic, backoff, throttling for batch workloads against Claude models
Use when building an automated test → issue → fix loop with Claude Code and GitHub issues — overnight auto-fixing, regression loops, self-healing CI.
Use when creating, editing, publishing, or deleting posts on Cyril's Workshop blog or the steponnopets.net devblog.
Use when writing or contributing a boofuzz network-protocol fuzzer in this repo — layout, formatting rules, and reading results.
Use when a task needs real-time control of a connected browser via the Browser Bridge Broker — submit JS jobs over HTTP that browsers eval and return.
Use when training a character LoRA (Chroma/Flux or Pony/SDXL) on a RunPod GPU and wiring it into the ComfyUI + pony_web render stack.
| name | Vram-GPU-OOM |
| description | Use when GPU services (Ollama, Whisper, ComfyUI/Flux, OCR) contend for VRAM on the RTX 3090 and hit CUDA OOM. |
Multiple services share one RTX 3090 (24GB). They coordinate without a central scheduler: everyone tries to load normally, catches OOM, waits for others to auto-unload, and retries.
On CUDA OOM: torch.cuda.empty_cache(), time.sleep(30), retry — up to 3 attempts, 30s apart. Re-raise non-OOM errors immediately and re-raise after the final attempt. Same idea in shell: loop a GPU command 3 times with a 30s sleep between failures. Alongside retry, configure every service to unload quickly when idle.
/etc/systemd/system/ollama.service.d/override.conf:
Environment="OLLAMA_KEEP_ALIVE=30s"
http://10.99.0.3:8765 — auto-unloads after idle (--auto-unload-minutes, default 5), does 3×/30s OOM retry, and exposes POST /request-unload and GET /status.For faster, more predictable starts, a service can call POST /request-unload on the others before loading a big model instead of relying on OOM-retry delays. The endpoint contracts (/request-unload, /status, the auto-unload background task), the coordinator usage pattern, and worked timelines are in reference-gpu-coordination.md. Helper script: request_gpu_unload.py in the OneCuriousRabbit repo.