Model fine-tuning, inference servers, workflow automation — hands-on GPU cloud notes from experiments we ran ourselves on gcube. Every benchmark and code sample is reproducible.
We verified multi-node distributed fine-tuning with Axolotl + DeepSpeed on two gcube RTX 5090 replicas. This troubleshooting log covers diagnosing an NCCL hang, fixing launcher and config issues, completing all 519 steps, and a before/after comparison.
We split a Korean ballad made with ACE-Step into vocal, instrumental and per-instrument stems with the open-source source separation model Demucs v4. On a gcube RTX 5060 8 GB, it takes 1.4 seconds for a 30-second track and 5.4 seconds for a 3-minute song, using about 1.5 GB of GPU memory.
We train Ultralytics YOLO26 on a construction site safety gear dataset in a gcube RTX 5060 8 GB container. In 8 minutes of training and 2.6 GB of VRAM, you get a model that spots whether workers are wearing helmets and vests.
We ran the open-source music generation model ACE-Step 1.5 on a gcube RTX 5060 8 GB and generated a 3-minute Korean song in 50.2 seconds with a peak of 4.6 GB VRAM. You can follow along from workload setup to generation code.
A step-by-step walkthrough of installing the MiniMax H3 video generation model on a gcube RTX 5090 workload and generating 5-second and 10-second videos.
Online FP8 quantization isn’t always faster. We compared it head to head with BF16 on an RTX 5090 and confirmed the results across two independent runs.
We served several fine-tuned LoRA adapters at once on a single GPU with vLLM, without merging them, and verified it on a gcube RTX 5090.
Deploy a 4-bit quantized Z-Image-Turbo as a gcube workload, then run automated bulk generation from conditions.json with web monitoring. It runs on an 8 GB GPU too.
We served Qwen3-8B with three engines on a single RTX 5090 and measured throughput and TTFT at concurrency levels of 8, 32 and 64. The data shows when to use vLLM, LMDeploy or Ollama.
We embedded the official gcube docs with BGE-M3, indexed them in Chroma, and connected them to Qwen2.5-7B served with vLLM. The entire pipeline runs in a notebook on a single RTX 5090, with no external APIs.
Deploy n8n as a gcube workload and build three workflows that automate everyday office tasks: PDF summaries, AI news briefings, and meeting notes.
Self-host the Tri-7B small LLM on a gcube RTX 5090 and build a meeting notes → action items pipeline with vLLM + n8n. A Jupyter notebook-based walkthrough that includes fine-tuning.
Hands-on notes from fine-tuning Llama 3.1 8B on a Korean dataset with Unsloth in gcube’s RTX 5090 environment.