GPU cloud hosting for people who actually ship models

# PyTorch training on H100
docker run --gpus all -v $(pwd):/workspace \
  nvcr.io/nvidia/pytorch:24.01-py3 \
  python train.py --fp16 --batch-size 32

# What you paste here runs the same in production.

# vLLM inference server
docker run --gpus all -p 8000:8000 \
  vllm/vllm-openai:latest \
  --model meta-llama/Llama-3-70b-hf \
  --tensor-parallel-size 2

# Scales to multi-GPU with --tensor-parallel-size

# SDXL training with Diffusers
accelerate launch train_text_to_image.py \
  --pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" \
  --train_data_dir="./data" \
  --resolution=1024 \
  --train_batch_size=1 \
  --gradient_accumulation_steps=4 \
  --mixed_precision="bf16"

# Full resolution, full precision training on L40S