CloudFlux
Production-ready templates

Pick a workload. We provision the stack.

Each template selects the right instance type, pre-caches the weights, and ships runnable code. No CUDA, driver or dependency setup.

  1. 1Launch an H100 with the Fine-tune starter. Llama is already cached.
  2. 2Upload your dataset (JSONL with prompt/response) to /data.
  3. 3Run the script below. Checkpoints save every 10%.
  4. 4Download the adapter from /out, or keep it for inference.
    5. Deploy this template
from unsloth import FastLanguageModel
from trl import SFTTrainer
from datasets import load_dataset

model, tok = FastLanguageModel.from_pretrained(
    "/cache/llama-3.1-8b", load_in_4bit=True)  # pre-warmed, no download
model = FastLanguageModel.get_peft_model(model, r=16)

ds = load_dataset("json", data_files="/data/train.jsonl")["train"]
SFTTrainer(model=model, tokenizer=tok, train_dataset=ds,
           args=dict(output_dir="/out", save_steps=0.1,
                     num_train_epochs=1)).train()