Аренда GPU для Hugging Face
запуск за 5 минут,
от 168 ₽/час

# Ubuntu 24.04, CUDA 12.x, Python 3.11+
python3 -m venv /opt/hf && source /opt/hf/bin/activate
pip install --upgrade pip
pip install 'transformers>=5.0' accelerate datasets peft trl \
  bitsandbytes diffusers[torch] hf_xet huggingface_hub

# Авторизация в Hub (token из huggingface.co/settings/tokens):
hf auth login

# Конфиг accelerate под одну A100 (bf16):
accelerate config default --mixed_precision bf16

# Ускорение скачивания крупных весов через Xet-бэкенд:
export HF_XET_HIGH_PERFORMANCE=1

# Trainer + LoRA fine-tuning через PEFT и TRL.
# Запускаем в venv из метода "Установка":
source /opt/hf/bin/activate

python3 - <<'PY'
from transformers import (AutoModelForCausalLM, AutoTokenizer,
                          TrainingArguments)
from peft import LoraConfig, get_peft_model
from trl import SFTTrainer
from datasets import load_dataset

model_id = 'Qwen/Qwen3-14B'
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype='bfloat16', device_map='auto')

lora = LoraConfig(r=16, lora_alpha=32, target_modules='all-linear',
                  lora_dropout=0.05, bias='none', task_type='CAUSAL_LM')
model = get_peft_model(model, lora)

ds = load_dataset('HuggingFaceH4/ultrachat_200k', split='train_sft')
args = TrainingArguments(output_dir='out', per_device_train_batch_size=4,
    gradient_accumulation_steps=4, num_train_epochs=1, bf16=True,
    learning_rate=2e-4, logging_steps=10, save_steps=500,
    max_length=2048)

trainer = SFTTrainer(model=model, args=args, train_dataset=ds,
    processing_class=tok)
trainer.train()
PY

# Запускаем в venv из метода "Установка":
source /opt/hf/bin/activate

# LLM-инференс через pipeline:
python3 - <<'PY'
from transformers import pipeline
pipe = pipeline('text-generation', model='meta-llama/Llama-3.1-8B-Instruct',
                torch_dtype='bfloat16', device_map='auto')
out = pipe('Объясни LoRA в двух предложениях.', max_new_tokens=200)
print(out[0]['generated_text'])
PY

# Image gen через diffusers:
python3 - <<'PY'
from diffusers import StableDiffusionXLPipeline
import torch
sdxl = StableDiffusionXLPipeline.from_pretrained(
    'stabilityai/stable-diffusion-xl-base-1.0',
    torch_dtype=torch.bfloat16, variant='fp16').to('cuda')
img = sdxl('cyberpunk samovar, 4k, octane render').images[0]
img.save('out.png')
PY

# Конфиг под FSDP (4 GPU, bf16, full shard):
accelerate config
# В диалоге: multi-GPU → 4 → no DeepSpeed → FSDP → FULL_SHARD → bf16

# Запуск тренировочного скрипта:
accelerate launch --num_processes 4 train.py \
  --model_name meta-llama/Llama-3.1-70B \
  --dataset HuggingFaceH4/ultrachat_200k \
  --bf16 --gradient_checkpointing

# Multi-node (2 ноды × 4 GPU):
accelerate launch --num_machines 2 --machine_rank 0 \
  --main_process_ip $MASTER_IP --main_process_port 29500 \
  --num_processes 8 train.py