Infrastructure Modal 高级用法

# Modal 高级用法指南 ## 多 GPU 训练 ### 单节点多 GPU python import modal app = modal.App("multi-gpu-training") image = modal.Image.debian_slim().pip_install("torch", "transformers", "accelerate") @app.function(gpu="H100:4", image=image, timeout=7200) def train_multi_gpu(): from accelerate import Accelerator accelerator = Accelerator() model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader) for batch in dataloader: outputs = model(**batch) loss = outputs.loss accelerator.backward(loss) optimizer.step() ### DeepSpeed 集成 python image = modal.Image.debian_slim().pip_install( "torch", "transformers", "deepspeed", "accelerate" ) @app.function(gpu="A100:8", image=image, timeout=14400) def deepspeed_train(config: dict): from transformers import Trainer, TrainingArguments args = TrainingArguments( output_dir="/outputs", deepspeed="ds_config.json", fp16=True, per_device_train_batch_size=4, gradient_accumulation_steps=4 ) trainer = Trainer(model=model, args=args, train_dataset=dataset) trainer.train() ### 多 GPU 注意事项 对于重新执行 Python 入口点的框架（如 PyTorch Lightning），请使用： - `ddp_spawn` 或 `ddp_notebook` 策略 - 将训练作为子进程运行以避免问题 python @app.function(gpu="H100:4") def train_with_subprocess(): import subprocess subprocess.run(["python", "-m", "torch.distributed.launch", "train.py"]) ## 高级容器配置 ### 用于缓存的多阶段构建 python # 阶段 1: 基础依赖 (已缓存) base_image = modal.Image.debian_slim().pip_install("torch", "numpy", "scipy") # 阶段 2: ML 库 (单独缓存) ml_image = base_image.pip_install("transformers", "datasets", "accelerate") # 阶段 3: 自定义代码 (更改时重建) final_image = ml_image.copy_local_dir("./src", "/app/src") ### 自定义 Dockerfile python image = modal.Image.from_dockerfile("./Dockerfile") ### 从 Git 安装 python image = modal.Image.debian_slim().pip_install( "git+https://github.com/huggingface/transformers.git@main" ) ### 使用 uv 加速安装 python image = modal.Image.debian_slim().uv_pip_install( "torch", "transformers", "accelerate" ) ## 高级类模式 ### 生命周期钩子 python @app.cls(gpu="A10G") class InferenceService: # ...