Training a Model with Limited Memory using Mixed Precision and Gradient Checkpointing

2025-12-24 · Source: MachineLearningMastery.com - Machinelearningmastery.com · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Intermediate, long

Summary

This article details techniques for training large language models in memory-constrained environments, focusing on mixed-precision training and gradient checkpointing. It explains various floating-point number types, including `float32`, `float16`, and `bfloat16`, highlighting `bfloat16`'s advantage in deep learning due to its `float32`-like dynamic range. The content describes how PyTorch's `torch.amp` sub-library automates mixed-precision training, dynamically casting tensors to lower precision where robust, and introduces `GradScaler` to prevent vanishing gradients. Additionally, it covers gradient checkpointing, a method that recomputes intermediate results during the backward pass to save memory at the cost of increased computation time, demonstrating its application to transformer blocks in a `LlamaModel`.

Key takeaway

For Deep Learning Engineers training large models on GPUs with limited VRAM, adopting mixed-precision training with `bfloat16` and implementing gradient checkpointing are crucial. These techniques allow you to fit larger models or batch sizes into memory, significantly improving training efficiency. Ensure you correctly integrate `torch.amp.GradScaler` and save its state, and consider `scaler.unscale_()` before gradient clipping for accurate updates.

Key insights

Mixed precision and gradient checkpointing enable training large models on memory-limited hardware.

Principles

• Dynamic range is critical for deep learning.
• Trade time for memory in gradient computation.
• Not all operations tolerate lower precision.

Method

Use `torch.amp.autocast()` for forward pass and `GradScaler` for backward pass in mixed-precision training. Implement `torch.utils.checkpoint.checkpoint()` to recompute intermediate activations during backpropagation.

In practice

• Set `torch.set_default_dtype(torch.bfloat16)` to save memory.
• Wrap forward pass with `torch.autocast()` for mixed precision.
• Save `scaler.state_dict()` when checkpointing models.

Topics

• Memory-constrained Training
• Mixed Precision Training
• Gradient Checkpointing
• Floating-point Data Types
• PyTorch

Best for: Deep Learning Engineer, Machine Learning Engineer, AI Engineer

Related on AIssential

• stochastic gradient descent #maths #statistics #datascience #machinelearning #dataanlysis — Faster iteration with noisy gradients often outperforms slow, precise convergence in machine learning.
• Resilient Large-Scale Training: Integrating TorchFT with TorchTitan on AMD GPUs — Decoupling parallelism from fault tolerance enables resilient, checkpoint-less distributed training for large AI models.
• Synthetic Benchmarks Overstate Forward-Forward Scaling: Real-Data Limits of Layer-Local Training — Layer-local Forward-Forward training, even at SOTA, struggles to scale to real-world data and offers no practical memory advantage over optimized backpropagation.
• Training Time Prediction for Mixed Precision-based Distributed Training — Precision settings significantly impact distributed deep learning training times, necessitating precision-aware prediction models.
• Part 12 -The 80GB Wall: GPU Infrastructure and Scheduling, Worked End to End — Infrastructure decisions regarding GPU memory, scheduling, and priority tiers are critical and must be made proactively to prevent costly LLM training failures.
• Train a Model Faster with torch.compile and Gradient Accumulation — Optimize PyTorch model training using `torch.compile()` for speed and gradient accumulation for larger effective batch sizes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by MachineLearningMastery.com - Machinelearningmastery.com.