training instability
In plain wordsThe tendency of very large training runs to go wrong — the error suddenly jumping or blowing up — which can waste weeks of expensive computing unless engineers prevent or repair it.
Erratic optimization behavior during training — loss spikesloss spike, n.: A sudden jump in a model's error partway through training. Sometimes the model recovers on its own; sometimes the run falls apart and engineers must rewind to an earlier saved copy., slow or sudden divergence to very large or NaN ('not a number') loss, or collapse of internal quantities — that interrupts progress or ruins a run. In large transformerstransformer, n.: The design behind today's chatbots and much of modern AI. It handles text by letting every word look at every other word and decide which of them matter for understanding it. recurring mechanisms include unbounded growth of attention logits (attention entropy collapse), drift of output logits, oversized Adam updates, loss of precision in bfloat16bfloat16, n.: A 16-bit number format invented at Google for AI. It keeps the huge range of the standard 32-bit format but drops precision, which neural networks can usually tolerate. or FP8FP8, abbr.: An 8-bit number format that uses a quarter of the memory of the standard 32-bit format, letting new AI chips train and run models roughly twice as fast as with 16 bits. arithmetic, and poorly scaled initialization; instability usually worsens with model size and learning ratelearning rate, n.: How big a step a model takes each time it adjusts its settings during training. Too large and learning goes haywire; too small and it crawls.. Practice combines prevention — warmuplearning rate warmup, n.: Starting training with tiny learning steps and gradually increasing them over the first stretch of training, so the model does not lurch wildly before it has found its footing., gradient clippinggradient clipping, n.: A safety limit on how big a single training update can be. If the computed adjustment is unusually large, it is scaled down so that one bad batch cannot wreck the model., pre-normalization with RMSNormRMSNorm, n.: A simpler, faster version of the layer that keeps a network's internal numbers at a steady scale. Most recent large language models use it., QK-normQK-norm, n.: A small change to how a language model pays attention — rescaling two sets of internal numbers before comparing them — that prevents a common way large models' training goes haywire., z-lossz-loss, n.: A small extra penalty added during training that stops the raw scores a language model produces before choosing a word from growing huge, which helps keep very large training runs from blowing up., weight decay, μPmaximal update parametrization, n.: A recipe for scaling a neural network's starting values and learning speeds with its size so that settings tuned on a small, cheap model work unchanged on a huge one.-style scaling and careful data filtering — with monitoring and recovery: frequent checkpointscheckpoint, n.: A saved snapshot of a model partway through or at the end of training, like a save point in a video game: training can resume from it after a crash, and finished snapshots are what people download and run., automatic rollback and skipping of offending batches. Because failures may appear only at scale, researchers look for small proxy models that reproduce them at high learning rates.
Training instability forced us to lower the peak learning rate by a third after week two.
Origin
Unstable training has been discussed since early neural networks (exploding gradients, 1990s). It became a central engineering topic with billion-parameter language models from 2019 onward, documented in reports on Megatron-LM, Gopher, OPT-175B, BLOOM and PaLM, and studied directly by Mitchell Wortsman and colleagues at Google (2023) using small-scale proxies.
Usage note
The exploding gradient problemexploding gradient problem, n.: A training failure in which the learning signal grows out of control as it passes back through a network, so the weights jump wildly and training breaks down. is one mechanism of instability, not a synonym. 'Divergence' here means the loss runs away, not a statistical divergence such as KL divergenceKullback–Leibler divergence, n.: A score for the cost of believing the wrong odds: on average, how much more surprised you will be by real outcomes than someone who knew the true probabilities. It is zero only when your odds are exactly right..
Related
- loss spikeloss spike, n.: A sudden jump in a model's error partway through training. Sometimes the model recovers on its own; sometimes the run falls apart and engineers must rewind to an earlier saved copy.
- z-lossz-loss, n.: A small extra penalty added during training that stops the raw scores a language model produces before choosing a word from growing huge, which helps keep very large training runs from blowing up.
- QK-normQK-norm, n.: A small change to how a language model pays attention — rescaling two sets of internal numbers before comparing them — that prevents a common way large models' training goes haywire.
- gradient clippinggradient clipping, n.: A safety limit on how big a single training update can be. If the computed adjustment is unusually large, it is scaled down so that one bad batch cannot wreck the model.
- exploding gradient problemexploding gradient problem, n.: A training failure in which the learning signal grows out of control as it passes back through a network, so the weights jump wildly and training breaks down.
- maximal update parametrizationmaximal update parametrization, n.: A recipe for scaling a neural network's starting values and learning speeds with its size so that settings tuned on a small, cheap model work unchanged on a huge one.
- mixed-precision trainingmixed-precision training, n.: Training a model with shorter, less exact numbers for most of the arithmetic and full-length numbers only where accuracy really matters, which saves a great deal of memory and lets modern chips train far faster.
Sources
- Wortsman, M., Liu, P. J., Xiao, L., Everett, K., Alemi, A. et al. (2023). Small-scale proxies for large-scale Transformer training instabilities. arXiv:2309.14322. arxiv.org/abs/2309.14322
Mentioned in
exploding gradient problemexploding gradient problem, n.: A training failure in which the learning signal grows out of control as it passes back through a network, so the weights jump wildly and training breaks down., gradient clippinggradient clipping, n.: A safety limit on how big a single training update can be. If the computed adjustment is unusually large, it is scaled down so that one bad batch cannot wreck the model., learning rate warmuplearning rate warmup, n.: Starting training with tiny learning steps and gradually increasing them over the first stretch of training, so the model does not lurch wildly before it has found its footing., normalization layernormalization layer, n.: A layer that keeps the numbers passing through a neural network within a steady range by shifting and resizing them, so that deep networks learn faster and are less likely to break down during training., numerical stabilitynumerical stability, n.: Whether a computer calculation stays accurate despite the tiny rounding errors in every step, instead of blowing up into nonsense such as infinities., loss spikeloss spike, n.: A sudden jump in a model's error partway through training. Sometimes the model recovers on its own; sometimes the run falls apart and engineers must rewind to an earlier saved copy., maximal update parametrizationmaximal update parametrization, n.: A recipe for scaling a neural network's starting values and learning speeds with its size so that settings tuned on a small, cheap model work unchanged on a huge one., QK-normQK-norm, n.: A small change to how a language model pays attention — rescaling two sets of internal numbers before comparing them — that prevents a common way large models' training goes haywire., RMSNormRMSNorm, n.: A simpler, faster version of the layer that keeps a network's internal numbers at a steady scale. Most recent large language models use it., silent data corruptionsilent data corruption, n.: When computer hardware quietly produces wrong numbers without crashing or raising any alarm. In huge AI training runs on thousands of chips, such hidden errors can damage a model before anyone notices., z-lossz-loss, n.: A small extra penalty added during training that stops the raw scores a language model produces before choosing a word from growing huge, which helps keep very large training runs from blowing up.