Frontier Lab Notes

Training Optimization and Stability Deep Dive

Training frontier-scale models is a controlled fight against instability, inefficiency, and bad extrapolation from small experiments.

Primary sources:

The Training Objective

For pretraining, the model minimizes average next-token cross-entropy:

L = - mean_t log P(x_t | x_<t)

At scale, the goal is not just low training loss. It is stable validation loss improvement under a fixed compute/data/hardware budget.

Optimizer

Most transformer LLMs use Adam or AdamW variants.

Adam tracks:

AdamW decouples weight decay from the adaptive update, making regularization behavior cleaner.

Learning Rate Schedule

Common pattern:

Why warmup matters:

Batch Size And Gradient Accumulation

Effective batch size:

microbatch_size * gradient_accumulation_steps * data_parallel_workers

Tradeoffs:

Mixed Precision

Training often uses BF16/FP16 with FP32 master states or carefully managed accumulations.

Why:

Risks:

BF16 is often preferred when hardware supports it because it has wider exponent range.

Gradient Clipping

Gradient clipping caps update magnitude:

if norm(g) > threshold:
    g = g * threshold / norm(g)

Why:

Initialization And Parameterization

At scale, hyperparameters that work for one model size may fail for another.

muP asks for parameterization and hyperparameter transfer rules that make small-model experiments more predictive of large-model behavior.

Research importance:

Normalization And Residual Stability

Pre-norm, RMSNorm, residual scaling, and initialization all affect deep training stability.

Failure signs:

Data Pipeline Stability

Training instability is not always math.

Watch for:

Distributed Training

Frontier models do not fit on one GPU.

Parallelism types:

Systems notes:

ZeRO-Style Memory Partitioning

ZeRO reduces memory by partitioning:

This makes larger models trainable across devices.

Training Run Monitoring

Track:

Research Questions

Related