← 🖥️ ML Infrastructure & ServingNEXT IN ML INFRASTRUCTURE & SERVINGMulti-LoRA Serving→
Core
Mixed-Precision Training
Mixed-precision training does most computation in 16-bit (FP16 or BF16) instead of 32-bit, roughly halving memory and speeding up training on modern GPUs, while keeping a few numerically-sensitive parts in FP32 for stability. BF16 is preferred over FP16 because it keeps FP32's exponent range, avoiding the overflow/underflow that FP16 needs loss scaling to handle. Applied-AI interviews probe it because it is standard practice for training at scale and a clean example of the precision-vs-stability trade-off.
a free account unlocks the core curriculum tier · no card
RELATED CONCEPTS
PRACTICE THIS IN REAL QUESTIONS
ML Infrastructure & GPUsExplain mixed-precision training: FP16 vs BF16, loss scaling, and where the numerics break.→ML Infrastructure & GPUsWhat is FP8 (and low-precision training/inference), and what are the tradeoffs?→LLM & GenAI FundamentalsWhat is FP8, and how does it differ from INT8 for LLM training and inference?→ML Infrastructure & GPUsYour GPUs sit at 40% utilization during training. How do you find and fix the bottleneck?→ML Infrastructure & GPUsYour INT4-quantized model lost too much accuracy. How do you recover it?→ML Infrastructure & GPUsYour large-model pretraining hits sudden loss spikes that don't recover. How do you stabilize it?→
COMPANIES THAT ASSUME THIS