#bf16

The latest AI and developer news about #bf16, with the original source and a short summary.

bf16 Fine-Tuning Freezes 95% of Weights, Masking Optimizer Instability

Using bf16 weights with AdamW without fp32 master copies causes large-magnitude weights to freeze, effectively acting as a sparse fine-tuner that masks underlying training instability.

·2026.11.02 09:00

How to compress LLM weights by 22% without quality loss

Unweight reduces LLM weights by 15~22% on H100 while keeping outputs bit-exact.

·2·2026.04.17 22:00

Efficient Rollout Scaling for Diffusion RL

NVIDIA Sol-RL reduces cost and accelerates convergence by separating FP4 rollouts from BF16 updates.

·2026.04.10 09:00

Lightweight Ad Engagement Prediction with Two-Tower Models Served on GPU

GPU serving based on MMOE-DCN improved both accuracy and latency for ad engagement prediction.

·2026.02.14 08:00