Train Models Faster with JAX and MaxText Using NVFP4 on NVIDIA Blackwell

来源：NVIDIA 技术博客·2026-06-09 02:18发布时间待核实

Pre-training frontier LLMs comes down to throughput. When training spans trillions of tokens across thousands of accelerators, every percentage point of step... Pre-training frontier LLMs comes down to throughput. When training spans trillions of tokens across thousands of accelerators, every percentage point of step time can add up to days of training and substantial compute costs.