bridge.perf_recipes.deepseek.gb300.deepseek_v4#
GB300 performance recipes for DeepSeek V4.
Module Contents#
Functions#
DeepSeek V4 Flash pretrain: 128× GB300, MXFP8. |
|
DeepSeek V4 Pro pretrain: 256× GB300, MXFP8, dev Megatron-Core required. |
API#
- bridge.perf_recipes.deepseek.gb300.deepseek_v4.deepseek_v4_flash_pretrain_128gpu_gb300_fp8mx_config() megatron.bridge.training.config.ConfigContainer#
DeepSeek V4 Flash pretrain: 128× GB300, MXFP8.
- bridge.perf_recipes.deepseek.gb300.deepseek_v4.deepseek_v4_pro_pretrain_256gpu_gb300_fp8mx_config() megatron.bridge.training.config.ConfigContainer#
DeepSeek V4 Pro pretrain: 256× GB300, MXFP8, dev Megatron-Core required.