bridge.perf_recipes.deepseek.b200.deepseek_v4#
B200 performance recipes for DeepSeek V4.
Module Contents#
Functions#
Apply the lower-memory B200 topology and communication settings. |
|
DeepSeek V4 Flash pretrain: 128× B200, MXFP8. |
API#
- bridge.perf_recipes.deepseek.b200.deepseek_v4._apply_deepseek_v4_b200_overrides(
- cfg: megatron.bridge.training.config.ConfigContainer,
Apply the lower-memory B200 topology and communication settings.
- bridge.perf_recipes.deepseek.b200.deepseek_v4.deepseek_v4_flash_pretrain_128gpu_b200_fp8mx_config() megatron.bridge.training.config.ConfigContainer#
DeepSeek V4 Flash pretrain: 128× B200, MXFP8.