bridge.perf_recipes.deepseek.b300.deepseek_v4#
B300 performance recipes for DeepSeek V4.
Module Contents#
Functions#
Apply a B300 topology scaled down from the larger DeepSeek V3 model. |
|
DeepSeek V4 Flash pretrain: 128× B300, MXFP8. |
API#
- bridge.perf_recipes.deepseek.b300.deepseek_v4._apply_deepseek_v4_b300_overrides(
- cfg: megatron.bridge.training.config.ConfigContainer,
Apply a B300 topology scaled down from the larger DeepSeek V3 model.
- bridge.perf_recipes.deepseek.b300.deepseek_v4.deepseek_v4_flash_pretrain_128gpu_b300_fp8mx_config() megatron.bridge.training.config.ConfigContainer#
DeepSeek V4 Flash pretrain: 128× B300, MXFP8.