`bridge.recipes.deepseek.gb300.deepseek_v4`#

GB300 recipes for DeepSeek V4 Pro.

Module Contents#

Functions#

`deepseek_v4_pro_pretrain_32gpu_gb300_bf16_config`	Return the DeepSeek-V4-Pro GB300 pre-training base config.
`deepseek_v4_pro_pretrain_32gpu_gb300_fp8mx_config`	Return the DeepSeek-V4-Pro Adam + MXFP8 pre-training config.

Data#

DEEPSEEK_V4_PRO_HF_PATH

API#

bridge.recipes.deepseek.gb300.deepseek_v4.DEEPSEEK_V4_PRO_HF_PATH#: ‘deepseek-ai/DeepSeek-V4-Pro’

bridge.recipes.deepseek.gb300.deepseek_v4.deepseek_v4_pro_pretrain_32gpu_gb300_bf16_config() → megatron.bridge.training.config.ConfigContainer#

Return the DeepSeek-V4-Pro GB300 pre-training base config.

DeepSeek-V4 still requires a compatible Megatron-Core development commit; the Megatron-Core commit pinned by Megatron Bridge main is not a supported runtime for this recipe.

bridge.recipes.deepseek.gb300.deepseek_v4.deepseek_v4_pro_pretrain_32gpu_gb300_fp8mx_config() → megatron.bridge.training.config.ConfigContainer#: Return the DeepSeek-V4-Pro Adam + MXFP8 pre-training config.