bridge.recipes.minimax.h100.minimax_m3#

Module Contents#

Functions#

minimax_m3_pretrain_256gpu_h100_bf16_config

Return a pre-training config for MiniMax-M3 (428B total, ~23B active).

minimax_m3_sft_128gpu_h100_bf16_config

MiniMax-M3 full SFT on packed (THD) sequences with Adam/bf16.

Data#

MINIMAX_M3_HF_PATH

__all__

API#

bridge.recipes.minimax.h100.minimax_m3.MINIMAX_M3_HF_PATH#

‘MiniMaxAI/MiniMax-M3’

bridge.recipes.minimax.h100.minimax_m3.minimax_m3_pretrain_256gpu_h100_bf16_config() → megatron.bridge.training.config.ConfigContainer#

Return a pre-training config for MiniMax-M3 (428B total, ~23B active).

MiniMax-M3 has 60 decoder layers (first 3 dense), 128 routed experts with top-4 routing plus one shared expert, and 4 KV heads. Recommended parallelism: TP=2, PP=4 (15 layers per stage), EP=32 (128 GPUs minimum; 256 GPUs gives DP=2).

bridge.recipes.minimax.h100.minimax_m3.minimax_m3_sft_128gpu_h100_bf16_config() → megatron.bridge.training.config.ConfigContainer#

MiniMax-M3 full SFT on packed (THD) sequences with Adam/bf16.

Same TP=2 / PP=4 / EP=32 layout as the pretrain config (128 GPUs minimum).

bridge.recipes.minimax.h100.minimax_m3.__all__#

[‘minimax_m3_pretrain_256gpu_h100_bf16_config’, ‘minimax_m3_sft_128gpu_h100_bf16_config’]