`bridge.recipes.kimi.h100.kimi_k2`#

Module Contents#

Functions#

`_get_kimi_k2_pipeline_layout`	Get pipeline layout for Kimi-K2 based on PP and VP size.
`_apply_kimi_k2_optimizer`	Apply Kimi K2 optimizer, scheduler, precision, and DDP settings.
`kimi_k2_pretrain_512gpu_h100_bf16_config`	Return a pre-training config for Kimi-K2 (1T).

Data#

API#

bridge.recipes.kimi.h100.kimi_k2._get_kimi_k2_pipeline_layout(pp_size: int, vp_size: int)#: Get pipeline layout for Kimi-K2 based on PP and VP size.

bridge.recipes.kimi.h100.kimi_k2._apply_kimi_k2_optimizer( cfg: megatron.bridge.training.config.ConfigContainer, optimizer_type: str, ) → None#: Apply Kimi K2 optimizer, scheduler, precision, and DDP settings.

bridge.recipes.kimi.h100.kimi_k2.kimi_k2_pretrain_512gpu_h100_bf16_config() → megatron.bridge.training.config.ConfigContainer#

Return a pre-training config for Kimi-K2 (1T).

Recommended parallelism: TP=2, PP=16, EP=32

bridge.recipes.kimi.h100.kimi_k2.__all__#: [‘kimi_k2_pretrain_512gpu_h100_bf16_config’]