bridge.recipes.muse_glimmer.h100.muse_glimmer#

Bounded H100 training recipes for Muse Glimmer 30B.

Module Contents#

Functions#

muse_glimmer_30b_pretrain_32gpu_h100_bf16_multimodal_config

Return a 100-step random-init multimodal pretraining config on 32 H100 GPUs.

muse_glimmer_30b_sft_32gpu_h100_bf16_config

Return a 100-step full multimodal SFT config on 32 H100 GPUs.

muse_glimmer_30b_sft_32gpu_h100_bf16_long_context_config

Return a 100-step packed 8K full SFT config with CP2 on 32 H100 GPUs.

muse_glimmer_30b_peft_8gpu_h100_bf16_config

Return a 100-step attention-projection LoRA config on 8 H100 GPUs.

Data#

API#

bridge.recipes.muse_glimmer.h100.muse_glimmer._MODEL_ID#

‘meta-models/Muse-Glimmer-30B’

bridge.recipes.muse_glimmer.h100.muse_glimmer._MODEL_REVISION#

‘f84ecc3a0ea984a4c04542a84269e3d065350a6e’

bridge.recipes.muse_glimmer.h100.muse_glimmer._CORD_V2_REVISION#

‘7f0115a4b758a71d6473b8d085751692da2fef98’

bridge.recipes.muse_glimmer.h100.muse_glimmer.muse_glimmer_30b_pretrain_32gpu_h100_bf16_multimodal_config() megatron.bridge.training.config.ConfigContainer#

Return a 100-step random-init multimodal pretraining config on 32 H100 GPUs.

bridge.recipes.muse_glimmer.h100.muse_glimmer.muse_glimmer_30b_sft_32gpu_h100_bf16_config() megatron.bridge.training.config.ConfigContainer#

Return a 100-step full multimodal SFT config on 32 H100 GPUs.

bridge.recipes.muse_glimmer.h100.muse_glimmer.muse_glimmer_30b_sft_32gpu_h100_bf16_long_context_config() megatron.bridge.training.config.ConfigContainer#

Return a 100-step packed 8K full SFT config with CP2 on 32 H100 GPUs.

bridge.recipes.muse_glimmer.h100.muse_glimmer.muse_glimmer_30b_peft_8gpu_h100_bf16_config(
peft_scheme: str | megatron.bridge.peft.base.PEFT = 'lora',
) megatron.bridge.training.config.ConfigContainer#

Return a 100-step attention-projection LoRA config on 8 H100 GPUs.

bridge.recipes.muse_glimmer.h100.muse_glimmer.__all__#

[‘muse_glimmer_30b_peft_8gpu_h100_bf16_config’, ‘muse_glimmer_30b_pretrain_32gpu_h100_bf16_multimoda…