bridge.recipes.qwen.gb200.qwen3_moe

GB200 recipes for Qwen3 MoE models.

Module Contents

Functions

qwen3_30b_a3b_pretrain_8gpu_gb200_fp8mx_config

Return a checkpointable Qwen3-30B-A3B MXFP8 pretraining config for eight GB200 GPUs.

Data

__all__

API

bridge.recipes.qwen.gb200.qwen3_moe.qwen3_30b_a3b_pretrain_8gpu_gb200_fp8mx_config() → megatron.bridge.training.config.ConfigContainer

Return a checkpointable Qwen3-30B-A3B MXFP8 pretraining config for eight GB200 GPUs.

The Blackwell topology, compute precision, dispatcher, overlap, kernels, batch sizes, and environment follow the corresponding flat performance recipe. The recipe keeps natural routing and safety checks enabled and runs 100 optimizer steps with MXFP8 parameter all-gather.

bridge.recipes.qwen.gb200.qwen3_moe.__all__

[‘qwen3_30b_a3b_pretrain_8gpu_gb200_fp8mx_config’]