`bridge.recipes.qwen_omni.h100.qwen3_omni`#

Qwen3-Omni thinker training recipes.

Module Contents#

`qwen3_omni_30b_a3b_sft_8gpu_h100_bf16_config`	Return an 8-GPU thinker-only SFT config for Qwen3-Omni 30B-A3B.
`qwen3_omni_30b_a3b_sft_8gpu_h100_bf16_hf_json_config`	Return an 8-GPU SFT config for JSONL loaded through Hugging Face datasets.

`_QWEN3_OMNI_HF_PATH`
`__all__`

bridge.recipes.qwen_omni.h100.qwen3_omni._QWEN3_OMNI_HF_PATH#: ‘Qwen/Qwen3-Omni-30B-A3B-Instruct’

bridge.recipes.qwen_omni.h100.qwen3_omni.qwen3_omni_30b_a3b_sft_8gpu_h100_bf16_config() → megatron.bridge.training.config.ConfigContainer#

Return an 8-GPU thinker-only SFT config for Qwen3-Omni 30B-A3B.

Recommended parallelism: TP=1, PP=1, EP=8.

bridge.recipes.qwen_omni.h100.qwen3_omni.qwen3_omni_30b_a3b_sft_8gpu_h100_bf16_hf_json_config() → megatron.bridge.training.config.ConfigContainer#: Return an 8-GPU SFT config for JSONL loaded through Hugging Face datasets.

bridge.recipes.qwen_omni.h100.qwen3_omni.__all__#: [‘qwen3_omni_30b_a3b_sft_8gpu_h100_bf16_config’, ‘qwen3_omni_30b_a3b_sft_8gpu_h100_bf16_hf_json_conf…