`bridge.recipes.qwen_vl.qwen3_vl`#

Qwen3-VL recipes with parameterless API.

This module provides pretrain, SFT, and PEFT configurations for Qwen3-VL models (8B, 30B-A3B, 235B-A22B).

Module Contents#

Classes#

Qwen3VLCommonKwargs

Typed options accepted by Qwen3-VL pretrain recipe helper.

Functions#

`_qwen3_vl_common`	Create a pre-training configuration for Qwen3-VL models.
`qwen3_vl_8b_pretrain_mock_config`	Return a pre-training config for Qwen3-VL 8B Instruct.
`qwen3_vl_30b_a3b_pretrain_mock_config`	Return a pre-training config for Qwen3-VL 30B-A3B (MoE).
`qwen3_vl_235b_a22b_pretrain_mock_config`	Return a pre-training config for Qwen3-VL 235B-A22B (MoE).
`_make_energon_dataset`	Create an EnergonProvider dataset config for Qwen3-VL recipes.
`qwen3_vl_8b_sft_config`	Return a full SFT config for Qwen3-VL 8B (dense model).
`qwen3_vl_30b_a3b_sft_config`	Return a full SFT config for Qwen3-VL 30B-A3B (MoE model).
`qwen3_vl_235b_a22b_sft_config`	Return a full SFT config for Qwen3-VL 235B-A22B (MoE model).
`qwen3_vl_8b_peft_config`	Return a PEFT config for Qwen3-VL 8B (dense model).
`qwen3_vl_30b_a3b_peft_config`	Return a PEFT config for Qwen3-VL 30B-A3B (MoE model).
`qwen3_vl_235b_a22b_peft_config`	Return a PEFT config for Qwen3-VL 235B-A22B (MoE model).
`qwen3_vl_8b_peft_energon_config`	Return a PEFT (LoRA/DoRA) config for Qwen3-VL 8B with Energon dataset.

API#

class bridge.recipes.qwen_vl.qwen3_vl.Qwen3VLCommonKwargs#

Bases: typing_extensions.TypedDict

Typed options accepted by Qwen3-VL pretrain recipe helper.

Initialization

Initialize self. See help(type(self)) for accurate signature.

hf_path: str#: None

tensor_model_parallel_size: int#: None

pipeline_model_parallel_size: int#: None

expert_model_parallel_size: int#: None

context_parallel_size: int#: None

sequence_parallel: bool#: None

seq_length: int#: None

train_iters: int#: None

global_batch_size: int#: None

micro_batch_size: int#: None

lr: float#: None

min_lr: float#: None

lr_warmup_iters: int#: None

lr_decay_iters: Optional[int]#: None

freeze_language_model: bool#: None

freeze_vision_model: bool#: None

freeze_vision_projection: bool#: None

precision_config: Optional[Union[megatron.bridge.training.mixed_precision.MixedPrecisionConfig, str]]#: None

comm_overlap_config: Optional[megatron.bridge.training.comm_overlap.CommOverlapConfig]#: None

moe_flex_dispatcher_backend: Optional[str]#: None

mock: bool#: None

bridge.recipes.qwen_vl.qwen3_vl._qwen3_vl_common( hf_path: str = 'Qwen/Qwen3-VL-8B-Instruct', *, tensor_model_parallel_size: int = 4, pipeline_model_parallel_size: int = 1, expert_model_parallel_size: int = 1, context_parallel_size: int = 1, sequence_parallel: bool = False, seq_length: int = 4096, train_iters: int = 300000, global_batch_size: int = 32, micro_batch_size: int = 2, lr: float = 0.0003, min_lr: float = 3e-05, lr_warmup_iters: int = 500, lr_decay_iters: Optional[int] = None, freeze_language_model: bool = True, freeze_vision_model: bool = True, freeze_vision_projection: bool = False, precision_config: Optional[Union[megatron.bridge.training.mixed_precision.MixedPrecisionConfig, str]] = 'bf16_mixed', comm_overlap_config: Optional[megatron.bridge.training.comm_overlap.CommOverlapConfig] = None, moe_flex_dispatcher_backend: Optional[str] = None, mock: bool = True, ) → megatron.bridge.training.config.ConfigContainer#

Create a pre-training configuration for Qwen3-VL models.

Uses MockVLMConversationProvider by default (mock=True).

bridge.recipes.qwen_vl.qwen3_vl.qwen3_vl_8b_pretrain_mock_config(

**user_kwargs: typing_extensions.Unpack[bridge.recipes.qwen_vl.qwen3_vl.Qwen3VLCommonKwargs],

) → megatron.bridge.training.config.ConfigContainer#

Return a pre-training config for Qwen3-VL 8B Instruct.

See _qwen3_vl_common for the full list of parameters.

bridge.recipes.qwen_vl.qwen3_vl.qwen3_vl_30b_a3b_pretrain_mock_config(

**user_kwargs: typing_extensions.Unpack[bridge.recipes.qwen_vl.qwen3_vl.Qwen3VLCommonKwargs],

) → megatron.bridge.training.config.ConfigContainer#

Return a pre-training config for Qwen3-VL 30B-A3B (MoE).

See _qwen3_vl_common for the full list of parameters.

bridge.recipes.qwen_vl.qwen3_vl.qwen3_vl_235b_a22b_pretrain_mock_config(

**user_kwargs: typing_extensions.Unpack[bridge.recipes.qwen_vl.qwen3_vl.Qwen3VLCommonKwargs],

) → megatron.bridge.training.config.ConfigContainer#

Return a pre-training config for Qwen3-VL 235B-A22B (MoE).

See _qwen3_vl_common for the full list of parameters.

bridge.recipes.qwen_vl.qwen3_vl._make_energon_dataset( hf_path: str, seq_length: int, micro_batch_size: int, global_batch_size: int, ) → megatron.bridge.data.energon.energon_provider.EnergonProvider#: Create an EnergonProvider dataset config for Qwen3-VL recipes.

bridge.recipes.qwen_vl.qwen3_vl.qwen3_vl_8b_sft_config() → megatron.bridge.training.config.ConfigContainer#

Return a full SFT config for Qwen3-VL 8B (dense model).

Default configuration: 1 node, 8 GPUs

TP=2, PP=1
LR=5e-6 (full SFT)
Sequence length: 4096

bridge.recipes.qwen_vl.qwen3_vl.qwen3_vl_30b_a3b_sft_config() → megatron.bridge.training.config.ConfigContainer#

Return a full SFT config for Qwen3-VL 30B-A3B (MoE model).

Default configuration: 4 nodes, 32 GPUs

TP=1, PP=1, EP=8
LR=5e-6 (full SFT)
Sequence length: 4096

bridge.recipes.qwen_vl.qwen3_vl.qwen3_vl_235b_a22b_sft_config() → megatron.bridge.training.config.ConfigContainer#

Return a full SFT config for Qwen3-VL 235B-A22B (MoE model).

Default configuration: 64 nodes, 512 GPUs

TP=4, PP=1, EP=32
LR=5e-6 (full SFT)
Sequence length: 4096

bridge.recipes.qwen_vl.qwen3_vl.qwen3_vl_8b_peft_config( peft_scheme: str | megatron.bridge.peft.base.PEFT = 'lora', ) → megatron.bridge.training.config.ConfigContainer#

Return a PEFT config for Qwen3-VL 8B (dense model).

Default configuration: 1 node, 8 GPUs

TP=1, PP=1
LR=1e-4 (PEFT)
Sequence length: 4096

Parameters:: peft_scheme – PEFT scheme - “lora”, “dora”, or a custom PEFT instance.

bridge.recipes.qwen_vl.qwen3_vl.qwen3_vl_30b_a3b_peft_config( peft_scheme: str | megatron.bridge.peft.base.PEFT = 'lora', ) → megatron.bridge.training.config.ConfigContainer#

Return a PEFT config for Qwen3-VL 30B-A3B (MoE model).

Default configuration: 1 node, 8 GPUs

TP=1, PP=1, EP=4
LR=1e-4 (PEFT)
Sequence length: 4096

Parameters:: peft_scheme – PEFT scheme - “lora”, “dora”, or a custom PEFT instance.

bridge.recipes.qwen_vl.qwen3_vl.qwen3_vl_235b_a22b_peft_config( peft_scheme: str | megatron.bridge.peft.base.PEFT = 'lora', ) → megatron.bridge.training.config.ConfigContainer#

Return a PEFT config for Qwen3-VL 235B-A22B (MoE model).

Default configuration: 8 nodes, 64 GPUs

TP=1, PP=1, EP=16
LR=1e-4 (PEFT)
Sequence length: 4096

Parameters:: peft_scheme – PEFT scheme - “lora”, “dora”, or a custom PEFT instance.

bridge.recipes.qwen_vl.qwen3_vl.qwen3_vl_8b_peft_energon_config( peft_scheme: str | megatron.bridge.peft.base.PEFT = 'lora', ) → megatron.bridge.training.config.ConfigContainer#

Return a PEFT (LoRA/DoRA) config for Qwen3-VL 8B with Energon dataset.

Same as qwen3_vl_8b_peft_config but uses EnergonProvider instead of HF dataset. Set the dataset path via CLI override: dataset.path=/path/to/energon/dataset

bridge.recipes.qwen_vl.qwen3_vl#

Module Contents#

Classes#

Functions#

API#

`bridge.recipes.qwen_vl.qwen3_vl`#