`bridge.recipes.qwen.qwen3_moe`#

Module Contents#

Classes#

`Qwen3MoeCommonKwargs`	Typed options accepted by Qwen3 MoE recipe helpers.
`Qwen3MoeFinetuneKwargs`	Typed options accepted by Qwen3 MoE finetuning recipe helper functions.

Functions#

`qwen3_30b_a3b_pretrain_config`	Return a pre-training config for Qwen3-30B-A3B MoE.
`qwen3_235b_a22b_pretrain_config`	Return a pre-training config for Qwen3-235B-A22B MoE.
`_qwen3_moe_common`	Create a pre-training configuration for Qwen3 MoE models using a given HuggingFace path.
`qwen3_30b_a3b_finetune_config`	Return a finetuning config for Qwen3-30B-A3B MoE.
`qwen3_235b_a22b_finetune_config`	Return a finetuning config for Qwen3-235B-A22B MoE.
`_qwen3_moe_finetune_common`	Create a finetuning configuration for Qwen3 MoE models using a given HuggingFace path.

API#

class bridge.recipes.qwen.qwen3_moe.Qwen3MoeCommonKwargs#

Bases: typing_extensions.TypedDict

Typed options accepted by Qwen3 MoE recipe helpers.

Initialization

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

hf_path: str#: None

dir: Optional[str]#: None

name: str#: None

data_paths: Optional[List[str]]#: None

data_args_path: Optional[str]#: None

train_data_path: Optional[List[str]]#: None

valid_data_path: Optional[List[str]]#: None

test_data_path: Optional[List[str]]#: None

per_split_data_args_path: Optional[str]#: None

mock: bool#: None

tensor_model_parallel_size: int#: None

pipeline_model_parallel_size: int#: None

pipeline_dtype: Optional[torch.dtype]#: None

virtual_pipeline_model_parallel_size: Optional[int]#: None

context_parallel_size: int#: None

expert_model_parallel_size: Optional[int]#: None

expert_tensor_parallel_size: int#: None

sequence_parallel: bool#: None

use_megatron_fsdp: bool#: None

enable_recompute: bool#: None

account_for_embedding_in_pipeline_split: bool#: None

account_for_loss_in_pipeline_split: bool#: None

train_iters: int#: None

global_batch_size: int#: None

micro_batch_size: int#: None

seq_length: int#: None

lr: float#: None

min_lr: float#: None

lr_warmup_iters: int#: None

lr_decay_iters: Optional[int]#: None

eval_interval: int#: None

save_interval: int#: None

use_null_tokenizer: 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: str | None#: None

class bridge.recipes.qwen.qwen3_moe.Qwen3MoeFinetuneKwargs#

Bases: typing_extensions.TypedDict

Typed options accepted by Qwen3 MoE finetuning recipe helper functions.

This is separate from Qwen3MoeCommonKwargs to avoid confusion - finetuning uses SQuAD dataset by default, not the data path fields.

Initialization

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

dir: Optional[str]#: None

name: str#: None

pretrained_checkpoint: Optional[str]#: None

peft: Union[str, megatron.bridge.peft.base.PEFT, None]#: None

packed_sequence: bool#: None

train_iters: int#: None

global_batch_size: Optional[int]#: None

micro_batch_size: int#: None

seq_length: Optional[int]#: None

eval_interval: int#: None

save_interval: int#: None

finetune_lr: Optional[float]#: None

min_lr: float#: None

lr_warmup_iters: int#: None

lr_decay_iters: Optional[int]#: None

wandb_project: Optional[str]#: None

wandb_entity: Optional[str]#: None

wandb_exp_name: Optional[str]#: None

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

bridge.recipes.qwen.qwen3_moe.qwen3_30b_a3b_pretrain_config(

**user_kwargs: typing_extensions.Unpack[bridge.recipes.qwen.qwen3_moe.Qwen3MoeCommonKwargs],

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

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

See _qwen3_moe_common for the full list of parameters.

bridge.recipes.qwen.qwen3_moe.qwen3_235b_a22b_pretrain_config(

**user_kwargs: typing_extensions.Unpack[bridge.recipes.qwen.qwen3_moe.Qwen3MoeCommonKwargs],

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

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

See _qwen3_moe_common for the full list of parameters.

bridge.recipes.qwen.qwen3_moe._qwen3_moe_common( hf_path: str, dir: Optional[str] = None, name: str = 'default', data_paths: Optional[List[str]] = None, data_args_path: Optional[str] = None, train_data_path: Optional[List[str]] = None, valid_data_path: Optional[List[str]] = None, test_data_path: Optional[List[str]] = None, per_split_data_args_path: Optional[str] = None, mock: bool = False, tensor_model_parallel_size: int = 4, pipeline_model_parallel_size: int = 2, pipeline_dtype: Optional[torch.dtype] = torch.bfloat16, virtual_pipeline_model_parallel_size: Optional[int] = None, context_parallel_size: int = 1, expert_model_parallel_size: Optional[int] = 4, expert_tensor_parallel_size: int = 1, sequence_parallel: bool = True, use_megatron_fsdp: bool = False, enable_recompute: bool = False, account_for_embedding_in_pipeline_split: bool = False, account_for_loss_in_pipeline_split: bool = False, train_iters: int = 300000, global_batch_size: int = 32, micro_batch_size: int = 2, seq_length: int = 4096, lr: float = 0.0003, min_lr: float = 3e-05, lr_warmup_iters: int = 500, lr_decay_iters: Optional[int] = None, eval_interval: int = 500, save_interval: int = 500, use_null_tokenizer: bool = False, 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, ) → megatron.bridge.training.config.ConfigContainer#

Create a pre-training configuration for Qwen3 MoE models using a given HuggingFace path.

Parameters:

hf_path (str) – HuggingFace model path (e.g., “Qwen/Qwen3-30B-A3B”, “Qwen/Qwen3-235B-A22B”).
dir (Optional[str]) – Base directory for saving logs and checkpoints.
name (str) – Name of the pre-training run.
data_paths (Optional[List[str]]) – List of paths to dataset files. If None, mock data will be used.
data_args_path (Optional[str]) – Path to file containing data arguments.
train_data_path (Optional[List[str]]) – List of training data paths.
valid_data_path (Optional[List[str]]) – List of validation data paths.
test_data_path (Optional[List[str]]) – List of test data paths.
per_split_data_args_path (Optional[str]) – Path to JSON file with per-split data configuration.
mock (bool) – Whether to use mock data. If True, ignores data_paths.
tensor_model_parallel_size (int) – Degree of tensor model parallelism.
pipeline_model_parallel_size (int) – Degree of pipeline model parallelism.
pipeline_dtype (Optional[torch.dtype]) – Data type for pipeline parallelism.
virtual_pipeline_model_parallel_size (Optional[int]) – Size of virtual pipeline parallelism.
context_parallel_size (int) – Degree of context parallelism to be passed to model_config.
expert_model_parallel_size (Optional[int]) – Degree of expert parallelism for MoE.
expert_tensor_parallel_size (int) – Expert tensor parallelism for MoE.
sequence_parallel (bool) – Whether to use sequence parallelism.
use_megatron_fsdp (bool) – Whether to use Megatron FSDP.
enable_recompute (bool) – Whether to enable recompute for memory optimization.
account_for_embedding_in_pipeline_split (bool) – Whether to account for embedding in pipeline split.
account_for_loss_in_pipeline_split (bool) – Whether to account for loss in pipeline split.
train_iters (int) – Total number of training iterations.
global_batch_size (int) – Global batch size for training.
micro_batch_size (int) – Micro batch size for training.
seq_length (int) – Sequence length for training data.
lr (float) – Learning rate.
min_lr (float) – Minimum learning rate for cosine decay.
lr_warmup_iters (int) – Number of warmup iterations for the learning rate.
lr_decay_iters (Optional[int]) – Number of iterations over which to decay the LR.
precision_config (Optional[Union[MixedPrecisionConfig, str]]) – Precision configuration for the model.
comm_overlap_config (Optional[CommOverlapConfig]) – Communication overlap configuration.
moe_flex_dispatcher_backend (str | None) – Token dispatcher type [deepep, hybridep].

Returns:

Configuration for pre-training.

Return type:

ConfigContainer

bridge.recipes.qwen.qwen3_moe.qwen3_30b_a3b_finetune_config(

**user_kwargs: typing_extensions.Unpack[bridge.recipes.qwen.qwen3_moe.Qwen3MoeFinetuneKwargs],

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

Return a finetuning config for Qwen3-30B-A3B MoE.

Default configuration: 1 node, 8 GPUs, LoRA

LoRA (default): TP=4, PP=1, EP=4, LR=1e-4, dim=8, alpha=16, target_modules=[‘linear_qkv’, ‘linear_proj’]
DoRA: TP=4, PP=1, EP=4, LR=1e-4, dim=8, alpha=16, target_modules=[‘linear_qkv’, ‘linear_proj’]
Full SFT (peft=None): TP=4, PP=2, EP=4, LR=5e-6, SP=True

Matches NeMo2 recipe at nemo/collections/llm/recipes/qwen3_30b_a3b.py

bridge.recipes.qwen.qwen3_moe.qwen3_235b_a22b_finetune_config(

**user_kwargs: typing_extensions.Unpack[bridge.recipes.qwen.qwen3_moe.Qwen3MoeFinetuneKwargs],

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

Return a finetuning config for Qwen3-235B-A22B MoE.

Default configuration: 8 nodes (LoRA) or 16 nodes (Full SFT), 8 GPUs per node

LoRA (default): TP=4, PP=4, EP=4, LR=1e-4, dim=8, alpha=16, target_modules=[‘linear_qkv’, ‘linear_proj’] Total: 64 GPUs (8 nodes)
DoRA: TP=4, PP=4, EP=4, LR=1e-4, dim=8, alpha=16, target_modules=[‘linear_qkv’, ‘linear_proj’] Total: 64 GPUs (8 nodes)
Full SFT (peft=None): TP=4, PP=16, EP=4, LR=5e-6, SP=True Total: 64 GPUs (8 nodes)

Matches NeMo2 recipe at nemo/collections/llm/recipes/qwen3_235b_a22b.py

Note: Uses account_for_embedding_in_pipeline_split and account_for_loss_in_pipeline_split for proper layer distribution in pipeline parallelism.

bridge.recipes.qwen.qwen3_moe._qwen3_moe_finetune_common( hf_path: str, dir: Optional[str] = None, name: str = 'default', pretrained_checkpoint: Optional[str] = None, packed_sequence: bool = False, train_iters: int = 100, global_batch_size: Optional[int] = None, micro_batch_size: int = 1, seq_length: Optional[int] = None, eval_interval: int = 50, save_interval: int = 100, finetune_lr: Optional[float] = None, min_lr: float = 0.0, lr_warmup_iters: int = 10, lr_decay_iters: Optional[int] = None, wandb_project: Optional[str] = None, wandb_entity: Optional[str] = None, wandb_exp_name: Optional[str] = None, precision_config: Optional[Union[megatron.bridge.training.mixed_precision.MixedPrecisionConfig, str]] = None, moe_flex_dispatcher_backend: Optional[str] = None, ) → megatron.bridge.training.config.ConfigContainer#

Create a finetuning configuration for Qwen3 MoE models using a given HuggingFace path.