Skip to content

qwen3 not support vpp #719

@haorannlp

Description

@haorannlp

set MP_VP=2, get


[rank3]: Traceback (most recent call last):
[rank3]:   File "/opt/tiger/PAI-Megatron-Patch-251027/examples/qwen3/pretrain_qwen.py", line 139, in <module>
[rank3]:     pretrain(
[rank3]:   File "/opt/tiger/Megatron-LM-250624/megatron/training/training.py", line 806, in pretrain
[rank3]:     model, optimizer, opt_param_scheduler = setup_model_and_optimizer(
[rank3]:                                             ^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank3]:   File "/opt/tiger/Megatron-LM-250624/megatron/training/training.py", line 1222, in setup_model_and_optimizer
[rank3]:     model = get_model(model_provider_func, model_type)
[rank3]:             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank3]:   File "/opt/tiger/Megatron-LM-250624/megatron/training/training.py", line 1033, in get_model
[rank3]:     model = build_model()
[rank3]:             ^^^^^^^^^^^^^
[rank3]:   File "/opt/tiger/Megatron-LM-250624/megatron/training/training.py", line 999, in build_model
[rank3]:     this_model = model_provider_func(
[rank3]:                  ^^^^^^^^^^^^^^^^^^^^
[rank3]: TypeError: model_provider() got an unexpected keyword argument 'vp_stage'

因为examples/qwen3/pretrain_qwen.py
def model_provider(pre_process=True, post_process=True) -> Union[GPTModel]:

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions