Skip to content

ModuleNotFoundError: No module named 'megatron.training.tokenizer' #1585

@pspdada

Description

@pspdada

When I use the latest version of megatraon NVIDIA Megatron Core 0.15.3 to run the script, there is an error.

ModuleNotFoundError: No module named 'megatron.training.tokenizer'

The logs:

  File "/slime/slime/backends/megatron_utils/arguments.py", line 4, in <module>
    from megatron.training.tokenizer.tokenizer import _vocab_size_with_padding
ModuleNotFoundError: No module named 'megatron.training.tokenizer'

The env:

pip show megatron-core
Name: megatron-core
Version: 0.15.3
Summary: Megatron Core - a library for efficient and scalable training of transformer based models
Home-page: https://github.com/NVIDIA/Megatron-LM/megatron/core
Author: 
Author-email: NVIDIA <nemo-toolkit@nvidia.com>
License: Apache 2.0
Location: /miniconda3/envs/psp-slime/lib/python3.11/site-packages
Requires: numpy, packaging, torch
Required-by:

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