Error message in checkpoint_{loader, saver}_megatron.py is usually a red herring

[Mittagskogel]

Megatron-DeepSpeed/tools/checkpoint_loader_megatron.py

Lines 28 to 39 in a7b7cb7

 try: 
     from megatron.arguments import parse_args, validate_args 
     from megatron.global_vars import set_args, set_global_variables 
     from megatron.checkpointing import load_args_from_checkpoint, load_checkpoint 
     from megatron.model import module 
     from megatron.core import mpu 
     from megatron.core.enums import ModelType 
     from megatron import fused_kernels 
 except ModuleNotFoundError: 
     print("Unable to import Megatron, please specify the path to Megatron using --megatron-path. Exiting.") 
     queue.put("exit") 
     exit(1) 

The problem here is usually not related to finding the megatron directory, but rather outdated megatron or missing subdependencies. Please remove the error message or at least print the underlying error and suggest specifying the megatron directory, instead of assuming that this is the root of the issue.