This PR adds functionality to output group information for large model execution, helping to track and manage task distribution during runtime.
New Functionality
parallelism_to_groups.json
Defines how tasks are grouped across various parallelism strategies (data, tensor, pipeline, etc.) for large model execution.
rank_to_parallelism_to_group_id.json
Maps device ranks to group IDs for different parallelism strategies.
rank_to_host_and_device.json
Provides mapping from device ranks to specific hardware (host IP, device ID, and GPU name).
Note
This PR enables the output of parallel group information for both decoder and encoder modes.
Usage Instructions
To enable the output of parallel group information during model training, add the following configuration to your training file:
analyze_save_dir: Specifies the directory where the group information will be saved. Replace group_info_output_path with your desired output path for storing the parallelism group details.
Description
This PR adds functionality to output group information for large model execution, helping to track and manage task distribution during runtime.
New Functionality
parallelism_to_groups.json
Defines how tasks are grouped across various parallelism strategies (data, tensor, pipeline, etc.) for large model execution.
rank_to_parallelism_to_group_id.json
Maps device ranks to group IDs for different parallelism strategies.
rank_to_host_and_device.json
Provides mapping from device ranks to specific hardware (host IP, device ID, and GPU name).
Note
This PR enables the output of parallel group information for both decoder and encoder modes.
Usage Instructions
To enable the output of parallel group information during model training, add the following configuration to your training file:
analyze_save_dir
: Specifies the directory where the group information will be saved. Replacegroup_info_output_path
with your desired output path for storing the parallelism group details.