[Bug Report] Wrong function call in model_parallel_v2 #4791
Comments
florianbodr
referenced
this issue
Nov 29, 2024
* Update example notebooks and related scripts for latest PT-2.2-TSM-2.2 release. Add FP8 training support on P5. * Add example notebook for accelerating Llama-v2 training with FP8 on P5. * Fix typo in version check * Update configurations. Revert jupyter notebook python version in metadata. Set activation_offloading=False for FP8 notebook. Explicitly enable use_smp_implementation in all SMP v2 notebooks. * Update FP8 notebook docs. * Set zipped_data=0 for use_fsx=False FP8 notebook. * Update compute_tflops() script. * Update minimum sagemaker pysdk version to `2.212`.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hi all,
There is a bug in the following py file: training/distributed_training/pytorch/model_parallel_v2/shared-scripts/logging_utils.py
Line 151 states the following:
avg_tflops = compute_tflops(avg_throughput, num_params, world_size, batch_seqlen)But the function definition in: training/distributed_training/pytorch/model_parallel_v2/shared-scripts/train_utils.py
at line 36 is the following:
def compute_tflops(args, global_batch_size, step_time, world_size):The arguments of the function call should be adapted, at least args shall be passed as the first one (or a new function as to be defined).
The text was updated successfully, but these errors were encountered: