Augustinas Malinauskas comments

Results 49 comments of


                                            Augustinas Malinauskas

Google Batch: Input files provided as output to the next process are null (mounted symlink not translated into real file on output)

Problem still persists, any solutions?

File exists: '/000000_epoch_shape' when using the ddp strategy from pytorch lightning

Also getting something similar ``` [rank1]: Traceback (most recent call last): [rank1]: File "/home/august/cfdx/ai/dna_fm/lightning_gym.py", line 59, in [rank1]: main() [rank1]: File "/home/august/cfdx/ai/dna_fm/lightning_gym.py", line 55, in main [rank1]: run_experiment(config) [rank1]: File...

Refactor spanner to avoid creating large array

Hi @XiaohanZhangCMU incredible job! I just tested and can confirm that this solves a problem with large number of shards. However, there may be another bottleneck with `StreamingDataLoader` on datasets...

Refactor spanner to avoid creating large array

Sharing profiler results on time to first batch

Refactor spanner to avoid creating large array

I see... So MosaicML currently is not suitable for datasets with large number of shards 😢. We're using litdata, but been thinking to migrate to MosaicML as has great features....

Refactor spanner to avoid creating large array

Thank you for the message. Will try on H100 - 8GPUs and report back the findings.

Refactor spanner to avoid creating large array

Hi guys, for me to load the dataset very large dataset (2.4B rows) even with the spanner fix takes 2min 28sec. When it comes to loading first batch ```python for...

Support for streaming tool call for Ollama

Implemented fix here https://github.com/openai/openai-go/pull/418

The config isn't consistent between chunks

Hi @bhimrazy @tchaton thank you for the reply. Would you say that as long I am saving data as torch.tensor or numpy array there should be no problems loading the...