Hi! Slides link in 6 item is broken. https://github.com/pytorch/workshops/blob/master/NeurIPS_2022/readme.md

Observed OOM when running 737M T5 model on V100s with 32GB RAM each. The OOM was at the start of Epoch 1 after model was init-ed.

File "/home/hybridai/.local/lib/python3.8/site-packages/transformers/models/t5/modeling_t5.py", line 514, in forward
    attn_weights = nn.functional.dropout(
  File "/home/hybridai/.local/lib/python3.8/site-packages/torch/nn/functional.py", line 1252, in dropout
    return _VF.dropout_(input, p, training) if inplace else _VF.dropout(input, p, training)
RuntimeError: CUDA out of memory. Tried to allocate 128.00 MiB (GPU 0; 31.75 GiB total capacity; 30.42 GiB already allocated; 12.75 MiB free; 30.58 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Changing the batch size to 4 didnt fix the OOM, but it reduced the "tried to allocate" part. Further reducing to 2 for batch size as well as validation results in training to start. GPU utilization is high (98%), however the GPU memory is showing 24GB each (per nvidia-smi).

@lessw2020 suggests the following fixes:

You are on V100 (so no BFloat16) and the code there does not have fp16 support added, so it's running in fp32 then, which is much more memory intensive.
I can add in the shardedgradscaler for you tomorrow to add FP16, and also update it to use FSDP activation checkpointing rather than HF (which we are finding is more performant).
For peak usage - I have a mem maximizer class that will show that for you.

in demo:PTMobileWalkthruIOS

"pod insatll" has been run under the right folder, the red of the "Pods" was fixed. But the file : "libPods-PTMobileDemo.a" remain to be red, I have done everything I can but it doesn't work.

I don't know if the file “libPods-PTMobileDemo.a" was forgotten to be uploaded to github.

Hoping to get help from you

TorchDynamo Deep Dive - NeurIPS 2022.pdf
I can't open it, it break down.

The current iOS implementation supports RGB model input - is there a way to use 1 channel input (grayscale)? I am unable to find any resources or examples online. Thank you!