Model hangs on eval

[GarrettMerz]

Hi! I'm running an enc-dec transformer with ROPE in the first self-attention layer of the encoder and decoder. I'm noticing that in the eval stage of my model, it hangs until my job times out after about 7 epochs; when running without this package, i.e. standard learnable positional encoding or shaw-style relative postion encoding, I do not notice this behavior. Are there any obvious places in this package that might lead to a memory leak or similar (i.e. does the register_buffer play nice with with_no_grad()?)

[GarrettMerz]

I'm wondering if this issue may be related: pytorch/pytorch#20275

[lucidrains]

hey Garrett at Madison! beautiful city, still have fond memories of it (worked at Epic Systems for a year right out of college)

yup, i think i may have an idea of what's wrong, will throw in a fix in an hour

[lucidrains]

@GarrettMerz basically i'm incorrectly caching by the sequence length, but it should cache the longest sequence length and slice out any subsequent calls with shorter ones

[lucidrains]

@GarrettMerz want to give 0.5.0 a try and see if it still hangs?

[GarrettMerz]

Updating this with results: this largely seems to fix things. I still see hanging behavior in cases where max output length is large and the model does not produce an EOS token before hitting the max length- i.e., if it gets stuck outputting something nonsensical like "++++++++++...", which may happen in early epochs: the length of this bad output is then cached, which causes ROPE to slow down a lot- but capping the max output length at a reasonable size generally seems to mitigate this, which is a good enough fix for now.

[lucidrains]

@GarrettMerz sounds good, as long as it does not hang anymore

best with your research and life out in the midwest

[GarrettMerz]

May need to reopen this, it seems that things are still hanging! I'm going to try to investigate more to figure out when specifically it might be happening- I'm going to use the relative position encoding in the encoder only (not the decoder) and see if that helps at all.

[lucidrains]

hmm, yea, i'll wait for more info from your end

you are the only one reporting this

[lucidrains]

@GarrettMerz could you try turning off cache altogether? https://github.com/lucidrains/rotary-embedding-torch/blob/main/rotary_embedding_torch/rotary_embedding_torch.py#L82 just to confirm that it is indeed caused by the freqs caching and not something on your end?

[jozhang97]

Hi, I was also encountering this issue with the latest commit (v0.5.3). On my end, I can confirm that this is caused by caching. Setting cache_if_possible=False worked for me.

[GarrettMerz]

I still seem to be encountering this issue in some cases- I'm investigating more to confirm it's not an implementation problem.

[lunixbochs]

I'm seeing consistent hangs with 8-gpu accelerate based DDP.
Hangs on NCCL comms after the first step if I'm using 3+ gpus.
My sequence lengths are both different per step and per accelerator.
No issue if cache is disabled, or with 2 gpus, or if I use a fixed sequence length.

[lucidrains]

@lunixbochs I see! thank you for this info

I'll try a way of standardizing the cache to same tensor shape across devices and ping you to give it a try when it is done

[lunixbochs]

the rope cache in fairseq seems to work fine: https://github.com/facebookresearch/fairseq/blob/920a548ca770fb1a951f7f4289b4d3a0c1bc226f/fairseq/modules/rotary_positional_embedding.py#L28

[lucidrains]

@lunixbochs sounds good, I'll take a look only if I can't figure it out

[lucidrains]

@lunixbochs want to try 0.8.2 and see if it resolves your issue?

[lunixbochs]

just updated, not hanging for me anymore. thanks!

[lucidrains]

if you see anything wrong with the loss curve, let me know, made some risky changes