Sentencepiece alternative

[Muennighoff]

Amazing work! Currently the spmtrain in build_tokenizer doesn't work, cuz I think it needs a local installation of sentencepiece to be able to use the command. Is there a specific reason you chose Google's sentencepiece over just:

from tokenizers import SentencePieceBPETokenizer
tokenizer = SentencePieceBPETokenizer()
tokenizer.train(files=paths, vocab_size=3000, special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"])
tokenizer.save(os.path.join(configs.data.path, "spiece.model"))

Also, I was wondering did fast attention work for you?

[mymusise]

Thanks for the feedback!

1. Yes, the installations of the sentencepiece is required. Try pip install -r requirements.txt to update your environment.
2. It's ok to use tokenizers.SentencePieceBPETokenizer, but both of them require the input sentence should be pretokenized in no-spaces languages such as Chinese or Japanese.
3. I forgot to add pretoken in build_tokenizer.py, I will update it later.
4. fast attention is still not available in this repo, hope norabelrose's work coming soon.

[Muennighoff]

Thanks for the feedback!

1. Yes, the installations of the sentencepiece is required. Try pip install -r requirements.txt to update your environment.
2. It's ok to use tokenizers.SentencePieceBPETokenizer, but both of them require the input sentence should be pretokenized in no-spaces languages such as Chinese or Japanese.
3. I forgot to add pretoken in build_tokenizer.py, I will update it later.
4. fast attention is still not available in this repo, hope norabelrose's work coming soon.

1-3: Great, thanks for the info. 4: Yeah that would be awesome. I've also been working on it on a fork from transformers, but it's still unable to converge somehow, so I was thinking of using the original tensorflow implementation from g-research as you did. But I guess it didnt work out for you?

[mymusise]

so I was thinking of using the original tensorflow implementation from g-research as you did. But I guess it didnt work out for you?

I try both of them two months ago but still have some problems. In my test:

• transformer.PerformerAttention works but it train and predict slower than the normal Transformer.
• the fast-attention of google train and predict faster, but the result of prediction is bad. It's confusing that the accuracy will stop at 66% with training a small corpus like dataset/test/raw.txt. Perhaps I build the model in the wrong way.

Hope it helps :) and wait for your good news :rocket:

[mymusise]

close it, reopen if it needs.

[Muennighoff]

close it, reopen if it needs.

BERT & DistilBERT Performers now work with pretrained transformers (https://colab.research.google.com/drive/1A9reiUZbA7DELuJ8keTo73sIQ4dJJVoT#scrollTo=F5k6jxicGf3E)

Still working on autoregressive models like gpt .. but it seems the gains won't be that big for them; will keep you updated!

[mymusise]

Great work! @Muennighoff After trying your code, BERT & DistilBERT Performers does work better than the normal Transformer. I tried GPT2 work with Performers a few days ago and got some clue, I think the reason of Performer will perform well in non-causal models like BERT and T5 but will have a negative effect in causal models like GPT, maybe is the matrix multiplication of softmax(Q @ K) @ CausalMask @ V is faster then the loop multiplication of _headwise_causal_numerator.

[Muennighoff]

Great work! @Muennighoff After trying your code, BERT & DistilBERT Performers does work better than the normal Transformer. I tried GPT2 work with Performers a few days ago and got some clue, I think the reason of Performer will perform well in non-causal models like BERT and T5 but will have a negative effect in causal models like GPT, maybe is the matrix multiplication of softmax(Q @ K) @ CausalMask @ V is faster then the loop multiplication of _headwise_causal_numerator.

Yeah the problem seems to be the loop multiplication - We could try the causal multiplication from lucidrains - it shouldn't be too difficult to copy it over.

I'm also still running into shape errors when trying GPT2 Performer - did you get it to work?

[mymusise]

Great work! @Muennighoff After trying your code, BERT & DistilBERT Performers does work better than the normal Transformer. I tried GPT2 work with Performers a few days ago and got some clue, I think the reason of Performer will perform well in non-causal models like BERT and T5 but will have a negative effect in causal models like GPT, maybe is the matrix multiplication of softmax(Q @ K) @ CausalMask @ V is faster then the loop multiplication of _headwise_causal_numerator.

Yeah the problem seems to be the loop multiplication - We could try the causal multiplication from luciddrains - it shouldn't be too difficult to copy it over.

I'm also still running into shape errors when trying GPT2 Performer - did you get it to work?

It looks so cool! I'll try it and keep you updated!

[mymusise]

@Muennighoff Hi, I got an OOM error after using the causal_linear_attention_noncuda function from luciddrains. I think it's a bad idea to generate a content matrix with torch.einsum('...nd,...ne->...nde', k, v). If the shape of Q, K, V is [B, H, L, D], and the context will have a huge size [B, H, L, D, D]. In many applications, the value of D is in [64-128]. Maybe it has a little deviation from the design of performer.

[Muennighoff]

@Muennighoff Hi, I got an OOM error after using the causal_linear_attention_noncuda function from luciddrains. I think it's a bad idea to generate a content matrix with torch.einsum('...nd,...ne->...nde', k, v). If the shape of Q, K, V is [B, H, L, D], and the context will have a huge size [B, H, L, D, D]. In many applications, the value of D is in [64-128]. Maybe it has a little deviation from the design of performer.

Yes i think that's why he also has the other version above (https://github.com/lucidrains/performer-pytorch/blob/3bff14e39284e7dc82952153099a63dcd3561dc0/performer_pytorch/performer_pytorch.py#L142) Did you try that one as well?

[mymusise]

Yes i think that's why he also has the other version above (https://github.com/lucidrains/performer-pytorch/blob/3bff14e39284e7dc82952153099a63dcd3561dc0/performer_pytorch/performer_pytorch.py#L142) Did you try that one as well?

I didn't try it, but I think maybe rewriting the loop multiplication with the basic API of CUDA, is a way to work out the performance. Some like fast_transformers.causal_product.CausalDotProduct do.

[JamesDeAntonis]

Hi all, I'm happy to join this conversation if there's anything to be done.

Is this too inefficient?

import numpy as np
  
q0 = np.array([[1, 2, 3], [4, 5, 6]])
k0 = np.array([[2, 3, 4], [5, 6, 7]])
v0 = np.array([[3, 4, 5, 6], [7, 8, 9, 10]])

q = np.expand_dims(q0, axis=1)
k = np.expand_dims(k0, axis=2)
v = np.expand_dims(v0, axis=2).transpose(0, 2, 1)

assert q.shape == (2, 1, 3)
assert k.shape == (2, 3, 1)
assert v.shape == (2, 1, 4)

kv = (k @ v)

assert kv.shape == (2, 3, 4)

kv_sum = np.cumsum(kv, axis=0)

assert kv_sum.shape == (2, 3, 4)

qkv = q @ kv_sum

assert qkv.shape == (2, 1, 4)

qkv = np.squeeze(qkv, axis=1)

assert (qkv == ((q0 @ k0.T) * np.array([[1, 0], [1, 1]])) @ v0).all()

[mymusise]

Welcome guy! @JamesDeAntonis

Is this too inefficient?

No, I think it can speed up the operation but will spend more memory. And I'm not sure the dimension of q0 (2, 3) mean (L, D)? A sequence with two tokens and each token has 3 dimensions?

If so, I think the code above is equal to torch.einsum('...nd,...ne->...nde', k, v) in a way.

As I comment above, it will spend a lot of memory.

If the shape of Q, K, V is [B, H, L, D], and the context will have a huge size [B, H, L, D, D]. In many applications, the value of D is in [64-128]. Maybe it has a little deviation from the design of performer.

[JamesDeAntonis]

And I'm not sure the dimension of q0 (2, 3) mean (L, D)? A sequence with two tokens and each token has 3 dimensions?

Yes, that's what I meant.

If the shape of Q, K, V is [B, H, L, D], and the context will have a huge size [B, H, L, D, D]. In many applications, the value of D is in [64-128]. Maybe it has a little deviation from the design of performer.

Don't you iterate across heads in the code, to save memory? This would mean [B, 1, L, D, D], whereas regular transformer has [B, H, L, L]. Generally speaking, I'm unsure what exactly you are trying to change.

[JamesDeAntonis]

Does this mean CausalDotProduct in fast-attention is what we want?

[mymusise]

Don't you iterate across heads in the code, to save memory?

Great, that's a good idea.

This would mean [B, 1, L, D, D], whereas regular transformer has [B, H, L, L]

Generally speaking, the value of D is in [64-128], if we have L = 2048, D = 128, then L*D*D=2048*128*128 is much bigger than the regular transformer L*L=2048*2048, if H=8, then H*L*L == L*D*D, it means we can save more memory when L>2048. But we still have to iterate each head if we do what you mention above. I‘m not sure I've understood your mind。

I'm unsure what exactly you are trying to change.

Actually, I've no idea to speed up the causal multiplication while saving memory. :mask:

[JamesDeAntonis]

isn't this only solvable by implementing the for-loop directly in the lower-level language? I imagine this is effectively what fast attention does

[mymusise]

Does this mean CausalDotProduct in fast-attention is what we want?

Yes, but I'd not test yea, cause I can't find the tensorflow version and I've not mastered C++. :joy:

[mymusise]

I imagine this is effectively what fast attention does

Yes, I have the same guess.

[JamesDeAntonis]

Can we talk over a call? I just emailed you

[JamesDeAntonis]

I was just reading the fast attention code, and I think it does exactly what we want. Typing is really the only reason the c++ code is torch-specific. Otherwise, all the logic should directly transfer to TF. My issue is that I don't know tf. I think we only need to change the #include line at the top of the c++ file, the typing in the c++ file, and the wrapper code in __init__.py, then it will be TF compatible

[mymusise]

@JamesDeAntonis Cool! I did not read the fast attention code seriously, sounds like it's worth trying.

[JamesDeAntonis]

I don't think I'm the guy to do this (I don't use c++ or tensorflow), but I think this is a pretty easy problem for someone who at least knows tensorflow

I started an implementation here. @mymusise will you take a look?

[mymusise]

Cool! You are so efficient! @JamesDeAntonis Actually, before converting to a tf version, I prefer to test fast-attention.CausalDotProduct in PyTorch first. At the same time, I will try to convert it base on your implementation.

Muennighoff seems to be solving the masking problem of the T5 decoder. @Muennighoff Did you try fast-attention.CausalDotProduct? Dose it work?

[JamesDeAntonis]

@ice-americano (who I work with) ran if and it seemed to work to some degree. Compared to regular attention, he was getting significant improvements in memory usage, but a noticeable slowdown. We're not sure why the slowdown is occurring

[Muennighoff]

@mymusise I made some experiments with the fast-attention.CausalDotProduct on the SST notebook (https://colab.research.google.com/drive/1A9reiUZbA7DELuJ8keTo73sIQ4dJJVoT#scrollTo=f_R7D-mZyRXm) --- & the CausalDotProduct itself seems to work, however I get a shape mismatch in the denominator/normalization, since the sequence length of the key may not be the same as the query in a t5 enc-dec; As the keys come from the encoder, but the queries from the decoder input;

If i force the shapes to fit or just skip the normalization, it goes through but always predicts the same value giving 50% accuracy whereas the equivalent transformer converges (check it under t5/pytorch/encdec in the notebook);

this might tho be unrelated to the CausalDotProduct and could e.g. be caused by the attention mask or sth else --- @JamesDeAntonis did @ice-americano get it to converge?

[JamesDeAntonis]

Yes, we got it to converge.

On our task, in relation to regular attention, it was (i) a step worse in terms of loss, (ii) noticeably lower memory utilization on long sequences, and (iii) slower (this is the most perplexing to us)