SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Add RWKV-4 (#22797)

Browse Source 
* First draft of RWKV-4

* Add support for generate

* Style post-rebase

* Properly use state

* Write doc

* Fix doc

* More math

* Add model to README, dummies and clean config

* Fix init

* multiple fixes:

- fix common tests
- fix configuraion default values
- add CI test for checking state computation
- fix some CI tests

* correct tokenizer

* some tweaks

- fix config docstring
- fix failing tests

* fix CI tests

- add output_attention / output_hidden_states
- override test_initialization
- fix failing CIs

* fix conversion script

- fix sharded case
- add new arguments

* add slow tests + more fixes on conversion script

* add another test

* final fixes

* change single name variable

* add mock attention mask for pipeline to work

* correct eos token id

* fix nits

* add checkpoints

* Apply suggestions from code review

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* add `tie_word_embeddings` in docstring

* change tensor name

* fix final nits

* Trigger CI

---------

Co-authored-by: younesbelkada <younesbelkada@gmail.com>
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
This commit is contained in:
Sylvain Gugger
2023-05-09 13:04:10 -04:00
committed by GitHub
parent 9a50cb6195
commit b4d4d6fe87
28 changed files with 2284 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docs/source/en/_toctree.yml
View File

@@ -399,6 +399,8 @@
title: RoCBert
- local: model_doc/roformer
title: RoFormer
- local: model_doc/rwkv
title: RWKV
- local: model_doc/splinter
title: Splinter
- local: model_doc/squeezebert

2
docs/source/en/index.mdx
View File

@@ -196,6 +196,7 @@ The documentation is organized into five sections:
1. **[RoBERTa-PreLayerNorm](model_doc/roberta-prelayernorm)** (from Facebook) released with the paper [fairseq: A Fast, Extensible Toolkit for Sequence Modeling](https://arxiv.org/abs/1904.01038) by Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, Michael Auli.
1. **[RoCBert](model_doc/roc_bert)** (from WeChatAI) released with the paper [RoCBert: Robust Chinese Bert with Multimodal Contrastive Pretraining](https://aclanthology.org/2022.acl-long.65.pdf) by HuiSu, WeiweiShi, XiaoyuShen, XiaoZhou, TuoJi, JiaruiFang, JieZhou.
1. **[RoFormer](model_doc/roformer)** (from ZhuiyiTechnology), released together with the paper [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/abs/2104.09864) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
1. **[RWKV](model_doc/rwkv)** (from Bo Peng), released on [this repo](https://github.com/BlinkDL/RWKV-LM) by Bo Peng.
1. **[SegFormer](model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo.
1. **[Segment Anything](model_doc/sam)** (from Meta AI) released with the paper [Segment Anything](https://arxiv.org/pdf/2304.02643v1.pdf) by Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alex Berg, Wan-Yen Lo, Piotr Dollar, Ross Girshick.
1. **[SEW](model_doc/sew)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi.
@@ -396,6 +397,7 @@ Flax), PyTorch, and/or TensorFlow.
| RoBERTa-PreLayerNorm | ❌ | ❌ | ✅ | ✅ | ✅ |
| RoCBert | ✅ | ❌ | ✅ | ❌ | ❌ |
| RoFormer | ✅ | ✅ | ✅ | ✅ | ✅ |
| RWKV | ❌ | ❌ | ✅ | ❌ | ❌ |
| SAM | ❌ | ❌ | ✅ | ❌ | ❌ |
| SegFormer | ❌ | ❌ | ✅ | ✅ | ❌ |
| SEW | ❌ | ❌ | ✅ | ❌ | ❌ |

129
docs/source/en/model_doc/rwkv.mdx Normal file
View File

@@ -0,0 +1,129 @@
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# RWKV
## Overview
The RWKV model was proposed in [this repo](https://github.com/BlinkDL/RWKV-LM)
It suggests a tweak in the traditional Transformer attention to make it linear. This way, the model can be used as recurrent network: passing inputs for timestamp 0 and timestamp 1 together is the same as passing inputs at timestamp 0, then inputs at timestamp 1 along with the state of timestamp 0 (see example below).
This can be more efficient than a regular Transformer and can deal with sentence of any length (even if the model uses a fixed context length for training).
This model was contributed by [sgugger](https://huggingface.co/sgugger).
The original code can be found [here](https://github.com/BlinkDL/RWKV-LM).
Example of use as an RNN:
```py
import torch
from transformers import AutoTokenizer, RwkvConfig, RwkvModel
model = RwkvModel.from_pretrained("sgugger/rwkv-430M-pile")
tokenizer = AutoTokenizer.from_pretrained("sgugger/rwkv-430M-pile")
inputs = tokenizer("This is an example.", return_tensors="pt")
# Feed everything to the model
outputs = model(inputs["input_ids"])
output_whole = outputs.last_hidden_state
outputs = model(inputs["input_ids"][:, :2])
output_one = outputs.last_hidden_state
# Using the state computed on the first inputs, we will get the same output
outputs = model(inputs["input_ids"][:, 2:], state=outputs.state)
output_two = outputs.last_hidden_state
torch.allclose(torch.cat([output_one, output_two], dim=1), output_whole, atol=1e-5)
```
## RwkvConfig
[[autodoc]] RwkvConfig
## RwkvModel
[[autodoc]] RwkvModel
- forward
## RwkvLMHeadModel
[[autodoc]] RwkvForCausalLM
- forward
## Rwkv attention and the recurrent formulas
In a traditional auto-regressive Transformer, attention is written as
$$O = \hbox{softmax}(QK^{T} / \sqrt{d}) V$$
with \\(Q\\), \\(K\\) and \\(V\\) are matrices of shape `seq_len x hidden_size` named query, key and value (they are actually bigger matrices with a batch dimension and an attention head dimension but we're only interested in the last two, which is where the matrix product is taken, so for the sake of simplicity we only consider those two). The product \\(QK^{T}\\) then has shape `seq_len x seq_len` and we can take the maxtrix product with \\(V\\) to get the output \\(O\\) of the same shape as the others.
Replacing the softmax by its value gives:
$$O_{i} = \frac{\sum_{j=1}^{i} e^{Q_{i} K_{j}^{T} / \sqrt{d}} V_{j}}{\sum_{j=1}^{i} e^{Q_{i} K_{j}^{T} / \sqrt{d}}}$$
Note that the entries in \\(QK^{T}\\) corresponding to \\(j > i\\) are masked (the sum stops at j) because the attention is not allowed to look at future tokens (only past ones).
In comparison, the RWKV attention is given by
$$O_{i} = \sigma(R_{i}) \frac{\sum_{j=1}^{i} e^{W_{i-j} + K_{j}} V_{j}}{\sum_{j=1}^{i} e^{W_{i-j} + K_{j}}}$$
where \\(R\\) is a new matrix called receptance by the author, \\(K\\) and \\(V\\) are still the key and value (\\(\sigma\\) here is the sigmoid function). \\(W\\) is a new vector that represents the position of the token and is given by
$$W_{0} = u \hbox{ and } W_{k} = (k-1)w \hbox{ for } k \geq 1$$
with \\(u\\) and \\(w\\) learnable parameters called in the code `time_first` and `time_decay` respectively. The numerator and denominator can both be expressed recursively. Naming them \\(N_{i}\\) and \\(D_{i}\\) we have:
$$N_{i} = e^{u + K_{i}} V_{i} + \hat{N}_{i} \hbox{ where } \hat{N}_{i} = e^{K_{i-1}} V_{i-1} + e^{w + K_{i-2}} V_{i-2} \cdots + e^{(i-2)w + K_{1}} V_{1}$$
so \\(\hat{N}_{i}\\) (called `numerator_state` in the code) satistfies
$$\hat{N}_{0} = 0 \hbox{ and } \hat{N}_{j+1} = e^{K_{j}} V_{j} + e^{w} \hat{N}_{j}$$
and
$$D_{i} = e^{u + K_{i}} + \hat{D}_{i} \hbox{ where } \hat{D}_{i} = e^{K_{i-1}} + e^{w + K_{i-2}} \cdots + e^{(i-2)w + K_{1}}$$
so \\(\hat{D}_{i}\\) (called `denominator_state` in the code) satistfies
$$\hat{D}_{0} = 0 \hbox{ and } \hat{D}_{j+1} = e^{K_{j}} + e^{w} \hat{D}_{j}$$
The actual recurrent formula used are a tiny bit more complex, as for numerical stability we don't want to compute exponentials of big numbers. Usually the softmax is not computed as is, but the exponential of the maximum term is divided of the numerator and denominator:
$$\frac{e^{x_{i}}}{\sum_{j=1}^{n} e^{x_{j}}} = \frac{e^{x_{i} - M}}{\sum_{j=1}^{n} e^{x_{j} - M}}$$
with \\(M\\) the maximum of all \\(x_{j}\\). So here on top of saving the numerator state (\\(\hat{N}\\)) and the denominator state (\\(\hat{D}\\)) we also keep track of the maximum of all terms encountered in the exponentials. So we actually use
$$\tilde{N}_{i} = e^{-M_{i}} \hat{N}_{i} \hbox{ and } \tilde{D}_{i} = e^{-M_{i}} \hat{D}_{i}$$
defined by the following recurrent formulas:
$$\tilde{N}_{0} = 0 \hbox{ and } \tilde{N}_{j+1} = e^{K_{j} - q} V_{j} + e^{w + M_{j} - q} \tilde{N}_{j} \hbox{ where } q = \max(K_{j}, w + M_{j})$$
and
$$\tilde{D}_{0} = 0 \hbox{ and } \tilde{D}_{j+1} = e^{K_{j} - q} + e^{w + M_{j} - q} \tilde{D}_{j} \hbox{ where } q = \max(K_{j}, w + M_{j})$$
and \\(M_{j+1} = q\\). With those, we can then compute
$$N_{i} = e^{u + K_{i} - q} V_{i} + e^{M_{i}} \tilde{N}_{i} \hbox{ where } q = \max(u + K_{i}, M_{i})$$
and
$$D_{i} = e^{u + K_{i} - q} + e^{M_{i}} \tilde{D}_{i} \hbox{ where } q = \max(u + K_{i}, M_{i})$$
which finally gives us
$$O_{i} = \sigma(R_{i}) \frac{N_{i}}{D_{i}}$$

2
docs/source/en/tasks/language_modeling.mdx
View File

@@ -33,8 +33,8 @@ You can finetune other architectures for causal language modeling following the
Choose one of the following architectures:
<!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
[BART](../model_doc/bart), [BERT](../model_doc/bert), [Bert Generation](../model_doc/bert-generation), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CodeGen](../model_doc/codegen), [CPM-Ant](../model_doc/cpmant), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [GIT](../model_doc/git), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT NeoX Japanese](../model_doc/gpt_neox_japanese), [GPT-J](../model_doc/gptj), [LLaMA](../model_doc/llama), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MVP](../model_doc/mvp), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Pegasus](../model_doc/pegasus), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [RWKV](../model_doc/rwkv), [Speech2Text2](../model_doc/speech_to_text_2), [Transformer-XL](../model_doc/transfo-xl), [TrOCR](../model_doc/trocr), [XGLM](../model_doc/xglm), [XLM](../model_doc/xlm), [XLM-ProphetNet](../model_doc/xlm-prophetnet), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod)
[BART](../model_doc/bart), [BERT](../model_doc/bert), [Bert Generation](../model_doc/bert-generation), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CodeGen](../model_doc/codegen), [CPM-Ant](../model_doc/cpmant), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [GIT](../model_doc/git), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT NeoX Japanese](../model_doc/gpt_neox_japanese), [GPT-J](../model_doc/gptj), [LLaMA](../model_doc/llama), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MVP](../model_doc/mvp), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Pegasus](../model_doc/pegasus), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [Speech2Text2](../model_doc/speech_to_text_2), [Transformer-XL](../model_doc/transfo-xl), [TrOCR](../model_doc/trocr), [XGLM](../model_doc/xglm), [XLM](../model_doc/xlm), [XLM-ProphetNet](../model_doc/xlm-prophetnet), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod)
<!--End of the generated tip-->