File size: 5,285 Bytes
922e22a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 |
---
datasets:
- EleutherAI/pile
---
# Model card for RWKV-4 | 7B parameters trained on Pile dataset
RWKV is a project led by [Bo Peng](https://github.com/BlinkDL). Learn more about the model architecture in the blogposts from Johan Wind [here](https://johanwind.github.io/2023/03/23/rwkv_overview.html) and [here](https://johanwind.github.io/2023/03/23/rwkv_details.html). Learn more about the project by joining the [RWKV discord server](https://discordapp.com/users/468093332535640064).
# Table of contents
0. [TL;DR](#TL;DR)
1. [Model Details](#model-details)
2. [Usage](#usage)
3. [Citation](#citation)
## TL;DR
Below is the description from the [original repository](https://github.com/BlinkDL/RWKV-LM)
> RWKV is an RNN with transformer-level LLM performance. It can be directly trained like a GPT (parallelizable). It's combining the best of RNN and transformer - great performance, fast inference, saves VRAM, fast training, "infinite" ctx_len, and free sentence embedding.
## Model Details
The details of the architecture can be found on the blogpost mentioned above and the Hugging Face blogpost of the integration.
## Usage
### Convert the raw weights to the HF format
You can use the [`convert_rwkv_checkpoint_to_hf.py`](https://github.com/huggingface/transformers/tree/main/src/transformers/models/rwkv/convert_rwkv_checkpoint_to_hf.py) script by specifying the repo_id of the original weights, the filename and the output directory. You can also optionally directly push the converted model on the Hub by passing `--push_to_hub` flag and `--model_name` argument to specify where to push the converted weights.
```bash
python convert_rwkv_checkpoint_to_hf.py --repo_id RAW_HUB_REPO --checkpoint_file RAW_FILE --output_dir OUTPUT_DIR --push_to_hub --model_name dummy_user/converted-rwkv
```
### Generate text
You can use the `AutoModelForCausalLM` and `AutoTokenizer` classes to generate texts from the model. Expand the sections below to understand how to run the model in different scenarios:
### Running the model on a CPU
<details>
<summary> Click to expand </summary>
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("RWKV/rwkv-4-7b-pile")
tokenizer = AutoTokenizer.from_pretrained("RWKV/rwkv-4-7b-pile")
prompt = "\nIn a shocking finding, scientist discovered a herd of dragons living in a remote, previously unexplored valley, in Tibet. Even more surprising to the researchers was the fact that the dragons spoke perfect Chinese."
inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(inputs["input_ids"], max_new_tokens=40)
print(tokenizer.decode(output[0].tolist(), skip_special_tokens=True))
```
### Running the model on a single GPU
<details>
<summary> Click to expand </summary>
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("RWKV/rwkv-4-7b-pile").to(0)
tokenizer = AutoTokenizer.from_pretrained("RWKV/rwkv-4-7b-pile")
prompt = "\nIn a shocking finding, scientist discovered a herd of dragons living in a remote, previously unexplored valley, in Tibet. Even more surprising to the researchers was the fact that the dragons spoke perfect Chinese."
inputs = tokenizer(prompt, return_tensors="pt").to(0)
output = model.generate(inputs["input_ids"], max_new_tokens=40)
print(tokenizer.decode(output[0].tolist(), skip_special_tokens=True))
```
</details>
</details>
### Running the model in half-precision, on GPU
<details>
<summary> Click to expand </summary>
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("RWKV/rwkv-4-7b-pile", torch_dtype=torch.float16).to(0)
tokenizer = AutoTokenizer.from_pretrained("RWKV/rwkv-4-7b-pile")
prompt = "\nIn a shocking finding, scientist discovered a herd of dragons living in a remote, previously unexplored valley, in Tibet. Even more surprising to the researchers was the fact that the dragons spoke perfect Chinese."
inputs = tokenizer(prompt, return_tensors="pt").to(0)
output = model.generate(inputs["input_ids"], max_new_tokens=40)
print(tokenizer.decode(output[0].tolist(), skip_special_tokens=True))
```
</details>
### Running the model multiple GPUs
<details>
<summary> Click to expand </summary>
```python
# pip install accelerate
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("RWKV/rwkv-4-7b-pile", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("RWKV/rwkv-4-7b-pile")
prompt = "\nIn a shocking finding, scientist discovered a herd of dragons living in a remote, previously unexplored valley, in Tibet. Even more surprising to the researchers was the fact that the dragons spoke perfect Chinese."
inputs = tokenizer(prompt, return_tensors="pt").to(0)
output = model.generate(inputs["input_ids"], max_new_tokens=40)
print(tokenizer.decode(output[0].tolist(), skip_special_tokens=True))
```
</details>
## Citation
If you use this model, please consider citing the original work, from the original repo [here](https://github.com/BlinkDL/ChatRWKV/) |