text-generation-webui/extensions/llava/README.md

49 lines
3.9 KiB
Markdown
Raw Normal View History

2023-04-23 23:32:22 +00:00
# LLaVA
## Description
Adds [LLaVA](https://github.com/haotian-liu/LLaVA) multimodality support to text-generation-webui.
https://user-images.githubusercontent.com/3718215/233817203-69b57e77-0c55-4fd6-b742-3204bb13b8fc.mp4
## Usage
To run this extension, download LLaVA weights, for example from [here](https://huggingface.co/wojtab/llava-13b-v0-4bit-128g), and then start server.py with `--extensions llava` argument.
When in ui, go to instruct mode, and select LLaVA template, you also should add `"\n###"` to "Custom stopping strings" in parameters tab.
Do note, that each image takes up 258 tokens, so adjust max_new_tokens to be at most 1700 (recommended value is between 200 to 500), so the images don't get truncated.
To send an image, just upload it to the extension field below chat, and send a prompt as always. The image will be added to the end of your message. If you wish to modify the placement, include a string `<image>` in your prompt.
Additionally, there is *Embed all images, not only the last one* checkbox. It modifies the image embeddings, by default (if it's unchecked), all but the most recent images have their embeddings empty, so they are not fed to the network. From initial testing, it seems as LLaVA considers the features in all images at the same time, so by default the extension skips previous images. If you want to include them anyway, just tick this checkbox.
## Extension config
This extension uses following parameters (from settings.json):
|Parameter|Description|
|---------|-----------|
|`llava-clip_bits`|Number of bits to load CLIP feature extractor in (either 32 or 16, default=32)|
|`llava-clip_device`|Torch device to run the extractor on, for example `cpu` or `cuda:0`, by default `cuda:0` if available|
|`llava-projector_bits`|Number of bits to load CLIP->LLaMA feature projector in (either 32 or 16, default=32)|
|`llava-projector_bits`|Torch device to run the CLIP->LLaMA feature projector on, for example `cpu` or `cuda:0`, by default `cuda:0` if available|
|`llava-add_all_images_to_prompt`|Default value of "Embed all images, not only the last one" checkbox|
## Technical description
### Original LLaVA
The default LLaVA implementation uses modified `transformers` library, however this extension forgoes this requirement. The transformers are modified in LLaVA in such a way, that the entire LLaVA model gets loaded, and the inference now looks as follows:
```
images --> CLIP --> projector --> input embeddings for images --> |
| --> LLaMA
prompt -------------------------> input embeddings for text ----> |
```
The images are represented in the prompt by the following token IDs:
- 32000 - `<im_patch>` - placeholder token for embeddings from projector
- 32001 - `<im_start>` - token marking start of an image
- 32002 - `<im_end>` - token marking end of an image
By default, image will be represented as `<im_start><im_patch>*256<im_end>`. The input embeddings for an image are converted with a single linear layer of the projector, then they are placed instead of `<im_patch>` tokens.
The concatenated prompt then gets fed to fine-tuned LLaMA.
### In this extension
Using default transformers, they only load the LLaMA part of LLaVA, ignoring the added projector weights, and not loading CLIP. We then reconstruct the `images -> CLIP -> projector` pipeline ourselves, then concatenate the input embeddings, and feed it to LLaMA loaded by transformers. This allows us to use normal flow from webui to load this model, and just hijack the model input with additional features.
Splitting it to 3 separate models, allows us to configure each of them, and to move them to different devices(for example we can run CLIP+projector on CPU and LLaMA on GPU). Also, it enables us to use 4-bit GPTQ quantization for LLaVA, massively cutting down the VRAM requirement (it should be possible to fit on 12GB of VRAM with full context size by moving CLIP and projector to CPU).