@huggingface/gguf
Version:
a GGUF parser that works on remotely hosted files
193 lines (137 loc) • 7.28 kB
Markdown
A GGUF parser that works on remotely hosted files.
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/gguf-spec.png"/>
Spec: https://github.com/ggerganov/ggml/blob/master/docs/gguf.md
Reference implementation (Python): https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/gguf/gguf_reader.py
## Install
```bash
npm install @huggingface/gguf
```
## Usage
### Basic usage
```ts
import { GGMLQuantizationType, gguf } from "@huggingface/gguf";
// remote GGUF file from https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF
const URL_LLAMA = "https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/resolve/191239b/llama-2-7b-chat.Q2_K.gguf";
const { metadata, tensorInfos } = await gguf(URL_LLAMA);
console.log(metadata);
// {
// version: 2,
// tensor_count: 291n,
// kv_count: 19n,
// "general.architecture": "llama",
// "general.file_type": 10,
// "general.name": "LLaMA v2",
// ...
// }
console.log(tensorInfos);
// [
// {
// name: "token_embd.weight",
// shape: [4096n, 32000n],
// dtype: GGMLQuantizationType.Q2_K,
// },
// ... ,
// {
// name: "output_norm.weight",
// shape: [4096n],
// dtype: GGMLQuantizationType.F32,
// }
// ]
```
```ts
// Reading a local file. (Not supported on browser)
const { metadata, tensorInfos } = await gguf(
'./my_model.gguf',
{ allowLocalFile: true },
);
```
You can get metadata with type information by setting `typedMetadata: true`. This provides both the original value and its GGUF data type:
```ts
import { GGMLQuantizationType, GGUFValueType, gguf } from "@huggingface/gguf";
const URL_LLAMA = "https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/resolve/191239b/llama-2-7b-chat.Q2_K.gguf";
const { metadata, typedMetadata } = await gguf(URL_LLAMA, { typedMetadata: true });
console.log(typedMetadata);
// {
// version: { value: 2, type: GGUFValueType.UINT32 },
// tensor_count: { value: 291n, type: GGUFValueType.UINT64 },
// kv_count: { value: 19n, type: GGUFValueType.UINT64 },
// "general.architecture": { value: "llama", type: GGUFValueType.STRING },
// "general.file_type": { value: 10, type: GGUFValueType.UINT32 },
// "general.name": { value: "LLaMA v2", type: GGUFValueType.STRING },
// "llama.attention.head_count": { value: 32, type: GGUFValueType.UINT32 },
// "llama.attention.layer_norm_rms_epsilon": { value: 9.999999974752427e-7, type: GGUFValueType.FLOAT32 },
// "tokenizer.ggml.tokens": { value: ["<unk>", "<s>", "</s>", ...], type: GGUFValueType.ARRAY, subType: GGUFValueType.STRING },
// "tokenizer.ggml.scores": { value: [0.0, -1000.0, -1000.0, ...], type: GGUFValueType.ARRAY, subType: GGUFValueType.FLOAT32 },
// ...
// }
// Access both value and type information
console.log(typedMetadata["general.architecture"].value); // "llama"
console.log(typedMetadata["general.architecture"].type); // GGUFValueType.STRING (8)
// For arrays, subType indicates the type of array elements
console.log(typedMetadata["tokenizer.ggml.tokens"].type); // GGUFValueType.ARRAY (9)
console.log(typedMetadata["tokenizer.ggml.tokens"].subType); // GGUFValueType.STRING (8)
```
By default, known fields in `metadata` are typed. This includes various fields found in [llama.cpp](https://github.com/ggerganov/llama.cpp), [whisper.cpp](https://github.com/ggerganov/whisper.cpp) and [ggml](https://github.com/ggerganov/ggml).
```ts
const { metadata, tensorInfos } = await gguf(URL_MODEL);
// Type check for model architecture at runtime
if (metadata["general.architecture"] === "llama") {
// "llama.attention.head_count" is a valid key for llama architecture, this is typed as a number
console.log(model["llama.attention.head_count"]);
// "mamba.ssm.conv_kernel" is an invalid key, because it requires model architecture to be mamba
console.log(model["mamba.ssm.conv_kernel"]); // error
}
```
Because GGUF format can be used to store tensors, we can technically use it for other usages. For example, storing [control vectors](https://github.com/ggerganov/llama.cpp/pull/5970), [lora weights](https://github.com/ggerganov/llama.cpp/pull/2632), etc.
In case you want to use your own GGUF metadata structure, you can disable strict typing by casting the parse output to `GGUFParseOutput<{ strict: false }>`:
```ts
const { metadata, tensorInfos }: GGUFParseOutput<{ strict: false }> = await gguf(URL_LLAMA);
```
This package provides a CLI equivalent to [`gguf_dump.py`](https://github.com/ggml-org/llama.cpp/blob/7a2c913e66353362d7f28d612fd3c9d51a831eda/gguf-py/gguf/scripts/gguf_dump.py) script. You can dump GGUF metadata and list of tensors using this command:
```bash
npx @huggingface/gguf my_model.gguf
```
Example for the output:
```
* Dumping 36 key/value pair(s)
Idx | Count | Value
----|--------|----------------------------------------------------------------------------------
1 | 1 | version = 3
2 | 1 | tensor_count = 292
3 | 1 | kv_count = 33
4 | 1 | general.architecture = "llama"
5 | 1 | general.type = "model"
6 | 1 | general.name = "Meta Llama 3.1 8B Instruct"
7 | 1 | general.finetune = "Instruct"
8 | 1 | general.basename = "Meta-Llama-3.1"
[]
* Dumping 292 tensor(s)
Idx | Num Elements | Shape | Data Type | Name
----|--------------|--------------------------------|-----------|--------------------------
1 | 64 | 64, 1, 1, 1 | F32 | rope_freqs.weight
2 | 525336576 | 4096, 128256, 1, 1 | Q4_K | token_embd.weight
3 | 4096 | 4096, 1, 1, 1 | F32 | blk.0.attn_norm.weight
4 | 58720256 | 14336, 4096, 1, 1 | Q6_K | blk.0.ffn_down.weight
[]
```
Alternatively, you can install this package as global, which will provide the `gguf-view` command:
```bash
npm i -g @huggingface/gguf
gguf-view my_model.gguf
```
The Hub supports all file formats and has built-in features for GGUF format.
Find more information at: http://hf.co/docs/hub/gguf.
- https://github.com/hyparam/hyllama by @platypii (MIT license)
- https://github.com/ahoylabs/gguf.js by @biw @dkogut1996 @spencekim (MIT license)
🔥❤️