93 lines
3.0 KiB
Markdown
93 lines
3.0 KiB
Markdown
|
## **UPDATE: Official version is out, use it instead: [https://huggingface.co/mistralai/Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)**
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
# mistral-7B-v0.1-hf
|
||
|
|
|
||
|
|
Huggingface compatible version of Mistral's 7B model: https://twitter.com/MistralAI/status/1706877320844509405
|
||
|
|
|
||
|
|
## Usage
|
||
|
|
|
||
|
|
### Load in bfloat16 (16GB VRAM or higher)
|
||
|
|
|
||
|
|
```python
|
||
|
|
import torch
|
||
|
|
from transformers import LlamaForCausalLM, LlamaTokenizer, pipeline, TextStreamer
|
||
|
|
|
||
|
|
tokenizer = LlamaTokenizer.from_pretrained("kittn/mistral-7B-v0.1-hf")
|
||
|
|
model = LlamaForCausalLM.from_pretrained(
|
||
|
|
"kittn/mistral-7B-v0.1-hf",
|
||
|
|
torch_dtype=torch.bfloat16,
|
||
|
|
device_map={"": 0}
|
||
|
|
)
|
||
|
|
|
||
|
|
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
|
||
|
|
|
||
|
|
pipe("Hi, my name", streamer=TextStreamer(tokenizer), max_new_tokens=128)
|
||
|
|
```
|
||
|
|
|
||
|
|
### Load in bitsandbytes nf4 (6GB VRAM or higher, maybe less with double_quant)
|
||
|
|
|
||
|
|
```python
|
||
|
|
import torch
|
||
|
|
from transformers import LlamaForCausalLM, LlamaTokenizer, pipeline, TextStreamer, BitsAndBytesConfig
|
||
|
|
|
||
|
|
tokenizer = LlamaTokenizer.from_pretrained("kittn/mistral-7B-v0.1-hf")
|
||
|
|
model = LlamaForCausalLM.from_pretrained(
|
||
|
|
"kittn/mistral-7B-v0.1-hf",
|
||
|
|
device_map={"": 0},
|
||
|
|
quantization_config=BitsAndBytesConfig(
|
||
|
|
load_in_4bit=True,
|
||
|
|
bnb_4bit_compute_dtype=torch.float16,
|
||
|
|
bnb_4bit_quant_type="nf4",
|
||
|
|
bnb_4bit_use_double_quant=False, # set to True to save more VRAM at the cost of some speed/accuracy
|
||
|
|
),
|
||
|
|
)
|
||
|
|
|
||
|
|
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
|
||
|
|
|
||
|
|
pipe("Hi, my name", streamer=TextStreamer(tokenizer), max_new_tokens=128)
|
||
|
|
```
|
||
|
|
|
||
|
|
### Load in bitsandbytes int8 (8GB VRAM or higher). Quite slow; not recommended.
|
||
|
|
|
||
|
|
```python
|
||
|
|
import torch
|
||
|
|
from transformers import LlamaForCausalLM, LlamaTokenizer, pipeline, TextStreamer, BitsAndBytesConfig
|
||
|
|
|
||
|
|
tokenizer = LlamaTokenizer.from_pretrained("kittn/mistral-7B-v0.1-hf")
|
||
|
|
model = LlamaForCausalLM.from_pretrained(
|
||
|
|
"kittn/mistral-7B-v0.1-hf",
|
||
|
|
device_map={"": 0},
|
||
|
|
quantization_config=BitsAndBytesConfig(
|
||
|
|
load_in_8bit=True,
|
||
|
|
),
|
||
|
|
)
|
||
|
|
|
||
|
|
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
|
||
|
|
|
||
|
|
pipe("Hi, my name", streamer=TextStreamer(tokenizer), max_new_tokens=128)
|
||
|
|
```
|
||
|
|
|
||
|
|
|
||
|
|
## Notes
|
||
|
|
|
||
|
|
* The original huggingface conversion script converts the model from bf16 to fp16 before saving it. This script doesn't
|
||
|
|
* The tokenizer is created with `legacy=False`, [more about this here](https://github.com/huggingface/transformers/pull/24565)
|
||
|
|
* Saved in safetensors format
|
||
|
|
|
||
|
|
## Conversion script [[link]](https://gist.github.com/sekstini/151d6946df1f6aa997b7cb15ee6f3be1)
|
||
|
|
|
||
|
|
Unlike [meta-llama/Llama-2-7b](https://huggingface.co/meta-llama/Llama-2-7b), this model uses GQA. This breaks some assumptions in the original conversion script, requiring a few changes.
|
||
|
|
|
||
|
|
Conversion script: [link](https://gist.github.com/sekstini/151d6946df1f6aa997b7cb15ee6f3be1)
|
||
|
|
|
||
|
|
Original conversion script: [link](https://github.com/huggingface/transformers/blob/946bac798caefada3f5f1c9fecdcfd587ed24ac7/src/transformers/models/llama/convert_llama_weights_to_hf.py)
|
||
|
|
|