license: other
license_name: qwen
license_link: >-
https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT
language:
- en
- zh
library_name: transformers
pipeline_tag: text-generation
inference: false
tags:
- mistral
- qwen
- qwen1.5
- qwen2
This is the Mistral version of Qwen1.5-14B-Chat model by Alibaba Cloud. The original codebase can be found at: (https://github.com/hiyouga/LLaMA-Factory/blob/main/tests/llamafy_qwen.py). I have made modifications to make it compatible with qwen1.5. This model is converted with https://github.com/Minami-su/character_AI_open/blob/main/mistral_qwen2.py
special
1.Before using this model, you need to modify modeling_mistral.py in transformers library
2.vim /root/anaconda3/envs/train/lib/python3.9/site-packages/transformers/models/mistral/modeling_mistral.py
3.find MistralAttention,
4.modify q,k,v,o bias=False ----->, bias=config.attention_bias
Before:
After:
Differences between qwen2 mistral and qwen2 llamafy
Compared to qwen2 llamafy,qwen2 mistral can use sliding window attention,qwen2 mistral is faster than qwen2 llamafy, and the context length is better
Usage:
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
tokenizer = AutoTokenizer.from_pretrained("Minami-su/Qwen1.5-14B-Chat_mistral")
model = AutoModelForCausalLM.from_pretrained("Minami-su/Qwen1.5-14B-Chat_mistral", torch_dtype="auto", device_map="auto")
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
messages = [
{"role": "user", "content": "Who are you?"}
]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt")
inputs = inputs.to("cuda")
generate_ids = model.generate(inputs,max_length=32768, streamer=streamer)
Test
load in 4bit
hf-causal (pretrained=Qwen1.5-14B-Chat), limit: None, provide_description: False, num_fewshot: 0, batch_size: 16
| Task |Version| Metric |Value | |Stderr|
|-------------|------:|--------|-----:|---|-----:|
|arc_challenge| 0|acc |0.4437|± |0.0145|
| | |acc_norm|0.4718|± |0.0146|
|truthfulqa_mc| 1|mc1 |0.4468|± |0.0174|
| | |mc2 |0.6310|± |0.0157|
|winogrande | 0|acc |0.6788|± |0.0131|
load in 4bit
hf-causal (pretrained=Qwen1.5-14B-Chat_mistral), limit: None, provide_description: False, num_fewshot: 0, batch_size: 16
| Task |Version| Metric |Value | |Stderr|
|-------------|------:|--------|-----:|---|-----:|
|arc_challenge| 0|acc |0.4445|± |0.0145|
| | |acc_norm|0.4718|± |0.0146|
|truthfulqa_mc| 1|mc1 |0.4468|± |0.0174|
| | |mc2 |0.6310|± |0.0157|
|winogrande | 0|acc |0.6788|± |0.0131|