OPEA/DeepSeek-V3-int4-sym-inc-cpu

Model Details

This model is an int4 model with group_size 128 and symmetric quantization of deepseek-ai/DeepSeek-V3 generated by intel/auto-round algorithm.

On CUDA devices, this model is prone to overflow caused by the INT4 kernel using the FP16 computation dtype. Additionally, loading the model in Transformers can be quite slow. Consider using an alternative serving framework capable of running INT4 models with the BF16 computation dtype.

Due to limited GPU resources, we have only tested a few prompts on a CPU backend with intel-extension-for-transformers . If this model does not meet your performance expectations, you may explore another quantized model in AWQ format, generated via AutoRound with different hyperparameters. This alternative model will be uploaded soon.

Please follow the license of the original model.

How To Use

INT4 Inference on CPU with ITREX(Recommended)

pip3 install auto-round (it will install intel-extension-for-pytorch and intel-extension-for-transformers both). For intel cpu, it will prioritize using intel-extension-for-pytorch , for other cpus, it will prioritize using intel-extension-for-transformers.

To make sure to use qbits with intel-extension-for-transformers, please uninstall intel-extension-for-pytorch

intel-extension-for-transformers: faster repacking, slower inference,higher accuracy

intel-extension-for-pytorch: much slower repacking, faster inference, lower accuracy

from auto_round import AutoRoundConfig ##must import for autoround format
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

quantized_model_dir = "OPEA/DeepSeek-V3-int4-sym-inc-cpu"
quantization_config = AutoRoundConfig(
    backend="cpu"
)
model = AutoModelForCausalLM.from_pretrained(
    quantized_model_dir,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    device_map="cpu",
    revision="8fe0735",##use autoround format, the only difference is config.json
    quantization_config = quantization_config, ##cpu only machine don't need to set this value
    
)

tokenizer = AutoTokenizer.from_pretrained(quantized_model_dir,trust_remote_code=True)
prompt = "There is a girl who likes adventure,"
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    model_inputs.input_ids,
    max_new_tokens=200,  ##change this to align with the official usage
    do_sample=False  ##change this to align with the official usage
)
generated_ids = [
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

prompt = "9.11和9.8哪个数字大"  

##INT4
"""要比较 **9.11** 和 **9.8** 的大小，可以按照以下步骤进行：

1. **比较整数部分**：
   - 两个数的整数部分都是 **9**，所以整数部分相同。

2. **比较小数部分**：
   - **9.11** 的小数部分是 **0.11**
   - **9.8** 的小数部分是 **0.8**（即 **0.80**）

3. **分析小数部分**：
   - **0.80** 大于 **0.11**

因此，**9.8** 大于 **9.11**。

最终答案：\boxed{9.8}

"""

prompt = "strawberry中有几个r?"
##INT4
"""
### 第一步：理解问题

首先，我需要明确问题的含义。问题是：“strawberry中有几个r？”。这里的“strawberry”是一个英文单词，意思是“草莓”。问题问的是这个单 词中有多少个字母“r”。

### 第二步：分解单词

为了找出“strawberry”中有多少个“r”，我需要将这个单词分解成单个字母。让我们逐个字母来看：

- s
- t
- r
- a
- w
- b
- e
- r
- r
- y

### 第三步：识别字母“r”

现在，我需要找出这些字母中哪些是“r”。让我们逐一检查：

1. s - 不是r
2. t - 不是r
3. r - 是r
4. a - 不是r
5. w - 不是r
6. b - 不是r
7. e - 不是r
8. r - 是r
"""

prompt = "How many r in strawberry."
##INT4 
"""The word "strawberry" contains **3 "r"s.
"""

prompt = "There is a girl who likes adventure,"
##INT4:
"""That's wonderful! A girl who loves adventure is likely curious, brave, and eager to explore the world around her. Here are some ideas to fuel her adventurous spirit:

### **Outdoor Adventures**

- **Hiking:** Explore local trails, national parks, or mountains.
- **Camping:** Spend a night under the stars and connect with nature.
- **Rock Climbing:** Challenge herself with bouldering or climbing walls.
- **Kayaking/Canoeing:** Paddle through rivers, lakes, or even the ocean.
- **Zip-lining:** Soar through the treetops for an adrenaline rush.

### **Travel Adventures**

- **Road Trips:** Plan a journey to new cities or scenic destinations.
- **Backpacking:** Travel light and explore different cultures and landscapes.
- **Volunteer Abroad:** Combine adventure with helping others in a new country.

### **Creative Adventures**

- **Photography:** Capture the beauty
"""

prompt = "Please give a brief introduction of DeepSeek company."
##INT4:
"""DeepSeek Artificial Intelligence Co., Ltd. (referred to as "DeepSeek" or "深度求索") , founded in 2023, is a Chinese company dedicated to making AGI a reality"""

INT4 Inference on CUDA(have not tested, maybe need 8X80G GPU)

Int4 kernel with BF16 computing dtype is required.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
quantized_model_dir = "OPEA/DeepSeek-V3-int4-sym-gptq-inc-preview"

model = AutoModelForCausalLM.from_pretrained(
    quantized_model_dir,
    torch_dtype=torch.float16,
    trust_remote_code=True,
    device_map="auto"
    
)

tokenizer = AutoTokenizer.from_pretrained(quantized_model_dir,trust_remote_code=True)
prompt = "There is a girl who likes adventure,"
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    model_inputs.input_ids,
    max_new_tokens=200,  ##change this to align with the official usage
    do_sample=False  ##change this to align with the official usage
)
generated_ids = [
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

Evaluate the model

we have no enough resource to evaluate the model

Generate the model

5*80G gpu is needed(could optimize), 1.4T cpu memory is needed

We discovered that the inputs and outputs of certain layers in this model are very large and even exceed the FP16 range when tested with a few prompts. It is recommended to exclude these layers from quantization—particularly the 'down_proj' in layer 60—and run them using BF16 precision instead. However, we have not implemented this in this int4 model as in cpu, the compute dtype for int4 is bf16 or FP32.

model.layers.60.mlp.experts.150.down_proj tensor(1144.) tensor(2122.9451)
model.layers.60.mlp.experts.231.down_proj tensor(25856.) tensor(12827.9980)
model.layers.60.mlp.shared_experts.down_proj tensor(1880.) tensor(3156.7344)
model.layers.60.mlp.experts.81.down_proj tensor(4416.) tensor(6124.6846)
model.layers.60.mlp.experts.92.down_proj tensor(107520.) tensor(50486.0781)
model.layers.59.mlp.experts.138.down_proj tensor(1568.) tensor(190.8769)
model.layers.60.mlp.experts.81.down_proj tensor(7360.) tensor(10024.4531)
model.layers.60.mlp.experts.92.down_proj tensor(116224.) tensor(55192.4180)

1 add meta data to bf16 model https://huggingface.co/opensourcerelease/DeepSeek-V3-bf16

import safetensors
from safetensors.torch import save_file
 
for i in range(1, 164):
    idx_str = "0" * (5-len(str(i))) + str(i)
    safetensors_path = f"model-{idx_str}-of-000163.safetensors"
    print(safetensors_path)
    tensors = dict()
    with safetensors.safe_open(safetensors_path, framework="pt") as f:
        for key in f.keys():
            tensors[key] = f.get_tensor(key)
    save_file(tensors, safetensors_path, metadata={'format': 'pt'})

2 replace the modeling_deepseek.py with the following file, basically align device and remove torch.no_grad as we need some tuning in AutoRound.

https://github.com/intel/auto-round/blob/deepseekv3/modeling_deepseek.py

3 tuning

git clone https://github.com/intel/auto-round.git && cd auto-round && git checkout deepseekv3

python3 -m auto_round --model  "/models/DeepSeek-V3-bf16/"  --group_size 128 --format "auto_gptq"  --iters 200 --devices 0,1,2,3,4 --nsamples 512 --batch_size 8 --seqlen 512   --low_gpu_mem_usage    --output_dir "tmp_autoround"  --disable_eval e 2>&1 | tee -a seekv3.txt

Ethical Considerations and Limitations

The model can produce factually incorrect output, and should not be relied on to produce factually accurate information. Because of the limitations of the pretrained model and the finetuning datasets, it is possible that this model could generate lewd, biased or otherwise offensive outputs.

Therefore, before deploying any applications of the model, developers should perform safety testing.

Caveats and Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.

Here are a couple of useful links to learn more about Intel's AI software:

• Intel Neural Compressor link

Disclaimer

The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please consult an attorney before using this model for commercial purposes.

Cite

@article{cheng2023optimize, title={Optimize weight rounding via signed gradient descent for the quantization of llms}, author={Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang and He, Xin and Lv, Kaokao and Liu, Yi}, journal={arXiv preprint arXiv:2309.05516}, year={2023} }

arxiv github