---
license: mit
library_name: transformers
base_model:
- deepseek-ai/DeepSeek-V3-0324
---
# huihui-ai/DeepSeek-V3-0324-bf16
This model converted from DeepSeek-V3-0324 to BF16.  
Therefore, we have provided only the command to convert from Windows and information related to ollama. 

The Windows environment is much faster than the WSL2(Ubuntu-22.04) environment, provided you have sufficient memory or virtual memory. The Linux environment hasn't been tested.

If you are in a Linux or WSL environment, please refer to [huihui-ai/DeepSeek-R1-bf16](https://huggingface.co/huihui-ai/DeepSeek-R1-bf16).

If needed, we can upload the bf16 version.

## FP8 to BF16
1. Download [deepseek-ai/DeepSeek-V3-0324](https://huggingface.co/deepseek-ai/DeepSeek-V3-0324) model, requires approximately 641GB of space.
```
cd /d C:\Users\admin\models
huggingface-cli download deepseek-ai/DeepSeek-V3-0324 --local-dir ./deepseek-ai/DeepSeek-V3-0324
```
2. Create the environment.
```
conda create -yn DeepSeek-V3-0324 python=3.10
conda activate DeepSeek-V3
pip install torch --index-url https://download.pytorch.org/whl/cu124
pip install -U triton-windows
pip install transformers==4.46.3
pip install safetensors==0.4.5
pip install sentencepiece
```
3. Convert to BF16, requires an additional approximately 1.3 TB of space.  
Here, you need to download the transformation code from the "inference" folder of [deepseek-ai/DeepSeek-V3](https://huggingface.co/deepseek-ai/DeepSeek-V3)
```
cd deepseek-ai/DeepSeek-V3/inference
python fp8_cast_bf16.py --input-fp8-hf-path C:/Users/admin/deepseek-ai/models/DeepSeek-V3-0324/ --output-bf16-hf-path C:/Users/admin/models/deepseek-ai/DeepSeek-V3-0324-bf16
```
## BF16 to f16.gguf
1. Use the [llama.cpp](https://github.com/ggerganov/llama.cpp) conversion program to convert DeepSeek-V3-0324-bf16 to gguf format, requires an additional approximately 1.3 TB of space.
```
python convert_hf_to_gguf.py C:/Users/admin/deepseek-ai/models/deepseek-ai/DeepSeek-V3-0324-bf16 --outfile C:/Users/admin/deepseek-ai/models/deepseek-ai/DeepSeek-V3-0324-bf16/ggml-model-f16.gguf --outtype f16
```
2. Use the [llama.cpp](https://github.com/ggerganov/llama.cpp) quantitative program to quantitative model (llama-quantize needs to be compiled.),
other [quant option](https://github.com/ggerganov/llama.cpp/blob/master/examples/quantize/quantize.cpp). 
Convert first Q2_K, requires an additional approximately 227 GB of space.
```
llama-quantize C:/Users/admin/deepseek-ai/models/deepseek-ai/DeepSeek-V3-0324-bf16/ggml-model-f16.gguf  C:/Users/admin/deepseek-ai/models/deepseek-ai/DeepSeek-V3-0324-bf16/ggml-model-Q2_K.gguf Q2_K
```
3. Use llama-cli to test.
```
llama-cli -m C:/Users/admin/deepseek-ai/models/deepseek-ai/DeepSeek-V3-0324-bf16/ggml-model-Q2_K.gguf -n 2048
```

## Use with ollama
**Note:** this model requires [Ollama 0.5.5](https://github.com/ollama/ollama/releases/tag/v0.5.5)  

### Modefile
```
FROM deepseek-ai/DeepSeek-V3-0324-bf16/ggml-model-Q2_K.gguf
TEMPLATE """{{- range $i, $_ := .Messages }}
{{- if eq .Role "user" }}<｜User｜>
{{- else if eq .Role "assistant" }}<｜Assistant｜>
{{- end }}{{ .Content }}
{{- if eq (len (slice $.Messages $i)) 1 }}
{{- if eq .Role "user" }}<｜Assistant｜>
{{- end }}
{{- else if eq .Role "assistant" }}<｜end▁of▁sentence｜><｜begin▁of▁sentence｜>
{{- end }}
{{- end }}"""
PARAMETER stop <｜begin▁of▁sentence｜>
PARAMETER stop <｜end▁of▁sentence｜>
PARAMETER stop <｜User｜>
PARAMETER stop <｜Assistant｜>
PARAMETER num_gpu 1
```

### Donation

If you like it, please click 'like' and follow us for more updates.  
You can follow [x.com/support_huihui](https://x.com/support_huihui) to get the latest model information from huihui.ai.

##### Your donation helps us continue our further development and improvement, a cup of coffee can do it.
- bitcoin:
```
  bc1qqnkhuchxw0zqjh2ku3lu4hq45hc6gy84uk70ge
```