Instructions to use meituan/DeepSeek-R1-Block-INT8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use meituan/DeepSeek-R1-Block-INT8 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="meituan/DeepSeek-R1-Block-INT8", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("meituan/DeepSeek-R1-Block-INT8", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("meituan/DeepSeek-R1-Block-INT8", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use meituan/DeepSeek-R1-Block-INT8 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "meituan/DeepSeek-R1-Block-INT8"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "meituan/DeepSeek-R1-Block-INT8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/meituan/DeepSeek-R1-Block-INT8

SGLang

How to use meituan/DeepSeek-R1-Block-INT8 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "meituan/DeepSeek-R1-Block-INT8" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "meituan/DeepSeek-R1-Block-INT8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "meituan/DeepSeek-R1-Block-INT8" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "meituan/DeepSeek-R1-Block-INT8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use meituan/DeepSeek-R1-Block-INT8 with Docker Model Runner:
```
docker model run hf.co/meituan/DeepSeek-R1-Block-INT8
```

Optimal `weight_block_size` for Intel AMX `amx_int8` `amx_tile`?

#17

by ubergarm - opened Mar 9, 2025

Discussion

ubergarm

Mar 9, 2025

在使用双路Intel Xeon NUMA节点的场景中，针对Intel(R) Xeon(R) 6980P平台，参考了关于双路Intel Xeon NUMA节点的讨论的建议，尝试采用int8量化以利用amx_tile和amx_int8 CPU指令集标志。

然而，当前配置中默认的weight_block_size设置为128x128。该参数值是否同时适用于GPU和CPU推理场景？

根据《Intel® 架构指令集扩展编程参考》第40页的说明，硬件对矩阵运算的维度限制如下：

位域07-00: tmul_maxk (行/列数最大值). 数值 = 16
位域23-08: tmul_maxn (列字节数最大值). 数值 = 64

实际限制可能取决于推理引擎（如ktransformers、llama.cpp等）是否会对模型权重矩阵进行分块处理。虽然可以通过实验验证，但若存在以下问题的明确结论将非常有帮助：

若已知Intel AMX CPU在推理过程中需要更小的权重分块尺寸（例如16x16），请提供指导建议。

祝好，感谢！

I have access to a dual socket Intel(R) Xeon(R) 6980P and received advice to try this int8 quant to take advantage of amx_tile and amx_int8 CPU flags from this post on Dual Intel Xeon NUMA Nodes.

However, I see the default weight_block_size is set to 128x128 in the configuration. Is this value appropriate for both GPU and CPU inferencing?

Refer to page 40 of Intel® Architecture Instruction Set Extensions Programming Reference suggests the tile max rows is 16 and tile max columns is 64 bytes:

Bits 07-00: tmul_maxk (rows or columns). Value = 16.
Bits 23-08: tmul_maxn (column bytes). Value = 64

It may depend on the inference engine e.g. ktransformers, llama.cpp, etc. if it splits the model weight tiles into smaller thread tiles etc.

I may just have to try it and see, but if anyone knows if Intel AMX CPU inference requires a smaller weight_block_size e.g. 16x16 please advise.

Cheers and thanks!

ubergarm

Mar 9, 2025

https://github.com/ikawrakow/ik_llama.cpp?tab=readme-ov-file#to-tile-or-not-to-tile

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment