Instructions to use XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash with Transformers:

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

pipe = pipeline("text-generation", model="XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash

SGLang

How to use XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash 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 "XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash" \
    --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": "XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash",
		"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 "XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash" \
        --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": "XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash with Docker Model Runner:
```
docker model run hf.co/XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash
```

What attention scheme for DFlash SWA? Non-Causal?

by benchislett - opened Jun 22

Discussion

benchislett

Jun 22

Hi, I'm the vllm maintainer implementing this in vLLM: https://github.com/vllm-project/vllm/pull/46104

Could you kindly disclose what attention mechanism should be used for the DFlash Sliding Window layers? Specifically: does it use non-causal attention (as is common in DFlash), or does it use causal-only attention? This changes how the attention pattern looks in the query block, and changes the kernel requirements (causal SWA has much better software support).

If non-causal SWA is used, is the window bi-directional? e.g. if a prefix token attends to some query token, does it always attend to the entire query? Or, could it attend to only a subset of the query tokens (those within the sliding window)?

benchislett

Jun 22

If there is a specforge recipe that I could use as a reference, that would be helpful to analyze the pattern.

csAugust

Jun 25

Yes, we use non-causal SWA attention for MiMo-v2.5 Pro DFlash, and the window is bi-directional. Prefix tokens are also limited in the sliding window. Currently we do not have a specforge-compatitble training codebase, but you may could refer to this: https://github.com/sgl-project/sglang/pull/27638

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