Instructions to use AIDC-AI/Marco-Mini-Instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AIDC-AI/Marco-Mini-Instruct with Transformers:

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

pipe = pipeline("text-generation", model="AIDC-AI/Marco-Mini-Instruct")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("AIDC-AI/Marco-Mini-Instruct")
model = AutoModelForCausalLM.from_pretrained("AIDC-AI/Marco-Mini-Instruct")
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 AIDC-AI/Marco-Mini-Instruct with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/AIDC-AI/Marco-Mini-Instruct

SGLang

How to use AIDC-AI/Marco-Mini-Instruct 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 "AIDC-AI/Marco-Mini-Instruct" \
    --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": "AIDC-AI/Marco-Mini-Instruct",
		"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 "AIDC-AI/Marco-Mini-Instruct" \
        --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": "AIDC-AI/Marco-Mini-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use AIDC-AI/Marco-Mini-Instruct with Docker Model Runner:
```
docker model run hf.co/AIDC-AI/Marco-Mini-Instruct
```

How to prevent repetitions?

by BingoBird - opened 1 day ago

Discussion

BingoBird

1 day ago

On llama.cpp (current build)

I tried q4_K_M
llama-server -m ./Marco-Mini-Instruct.Q4_K_M.gguf --no-webui --ctx-checkpoints 4 -b 512 -ub 1024 --ctx-size 16834 --cache-type-k q8_0 --cache-type-v q8_0 -t 4 --parallel 1 --spec-type ngram-mod --spec-ngram-mod-n-match 40 --spec-ngram-mod-n-min 0 --spec-ngram-mod-n-max 16 -fa on --no-mmproj --fit on --no-warmup

and q6_K
/usr/local/bin/llama-server -m /media/sda3/Models/Marco-Nano-Instruct.i1-Q6_K.gguf --no-webui --ctx-checkpoints 4 -b 512 -ub 1024 --ctx-size 16834 --cache-type-k q8_0 --cache-type-v q8_0 -t 4 --parallel 1 --spec-type ngram-mod --spec-ngram-mod-n-match 40 --spec-ngram-mod-n-min 0 --spec-ngram-mod-n-max 16 -fa on --no-mmproj --fit on --no-warmup

Very often i get repetition loops.

    # Use a more robust method: use a color histogram and threshold
    # For simplicity, use a threshold on the L channel
    # But we need to use a color space that is perceptually uniform
    # Use a more robust method: use a color histogram and threshold
    # For simplicity, use a threshold on the L channel
    # But we need to use a color space that is perceptually uniform
    # Use a more robust method: use a color histogram and threshold
    # For simplicity, use a threshold on the L channel

I use these settings:

parameters="$(jq -nR '{
        temperature: 1,
        top_p: 1,
        top_k: 1,
        min_p: 0.0,
        repeat_penalty: 1.0,
        presence_penalty: 0.5,
        jinja: true,
        flash_attn: true,
        cont_batching: true,
        repeat_last_n: 0,
        penalize_nl: false,
        n_predict: -1,
        stream: true,
        cache_prompt: true,
        thinking_budget_tokens: 240,
}')"

Is there a way to avoid the loops?

Thanks for the interesting model.

BingoBird

1 day ago

•

edited 1 day ago

I am now testing with fewer parameters and seem to have much better success:

parameters="$(jq -nR '{
        jinja: true,
        flash_attn: true,
        cont_batching: true,
        repeat_last_n: 0,
        penalize_nl: false,
        n_predict: -1,
        stream: true,
        cache_prompt: true,
        thinking_budget_tokens: 240,
}')"

Performance on ryzen 3500u laptop (+ vulkan on Vega8) is approxL: pp:44 tg: 13.5
It is one of the smartest 10+t/s models to have run on this laptop. Congratulations.

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