Instructions to use dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4 with Transformers:

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

pipe = pipeline("text-generation", model="dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4", 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("dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4", trust_remote_code=True, device_map="auto")
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 Settings

vLLM

How to use dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4

SGLang

How to use dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4 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 "dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4" \
    --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": "dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4",
		"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 "dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4" \
        --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": "dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4 with Docker Model Runner:
```
docker model run hf.co/dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4
```

Issues for GB10 users

by scottgl - opened Apr 16

Discussion

scottgl

Apr 16

Primary issue: Quantization format incompatible with GB10 (SM12.1)

The model uses W4A16 NVFP4 (weights in FP4, activations in FP16) — the Blackwell B100/B200 native microscaling format. This is great for B100/B200 but doesn't work on GB10:

GB10 (SM12.1) supports NVFP4 through the W4A4 path (both weights AND activations FP4), which routes through the CompressedTensorsW4A4Fp4 kernel with Marlin fallback. The W4A16 format falls between both supported kernel paths:

CompressedTensorsW4A4Fp4 — rejects it (expects non-null activations)
CompressedTensorsWNA16 / Marlin — rejects it (NVFP4 microscaling packing incompatible with Marlin's integer pack layout)
The model card mentions "functional via upcasting on Hopper/Ampere" but that upcasting path isn't implemented for GB10 in SGLang.

Secondary issue: Even if loaded, won't fit in 128GB

The packed FP4 weights are ~75GB. Dequanting to BF16 at load time = ~302GB. FP8 = ~150GB. Both exceed the 128GB unified memory limit. The whole point of FP4 is fitting in 128GB, but the W4A4 format is what actually achieves that on GB10.

What would work for GB10:
Re-quantize the BF16 source (dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B) using LLM Compressor with the saricles W4A4 recipe (same one used for saricles/MiniMax-M2.5-REAP-172B-A10B-NVFP4-GB10). Output would be ~80GB W4A4 NVFP4 — fits GB10 comfortably with ~48GB KV headroom.

dervig

Owner Apr 16

Hi! Producing the W4A4 variant now via llmcompressor (same recipe as the saricles GB10 model). ETA within a few hours, will land at dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4-GB10.

dervig

Owner Apr 16

https://huggingface.co/dervig/m51Lab-MiniMax-M2.7-REAP-139B-A10B-NVFP4-GB10

dervig changed discussion status to closed Apr 16

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