Add text-only Llama 4 support to mini-sglang.

[sheepfish5]

Summary

This PR adds support for Llama 4 to mini-sglang.

Llama 4 currently has two public model variants:

• Scout: about 109B total parameters
• Maverick: about 400B total parameters

This PR implements the text-only Llama 4 model path.
Although Llama 4 natively supports image input, multimodal/image support is not included in this PR.

What’s included

• Added Llama 4 model support in mini-sglang
• Implemented the text model path for Llama 4

Tested model

I tested this implementation with the Hugging Face model:

• meta-llama/Llama-4-Scout-17B-16E-Instruct

Runtime / memory usage

Test environment:

• 4 x 96GB GPUs

Observed memory usage:

• about 58GB VRAM per GPU on average

Run command:

python3 -m minisgl --model-path "./models" --shell --attention-backend fi --tp 4

Notes

• This PR is for text inference only
• Image / multimodal support for Llama 4 is not implemented here