Blog post: Granite Switch in Mellea

[planetf1]

A blog post is in flight that introduces Granite Switch as a delivery mechanism for Mellea intrinsics — single vLLM-served checkpoint, no adapter weights to wrangle. Draft to land as a PR against content/blogs/granite-switch.md. Tagging here for visibility on the public side.

[planetf1]

Note for blog post author: the draft should include (or link to) vLLM setup instructions for granite-switch. Currently the mellea docs assume vLLM is already running and say nothing about the granite-switch package that must be installed in the vLLM environment.

Without it, vLLM crashes with:

ValueError: Model architectures ['GraniteSwitchForCausalLM'] are not supported for now.

The install step that needs documenting:

# vLLM 0.20+ (CUDA 13+)
pip install "granite-switch[vllm20]"

# vLLM 0.19.x (CUDA 12.x)
pip install "granite-switch[vllm]"

Once installed, no special vLLM flags are needed — adapters activate via control tokens in the chat template. This gap should be covered in the blog post or in a linked setup guide. Separately, a docs issue should be filed in the main mellea repo to add a "Setting up a granite-switch vLLM server" section to the server setup docs.