Update Mellum2 Instruct serve command

[haic0]

Summary

• Add trust-remote-code to the Mellum2 Instruct recipe and show the TP=1 launch command.
• Aligns the recipe launch guidance with the provided vLLM serve command.

Test plan

• Ran node scripts/build-recipes-api.mjs on the complete validated recipe update set.

Made with Cursor

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview, Comment	Jun 29, 2026 1:44pm

[gemini-code-assist]

Code Review

This pull request updates the configuration for the Mellum2-12B-A2.5B-Instruct model, adding hardware overrides and replacing the --max-model-len argument with --trust-remote-code in both the base arguments and the guide commands. The reviewer feedback correctly points out that --max-model-len should be retained alongside --trust-remote-code to prevent potential out-of-memory errors during KV cache allocation due to the model's large context length.

[gemini-code-assist]

It looks like --max-model-len was accidentally replaced by --trust-remote-code instead of appending it. For models with large context lengths like Mellum2 (131,072), explicitly setting --max-model-len is important to prevent out-of-memory errors during KV cache allocation in vLLM. We should keep both arguments.

    - "--max-model-len"
    - "131072"
    - "--trust-remote-code"

[gemini-code-assist]

To align with the base_args and ensure the launch command is robust, we should retain the --max-model-len 131072 parameter in the guide's serving commands.

    --max-model-len 131072 \
    --trust-remote-code \
    --tensor-parallel-size 1

[gemini-code-assist]

Similarly, we should retain --max-model-len 131072 in the tool-calling launch command example.

    --max-model-len 131072 \
    --trust-remote-code \
    --tensor-parallel-size 1 \