[NV] Add MiniMax M3 Blackwell marlin TP8 recipe

[jasonlizhengjian]

Summary

Adds --moe-backend marlin only to the MiniMax-M3 MXFP8 Blackwell TP8 recipe path.

Changes

• Adds variant-scoped strategy/hardware override support in command synthesis.
• Moves the Blackwell single_node_tp marlin override under the mxfp8 variant so the BF16 default path is unaffected.
• Updates only the MiniMax-M3 MXFP8 TP8 guide command to include --moe-backend marlin.

Validation

• git diff --check
• node scripts/build-recipes-api.mjs
• Confirmed generated BF16 B200 single_node_tp commands do not include --moe-backend marlin.
• Confirmed generated MXFP8 B200 single_node_tp commands include --moe-backend marlin.
• Confirmed generated MXFP8 H200, MXFP8 MI355X, and B200 single_node_tep commands do not include --moe-backend marlin.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview, Comment	Jun 16, 2026 8:08pm

[gemini-code-assist]

Code Review

This pull request introduces strategy overrides and guide documentation for running the MiniMax-M3 model on NVIDIA Blackwell hardware using the Marlin MoE backend. However, the reviewer identified critical issues where the Marlin MoE backend is applied to the unquantized BF16 model (both in the global strategy overrides and the guide commands), which will cause runtime failures since Marlin MoE kernels do not support unquantized BF16 weights.

[gemini-code-assist]

Applying --moe-backend marlin globally under strategy_overrides for Blackwell will affect all variants, including the default BF16 model (MiniMaxAI/MiniMax-M3). Since Marlin MoE kernels in vLLM do not support unquantized BF16 weights, running the default BF16 recipe on Blackwell will fail at runtime.

If the schema supports it, please move this override under the mxfp8 variant configuration, or remove it from the global strategy_overrides to avoid breaking the BF16 path.

[gemini-code-assist]

This guide section attempts to run the BF16 model (MiniMaxAI/MiniMax-M3) with --moe-backend marlin. Since Marlin MoE kernels do not support unquantized BF16 weights, this command will fail at runtime.

Since the BF16 model on Blackwell runs with the same parameters as Hopper, this section is redundant. Please remove this section or merge it with the Hopper section above.