Add GLM-5.2 MXFP4 recipe support

[xiao-llm]

Summary

• Add GLM-5.2 MXFP4 as a recipe variant alongside the existing FP8 and BF16 options.
• Configure the MXFP4 variant for the AMD Quark checkpoint with --quantization quark.
• Restrict the MXFP4 variant to supported MI355X hardware, following the MiniMax-M3 MXFP4 recipe pattern.
• Set the MXFP4 generated command to the validated TP8 path and include the MI355X-specific --trust-remote-code override.
• Add ROCm/AITER serving guidance for GLM-5.2 MXFP4, including recommended MI355X launch flags and KV-cache settings.
• Document MTP usage caveats for MXFP4, including the need for a vLLM build that honors Quark exclude entries for the unquantized MTP layer.

PR Link for mxfp4 MTP support

Motivation

GLM-5.2 has an AMD Quark MXFP4 checkpoint that significantly reduces HBM footprint compared with the FP8 and BF16 variants. This recipe update makes that checkpoint selectable in the recipe UI while preventing unsupported hardware combinations from being generated.

The MTP note is tied to the corresponding vLLM Quark loading fix: vLLM PR #46757.

Test Plan

• Verified the GLM-5.2 recipe YAML parses successfully.
• Compared the MXFP4 hardware gating pattern with models/MiniMaxAI/MiniMax-M3.yaml.
• Confirmed the generated diff updates the GLM-5.2 recipe metadata, variant config, and guide text only.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview, Comment	Jun 26, 2026 6:30pm

[gemini-code-assist]

Code Review

This pull request introduces support for the AMD Quark MXFP4 quantized variant of the GLM-5.2 model, optimized for AMD MI350 and MI355X GPUs. It adds the mxfp4 variant configuration to the model YAML and updates the documentation with prerequisites, environment variables, and run commands for utilizing ROCm AITER kernels. The feedback suggests updating the mxfp4 variant metadata to include --trust-remote-code in extra_args and the required environment variables in extra_env to ensure consistency with the guide and support automated deployment tools.

[gemini-code-assist]

To ensure consistency with the launch instructions in the guide and to support automated deployment tools that parse this recipe, the mxfp4 variant metadata should include --trust-remote-code in extra_args and the required environment variables (VLLM_ROCM_USE_AITER_FP8BMM: "0" and VLLM_ROCM_USE_AITER_FP4BMM: "0") in extra_env.

  mxfp4:
    model_id: "amd/GLM-5.2-MXFP4"
    precision: mxfp4
    vram_minimum_gb: 446
    description: "AMD Quark MXFP4 checkpoint — MoE weights quantized for MI350 / MI355X"
    extra_args:
      - "--quantization"
      - "quark"
      - "--trust-remote-code"
    extra_env:
      VLLM_ROCM_USE_AITER_FP8BMM: "0"
      VLLM_ROCM_USE_AITER_FP4BMM: "0"

[xiao-llm]

Setting them to 0 keeps serving on the known-good kernels and avoids accidental dispatch into experimental/unsupported FP8/FP4 BMM implementations.

[BowenBao]

related to gemini's comment, are these needed?

[xiao-llm]

Will remove it after verified, its a extra condition to keep it safe.