[AMD] GLM 5.2 MXFP4 SGLANG COOKBOOK

[ajith-sirra-amd]

Motivation

AMD publishes a Quark-quantized MXFP4 build of GLM-5.2 for MI355X (amd/GLM-5.2-MXFP4), alongside the existing FP8/BF16 (all AMD SKUs) and NVFP4 (NVIDIA Blackwell) options already documented in the GLM-5.2 cookbook. This PR adds MXFP4 as a selectable quantization in the GLM-5.2 cookbook so users deploying on MI355X (gfx950) have a documented, ready-to-run recipe, consistent with the precedent already set for amd/GLM-5.1-MXFP4.

Modifications

• docs/src/snippets/configs/zai-org/glm-5.2.jsx:
  • Added mxfp4 to the quantizations list and mapped default|mxfp4 → amd/GLM-5.2-MXFP4 in modelNames.
  • Pinned a dedicated, newer Docker image for the MI355X + MXFP4 combination (dockerImages["mi355x|mxfp4"] = lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260728), separate from the existing MI355X FP8/BF16 image.
  • Added three new deployment cells for hw: "mi355x", quant: "mxfp4" (low-latency, balanced, high-throughput), mirroring the FP8/BF16 MI355X recipes but at tp=4 (4-bit MoE weights fit a 4-GPU slice), with --trust-remote-code (required for Quark's custom quant config) and --kv-cache-dtype fp8_e4m3, carried over from the validated amd/GLM-5.1-MXFP4 MI355X recipe (same DSA architecture family).
  • All three new cells are marked verified: false — not yet benchmarked on GLM-5.2. MXFP4 is gfx950-only and not exposed for MI300X/MI325X (gfx942).
• docs/cookbook/autoregressive/GLM/GLM-5.2.mdx:
  • Updated the model introduction to mention the AMD MXFP4 build and its unverified status.
  • Added an amd/GLM-5.2-MXFP4 row to the model variants table and a corresponding resource link.
  • Added a new "MI355X MXFP4 (gfx950-only)" configuration tip explaining the tp=4, --trust-remote-code, and --kv-cache-dtype fp8_e4m3 requirements and the unverified status.

No runtime/engine code is touched — this is a docs/cookbook-only change (config + Mintlify docs).

Accuracy Tests

Speed Tests and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests. (N/A — docs/cookbook config only)
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed. (pending — recipe currently marked verified: false)
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): ✅ Run #31487083615
Latest PR Test (Extra): ❌ Run #31487083604

[ajith-sirra-amd]

/tag-and-rerun-ci extra

[seungrokj]

@sogalin Hi Soga, can you plz review this We need this merged to push SGL perf data to SA.

[functionstackx]

+viz @HaiShaw

[1am9trash]

LGTM