[Raiden-Makoto]

Motivation

Follow-up to #30265 / #29781 (supersedes #30379). #30265 merged with two gaps flagged in its own review:

1. The GLM-5.2 Quark-MXFP4 MTP path has no PR-CI coverage — the only AMD GLM-MXFP4 test is nightly=True and does not enable MTP; the GLM MTP e2e test is CUDA-only + FP8, not Quark MXFP4.
2. GlmMoeDsaForCausalLMNextN._resolve_nextn_quant_config remaps layer_quant_config keys by mutating the shared quant_config in place, which can corrupt the main model's per-layer scheme selection (gemini-code-assist review on [AMD] Fix GLM-5.2 MTP Quark excludes #30265).

The fix and the CPU test here originate from #30379 (@fanxingran, co-authored).

Modifications

• python/sglang/srt/models/glm4_moe.py: _resolve_nextn_quant_config copy.copys quant_config (and its nested dict) before any remap, and only copies when a remap/exclude is actually needed — no shared-state mutation.
• python/sglang/srt/models/deepseek_nextn.py: docstring on the base _resolve_nextn_quant_config documenting the subclass override contract.
• test/registered/unit/models/test_nextn_quark_exclude.py (new, CPU CI base-a-test-cpu): base DeepseekV3ForCausalLMNextN passthrough (non-quark / None), GLM-5.2 MTP exclude_layers + layer_quant_config remap, coarse fused-expert prefix, and the no-mutation guarantee.
• test/registered/amd/accuracy/mi35x/test_glm52_mxfp4_mtp_eval_mi35x.py (new, nightly nightly-amd-8-gpu-mi35x-glm52-mxfp4-mtp): 8-GPU GLM-5.2-MXFP4 EAGLE-MTP GSM8K eval — GPU coverage that the bf16 nextn layer loads under Quark MXFP4 and MTP accepts drafts. Wired into nightly-test-amd.yml and nightly-test-amd-rocm720.yml.

Accuracy Tests

CPU unit test test/registered/unit/models/test_nextn_quark_exclude.py on this branch:

6 passed, 5 warnings in 10.64s

test_layer_quant_config_keys_remapped_without_mutation fails on current main (exercises the #30265 in-place mutation); green after this change. pre-commit run passes all hooks (isort/ruff/black/yaml/duplicate-job-name/registry validation).

Nightly GPU eval thresholds are the reported amd/GLM-5.2-MXFP4 MI355x numbers: GSM8K accuracy floor 0.935 (#29781 measured 0.950 at EAGLE 5/1/6, the fixture's config) and accept-length floor 5.0 (#30265 measured 5.94).

Speed Benchmarks

N/A — quant-config resolution runs once at model construction; no inference-path change.

Checklist

• CPU unit test added and registered (base-a-test-cpu)
• Nightly GPU test added and wired into the AMD nightly workflows
• Co-authored with the original author ([AMD] Fix Quark MXFP4 MTP load when nextn layer is bf16 (GLM-5.2): de… #30379)

---

CI States

Latest PR Test (Base): 🚫 Run #33544291185
Latest PR Test (Extra): ❌ Run #33544290924
Latest PR Test (AMD ROCm 7.2): ❌ Run #33544291139

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[HaiShaw]

/tag-and-rerun-ci

[Raiden-Makoto]

@amd-bot ci-status

[Raiden-Makoto]

CI tests failed due to rate limit.

[Raiden-Makoto]

/rerun-failed-ci

[Raiden-Makoto]

/rerun-failed-ci

[Raiden-Makoto]

/rerun-failed-ci