[usr-bin-roygbiv]

Motivation

Fixes #36599 and addresses the draft-load failure described in #36653. DeepseekModelNextN and BailingMoEModelNextN currently force modelopt_fp4 quantization to None. That makes quantized GLM-5.3-Flash NextN/MTP weights load into BF16 modules and fail with a packed-versus-unpacked shape mismatch.

Modifications

• Remove the blanket modelopt_fp4 to None override from the DeepSeek and Bailing NextN paths while preserving the legacy GLM-4.x and GLM-OCR overrides.
• Resolve ModelOpt FP4 MTP exclusions after loader name mapping, including raw and already-remapped wildcard/exact forms, the canonical legacy layer-0 layout, coarse FusedMoE prefixes, and Bailing hybrid runtime prefixes.
• Retain ModelOpt FP4 for checkpoints with quantized MTP weights and exclude only explicitly unquantized MTP modules. This is the supported checkpoint contract.
• Copy quantization configs before changing per-layer exclusions and normalize optional packed-module mappings.
• Add coverage for mapper-before-resolver ordering, production wildcard spelling, exact parent-module matching, ModelOpt consumer behavior, Bailing hybrid prefixes, legacy layer-0 layouts, and configuration immutability.

Accuracy Tests

CUDA_HOME=/usr PYTHONPATH=python:. uv run --with-editable ./python/ --with pytest pytest test/registered/unit/models/test_nextn_modelopt_quant.py test/registered/unit/layers/quantization/test_modelopt_nvfp4.py test/registered/unit/layers/quantization/test_modelopt_nvfp4_moe_scales.py -v

Result: 75 passed, 15 warnings, 30 subtests passed in 13.77s.

The unit coverage exercises loader mapper ordering and the downstream ModelOpt get_quant_method decisions. No full ModelOpt FP4 checkpoint load, speculative-decoding run, or model-output accuracy run was performed locally.

Speed Tests and Profiling

This changes NextN module construction and checkpoint weight-loading precision selection; it does not add a kernel benchmark. No throughput benchmark or profiler run was performed.

Relationship to Other Work

• [Fix] NextN: derive the quant-config rename mapper from the checkpoint's MTP layer index #34906 derives the NextN rename mapper from checkpoint metadata and removes the static mapper used by this branch. If [Fix] NextN: derive the quant-config rename mapper from the checkpoint's MTP layer index #34906 lands first, rebase and switch the Quark path to its mapper API.
• [Fix] [AMD] GLM5: Fix MTP layer_quant_config in-place mutation and add nextn Quark-exclude unit test #30565 changes the same NextN resolver regions and overlaps on copy-before-mutate behavior. Rebase and resolve the overlap after either change lands.
• fix(modelopt): expand is_layer_excluded for fused and model.-prefixed names #36761 changes ModelOpt exclusion matching in the quantization layer. Coordinate its matcher semantics with this model-level mapping.
• [Bug] ModelOpt FP4: is_layer_excluded misses fused module names and model.-prefixed names — mixed-precision NVFP4 checkpoints crash at load (GLM-5.3-Flash) #36596, [Bug] NVFP4 MoE + EP>1: globally-loaded input scales not sliced to local experts in non-CuteDSL branch (_compute_gemm1_alphas shape mismatch) #36597, and [Bug] Disabling DSA via index_topk override with index_share_for_mtp_iteration=true asserts in scheduler (get_dsa_index_topk) #36598 are related GLM-5.3 bring-up issues but are not implemented by this PR.
• Full GLM-5.3-Flash NVFP4 NextN validation also depends on the vision/model support in GLM-5.3-Flash support #36507.

The branch starts from an older main ancestor (88cf5c9541); the intervening upstream commits do not touch this PR's changed paths. The branch should be rebased after the overlapping PRs above land.

Checklist

• Format code according to the pre-commit format guidance.
• Add unit tests according to the unit-test guidance.
• Update documentation according to the documentation guidance; no user-facing documentation change is needed for this internal loader behavior.
• Provide full checkpoint accuracy and speed benchmark results; no full checkpoint or benchmark run was performed locally.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #33458470679
Latest PR Test (Extra): ❌ Run #33458470537
Latest PR Test (AMD ROCm 7.2): ❌ Run #33458470657

[usr-bin-roygbiv]

/tag-and-rerun-ci

[usr-bin-roygbiv]

Hi @HaiShaw — could you add the run-ci label and rerun CI on this PR? I tried /tag-and-rerun-ci, but the author account can't add labels, so the gate stops before substantive tests run. Local unit tests and lint are passing. Thanks!