fix(modelopt): expand exclusions nested under model.language_model.*

[divyvasal]

Motivation

Follow-up to #36596. #36761 fixes the fused-module and model.-prefixed vision-tower halves; this fixes the third naming form, which neither the shipped matcher nor that PR handles: exclusions nested under model.language_model.*.

apply_weight_name_mapper strips language_model. only when it is the leading segment. ModelOpt writes some vision-language checkpoints with the decoder nested one level in — model.language_model.layers.N.* — while sglang builds it as model.layers.N.*. The strip never fires, every layer the checkpoint keeps in BF16 is built as an FP4 module, and the load dies on a packed-versus-unpacked shape mismatch.

Hit on nvidia/GLM-5.3-Flash-NVFP4 (the official NVIDIA checkpoint; #36596 was filed against LibertAIDAI's, which is spelled differently) at TP8 on 8xB200, FlashInfer 0.6.17:

AssertionError: target.shape=torch.Size([4096, 256]),
                loaded_weight.shape=torch.Size([4096, 512])

the FP4-packed parameter against the BF16 tensor the checkpoint actually ships.

Its exclusion list uses this form throughout — model.language_model.layers.N.self_attn*, ...mlp.gate, ...mlp.shared_experts* — so on this checkpoint the two PRs are both required and neither is sufficient. Measured against that exclusion list with the matcher's shipped behaviour, #36761's logic, and this change:

built prefix	shipped	#36761	this PR
visual.blocks.10.attn.proj	miss	fixed	miss
model.layers.0.self_attn.kv_b_proj	miss	miss	fixed
model.layers.10.mlp.gate	miss	miss	fixed
model.layers.10.mlp.shared_experts.gate_up_proj	miss	miss	fixed

Modifications

python/sglang/srt/layers/quantization/modelopt_quant.py: in apply_weight_name_mapper, expand model.language_model.<rest> to model.<rest> alongside the existing leading-segment strip. Same place, same shape, and the expansion is exact rather than a wildcard, so it cannot widen a match.

Accuracy

No numerical change. This only affects which modules are built quantized; it restores the checkpoint's own intent, which is that these layers stay BF16.

Checklist

• Unit tests in test/registered/unit/layers/quantization/test_modelopt_nvfp4.py: the interior form matches, the checkpoint's own spelling still matches, a neighbouring layer index does not match, the leading strip is unchanged, and names without the segment are left untouched.
• Verified against the unmodified hf_quant_config.json from nvidia/GLM-5.3-Flash-NVFP4, in the sglang image, on 8xB200.
• With this and fix(modelopt): expand is_layer_excluded for fused and model.-prefixed names #36761 applied, the checkpoint loads and serves.

---

CI States

Latest PR Test (Base): ❌ Run #34433045876
Latest PR Test (Extra): ❌ Run #34433045790
Latest PR Test (AMD ROCm 10): ❌ Run #34433045968

[divyvasal]

Closing as redundant — superseded by #38621, which merged on 2026-09-09 and which I missed when I opened this.

#38621 handles the same model.language_model.* case, and the vision-tower rename from #36596 as well, in a better place: a model-scoped WeightsMapper on Glm5NextForConditionalGeneration rather than a special case in the generic matcher.

hf_to_sglang_mapper = WeightsMapper(orig_to_new_substr={
    "model.language_model.": "model.", "model.visual": "visual"})

Confirmed on our side: with #38621 applied to the 2026-09-03 lmsysorg/sglang:glm-5.3-flash image, nvidia/GLM-5.3-Flash-NVFP4 loads and serves on 8×B300 with an unmodified hf_quant_config.json (132 exclusion entries, no hand-patching), and --moe-runner-backend flashinfer_cutlass works — both of which previously failed for us.

Thanks to the author of #38621. Sorry for the duplicate noise.