fix(image): patch sglang for GLM-5.2 MTP nextn quark-exclude (backport #30265)

[jiejingzhangamd]

Problem

GLM-5.2 EAGLE/MTP crashes at draft weight-load on our sglang v0.5.15.post1 base:

AssertionError: param.shape=[6144,6144] uint8 vs loaded_weight.shape=[6144,12288] bf16
  at linear.py weight_loader (via models/deepseek_nextn.py)

GLM-5.2's MTP layer (index = num_hidden_layers) is entirely bf16/unquantized — the quark quantization_config lists model.layers.<N>.eh_proj (and the layer's other submodules) in exclude. sglang's DeepseekV3ForCausalLMNextN disables nextn quant only when the bare layer prefix model.layers.<N> is in exclude_layers, but the exclude entries are submodule-level (...<N>.eh_proj), so should_ignore_layer() returns False, eh_proj is built as an MXFP4 (uint8) param, and the bf16 checkpoint weight won't load.

Fix

Upstream fixed this in sgl-project/sglang#30265 (merged 2026-07-08, a dedicated GlmMoeDsaForCausalLMNextN class), but our pinned base v0.5.15.post1 predates it. This adds a temporary, self-locating, idempotent Python patch that probes the eh_proj submodule (an exact exclude entry) instead of the bare layer → nextn_quant_config=None → the whole bf16 MTP layer builds bf16.

• deploy/docker/patches/sglang/patch_glm52_nextn_quark_exclude.py — the patch (no-op if the anchor is absent, i.e. once sglang carries #30265).
• Dockerfile.sglang — a patch loop after pip install (mirrors the vLLM patch loop).

Verification

Coherent GLM-5.2 MTP output on the v0.5.15.post1 base (single-stream accept len ~1.85–2.6, ~1.3–2× decode speedup), and the fix holds across single-node + 1P1D PD.

Removal

Temporary — drop deploy/docker/patches/sglang/ and the loop when we bump the base sglang to a build that includes #30265 (the patch already no-ops then). Tracking the broader GLM-5.x ROCm EAGLE enablement upstream in #32175 (still open).

🤖 Generated with Claude Code