Fix disagg PP MTP for GLM-5.2

[nvpohanh]

Supersedes #39052 while @nvjullin is OOTO. The original two commits are preserved with Julien as their author; this branch rebases the change onto current main and addresses every review comment on the original PR.

Motivation

GLM-5.2 could not run disaggregated PP prefill with MTP because:

1. PP + speculative decoding was rejected outside NPU even though the prefill path already forwards PP proxy tensors from stages that do not host the draft.
2. The NextN draft is hosted on the last PP stage and borrows the target embedding, but only the first PP stage materialized and loaded embed_tokens.
3. The PP ring dropped the draft model's DSA IndexShare seed (dsa_topk_indices), which could make decode ranks diverge between eager and CUDA-graph paths.
4. NIXL needed to map each heterogeneous PP prefill stage to the corresponding layer span in an unpipelined decode worker's MLA pool.

Modifications

• Allow single-layer EAGLE with PP only for disaggregated prefill on the DeepSeek/GLM model path that supplies the last-stage draft embedding. Decode/null PP + speculative decoding and unsupported model families remain rejected.
• Materialize and load embed_tokens on the last DeepSeek/GLM PP stage when speculative decoding is enabled.
• Preserve dsa_topk_indices through PP serialization and reconstruction.
• Map heterogeneous plain-MLA NIXL transfers through one destination-index dispatcher, while retaining layer-ID/positional pairing for other layouts.
• Add regression coverage for PP/MLA transfer mapping, PP/EAGLE validation, embedding ownership, and DSA seed round trips (including the None case).

Accuracy Tests

Negative control on the exact parent commit (f539c1fc65e3a0de79b2a0d08805b096fdc85865): launching the same GLM-5.2-NVFP4 PP4 disaggregated-prefill configuration with EAGLE steps 3 on a GB300 failed before server readiness in check_server_args:

AssertionError: Pipeline parallelism is not compatible with overlap schedule, speculative decoding

This confirms that PP4 prefill with EAGLE did not start before this PR; the failure is the pre-change non-NPU PP + speculative-decoding restriction removed by this change.

GLM-5.2-NVFP4 on GB300 with Dynamo router:

• Prefill: 1 worker, PP4, EAGLE steps 3
• Decode: 1 worker, TP4, EAGLE steps 3
• GSM8K: 1,319-example dataset, 20-shot (1,299 scored examples), --max-tokens 16384
• Model revision: 53e0691e21895a3863a606dfd12910c69eba94ab
• Hardware: 2 GB300 nodes in one NVL72 segment

configuration	GSM8K score	decode acceptance length
PP4 prefill + TP4 decode, steps 3/3	94.69% (1230/1299)	3.745 mean (3.00-4.00)

This is within 0.76 percentage points of the original PR's 95.45% steps1/steps3 result. The decode acceptance length is the mean of 143 scheduler samples and is higher than the original PR's reported 3.589.

Tests

• BLACK_NUM_WORKERS=1 SKIP=no-commit-to-branch pre-commit run --all-files --show-diff-on-failure
• Targeted PP/MLA, PP/EAGLE, embedding-ownership, and PP DSA round-trip unit tests on the GB300 validation container: 28 passed, 264 deselected, 32 subtests passed.

Checklist

• Format code with pre-commit.
• Add unit tests.
• Update documentation (not applicable; no user-facing interface change).
• Provide accuracy results.
• Follow the SGLang code-style guidance.

---

CI States

Latest PR Test (Base): ✅ Run #35381952475
Latest PR Test (Extra): ❌ Run #35381952439
Latest PR Test (AMD ROCm 10): ❌ Run #35381952427

[YAMY1234]

LGTM. Very clean and clear implementation.

[YAMY1234]

/tag-and-rerun-ci

[YAMY1234]

FYI tracking issue for PP x speculative decoding, including the suggestion to fold the Qwen3_5 architectures from #39602 into this allowlist: #39634

[YAMY1234]

All NV CI Pass: https://github.com/sgl-project/sglang/actions/runs/35381952475/job/105916126771?pr=39378