[Fix] Fix Qwen3.5 MTP startup with HiCache

[DarkraiHL]

Motivation

Qwen3.5 conditional-generation checkpoints store language-model attributes in
the nested text_config. During MTP draft remapping, SGLang sets
num_nextn_predict_layers = 1 only on the parent Hugging Face config. However,
ModelConfig.num_nextn_predict_layers is derived from hf_text_config, so the
draft depth remains None.

With NEXTN and HiCache enabled, the draft cache is therefore classified as a
sidecar instead of a packed MTP cache, and scheduler initialization fails with:

File "sglang/srt/mem_cache/hybrid_cache/hybrid_pool_assembler.py", line 936, in build_full_draft_pools
  if pool.layer_num == 0:
AttributeError: 'HybridLinearKVPool' object has no attribute 'layer_num'

This regression was exposed by
#30393, which introduced
packed-versus-sidecar HiCache draft routing. The original Qwen3.5 support in
#18489 normalized the MTP
depth only on the parent config. This PR completes that normalization without
changing the HiCache routing behavior.

Minimal reproduction

Tested with Qwen/Qwen3.5-397B-A17B-FP8, TP=8, on one node with 8 NVIDIA H20
96 GB GPUs (Python 3.12.3, PyTorch 2.13.0+cu130, CUDA 13.0,
sglang-kernel 0.4.6.post1+cu130):

python3 -m sglang.launch_server \
  --model-path Qwen/Qwen3.5-397B-A17B-FP8 \
  --trust-remote-code \
  --tp-size 8 \
  --mem-fraction-static 0.8 \
  --speculative-algo NEXTN \
  --enable-hierarchical-cache \
  --hicache-size 8

No L3 storage backend (including Mooncake) or custom HiCache I/O option is
required to reproduce the failure.

Modifications

• Synchronize the normalized Qwen3.5 MTP draft depth to hf_text_config.
• Add a focused regression test for Qwen3.5 MoE draft remapping.

 self.hf_config.num_nextn_predict_layers = 1
+self.hf_text_config.num_nextn_predict_layers = 1

Accuracy Tests

• Focused unit tests: 4 tests and 4 subtests passed.
• Minimal 8-H20 startup A/B with the same model and launch command:
  • Unpatched: target and MTP weights load, then scheduler initialization fails
    with the exception above; the server never becomes healthy.
  • Patched: /health returns HTTP 200, all 8 TP schedulers remain alive, and a
    completion request succeeds.
• Mooncake smoke test on the patched source: 3/3 usage, 4/4 text, and 4/4
  uncached multimodal requests returned HTTP 200.
• Independent AMD MI355X/ROCm reproduction: #34611
  reports the same stack trace, and
  SemiAnalysisAI/InferenceX#2582
  applies this exact fix to its Qwen3.5 NEXTN + HiCache benchmark.

Speed Tests and Profiling

Not applicable. This change only synchronizes a configuration attribute during
startup and does not modify the inference hot path.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci.
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): ✅ Run #31775596976
Latest PR Test (Extra): ✅ Run #31775596847

[1e4ves]

thanks！But sidecar should not fail as a fallback path, so there are also bugs here. let me fix it.

[DarkraiHL]

@hnyls2002 Could you take a look when convenient? This is a small, focused fix for the Qwen3.5 NEXTN + HiCache startup failure, with a regression test. Thanks!

[huangtingwei9988]

/tag-and-rerun-ci

[DarkraiHL]

@1e4ves Thanks for catching the fallback-path issue. I’ve incorporated your HybridLinearKVPool unwrapping fix, added focused regression coverage.

[YAMY1234]

/rerun-failed-ci