fix(pp): generate MTP draft tree after prefill

[lmyybh]

Motivation

	Dummy bootstrap	Real draft tree
After prefill	Builds a dummy tree by repeating the bonus token	Runs draft() and builds a real tree
First decode step	Mainly serves as a bootstrap step	Can verify real draft tokens immediately
Extra work	No extra draft call	One extra draft call after prefill
Implementation	Simpler	More state handling between prefill and decode

Speed Tests and Profiling

• environment: 16 x H800
• model: GLM-5.2-FP8 (TP8 CP8 PP2)

sglang command

export SGLANG_ENABLE_UNIFIED_RADIX_TREE=1
export SGLANG_PP_LAYER_PARTITION=40,38
sglang serve \
    --model-path GLM-5.2-FP8 \
    --tp 8 \
    --pp-size 2 \
    --enable-prefill-cp \
    --cp-strategy interleave \
    --attn-cp-size 8 \
    --trust-remote-code \
    --disable-overlap-schedule \
    --nnodes 2 \
    --dist-init-addr ${ip}:5000 \
    --node-rank ${node} \
    --kv-cache-dtype fp8_e4m3 \
    --enable-cache-report \
    --reasoning-parser glm45 \
    --tool-call-parser glm47 \
    --mem-fraction-static 0.8 \
    --cuda-graph-max-bs-decode 64 \
    --port 8000 \
    --speculative-algorithm EAGLE \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4

benchmark command

python -m sglang.benchmark.serving \
    --backend sglang-oai-chat \
    --port 8000 \
    --model GLM-5.2-FP8 \
    --dataset-name sharegpt \
    --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json \
    --request-rate inf \
    --max-concurrency 16 \
    --num-prompts 100 \
    --warmup-requests 10 \
    --flush-cache

metric	prefill_dummy	prefill_draft	change
output_throughput (tok/s)	474.836	487.109	+2.58%
mean_ttft (ms)	349.051	345.166	-1.11%
mean_tpot (ms)	30.241	27.745	-8.25%
mean_itl (ms)	29.109	28.303	-2.77%
accept_length	2.954	2.997	+1.45%

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 #30509714031
Latest PR Test (Extra): ⚠️ Run #30509713874

[lmyybh]

@liuqi-1

[liuqi-1]

@lmyybh I tested your code using GLM5.2 (PP2+TP8). An error about "pool memory leak" occurs.

[2026-08-01 12:10:03 PP1 ATTN_CP0 TP0] Scheduler hit an exception: Traceback (most recent call last):
  File "/usr/local/lib/python3.12/dist-packages/sglang/srt/managers/scheduler.py", line 4784, in run_scheduler_process
    scheduler.run_event_loop()
  File "/usr/local/lib/python3.12/dist-packages/sglang/srt/managers/scheduler.py", line 1571, in run_event_loop
    dispatch_event_loop(self)
  File "/usr/local/lib/python3.12/dist-packages/sglang/srt/managers/scheduler.py", line 4631, in dispatch_event_loop
    scheduler.event_loop_pp()
  File "/usr/local/lib/python3.12/dist-packages/torch/utils/_contextlib.py", line 124, in decorate_context
    return func(*args, **kwargs)
  File "/usr/local/lib/python3.12/dist-packages/sglang/srt/managers/scheduler_pp_mixin.py", line 174, in event_loop_pp
    self.on_idle()
  File "/usr/local/lib/python3.12/dist-packages/sglang/srt/managers/scheduler.py", line 3805, in on_idle
    self.invariant_checker._report_leak("pool", "\n".join(messages))
  File "/usr/local/lib/python3.12/dist-packages/sglang/srt/managers/scheduler_components/invariant_checker.py", line 428, in _report_leak
    raise_error_or_warn(
  File "/usr/local/lib/python3.12/dist-packages/sglang/srt/utils/common.py", line 4500, in raise_error_or_warn
    raise ValueError(message)
ValueError: pool memory leak detected! [full] total=2107968, available=2107712, evictable=0, protected=0, session_held=0, uncached=0

The test cli is:

NCCL_DEBUG_SUBSYS="" \
TORCHINDUCTOR_CACHE_DIR=/home/admin/inductor_root_cache \
SGLANG_PP_SKIP_PURE_CHUNKED_OUTPUT_COMM=1 \
SGLANG_PP_LAYER_PARTITION=40,38 \
nohup python -m sglang.launch_server \
  --model-path /home/models/GLM-5.2-FP8 \
  --host 0.0.0.0 --port 8076 \
  --attention-backend dsa \
  --dsa-prefill-backend flashmla_kv \
  --dsa-decode-backend flashmla_kv \
  --kv-cache-dtype fp8_e4m3 \
  --mem-fraction-static 0.83 \
  --tp-size 8 --pp-size 2 \
  --nnodes 2 --node-rank 0 \
  --dist-init-addr 10.13.2.89:25000 \
  --chunked-prefill-size 8192 \
  --enable-nsa-prefill-context-parallel \
  --nsa-prefill-cp-mode round-robin-split \
  --disable-radix-cache \
  --reasoning-parser glm45 \
  --tool-call-parser glm47 \
  --speculative-algo EAGLE \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4 \
  
  python3 -m sglang.test.run_eval \
  --base-url http://localhost:8076 \
  --eval-name gsm8k \
  --num-threads 64 \
  --max-tokens 2048 \
  --temperature 0.0 \
  --num-examples 200 \
  --num-shots 5 \
  --gsm8k-data-path '/home/shared/liuqi/gsm8k_test.jsonl'

I have updated my branch onto the main. Could you fix your code, and rebase your code onto my branch?