GLM-5.2-DSpark

GLM-5.2-DSpark

This repository releases a DSpark draft model for GLM-5.2, trained and published by AlayaNeW. It is a speculator, not a standalone LLM: pair it with a GLM-5.2 target (BF16 or FP8) in SGLang or vLLM.

• Hugging Face: AlayaNeW/GLM-5.2-DSpark
• Target model: zai-org/GLM-5.2 / zai-org/GLM-5.2-FP8
• Algorithm: DSpark (parallel block draft + Markov head + confidence head)

Architecture

The training data is a mixture of code, reasoning, Chinese alignment, and long-context corpora. Answers were regenerated by GLM-5.2 to build the target cache. The draft was trained from scratch at 8k sequence length using DeepSpec。

Quick Start

Requires a recent SGLang or vLLM build with native DSpark support. This draft is for GLM-5.2 only.

Serve with SGLang

export SGLANG_ENABLE_SPEC_V2=1

sglang serve \
  --model-path zai-org/GLM-5.2-FP8 \
  --trust-remote-code \
  --tp-size 8 \
  --reasoning-parser glm45 \
  --tool-call-parser glm47 \
  --context-length 202752 \
  --mem-fraction-static 0.85 \
  --cuda-graph-max-bs 256 \
  --disable-shared-experts-fusion \
  --host 0.0.0.0 \
  --port 8100 \
  --speculative-algorithm DSPARK \
  --speculative-draft-model-path AlayaNeW/GLM-5.2-DSpark \
  --speculative-dspark-block-size 8

Serve with vLLM

vllm serve zai-org/GLM-5.2-FP8 \
  --trust-remote-code \
  --tensor-parallel-size 8 \
  --tool-call-parser glm47 \
  --enable-auto-tool-choice \
  --reasoning-parser glm45 \
  --max-model-len 202752 \
  --max-num-seqs 256 \
  --gpu-memory-utilization 0.85 \
  --max-cudagraph-capture-size 256 \
  --host 0.0.0.0 \
  --port 8100 \
  --served-model-name GLM-5.2 \
  --speculative-config '{
    "method": "dspark",
    "model": "AlayaNeW/GLM-5.2-DSpark",
    "num_speculative_tokens": 8,
    "draft_sample_method": "probabilistic"
  }'

This draft includes a confidence head. vLLM can enable adaptive verification (load-aware verify budget; losslessness unchanged):

  --speculative-config '{
    "method": "dspark",
    "model": "AlayaNeW/GLM-5.2-DSpark",
    "num_speculative_tokens": 8,
    "draft_sample_method": "probabilistic",
    "enable_adaptive_verification": true
  }'

Call the server

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8100/v1", api_key="EMPTY")
resp = client.chat.completions.create(
    model="GLM-5.2",
    messages=[{"role": "user", "content": "Write a Python function that checks if a number is prime."}],
    max_tokens=1024,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)
print(resp.choices[0].message.content)

Evaluation

• Runtime: SGLang on 8 NVIDIA B200 GPUs (TP8)
• Speculation block size: 8 (--speculative-dspark-block-size 8, --speculative-num-draft-tokens 9 = block size + 1)
• Sampling: temperature 0.7, top-p 0.95 (thinking disabled)

Acceptance length is the per-request mean of completion tokens divided by verification steps. Higher is better.

Speed-Bench: Low Entropy

Speed-Bench: Qualitative

Citation

If you use this draft model, please cite AlayaNeW/GLM-5.2-DSpark:

@misc{alayanew2026glm52dspark,
  title={GLM-5.2-DSpark},
  author={AlayaNeW},
  year={2026},
  howpublished={\url{https://huggingface.co/AlayaNeW/GLM-5.2-DSpark}}
}