Modelpublic

AlayaNeW/GLM-5.2-DSpark

sourceHugging Facemitupdated 21d agoView on Hugging Face

1likes498downloads

Model Card

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

Field	Value
Architecture	`Qwen3DSparkModel` (`qwen3_dspark`)
Draft layers	5 (4× sliding attention + 1× full attention)
Hidden size / head dim	6144 / 128
Attention	MHA, 64 Q heads, 64 KV heads
Sliding window	1024
Block size $\gamma$	8
Target layers (KV inject)	`[1, 20, 38, 56, 75]` (GLM-5.2 has 78 layers)
Markov head	vanilla, rank 512
Confidence head	enabled (with Markov features)
Vocab	154880 (same as GLM-5.2)
Precision	bfloat16

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

bash

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

bash

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):

bash

  --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

python

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

Context Len	AL	Pos 0	Pos 1	Pos 2	Pos 3	Pos 4	Pos 5	Pos 6	Pos 7
1k	4.75	84.28	68.72	55.81	45.87	38.09	31.93	26.92	22.64
2k	4.65	83.64	68.00	54.98	44.79	36.68	30.42	25.46	21.44
8k	4.80	84.33	69.57	57.15	47.40	39.39	32.68	27.13	22.37
16k	4.74	84.25	69.18	56.59	46.36	38.09	31.65	26.26	21.70
32k	4.52	83.30	67.37	53.97	43.12	34.85	28.13	22.76	18.37

Speed-Bench: Qualitative

Category	AL	Pos 0	Pos 1	Pos 2	Pos 3	Pos 4	Pos 5	Pos 6	Pos 7
coding	4.34	84.21	66.83	51.88	40.30	31.43	24.66	19.54	15.67
stem	3.60	76.62	56.13	40.20	29.33	21.70	15.98	11.66	8.36
humanities	3.54	76.44	55.38	39.23	28.25	20.62	15.22	11.12	8.09
math	3.72	77.59	57.41	42.00	30.99	23.25	17.50	13.29	9.96
multilingual	5.05	84.59	70.43	59.02	50.43	43.45	37.50	32.31	27.60
qa	3.71	76.26	55.80	41.32	31.14	23.98	18.26	13.96	10.72
rag	5.35	87.21	74.31	64.01	55.02	47.85	41.40	35.39	30.29
roleplay	2.80	67.73	41.69	25.42	16.09	11.13	7.80	5.82	4.53
reasoning	4.19	80.65	62.19	47.78	37.65	29.79	24.06	19.99	16.51
summarization	4.53	84.10	67.37	54.29	43.26	34.78	28.23	22.50	18.14
writing	3.27	72.69	49.30	33.60	23.86	17.54	13.19	10.04	7.65

Citation

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

bibtex

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