harrypart's picture
model card: trim intro
83609f8 verified
|
Raw
History Blame
4.89 kB
metadata
license: mit
library_name: sglang
pipeline_tag: image-text-to-text
tags:
  - multimodal
  - vision-language
  - glm
  - sglang
base_model:
  - zai-org/GLM-5.2

GLM-5.2-Vision (NVFP4)

GLM-5.2 with sight. A vision-language model that bolts the MoonViT vision encoder from Kimi-K2.6 onto GLM-5.2 through a trained PatchMerger projector.

GLM-5.2 is a strong open reasoning model with no vision input. This checkpoint adds it, without touching a single GLM weight: the text backbone and the vision tower are both frozen and byte-identical to their upstream releases. The only newly-trained parameters are the 49.5M-parameter projector that maps MoonViT's 1152-dim patch embeddings into GLM's 6144-dim token space.

Component Detail
Text backbone GLM-5.2 (744B total / A40B active, MoE + MLA + DSA sparse attention) — frozen
Vision tower MoonViT-3d from Kimi-K2.6, 27 layers, 1152-dim — frozen
Projector PatchMerger MLP (pre_norm → linear_1 → GELU → linear_2), 1152→4608→6144 — trained
Text weights NVFP4, from nvidia/GLM-5.2-NVFP4
Size ~466 GB
Hardware 8×B200, or 4×B200 at 256k context — Blackwell only
Image tokens up to 4096 per image (16384 MoonViT patches, 2×2 merge)
Max context 1048576 (1M tokens)

Quickstart

SGLang needs a small out-of-tree plugin, because Glm5vForConditionalGeneration is not yet an upstream architecture. It ships inside this repo, so there is nothing else to clone:

hf download baseten/GLM-5.2-Vision-NVFP4 --include 'plugins/*' --local-dir ./glm5v
pip install ./glm5v/plugins

SGLang

export SGLANG_EXTERNAL_MODEL_PACKAGE=sglang_glm5v
export SGLANG_EXTERNAL_MM_PROCESSOR_PACKAGE=sglang_glm5v
export SGLANG_EXTERNAL_MM_MODEL_ARCH=Glm5vForConditionalGeneration
python -m sglang_glm5v.patch          # registers the arch on SGLang's DSA/MLA paths

8×B200 — full 1M context

python -m sglang.launch_server \
  --model-path baseten/GLM-5.2-Vision-NVFP4 --trust-remote-code \
  --tp-size 8 \
  --quantization modelopt_fp4 \
  --disable-shared-experts-fusion --disable-flashinfer-autotune \
  --attention-backend dsa --mm-attention-backend sdpa \
  --kv-cache-dtype fp8_e4m3 --page-size 64 \
  --mem-fraction-static 0.85 \
  --context-length 1048576 \
  --reasoning-parser glm45 --tool-call-parser glm47 \
  --served-model-name glm-5.2-vision \
  --port 30000

4×B200 — 256k context

Same command with --tp-size 4, a higher memory fraction, and a smaller context:

python -m sglang.launch_server \
  --model-path baseten/GLM-5.2-Vision-NVFP4 --trust-remote-code \
  --tp-size 4 \
  --quantization modelopt_fp4 \
  --disable-shared-experts-fusion --disable-flashinfer-autotune \
  --attention-backend dsa --mm-attention-backend sdpa \
  --kv-cache-dtype fp8_e4m3 --page-size 64 \
  --mem-fraction-static 0.90 \
  --context-length 262144 \
  --reasoning-parser glm45 --tool-call-parser glm47 \
  --served-model-name glm-5.2-vision \
  --port 30000

Query it

Standard OpenAI multimodal messages — the image is delivered as image_url:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:30000/v1", api_key="none")
r = client.chat.completions.create(
    model="glm-5.2-vision",
    messages=[{"role": "user", "content": [
        {"type": "image_url", "image_url": {"url": "https://ultralytics.com/images/bus.jpg"}},
        {"type": "text", "text": "Describe this image in detail."},
    ]}],
    temperature=1.0, top_p=0.95, max_tokens=512,
)
print(r.choices[0].message.content)

GLM-5.2 is a reasoning model: with --reasoning-parser glm45 the chain of thought arrives in message.reasoning_content and the answer in message.content.

Deploy on Baseten

Ready-to-push Truss configs ship in this repo under truss/:

hf download baseten/GLM-5.2-Vision-NVFP4 --include 'truss/*' --local-dir ./glm5v
cd glm5v/truss
Config GPUs Context
config_nvfp4.yaml 8×B200 1M
config_nvfp4_4gpu.yaml 4×B200 256k
truss push --config config_nvfp4.yaml      --remote <your-remote>   # 8 GPU
truss push --config config_nvfp4_4gpu.yaml --remote <your-remote>   # 4 GPU

License

MIT, following both parents: GLM-5.2 (MIT) and Kimi-K2.6 (Modified MIT). The projector weights are released under MIT. Redistributed upstream weights remain under their original terms.

Acknowledgements

Built on Z.ai's GLM-5.2 and Moonshot AI's Kimi-K2.6. Neither team was involved in this work; please do not direct issues with this checkpoint to them.