model card: sglang only

README.md

@@ -7,7 +7,6 @@ tags:
   - vision-language
   - glm
   - sglang
-  - vllm
 base_model:
   - zai-org/GLM-5.2
 ---
@@ -17,7 +16,7 @@ base_model:
 **GLM-5.2 with sight.** A vision-language model that bolts the MoonViT vision encoder from
 [Kimi-K2.6](https://huggingface.co/moonshotai/Kimi-K2.6) onto
 [GLM-5.2](https://huggingface.co/zai-org/GLM-5.2) through a trained PatchMerger projector —
-shipped as **one self-contained checkpoint** you can hand straight to SGLang or vLLM.
+shipped as **one self-contained checkpoint** you can hand straight to SGLang.
 
 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
@@ -45,8 +44,8 @@ At ~757 GB the weights do not fit on four GPUs. For a 4-GPU deployment use
 
 ## Quickstart
 
-Both engines need 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:
+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:
 
 ```bash
 hf download baseten/GLM-5.2-Vision-FP8 --include 'plugins/*' --local-dir ./glm5v
@@ -77,19 +76,6 @@ python -m sglang.launch_server \
   --port 30000
 ```
 
-Measured KV budget at these settings: ~1.16M tokens.
-
-### vLLM
-
-The pip install registers a `vllm.general_plugins` entry point, so no flags and no env vars:
-
-```bash
-vllm serve baseten/GLM-5.2-Vision-FP8 --trust-remote-code --tensor-parallel-size 8
-```
-
-> Note: the vLLM path is wired up but has had less testing than SGLang here. If you hit
-> trouble, SGLang is the better-trodden route.
-
 ### Query it
 
 Standard OpenAI multimodal messages — the image is delivered as `image_url`: