sglang plugin

plugins/pyproject.toml

@@ -5,16 +5,10 @@ build-backend = "setuptools.build_meta"
 [project]
 name = "glm5v-serve"
 version = "0.1.0"
-description = "Serve GLM-5.2-Vision (GLM-5.2 text + MoonViT vision + PatchMerger projector) on vLLM and SGLang"
+description = "Serve GLM-5.2-Vision (GLM-5.2 text + MoonViT vision + PatchMerger projector) on SGLang"
 readme = "README.md"
 requires-python = ">=3.10"
 license = { text = "Apache-2.0" }
 
-# vLLM discovers plugins through this entry-point group and calls them during
-# engine init — so `pip install glm5v-serve` is all that stands between a stock
-# `vllm serve baseten/GLM-5.2-Vision-FP8` and a working VLM. No flags, no patches.
-[project.entry-points."vllm.general_plugins"]
-glm5v = "vllm_glm5v.register:register"
-
 [tool.setuptools]
-packages = ["vllm_glm5v", "sglang_glm5v"]
+packages = ["sglang_glm5v"]

plugins/sglang_glm5v/glm5v.py

@@ -2,8 +2,7 @@
 """glm5v on SGLang: MoonViT (Kimi-K2.5) vision tower + trained PatchMerger
 projector + GLM-5.2 (GlmMoeDsa) text decoder.
 
-Mirror of ``vllm_glm5v/glm5v.py`` for SGLang. It is SGLang's in-tree
-``KimiK25ForConditionalGeneration`` with the same three swaps we made on vLLM:
+It is SGLang's in-tree ``KimiK25ForConditionalGeneration`` with three swaps:
 
   1. text backbone ``DeepseekV3ForCausalLM`` -> ``GlmMoeDsaForCausalLM``
      (GLM-5.2; SGLang ships it natively in ``glm4_moe.py``, DSA path included).

plugins/sglang_glm5v/processor.py

@@ -6,7 +6,7 @@ placeholder: ``<|image|>`` (id from ``config.media_placeholder_token_id``,
 154854) instead of Kimi's ``<|media_pad|>``. The GLM chat template
 (``chat_template.jinja`` in the assembled checkpoint) wraps each image as
 ``<|begin_of_image|><|image|><|end_of_image|>``; only the inner ``<|image|>``
-run is expanded to per-patch tokens, exactly as in the vLLM deployment.
+run is expanded to per-patch tokens.
 
 All preprocessing (NaViT resize math, GPU patchify, grid_thw handling) is
 inherited unchanged — it is the same MoonViT pipeline the projector was