fix: install latest transformers for glm_moe_dsa support in patched image

Use inline Dockerfile to pip install transformers from source on top of
glm5-hopper-patched, fixing the missing glm_moe_dsa architecture error.
Also adds --max-running-requests 16 to prevent hanging under load.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: Evrard-Nil

GLM-5.yaml

@@ -65,7 +65,10 @@ services:
 
   glm:
     <<: *vllm-common
-    image: lmsysorg/sglang:glm5-hopper-patched@sha256:abf8deb5e81cd7f942be8be10b1a92d4360d2f0a245b50ca8d9e27e9c05a98d6
+    build:
+      dockerfile_inline: |
+        FROM lmsysorg/sglang:glm5-hopper-patched@sha256:abf8deb5e81cd7f942be8be10b1a92d4360d2f0a245b50ca8d9e27e9c05a98d6
+        RUN uv pip install git+https://github.com/huggingface/transformers.git --system --break-system-packages -i https://mirrors.aliyun.com/pypi/simple/ 2>/dev/null || true
     container_name: glm
     command: >
       sglang serve