[Diffusion] Is there support for /metrics endpoint in SGLang Diffusion (Qwen-Image)

[capri-xiyue]

I am deploying a diffusion model (Qwen/Qwen-Image-2512) using the sglang[diffusion] package on Kubernetes . However I am unable to access the /metrics endpoint when running in diffusion mode.

Environment:

• Image: lmsysorg/sglang:nightly-dev-20260209-bf89cc38
• Command:

            sglang serve \
              --model-path Qwen/Qwen-Image-2512 \
              --host 0.0.0.0 \
              --port 30000 \
              --dtype bfloat16 \
              --tp-size 1 \
              --dit-layerwise-offload \
              --text-encoder-cpu-offload \
              --enable-metrics \
              --mem-fraction-static 0.6

• Hardware: NVIDIA H100 (80GB)

The Issue: Even with --enable-metrics specified in the launch arguments, requests to the /metrics endpoint return a 404 Not Found.

Logs:

[2026-02-11 00:45:07]"GET /metrics HTTP/1.1" 404 Not Found

Questions:

Is the /metrics endpoint currently supported for the diffusion runtime ?

If not, is there a planned roadmap to bring the Prometheus metrics (like sglang:xxxx) to the diffusion-side of the project?

Reproducible Deployment Snippet:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: sglang-qwen-diffusion
spec:
  replicas: 1
  selector:
    matchLabels:
      app: sglang-diffusion
  template:
    metadata:
      labels:
        app: sglang-diffusion
    spec:
      containers:
      - name: sglang-runtime
        image: lmsysorg/sglang:nightly-dev-20260209-bf89cc38
        command: ["/bin/bash", "-c"]
        args:
          - |
            # 1. FIX: Corrected pip flag for pre-releases
            pip install --upgrade --pre "sglang[diffusion]"
            
            # 2. Launch with even more conservative memory settings
            sglang serve \
              --model-path Qwen/Qwen-Image-2512 \
              --host 0.0.0.0 \
              --port 30000 \
              --dtype bfloat16 \
              --tp-size 1 \
              --dit-layerwise-offload \
              --text-encoder-cpu-offload \
              --enable-metrics \
              --mem-fraction-static 0.6
        env:
        - name: HF_TOKEN
          valueFrom:
            secretKeyRef:
              name: hf-token
              key: token
        - name: FLASHINFER_DISABLE_VERSION_CHECK 
          value: "1"
        - name: SGLANG_USE_CACHE_DIT
          value: "1"
        resources:
          limits:
            nvidia.com/gpu: 1
            cpu: "20"
            memory: "128Gi" # Essential for text-encoder offloading
          requests:
            nvidia.com/gpu: 1
            cpu: "10"
            memory: "96Gi"
        volumeMounts:
        - name: shm
          mountPath: /dev/shm
      volumes:
      - name: shm
        emptyDir:
          medium: Memory
          sizeLimit: 32Gi
      tolerations:
      - key: "nvidia.com/gpu"
        operator: "Equal"
        value: "present"
        effect: "NoSchedule"
---
apiVersion: v1
kind: Service
metadata:
  name: sglang-service
spec:
  selector:
    app: sglang-diffusion
  ports:
  - protocol: TCP
    port: 80
    targetPort: 30000
  type: LoadBalancer

Thank you for the help!