Feature Request: Allow Master-Free Mode (Headless Dispatch Only)

[IT-BillDeng]

---

🚀 Feature Request: Allow Master-Free Mode (Headless Dispatch Only)

Environment Overview

• Local machine: NVIDIA RTX 5080 (single GPU), used mainly for workflow design and testing.

• Remote server: 4 × RTX 2080 Ti 22 GB cards, running as distributed Workers for heavy generation tasks.

• Deployment: Docker-based ComfyUI-Distributed setup.

  • Local PC runs as current “Master”.
  • Remote server runs 4–5 Workers (--listen 0.0.0.0 --enable-cors-header).

---

🧩 Current Behavior

At present, ComfyUI-Distributed requires the Master process to remain active and ticked (grey-checked) even if all computation is delegated to Workers.
If the Master is closed or unavailable, the distributed queue cannot aggregate Worker results or coordinate execution.
This behavior effectively means:

• The Master must have the same Python + Torch environment.
• It must stay powered on during all remote jobs.
• Results cannot be returned to a “headless” or non-compute node.

---

💡 Requested Feature

Provide an option to disable Master participation entirely, such that:

1. Master can run in “dispatch-only / coordinator mode”, or even as a stateless lightweight service.

2. After job dispatch, Workers could:

   • Write outputs directly to shared storage (e.g., NFS/S3).
   • Or send completion callbacks to an API endpoint rather than requiring a live Master aggregator.

3. Allow running without GPU / PyTorch on the Master, useful for cloud-hosted or laptop-based coordination.

---

🧭 Why This Matters

This feature would unlock much more flexible usage patterns:

Scenario	Benefit
Local PC (5080) for workflow debugging → then delegate final high-res render to remote cluster	Can shut down local PC once job is dispatched
Completely non-compute orchestration node (e.g., cloud VPS)	Enables centralized scheduling across multiple GPU servers
Multi-user lab setups	One lightweight web Master per user, many shared Workers underneath

Currently, the Master still needs to stay online and have Torch installed, otherwise results aggregation fails.

---

⚙️ Possible Implementations

• Add a --no-master-runtime or --dispatch-only flag.
• Allow Worker nodes to write outputs to shared storage and notify via REST WebSocket API.
• Introduce an optional lightweight “Result Collector” microservice to replace the Master aggregation process.

---

🙏 Thanks

This project is already a huge improvement for multi-GPU / distributed workflows.
Adding this “headless Master” mode would make it ideal for hybrid local+remote setups like mine, enabling better power management and flexibility.

---