ROCm/FastFlowLM v0.9.22

🚀 FastFlowLM v0.9.22 — Qwen Series Speed Boost & Custom IP Binding

on GitHub

FastFlowLM v0.9.22 delivers decoding optimizations across the entire Qwen model family and adds more flexible server deployment options.

⚡ Decoding Acceleration: Qwen Series

This update brings decoding performance improvements to all models in Qwen3 family:

• qwen3:0.6b (reasoning) reaching 80 tps
• qwen3:1.7b (reasoning) reaching 43 tps
• qwen3:4b (reasoning) reaching 21 tps
• qwen3:8b (reasoning) reaching 12.5 tps
• deepseek-r1-0528:8b (reasoning) reaching 12.5 tps
• qwen3-it:4b reaching 21 tps
• qwen3-tk:4b (reasoning) reaching 21 tps
• qwen3vl-it:4b (vison) reaching 21 tps

Expect smoother token generation and reduced latency across all context lengths.

🌐 IP Binding for Serve Mode

You can now bind flm serve to a specific IP address:

flm serve --host 127.0.0.1

Use cases include:

• Localhost-only testing
• Binding to specific network interfaces
• Secure multi-device deployments

🤖 Flm-Compagnon (flm-manager)

Flm Compagnon (flm-manager) is an independent project done by @julienM77 It’s intuitive, user-friendly, and we absolutely love it!

Brief: Flm Compagnon is a modern GUI designed to accompany and manage the FastFlowLM (FLM) project. It offers a smooth user experience to interact with your local AI models, monitor the server, and manage your configurations.

https://github.com/julienM77/Flm-Companion

This release improves both backend speed and frontend deployment control.