ROCm/FastFlowLM v0.1.6

## 🧠 FastFlowLM v0.1.6 — Qwen3 Support, Flash Attention Optimization & Bug Fixes

on GitHub

FastFlowLM v0.1.6 brings support for the Qwen3 model family, improved precision, and several important bug fixes and usability enhancements.

✨ What's New

1. Qwen3:4B Support with Think Mode

• Added support for the Qwen3:4B model.
• You can now toggle think mode using:
  • CLI: Type /think to toggle on/off interactively.
  • Server mode: Set the "think" flag in the request payload.

2. Unified Code Architecture

• Refactored model handling logic to improve code reuse and simplify support for multiple model families.

3. Configurable Serve Port via Environment Variable

• You can now customize the server port using the FLM_SERVE_PORT environment variable.
  • Default port remains 11434.

4. Improved Flash Attention Precision

• Optimized computation paths for flash attention to enhance numerical stability and inference accuracy.

5. Bug Fixes

• 🐞 Fixed sampler not resetting correctly when switching models.
• 🐞 Fixed hyperparameters being overwritten by defaults during model load.
• 🐞 Fixed several typos across CLI messages and logs.

6. Version Info Command

• Added flm version command to print the current version of FastFlowLM.

✅ Usage Recap

Interactive CLI:

flm run qwen3
flm run llama3.2 prompt.txt

Server Mode (customizable via FLM_SERVE_PORT)

This release paves the way for seamless support of multi-family models with increased precision and flexibility.