ROCm/FastFlowLM v0.9.4

🔧 FastFlowLM v0.9.4 — Vision Stability Patch

on GitHub

FastFlowLM v0.9.4 is a minor update focused on improving the numerical stability of the vision pipeline for Gemma3.

🔍 What's Improved

🧠 1. Gemma3 Vision Stability

• Enhanced numerical stability for the vision-language model (VLM) in gemma3:4b.
• This improves consistency during image processing and response generation.

⚙️ 2. NPU Performance Mode (New Default)

• FastFlowLM now defaults to NPU performance mode for improved runtime efficiency.
• You can override the mode with the --pmode flag:

flm run llama3.2 --pmode turbo
flm run gemma3 [filename] --pmode turbo
flm run gemma3 [filename] 4096 --pmode turbo
flm serve --pmode turbo
flm serve qwen3 --pmode turbo

🔏 3. Official Code Signing

• FastFlowLM binaries are now officially signed.
• This eliminates false virus alarms and enhances installation safety and system trust.

No user-facing changes were made in this release beyond improved stability. Stay tuned for more performance and compatibility upgrades in upcoming versions.