Running Qwen3.5-122B on Mac Studio 96GB: Fixed 3 bugs that made long-context inference usable

Hey everyone, I recently switched from DS4 Flash to Qwen3.5-122B on my M3 Ultra Mac Studio for long-context agentic coding. While the model fit better, I hit a wall where follow-up messages took 3-5 minutes to start generating (cold fills) despite having a "warm" context. Turns out the issue wasn't the model, but three specific bugs in my serving stack (qMLX fork of rapid-mlx): Prompt Instability: A unique message ID in the system prompt broke byte-exact KV cache matching, forcing a full re-compute on every