Fixed three bugs that made Qwen3.5-122B a daily driver on Mac Studio (mrzk.io)

A developer has made significant improvements to the Qwen 3.5-122B model to enhance its usability for long-context programming on a Mac Studio Ultra. After experiencing frustrating latencies with prior models like DS4 Flash, which caused delays of up to five minutes during follow-up queries, the developer opted to fix three critical bugs in their serving stack. This transition to Qwen 3.5 aligns more closely with their need for near-instantaneous responses while handling extensive token conversations, essential for maintaining a productive coding flow. The updates address issues related to cache management, especially concerning a hybrid attention architecture that previously led to inefficient memory usage and lengthy processing times. Key fixes included stabilizing the system prompt to prevent unnecessary recomputations, ensuring streamed replies were correctly stored during interruptions, and refining checkpoint storage to maintain critical data amidst excessive irrelevant writes. As a result, prefill times dramatically decreased from minutes to mere seconds, allowing the model to efficiently utilize its 96GB unified memory and improving the interaction speed to match the developer's workflow needs. This development not only enhances local inference capabilities but also illustrates the potential of consumer-grade hardware to run advanced AI models effectively.