A developer running Qwen3.5-122B on an M3 Ultra Mac Studio identified three specific bugs in the qMLX serving stack that caused severe latency during long-context agentic coding.
The issues included prompt instability breaking KV cache matching, interrupted streaming replies causing history divergence, and checkpoint poisoning triggering aggressive eviction.
Fixing these bugs reduced prefill time from minutes to sub-seconds, allowing 53k tokens to be cached with only 33 prefilled. The author open-sourced the qMLX fork and a benchmark script, noting that hybrid attention optimizations specific to Qwen will likely cause divergence from upstream stacks.