Marzukia's qMLX fork fixes 3 bugs making Qwen3.5-122B long-context usable on Mac Studio

Running large language models locally for long-context agentic coding has long been a pain point — high memory requirements, agonizing cold starts, and unreliable caching. Developer marzukia recently switched from DS4 Flash to Qwen3.5-122B on an M3 Ultra Mac Studio with 96GB of unified memory, expecting better context handling. Instead, they hit 3–5 minute delays on every follow-up message despite a 'warm' context. After digging into the qMLX serving stack (a fork of rapid-mlx), they uncovered three distinct bugs that were destroying KV cache efficiency.

The first bug, prompt instability, stemmed from a unique message ID embedded in the system prompt that broke byte-exact KV cache matching — forcing a full re-computation on every turn. The second, interrupt path, meant streaming replies weren't persisted when a generation was interrupted, causing history divergence on the next call. Finally, checkpoint poison: a background writer created unmatchable checkpoint files that crowded out valid ones, triggering aggressive eviction of cached tokens. After fixing all three, prefill time for 168k-token sessions dropped to 2.6–4.8 seconds with delta prefills as low as 67 tokens. The developer open-sourced the fork (qMLX) and a benchmark script (bench_qmlx.py) that separates prefill and decode metrics, noting that the hybrid attention optimizations are highly specific to Qwen and unlikely to be merged upstream.