Qwen3.8 27B

User runs SWEBench locally with llama.cpp and an Unsloth Qwen3.8 27B IQ4_NL GGUF on an RTX 3090, at 229,376 context with a q8_0 KV cache. Setup uses --jinja, -ngl 999, flash attention, mmap load mode, and a bf16 mmproj; speculative decoding is disabled. The 500-test run took 8 days and produced 9 LimitExceeded and 103 TimeoutExpired errors, leaving 388 completed tests with 339 resolved and 48 unresolved, an 87.5969% resolve rate. User asks how to reduce errors, improve success rate, and speed up the run without sacrificing context.