llm-kb
← Back to social
Community

Developer fixes 3 bugs in qMLX serving stack to make Qwen3.5-122B long-context inference usable on Mac Studio

A developer running Qwen3.5-122B on an M3 Ultra Mac Studio with 96GB RAM identified and fixed three bugs in the qMLX fork of rapid-mlx that caused 3-5 minute cold fills during long-context agentic coding. The fixes addressed prompt instability from a unique message ID breaking KV cache matching, an interrupt path that failed to persist streaming replies, and a third unspecified bug, making long-context inference practical.

3 engagement·1 source·Mon, Jul 13, 2026, 12:47 AM
The developer switched from DS4 Flash to Qwen3.5-122B for long-context agentic coding on an M3 Ultra Mac Studio. Despite the model fitting better, follow-up messages took 3-5 minutes to start generating (cold fills) even with a 'warm' context. The root cause was three bugs in the qMLX fork of rapid-mlx: (1) Prompt Instability: a unique message ID in the system prompt broke byte-exact KV cache matching, forcing full re-computation on every turn; (2) Interrupt Path: streaming replies were not persisted when generation was interrupted; (3) a third bug not detailed. After fixing these, long-context inference became usable.

Entities

Qwen3.5-122B(model)Mac Studio(tool)qMLX(tool)rapid-mlx(tool)M3 Ultra(concept)

Related

CommunitySun, Jul 12, 2026, 07:16 AM

User seeks help tuning llama-server cache on Strix Halo for Qwen 3.5 122B

A user on Reddit reports performance issues with large models (e.g., Qwen 3.5 122B) on a Strix Halo system, where a full cache miss at 100k context causes 10-20 minutes of prompt processing time. They have configured --cache-ram 16384 to increase available VRAM for cache, but seek further tuning advice.

2 engagement·1 source·reddit
CommunitySat, Jul 11, 2026, 06:03 PM

User seeks to extend Qwen 3.6 27B context window beyond 100k tokens

A user reports running Qwen 3.6 27B (Q8_0) at 100k context length but finds reliability insufficient. They ask the community for techniques beyond KV cache quantization to improve stability at longer contexts.

20 engagement·1 source·reddit
CommunitySun, Jul 12, 2026, 12:24 PM

Users report tool-call failures and looping in Qwen3.6-27B

A Reddit user reports persistent tool-call failures and looping behavior when using Qwen3.6-27B as a local model. The user describes needing constant monitoring to prevent the model from entering infinite loops or hallucinating tool calls, and has developed an extensible workaround to mitigate these issues.

18 engagement·2 sources·reddit
IncidentSun, Jul 12, 2026, 08:04 PM

User reports character encoding issues with LM Studio, Zoo, and Qwen 3.6 pipeline

A user running an Unsloth quant of Qwen3.6-35B-A3B on a Legion 7i laptop with a 5080 GPU reports two issues in the LM Studio to Zoo (formerly Roo) pipeline: character encoding mismatches causing em-dashes to appear as random characters, and a second unspecified problem. The user achieves ~80 tk/s by offloading all experts to CPU.

3 engagement·1 source·reddit
BenchmarkSun, Jul 12, 2026, 10:47 AM

User benchmarks 4x RTX 5060 Ti with SGLang for Qwen3.6 27B, finds better concurrency than vLLM

A user shared benchmark results showing that SGLang handles higher concurrency better than vLLM when running Qwen3.6 27B (INT8 with bf16 KV cache) on a 4x RTX 5060 Ti (64GB VRAM) setup. The test achieved 200 successful requests at 8 concurrency over 348.87 seconds, processing 61,870 input tokens and generating 44,525 tokens. This provides a practical reference for others considering multi-GPU configurations with these consumer cards.

3 engagement·1 source·reddit