Qwen3.6 Model Performance Degradation Bug

A bug has been reported in llama.cpp where, when using the Qwen3.6-35B-A3B model, the last portion of the KV cache is dropped by an amount equal to the batch size. Because the dropped tokens must be reprocessed every turn, the performance degradation becomes more severe as the context grows longer.

Key symptoms:

• With a 150k context and a batch size of 4096, prefill time increases roughly 10x, from 317ms to 3,074ms.
• This issue has only been confirmed with the Qwen3.6-35B-A3B model, and does not reproduce with other models such as Gemma.

Workaround:

• Use build b9222, or
• Roll back to commit ccee426 and build directly from source.