Eval bug: Vulkan Backend `no-kv-offload` on Qwen3-Coder-Next and Qwen3.6-35B-A3B produces gibberish output

当你在 llama.cpp 的 Vulkan 后端下对 Qwen3-Coder-Next、Qwen3.6-35B-A3B 这类 MoE 模型启用 no-kv-offload 时,模型输出会变成乱码。优先排查是否关闭了 KV Cache 的 GPU 卸载,并尝试将 KV Cache 也放到 GPU 上

![[Bug]: Composite VLM wrapper (Mistral3ForConditionalGeneration) resolves tie_word_embeddings from the wrong (top-level) config, silently dis](https://www.chat-gpts.plus/wp-content/uploads/2026/08/51063-e1d92b3e-768x403.jpg)

![[Question]: When parsing a document under the maximum token limit preset by the model, an error is reported indicating that the input tokens](https://www.chat-gpts.plus/wp-content/uploads/2026/08/6291-08cae5a7-768x403.jpg)




![[Bug]: Structured-output scheduler can keep advancing a terminated xgrammar matcher](https://www.chat-gpts.plus/wp-content/uploads/2026/08/42619-81e0022e-768x403.jpg)