APIEmptyResponseError

你运行的 Kimi Code 版本是？

0.38.0

你使用的是哪个开放平台/订阅？

本地通过华为的NPU部署的量化版本模型

你使用的是哪个模型？

kimi-k2.7

你的电脑平台是？

linux x64

你遇到了什么问题？

Error: [provider.api_error] The API returned a response containing only thinking content without any text or tool calls. This usually indicates the stream
was interrupted or the output token budget was exhausted during reasoning. Provider stop details: finishReason=tool_calls, rawFinishReason=tool_use.
Provider: anthropic, model: kimi-k27

复现步骤？

进入一个项目目录，运行kimi 让他分析整个项目，他会调用工具读取文件，但是一会之后就会出现

kimi-debug-session_-20260821-014913.zip

期望的行为是什么？

No response

补充信息

exec vllm serve /data/models/Kimi-K2.7-Code-w4a8
--host 0.0.0.0
--port 8000
--served-model-name kimi-k27
--trust-remote-code
--quantization ascend
--async-scheduling
--tensor-parallel-size 8
--data-parallel-size 2
--data-parallel-size-local 1
--data-parallel-start-rank 0
--data-parallel-address 172.16.0.2
--data-parallel-rpc-port 2358
--enable-expert-parallel
--mm-encoder-tp-mode data
--mm-processor-cache-gb 2
--max-num-seqs 8
--max-model-len 196608
--max-num-batched-tokens 4096
--gpu-memory-utilization 0.9
--enable-prefix-caching
--reasoning-parser kimi_k2
--enable-auto-tool-choice
--tool-call-parser kimi_k2
--limit-mm-per-prompt '{"image":1}'
--override-generation-config '{"max_tokens":24576}'
--compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[8,16]}'
--additional-config '{"ascend_scheduler_config":{"enabled":false},"torchair_graph_config":{"enabled":false},"enable_cpu_binding":true}'
2>&1 | tee "$LOGDIR/serve_k27_dp_master.log"

Contribution

• 我愿意自己提交修复此 bug 的 PR（请先等待维护者在本 issue 中批准）

No repository file or test is named. Start by reproducing the failure with the provided vLLM command and inspect the attached debug session, then trace the provider handling for responses containing only thinking content and tool-call stop details. Done means the reported APIEmptyResponseError path is correctly handled or produces a verified fix with regression coverage.