kvcache-ai/Mooncake

Fork 0

mirror of https://github.com/kvcache-ai/Mooncake.git synced 2026-11-05 01:17:55 +00:00

[GH-ISSUE #3048] [Bug]: 有人遇到过模型答非所问吗 #782

New issue

Open

opened 2026-10-29 21:19:28 +00:00 by bsakaguchi · 7 comments

bsakaguchi commented

2026-10-29 21:19:28 +00:00

Owner

Originally created by @longyuetang on GitHub (Jul 22, 2026).
Original GitHub issue: https://github.com/kvcache-ai/Mooncake/issues/3048

Originally assigned to: @Lin-z-w on GitHub.

Bug Report

vllm版本：v0.23.0
GPU：NVIDIA H20
模型：deepseek v4 flash
mooncake版本：0.3.10.post1
配置上mooncake store作为kvcache卸载，出现了大量的答非所谓，重启ai-coding agent又没出现了
另外的vllm实例没有配置mooncake就没有出，该怎么排查啊

Before submitting...

Ensure you searched for relevant issues and read the [documentation]

Originally created by @longyuetang on GitHub (Jul 22, 2026). Original GitHub issue: https://github.com/kvcache-ai/Mooncake/issues/3048 Originally assigned to: @Lin-z-w on GitHub. ### Bug Report vllm版本：v0.23.0 GPU：NVIDIA H20 模型：deepseek v4 flash mooncake版本：0.3.10.post1 配置上mooncake store作为kvcache卸载，出现了大量的答非所谓，重启ai-coding agent又没出现了另外的vllm实例没有配置mooncake就没有出，该怎么排查啊 ### Before submitting... - [ ] Ensure you searched for relevant issues and read the [documentation]

bsakaguchi added the

stale

bug

labels

2026-10-29 21:19:28 +00:00

bsakaguchi commented

2026-10-29 21:19:28 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Jul 22, 2026):

Thanks for opening this issue, @longyuetang!

Field	Value
Issue	#3048
GitHub user ID	`13553842`
Reporter	@longyuetang

A maintainer will triage this when possible. To help us respond faster, please include:

Mooncake version or commit SHA
Environment (OS, CUDA/driver, RDMA stack if relevant)
Steps to reproduce and expected vs. actual behavior

Useful links: Documentation · Contributing guide

This message was posted automatically by the issue bot.

@github-actions[bot] commented on GitHub (Jul 22, 2026): Thanks for opening this issue, @longyuetang! | Field | Value | |-------|-------| | **Issue** | #3048 | | **GitHub user ID** | `13553842` | | **Reporter** | @longyuetang | A maintainer will triage this when possible. To help us respond faster, please include: - Mooncake version or commit SHA - Environment (OS, CUDA/driver, RDMA stack if relevant) - Steps to reproduce and expected vs. actual behavior Useful links: [Documentation](https://kvcache-ai.github.io/Mooncake/) · [Contributing guide](https://github.com/kvcache-ai/Mooncake/blob/main/CONTRIBUTING.md) > This message was posted automatically by the issue bot.

bsakaguchi commented

2026-10-29 21:19:29 +00:00

Author

Owner

@longyuetang commented on GitHub (Jul 22, 2026):

mooncake master大量打印key已经存在：

@longyuetang commented on GitHub (Jul 22, 2026): mooncake master大量打印key已经存在： <img width="804" height="96" alt="Image" src="https://github.com/user-attachments/assets/df03aa03-e409-4fa0-9b9c-94a5fcd75501" />

bsakaguchi commented

2026-10-29 21:19:29 +00:00

Author

Owner

@Lin-z-w commented on GitHub (Jul 23, 2026):

mooncake master大量打印key已经存在：

请问是否能够提供完整 vLLM 启动参数、kv_transfer_config 和 Mooncake 配置或者是答非所问发生时的对应日志？

object_already_exists 本身不等于故障，vLLM 会先检查 key、再写入；多个实例并发保存相同前缀时可能同时判断“不存在”，随后其中一个写入成功，其他写入触发该日志。Mooncake v0.3.10.post1 会把这种重复写视为成功，这是内容寻址去重的正常路径。

@Lin-z-w commented on GitHub (Jul 23, 2026): > mooncake master大量打印key已经存在： > > <img alt="Image" width="804" height="96" src="https://private-user-images.githubusercontent.com/13553842/625033827-df03aa03-e409-4fa0-9b9c-94a5fcd75501.png?jwt=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.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.xncScuBbz2zkbtubisgnMNe-1YLlEeVUmL_0uUFoYXI"> 请问是否能够提供完整 vLLM 启动参数、kv_transfer_config 和 Mooncake 配置或者是答非所问发生时的对应日志？ object_already_exists 本身不等于故障，vLLM 会先检查 key、再写入；多个实例并发保存相同前缀时可能同时判断“不存在”，随后其中一个写入成功，其他写入触发该日志。Mooncake v0.3.10.post1 会把这种重复写视为成功，这是内容寻址去重的正常路径。

bsakaguchi commented

2026-10-29 21:19:29 +00:00

Author

Owner

@longyuetang commented on GitHub (Jul 24, 2026):

vllm的启动参数：
vllm serve /data/models/DeepSeek-V4-Flash/ --trust-remote-code --max-num-batched-tokens 8192 --max-num-seqs 128 --tensor-parallel-size 4 --port 8900 --enable-chunked-prefill --enable-prefix-caching --quantization fp8 --tokenizer-mode deepseek_v4 --tool-call-parser deepseek_v4 --enable-auto-tool-choice --reasoning-parser deepseek_v4 --enable-prefix-caching --served-model-name v4_flash --max-model-len 262144 --block-size 256 –kv-transfer-config '{"kv_connector:MooncakeStoreConnector", "kv_role":"kv_both"}'

mooncake配置文件：
{
“mode”: “embedded”,
“metadata_server”: “P2PHANDSHAKE”,
“master_server_address”: “localhost:50063”,
“global_segment_size”: “50GB”,
“local_buffer_size”: “4GB”,
“protocol”: “rdma”,
“device_name”: “mlx5_11”,
“enable_offload”: “false”
}
mooncake_master没有任何异常日志，除了object_already_exists
vllm日志只有 400 Bad Request的打印，其他没有任何异常打印
@Lin-z-w

@longyuetang commented on GitHub (Jul 24, 2026): vllm的启动参数： vllm serve /data/models/DeepSeek-V4-Flash/ --trust-remote-code --max-num-batched-tokens 8192 --max-num-seqs 128 --tensor-parallel-size 4 --port 8900 --enable-chunked-prefill --enable-prefix-caching --quantization fp8 --tokenizer-mode deepseek_v4 --tool-call-parser deepseek_v4 --enable-auto-tool-choice --reasoning-parser deepseek_v4 --enable-prefix-caching --served-model-name v4_flash --max-model-len 262144 --block-size 256 –kv-transfer-config '{"kv_connector:MooncakeStoreConnector", "kv_role":"kv_both"}' mooncake配置文件： { “mode”: “embedded”, “metadata_server”: “P2PHANDSHAKE”, “master_server_address”: “localhost:50063”, “global_segment_size”: “50GB”, “local_buffer_size”: “4GB”, “protocol”: “rdma”, “device_name”: “mlx5_11”, “enable_offload”: “false” } mooncake_master没有任何异常日志，除了object_already_exists vllm日志只有 400 Bad Request的打印，其他没有任何异常打印 @Lin-z-w

bsakaguchi commented

2026-10-29 21:19:30 +00:00

Author

Owner

@Lin-z-w commented on GitHub (Jul 24, 2026):

vllm的启动参数： vllm serve /data/models/DeepSeek-V4-Flash/ --trust-remote-code --max-num-batched-tokens 8192 --max-num-seqs 128 --tensor-parallel-size 4 --port 8900 --enable-chunked-prefill --enable-prefix-caching --quantization fp8 --tokenizer-mode deepseek_v4 --tool-call-parser deepseek_v4 --enable-auto-tool-choice --reasoning-parser deepseek_v4 --enable-prefix-caching --served-model-name v4_flash --max-model-len 262144 --block-size 256 –kv-transfer-config '{"kv_connector:MooncakeStoreConnector", "kv_role":"kv_both"}'

mooncake配置文件： { “mode”: “embedded”, “metadata_server”: “P2PHANDSHAKE”, “master_server_address”: “localhost:50063”, “global_segment_size”: “50GB”, “local_buffer_size”: “4GB”, “protocol”: “rdma”, “device_name”: “mlx5_11”, “enable_offload”: “false” } mooncake_master没有任何异常日志，除了object_already_exists vllm日志只有 400 Bad Request的打印，其他没有任何异常打印 @Lin-z-w

感谢补充。你贴出的命令中，–kv-transfer-config 使用了 Unicode 长横线，JSON 里的 kv_connector 也存在语法错误；配置文件中的引号看起来还是中文弯引号。如果实际命令/文件确实如此，程序应该在启动阶段直接报参数或 JSON 解析错误并退出，不会正常运行后再导致答非所问，因此更可能是粘贴时格式发生了变化。正确参数应类似：

--kv-transfer-config '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both"}'

麻烦直接提供启动脚本原文，或运行中进程的 /proc/<pid>/cmdline，避免手工转写。另外，目前最关键的异常是 vLLM 返回的 400 Bad Request，请提供问题请求对应的完整 400 响应 body 及时间/request ID。400 通常意味着请求在进入推理前就因上下文长度、tool message 或请求格式等原因被拒绝；重启 agent 后恢复，也可能是因为会话历史被清空。object_already_exists 本身仍可能只是正常的对象去重，暂时不足以说明 KV 数据发生了损坏。

@Lin-z-w commented on GitHub (Jul 24, 2026): > vllm的启动参数： vllm serve /data/models/DeepSeek-V4-Flash/ --trust-remote-code --max-num-batched-tokens 8192 --max-num-seqs 128 --tensor-parallel-size 4 --port 8900 --enable-chunked-prefill --enable-prefix-caching --quantization fp8 --tokenizer-mode deepseek_v4 --tool-call-parser deepseek_v4 --enable-auto-tool-choice --reasoning-parser deepseek_v4 --enable-prefix-caching --served-model-name v4_flash --max-model-len 262144 --block-size 256 –kv-transfer-config '{"kv_connector:MooncakeStoreConnector", "kv_role":"kv_both"}' > > mooncake配置文件： { “mode”: “embedded”, “metadata_server”: “P2PHANDSHAKE”, “master_server_address”: “localhost:50063”, “global_segment_size”: “50GB”, “local_buffer_size”: “4GB”, “protocol”: “rdma”, “device_name”: “mlx5_11”, “enable_offload”: “false” } mooncake_master没有任何异常日志，除了object_already_exists vllm日志只有 400 Bad Request的打印，其他没有任何异常打印 [@Lin-z-w](https://github.com/Lin-z-w) 感谢补充。你贴出的命令中，`–kv-transfer-config` 使用了 Unicode 长横线，JSON 里的 `kv_connector` 也存在语法错误；配置文件中的引号看起来还是中文弯引号。如果实际命令/文件确实如此，程序应该在启动阶段直接报参数或 JSON 解析错误并退出，不会正常运行后再导致答非所问，因此更可能是粘贴时格式发生了变化。正确参数应类似： ```bash --kv-transfer-config '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both"}' ``` 麻烦直接提供启动脚本原文，或运行中进程的 `/proc/<pid>/cmdline`，避免手工转写。另外，目前最关键的异常是 vLLM 返回的 `400 Bad Request`，请提供问题请求对应的完整 400 响应 body 及时间/request ID。400 通常意味着请求在进入推理前就因上下文长度、tool message 或请求格式等原因被拒绝；重启 agent 后恢复，也可能是因为会话历史被清空。`object_already_exists` 本身仍可能只是正常的对象去重，暂时不足以说明 KV 数据发生了损坏。

bsakaguchi commented

2026-10-29 21:19:30 +00:00

Author

Owner

@longyuetang commented on GitHub (Jul 27, 2026):

@Lin-z-w 多谢回复，确实是我转写错了，正确的就是你纠正后的，抱歉。我也怀疑可能是这个400，如果复现了就追踪一下这个Request

@longyuetang commented on GitHub (Jul 27, 2026): @Lin-z-w 多谢回复，确实是我转写错了，正确的就是你纠正后的，抱歉。我也怀疑可能是这个400，如果复现了就追踪一下这个Request

bsakaguchi commented

2026-10-29 21:19:30 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Oct 25, 2026):

This issue has had no activity for 90 days and will be closed in 7 days if there is no further activity. Please comment or react if it should stay open.

@github-actions[bot] commented on GitHub (Oct 25, 2026): This issue has had no activity for 90 days and will be closed in 7 days if there is no further activity. Please comment or react if it should stay open.