mirror of
https://github.com/kvcache-ai/ktransformers.git
synced 2026-11-05 01:07:28 +00:00
[GH-ISSUE #331] 推理死循环 #212
Labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
kvcache-ai/ktransformers#212
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @maaaxinfinity on GitHub (Feb 15, 2025).
Original GitHub issue: https://github.com/kvcache-ai/ktransformers/issues/331
使用0.21版本
在尝试默认预设词进行回答后 回答陷入死循环
使用我自己的提示词
同样陷入死循环
pastebin
模型使用671b Q4量化
@maaaxinfinity commented on GitHub (Feb 15, 2025):
我的启动参数 numactl -N 1 -m 1 python3 ./ktransformers/local_chat.py --model_path ./DeepSeek-R1 --gguf_path /mnt/model/DeepSeek-R1 --cpu_infer 18 --cache_lens 1536 --max_new_tokens 4096 --optimize_config_path ./ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat.yaml --force_think true
@3wweiweiwu commented on GitHub (Feb 15, 2025):
我也遇到了同样的问题
Version: 0.21
OS: Ubuntu 22
Compilation: USE_NUMA=1
CPU: AMD EPYC 74F3 36-Core Processor
Memory: 880G
Command:
Workaround: Revert to 0.2. But this degrade performance from 7token/s to 1token/s
@DotIN13 commented on GitHub (Feb 15, 2025):
同样的问题,Llama.cpp没有类似的问题
@KMSorSMS commented on GitHub (Feb 16, 2025):
是MLA部分算子的问题,新commit的版本先默认取消了,应该就不会出这个问题,我们后续会更正算子
@ToviHe commented on GitHub (Feb 17, 2025):
想咨询下,目前使用哪个版本不会出现死循环?我们使用的也是0.2.1
@maaaxinfinity commented on GitHub (Feb 17, 2025):
官方的回复👇
关于0.2.1版本模型降智的问题,我们已经找到bug,大家可以pull fix_precision_MLA分支(PR 413)并在local_chat里玩耍了,我们会在server侧进一步完善并测试MMLU等benchmark后合入主线,非常抱歉之前测试不完全给大家带来了困扰。此次修改也支持了flashinfer MLA kernel,现在的flashinfer还未优化,性能并不比triton实现快(不安装flashinfer就不会启用)。
@3wweiweiwu commented on GitHub (Feb 17, 2025):
You can use 0.2.0
Or you can pull latest code from main branch. It is fixed there also.