[GH-ISSUE #1650] DeepSeekV3.2 乱码 #1062

New issue

Open

opened 2026-10-29 14:48:05 +00:00 by bsakaguchi · 2 comments

bsakaguchi commented

2026-10-29 14:48:05 +00:00

Owner

Originally created by @ErvinXie on GitHub (Nov 28, 2025).
Original GitHub issue: https://github.com/kvcache-ai/ktransformers/issues/1650

Originally assigned to: @ovowei on GitHub.

Reminder

I have read the above rules and searched the existing issues.

System Info

AMX_METHOD=AMXINT8 SGLANG_ENABLE_JIT_DEEPGEMM=0 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True python -m sglang.launch_server --host 0.0.0.0 --port 60000 --model /mnt/ktd/v3.2exp/ --kt-weight-path ~/v3.2cpu/ --kt-cpuinfer 116 --kt-threadpool-count 2 --kt-num-gpu-experts 5 --kt-method AMXINT8 --attention-backend flashinfer --trust-remote-code --mem-fraction-static 0.5 --chunked-prefill-size 1024 --max-running-requests 4 --enable-mixed-chunk --tensor-parallel-size 1 --enable-p2p-check --disable-shared-experts-fusion --served-model-name kimi_k2 --tool-call-parser deepseekv3 --kt-max-deferred-experts-per-token 7

运行命令是这个。

Originally created by @ErvinXie on GitHub (Nov 28, 2025). Original GitHub issue: https://github.com/kvcache-ai/ktransformers/issues/1650 Originally assigned to: @ovowei on GitHub. ### Reminder - [x] I have read the above rules and searched the existing issues. ### System Info <img width="1241" height="944" alt="Image" src="https://github.com/user-attachments/assets/a27536ed-a556-46b0-a4b5-7df206d58117" /> `AMX_METHOD=AMXINT8 SGLANG_ENABLE_JIT_DEEPGEMM=0 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True python -m sglang.launch_server --host 0.0.0.0 --port 60000 --model /mnt/ktd/v3.2exp/ --kt-weight-path ~/v3.2cpu/ --kt-cpuinfer 116 --kt-threadpool-count 2 --kt-num-gpu-experts 5 --kt-method AMXINT8 --attention-backend flashinfer --trust-remote-code --mem-fraction-static 0.5 --chunked-prefill-size 1024 --max-running-requests 4 --enable-mixed-chunk --tensor-parallel-size 1 --enable-p2p-check --disable-shared-experts-fusion --served-model-name kimi_k2 --tool-call-parser deepseekv3 --kt-max-deferred-experts-per-token 7` 运行命令是这个。 <img width="1034" height="970" alt="Image" src="https://github.com/user-attachments/assets/595c2583-f0ca-46da-8094-b65b96b2ac47" />

bsakaguchi added the

bug

label

2026-10-29 14:48:05 +00:00

bsakaguchi commented

2026-10-29 14:48:06 +00:00

Author

Owner

@shrould8888 commented on GitHub (Dec 1, 2025):

今天发布了DeepSeek-V3.2正式版和DeepSeek-V3.2-Speciale，用ktransformer 应该怎样跑（指令是什么，我的是XEON5+768GB+双4090）

@shrould8888 commented on GitHub (Dec 1, 2025): 今天发布了DeepSeek-V3.2正式版和DeepSeek-V3.2-Speciale，用ktransformer 应该怎样跑（指令是什么，我的是XEON5+768GB+双4090）

bsakaguchi commented

2026-10-29 14:48:06 +00:00

Author

Owner

@ovowei commented on GitHub (Dec 2, 2025):

https://github.com/kvcache-ai/ktransformers/blob/main/doc/en/kt-kernel/deepseek-v3.2-sglang-tutorial.md

@ovowei commented on GitHub (Dec 2, 2025): https://github.com/kvcache-ai/ktransformers/blob/main/doc/en/kt-kernel/deepseek-v3.2-sglang-tutorial.md