[GH-ISSUE #1761] 以原精度运行M2.1时占用内存异常，达到366G #1111

New issue

Open

opened 2026-10-29 14:48:41 +00:00 by bsakaguchi · 2 comments

bsakaguchi commented

2026-10-29 14:48:41 +00:00

Owner

Originally created by @CodeZ-Hao on GitHub (Dec 27, 2025).
Original GitHub issue: https://github.com/kvcache-ai/ktransformers/issues/1761

Originally assigned to: @ouqingliang on GitHub.

Reminder

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

System Info

GPU: 4080 32G
MEM: 384GB
CPU: 8461V*1

KT版本: 0.5.0 commit: 63796374c1
sglang: kvcache-ai/sglang.git main分支 eaf832efc2ecbdd0b12f691fdf2608d4baf5b3c5
cuda版本: 12.6

Reproduction

启动命令：

CUDA_VISIBLE_DEVICES=0 python3 -m sglang.launch_server --host 0.0.0.0 --port 6399 --model /models/MiniMax/MiniMax-M2.1 --kt-weight-path /models/MiniMax/MiniMax-M2.1 --kt-cpuinfer 46 --kt-threadpool-count 1 --kt-num-gpu-experts 2 --kt-method FP8 --kt-gpu-prefill-token-threshold 100 --attention-backend flashinfer --trust-remote-code --mem-fraction-static 0.98 --chunked-prefill-size 32768 --max-running-requests 1 --max-total-tokens 65535 --served-model-name MiniMax-M2 --disable-shared-experts-fusion --tool-call-parser minimax-m2 --reasoning-parser minimax-append-think

异常描述

服务启动后，使用htop查看内存占用，发现sglang进程占用内存达到366G，远超模型大小。
且确认在停止服务后，内存占用为0.3%，因此可确认没有其他进程占用内存，确实是sglang占用。

预期现象

预期占用应当在240G左右

Others

No response

Originally created by @CodeZ-Hao on GitHub (Dec 27, 2025). Original GitHub issue: https://github.com/kvcache-ai/ktransformers/issues/1761 Originally assigned to: @ouqingliang on GitHub. ### Reminder - [x] I have read the above rules and searched the existing issues. ### System Info GPU: 4080 32G MEM: 384GB CPU: 8461V*1 KT版本: 0.5.0 commit: 63796374c110e14464c5540324a21a658c5f1cf8 sglang: kvcache-ai/sglang.git main分支 eaf832efc2ecbdd0b12f691fdf2608d4baf5b3c5 cuda版本: 12.6 ### Reproduction ### 启动命令： ```shell CUDA_VISIBLE_DEVICES=0 python3 -m sglang.launch_server --host 0.0.0.0 --port 6399 --model /models/MiniMax/MiniMax-M2.1 --kt-weight-path /models/MiniMax/MiniMax-M2.1 --kt-cpuinfer 46 --kt-threadpool-count 1 --kt-num-gpu-experts 2 --kt-method FP8 --kt-gpu-prefill-token-threshold 100 --attention-backend flashinfer --trust-remote-code --mem-fraction-static 0.98 --chunked-prefill-size 32768 --max-running-requests 1 --max-total-tokens 65535 --served-model-name MiniMax-M2 --disable-shared-experts-fusion --tool-call-parser minimax-m2 --reasoning-parser minimax-append-think ``` ### 异常描述服务启动后，使用htop查看内存占用，发现sglang进程占用内存达到366G，远超模型大小。且确认在停止服务后，内存占用为0.3%，因此可确认没有其他进程占用内存，确实是sglang占用。 ### 预期现象预期占用应当在240G左右 ### Others _No response_

bsakaguchi added the

bug

label

2026-10-29 14:48:41 +00:00

bsakaguchi commented

2026-10-29 14:48:41 +00:00

Author

Owner

@ErvinXie commented on GitHub (Dec 28, 2025):

This is really not the expected behavior, we also get some feedback from our wechat group. We are going to figure out what is going on.

@ErvinXie commented on GitHub (Dec 28, 2025): This is really not the expected behavior, we also get some feedback from our wechat group. We are going to figure out what is going on.

bsakaguchi commented

2026-10-29 14:48:41 +00:00

Author

Owner

@ovowei commented on GitHub (Jan 22, 2026):

@ouqingliang

@ovowei commented on GitHub (Jan 22, 2026): @ouqingliang