deepseek-ai/DeepEP

Fork 0

mirror of https://github.com/deepseek-ai/DeepEP.git synced 2026-11-05 00:47:32 +00:00

[GH-ISSUE #216] Is NVLink necessary? #170

New issue

Closed

opened 2026-10-26 09:55:42 +00:00 by bsakaguchi · 3 comments

bsakaguchi commented

2026-10-26 09:55:42 +00:00

Owner

Originally created by @Xiaofei-fei on GitHub (Jun 17, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/216

Hi teams，
I would like to deploy and run DeepEP on an L20 device. However, I noticed that your repository mentions that NVLink-enabled devices are required when using NVSHMEM. The L20 does not support NVLink, and according to NVIDIA's official documentation, the native NVSHMEM only requires peer-to-peer (P2P) access between devices.

Could you please clarify why NVLink is considered mandatory for DeepEP? Additionally, could you provide any guidance or suggestions on how to run DeepEP properly on L20 devices without NVLink support?

Originally created by @Xiaofei-fei on GitHub (Jun 17, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/216 Hi teams， I would like to deploy and run DeepEP on an L20 device. However, I noticed that your repository mentions that NVLink-enabled devices are required when using NVSHMEM. The L20 does not support NVLink, and according to NVIDIA's official documentation, the native NVSHMEM only requires peer-to-peer (P2P) access between devices. Could you please clarify why NVLink is considered mandatory for DeepEP? Additionally, could you provide any guidance or suggestions on how to run DeepEP properly on L20 devices without NVLink support?

bsakaguchi closed this issue

2026-10-26 09:55:42 +00:00

bsakaguchi commented

2026-10-26 09:55:44 +00:00

Author

Owner

@LyricZhao commented on GitHub (Jun 18, 2025):

Could you please clarify why NVLink is considered mandatory for DeepEP?

For low-latency mode, it is not mandatory. For normal mode, NVLink is a must, because it is designed with V3 models to leverage our cluster's network settings.

Additionally, could you provide any guidance or suggestions on how to run DeepEP properly on L20 devices without NVLink support?

You can use low-latency mode, or you can refer to some other EP impls, e.g. https://github.com/ppl-ai/pplx-kernels. Or if you strongly want to use normal mode, DeepEP must have a refactor.

@LyricZhao commented on GitHub (Jun 18, 2025): > Could you please clarify why NVLink is considered mandatory for DeepEP? For low-latency mode, it is not mandatory. For normal mode, NVLink is a must, because it is designed with V3 models to leverage our cluster's network settings. > Additionally, could you provide any guidance or suggestions on how to run DeepEP properly on L20 devices without NVLink support? You can use low-latency mode, or you can refer to some other EP impls, e.g. https://github.com/ppl-ai/pplx-kernels. Or if you strongly want to use normal mode, DeepEP must have a refactor.

bsakaguchi commented

2026-10-26 09:55:44 +00:00

Author

Owner

@lnykww commented on GitHub (Jul 25, 2025):

Could you please clarify why NVLink is considered mandatory for DeepEP?

For low-latency mode, it is not mandatory. For normal mode, NVLink is a must, because it is designed with V3 models to leverage our cluster's network settings.

Additionally, could you provide any guidance or suggestions on how to run DeepEP properly on L20 devices without NVLink support?

You can use low-latency mode, or you can refer to some other EP impls, e.g. https://github.com/ppl-ai/pplx-kernels. Or if you strongly want to use normal mode, DeepEP must have a refactor.

@LyricZhao Hi，If DeepEP is to support PCIe P2P, apart from NVLink forwarding, what other aspects need to be re-evaluated?

TY！

@lnykww commented on GitHub (Jul 25, 2025): > > Could you please clarify why NVLink is considered mandatory for DeepEP? > > For low-latency mode, it is not mandatory. For normal mode, NVLink is a must, because it is designed with V3 models to leverage our cluster's network settings. > > > Additionally, could you provide any guidance or suggestions on how to run DeepEP properly on L20 devices without NVLink support? > > You can use low-latency mode, or you can refer to some other EP impls, e.g. https://github.com/ppl-ai/pplx-kernels. Or if you strongly want to use normal mode, DeepEP must have a refactor. @LyricZhao Hi，If DeepEP is to support PCIe P2P, apart from NVLink forwarding, what other aspects need to be re-evaluated? TY！

bsakaguchi commented

2026-10-26 09:55:44 +00:00

Author

Owner

@Deeperfinder commented on GitHub (Aug 14, 2025):

+1 ， There are many scenes here that only have PCIe . @LyricZhao

@Deeperfinder commented on GitHub (Aug 14, 2025): +1 ， There are many scenes here that only have PCIe . @LyricZhao

No labels

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

deepseek-ai/DeepEP#170

No description provided.

Rows
Columns