[GH-ISSUE #7] Ada Lovelace support

bsakaguchi commented

2026-10-26 09:52:18 +00:00

Owner

Originally created by @xutizhou on GitHub (Feb 25, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/7

Hi team,
Thank you for your excellent work, I wonder if this repo could support Ada Lovelace architecture such as L20 GPU.

Thanks

Originally created by @xutizhou on GitHub (Feb 25, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/7 Hi team, Thank you for your excellent work, I wonder if this repo could support Ada Lovelace architecture such as L20 GPU. Thanks

bsakaguchi commented

2026-10-26 09:52:18 +00:00

Author

Owner

@NorthSecond commented on GitHub (Feb 25, 2025):

First of all, I'm not a member of the team.

In my understanding, as long as you have cluster environments with RDMA (usually IB NICs and the corresponding software environment ), NVLink between GPUs, and those environments meet the NVSHMEM requirements, it may be usable.

@NorthSecond commented on GitHub (Feb 25, 2025): > First of all, I'm not a member of the team. In my understanding, as long as you have cluster environments with RDMA (usually IB NICs and the corresponding software environment ), NVLink between GPUs, and those environments meet the NVSHMEM requirements, it may be usable.

bsakaguchi commented

2026-10-26 09:52:19 +00:00

Author

Owner

@haswelliris commented on GitHub (Feb 25, 2025):

Could you please confirm whether the Ada Lovelace architecture GPUs support GPU Direct RDMA (GDR) and GPU Direct Async (IBGDA)? If so, DeepEP should also be able to run on this architecture.

@haswelliris commented on GitHub (Feb 25, 2025): Could you please confirm whether the Ada Lovelace architecture GPUs support GPU Direct RDMA (GDR) and GPU Direct Async (IBGDA)? If so, DeepEP should also be able to run on this architecture.

bsakaguchi commented

2026-10-26 09:52:20 +00:00

Author

Owner

@wangzhen2271 commented on GitHub (Mar 11, 2025):

First of all, I'm not a member of the team.

In my understanding, as long as you have cluster environments with RDMA (usually IB NICs and the corresponding software environment ), NVLink between GPUs, and those environments meet the NVSHMEM requirements, it may be usable.

Can not work. NVSHMEM does not rely on NVLink. I've tried it on one node with 8 L20 cards. It just won't run successfully. After running for a while, it will report an error. It seems that a certain kernel execution has gone wrong. Can Lyric Zhao give me some hints?

@wangzhen2271 commented on GitHub (Mar 11, 2025): > > First of all, I'm not a member of the team. > > In my understanding, as long as you have cluster environments with RDMA (usually IB NICs and the corresponding software environment ), NVLink between GPUs, and those environments meet the NVSHMEM requirements, it may be usable. Can not work. NVSHMEM does not rely on NVLink. I've tried it on one node with 8 L20 cards. It just won't run successfully. After running for a while, it will report an error. It seems that a certain kernel execution has gone wrong. Can Lyric Zhao give me some hints? ![Image](https://github.com/user-attachments/assets/0ac31b46-8f7d-46a3-b37f-1cb732e05d3e)

bsakaguchi commented

2026-10-26 09:52:20 +00:00

Author

Owner

@Xiaofei-fei commented on GitHub (Jun 10, 2025):

First of all, I'm not a member of the team.

In my understanding, as long as you have cluster environments with RDMA (usually IB NICs and the corresponding software environment ), NVLink between GPUs, and those environments meet the NVSHMEM requirements, it may be usable.

Can not work. NVSHMEM does not rely on NVLink. I've tried it on one node with 8 L20 cards. It just won't run successfully. After running for a while, it will report an error. It seems that a certain kernel execution has gone wrong. Can Lyric Zhao give me some hints?

Hi，I wonder have you successfully deployed deepep on L20?

@Xiaofei-fei commented on GitHub (Jun 10, 2025): > > > First of all, I'm not a member of the team. > > > > > > In my understanding, as long as you have cluster environments with RDMA (usually IB NICs and the corresponding software environment ), NVLink between GPUs, and those environments meet the NVSHMEM requirements, it may be usable. > > Can not work. NVSHMEM does not rely on NVLink. I've tried it on one node with 8 L20 cards. It just won't run successfully. After running for a while, it will report an error. It seems that a certain kernel execution has gone wrong. Can Lyric Zhao give me some hints? > > ![Image](https://github.com/user-attachments/assets/0ac31b46-8f7d-46a3-b37f-1cb732e05d3e) Hi，I wonder have you successfully deployed deepep on L20?

bsakaguchi commented

2026-10-26 09:52:20 +00:00

Author

Owner

@MengYu10151 commented on GitHub (Aug 19, 2025):

First of all, I'm not a member of the team.

In my understanding, as long as you have cluster environments with RDMA (usually IB NICs and the corresponding software environment ), NVLink between GPUs, and those environments meet the NVSHMEM requirements, it may be usable.

Can not work. NVSHMEM does not rely on NVLink. I've tried it on one node with 8 L20 cards. It just won't run successfully. After running for a while, it will report an error. It seems that a certain kernel execution has gone wrong. Can Lyric Zhao give me some hints?

Hi，I wonder have you successfully deployed deepep on L20?
Hi, @Xiaofei-fei I met the same issue as u, have you deployed it successfully?

@MengYu10151 commented on GitHub (Aug 19, 2025): > > > > First of all, I'm not a member of the team. > > > > > > > > > In my understanding, as long as you have cluster environments with RDMA (usually IB NICs and the corresponding software environment ), NVLink between GPUs, and those environments meet the NVSHMEM requirements, it may be usable. > > > > > > Can not work. NVSHMEM does not rely on NVLink. I've tried it on one node with 8 L20 cards. It just won't run successfully. After running for a while, it will report an error. It seems that a certain kernel execution has gone wrong. Can Lyric Zhao give me some hints? > > ![Image](https://github.com/user-attachments/assets/0ac31b46-8f7d-46a3-b37f-1cb732e05d3e) > > Hi，I wonder have you successfully deployed deepep on L20? Hi, @Xiaofei-fei I met the same issue as u, have you deployed it successfully?

bsakaguchi commented

2026-10-26 09:52:20 +00:00

Author

Owner

@Xiaofei-fei commented on GitHub (Aug 19, 2025):

First of all, I'm not a member of the team.

In my understanding, as long as you have cluster environments with RDMA (usually IB NICs and the corresponding software environment ), NVLink between GPUs, and those environments meet the NVSHMEM requirements, it may be usable.

Can not work. NVSHMEM does not rely on NVLink. I've tried it on one node with 8 L20 cards. It just won't run successfully. After running for a while, it will report an error. It seems that a certain kernel execution has gone wrong. Can Lyric Zhao give me some hints?

Hi，I wonder have you successfully deployed deepep on L20?
Hi, @Xiaofei-fei I met the same issue as u, have you deployed it successfully?

We have resolved most of the issues in intranode mode and can now run together with sglang, but some problems are still being worked on.

@Xiaofei-fei commented on GitHub (Aug 19, 2025): > > > > > First of all, I'm not a member of the team. > > > > > > > > > > > > In my understanding, as long as you have cluster environments with RDMA (usually IB NICs and the corresponding software environment ), NVLink between GPUs, and those environments meet the NVSHMEM requirements, it may be usable. > > > > > > > > > Can not work. NVSHMEM does not rely on NVLink. I've tried it on one node with 8 L20 cards. It just won't run successfully. After running for a while, it will report an error. It seems that a certain kernel execution has gone wrong. Can Lyric Zhao give me some hints? > > > ![Image](https://github.com/user-attachments/assets/0ac31b46-8f7d-46a3-b37f-1cb732e05d3e) > > > > > > Hi，I wonder have you successfully deployed deepep on L20? > > Hi, [@Xiaofei-fei](https://github.com/Xiaofei-fei) I met the same issue as u, have you deployed it successfully? We have resolved most of the issues in intranode mode and can now run together with sglang, but some problems are still being worked on.

bsakaguchi commented

2026-10-26 09:52:20 +00:00

Author

Owner

@Xiaofei-fei commented on GitHub (Sep 8, 2025):

First of all, I'm not a member of the team.

In my understanding, as long as you have cluster environments with RDMA (usually IB NICs and the corresponding software environment ), NVLink between GPUs, and those environments meet the NVSHMEM requirements, it may be usable.

Can not work. NVSHMEM does not rely on NVLink. I've tried it on one node with 8 L20 cards. It just won't run successfully. After running for a while, it will report an error. It seems that a certain kernel execution has gone wrong. Can Lyric Zhao give me some hints?

Hi，I wonder have you successfully deployed deepep on L20?
Hi, @Xiaofei-fei I met the same issue as u, have you deployed it successfully?

btw，I noticed your technical talk on deploying DeepEP on PCIe GPUs, and I am very interested in the idea of merging low-latency and normal-related kernels. Could you please provide a contact so that I can discuss the technical details further?

@Xiaofei-fei commented on GitHub (Sep 8, 2025): > > > > > First of all, I'm not a member of the team. > > > > > > > > > > > > In my understanding, as long as you have cluster environments with RDMA (usually IB NICs and the corresponding software environment ), NVLink between GPUs, and those environments meet the NVSHMEM requirements, it may be usable. > > > > > > > > > Can not work. NVSHMEM does not rely on NVLink. I've tried it on one node with 8 L20 cards. It just won't run successfully. After running for a while, it will report an error. It seems that a certain kernel execution has gone wrong. Can Lyric Zhao give me some hints? > > > ![Image](https://github.com/user-attachments/assets/0ac31b46-8f7d-46a3-b37f-1cb732e05d3e) > > > > > > Hi，I wonder have you successfully deployed deepep on L20? > > Hi, [@Xiaofei-fei](https://github.com/Xiaofei-fei) I met the same issue as u, have you deployed it successfully? btw，I noticed your technical talk on deploying DeepEP on PCIe GPUs, and I am very interested in the idea of merging low-latency and normal-related kernels. Could you please provide a contact so that I can discuss the technical details further?

bsakaguchi commented

2026-10-26 09:52:20 +00:00

Author

Owner

@MengYu10151 commented on GitHub (Sep 8, 2025):

First of all, I'm not a member of the team.

In my understanding, as long as you have cluster environments with RDMA (usually IB NICs and the corresponding software environment ), NVLink between GPUs, and those environments meet the NVSHMEM requirements, it may be usable.

Can not work. NVSHMEM does not rely on NVLink. I've tried it on one node with 8 L20 cards. It just won't run successfully. After running for a while, it will report an error. It seems that a certain kernel execution has gone wrong. Can Lyric Zhao give me some hints?

Hi，I wonder have you successfully deployed deepep on L20?
Hi, @Xiaofei-fei I met the same issue as u, have you deployed it successfully?

btw，I noticed your technical talk on deploying DeepEP on PCIe GPUs, and I am very interested in the idea of merging low-latency and normal-related kernels. Could you please provide a contact so that I can discuss the technical details further?

Really appreciate for your attention to our work！Actually we‘ve already submit a PR to support normal mode w/o NVL https://github.com/deepseek-ai/DeepEP/pull/375 ，and you can contact me via wechat misty10151，thx：）

@MengYu10151 commented on GitHub (Sep 8, 2025): > > > > > > First of all, I'm not a member of the team. > > > > > > > > > > > > > > > In my understanding, as long as you have cluster environments with RDMA (usually IB NICs and the corresponding software environment ), NVLink between GPUs, and those environments meet the NVSHMEM requirements, it may be usable. > > > > > > > > > > > > Can not work. NVSHMEM does not rely on NVLink. I've tried it on one node with 8 L20 cards. It just won't run successfully. After running for a while, it will report an error. It seems that a certain kernel execution has gone wrong. Can Lyric Zhao give me some hints? > > > > ![Image](https://github.com/user-attachments/assets/0ac31b46-8f7d-46a3-b37f-1cb732e05d3e) > > > > > > > > > Hi，I wonder have you successfully deployed deepep on L20? > > > Hi, [@Xiaofei-fei](https://github.com/Xiaofei-fei) I met the same issue as u, have you deployed it successfully? > > btw，I noticed your technical talk on deploying DeepEP on PCIe GPUs, and I am very interested in the idea of merging low-latency and normal-related kernels. Could you please provide a contact so that I can discuss the technical details further? Really appreciate for your attention to our work！Actually we‘ve already submit a PR to support normal mode w/o NVL [https://github.com/deepseek-ai/DeepEP/pull/375](url) ，and you can contact me via wechat misty10151，thx：）

bsakaguchi commented

2026-10-26 09:52:20 +00:00

Author

Owner

@0z5a commented on GitHub (Sep 19, 2026):

I’d like to investigate a bounded Ada/L20 compatibility pass here.
@haswelliris

I noticed that the non-NVLink normal-mode implementation is already present in hybrid-ep through #375, so I am not proposing another PCIe backend.

@0z5a commented on GitHub (Sep 19, 2026): I’d like to investigate a bounded Ada/L20 compatibility pass here. @haswelliris I noticed that the non-NVLink normal-mode implementation is already present in hybrid-ep through #375, so I am not proposing another PCIe backend.

bsakaguchi referenced this issue

2026-10-26 09:53:39 +00:00

[GH-ISSUE #74] test_internode.py hangs at internode.cu:dispatch(...) #68

bsakaguchi referenced this issue

2026-10-26 09:56:29 +00:00

[GH-ISSUE #270] A timeout error occurred when running test_internode.py on four devices. #200

bsakaguchi referenced this issue

2026-10-26 09:56:40 +00:00

[GH-ISSUE #290] Running the test case on a 2-node H20 3E machine failed with internode.cu:1029 'unspecified launch failure' #208

Rows
Columns

[GH-ISSUE #7] Ada Lovelace support #7