[GH-ISSUE #790] [Feature] Native Windows support for multi-node EP on DGX Station GB300 #422

New issue

Open

opened 2026-10-26 10:00:24 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 10:00:24 +00:00

Owner

Originally created by @luhan2017 on GitHub (Oct 22, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/790

I'm from Microsoft, and we're working with NVIDIA on native Windows support for vLLM on NVIDIA GPUs, including Windows ARM64 systems such as the DGX Station GB300. Since each GB300 station has a single GPU, running large MoE models like DeepSeek would involve expert parallelism across multiple GB300 nodes, so DeepEP would naturally be an important piece for us.

We're aware that DeepEP is designed for Linux today and relies on NCCL and RDMA networking, which aren't yet available on native Windows. NVIDIA is looking into the NCCL side, so we simply wanted to give you an early heads-up that this work is starting.

If you happen to have any thoughts, such as things you think could be tricky on Windows or anything we should keep in mind, we'd be very grateful to hear them whenever convenient.

Summary

Context: Microsoft and NVIDIA are bringing vLLM to native Windows, including multi-node expert parallelism across DGX Station GB300 systems.
Dependency: This would rely on NCCL and RDMA networking on Windows, which NVIDIA is looking into.
Ask: Just a heads-up, with no action needed. Any thoughts or concerns you'd like to share would be very welcome.

Thanks,
Lu Han, Microsoft

Originally created by @luhan2017 on GitHub (Oct 22, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/790 I'm from Microsoft, and we're working with NVIDIA on native Windows support for vLLM on NVIDIA GPUs, including Windows ARM64 systems such as the DGX Station GB300. Since each GB300 station has a single GPU, running large MoE models like DeepSeek would involve expert parallelism across multiple GB300 nodes, so DeepEP would naturally be an important piece for us. We're aware that DeepEP is designed for Linux today and relies on NCCL and RDMA networking, which aren't yet available on native Windows. NVIDIA is looking into the NCCL side, so we simply wanted to give you an early heads-up that this work is starting. If you happen to have any thoughts, such as things you think could be tricky on Windows or anything we should keep in mind, we'd be very grateful to hear them whenever convenient. ### Summary - **Context:** Microsoft and NVIDIA are bringing vLLM to native Windows, including multi-node expert parallelism across DGX Station GB300 systems. - **Dependency:** This would rely on NCCL and RDMA networking on Windows, which NVIDIA is looking into. - **Ask:** Just a heads-up, with no action needed. Any thoughts or concerns you'd like to share would be very welcome. Thanks, Lu Han, Microsoft

bsakaguchi referenced this issue

2026-10-27 21:04:02 +00:00

[PR #422] [MERGED] Make dtype of topk_idx configurable #559