[GH-ISSUE #252] About the concept of channel in DeepEP and NCCL #187

New issue

Open

opened 2026-10-26 09:56:11 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 09:56:11 +00:00

Owner

Originally created by @lucky0218 on GitHub (Jun 24, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/252

I'm confused about the concept of ‘channels’ used in both DeepEP and NCCL. Why do we isolate communication resources using channels? In other words, why not configure a single large channel to handle all data transfers by allocating more resources to it?

What specific resource isolation exists between different channels? What benefits do channels provide?

Additionally, why doesn’t the LL (Low Latency) mode require channel partitioning?

Any clarification you could provide on these points would be incredibly helpful for my understanding. Thank you for your time on the project.

Originally created by @lucky0218 on GitHub (Jun 24, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/252 I'm confused about the concept of ‘channels’ used in both DeepEP and NCCL. Why do we isolate communication resources using channels? In other words, why not configure a single large channel to handle all data transfers by allocating more resources to it? What specific resource isolation exists between different channels? What benefits do channels provide? Additionally, why doesn’t the LL (Low Latency) mode require channel partitioning? Any clarification you could provide on these points would be incredibly helpful for my understanding. Thank you for your time on the project.

bsakaguchi commented

2026-10-26 09:56:12 +00:00

Author

Owner

@LyricZhao commented on GitHub (Jun 27, 2025):

In a single channel, the tokens are sent in a specific order.

If multiple SMs share the same channel, you have to do extra synchronization between SMs, which is inefficient.

@LyricZhao commented on GitHub (Jun 27, 2025): In a single channel, the tokens are sent in a specific order. If multiple SMs share the same channel, you have to do extra synchronization between SMs, which is inefficient.

bsakaguchi referenced this issue

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

[PR #187] [CLOSED] Support other ways of fp8 quantization #453