[PR #434] Fix get_nvl_buffer_size_hint and get_rdma_buffer_size_hint #567

Open
opened 2026-10-27 21:04:07 +00:00 by bsakaguchi · 0 comments
Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/434
Author: @yuantailing
Created: 9/26/2025
Status: 🔄 Open

Base: main ← Head: fix_buffer_size


📝 Commits (1)

  • 74cfcc2 Fix get_nvl_buffer_size_hint and get_rdma_buffer_size_hint

📊 Changes

1 file changed (+5 additions, -3 deletions)

View changed files

📝 csrc/config.hpp (+5 -3)

📄 Description

If I understand correctly,

  1. L62 is corresponding to github.com/deepseek-ai/DeepEP@e02e4d2e1f/csrc/kernels/internode.cu (L416)
  2. L67 is corresponding to github.com/deepseek-ai/DeepEP@e02e4d2e1f/csrc/kernels/internode.cu (L579-L581) , and it is also counted as sizeof(int) in get_num_bytes_per_token
  3. L70 is corresponding to int4 alignment: github.com/deepseek-ai/DeepEP@e02e4d2e1f/csrc/kernels/internode.cu (L45-L47)
  4. L94 is similar to L67
  5. L98 is corresponding to github.com/deepseek-ai/DeepEP@e02e4d2e1f/csrc/kernels/internode.cu (L417-L418)

Before change:

>>> deep_ep.Config(24, 8, 512, 16, 128).get_nvl_buffer_size_hint(7168, 16)
453380736
>>> deep_ep.Config(24, 8, 512, 16, 128).get_rdma_buffer_size_hint(7168, 16)
56774016

After change:

>>> deep_ep.Config(24, 8, 512, 16, 128).get_nvl_buffer_size_hint(7168, 16)
429000960
>>> deep_ep.Config(24, 8, 512, 16, 128).get_rdma_buffer_size_hint(7168, 16)
53629056

🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.

## 📋 Pull Request Information **Original PR:** https://github.com/deepseek-ai/DeepEP/pull/434 **Author:** [@yuantailing](https://github.com/yuantailing) **Created:** 9/26/2025 **Status:** 🔄 Open **Base:** `main` ← **Head:** `fix_buffer_size` --- ### 📝 Commits (1) - [`74cfcc2`](https://github.com/deepseek-ai/DeepEP/commit/74cfcc2c19ce599e608511227deb0db3621ff28c) Fix get_nvl_buffer_size_hint and get_rdma_buffer_size_hint ### 📊 Changes **1 file changed** (+5 additions, -3 deletions) <details> <summary>View changed files</summary> 📝 `csrc/config.hpp` (+5 -3) </details> ### 📄 Description If I understand correctly, 1. L62 is corresponding to https://github.com/deepseek-ai/DeepEP/blob/e02e4d2e1fbfdf09e02e870b6acc5831cbd11e39/csrc/kernels/internode.cu#L416 2. L67 is corresponding to https://github.com/deepseek-ai/DeepEP/blob/e02e4d2e1fbfdf09e02e870b6acc5831cbd11e39/csrc/kernels/internode.cu#L579-L581 , and it is also counted as `sizeof(int)` in `get_num_bytes_per_token` 3. L70 is corresponding to int4 alignment: https://github.com/deepseek-ai/DeepEP/blob/e02e4d2e1fbfdf09e02e870b6acc5831cbd11e39/csrc/kernels/internode.cu#L45-L47 4. L94 is similar to L67 5. L98 is corresponding to https://github.com/deepseek-ai/DeepEP/blob/e02e4d2e1fbfdf09e02e870b6acc5831cbd11e39/csrc/kernels/internode.cu#L417-L418 Before change: ``` >>> deep_ep.Config(24, 8, 512, 16, 128).get_nvl_buffer_size_hint(7168, 16) 453380736 >>> deep_ep.Config(24, 8, 512, 16, 128).get_rdma_buffer_size_hint(7168, 16) 56774016 ``` After change: ``` >>> deep_ep.Config(24, 8, 512, 16, 128).get_nvl_buffer_size_hint(7168, 16) 429000960 >>> deep_ep.Config(24, 8, 512, 16, 128).get_rdma_buffer_size_hint(7168, 16) 53629056 ``` --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>
Sign in to join this conversation.
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#567
No description provided.