[PR #746] [MERGED] [HybridEP] Keep overflow flag last in ragged permute handles #751

New issue

Closed

opened 2026-10-27 21:05:53 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-27 21:05:53 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/746
Author: @Autumn1998
Created: 9/9/2026
Status: ✅ Merged
Merged: 9/11/2026
Merged by: @jershi425

Base: hybrid-ep ← Head: fix/hybridep-overflow-handle

📝 Commits (2)

e65c2a9 [HybridEP] Keep overflow flag last in ragged permute handles
56158d0 [HybridEP] Remove extra handle layout assertions from ragged tests

📊 Changes

2 files changed (+7 additions, -7 deletions)

View changed files

📝 deep_ep/hybrid_ep_buffer.py (+5 -5)
📝 docs/Hybrid-EP_Implementation.md (+2 -2)

📄 Description

Preserve downstream handle[-1] overflow checks by placing the valid-token count before the overflow flag. Update cached dispatch and combine unpacking, document the layout, and assert the contract in the ragged regression test.

_{🔄 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/746 **Author:** [@Autumn1998](https://github.com/Autumn1998) **Created:** 9/9/2026 **Status:** ✅ Merged **Merged:** 9/11/2026 **Merged by:** [@jershi425](https://github.com/jershi425) **Base:** `hybrid-ep` ← **Head:** `fix/hybridep-overflow-handle` --- ### 📝 Commits (2) - [`e65c2a9`](https://github.com/deepseek-ai/DeepEP/commit/e65c2a9088b757a6c4d73793632f21421673bcfd) [HybridEP] Keep overflow flag last in ragged permute handles - [`56158d0`](https://github.com/deepseek-ai/DeepEP/commit/56158d03cb63ac3ed222ccc6a76c313c36b79600) [HybridEP] Remove extra handle layout assertions from ragged tests ### 📊 Changes **2 files changed** (+7 additions, -7 deletions) <details> <summary>View changed files</summary> 📝 `deep_ep/hybrid_ep_buffer.py` (+5 -5) 📝 `docs/Hybrid-EP_Implementation.md` (+2 -2) </details> ### 📄 Description Preserve downstream handle[-1] overflow checks by placing the valid-token count before the overflow flag. Update cached dispatch and combine unpacking, document the layout, and assert the contract in the ragged regression test. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-27 21:05:53 +00:00