[PR #674] [MERGED] Merge some updates #698

New issue

Closed

opened 2026-10-27 21:05:23 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-27 21:05:23 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/674
Author: @sphish
Created: 7/3/2026
Status: ✅ Merged
Merged: 7/3/2026
Merged by: @sphish

Base: main ← Head: update-main

📝 Commits (1)

cec9a2a Update.

📊 Changes

26 files changed (+731 additions, -443 deletions)

View changed files

📝 csrc/elastic/buffer.hpp (+104 -65)
📝 csrc/indexing/main.cu (+0 -1)
📝 csrc/jit/handle.hpp (+3 -3)
📝 csrc/kernels/backend/nccl.cu (+5 -8)
📝 csrc/kernels/backend/symmetric.hpp (+0 -3)
📝 csrc/kernels/elastic/barrier.hpp (+8 -4)
📝 csrc/kernels/elastic/dispatch.hpp (+21 -93)
📝 csrc/kernels/elastic/engram.hpp (+21 -6)
📝 deep_ep/__init__.py (+1 -1)
📝 deep_ep/buffers/elastic.py (+207 -28)
📝 deep_ep/include/deep_ep/common/compiled.cuh (+2 -0)
📝 deep_ep/include/deep_ep/common/handle.cuh (+5 -2)
📝 deep_ep/include/deep_ep/common/ptx.cuh (+21 -0)
📝 deep_ep/include/deep_ep/impls/barrier.cuh (+18 -4)
📝 deep_ep/include/deep_ep/impls/combine.cuh (+10 -4)
📝 deep_ep/include/deep_ep/impls/dispatch.cuh (+6 -0)
📝 deep_ep/include/deep_ep/impls/dispatch_copy_epilogue.cuh (+118 -17)
➖ deep_ep/include/deep_ep/impls/dispatch_deterministic_prologue.cuh (+0 -143)
📝 deep_ep/include/deep_ep/impls/engram_fetch.cuh (+31 -10)
📝 deep_ep/include/deep_ep/impls/hybrid_combine.cuh (+12 -6)

...and 6 more files

📄 Description

Bug Fixes

Fix a data race in the PP TMA pipeline.
Fix a potential hang in Engram fetch (added periodic flush).
Register symmetric memory with strict order flag.
Fix hybrid mode multi-node test synchronization.

Features

Add FP8 support for Engram.
Refactor deterministic dispatch.
Support do_expand for cache mode.

_{🔄 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/674 **Author:** [@sphish](https://github.com/sphish) **Created:** 7/3/2026 **Status:** ✅ Merged **Merged:** 7/3/2026 **Merged by:** [@sphish](https://github.com/sphish) **Base:** `main` ← **Head:** `update-main` --- ### 📝 Commits (1) - [`cec9a2a`](https://github.com/deepseek-ai/DeepEP/commit/cec9a2acc35e38234c4fdfb677d306d6778b5a70) Update. ### 📊 Changes **26 files changed** (+731 additions, -443 deletions) <details> <summary>View changed files</summary> 📝 `csrc/elastic/buffer.hpp` (+104 -65) 📝 `csrc/indexing/main.cu` (+0 -1) 📝 `csrc/jit/handle.hpp` (+3 -3) 📝 `csrc/kernels/backend/nccl.cu` (+5 -8) 📝 `csrc/kernels/backend/symmetric.hpp` (+0 -3) 📝 `csrc/kernels/elastic/barrier.hpp` (+8 -4) 📝 `csrc/kernels/elastic/dispatch.hpp` (+21 -93) 📝 `csrc/kernels/elastic/engram.hpp` (+21 -6) 📝 `deep_ep/__init__.py` (+1 -1) 📝 `deep_ep/buffers/elastic.py` (+207 -28) 📝 `deep_ep/include/deep_ep/common/compiled.cuh` (+2 -0) 📝 `deep_ep/include/deep_ep/common/handle.cuh` (+5 -2) 📝 `deep_ep/include/deep_ep/common/ptx.cuh` (+21 -0) 📝 `deep_ep/include/deep_ep/impls/barrier.cuh` (+18 -4) 📝 `deep_ep/include/deep_ep/impls/combine.cuh` (+10 -4) 📝 `deep_ep/include/deep_ep/impls/dispatch.cuh` (+6 -0) 📝 `deep_ep/include/deep_ep/impls/dispatch_copy_epilogue.cuh` (+118 -17) ➖ `deep_ep/include/deep_ep/impls/dispatch_deterministic_prologue.cuh` (+0 -143) 📝 `deep_ep/include/deep_ep/impls/engram_fetch.cuh` (+31 -10) 📝 `deep_ep/include/deep_ep/impls/hybrid_combine.cuh` (+12 -6) _...and 6 more files_ </details> ### 📄 Description ## Bug Fixes - Fix a data race in the PP TMA pipeline. - Fix a potential hang in Engram fetch (added periodic flush). - Register symmetric memory with strict order flag. - Fix hybrid mode multi-node test synchronization. ## Features - Add FP8 support for Engram. - Refactor deterministic dispatch. - Support `do_expand` for cache mode. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-27 21:05:23 +00:00