[PR #532] Fix TMA assertion for non-32-divisible hidden sizes in combine_token #615

New issue

Open

opened 2026-10-27 21:04:33 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-27 21:04:33 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepEP/pull/532
Author: @yurekami
Created: 12/25/2025
Status: 🔄 Open

Base: main ← Head: fix/tma-hidden-size-compatibility

📝 Commits (1)

6e98f65 Fix TMA assertion for non-32-divisible hidden sizes in combine_token

📊 Changes

1 file changed (+37 additions, -15 deletions)

View changed files

📝 csrc/kernels/internode.cu (+37 -15)

📄 Description

Summary

Add runtime check to use TMA only when hidden_int4 % 32 == 0
For hidden sizes like 2880 (hidden_int4=180), fall back to non-TMA path
Preserves TMA performance optimization for compatible hidden sizes

Test plan

Run python tests/test_internode.py --hidden 2880 in multi-node setup
Verify no assertion failures for GPTOSS hidden size
Verify TMA still works for standard hidden sizes (e.g., 7168)

Fixes #506

🤖 Generated with Claude Code

_{🔄 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/532 **Author:** [@yurekami](https://github.com/yurekami) **Created:** 12/25/2025 **Status:** 🔄 Open **Base:** `main` ← **Head:** `fix/tma-hidden-size-compatibility` --- ### 📝 Commits (1) - [`6e98f65`](https://github.com/deepseek-ai/DeepEP/commit/6e98f655867080c6127e8d6892cc6f04934b982e) Fix TMA assertion for non-32-divisible hidden sizes in combine_token ### 📊 Changes **1 file changed** (+37 additions, -15 deletions) <details> <summary>View changed files</summary> 📝 `csrc/kernels/internode.cu` (+37 -15) </details> ### 📄 Description ## Summary - Add runtime check to use TMA only when hidden_int4 % 32 == 0 - For hidden sizes like 2880 (hidden_int4=180), fall back to non-TMA path - Preserves TMA performance optimization for compatible hidden sizes ## Test plan - [ ] Run `python tests/test_internode.py --hidden 2880` in multi-node setup - [ ] Verify no assertion failures for GPTOSS hidden size - [ ] Verify TMA still works for standard hidden sizes (e.g., 7168) Fixes #506 🤖 Generated with [Claude Code](https://claude.com/claude-code) --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi added the

pull-request

label

2026-10-27 21:04:33 +00:00

bsakaguchi referenced this issue

2026-10-27 21:06:13 +00:00

[PR #771] Add official PyTorch-based Dockerfile for DeepEP #772