[PR #422] [Feat] add support sm90 fp8 mega moe #434

New issue

Open

opened 2026-10-26 21:17:45 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 21:17:45 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepGEMM/pull/422
Author: @lengrongfu
Created: 8/27/2026
Status: 🔄 Open

Base: nv_dev ← Head: claude/pr36-mega-moe-port

📝 Commits (3)

61897f9 Add SM90 (Hopper) MegaMoE support
ae4228a Fix missing 'template' disambiguator in grid_sync dependent-type call
11885a0 add hidden size check to 256

📊 Changes

11 files changed (+3430 additions, -8 deletions)

View changed files

📝 csrc/apis/mega.hpp (+225 -0)
➕ csrc/jit_kernels/heuristics/sm90_mega_moe.hpp (+236 -0)
➕ csrc/jit_kernels/impls/sm90_fp8_mega_moe.hpp (+307 -0)
📝 csrc/utils/layout.hpp (+4 -3)
📝 deep_gemm/__init__.py (+5 -0)
📝 deep_gemm/include/deep_gemm/comm/barrier.cuh (+20 -5)
➕ deep_gemm/include/deep_gemm/impls/sm90_fp8_mega_moe.cuh (+1936 -0)
➕ deep_gemm/include/deep_gemm/layout/sm90_mega_moe.cuh (+179 -0)
➕ deep_gemm/include/deep_gemm/scheduler/sm90_mega_moe.cuh (+235 -0)
📝 deep_gemm/mega/__init__.py (+111 -0)
➕ tests/test_mega_moe_hopper.py (+172 -0)

📄 Description

Ref: https://github.com/sgl-project/DeepGEMM/pull/36

vllm support deepseek-v4-flash-fp8 model PR: https://github.com/vllm-project/vllm/pull/53527

_{🔄 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/DeepGEMM/pull/422 **Author:** [@lengrongfu](https://github.com/lengrongfu) **Created:** 8/27/2026 **Status:** 🔄 Open **Base:** `nv_dev` ← **Head:** `claude/pr36-mega-moe-port` --- ### 📝 Commits (3) - [`61897f9`](https://github.com/deepseek-ai/DeepGEMM/commit/61897f9b5aa1ce5524cfdcaaa03c8955a8ab645a) Add SM90 (Hopper) MegaMoE support - [`ae4228a`](https://github.com/deepseek-ai/DeepGEMM/commit/ae4228ad7acb34f39f3daa7a0633333d044b9921) Fix missing 'template' disambiguator in grid_sync dependent-type call - [`11885a0`](https://github.com/deepseek-ai/DeepGEMM/commit/11885a0e83738b68f226a80bffe1efdc83dc5364) add hidden size check to 256 ### 📊 Changes **11 files changed** (+3430 additions, -8 deletions) <details> <summary>View changed files</summary> 📝 `csrc/apis/mega.hpp` (+225 -0) ➕ `csrc/jit_kernels/heuristics/sm90_mega_moe.hpp` (+236 -0) ➕ `csrc/jit_kernels/impls/sm90_fp8_mega_moe.hpp` (+307 -0) 📝 `csrc/utils/layout.hpp` (+4 -3) 📝 `deep_gemm/__init__.py` (+5 -0) 📝 `deep_gemm/include/deep_gemm/comm/barrier.cuh` (+20 -5) ➕ `deep_gemm/include/deep_gemm/impls/sm90_fp8_mega_moe.cuh` (+1936 -0) ➕ `deep_gemm/include/deep_gemm/layout/sm90_mega_moe.cuh` (+179 -0) ➕ `deep_gemm/include/deep_gemm/scheduler/sm90_mega_moe.cuh` (+235 -0) 📝 `deep_gemm/mega/__init__.py` (+111 -0) ➕ `tests/test_mega_moe_hopper.py` (+172 -0) </details> ### 📄 Description Ref: https://github.com/sgl-project/DeepGEMM/pull/36 vllm support deepseek-v4-flash-fp8 model PR: https://github.com/vllm-project/vllm/pull/53527 --- <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

2026-10-26 21:17:45 +00:00

bsakaguchi referenced this issue

2026-10-26 21:17:57 +00:00

[PR #434] fix: derive the SM120 fp8 mqa_logits swizzle mode from head_dim #445