[PR #319] [CLOSED] Implement SM120 kernel for tf32_hc_prenorm_gemm #363

Closed
opened 2026-10-26 21:16:49 +00:00 by bsakaguchi · 0 comments
Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepGEMM/pull/319
Author: @ergodic-flow
Created: 4/25/2026
Status: ❌ Closed

Base: main ← Head: feature/sm120-support-v0


📝 Commits (1)

  • 872b23e Implement SM120 kernel for tf32_hc_prenorm_gemm

📊 Changes

6 files changed (+750 additions, -0 deletions)

View changed files

📝 csrc/apis/hyperconnection.hpp (+3 -0)
📝 csrc/indexing/main.cu (+1 -0)
➕ csrc/jit_kernels/heuristics/sm120.hpp (+243 -0)
➕ csrc/jit_kernels/impls/sm120_tf32_hc_prenorm_gemm.hpp (+153 -0)
➕ deep_gemm/include/deep_gemm/impls/sm120_tf32_hc_prenorm_gemm.cuh (+288 -0)
➕ deep_gemm/include/deep_gemm/mma/sm120.cuh (+62 -0)

📄 Description

This change introduces a new kernel for SM120 architecture (RTX 5090 and 6000 Pro) specifically for tf32_hc_prenorm_gemm as mentioned in issue #317 and issue #236 .

I primarily copied what exists for sm90, and made some adaptions until tests passed. The meat of the change will be in deep_gemm/include/deep_gemm/impls/sm120_tf32_hc_prenorm_gemm.cuh and a small shim around PTX in deep_gemm/include/deep_gemm/mma/sm120.cuh. As far as I can tell, the test test_hyperconnection.py appears to be passing, though the rest of the test suite is not working for me on SM120.

This is a first attempt, and I am very willing to keep working at it, so please take a look if / when you have time. Supporting a new arch is a lot of work, and I hope we the community can help! Thanks, and congrats on your recent release. If the PR is not to your liking, please close it!


🔄 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/319 **Author:** [@ergodic-flow](https://github.com/ergodic-flow) **Created:** 4/25/2026 **Status:** ❌ Closed **Base:** `main` ← **Head:** `feature/sm120-support-v0` --- ### 📝 Commits (1) - [`872b23e`](https://github.com/deepseek-ai/DeepGEMM/commit/872b23ec5f087fc1063698a02bdb5a1ab9d81b74) Implement SM120 kernel for tf32_hc_prenorm_gemm ### 📊 Changes **6 files changed** (+750 additions, -0 deletions) <details> <summary>View changed files</summary> 📝 `csrc/apis/hyperconnection.hpp` (+3 -0) 📝 `csrc/indexing/main.cu` (+1 -0) ➕ `csrc/jit_kernels/heuristics/sm120.hpp` (+243 -0) ➕ `csrc/jit_kernels/impls/sm120_tf32_hc_prenorm_gemm.hpp` (+153 -0) ➕ `deep_gemm/include/deep_gemm/impls/sm120_tf32_hc_prenorm_gemm.cuh` (+288 -0) ➕ `deep_gemm/include/deep_gemm/mma/sm120.cuh` (+62 -0) </details> ### 📄 Description This change introduces a new kernel for SM120 architecture (RTX 5090 and 6000 Pro) specifically for `tf32_hc_prenorm_gemm` as mentioned in issue #317 and issue #236 . I primarily copied what exists for `sm90`, and made some adaptions until tests passed. The meat of the change will be in `deep_gemm/include/deep_gemm/impls/sm120_tf32_hc_prenorm_gemm.cuh` and a small shim around PTX in `deep_gemm/include/deep_gemm/mma/sm120.cuh`. As far as I can tell, the test `test_hyperconnection.py` appears to be passing, though the rest of the test suite is not working for me on SM120. This is a first attempt, and I am very willing to keep working at it, so please take a look if / when you have time. Supporting a new arch is a lot of work, and I hope we the community can help! Thanks, and congrats on your recent release. If the PR is not to your liking, please close it! --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>
bsakaguchi 2026-10-26 21:16:49 +00:00
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/DeepGEMM#363
No description provided.