[PR #86] [MERGED] Use swizzling instead of padding #240

New issue

Closed

opened 2026-10-26 21:15:12 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 21:15:12 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepGEMM/pull/86
Author: @LyricZhao
Created: 4/14/2025
Status: ✅ Merged
Merged: 4/14/2025
Merged by: @LyricZhao

Base: main ← Head: swizzle-d

📝 Commits (10+)

6078b25 Add swizzling params
93c92c2 Add TMA D descriptor
76804c0 Always use STSMx2
4c11141 Swizzling draft
23d4289 Compatible with padding
5ff0eb2 Fix bugs
9406e2a Optimize swizzle performance
6366f5a Optimize expression
f8797d3 Optimize TMA issues
b699750 Fix README

📊 Changes

6 files changed (+147 additions, -104 deletions)

View changed files

📝 README.md (+2 -51)
📝 deep_gemm/include/deep_gemm/fp8_gemm.cuh (+93 -29)
📝 deep_gemm/jit/compiler.py (+1 -1)
📝 deep_gemm/jit_kernels/gemm.py (+35 -13)
📝 deep_gemm/jit_kernels/m_grouped_gemm.py (+12 -8)
📝 indexing/main.cu (+4 -2)

📄 Description

Remove performance report on README for maintenance simplicity
Peak TFLOPS 1503 -> 1519 TFLOPS, the original 1520 is on another "good" GPU
- (m= 4096, n=24576, k= 1536): 1212 -> 1270 TFLOPS
- (m= 4096, n=32768, k= 512): 775 -> 836 TFLOPS
- (m= 4096, n= 7168, k=16384): 1503 -> 1519 TFLOPS
- (m= 4096, n= 4096, k= 7168): 1410 -> 1439 TFLOPS
- (m= 4096, n= 7168, k= 2048): 1168 -> 1215 TFLOPS

_{🔄 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/86 **Author:** [@LyricZhao](https://github.com/LyricZhao) **Created:** 4/14/2025 **Status:** ✅ Merged **Merged:** 4/14/2025 **Merged by:** [@LyricZhao](https://github.com/LyricZhao) **Base:** `main` ← **Head:** `swizzle-d` --- ### 📝 Commits (10+) - [`6078b25`](https://github.com/deepseek-ai/DeepGEMM/commit/6078b25424699b9ea161aa3f34070ee8c9f25a41) Add swizzling params - [`93c92c2`](https://github.com/deepseek-ai/DeepGEMM/commit/93c92c2c89a22c012129efc85cb5e6b5c3314bc7) Add TMA D descriptor - [`76804c0`](https://github.com/deepseek-ai/DeepGEMM/commit/76804c096ddffe4a9ae1dbdf351f02b9805ac3d8) Always use STSMx2 - [`4c11141`](https://github.com/deepseek-ai/DeepGEMM/commit/4c111418a2e5a87a67678f026f602e34264ec6e2) Swizzling draft - [`23d4289`](https://github.com/deepseek-ai/DeepGEMM/commit/23d4289365f869062d01cbc65a453e61cfec7ae1) Compatible with padding - [`5ff0eb2`](https://github.com/deepseek-ai/DeepGEMM/commit/5ff0eb24b5608900a3d47306f1d68d0b0aaca45a) Fix bugs - [`9406e2a`](https://github.com/deepseek-ai/DeepGEMM/commit/9406e2a3a1b96c9eb8d7df7386f9a717b593b7e2) Optimize swizzle performance - [`6366f5a`](https://github.com/deepseek-ai/DeepGEMM/commit/6366f5ad1a88450f7dbee87223b130bbb330b9a2) Optimize expression - [`f8797d3`](https://github.com/deepseek-ai/DeepGEMM/commit/f8797d3c127fb7d5402f6eb022b51a32547cecee) Optimize TMA issues - [`b699750`](https://github.com/deepseek-ai/DeepGEMM/commit/b699750a4a7075cad00b9597dc3ab8f40c53e867) Fix README ### 📊 Changes **6 files changed** (+147 additions, -104 deletions) <details> <summary>View changed files</summary> 📝 `README.md` (+2 -51) 📝 `deep_gemm/include/deep_gemm/fp8_gemm.cuh` (+93 -29) 📝 `deep_gemm/jit/compiler.py` (+1 -1) 📝 `deep_gemm/jit_kernels/gemm.py` (+35 -13) 📝 `deep_gemm/jit_kernels/m_grouped_gemm.py` (+12 -8) 📝 `indexing/main.cu` (+4 -2) </details> ### 📄 Description - Remove performance report on README for maintenance simplicity - Peak TFLOPS 1503 -> 1519 TFLOPS, the original 1520 is on another "good" GPU - `(m= 4096, n=24576, k= 1536)`: 1212 -> 1270 TFLOPS - `(m= 4096, n=32768, k= 512)`: 775 -> 836 TFLOPS - `(m= 4096, n= 7168, k=16384)`: 1503 -> 1519 TFLOPS - `(m= 4096, n= 4096, k= 7168)`: 1410 -> 1439 TFLOPS - `(m= 4096, n= 7168, k= 2048)`: 1168 -> 1215 TFLOPS --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 21:15:12 +00:00