[PR #278] [CLOSED] no block quant #338

Closed
opened 2026-10-26 21:16:27 +00:00 by bsakaguchi · 0 comments
Owner

📋 Pull Request Information

Original PR: https://github.com/deepseek-ai/DeepGEMM/pull/278
Author: @PatrykSaffer
Created: 1/13/2026
Status: ❌ Closed

Base: main ← Head: no-block-quant


📝 Commits (2)

📊 Changes

8 files changed (+90 additions, -125 deletions)

View changed files

📝 csrc/apis/gemm.hpp (+2 -9)
📝 csrc/jit_kernels/heuristics/sm90.hpp (+6 -12)
📝 csrc/jit_kernels/impls/sm90_fp8_gemm_1d2d.hpp (+5 -11)
📝 csrc/utils/layout.hpp (+0 -24)
📝 deep_gemm/include/deep_gemm/impls/sm90_fp8_gemm_1d2d.cuh (+28 -62)
📝 deep_gemm/utils/math.py (+24 -0)
📝 tests/generators.py (+25 -4)
📝 tests/test_fp8.py (+0 -3)

📄 Description

from

Perf (num_groups=4, m=35456, n= 4096, k= 7168, 1D2D, layout=NT): 1458 us | 1428 TFLOPS | 460 GB/s
Perf (num_groups=4, m=36096, n= 7168, k= 2048, 1D2D, layout=NT): 784 us | 1352 TFLOPS | 832 GB/s
Perf (num_groups=8, m=32384, n= 4096, k= 7168, 1D2D, layout=NT): 1346 us | 1413 TFLOPS | 549 GB/s
Perf (num_groups=8, m=31232, n= 7168, k= 2048, 1D2D, layout=NT): 683 us | 1342 TFLOPS | 924 GB/s

no B block quant

Perf (num_groups=4, m=35456, n= 4096, k= 7168, 1D2D, layout=NT): 1449 us | 1437 TFLOPS | 462 GB/s
Perf (num_groups=4, m=36096, n= 7168, k= 2048, 1D2D, layout=NT): 774 us | 1369 TFLOPS | 843 GB/s
Perf (num_groups=8, m=32384, n= 4096, k= 7168, 1D2D, layout=NT): 1335 us | 1424 TFLOPS | 554 GB/s
Perf (num_groups=8, m=31232, n= 7168, k= 2048, 1D2D, layout=NT): 673 us | 1362 TFLOPS | 937 GB/s

static A quant

Perf (num_groups=4, m=35456, n= 4096, k= 7168, 1D2D, layout=NT): 1324 us | 1572 TFLOPS | 500 GB/s
Perf (num_groups=4, m=36096, n= 7168, k= 2048, 1D2D, layout=NT): 707 us | 1498 TFLOPS | 919 GB/s
Perf (num_groups=8, m=32384, n= 4096, k= 7168, 1D2D, layout=NT): 1218 us | 1561 TFLOPS | 601 GB/s
Perf (num_groups=8, m=31232, n= 7168, k= 2048, 1D2D, layout=NT): 616 us | 1488 TFLOPS | 1021 GB/s


🔄 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/278 **Author:** [@PatrykSaffer](https://github.com/PatrykSaffer) **Created:** 1/13/2026 **Status:** ❌ Closed **Base:** `main` ← **Head:** `no-block-quant` --- ### 📝 Commits (2) - [`ab02dd5`](https://github.com/deepseek-ai/DeepGEMM/commit/ab02dd5810b29a127b5af09f9a7e06822f74697e) no block quant - [`3b1c559`](https://github.com/deepseek-ai/DeepGEMM/commit/3b1c55953dadcb07087484f29b337f22428067c0) 0d0d ### 📊 Changes **8 files changed** (+90 additions, -125 deletions) <details> <summary>View changed files</summary> 📝 `csrc/apis/gemm.hpp` (+2 -9) 📝 `csrc/jit_kernels/heuristics/sm90.hpp` (+6 -12) 📝 `csrc/jit_kernels/impls/sm90_fp8_gemm_1d2d.hpp` (+5 -11) 📝 `csrc/utils/layout.hpp` (+0 -24) 📝 `deep_gemm/include/deep_gemm/impls/sm90_fp8_gemm_1d2d.cuh` (+28 -62) 📝 `deep_gemm/utils/math.py` (+24 -0) 📝 `tests/generators.py` (+25 -4) 📝 `tests/test_fp8.py` (+0 -3) </details> ### 📄 Description from > Perf (num_groups=4, m=35456, n= 4096, k= 7168, 1D2D, layout=NT): 1458 us | 1428 TFLOPS | 460 GB/s > Perf (num_groups=4, m=36096, n= 7168, k= 2048, 1D2D, layout=NT): 784 us | 1352 TFLOPS | 832 GB/s > Perf (num_groups=8, m=32384, n= 4096, k= 7168, 1D2D, layout=NT): 1346 us | 1413 TFLOPS | 549 GB/s > Perf (num_groups=8, m=31232, n= 7168, k= 2048, 1D2D, layout=NT): 683 us | 1342 TFLOPS | 924 GB/s no B block quant > Perf (num_groups=4, m=35456, n= 4096, k= 7168, 1D2D, layout=NT): 1449 us | 1437 TFLOPS | 462 GB/s > Perf (num_groups=4, m=36096, n= 7168, k= 2048, 1D2D, layout=NT): 774 us | 1369 TFLOPS | 843 GB/s > Perf (num_groups=8, m=32384, n= 4096, k= 7168, 1D2D, layout=NT): 1335 us | 1424 TFLOPS | 554 GB/s > Perf (num_groups=8, m=31232, n= 7168, k= 2048, 1D2D, layout=NT): 673 us | 1362 TFLOPS | 937 GB/s static A quant > Perf (num_groups=4, m=35456, n= 4096, k= 7168, 1D2D, layout=NT): 1324 us | 1572 TFLOPS | 500 GB/s > Perf (num_groups=4, m=36096, n= 7168, k= 2048, 1D2D, layout=NT): 707 us | 1498 TFLOPS | 919 GB/s > Perf (num_groups=8, m=32384, n= 4096, k= 7168, 1D2D, layout=NT): 1218 us | 1561 TFLOPS | 601 GB/s > Perf (num_groups=8, m=31232, n= 7168, k= 2048, 1D2D, layout=NT): 616 us | 1488 TFLOPS | 1021 GB/s --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>
bsakaguchi 2026-10-26 21:16:27 +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#338
No description provided.