Skip to content

[GEMM][Config] add a8w8 block scale tuned config for deepseek-v3 #1310

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
valarLip merged 12 commits into from
Nov 10, 2025
Merged

[GEMM][Config] add a8w8 block scale tuned config for deepseek-v3 #1310

Changes from 8 commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
a0a2778
add a8w8 gemm tuned config with block scale for deepseek-v3 shapes
Oct 31, 2025
32c5023
Merge branch 'main' into guanbao/add_ds3_tuned_gemm_config
gbyu-amd Oct 31, 2025
ae47171
Merge branch 'main' into guanbao/add_ds3_tuned_gemm_config
gbyu-amd Nov 3, 2025
a00e838
Merge branch 'main' into guanbao/add_ds3_tuned_gemm_config
gbyu-amd Nov 3, 2025
1557e37
reorganize the config files
Nov 3, 2025
99f6129
revert unnecessnary changes
Nov 3, 2025
fba2834
add headers
Nov 3, 2025
28e3e73
Merge branch 'main' into guanbao/add_ds3_tuned_gemm_config
gbyu-amd Nov 4, 2025
9604376
move ds configs to specific model config file
Nov 5, 2025
3c18fc6
Merge branch 'main' into guanbao/add_ds3_tuned_gemm_config
gbyu-amd Nov 5, 2025
233931a
Merge branch 'main' into guanbao/add_ds3_tuned_gemm_config
gbyu-amd Nov 6, 2025
4776b0e
Merge branch 'main' into guanbao/add_ds3_tuned_gemm_config
gbyu-amd Nov 9, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
28 changes: 28 additions & 0 deletions aiter/configs/model_configs/a8w8_blockscale_tuned_gemm_ds_v3.csv
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,28 @@
cu_num,M,N,K,kernelId,splitK,us,kernelName,tflops,bw,errRatio
256,32768,7168,2048,0,0,661.7283,a8w8_blockscale_1x128x128_256x128x128x128_16x16_32x32_8x32x1_8x32x1_1x32x1x8_8_1x1_intrawave_v3,1453.88,833.5,0.0
256,1,2112,7168,8,0,23.2553,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,1.3,651.47,0.0
256,2,2112,7168,8,0,23.2237,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,2.61,652.85,0.0
256,4,2112,7168,8,0,23.3083,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,5.2,651.46,0.0
256,8,2112,7168,8,0,23.3352,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,10.38,652.66,0.0
256,16,2112,7168,8,0,20.4748,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,23.66,748.29,0.0
256,32,2112,7168,8,0,20.1767,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,48.02,768.38,0.0
256,64,2112,7168,8,0,19.9394,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,97.18,795.81,0.0
256,128,2112,7168,8,0,20.3542,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,190.4,815.41,0.0
256,256,2112,7168,18,0,22.1943,a8w8_blockscale_1x128x128_256x64x64x256_16x16_32x32_16x16x1_16x16x1_1x32x1x8_8_1x1_intrawave_v1,349.24,813.5,0.0
256,512,2112,7168,18,0,28.9658,a8w8_blockscale_1x128x128_256x64x64x256_16x16_32x32_16x16x1_16x16x1_1x32x1x8_8_1x1_intrawave_v1,535.19,724.01,0.0
256,1024,2112,7168,18,0,44.9407,a8w8_blockscale_1x128x128_256x64x64x256_16x16_32x32_16x16x1_16x16x1_1x32x1x8_8_1x1_intrawave_v1,689.89,596.44,0.0
256,2048,2112,7168,18,0,75.3204,a8w8_blockscale_1x128x128_256x64x64x256_16x16_32x32_16x16x1_16x16x1_1x32x1x8_8_1x1_intrawave_v1,823.26,510.75,0.0
256,4096,2112,7168,0,0,124.5688,a8w8_blockscale_1x128x128_256x128x128x128_16x16_32x32_8x32x1_8x32x1_1x32x1x8_8_1x1_intrawave_v3,995.57,496.11,0.0
256,8192,2112,7168,0,0,199.4594,a8w8_blockscale_1x128x128_256x128x128x128_16x16_32x32_8x32x1_8x32x1_1x32x1x8_8_1x1_intrawave_v3,1243.53,543.78,0.0
256,16384,2112,7168,0,0,357.2144,a8w8_blockscale_1x128x128_256x128x128x128_16x16_32x32_8x32x1_8x32x1_1x32x1x8_8_1x1_intrawave_v3,1388.71,564.89,0.0
256,32768,2112,7168,0,0,711.343,a8w8_blockscale_1x128x128_256x128x128x128_16x16_32x32_8x32x1_8x32x1_1x32x1x8_8_1x1_intrawave_v3,1394.74,546.05,0.0
256,1,3072,1536,8,0,7.1566,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,1.32,660.41,0.0
256,2,3072,1536,8,0,7.4571,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,2.53,634.82,0.0
256,4,3072,1536,8,0,7.7929,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,4.84,609.44,0.0
256,8,3072,1536,8,0,8.1011,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,9.32,590.05,0.0
256,32768,3072,1536,0,0,229.3125,a8w8_blockscale_1x128x128_256x128x128x128_16x16_32x32_8x32x1_8x32x1_1x32x1x8_8_1x1_intrawave_v3,1348.54,1118.02,0.0
256,1,4096,512,8,0,3.2195,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,1.3,654.09,0.0
256,2,4096,512,8,0,3.2712,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,2.56,646.42,0.0
256,4,4096,512,8,0,3.2431,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,5.17,657.39,0.0
256,8,4096,512,8,0,3.8036,a8w8_blockscale_1x128x128_256x16x64x256_16x16_16x16_16x16x1_16x16x1_1x16x1x16_4_1x1_intrawave_v1,8.82,569.67,0.0
Comment thread
valarLip marked this conversation as resolved.
Outdated
256,32768,4096,512,0,0,163.1056,a8w8_blockscale_1x128x128_256x128x128x128_16x16_32x32_8x32x1_8x32x1_1x32x1x8_8_1x1_intrawave_v3,842.64,1761.5,0.0