minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Delete sgl-kernel AOT router GEMM and fused A GEMM (#30280)

Browse Source 
Co-authored-by: Brayden Zhong <brayden@radixark.ai>
Co-authored-by: root <root@sgl-b300-inference.datacrunch.io>
This commit is contained in:
Brayden Zhong
2026-07-22 08:44:59 +08:00
committed by GitHub
co-authored by Brayden Zhong root
parent 8ae0eb83fc
commit 03342e7732
19 changed files with 10 additions and 1834 deletions
Download Patch File Download Diff File Expand all files Collapse all files
sgl-kernel/tests/test_dsv3_fused_a_gemm.py
-34
View File
@@ -1,34 +0,0 @@
import sys
import pytest
import torch
import torch.nn.functional as F
from sgl_kernel import dsv3_fused_a_gemm
@pytest.mark.parametrize("num_tokens", [1, 8, 15, 16])
def test_dsv3_fused_a_gemm(num_tokens):
kHdIn = 7168
kHdOut = 2112
mat_a = torch.randn(
(num_tokens, kHdIn), dtype=torch.bfloat16, device="cuda"
).contiguous()
mat_b = torch.randn((kHdOut, kHdIn), dtype=torch.bfloat16, device="cuda").transpose(
0, 1
)
output = torch.empty(
(num_tokens, kHdOut), dtype=torch.bfloat16, device="cuda"
).contiguous()
ref = F.linear(mat_a, mat_b.T)
output = dsv3_fused_a_gemm(mat_a, mat_b)
assert torch.allclose(
output, ref, rtol=1e-2, atol=1e-3
), "Fused GEMM output mismatch with torch.nn.functional.linear reference"
if __name__ == "__main__":
sys.exit(pytest.main([__file__]))