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

Refining fused A GEMM dispatch (#31241)

Browse Source 
Co-authored-by: root <root@sgl-b300-inference.datacrunch.io>
Co-authored-by: Brayden Zhong <brayden@radixark.ai>
This commit is contained in:
Brayden Zhong
2026-07-16 16:57:01 +08:00
committed by GitHub
co-authored by root Brayden Zhong
parent a798a2aeea
commit e5f9804e26
5 changed files with 66 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/jit/test_dsv3_fused_a_gemm.py
+6 -4
View File
@@ -12,10 +12,12 @@ from sglang.test.ci.ci_register import register_cuda_ci
register_cuda_ci(est_time=30, stage="base-b-kernel-unit", runner_config="1-gpu-large")
# hd_in must be a multiple of 256; 6144/7168 cover the real fused-A shapes.
HD_INS = [6144, 7168]
# hd_out must be a multiple of 16; 2112 and 2624 cover real fused-A variants.
HD_OUTS = [2112, 2624]
# hd_in must be a multiple of 256; 2048/6144/7168 cover the real fused-A shapes
# (2048 is q_b_proj TP4/TP8, 6144/7168 are qkv_a).
HD_INS = [2048, 6144, 7168]
# hd_out must be a multiple of 16; 2048/2112/2624/4096 cover real fused-A variants
# (2048/4096 are q_b_proj TP8/TP4, 2112/2624 are qkv_a).
HD_OUTS = [2048, 2112, 2624, 4096]
NUM_TOKENS = get_ci_test_range(list(range(1, 17)), [1, 8, 16])