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

enable ut test for xpu devices (#11712)

Browse Source 
Co-authored-by: jundu <jun.du@intel.com>
Co-authored-by: Gao, Pengfei <pengfei.gao@intel.com>
This commit is contained in:
DiweiSun
2026-02-03 11:15:14 -08:00
committed by GitHub
co-authored by jundu Gao, Pengfei
parent 0a6925639b
commit 495290aefd
20 changed files with 237 additions and 151 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/manual/test_triton_moe_wna16.py
+17 -10
View File
@@ -7,6 +7,7 @@ from sglang.srt.layers.activation import SiluAndMul
from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_moe
from sglang.srt.layers.moe.topk import TopKConfig, select_experts
from sglang.srt.server_args import ServerArgs, set_global_server_args_for_scheduler
from sglang.srt.utils import get_device
NUM_EXPERTS = [8, 64]
TOP_KS = [2, 6]
@@ -159,10 +160,10 @@ def test_fused_moe_wn16(
weight_bits: int,
):
print(m, n, k, e, topk, dtype, group_size, has_zp, weight_bits)
a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10
w2 = torch.randn((e, k, n), device="cuda", dtype=dtype) / 10
score = torch.randn((m, e), device="cuda", dtype=dtype)
a = torch.randn((m, k), device=get_device(), dtype=dtype) / 10
w1 = torch.randn((e, 2 * n, k), device=get_device(), dtype=dtype) / 10
w2 = torch.randn((e, k, n), device=get_device(), dtype=dtype) / 10
score = torch.randn((m, e), device=get_device(), dtype=dtype)
if weight_bits == 4:
pack_factor = 2
@@ -174,16 +175,22 @@ def test_fused_moe_wn16(
w1_ref = w1.clone()
w2_ref = w2.clone()
w1_qweight = torch.empty(
(e, 2 * n, k // pack_factor), device="cuda", dtype=torch.uint8
(e, 2 * n, k // pack_factor), device=get_device(), dtype=torch.uint8
)
w2_qweight = torch.empty((e, k, n // pack_factor), device="cuda", dtype=torch.uint8)
w1_scales = torch.empty((e, 2 * n, k // group_size), device="cuda", dtype=dtype)
w2_scales = torch.empty((e, k, n // group_size), device="cuda", dtype=dtype)
w2_qweight = torch.empty(
(e, k, n // pack_factor), device=get_device(), dtype=torch.uint8
)
w1_scales = torch.empty(
(e, 2 * n, k // group_size), device=get_device(), dtype=dtype
)
w2_scales = torch.empty((e, k, n // group_size), device=get_device(), dtype=dtype)
w1_qzeros = torch.empty(
(e, 2 * n // pack_factor, k // group_size), device="cuda", dtype=torch.uint8
(e, 2 * n // pack_factor, k // group_size),
device=get_device(),
dtype=torch.uint8,
)
w2_qzeros = torch.empty(
(e, k // pack_factor, n // group_size), device="cuda", dtype=torch.uint8
(e, k // pack_factor, n // group_size), device=get_device(), dtype=torch.uint8
)
for i in range(e * 2):