EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Restruct sgl-kernel benchmark (#10861)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2025-09-25 07:45:25 +08:00
committed by GitHub
parent 7a06ef984d
commit c4e314f986
27 changed files with 425 additions and 319 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/layers/quantization/w4afp8.py
View File

@@ -19,10 +19,6 @@ from sglang.srt.layers.quantization.unquant import UnquantizedLinearMethod
from sglang.srt.layers.quantization.utils import is_layer_skipped
from sglang.srt.utils import is_npu, set_weight_attrs
_is_npu = is_npu()
if not _is_npu:
from sglang.srt.layers.moe.cutlass_w4a8_moe import cutlass_w4a8_moe
if TYPE_CHECKING:
from sglang.srt.layers.moe import MoeRunnerConfig
from sglang.srt.layers.moe.ep_moe.layer import EPMoE

4
python/sglang/test/test_block_fp8.py
View File

@@ -621,11 +621,11 @@ class TestW8A8BlockFP8BatchedDeepGemm(CustomTestCase):
w_s,
)
from deep_gemm import m_grouped_gemm_fp8_fp8_bf16_nt_masked
from deep_gemm import fp8_m_grouped_gemm_nt_masked
with torch.inference_mode():
ref_out = torch_w8a8_block_fp8_bmm(a, a_s, w, w_s, block_size, dtype)
m_grouped_gemm_fp8_fp8_bf16_nt_masked(lhs, rhs, oe, masked_m, expected_m)
fp8_m_grouped_gemm_nt_masked(lhs, rhs, oe, masked_m, expected_m)
out = oe[:, :M, :]
self.assertTrue(