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
python/sglang/jit_kernel/dsv3_fused_a_gemm.py
+1 -1
View File
@@ -1,7 +1,7 @@
"""
JIT kernel for DeepSeek V3 fused QKV-A GEMM (min-latency).
Replaces the AOT sgl_kernel.dsv3_fused_a_gemm for SM90+ (Hopper) GPUs.
Runtime-compiled CUDA C++ kernel for SM90+ (Hopper) GPUs.
Shapes: hd_in a multiple of 256, hd_out a multiple of 16, num_tokens 1-16, bfloat16.
"""
python/sglang/jit_kernel/dsv3_router_gemm.py
+1 -1
View File
@@ -1,7 +1,7 @@
"""
JIT kernel for DeepSeek V3 router GEMM.
Replaces the AOT sgl_kernel.dsv3_router_gemm for SM90+ (Hopper) GPUs.
Runtime-compiled CUDA C++ kernel for SM90+ (Hopper) GPUs.
Supports num_experts in {256, 384}, hidden_dim a multiple of 1024, num_tokens 1-16.
"""
python/sglang/jit_kernel/fused_a_gemm.py
+2 -6
View File
@@ -1,8 +1,7 @@
"""Unified entry point for the DeepSeek-V3 fused QKV-A GEMM.
Dispatches to one of three interchangeable implementations via ``backend``:
Dispatches to one of two interchangeable implementations via ``backend``:
- ``"aot"``: prebuilt ``sgl_kernel.dsv3_fused_a_gemm`` (CUDA C++).
- ``"jit"``: runtime-compiled CUDA C++ (``sglang.jit_kernel.dsv3_fused_a_gemm``).
- ``"cutedsl"``: CuTe DSL (``sglang.jit_kernel.cutedsl_dsv3_fused_a_gemm``).
- ``"auto"``: CuTe DSL on SM120+, otherwise the JIT kernel.
@@ -21,7 +20,6 @@ from sglang.srt.utils.common import get_device_sm, is_cuda, is_sm120_supported
class FusedAGemmBackend(str, Enum):
AUTO = "auto"
AOT = "aot"
JIT = "jit"
CUTEDSL = "cutedsl"
@@ -70,9 +68,7 @@ def dsv3_fused_a_gemm(
if backend == FusedAGemmBackend.AUTO:
backend = _AUTO_BACKEND
if backend == FusedAGemmBackend.AOT:
from sgl_kernel import dsv3_fused_a_gemm as impl
elif backend == FusedAGemmBackend.JIT:
if backend == FusedAGemmBackend.JIT:
from sglang.jit_kernel.dsv3_fused_a_gemm import dsv3_fused_a_gemm as impl
else:
from sglang.jit_kernel.cutedsl_dsv3_fused_a_gemm import (