chore: upgrade sgl-kernel 0.0.8.post3 (#5342)

Author: zhyncs

python/pyproject.toml

@@ -47,7 +47,7 @@ runtime_common = [
 
 srt = [
     "sglang[runtime_common]",
-    "sgl-kernel==0.0.8",
+    "sgl-kernel==0.0.8.post3",
     "flashinfer_python==0.2.3",
     "torch==2.5.1",
     "torchvision==0.20.1",

python/sglang/srt/layers/quantization/fp8_kernel.py

@@ -41,7 +41,7 @@
 
 _is_cuda = is_cuda()
 if _is_cuda:
-    import deep_gemm  # `pip install "sgl-kernel>=0.0.4.post3"`
+    import deep_gemm
     from sgl_kernel import sgl_per_token_group_quant_fp8, sgl_per_token_quant_fp8
 
     sm_version = get_device_sm()

scripts/ci_install_dependency.sh

@@ -20,7 +20,7 @@ pip install --upgrade pip
 
 # Install flashinfer and sgl-kernel
 pip install flashinfer_python==0.2.3 --find-links ${FLASHINFER_REPO} --no-cache-dir
-pip install sgl-kernel==0.0.8 --no-cache-dir
+pip install sgl-kernel==0.0.8.post3 --no-cache-dir
 
 # Install the main package
 pip install -e "python[all]" --find-links ${FLASHINFER_REPO}