Update sgl-kernel and remove fast hadamard depedency (#11844)

docker/Dockerfile

@@ -9,7 +9,7 @@ ARG DEEPEP_COMMIT=9af0e0d0e74f3577af1979c9b9e1ac2cad0104ee
 ARG FLASHMLA_COMMIT=1408756a88e52a25196b759eaf8db89d2b51b5a1
 ARG FAST_HADAMARD_TRANSFORM_COMMIT=7fd811c2b47f63b0b08d2582619f939e14dad77c
 ARG CMAKE_BUILD_PARALLEL_LEVEL=2
-ARG SGL_KERNEL_VERSION=0.3.15
+ARG SGL_KERNEL_VERSION=0.3.16.post3
 ENV DEBIAN_FRONTEND=noninteractive \
     CUDA_HOME=/usr/local/cuda \
     GDRCOPY_HOME=/usr/src/gdrdrv-2.4.4/ \
@@ -152,14 +152,6 @@ RUN if [ "$TARGETARCH" = "amd64" ]; then \
       pip install -v . ; \
     fi
 
-# Install fast-hadamard-transform
-RUN if [ "$TARGETARCH" = "amd64" ]; then \
-      git clone https://github.com/Dao-AILab/fast-hadamard-transform && \
-      cd fast-hadamard-transform && \
-      git checkout ${FAST_HADAMARD_TRANSFORM_COMMIT} && \
-      pip install . ; \
-    fi
-
 # Python tools
 RUN python3 -m pip install --no-cache-dir \
     datamodel_code_generator \

python/pyproject.toml

@@ -55,7 +55,7 @@ dependencies = [
   "scipy",
   "sentencepiece",
   "setproctitle",
-  "sgl-kernel==0.3.15",
+  "sgl-kernel==0.3.16.post3",
   "soundfile==0.13.1",
   "tiktoken",
   "timm==1.0.16",

python/sglang/srt/entrypoints/engine.py

@@ -693,7 +693,7 @@ def _set_envs_and_config(server_args: ServerArgs):
     if _is_cuda and not get_bool_env_var("SGLANG_SKIP_SGL_KERNEL_VERSION_CHECK"):
         assert_pkg_version(
             "sgl-kernel",
-            "0.3.15",
+            "0.3.16.post3",
             "Please reinstall the latest version with `pip install sgl-kernel --force-reinstall`",
         )
 

python/sglang/srt/layers/attention/nsa/nsa_indexer.py

@@ -74,7 +74,7 @@ class BaseIndexerMetadata(ABC):
 
 def rotate_activation(x: torch.Tensor) -> torch.Tensor:
     assert x.dtype == torch.bfloat16
-    from fast_hadamard_transform import hadamard_transform
+    from sgl_kernel import hadamard_transform
 
     hidden_size = x.size(-1)
     assert (

scripts/ci/ci_install_dependency.sh

@@ -96,14 +96,6 @@ if [ "$RUN_DEEPSEEK_V32" = "1" ]; then
     git submodule update --init --recursive
     FLASH_MLA_DISABLE_SM100=${FLASH_MLA_DISABLE_SM100} $PIP_CMD install -v . $PIP_INSTALL_SUFFIX --no-build-isolation
     cd ..
-
-    # Install fast-hadamard-transform
-    FAST_HADAMARD_TRANSFORM_COMMIT="7fd811c2b47f63b0b08d2582619f939e14dad77c"
-    git clone https://github.com/Dao-AILab/fast-hadamard-transform
-    cd fast-hadamard-transform
-    git checkout ${FAST_HADAMARD_TRANSFORM_COMMIT}
-    $PIP_CMD install . $PIP_INSTALL_SUFFIX --no-build-isolation
-    cd ..
 fi
 
 # Show current packages