chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

bump fa4 to official released fa4 pkg (#20303)

Browse Source
This commit is contained in:
Rain Jiang
2026-03-17 17:22:56 -07:00
committed by GitHub
parent c77d7c629e
commit cb1e63aba4
4 changed files with 11 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
docker/Dockerfile
View File

@@ -321,22 +321,21 @@ RUN --mount=type=cache,target=/root/.cache/pip \
&& rm -rf /root/.cargo /root/.rustup target dist ~/.cargo \ && rm -rf /root/.cargo /root/.rustup target dist ~/.cargo \
&& sed -i '/\.cargo\/env/d' /root/.profile /root/.bashrc 2>/dev/null || true && sed -i '/\.cargo\/env/d' /root/.profile /root/.bashrc 2>/dev/null || true
# Pin nvidia-cutlass-dsl to 4.3.5
RUN --mount=type=cache,target=/root/.cache/pip \ RUN --mount=type=cache,target=/root/.cache/pip \
python3 -m pip uninstall nvidia-cutlass-dsl-libs-base; \ python3 -m pip install "nvidia-cutlass-dsl>=4.4.1" "nvidia-cutlass-dsl-libs-base>=4.4.1" --force-reinstall --no-deps;
python3 -m pip install nvidia-cutlass-dsl==4.3.5 --force-reinstall;
# Patching packages for CUDA 12/13 compatibility # Patching packages for CUDA 12/13 compatibility
# TODO: Remove when torch version covers these packages # TODO: Remove when torch version covers these packages
RUN --mount=type=cache,target=/root/.cache/pip if [ "${CUDA_VERSION%%.*}" = "12" ]; then \ RUN --mount=type=cache,target=/root/.cache/pip if [ "${CUDA_VERSION%%.*}" = "12" ]; then \
python3 -m pip install nvidia-nccl-cu12==2.28.3 --force-reinstall --no-deps ; \ python3 -m pip install nvidia-nccl-cu12==2.28.3 --force-reinstall --no-deps ; \
python3 -m pip install nvidia-cudnn-cu12==9.16.0.29 --force-reinstall --no-deps ; \ python3 -m pip install nvidia-cudnn-cu12==9.16.0.29 --force-reinstall --no-deps ; \
python3 -m pip install cuda-python==12.9 ; \
elif [ "${CUDA_VERSION%%.*}" = "13" ]; then \ elif [ "${CUDA_VERSION%%.*}" = "13" ]; then \
python3 -m pip install nvidia-nccl-cu13==2.28.3 --force-reinstall --no-deps ; \ python3 -m pip install nvidia-nccl-cu13==2.28.3 --force-reinstall --no-deps ; \
python3 -m pip install nvidia-cudnn-cu13==9.16.0.29 --force-reinstall --no-deps ; \ python3 -m pip install nvidia-cudnn-cu13==9.16.0.29 --force-reinstall --no-deps ; \
python3 -m pip install nvidia-cublas==13.1.0.3 --force-reinstall --no-deps ; \ python3 -m pip install nvidia-cublas==13.1.0.3 --force-reinstall --no-deps ; \
python3 -m pip install nixl-cu13 --no-deps ; \ python3 -m pip install nixl-cu13 --no-deps ; \
python3 -m pip install cuda-python==13.1.1 ; \ python3 -m pip install cuda-python==13.2.0 ; \
fi fi
# Install development tools # Install development tools

7
python/pyproject.toml
View File

@@ -38,7 +38,7 @@ dependencies = [
"msgspec", "msgspec",
"ninja", "ninja",
"numpy", "numpy",
"nvidia-cutlass-dsl==4.3.5", "nvidia-cutlass-dsl>=4.4.1",
"nvidia-ml-py", "nvidia-ml-py",
"openai-harmony==0.0.4", "openai-harmony==0.0.4",
"openai==2.6.1", "openai==2.6.1",
@@ -54,12 +54,12 @@ dependencies = [
"pydantic", "pydantic",
"python-multipart", "python-multipart",
"pyzmq>=25.1.2", "pyzmq>=25.1.2",
"quack-kernels==0.2.4", "quack-kernels>=0.3.0",
"requests", "requests",
"scipy", "scipy",
"sentencepiece", "sentencepiece",
"setproctitle", "setproctitle",
"sgl-fa4==4.0.5", "flash-attn-4>=4.0.0b4",
"sglang-kernel==0.4.0", "sglang-kernel==0.4.0",
"soundfile==0.13.1", "soundfile==0.13.1",
"tiktoken", "tiktoken",
@@ -77,7 +77,6 @@ dependencies = [
"uvloop", "uvloop",
"watchfiles", "watchfiles",
"xgrammar==0.1.27", "xgrammar==0.1.27",
"smg-grpc-servicer>=0.5.0", "smg-grpc-servicer>=0.5.0",
] ]

4
python/sglang/jit_kernel/flash_attention_v4.py
View File

@@ -5,7 +5,7 @@ from typing import Callable, Optional, Tuple, Union
import torch import torch
try: try:
from sgl_fa4.cute import flash_attn_varlen_func as _flash_attn_varlen_func from flash_attn.cute import flash_attn_varlen_func as _flash_attn_varlen_func
except Exception as _e: # pragma: no cover except Exception as _e: # pragma: no cover
_flash_attn_varlen_func = None _flash_attn_varlen_func = None
_flash_attn_import_error = _e _flash_attn_import_error = _e
@@ -44,7 +44,7 @@ def flash_attn_varlen_func(
if _flash_attn_varlen_func is None: # pragma: no cover if _flash_attn_varlen_func is None: # pragma: no cover
raise ImportError( raise ImportError(
"Vendored FlashAttention CUTE is not available (cannot import " "Vendored FlashAttention CUTE is not available (cannot import "
"sgl_fa4.cute). Please check your source tree." "flash_attn.cute). Please check your source tree."
) from _flash_attn_import_error ) from _flash_attn_import_error
q, k, v = [_maybe_contiguous(t) for t in (q, k, v)] q, k, v = [_maybe_contiguous(t) for t in (q, k, v)]

18
scripts/ci/cuda/ci_install_dependency.sh
View File

@@ -117,7 +117,7 @@ else
fi fi
# Clean up existing installations # Clean up existing installations
$PIP_UNINSTALL_CMD sgl-kernel sglang-kernel sglang $PIP_UNINSTALL_SUFFIX || true $PIP_UNINSTALL_CMD sgl-kernel sglang-kernel sglang sgl-fa4 flash-attn-4 $PIP_UNINSTALL_SUFFIX || true
# Keep flashinfer packages installed if version matches to avoid re-downloading: # Keep flashinfer packages installed if version matches to avoid re-downloading:
# - flashinfer-cubin: 150+ MB, plus extra cubins from ci_download_flashinfer_cubin.sh # - flashinfer-cubin: 150+ MB, plus extra cubins from ci_download_flashinfer_cubin.sh
@@ -178,17 +178,6 @@ fi
# Install router for pd-disagg test # Install router for pd-disagg test
$PIP_CMD install sglang-router $PIP_INSTALL_SUFFIX $PIP_CMD install sglang-router $PIP_INSTALL_SUFFIX
# Remove flash_attn folder to avoid conflicts
PYTHON_LIB_PATH=$(python3 -c "import site; print(site.getsitepackages()[0])")
FLASH_ATTN_PATH="${PYTHON_LIB_PATH}/flash_attn"
if [ -d "$FLASH_ATTN_PATH" ]; then
echo "Directory $FLASH_ATTN_PATH exists. Removing..."
rm -rf "$FLASH_ATTN_PATH"
else
echo "Directory $FLASH_ATTN_PATH does not exist."
fi
# Install sgl-kernel # Install sgl-kernel
SGL_KERNEL_VERSION_FROM_KERNEL=$(grep -Po '(?<=^version = ")[^"]*' sgl-kernel/pyproject.toml) SGL_KERNEL_VERSION_FROM_KERNEL=$(grep -Po '(?<=^version = ")[^"]*' sgl-kernel/pyproject.toml)
SGL_KERNEL_VERSION_FROM_SRT=$(grep -Po -m1 '(?<=sglang-kernel==)[0-9A-Za-z\.\-]+' python/pyproject.toml) SGL_KERNEL_VERSION_FROM_SRT=$(grep -Po -m1 '(?<=sglang-kernel==)[0-9A-Za-z\.\-]+' python/pyproject.toml)
@@ -329,9 +318,8 @@ fi
# Download flashinfer cubins if the local set is incomplete # Download flashinfer cubins if the local set is incomplete
bash "${SCRIPT_DIR}/ci_download_flashinfer_cubin.sh" bash "${SCRIPT_DIR}/ci_download_flashinfer_cubin.sh"
# Clean nvidia-cutlass-dsl-libs-base for cutedsl lower than 0.4.4 # Force reinstall nvidia-cutlass-dsl to avoid potential version mismatch issues
$PIP_UNINSTALL_CMD nvidia-cutlass-dsl-libs-base $PIP_UNINSTALL_SUFFIX || true $PIP_CMD install "nvidia-cutlass-dsl>=4.4.1" "nvidia-cutlass-dsl-libs-base>=4.4.1" --no-deps --force-reinstall $PIP_INSTALL_SUFFIX || true
$PIP_CMD install nvidia-cutlass-dsl==4.3.5 --force-reinstall $PIP_INSTALL_SUFFIX || true
# Show current packages # Show current packages
$PIP_CMD list $PIP_CMD list