chore: vendor sglang v0.5.10 snapshot

2026-04-24 12:29:36 +00:00
parent 78f0d15221
commit bded08301f
4308 changed files with 1200894 additions and 2 deletions
--- a/third_party/sglang/3rdparty/amd/wheel/README.md
+++ b/third_party/sglang/3rdparty/amd/wheel/README.md
@@ -0,0 +1,97 @@
+# sglang-kernel (prior sgl-kernel)
+
+Building and releasing `sglang-kernel` as a wheel is a part of the release workflow. Check [release-whl-kernel.yml](https://github.com/sgl-project/sglang/blob/main/.github/workflows/release-whl-kernel.yml) for details.
+
+# sglang
+
+`3rdparty/amd/wheel/sglang/pyproject.toml` is the AMD-specific pyproject for building the `amd-sglang` wheel. It extends `python/pyproject_other.toml` with two ROCm-version extras (`rocm700`, `rocm720`) that pin the matching torch/triton/torchaudio/torchvision/`sglang-kernel` wheels, and renames the package to `amd-sglang`.
+
+## Operation to build sglang wheel
+
+```
+$ git clone https://github.com/sgl-project/sglang.git && cd sglang
+$ cp 3rdparty/amd/wheel/sglang/pyproject.toml python/pyproject.toml
+$ cd python && python -m build
+```
+
+## Installation
+
+### v0.5.9
+
+ROCm 7.0.0:
+```
+pip uninstall sglang-kernel sglang amd-sglang
+pip install "amd-sglang[all-hip,rocm700]" -i https://pypi.amd.com/rocm-7.0.0/simple --extra-index-url https://pypi.org/simple
+```
+
+ROCm 7.2.0:
+```
+pip uninstall sglang-kernel sglang amd-sglang
+pip install "amd-sglang[all-hip,rocm720]" -i https://pypi.amd.com/rocm-7.2.0/simple --extra-index-url https://pypi.org/simple
+```
+
+Note: You must resolve the two dependencies, AITER and triton, below.  Others are optional depending on your applications.
+
+## Manual Dependency Resolution
+
+### Resolving AITER
+
+[AITER](https://github.com/ROCm/aiter) is a fundamental dependency. Wheel-izing it is ongoing.
+Until we can pin it reliably, install it manually (typically following the [ROCm docker recipe](https://github.com/sgl-project/sglang/blob/main/docker/rocm.Dockerfile#L106).
+
+### Revolving triton
+
+To avoid known issues in triton 3.5.1 installed by default, we recommend upgrading triton after installation.  In ROCm 7.0.0 environment,
+```
+pip install triton==3.6.0
+```
+or ROCm 7.2.0,
+```
+pip install https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/triton-3.6.0%2Brocm7.2.0.gitba5c1517-cp310-cp310-linux_x86_64.whl
+```
+
+#### `torch._inductor.exc.InductorError: AttributeError: 'KernelMetadata' object has no attribute 'cluster_dims'`
+
+After upgrading, you may hit this error during inference when PyTorch Inductor interacts with Triton metadata.
+
+A pragmatic workaround is to guard the metadata access in Inductor's Triton heuristics so it only reads `cluster_dims` when the attribute exists:
+
+```diff
+--- a/opt/venv/lib/python3.10/site-packages/torch/_inductor/runtime/triton_heuristics.py
+++ b/opt/venv/lib/python3.10/site-packages/torch/_inductor/runtime/triton_heuristics.py
+@@ -1759,6 +1759,8 @@
+                 else (
+                     (binary.metadata.num_ctas, *binary.metadata.cluster_dims)
+                     if hasattr(binary, "metadata")
+                    and hasattr(binary.metadata, "num_ctas")
+                    and hasattr(binary.metadata, "cluster_dims")
+                     else ()
+                 )
+             ),
+```
+
+### Resolving Dependencies for Distributed Inference
+
+#### sgl-model-gateway
+
+Install sgl-model-gateway as follows:
+
+```
+$ apt install openssl libssl-dev protobuf
+$ export PATH="/$HOME/.cargo/bin:${PATH}" \
+  && curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y \
+  && rustc --version && cargo --version # Prepare for a rust toolchain
+$ python3 -m pip install --no-cache-dir setuptools-rust \
+  && cd /sgl-workspace/sglang/sgl-model-gateway/bindings/python \
+  && cargo build --release \
+  && python3 -m pip install --no-cache-dir . \
+  && rm -rf /root/.cache # Build and install sgl-model-gateway
+```
+
+#### [Mori](https://github.com/sgl-project/sglang/blob/main/docker/rocm.Dockerfile#L381)
+
+### Resolving Dependencies for DeepSeek-V3.2
+
+#### [TileLang](https://github.com/sgl-project/sglang/blob/main/docker/rocm.Dockerfile#L216)
+
+#### [FHT (fast-hadamard-transform)](https://github.com/sgl-project/sglang/blob/main/docker/rocm.Dockerfile#L300)
--- a/third_party/sglang/3rdparty/amd/wheel/sgl-kernel/CMakeLists_rocm.txt
+++ b/third_party/sglang/3rdparty/amd/wheel/sgl-kernel/CMakeLists_rocm.txt
@@ -0,0 +1,161 @@
+cmake_minimum_required(VERSION 3.24 FATAL_ERROR)
+project(sgl_kernel LANGUAGES CXX)
+
+# Cmake
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_POSITION_INDEPENDENT_CODE ON)
+set(CMAKE_SHARED_LIBRARY_PREFIX "")
+
+set(CMAKE_COLOR_DIAGNOSTICS ON)
+set(CMAKE_VERBOSE_MAKEFILE ON CACHE BOOL "ON")
+
+# Python / Torch
+find_package(Python COMPONENTS Interpreter Development.Module ${SKBUILD_SABI_COMPONENT} REQUIRED)
+
+execute_process(
+  COMMAND ${Python_EXECUTABLE} -c "import torch; print(torch.utils.cmake_prefix_path)"
+  OUTPUT_VARIABLE TORCH_PY_PREFIX
+  OUTPUT_STRIP_TRAILING_WHITESPACE
+)
+
+set(Torch_DIR "${TORCH_PY_PREFIX}/Torch")
+list(APPEND CMAKE_PREFIX_PATH "${TORCH_PY_PREFIX}/Torch")
+find_package(Torch REQUIRED)
+
+execute_process(
+  COMMAND ${Python_EXECUTABLE} -c "import torch; print(int(torch._C._GLIBCXX_USE_CXX11_ABI))"
+  OUTPUT_VARIABLE TORCH_CXX11_ABI
+  OUTPUT_STRIP_TRAILING_WHITESPACE
+)
+if(TORCH_CXX11_ABI STREQUAL "0")
+  add_compile_definitions(_GLIBCXX_USE_CXX11_ABI=0)
+else()
+  add_compile_definitions(_GLIBCXX_USE_CXX11_ABI=1)
+endif()
+
+# ROCm/HIP
+enable_language(HIP)
+list(APPEND CMAKE_PREFIX_PATH "/opt/rocm/lib/cmake/hip-lang")
+find_package(hip REQUIRED CONFIG)
+
+# Determine AMDGPU target from environment variable or default to gfx942
+set(AMDGPU_TARGET_ENV "$ENV{AMDGPU_TARGET}")
+
+if(AMDGPU_TARGET_ENV)
+  # Use environment variable if specified
+  set(AMDGPU_TARGETS "${AMDGPU_TARGET_ENV}")
+  message(STATUS "Using AMDGPU_TARGET from environment: ${AMDGPU_TARGETS}")
+else()
+  # Default to gfx942 only
+  set(AMDGPU_TARGETS "gfx942")
+  message(STATUS "AMDGPU_TARGET not set, defaulting to gfx942")
+endif()
+
+# Set HIP architectures
+set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
+
+# FP8 macro selection
+# Always define HIP_FP8_TYPE_FNUZ=1 (for gfx942 and host compilation)
+# Additionally define HIP_FP8_TYPE_E4M3=1 when building for gfx950
+# The existing utils.h logic will pick the right one based on architecture
+set(SGL_FP8_MACROS "-DHIP_FP8_TYPE_FNUZ=1")
+
+if(AMDGPU_TARGETS MATCHES "gfx950")
+  list(APPEND SGL_FP8_MACROS "-DHIP_FP8_TYPE_E4M3=1")
+  message(STATUS "Multi-arch build: Enabling both HIP_FP8_TYPE_FNUZ (gfx942) and HIP_FP8_TYPE_E4M3 (gfx950)")
+elseif(AMDGPU_TARGETS MATCHES "gfx942")
+  message(STATUS "Single-arch build: Enabling HIP_FP8_TYPE_FNUZ for gfx942")
+else()
+  message(FATAL_ERROR "Unsupported AMDGPU_TARGET '${AMDGPU_TARGETS}'. Expected 'gfx942' or 'gfx950' or both.")
+endif()
+
+# TopK dynamic smem bytes
+# Dynamic shared-memory budget for the TopK kernels.
+# - gfx942 (MI300/MI325): LDS is typically 64KB per workgroup -> keep dynamic smem <= ~48KB
+#   (leaves room for static shared allocations in the kernel).
+# - gfx95x (MI350): LDS is larger (e.g. 160KB per CU) -> allow the original 128KB dynamic smem.
+if(AMDGPU_TARGET_ONE STREQUAL "gfx942")
+  math(EXPR SGL_TOPK_DYNAMIC_SMEM_BYTES "48 * 1024")
+else()
+  math(EXPR SGL_TOPK_DYNAMIC_SMEM_BYTES "32 * 1024 * 4")
+endif()
+
+set(SGL_TOPK_MACROS "-DSGL_TOPK_DYNAMIC_SMEM_BYTES=${SGL_TOPK_DYNAMIC_SMEM_BYTES}")
+
+# Paths / includes
+set(PROJ_ROOT ${CMAKE_CURRENT_LIST_DIR})
+set(SGL_INCLUDE_DIRS
+  ${PROJ_ROOT}/include
+  ${PROJ_ROOT}/include/impl
+  ${PROJ_ROOT}/csrc
+  ${TORCH_INCLUDE_DIRS}
+)
+
+# Platform-specific library directory
+set(PLAT_LIB_DIR "/usr/lib/x86_64-linux-gnu")
+link_directories(${PLAT_LIB_DIR})
+
+# Sources
+set(SOURCES
+${PROJ_ROOT}/csrc/allreduce/custom_all_reduce.hip
+${PROJ_ROOT}/csrc/allreduce/deterministic_all_reduce.hip
+${PROJ_ROOT}/csrc/allreduce/quick_all_reduce.hip
+${PROJ_ROOT}/csrc/common_extension_rocm.cc
+${PROJ_ROOT}/csrc/elementwise/activation.hip
+${PROJ_ROOT}/csrc/elementwise/pos_enc.hip
+${PROJ_ROOT}/csrc/elementwise/topk.hip
+${PROJ_ROOT}/csrc/grammar/apply_token_bitmask_inplace_hip.hip
+${PROJ_ROOT}/csrc/kvcacheio/transfer.hip
+${PROJ_ROOT}/csrc/memory/weak_ref_tensor.cpp
+${PROJ_ROOT}/csrc/moe/moe_align_kernel.hip
+${PROJ_ROOT}/csrc/moe/moe_topk_softmax_kernels.hip
+${PROJ_ROOT}/csrc/moe/moe_topk_sigmoid_kernels.hip
+${PROJ_ROOT}/csrc/speculative/eagle_utils.hip
+)
+set_source_files_properties(
+  ${SOURCES}
+  PROPERTIES
+    LANGUAGE HIP
+)
+
+# Compile / Link flags
+add_compile_options($<$<COMPILE_LANGUAGE:CXX>:-O3>)
+
+set(SGL_HIP_FLAGS
+  -DNDEBUG
+  -DOPERATOR_NAMESPACE=sgl_kernel
+  -O3
+  -std=c++17
+  -DENABLE_BF16
+  -DENABLE_FP8
+  ${SGL_FP8_MACROS}
+  -Wno-pass-failed
+  -Wundefined-internal
+  ${SGL_TOPK_MACROS}
+)
+
+# Python extension
+Python_add_library(common_ops MODULE USE_SABI ${SKBUILD_SABI_VERSION} WITH_SOABI ${SOURCES})
+target_include_directories(common_ops PRIVATE ${SGL_INCLUDE_DIRS})
+
+# Apply per-language flags
+target_compile_options(common_ops PRIVATE
+  $<$<COMPILE_LANGUAGE:HIP>:${SGL_HIP_FLAGS}>
+)
+
+target_link_libraries(common_ops PRIVATE
+  ${TORCH_LIBRARIES}
+  hip::device
+  hip::host
+  hiprtc
+  amdhip64
+)
+
+target_link_options(common_ops PRIVATE
+  "SHELL:-Wl,-rpath,'\$ORIGIN/../../torch/lib'"
+)
+
+install(TARGETS common_ops
+  LIBRARY DESTINATION sgl_kernel
+)
--- a/third_party/sglang/3rdparty/amd/wheel/sgl-kernel/build_rocm.sh
+++ b/third_party/sglang/3rdparty/amd/wheel/sgl-kernel/build_rocm.sh
@@ -0,0 +1,50 @@
+#!/bin/bash
+set -euo pipefail
+
+ROCM_VERSION=${1:-}
+
+if [[ "${ROCM_VERSION}" == "700" ]]; then
+  IMAGE="lmsysorg/sglang:v0.5.8.post1-rocm700-mi35x"
+elif [[ "${ROCM_VERSION}" == "720" ]]; then
+  IMAGE="rocm/pytorch:rocm7.2_ubuntu22.04_py3.10_pytorch_release_2.9.1"
+else
+  echo "ERROR: Unsupported ROCM_VERSION='${ROCM_VERSION}'. Only '700' and '720' are supported." >&2
+  exit 1
+fi
+
+PYTHON_ROOT_PATH="/opt/venv/bin"
+AMDGPU_TARGET="gfx942;gfx950"
+
+# Pull and run the latest image
+echo "Pulling Docker image: ${IMAGE}"
+docker pull "${IMAGE}"
+
+docker run --rm \
+  -v $(pwd):/sgl-kernel \
+  -e AMDGPU_TARGET="${AMDGPU_TARGET}" \
+  -e PYTORCH_ROCM_ARCH="${AMDGPU_TARGET}" \
+  ${IMAGE} \
+  bash -c "
+  # Install torch, triton, and friends, depending on the ROCm version
+  if [[ "${ROCM_VERSION}" == "700" ]]; then
+    ${PYTHON_ROOT_PATH}/pip install https://repo.radeon.com/rocm/manylinux/rocm-rel-7.0.2/torch-2.9.1.dev20251204%2Brocm7.0.2.lw.git351ff442-cp310-cp310-linux_x86_64.whl https://repo.radeon.com/rocm/manylinux/rocm-rel-7.0.2/triton-3.5.1%2Brocm7.0.2.gita272dfa8-cp310-cp310-linux_x86_64.whl https://repo.radeon.com/rocm/manylinux/rocm-rel-7.0.2/torchaudio-2.9.0%2Brocm7.0.2.gite3c6ee2b-cp310-cp310-linux_x86_64.whl https://repo.radeon.com/rocm/manylinux/rocm-rel-7.0.2/torchvision-0.24.0%2Brocm7.0.2.gitb919bd0c-cp310-cp310-linux_x86_64.whl
+  elif [[ "${ROCM_VERSION}" == "720" ]]; then
+    ${PYTHON_ROOT_PATH}/pip install https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/torch-2.9.1%2Brocm7.2.0.lw.git7e1940d4-cp310-cp310-linux_x86_64.whl https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/triton-3.5.1%2Brocm7.2.0.gita272dfa8-cp310-cp310-linux_x86_64.whl https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/torchaudio-2.9.0%2Brocm7.2.0.gite3c6ee2b-cp310-cp310-linux_x86_64.whl https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/torchvision-0.24.0%2Brocm7.2.0.gitb919bd0c-cp310-cp310-linux_x86_64.whl
+  fi
+  # Install CMake (version >= 3.26) - Robust Installation
+  export CMAKE_VERSION_MAJOR=3.31
+  export CMAKE_VERSION_MINOR=1
+  echo \"Downloading CMake from: https://cmake.org/files/v\${CMAKE_VERSION_MAJOR}/cmake-\${CMAKE_VERSION_MAJOR}.\${CMAKE_VERSION_MINOR}-linux-x86_64.tar.gz\"
+  wget https://cmake.org/files/v\${CMAKE_VERSION_MAJOR}/cmake-\${CMAKE_VERSION_MAJOR}.\${CMAKE_VERSION_MINOR}-linux-x86_64.tar.gz
+  tar -xzf cmake-\${CMAKE_VERSION_MAJOR}.\${CMAKE_VERSION_MINOR}-linux-x86_64.tar.gz
+  mv cmake-\${CMAKE_VERSION_MAJOR}.\${CMAKE_VERSION_MINOR}-linux-x86_64 /opt/cmake
+  export PATH=/opt/cmake/bin:\$PATH
+
+  ${PYTHON_ROOT_PATH}/pip install --no-cache-dir ninja setuptools wheel numpy uv scikit-build-core && \
+
+  cd /sgl-kernel && \
+  rm -rf CMakeLists.txt && mv CMakeLists_rocm.txt CMakeLists.txt && \
+  ${PYTHON_ROOT_PATH}/python rocm_hipify.py && \
+  ${PYTHON_ROOT_PATH}/python -m uv build --wheel -Cbuild-dir=build . --color=always --no-build-isolation && \
+  ./rename_wheels_rocm.sh
+"
--- a/third_party/sglang/3rdparty/amd/wheel/sgl-kernel/rename_wheels_rocm.sh
+++ b/third_party/sglang/3rdparty/amd/wheel/sgl-kernel/rename_wheels_rocm.sh
@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+set -ex
+
+WHEEL_DIR="dist"
+
+wheel_files=($WHEEL_DIR/*.whl)
+for wheel in "${wheel_files[@]}"; do
+    intermediate_wheel="${wheel/linux/manylinux2014}"
+    [[ "$intermediate_wheel" == *"+rocm"* ]] && continue
+
+    # Extract the current python version from the wheel name
+    if [[ $intermediate_wheel =~ -cp([0-9]+)- ]]; then
+        cp_version="${BASH_REMATCH[1]}"
+    else
+        echo "Could not extract Python version from wheel name: $intermediate_wheel"
+        continue
+    fi
+
+    # Detect ROCm version and add appropriate suffix
+    ver_abrv=$(realpath /opt/rocm-* | sed -e 's/.*-//' -e 's/\.//g')
+    new_wheel=${intermediate_wheel/-cp${cp_version}/+rocm${ver_abrv}-cp${cp_version}}
+
+    if [[ "$wheel" != "$new_wheel" ]]; then
+        echo "Renaming $wheel to $new_wheel"
+        mv -- "$wheel" "$new_wheel"
+    fi
+done
+echo "Wheel renaming completed."
--- a/third_party/sglang/3rdparty/amd/wheel/sgl-kernel/rocm_hipify.py
+++ b/third_party/sglang/3rdparty/amd/wheel/sgl-kernel/rocm_hipify.py
@@ -0,0 +1,41 @@
+from pathlib import Path
+
+import torch
+from torch.utils.cpp_extension import CUDAExtension
+
+root = Path(__file__).parent.resolve()
+
+include_dirs = [
+    root / "include",
+    root / "include" / "impl",
+    root / "csrc",
+]
+
+sources = [
+    "csrc/allreduce/custom_all_reduce.hip",
+    "csrc/allreduce/deterministic_all_reduce.hip",
+    "csrc/allreduce/quick_all_reduce.cu",
+    "csrc/common_extension_rocm.cc",
+    "csrc/elementwise/activation.cu",
+    "csrc/elementwise/pos_enc.cu",
+    "csrc/elementwise/topk.cu",
+    "csrc/grammar/apply_token_bitmask_inplace_cuda.cu",
+    "csrc/kvcacheio/transfer.cu",
+    "csrc/memory/weak_ref_tensor.cpp",
+    "csrc/moe/moe_align_kernel.cu",
+    "csrc/moe/moe_topk_softmax_kernels.cu",
+    "csrc/moe/moe_topk_sigmoid_kernels.cu",
+    "csrc/speculative/eagle_utils.cu",
+]
+
+libraries = ["hiprtc", "amdhip64", "c10", "torch", "torch_python"]
+
+ext_modules = [
+    CUDAExtension(
+        name="sgl_kernel.common_ops",
+        sources=sources,
+        include_dirs=include_dirs,
+        libraries=libraries,
+        py_limited_api=False,
+    ),
+]
--- a/third_party/sglang/3rdparty/amd/wheel/sglang/pyproject.toml
+++ b/third_party/sglang/3rdparty/amd/wheel/sglang/pyproject.toml
@@ -0,0 +1,213 @@
+[build-system]
+requires = ["setuptools>=61.0", "setuptools-scm>=8.0", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "amd-sglang"
+dynamic = ["version"]
+description = "SGLang is a fast serving framework for large language models and vision language models."
+readme = "README.md"
+requires-python = ">=3.10"
+license = { file = "LICENSE" }
+classifiers = [
+  "Programming Language :: Python :: 3",
+  "License :: OSI Approved :: Apache Software License",
+]
+dependencies = ["aiohttp", "requests", "tqdm", "numpy", "IPython", "setproctitle"]
+
+[project.optional-dependencies]
+runtime_common = [
+  "IPython",
+  "aiohttp",
+  "anthropic>=0.20.0",
+  "blobfile==3.0.0",
+  "build",
+  "compressed-tensors",
+  "decord2",
+  "datasets",
+  "einops",
+  "fastapi",
+  "gguf",
+  "hf_transfer",
+  "huggingface_hub",
+  "interegular",
+  "llguidance>=0.7.11,<0.8.0",
+  "modelscope",
+  "msgspec",
+  "ninja",
+  "numpy",
+  "openai-harmony==0.0.4",
+  "openai==2.6.1",
+  "orjson",
+  "outlines==0.1.11",
+  "packaging",
+  "partial_json_parser",
+  "pillow",
+  "prometheus-client>=0.20.0",
+  "psutil",
+  "py-spy",
+  "pybase64",
+  "pydantic",
+  "python-multipart",
+  "pyzmq>=25.1.2",
+  "requests",
+  "scipy",
+  "sentencepiece",
+  "setproctitle",
+  "soundfile==0.13.1",
+  "tiktoken",
+  "timm==1.0.16",
+  "torchao==0.9.0",
+  "tqdm",
+  "transformers==4.57.1",
+  "uvicorn",
+  "uvloop",
+  "xgrammar==0.1.32",
+  "smg-grpc-servicer>=0.5.0",
+]
+
+# ROCm specific packages (https://repo.radeon.com/rocm/manylinux/)
+# Existing practice for daily rocm700 docker images relies on 700-rc
+# versions of software that are not public available. Here we pin some
+# from rocm702 as the closest set as daily rocm700 images.
+rocm700 = [
+  "torch @ https://repo.radeon.com/rocm/manylinux/rocm-rel-7.0.2/torch-2.9.1.dev20251204%2Brocm7.0.2.lw.git351ff442-cp310-cp310-linux_x86_64.whl",
+  "triton @ https://repo.radeon.com/rocm/manylinux/rocm-rel-7.0.2/triton-3.5.1%2Brocm7.0.2.gita272dfa8-cp310-cp310-linux_x86_64.whl",
+  "torchaudio @ https://repo.radeon.com/rocm/manylinux/rocm-rel-7.0.2/torchaudio-2.9.0%2Brocm7.0.2.gite3c6ee2b-cp310-cp310-linux_x86_64.whl",
+  "torchvision @ https://repo.radeon.com/rocm/manylinux/rocm-rel-7.0.2/torchvision-0.24.0%2Brocm7.0.2.gitb919bd0c-cp310-cp310-linux_x86_64.whl",
+  "mooncake-transfer-engine-non-cuda==0.3.8.post1",
+  "sglang-kernel @ https://github.com/sgl-project/whl/releases/download/v0.4.0/sglang_kernel-0.4.0+rocm700-cp310-abi3-manylinux2014_x86_64.whl",
+]
+
+rocm720 = [
+  "torch @ https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/torch-2.9.1%2Brocm7.2.0.lw.git7e1940d4-cp310-cp310-linux_x86_64.whl",
+  "triton @ https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/triton-3.5.1%2Brocm7.2.0.gita272dfa8-cp310-cp310-linux_x86_64.whl",
+  "torchaudio @ https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/torchaudio-2.9.0%2Brocm7.2.0.gite3c6ee2b-cp310-cp310-linux_x86_64.whl",
+  "torchvision @ https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/torchvision-0.24.0%2Brocm7.2.0.gitb919bd0c-cp310-cp310-linux_x86_64.whl",
+  "mooncake-transfer-engine-non-cuda==0.3.8.post1",
+  "sglang-kernel @ https://github.com/sgl-project/whl/releases/download/v0.4.0/sglang_kernel-0.4.0+rocm720-cp310-abi3-manylinux2014_x86_64.whl",
+]
+
+# HIP (Heterogeneous-computing Interface for Portability) for AMD
+# Install with one of:
+#   pip install "amd-sglang[srt_hip,rocm700]"
+#   pip install "amd-sglang[srt_hip,rocm720]"
+srt_hip = [
+  "amd-sglang[runtime_common]",
+  "petit_kernel==0.0.2",
+  "wave-lang==3.8.2",
+]
+
+diffusion_hip = [
+  "PyYAML==6.0.1",
+  "cloudpickle",
+  "diffusers==0.37.0",
+  "imageio==2.36.0",
+  "imageio-ffmpeg==0.5.1",
+  "moviepy>=2.0.0",
+  "opencv-python-headless==4.10.0.84",
+  "remote-pdb",
+  "st_attn==0.0.7",
+  "vsa==0.0.4",
+  "runai_model_streamer>=0.15.5",
+  "cache-dit==1.1.8",
+  "addict",
+]
+
+# For Intel Gaudi(device : hpu) follow the installation guide
+# https://docs.vllm.ai/en/latest/getting_started/gaudi-installation.html
+srt_hpu = ["sglang[runtime_common]"]
+
+# https://docs.sglang.io/platforms/mthreads_gpu.md
+srt_musa = [
+  "sglang[runtime_common]",
+  "torch",
+  "torch_musa",
+  "torchada>=0.1.45",
+  "mthreads-ml-py",
+  "numpy<2.0",
+]
+
+diffusion_musa = [
+  "PyYAML==6.0.1",
+  "cloudpickle",
+  "diffusers==0.37.0",
+  "imageio==2.36.0",
+  "imageio-ffmpeg==0.5.1",
+  "moviepy>=2.0.0",
+  "opencv-python-headless==4.10.0.84",
+  "remote-pdb",
+  "st_attn==0.0.7",
+  "vsa==0.0.4",
+  "runai_model_streamer>=0.15.5",
+  "cache-dit==1.1.8",
+  "addict",
+]
+
+tracing = [
+  "opentelemetry-sdk",
+  "opentelemetry-api",
+  "opentelemetry-exporter-otlp",
+  "opentelemetry-exporter-otlp-proto-grpc",
+]
+
+test = [
+  "accelerate",
+  "expecttest",
+  "gguf",
+  "jsonlines",
+  "matplotlib",
+  "pandas",
+  "peft",
+  "pytest",
+  "sentence_transformers",
+  "tabulate",
+]
+
+all_hip = ["amd-sglang[srt_hip]", "amd-sglang[diffusion_hip]"]
+all_hpu = ["sglang[srt_hpu]"]
+all_musa = ["sglang[srt_musa]", "sglang[diffusion_musa]"]
+
+dev_hip = ["amd-sglang[all_hip]", "amd-sglang[test]"]
+dev_hpu = ["sglang[all_hpu]", "sglang[test]"]
+dev_musa = ["sglang[all_musa]", "sglang[test]"]
+
+[project.urls]
+"Homepage" = "https://github.com/sgl-project/sglang"
+"Bug Tracker" = "https://github.com/sgl-project/sglang/issues"
+
+[project.scripts]
+sglang = "sglang.cli.main:main"
+
+[tool.setuptools.package-data]
+"sglang" = [
+  "srt/**/*",
+  "jit_kernel/**/*",
+]
+
+[tool.setuptools.packages.find]
+exclude = [
+  "assets*",
+  "benchmark*",
+  "docs*",
+  "dist*",
+  "playground*",
+  "scripts*",
+  "tests*",
+]
+
+[tool.wheel]
+exclude = [
+  "assets*",
+  "benchmark*",
+  "docs*",
+  "dist*",
+  "playground*",
+  "scripts*",
+  "tests*",
+]
+
+[tool.setuptools_scm]
+root = ".."
+version_file = "sglang/_version.py"
+git_describe_command = ["git", "describe", "--tags", "--long", "--match", "v*"]