diff --git a/.github/workflows/release-whl-kernel.yml b/.github/workflows/release-whl-kernel.yml
index 9e6cb3717..3a19fa260 100644
--- a/.github/workflows/release-whl-kernel.yml
+++ b/.github/workflows/release-whl-kernel.yml
@@ -18,6 +18,7 @@ on:
           - 'cu129'
           - 'cu130'
           - 'rocm700'
+          - 'rocm720'
           - 'musa43'
       tag_name:
         type: string
@@ -223,15 +224,15 @@ jobs:
           git commit -m "update whl index"
           git push
 
-  build-rocm700:
+  build-rocm-matrix:
     if: |
       github.repository == 'sgl-project/sglang' &&
-      (github.event.inputs.target == 'all' || github.event.inputs.target == 'rocm700')
+      (github.event.inputs.target == 'all' || github.event.inputs.target == 'rocm700' || github.event.inputs.target == 'rocm720')
     runs-on: amd-docker-scale
     strategy:
       matrix:
         python-version: ["3.10"]
-        rocm-version: ["700"]
+        rocm-version: ["700", "720"]
     steps:
       - uses: actions/checkout@v4
         with:
@@ -245,7 +246,7 @@ jobs:
 
       - name: Build wheels
         run: |
-          cp 3rdparty/amd/sgl-kernel/* sgl-kernel/
+          cp 3rdparty/amd/wheel/sgl-kernel/* sgl-kernel/
           cd sgl-kernel
           chmod +x ./build_rocm.sh
           ./build_rocm.sh "${{ matrix.rocm-version }}"
@@ -257,7 +258,7 @@ jobs:
           path: sgl-kernel/dist/*
 
   release-rocm700:
-    needs: build-rocm700
+    needs: build-rocm-matrix
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4
@@ -269,7 +270,7 @@ jobs:
         with:
           path: sgl-kernel/dist/
           merge-multiple: true
-          pattern: wheel-*
+          pattern: wheel-*-rocm700
 
       - name: Set tag name
         id: set_tag_name
@@ -307,6 +308,57 @@ jobs:
           git commit -m "update whl index"
           git push
 
+  release-rocm720:
+    needs: build-rocm-matrix
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.pr_number && format('refs/pull/{0}/head', inputs.pr_number) || '' }}
+
+      - name: Download artifacts
+        uses: actions/download-artifact@v4
+        with:
+          path: sgl-kernel/dist/
+          merge-multiple: true
+          pattern: wheel-*-rocm720
+
+      - name: Set tag name
+        id: set_tag_name
+        run: |
+          if [ -z "${{ inputs.tag_name }}" ]; then
+            TAG_NAME="v$(cat sgl-kernel/python/sgl_kernel/version.py | cut -d'"' -f2)"
+            echo "tag_name=$TAG_NAME" >> $GITHUB_OUTPUT
+          else
+            echo "tag_name=${{ inputs.tag_name }}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Release
+        uses: softprops/action-gh-release@v2
+        with:
+          tag_name: ${{ steps.set_tag_name.outputs.tag_name }}
+          repository: sgl-project/whl
+          token: ${{ secrets.GH_PAT_FOR_WHL_RELEASE }}
+          files: |
+            sgl-kernel/dist/*
+
+      - name: Clone wheel index
+        run: git clone https://oauth2:${WHL_TOKEN}@github.com/sgl-project/whl.git sgl-whl
+        env:
+          WHL_TOKEN: ${{ secrets.GH_PAT_FOR_WHL_RELEASE }}
+
+      - name: Update wheel index
+        run: python3 scripts/update_kernel_whl_index.py --rocm 720
+
+      - name: Push wheel index
+        run: |
+          cd sgl-whl
+          git config --local user.name "sglang-bot"
+          git config --local user.email "sglangbot@gmail.com"
+          git add -A
+          git commit -m "update whl index"
+          git push
+
   build-musa43:
     if: |
       github.repository == 'sgl-project/sglang' &&
diff --git a/3rdparty/amd/sgl-kernel/build_rocm.sh b/3rdparty/amd/sgl-kernel/build_rocm.sh
deleted file mode 100755
index 1022d8bb5..000000000
--- a/3rdparty/amd/sgl-kernel/build_rocm.sh
+++ /dev/null
@@ -1,123 +0,0 @@
-#!/bin/bash
-set -euo pipefail
-ROCM_VERSION=$1
-
-PYTHON_ROOT_PATH="/opt/venv/bin"
-AMDGPU_TARGET="gfx942;gfx950"
-
-echo "Python root path is: $PYTHON_ROOT_PATH"
-
-# Get version from git tags
-SGLANG_VERSION="v0.5.6"   # Default version, will be overridden if git tags are found
-
-# Fetch tags from origin to ensure we have the latest
-if git fetch --tags origin; then
-  # Get the latest version tag sorted by version number (e.g., v0.5.7)
-  VERSION_FROM_TAG=$(git tag -l 'v[0-9]*' --sort=-v:refname | head -1)
-  if [ -n "$VERSION_FROM_TAG" ]; then
-    SGLANG_VERSION="$VERSION_FROM_TAG"
-    echo "Using SGLang version from git tags: $SGLANG_VERSION"
-  else
-    echo "Warning: No version tags found; using default $SGLANG_VERSION" >&2
-  fi
-else
-  echo "Warning: Failed to fetch tags from origin; using default $SGLANG_VERSION" >&2
-fi
-
-# Default base tags (can be overridden by command line arguments)
-DEFAULT_MI30X_BASE_TAG="${SGLANG_VERSION}-rocm700-mi30x"
-DEFAULT_MI35X_BASE_TAG="${SGLANG_VERSION}-rocm700-mi35x"
-
-# Parse command line arguments
-MI30X_BASE_TAG="${DEFAULT_MI30X_BASE_TAG}"
-MI35X_BASE_TAG="${DEFAULT_MI35X_BASE_TAG}"
-
-# Detect GPU architecture from the Kubernetes runner hostname
-HOSTNAME_VALUE=$(hostname)
-GPU_ARCH="mi30x"   # default
-
-# Host names look like: linux-mi35x-gpu-1-xxxxx-runner-zzzzz
-if [[ "${HOSTNAME_VALUE}" =~ ^linux-(mi[0-9]+[a-z]*)-gpu-[0-9]+ ]]; then
-  GPU_ARCH="${BASH_REMATCH[1]}"
-  echo "Detected GPU architecture from hostname: ${GPU_ARCH}"
-else
-  echo "Warning: could not parse GPU architecture from '${HOSTNAME_VALUE}', defaulting to ${GPU_ARCH}"
-fi
-
-case "${GPU_ARCH}" in
-  mi35x)
-    echo "Runner uses ${GPU_ARCH}; will fetch mi35x image."
-    ;;
-  mi30x|mi300|mi325)
-    echo "Runner uses ${GPU_ARCH}; will fetch mi30x image."
-    GPU_ARCH="mi30x"
-    ;;
-  *)
-    echo "Runner architecture '${GPU_ARCH}' unrecognised; defaulting to mi30x image." >&2
-    GPU_ARCH="mi30x"
-    ;;
-esac
-
-if [[ -f /etc/podinfo/gha-render-devices ]]; then
-  DEVICE_FLAG=$(cat /etc/podinfo/gha-render-devices)
-else
-  DEVICE_FLAG="--device /dev/dri"
-fi
-
-# Find the latest image
-find_latest_image() {
-  local gpu_arch=$1
-  local base_tag days_back image_tag
-
-  case "${gpu_arch}" in
-      mi30x) base_tag="${MI30X_BASE_TAG}" ;;
-      mi35x) base_tag="${MI35X_BASE_TAG}" ;;
-      *)     echo "Error: unsupported GPU architecture '${gpu_arch}'" >&2; return 1 ;;
-  esac
-
-  for days_back in {0..6}; do
-    image_tag="${base_tag}-$(date -d "${days_back} days ago" +%Y%m%d)"
-    echo "Checking for image: rocm/sgl-dev:${image_tag}" >&2
-    if docker manifest inspect "rocm/sgl-dev:${image_tag}" >/dev/null 2>&1; then
-      echo "Found available image: rocm/sgl-dev:${image_tag}" >&2
-      echo "rocm/sgl-dev:${image_tag}"
-      return 0
-    fi
-  done
-
-  echo "Error: no ${gpu_arch} image found in the last 7 days for base ${base_tag}" >&2
-  echo "Using hard-coded fallback…" >&2
-  if [[ "${gpu_arch}" == "mi35x" ]]; then
-    echo "rocm/sgl-dev:v0.5.3-rocm700-mi35x-20251009"
-  else
-    echo "rocm/sgl-dev:v0.5.3-rocm700-mi30x-20251009"
-  fi
-}
-
-# Pull and run the latest image
-IMAGE=$(find_latest_image "${GPU_ARCH}")
-echo "Pulling Docker image: ${IMAGE}"
-docker pull "${IMAGE}"
-
-docker run --rm \
-  -v $(pwd):/sgl-kernel \
-  -e AMDGPU_TARGET="${AMDGPU_TARGET}" \
-  ${IMAGE} \
-  bash -c "
-  # 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 && \
-  ${PYTHON_ROOT_PATH}/python -m uv build --wheel -Cbuild-dir=build . --color=always --no-build-isolation && \
-  ./rename_wheels_rocm.sh
-"
diff --git a/3rdparty/amd/wheel/README.md b/3rdparty/amd/wheel/README.md
new file mode 100644
index 000000000..53d118e44
--- /dev/null
+++ b/3rdparty/amd/wheel/README.md
@@ -0,0 +1,97 @@
+# sgl-kernel
+
+Building and releasing sgl-kernel as a wheel is a part of the release workflow.  Check [release-whl-kernel.yml](https://github.com/sgl-project/sglang/blob/main/.github/workflows/release-whl-kernel.yml) for details.
+
+# sglang
+
+`3rdparty/amd/wheel/sglang/pyproject.toml` is the AMD-specific pyproject for building the `amd-sglang` wheel.  It extends `python/pyproject_other.toml` with two ROCm-version extras (`rocm700`, `rocm720`) that pin the matching torch/triton/torchaudio/torchvision/sgl-kernel wheels, and renames the package to `amd-sglang`.
+
+## Operation to build sglang wheel
+
+```
+$ 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 sgl-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 sgl-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)
diff --git a/3rdparty/amd/sgl-kernel/CMakeLists_rocm.txt b/3rdparty/amd/wheel/sgl-kernel/CMakeLists_rocm.txt
similarity index 97%
rename from 3rdparty/amd/sgl-kernel/CMakeLists_rocm.txt
rename to 3rdparty/amd/wheel/sgl-kernel/CMakeLists_rocm.txt
index 86ba249c1..eb9f1e40d 100644
--- a/3rdparty/amd/sgl-kernel/CMakeLists_rocm.txt
+++ b/3rdparty/amd/wheel/sgl-kernel/CMakeLists_rocm.txt
@@ -36,6 +36,7 @@ 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
@@ -106,6 +107,7 @@ ${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
 ${PROJ_ROOT}/csrc/moe/moe_topk_sigmoid_kernels.hip
diff --git a/3rdparty/amd/wheel/sgl-kernel/build_rocm.sh b/3rdparty/amd/wheel/sgl-kernel/build_rocm.sh
new file mode 100755
index 000000000..4347737fa
--- /dev/null
+++ b/3rdparty/amd/wheel/sgl-kernel/build_rocm.sh
@@ -0,0 +1,50 @@
+#!/bin/bash
+set -euo pipefail
+
+ROCM_VERSION=${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 && \
+  ${PYTHON_ROOT_PATH}/python -m uv build --wheel -Cbuild-dir=build . --color=always --no-build-isolation && \
+  ./rename_wheels_rocm.sh
+"
diff --git a/3rdparty/amd/sgl-kernel/rename_wheels_rocm.sh b/3rdparty/amd/wheel/sgl-kernel/rename_wheels_rocm.sh
similarity index 75%
rename from 3rdparty/amd/sgl-kernel/rename_wheels_rocm.sh
rename to 3rdparty/amd/wheel/sgl-kernel/rename_wheels_rocm.sh
index 691407a3e..3c8fb2a63 100755
--- a/3rdparty/amd/sgl-kernel/rename_wheels_rocm.sh
+++ b/3rdparty/amd/wheel/sgl-kernel/rename_wheels_rocm.sh
@@ -6,6 +6,7 @@ WHEEL_DIR="dist"
 wheel_files=($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 [[ $intermediate_wheel =~ -cp([0-9]+)- ]]; then
@@ -16,11 +17,8 @@ for wheel in "${wheel_files[@]}"; do
     fi
 
     # Detect ROCm version and add appropriate suffix
-    if ls /opt | grep -q "7.0"; then
-        new_wheel="${intermediate_wheel/-cp${cp_version}/+rocm700-cp${cp_version}}"
-    else
-        new_wheel="$intermediate_wheel"
-    fi
+    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"
diff --git a/3rdparty/amd/sgl-kernel/rocm_hipify.py b/3rdparty/amd/wheel/sgl-kernel/rocm_hipify.py
similarity index 96%
rename from 3rdparty/amd/sgl-kernel/rocm_hipify.py
rename to 3rdparty/amd/wheel/sgl-kernel/rocm_hipify.py
index 8373f741d..7383408be 100644
--- a/3rdparty/amd/sgl-kernel/rocm_hipify.py
+++ b/3rdparty/amd/wheel/sgl-kernel/rocm_hipify.py
@@ -21,6 +21,7 @@ sources = [
     "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",
diff --git a/3rdparty/amd/wheel/sglang/pyproject.toml b/3rdparty/amd/wheel/sglang/pyproject.toml
new file mode 100644
index 000000000..6c2fc153a
--- /dev/null
+++ b/3rdparty/amd/wheel/sglang/pyproject.toml
@@ -0,0 +1,215 @@
+[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==0.1.27",
+  "smg-grpc-proto>=0.3.3",
+  "grpcio>=1.78.0",
+  "grpcio-reflection>=1.78.0",
+]
+
+# 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",
+  "sgl-kernel @ https://github.com/sgl-project/whl/releases/download/v0.3.21/sgl_kernel-0.3.21+rocm700-cp310-abi3-manylinux2014_x86_64.whl#sha256=2d56ee97d8e6dae29d35a2524de4bffb8e07752bcc74e73313462496427832bf",
+]
+
+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",
+  "sgl-kernel @ https://github.com/sgl-project/whl/releases/download/v0.3.21/sgl_kernel-0.3.21+rocm720-cp310-abi3-manylinux2014_x86_64.whl#sha256=91a3f50ef0b9d2a140dbf6e2aabff45a50c45a4782b9215997bce6f71749180f",
+]
+
+# 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==0.36.0",
+  "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>=0.1.25",
+  "mthreads-ml-py",
+  "numpy<2.0",
+]
+
+diffusion_musa = [
+  "PyYAML==6.0.1",
+  "cloudpickle",
+  "diffusers==0.36.0",
+  "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*"]