diff --git a/.github/workflows/pr-test-amd-extra.yml b/.github/workflows/pr-test-amd-extra.yml
index fb964f03bfdd..029a51970a0f 100644
--- a/.github/workflows/pr-test-amd-extra.yml
+++ b/.github/workflows/pr-test-amd-extra.yml
@@ -37,18 +37,24 @@ on:
           - mi300
           - mi325
       rocm_version:
-        description: 'ROCm container variant (rocm720 = ROCm 7.2.0, the PR default; rocm700 = ROCm 7.0.0)'
+        description: 'ROCm container variant (ROCm 7.2.0 by default)'
         required: false
         type: choice
         default: rocm720
         options:
           - rocm720
+          - rocm724
           - rocm700
       aiter_ref:
         description: 'Override AITER commit (optional, leave empty to use Dockerfile default)'
         required: false
         type: string
         default: ''
+      amd_ci_image:
+        description: 'Override AMD CI Docker image (optional, e.g. rocm/sgl-dev:tag)'
+        required: false
+        type: string
+        default: ''
       continue_on_error:
         description: 'Continue on error (do not fail the workflow on test failures)'
         required: false
@@ -67,7 +73,7 @@ on:
         type: string
         default: mi300
       rocm_version:
-        description: 'ROCm container variant (rocm720 = ROCm 7.2.0, the PR default; rocm700 = ROCm 7.0.0)'
+        description: 'ROCm container variant (ROCm 7.2.0 by default)'
         required: false
         type: string
         default: rocm720
@@ -76,6 +82,11 @@ on:
         required: false
         type: string
         default: ''
+      amd_ci_image:
+        description: 'Override AMD CI Docker image (optional, e.g. rocm/sgl-dev:tag)'
+        required: false
+        type: string
+        default: ''
       continue_on_error:
         description: 'Continue on error (do not fail the workflow on test failures)'
         required: false
@@ -84,6 +95,7 @@ on:
 
 env:
   AITER_COMMIT_OVERRIDE: ${{ inputs.aiter_ref }}
+  AMD_CI_IMAGE: ${{ inputs.amd_ci_image }}
   DOCKERHUB_AMD_USERNAME: ${{ secrets.DOCKERHUB_AMD_USERNAME }}
   DOCKERHUB_AMD_TOKEN: ${{ secrets.DOCKERHUB_AMD_TOKEN }}
 
@@ -145,7 +157,7 @@ jobs:
 
       - name: Start CI container
         # `inputs` is empty on pull_request events, so fall back explicitly to
-        # the ROCm 7.2 PR default. Reusable callers can still request ROCm 7.0.
+        # the ROCm 7.2 PR default. Reusable callers can request another variant.
         run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version ${{ inputs.rocm_version || 'rocm720' }}
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
diff --git a/.github/workflows/pr-test-amd-rocm720.yml b/.github/workflows/pr-test-amd-rocm720.yml
index 699d612114b6..ec6ac69572ea 100644
--- a/.github/workflows/pr-test-amd-rocm720.yml
+++ b/.github/workflows/pr-test-amd-rocm720.yml
@@ -16,6 +16,14 @@ on:
       - "docker/rocm.Dockerfile"
   workflow_dispatch:
     inputs:
+      rocm_version:
+        description: 'ROCm image version'
+        required: false
+        type: choice
+        default: rocm720
+        options:
+          - rocm720
+          - rocm724
       target_stage_select:
         description: "Select a stage to run from dropdown (leave empty for auto-detect)"
         required: false
@@ -57,6 +65,11 @@ on:
         required: false
         type: string
         default: ''
+      amd_ci_image:
+        description: 'Override AMD CI Docker image (optional, e.g. rocm/sgl-dev:tag)'
+        required: false
+        type: string
+        default: ''
       continue_on_error:
         description: 'Continue on error (do not fail the workflow on test failures)'
         required: false
@@ -77,6 +90,11 @@ on:
         default: false
   workflow_call:
     inputs:
+      rocm_version:
+        description: 'ROCm image version'
+        required: false
+        type: string
+        default: rocm720
       ref:
         description: 'Git ref (branch, tag, or SHA) to test. If not provided, uses the default branch.'
         required: false
@@ -97,6 +115,11 @@ on:
         required: false
         type: string
         default: ''
+      amd_ci_image:
+        description: 'Override AMD CI Docker image (optional, e.g. rocm/sgl-dev:tag)'
+        required: false
+        type: string
+        default: ''
       continue_on_error:
         description: 'Continue on error (do not fail the workflow on test failures)'
         required: false
@@ -116,6 +139,8 @@ permissions:
 
 env:
   AITER_COMMIT_OVERRIDE: ${{ inputs.aiter_ref }}
+  AMD_CI_IMAGE: ${{ inputs.amd_ci_image }}
+  ROCM_VERSION: ${{ inputs.rocm_version || 'rocm720' }}
   DOCKERHUB_AMD_USERNAME: ${{ secrets.DOCKERHUB_AMD_USERNAME }}
   DOCKERHUB_AMD_TOKEN: ${{ secrets.DOCKERHUB_AMD_TOKEN }}
 
@@ -225,8 +250,9 @@ jobs:
     with:
       ref: ${{ inputs.pr_head_sha || inputs.ref || '' }}
       runner_arch: ${{ inputs.runner_arch || 'mi300' }}
-      rocm_version: rocm720
+      rocm_version: ${{ inputs.rocm_version || 'rocm720' }}
       aiter_ref: ${{ inputs.aiter_ref }}
+      amd_ci_image: ${{ inputs.amd_ci_image }}
       continue_on_error: true
     secrets: inherit
 
@@ -255,7 +281,7 @@ jobs:
         run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -313,7 +339,7 @@ jobs:
         run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -354,7 +380,7 @@ jobs:
         run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -391,7 +417,7 @@ jobs:
         run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -428,7 +454,7 @@ jobs:
         run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -496,7 +522,7 @@ jobs:
         run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -532,7 +558,7 @@ jobs:
         run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -571,7 +597,7 @@ jobs:
         run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -612,7 +638,7 @@ jobs:
         run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -653,7 +679,7 @@ jobs:
         run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -702,7 +728,7 @@ jobs:
           pattern: wheel-python3.10-cuda12.9
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -833,7 +859,7 @@ jobs:
           pattern: wheel-python3.10-cuda12.9
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -963,7 +989,7 @@ jobs:
           pattern: wheel-python3.10-cuda12.9
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -1040,7 +1066,7 @@ jobs:
         run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -1099,7 +1125,7 @@ jobs:
         run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -1145,7 +1171,7 @@ jobs:
         run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
 
       - name: Start CI container
-        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
@@ -1189,7 +1215,7 @@ jobs:
       - name: Setup docker (ROCm 7.2)
         run: |
           touch github_summary.md
-          bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+          bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
           ENABLE_CACHE_HOST: "1"
@@ -1243,7 +1269,7 @@ jobs:
       - name: Setup docker (ROCm 7.2)
         run: |
           touch github_summary.md
-          bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
+          bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
           ENABLE_CACHE_HOST: "1"
@@ -1336,7 +1362,7 @@ jobs:
           echo "=== Host RDMA Check Complete ==="
 
       - name: Start Special Container
-        run: bash scripts/ci/amd/amd_ci_start_container_disagg.sh --rocm-version rocm720
+        run: bash scripts/ci/amd/amd_ci_start_container_disagg.sh --rocm-version "$ROCM_VERSION"
         env:
           GITHUB_WORKSPACE: ${{ github.workspace }}
 
diff --git a/.github/workflows/release-docker-amd-rocm720-nightly.yml b/.github/workflows/release-docker-amd-rocm720-nightly.yml
index 8126ed7d66ed..b87edbad97cd 100644
--- a/.github/workflows/release-docker-amd-rocm720-nightly.yml
+++ b/.github/workflows/release-docker-amd-rocm720-nightly.yml
@@ -10,6 +10,17 @@ on:
         options:
           - 'all'
           - publish
+      gpu_arch:
+        description: 'Select which ROCm 7.2 GPU arch to build'
+        required: false
+        type: choice
+        default: 'all'
+        options:
+          - 'all'
+          - gfx942-rocm724
+          - gfx950-rocm724
+          - gfx942-rocm720
+          - gfx950-rocm720
   schedule:
     - cron: '0 12 * * *'
 
@@ -29,7 +40,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        gpu_arch: ['gfx942-rocm720', 'gfx950-rocm720']
+        gpu_arch: ${{ fromJson((github.event_name == 'workflow_dispatch' && inputs.gpu_arch != 'all' && inputs.gpu_arch != '') && format('["{0}"]', inputs.gpu_arch) || '["gfx942-rocm720", "gfx950-rocm720", "gfx942-rocm724", "gfx950-rocm724"]') }}
         build_type: ['all']
     steps:
       - name: Checkout repository
@@ -81,7 +92,11 @@ jobs:
           echo "Version: ${version}"
           echo "Pretend version: ${pretend_version}"
 
-          if [ "${{ matrix.gpu_arch }}" = "gfx942-rocm720" ]; then
+          if [ "${{ matrix.gpu_arch }}" = "gfx942-rocm724" ]; then
+            rocm_tag="rocm724-mi30x"
+          elif [ "${{ matrix.gpu_arch }}" = "gfx950-rocm724" ]; then
+            rocm_tag="rocm724-mi35x"
+          elif [ "${{ matrix.gpu_arch }}" = "gfx942-rocm720" ]; then
             rocm_tag="rocm720-mi30x"
           elif [ "${{ matrix.gpu_arch }}" = "gfx950-rocm720" ]; then
             rocm_tag="rocm720-mi35x"
@@ -141,7 +156,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        gpu_arch: ['gfx942-rocm720', 'gfx950-rocm720']
+        gpu_arch: ${{ fromJson((github.event_name == 'workflow_dispatch' && inputs.gpu_arch != 'all' && inputs.gpu_arch != '') && format('["{0}"]', inputs.gpu_arch) || '["gfx942-rocm720", "gfx950-rocm720", "gfx942-rocm724", "gfx950-rocm724"]') }}
     steps:
       - name: Download image tag artifact
         uses: actions/download-artifact@v4
diff --git a/.github/workflows/release-docker-amd.yml b/.github/workflows/release-docker-amd.yml
index 6920338014f6..e5cd66ca35b1 100644
--- a/.github/workflows/release-docker-amd.yml
+++ b/.github/workflows/release-docker-amd.yml
@@ -8,6 +8,16 @@ on:
       version:
         description: 'Version to build (without v prefix, e.g., 0.5.7)'
         required: true
+      rocm_version:
+        description: 'ROCm version to build'
+        required: false
+        type: choice
+        default: 'all'
+        options:
+          - 'all'
+          - rocm724
+          - rocm720
+          - rocm700
 
 jobs:
   publish:
@@ -15,8 +25,11 @@ jobs:
     runs-on: amd-docker-scale
     environment: 'prod'
     strategy:
+      # Six flavors publish independently: a failure in one must not cancel the
+      # others mid-push and leave the release tag with a partial image set.
+      fail-fast: false
       matrix:
-        rocm_version: ['rocm700', 'rocm720']
+        rocm_version: ${{ fromJson((github.event_name == 'workflow_dispatch' && inputs.rocm_version != 'all' && inputs.rocm_version != '') && format('["{0}"]', inputs.rocm_version) || '["rocm700", "rocm720", "rocm724"]') }}
         gpu_arch: ['gfx942', 'gfx950']
         build_type: ['all']
     steps:
@@ -66,12 +79,12 @@ jobs:
               echo "Unsupported gfx arch"
               exit 1
             fi
-          elif [ "${{ matrix.rocm_version }}" = "rocm720" ]; then
+          elif [ "${{ matrix.rocm_version }}" = "rocm720" ] || [ "${{ matrix.rocm_version }}" = "rocm724" ]; then
             gpu_arch_suffix="-${{ matrix.rocm_version }}"
             if [ "${{ matrix.gpu_arch }}" = "gfx942" ]; then
-              rocm_tag="rocm720-mi30x"
+              rocm_tag="${{ matrix.rocm_version }}-mi30x"
             elif [ "${{ matrix.gpu_arch }}" = "gfx950" ]; then
-              rocm_tag="rocm720-mi35x"
+              rocm_tag="${{ matrix.rocm_version }}-mi35x"
             else
               echo "Unsupported gfx arch"
               exit 1
diff --git a/docker/rocm.Dockerfile b/docker/rocm.Dockerfile
index b0123cef9562..0e6152588db7 100644
--- a/docker/rocm.Dockerfile
+++ b/docker/rocm.Dockerfile
@@ -1,8 +1,15 @@
 # Usage (to build SGLang ROCm docker image):
 #   docker build --build-arg SGL_BRANCH=v0.5.10.post1 --build-arg GPU_ARCH=gfx942 -t v0.5.10.post1-rocm700-mi30x -f rocm.Dockerfile .
 #   docker build --build-arg SGL_BRANCH=v0.5.10.post1 --build-arg GPU_ARCH=gfx942-rocm720 -t v0.5.10.post1-rocm720-mi30x -f rocm.Dockerfile .
+#   docker build --build-arg SGL_BRANCH=v0.5.10.post1 --build-arg GPU_ARCH=gfx942-rocm724 -t v0.5.10.post1-rocm724-mi30x -f rocm.Dockerfile .
 #   docker build --build-arg SGL_BRANCH=v0.5.10.post1 --build-arg GPU_ARCH=gfx950 -t v0.5.10.post1-rocm700-mi35x -f rocm.Dockerfile .
 #   docker build --build-arg SGL_BRANCH=v0.5.10.post1 --build-arg GPU_ARCH=gfx950-rocm720 -t v0.5.10.post1-rocm720-mi35x -f rocm.Dockerfile .
+#   docker build --build-arg SGL_BRANCH=v0.5.10.post1 --build-arg GPU_ARCH=gfx950-rocm724 -t v0.5.10.post1-rocm724-mi35x -f rocm.Dockerfile .
+#
+# Flavor notes:
+#   GPU_ARCH=*-rocm724 is built on a Python 3.12 base and upgrades the stack to
+#   torch 2.11 (+torchvision 0.26 / torchaudio 2.11) and Triton 3.7.
+#   The ROCm 7.2.0 flavors remain on Python 3.10 and torch 2.9.1.
 
 # Usage (to build SGLang ROCm + Mori docker image):
 # remove --build-arg NIC_BACKEND=ainic since new MoRI JIT will do NIC auto detection on target
@@ -23,8 +30,10 @@
 # Default base images
 ARG BASE_IMAGE_942="rocm/sgl-dev:rocm7-vllm-20250904"
 ARG BASE_IMAGE_942_ROCM720="rocm/pytorch:rocm7.2_ubuntu22.04_py3.10_pytorch_release_2.9.1"
+ARG BASE_IMAGE_942_ROCM724="rocm/pytorch:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0"
 ARG BASE_IMAGE_950="rocm/sgl-dev:rocm7-vllm-20250904"
 ARG BASE_IMAGE_950_ROCM720="rocm/pytorch:rocm7.2_ubuntu22.04_py3.10_pytorch_release_2.9.1"
+ARG BASE_IMAGE_950_ROCM724="rocm/pytorch:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0"
 
 # This is necessary for scope purpose
 ARG GPU_ARCH=gfx950
@@ -49,6 +58,30 @@ ENV BUILD_AITER_ALL="1"
 ENV BUILD_MOONCAKE="1"
 ENV AITER_COMMIT_DEFAULT="d9e5ef7ce08ee7045d583aed768cff41aa9210fe"
 
+# ===============================
+# Base image 942 with rocm724 and args (Python 3.12 + torch 2.11)
+FROM $BASE_IMAGE_942_ROCM724 AS gfx942-rocm724
+ENV BUILD_VLLM="0"
+ENV BUILD_TRITON="1"
+ENV BUILD_LLVM="0"
+ENV BUILD_AITER_ALL="1"
+ENV BUILD_MOONCAKE="1"
+ENV AITER_COMMIT_DEFAULT="d9e5ef7ce08ee7045d583aed768cff41aa9210fe"
+# Pin the ROCm torch stack for every pip invocation in this flavor. The file is
+# filled in after the torch 2.11 upgrade below; it must already exist (empty is
+# valid) because pip reads PIP_CONSTRAINT from the first pip call onwards.
+# Deliberately still set in the shipped image, not just during the build: a later
+# `pip install` that resolves torch would otherwise pull the PyPI CUDA build over
+# this ROCm one, and the constraint turns that into a resolution error instead.
+# It names only torch / torchvision / torchaudio, so nothing else is constrained.
+ENV PIP_CONSTRAINT="/etc/sglang/constraints/torch-rocm.txt"
+RUN mkdir -p /etc/sglang/constraints && : > /etc/sglang/constraints/torch-rocm.txt
+# Work around ROCM-21485: the CUDA/ROCm IPC path leaks GPU memory (a freed IPC
+# block is not returned to the driver). Legacy IPC mode releases it. Verified on
+# ROCm 7.2.1 and 7.2.4; scoped to this flavor so rocm700 / rocm720 keep current
+# IPC behavior.
+ENV HSA_ENABLE_IPC_MODE_LEGACY=1
+
 # ===============================
 # Base image 950 and args
 FROM $BASE_IMAGE_950 AS gfx950
@@ -69,6 +102,30 @@ ENV BUILD_AITER_ALL="1"
 ENV BUILD_MOONCAKE="1"
 ENV AITER_COMMIT_DEFAULT="d9e5ef7ce08ee7045d583aed768cff41aa9210fe"
 
+# ===============================
+# Base image 950 with rocm724 and args (Python 3.12 + torch 2.11)
+FROM $BASE_IMAGE_950_ROCM724 AS gfx950-rocm724
+ENV BUILD_VLLM="0"
+ENV BUILD_TRITON="1"
+ENV BUILD_LLVM="0"
+ENV BUILD_AITER_ALL="1"
+ENV BUILD_MOONCAKE="1"
+ENV AITER_COMMIT_DEFAULT="d9e5ef7ce08ee7045d583aed768cff41aa9210fe"
+# Pin the ROCm torch stack for every pip invocation in this flavor. The file is
+# filled in after the torch 2.11 upgrade below; it must already exist (empty is
+# valid) because pip reads PIP_CONSTRAINT from the first pip call onwards.
+# Deliberately still set in the shipped image, not just during the build: a later
+# `pip install` that resolves torch would otherwise pull the PyPI CUDA build over
+# this ROCm one, and the constraint turns that into a resolution error instead.
+# It names only torch / torchvision / torchaudio, so nothing else is constrained.
+ENV PIP_CONSTRAINT="/etc/sglang/constraints/torch-rocm.txt"
+RUN mkdir -p /etc/sglang/constraints && : > /etc/sglang/constraints/torch-rocm.txt
+# Work around ROCM-21485: the CUDA/ROCm IPC path leaks GPU memory (a freed IPC
+# block is not returned to the driver). Legacy IPC mode releases it. Verified on
+# ROCm 7.2.1 and 7.2.4; scoped to this flavor so rocm700 / rocm720 keep current
+# IPC behavior.
+ENV HSA_ENABLE_IPC_MODE_LEGACY=1
+
 # Local source stage: with BRANCH_TYPE=local the build context is copied here and
 # used instead of git clone (mirrors docker/Dockerfile's local_src stage).
 FROM scratch AS local_src
@@ -80,6 +137,11 @@ FROM ${GPU_ARCH}
 
 # This is necessary for scope purpose, again
 ARG GPU_ARCH=gfx950
+# ARG is build-time only. Stamp the stage name (gfx950-rocm724, gfx942, ...)
+# so CI can read which AITER_COMMIT_DEFAULT block to use instead of guessing
+# from torch or HIP — 720 may also ship torch 2.11 later, and both 7.2 flavors
+# report HIP 7.2*.
+ENV GPU_ARCH=${GPU_ARCH}
 ENV GPU_ARCH_LIST=${GPU_ARCH%-*}
 ENV PYTORCH_ROCM_ARCH=gfx942;gfx950
 
@@ -91,6 +153,22 @@ ARG BRANCH_TYPE=remote
 # Version override for setuptools_scm (used in nightly builds)
 ARG SETUPTOOLS_SCM_PRETEND_VERSION=""
 
+# ROCm 7.2 Triton (BUILD_TRITON=1 stages only). Both wheels are the same
+# upstream revision, triton-lang/triton@89002410. AITER only requires
+# triton>=3.6.0 and treats the base image as the owner of the version, so the
+# choice is ours; bump these together after checking the index.
+ARG TRITON_INDEX_URL="https://pypi.amd.com/triton/release/rocm-7.2.0/simple/"
+ARG TRITON_VERSION="3.7.0+amd.rocm7.2.0.git89002410"
+ARG TRITON_KERNELS_VERSION="1.0.0+amd.rocm7.2.0.git89002410"
+
+# ROCm 7.2.4 torch upgrade pins (Python 3.12). Torch 2.11 for ROCm 7.2 is only
+# published on the PyTorch Foundation index; AMD's repo.radeon.com wheels top
+# out at torch 2.10.
+ARG TORCH_ROCM_INDEX_URL="https://download.pytorch.org/whl/rocm7.2"
+ARG TORCH_ROCM_VERSION="2.11.0+rocm7.2"
+ARG TORCHVISION_ROCM_VERSION="0.26.0+rocm7.2"
+ARG TORCHAUDIO_ROCM_VERSION="2.11.0+rocm7.2"
+
 ARG AITER_REPO="https://github.com/ROCm/aiter.git"
 ARG AITER_COMMIT=""
 ENV AITER_COMMIT="${AITER_COMMIT:-${AITER_COMMIT_DEFAULT}}"
@@ -131,14 +209,16 @@ ARG UBUNTU_CODENAME=jammy
 # Optional Ubuntu mirror override + apt hardening.
 # - UBUNTU_MIRROR is empty by default (no behaviour change for local builds).
 #   When set (typically in CI), all http://*archive.ubuntu.com and
-#   http://*security.ubuntu.com entries in /etc/apt/sources.list are rewritten
-#   to point at the given base URL, e.g.
+#   http://*security.ubuntu.com entries in every /etc/apt source file are
+#   rewritten to point at the given base URL, e.g.
 #     --build-arg UBUNTU_MIRROR=https://archive.ubuntu.com
 #     --build-arg UBUNTU_MIRROR=https://tw.archive.ubuntu.com
 #     --build-arg UBUNTU_MIRROR=http://internal-cache.example.com
 #   This mirrors the pattern already used in docker/Dockerfile (NVIDIA) and
 #   docker/npu.Dockerfile, and lets CI runners that cannot reach Canonical's
-#   port-80 mirror IPs still complete `apt-get update`.
+#   port-80 mirror IPs still complete `apt-get update`. Every file, not just
+#   sources.list: the noble base used by rocm724 keeps its URIs in the deb822
+#   /etc/apt/sources.list.d/ubuntu.sources instead.
 # - The 80-net-hardening apt config adds retries + per-request timeout so that
 #   transient mirror flakes don't immediately fail a build (apt's default is 0
 #   retries).
@@ -146,21 +226,40 @@ ARG UBUNTU_MIRROR=
 USER root
 
 RUN if [ -n "$UBUNTU_MIRROR" ]; then \
-        sed -i "s|http://[^[:space:]/]*archive.ubuntu.com|$UBUNTU_MIRROR|g" /etc/apt/sources.list && \
-        sed -i "s|http://[^[:space:]/]*security.ubuntu.com|$UBUNTU_MIRROR|g" /etc/apt/sources.list; \
+        find /etc/apt -type f \( -name '*.list' -o -name '*.sources' \) \
+          -exec sed -i \
+            -e "s|http://[^[:space:]/]*archive.ubuntu.com|$UBUNTU_MIRROR|g" \
+            -e "s|http://[^[:space:]/]*security.ubuntu.com|$UBUNTU_MIRROR|g" \
+            {} + ; \
     fi && \
     printf 'Acquire::Retries "5";\nAcquire::http::Timeout "30";\nAcquire::https::Timeout "30";\n' \
         > /etc/apt/apt.conf.d/80-net-hardening
 
-# Fix hipDeviceGetName returning empty string in ROCm 7.0 docker images.
-# The ROCm 7.0 base image is missing libdrm-amdgpu-common which provides the
-# amdgpu.ids device-ID-to-marketing-name mapping file.
-# ROCm 7.2 base images already ship these packages, so this step is skipped.
+# Fix hipDeviceGetName returning empty / generic names.
+# amdgpu.ids maps PCI IDs to marketing names. The ROCm 7.0 base is missing it.
+# The 7.2.0 base was built with amdgpu-install and already has it. The 7.2.4
+# ubuntu24.04 base installs the `rocm` apt metapackage and does not; noble's
+# distro table has MI300 (74A*) but no MI355X (75A3), so gfx950-rocm724 would
+# otherwise report "AMD Radeon Graphics" and miss every name-keyed config.
 # See https://github.com/ROCm/ROCm/issues/5992
 RUN set -eux; \
     case "${GPU_ARCH}" in \
+      *rocm724*) \
+        echo "ROCm 7.2.4 (GPU_ARCH=${GPU_ARCH}): installing libdrm-amdgpu from graphics/7.2.4 noble"; \
+        curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key \
+          | gpg --dearmor -o /etc/apt/keyrings/amdgpu-graphics.gpg \
+        && echo 'deb [arch=amd64,i386 signed-by=/etc/apt/keyrings/amdgpu-graphics.gpg] https://repo.radeon.com/graphics/7.2.4/ubuntu noble main' \
+          > /etc/apt/sources.list.d/amdgpu-graphics.list \
+        && apt-get update \
+        && apt-get install -y --no-install-recommends \
+             libdrm-amdgpu-common \
+             libdrm-amdgpu-amdgpu1 \
+             libdrm2-amdgpu \
+        && rm -rf /var/lib/apt/lists/* \
+        && cp /opt/amdgpu/share/libdrm/amdgpu.ids /usr/share/libdrm/amdgpu.ids; \
+        ;; \
       *rocm720*) \
-        echo "ROCm 7.2 (GPU_ARCH=${GPU_ARCH}): libdrm-amdgpu packages already present, skipping"; \
+        echo "ROCm 7.2.0 (GPU_ARCH=${GPU_ARCH}): libdrm-amdgpu packages already present, skipping"; \
         ;; \
       *) \
         echo "ROCm 7.0 (GPU_ARCH=${GPU_ARCH}): installing libdrm-amdgpu packages"; \
@@ -187,7 +286,7 @@ RUN apt-get purge -y sccache; python -m pip uninstall -y sccache; rm -f "$(which
 # The ROCm 7.2 base image (rocm/pytorch) does not pre-install this package.
 RUN set -eux; \
     case "${GPU_ARCH}" in \
-      *rocm720*) \
+      *rocm720*|*rocm724*) \
         echo "ROCm 7.2 flavor detected from GPU_ARCH=${GPU_ARCH}"; \
         cd /opt/rocm/share/amd_smi \
         && python3 -m pip install --no-cache-dir . \
@@ -197,6 +296,33 @@ RUN set -eux; \
         ;; \
     esac
 
+# -----------------------
+# ROCm 7.2.4: upgrade torch 2.10 -> 2.11 (+ vision/audio), which pulls triton-rocm 3.6.0.
+# Done here, before AITER / sgl-kernel, so those extensions build against torch 2.11's ABI.
+RUN case "${GPU_ARCH}" in \
+      *-rocm724) \
+        python3 -m pip install --no-cache-dir --index-url "${TORCH_ROCM_INDEX_URL}" \
+            "torch==${TORCH_ROCM_VERSION}" \
+            "torchvision==${TORCHVISION_ROCM_VERSION}" \
+            "torchaudio==${TORCHAUDIO_ROCM_VERSION}" \
+        ;; \
+      *) \
+        echo "Not a ROCm 7.2.4 flavor (GPU_ARCH=${GPU_ARCH}), keep base torch/triton"; \
+        ;; \
+    esac
+
+# Populate the PIP_CONSTRAINT file, which only the rocm724 stages define, so that
+# resolving AITER and SGLang dependencies cannot replace the torch stack above.
+# Triton is left out: the BUILD_TRITON step installs it later.
+RUN case "${GPU_ARCH}" in \
+      *-rocm724) \
+        python3 -m pip freeze \
+          | grep -E '^(torch|torchvision|torchaudio)(==| @ )' \
+          > /etc/sglang/constraints/torch-rocm.txt \
+        && cat /etc/sglang/constraints/torch-rocm.txt \
+        ;; \
+    esac
+
 WORKDIR /sgl-workspace
 
 # -----------------------
@@ -220,7 +346,7 @@ RUN if [ "$BUILD_LLVM" = "1" ]; then \
 
 ENV SETUPTOOLS_SCM_PRETEND_VERSION=
 # Compile AITER against the base image's Triton; the Triton step at the end of
-# this file swaps in AITER's own pin afterwards.
+# this file installs the pinned one afterwards.
 ENV AITER_USE_SYSTEM_TRITON=1
 RUN pip uninstall -y aiter
 # Use `checkout -f` so the smudge-filter-induced "dirty" working tree from
@@ -248,6 +374,28 @@ RUN cd aiter \
         fi \
       && echo "export PYTHONPATH=/sgl-workspace/aiter:\${PYTHONPATH}" >> /etc/bash.bashrc
 
+# torch 2.11 Dynamo may pass a base torch.Stream; drop after ROCm/aiter#4817.
+RUN python3 <<'PY'
+from pathlib import Path
+p = Path("/sgl-workspace/aiter/csrc/cpp_itfs/torch_utils.py")
+s = p.read_text()
+old = """        elif isinstance(arg, torch.cuda.Stream):
+            c_args.append(ctypes.cast(arg.cuda_stream, ctypes.c_void_p))
+"""
+new = """        elif isinstance(arg, torch.Stream):
+            handle = getattr(arg, "cuda_stream", None)
+            if handle is None:
+                handle = torch.cuda.Stream(
+                    stream_id=arg.stream_id,
+                    device_index=arg.device_index,
+                    device_type=arg.device_type,
+                ).cuda_stream
+            c_args.append(ctypes.cast(handle, ctypes.c_void_p))
+"""
+if old in s:
+    p.write_text(s.replace(old, new))
+PY
+
 # -----------------------
 # Build Mooncake
 ENV PATH=$PATH:/usr/local/go/bin
@@ -319,14 +467,30 @@ RUN if [ "$BRANCH_TYPE" = "local" ]; then \
     && AMDGPU_TARGET=$GPU_ARCH_LIST python setup_rocm.py install \
     && cd ../../../.. \
     && rm -rf python/pyproject.toml && mv python/pyproject_other.toml python/pyproject.toml \
+    # srt_hip pins compressed-tensors==0.15.0, which requires torch<2.11 and so
+    # cannot be satisfied on the ROCm 7.2.4 torch 2.11 stack. The *_rocm724 extras
+    # carry a 0.16.0 pin instead; all other flavors keep the extras they used before.
+    && case "${GPU_ARCH}" in \
+         *-rocm724) srt_extras="srt_hip_rocm724,diffusion_hip"; all_extras="all_hip_rocm724" ;; \
+         *) srt_extras="srt_hip,diffusion_hip"; all_extras="all_hip" ;; \
+       esac \
     && if [ "$BUILD_TYPE" = "srt" ]; then \
-         export SETUPTOOLS_SCM_PRETEND_VERSION="${SETUPTOOLS_SCM_PRETEND_VERSION}" && python -m pip --no-cache-dir install -e "python[srt_hip,diffusion_hip]"; \
+         export SETUPTOOLS_SCM_PRETEND_VERSION="${SETUPTOOLS_SCM_PRETEND_VERSION}" && python -m pip --no-cache-dir install -e "python[${srt_extras}]"; \
        else \
-         export SETUPTOOLS_SCM_PRETEND_VERSION="${SETUPTOOLS_SCM_PRETEND_VERSION}" && python -m pip --no-cache-dir install -e "python[all_hip]"; \
+         export SETUPTOOLS_SCM_PRETEND_VERSION="${SETUPTOOLS_SCM_PRETEND_VERSION}" && python -m pip --no-cache-dir install -e "python[${all_extras}]"; \
        fi
 
 RUN python -m pip cache purge
 
+RUN if [ "${GPU_ARCH##*-}" = "rocm724" ]; then \
+      python3 -m pip check \
+      && python3 -c "import torch, torchaudio, torchvision, triton; expected={'torch':'2.11.','torchaudio':'2.11.','torchvision':'0.26.'}; actual={'torch':torch.__version__,'torchaudio':torchaudio.__version__,'torchvision':torchvision.__version__,'triton':triton.__version__}; assert torch.version.hip, actual; assert all(actual[name].startswith(version) for name, version in expected.items()), actual; print('Validated ROCm stack:', actual, 'HIP', torch.version.hip)" \
+      && if pip list --format=freeze | grep -Eq '^nvidia-.*-cu[0-9]+'; then \
+           echo "ERROR: NVIDIA CUDA runtime packages were installed into the ROCm image"; \
+           exit 1; \
+         fi; \
+    fi
+
 # Copy config files to support MI300X in virtualized environments (MI300X_VF).  Symlinks will not be created in image build.
 RUN find /sgl-workspace/sglang/python/sglang/srt/layers/quantization/configs/ \
          /sgl-workspace/sglang/python/sglang/srt/layers/moe/fused_moe_triton/configs/ \
@@ -536,6 +700,13 @@ RUN /bin/bash -lc 'set -euo pipefail; \
   apt-get update && apt-get install -y --no-install-recommends \
       build-essential autoconf automake libtool pkg-config git \
       libibverbs-dev librdmacm-dev rdma-core && rm -rf /var/lib/apt/lists/*; \
+  # Mooncake's dependencies.sh apt-installs Ubuntu's libabsl-dev (20220623 on
+  # the noble base used by rocm724). NIXL's meson then finds absl_base but no
+  # absl_log and refuses to fall back to its bundled Abseil -- "that would
+  # result in a mix of Abseil versions at runtime" -- so nixl fails at metadata
+  # generation. Drop just the -dev package (headers and pkg-config files); the
+  # runtime library that already-built components link against stays in place.
+  case "${GPU_ARCH}" in *-rocm724) apt-get remove -y libabsl-dev ;; esac; \
   pip install --no-cache-dir meson ninja pybind11 meson-python patchelf pyyaml; \
   git clone --depth=1 -b "${UCX_BRANCH}" "${UCX_REPO}" /sgl-workspace/ucx; \
   cd /sgl-workspace/ucx && ./autogen.sh && mkdir build && cd build && \
@@ -624,50 +795,30 @@ RUN cd /tmp/whl \
         ;; \
     esac
 
-# -----------------------
-# Hot patch: transformers dynamic_module_utils symlink bug (v5.12.1).
-# _compute_local_source_files_hash calls Path(...).resolve() on custom-code
-# module files, following the HF-cache snapshots/<hash>/x.py -> blobs/<blob>
-# symlink. trust_remote_code models whose custom code uses relative imports
-# (e.g. Kimi-K2.6's kimi_k25_vision_processing.py: `from .media_utils import`)
-# then crash with FileNotFoundError: .../blobs/<name>.py at processor init.
-# Mirrors upstream transformers PR #46618 (merged, not yet released): drop the
-# .resolve() on the module file and its relative-import sources so the snapshot
-# .py names (not the blob targets) are used. Self-skips once transformers ships
-# the fix; fails the build loudly if the pattern is present but unpatched.
-RUN python3 - <<'PY'
-import pathlib
-import transformers.dynamic_module_utils as m
-
-MARKS = ["Path(resolved_module_file).resolve()", "Path(source_file).resolve()"]
-path = pathlib.Path(m.__file__)
-src = path.read_text()
-if not any(mark in src for mark in MARKS):
-    print("transformers dynamic_module_utils already fixed; no patch needed")
-else:
-    patched = (
-        src.replace("Path(resolved_module_file).resolve()", "Path(resolved_module_file)")
-           .replace("Path(source_file).resolve()", "Path(source_file)")
-    )
-    assert patched != src, "FATAL: transformers symlink patch matched nothing"
-    path.write_text(patched)
-    print("patched transformers dynamic_module_utils.py (symlink hash fix)")
-PY
+# transformers 5.12.1: don't follow HF-cache symlinks when hashing custom modules
+# (transformers#46618, not yet released).
+RUN python3 -c "from pathlib import Path; import transformers.dynamic_module_utils as m; p=Path(m.__file__); t=p.read_text(); p.write_text(t.replace('Path(resolved_module_file).resolve()','Path(resolved_module_file)').replace('Path(source_file).resolve()','Path(source_file)'))"
 
 # -----------------------
-# Install the Triton AITER pins, replacing the base image's. No version check
-# on purpose: the pin is AITER's to move, and its installer enforces a floor.
+# Install AMD's ROCm Triton, replacing the base image's. The local version is
+# part of the pin: `==3.7.0` alone would accept any revision the index later
+# publishes under that number, and pip would choose between them by lexical
+# order of the git hash rather than by date.
 #
 # Keep this last. Base ROCm Torch pins triton==3.5.1 and the torch patch above
 # is what drops that pin, so installing Triton any earlier lets the next pip
 # install pull CUDA torch instead. The hip check below is the tripwire.
+# torch 2.11 names this `triton-rocm`; uninstall it so the pin is the only Triton.
 RUN if [ "$BUILD_TRITON" = "1" ]; then \
-        cd /sgl-workspace/aiter \
-     && test -f .github/scripts/install_triton.sh \
-     && PIP_NO_CACHE_DIR=1 bash .github/scripts/install_triton.sh \
+        pip uninstall -y triton-rocm || true \
+     && PIP_NO_CACHE_DIR=1 pip install --extra-index-url ${TRITON_INDEX_URL} \
+            "triton==${TRITON_VERSION}" "triton-kernels==${TRITON_KERNELS_VERSION}" \
      && python3 -c "import torch; from importlib.metadata import version; v = version('triton'); k = version('triton-kernels'); assert torch.version.hip is not None, torch.__version__; print(f'[Triton] ROCm Torch {torch.__version__}, Triton {v}, triton-kernels {k}')"; \
     fi
 
+# torch 2.11 still Requires-Dist: triton-rocm after the swap above.
+RUN case "${GPU_ARCH}" in *-rocm724) python3 -c "import pathlib,re,importlib.metadata as m; p=pathlib.Path(m.distribution('torch')._path)/'METADATA'; v=m.version('triton'); t,n=re.subn(r'^Requires-Dist: (?:triton|triton-rocm)==[^ ;]+', 'Requires-Dist: triton=='+v, p.read_text(), count=1, flags=re.M); assert n==1, n; p.write_text(t)" ;; esac
+
 # -----------------------
 # Performance environment variable.
 
diff --git a/python/pyproject_other.toml b/python/pyproject_other.toml
index abd14689af00..e4952cecdbd1 100755
--- a/python/pyproject_other.toml
+++ b/python/pyproject_other.toml
@@ -107,6 +107,10 @@ tracing = [
 
 # HIP (Heterogeneous-computing Interface for Portability) for AMD
 # => base docker rocm/vllm-dev:20250114, not from public vllm whl
+#
+# srt_hip_rocm724 below is the torch 2.11 variant of this list and differs only
+# in its compressed-tensors and torch pins. A dependency added here has to be
+# added there too; nothing enforces that.
 srt_hip = [
   # Pin to 0.15.0: 0.16.0 needs torch>=2.10 (incompatible with ROCm torch
   # 2.9.1). An open-ended `<0.16.0` made pip backtrack into an unbuildable
@@ -118,6 +122,17 @@ srt_hip = [
   "wave-lang==3.8.2",
 ]
 
+# Kept in step with srt_hip above by hand; only the two pins below differ.
+srt_hip_rocm724 = [
+  # ROCm 7.2.4 uses torch 2.11. compressed-tensors 0.15.0 requires
+  # torch<2.11, while 0.16.0 supports torch>=2.10.
+  "compressed-tensors==0.16.0",
+  "petit_kernel==0.0.2",
+  "sglang[runtime_common]",
+  "torch==2.11.0",
+  "wave-lang==3.8.2",
+]
+
 diffusion_hip = [
   "cache-dit==1.3.0",
   "peft>=0.18.0,<0.19.0", # Pin to <0.19.0 due to torchao incompatibility
@@ -189,11 +204,13 @@ test = [
 ]
 
 all_hip = ["sglang[diffusion_hip]", "sglang[srt_hip]", "sglang[tracing]"]
+all_hip_rocm724 = ["sglang[diffusion_hip]", "sglang[srt_hip_rocm724]", "sglang[tracing]"]
 all_hpu = ["sglang[srt_hpu]"]
 all_musa = ["sglang[diffusion_musa]", "sglang[srt_musa]"]
 all_mps = ["sglang[diffusion_mps]", "sglang[srt_mps]"]
 
 dev_hip = ["sglang[all_hip]", "sglang[test]"]
+dev_hip_rocm724 = ["sglang[all_hip_rocm724]", "sglang[test]"]
 dev_hpu = ["sglang[all_hpu]", "sglang[test]"]
 dev_musa = ["sglang[all_musa]", "sglang[test]"]
 dev_mps = ["sglang[all_mps]", "sglang[test]"]
diff --git a/scripts/ci/amd/amd_ci_install_dependency.sh b/scripts/ci/amd/amd_ci_install_dependency.sh
index 999b037aa225..b5ad5c231c81 100755
--- a/scripts/ci/amd/amd_ci_install_dependency.sh
+++ b/scripts/ci/amd/amd_ci_install_dependency.sh
@@ -30,7 +30,6 @@ EXTRAS="dev_hip,tracing"
 if [ -n "$OPTIONAL_DEPS" ]; then
     EXTRAS="dev_hip,tracing,${OPTIONAL_DEPS}"
 fi
-echo "Installing python extras: [${EXTRAS}]"
 
 # Host names look like: linux-mi35x-gpu-1-xxxxx-runner-zzzzz
 if [[ "${HOSTNAME_VALUE}" =~ ^linux-(mi[0-9]+[a-z]*)-gpu-[0-9]+ ]]; then
@@ -40,7 +39,56 @@ else
   echo "Warning: could not parse GPU architecture from '${HOSTNAME_VALUE}', defaulting to ${GPU_ARCH}"
 fi
 
+# Identify the Dockerfile stage that built this image. Both the python extras
+# below and the AITER pin lookup further down need the flavor, and they have to
+# agree, so detect it once here.
+#
+# Prefer GPU_ARCH stamped into the image (gfx950-rocm724, gfx942, ...).
+# Images built before that ENV existed: 724 stages already set
+# PIP_CONSTRAINT and HSA_ENABLE_IPC_MODE_LEGACY; remaining HIP 7.2* is
+# 720; else 7.0. Do not key off torch 2.11 — 720 may ship that later.
+IMAGE_TORCH_VERSION=$(docker exec ci_sglang python3 -c 'import torch; print(torch.__version__)')
+IMAGE_HIP_VERSION=$(docker exec ci_sglang python3 -c 'import torch; print(torch.version.hip or "")')
+IMAGE_GPU_ARCH=$(docker exec ci_sglang printenv GPU_ARCH 2>/dev/null || true)
+if [[ "${IMAGE_GPU_ARCH}" =~ ^(gfx942|gfx950)(-rocm720|-rocm724)?$ ]]; then
+    echo "[CI-IMAGE] Image GPU_ARCH=${IMAGE_GPU_ARCH}"
+    case "${IMAGE_GPU_ARCH}" in
+        *-rocm724) IMAGE_BASE_ARG_SUFFIX="_ROCM724"; IMAGE_STAGE_SUFFIX="-rocm724" ;;
+        *-rocm720) IMAGE_BASE_ARG_SUFFIX="_ROCM720"; IMAGE_STAGE_SUFFIX="-rocm720" ;;
+        *)         IMAGE_BASE_ARG_SUFFIX=""; IMAGE_STAGE_SUFFIX="" ;;
+    esac
+    IMAGE_GFX="${IMAGE_GPU_ARCH%-*}"
+else
+    IMAGE_PIP_CONSTRAINT=$(docker exec ci_sglang printenv PIP_CONSTRAINT 2>/dev/null || true)
+    IMAGE_HSA_LEGACY=$(docker exec ci_sglang printenv HSA_ENABLE_IPC_MODE_LEGACY 2>/dev/null || true)
+    if [[ -n "${IMAGE_PIP_CONSTRAINT}" || "${IMAGE_HSA_LEGACY}" == "1" ]]; then
+        IMAGE_BASE_ARG_SUFFIX="_ROCM724"
+        IMAGE_STAGE_SUFFIX="-rocm724"
+    elif [[ "${IMAGE_HIP_VERSION}" == 7.2* ]]; then
+        IMAGE_BASE_ARG_SUFFIX="_ROCM720"
+        IMAGE_STAGE_SUFFIX="-rocm720"
+    else
+        IMAGE_BASE_ARG_SUFFIX=""
+        IMAGE_STAGE_SUFFIX=""
+    fi
+    if [[ "${GPU_ARCH}" == "mi35x" ]]; then
+        IMAGE_GFX="gfx950"
+    else
+        IMAGE_GFX="gfx942"
+    fi
+    echo "[CI-IMAGE] Image has no GPU_ARCH stamp; inferred ${IMAGE_GFX}${IMAGE_STAGE_SUFFIX} (PIP_CONSTRAINT='${IMAGE_PIP_CONSTRAINT}', HSA_ENABLE_IPC_MODE_LEGACY='${IMAGE_HSA_LEGACY}', HIP=${IMAGE_HIP_VERSION})"
+    unset IMAGE_PIP_CONSTRAINT IMAGE_HSA_LEGACY
+fi
+unset IMAGE_GPU_ARCH
+
 # Install the required dependencies in CI.
+# ROCm 7.2.4 images ship torch 2.11, which srt_hip cannot satisfy (it pins
+# compressed-tensors 0.15.0, requiring torch<2.11). Select the rocm724 extras.
+if [[ "${IMAGE_STAGE_SUFFIX}" == "-rocm724" ]]; then
+  EXTRAS="${EXTRAS/dev_hip/dev_hip_rocm724}"
+fi
+echo "Image torch ${IMAGE_TORCH_VERSION}, HIP ${IMAGE_HIP_VERSION}; installing python extras: [${EXTRAS}]"
+
 # Fix permissions on pip cache, ignore errors from concurrent access or missing temp files
 docker exec ci_sglang chown -R root:root /sgl-data/pip-cache 2>/dev/null || true
 docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache --upgrade pip
@@ -221,6 +269,8 @@ if docker exec ci_sglang test -d /sgl-workspace/mori; then
     cd /sgl-workspace/mori
     git checkout '${MORI_COMMIT}'
     git submodule update --init --recursive
+    apt-get update
+    apt-get install -y --no-install-recommends libgrpc++-dev 2>/dev/null || true
     python3 setup.py develop
     python3 -c 'import os, torch; print(os.path.join(os.path.dirname(torch.__file__), \"lib\"))' > /etc/ld.so.conf.d/torch.conf
     ldconfig
@@ -248,35 +298,23 @@ DOCKERFILE="docker/rocm.Dockerfile"
 GPU_ARCH="${GPU_ARCH:-mi30x}"
 echo "[CI-AITER-CHECK] Runner GPU_ARCH=${GPU_ARCH}"
 
-# ROCm 7.0 keeps the Triton its base image ships; later ROCm images run on the
-# Triton AITER pins, so a rebuilt AITER has to bring its own along.
-IMAGE_HIP_VERSION=$(docker exec ci_sglang python3 -c 'import torch; print(torch.version.hip or "")')
-case "${IMAGE_HIP_VERSION}" in
-    7.0*) INSTALL_AITER_TRITON="false" ;;
-    7.*)  INSTALL_AITER_TRITON="true" ;;
-    *)
-        echo "[CI-AITER-CHECK] ERROR: Unsupported or empty HIP version: '${IMAGE_HIP_VERSION}'"
-        exit 1
-        ;;
-esac
-echo "[CI-AITER-CHECK] Container HIP=${IMAGE_HIP_VERSION}, install AITER's Triton on rebuild=${INSTALL_AITER_TRITON}"
+# Image owns Triton (pinned in docker/rocm.Dockerfile). Rebuild AITER against it.
 
 #############################################
-# 1. Extract AITER_COMMIT from correct Dockerfile block
+# 1. Extract AITER_COMMIT from the Dockerfile stage that built this image, as
+# identified near the top of this script.
 #############################################
-if [[ "${GPU_ARCH}" == "mi35x" ]]; then
-    echo "[CI-AITER-CHECK] Using gfx950 block from Dockerfile..."
-    REPO_AITER_COMMIT=$(grep -F -A20 'FROM $BASE_IMAGE_950 AS gfx950' docker/rocm.Dockerfile \
-                        | grep 'AITER_COMMIT_DEFAULT=' \
-                        | head -n1 \
-                        | sed 's/.*AITER_COMMIT_DEFAULT="\([^"]*\)".*/\1/')
+if [[ "${IMAGE_GFX}" == "gfx950" ]]; then
+    _from_line="FROM \$BASE_IMAGE_950${IMAGE_BASE_ARG_SUFFIX} AS gfx950${IMAGE_STAGE_SUFFIX}"
 else
-    echo "[CI-AITER-CHECK] Using gfx942 block from Dockerfile..."
-    REPO_AITER_COMMIT=$(grep -F -A20 'FROM $BASE_IMAGE_942 AS gfx942' docker/rocm.Dockerfile \
-                        | grep 'AITER_COMMIT_DEFAULT=' \
-                        | head -n1 \
-                        | sed 's/.*AITER_COMMIT_DEFAULT="\([^"]*\)".*/\1/')
+    _from_line="FROM \$BASE_IMAGE_942${IMAGE_BASE_ARG_SUFFIX} AS gfx942${IMAGE_STAGE_SUFFIX}"
 fi
+echo "[CI-AITER-CHECK] Using ${_from_line} from Dockerfile..."
+REPO_AITER_COMMIT=$(grep -F -A20 "${_from_line}" docker/rocm.Dockerfile \
+                    | grep 'AITER_COMMIT_DEFAULT=' \
+                    | head -n1 \
+                    | sed 's/.*AITER_COMMIT_DEFAULT="\([^"]*\)".*/\1/')
+unset _from_line
 
 
 if [[ -z "${REPO_AITER_COMMIT}" ]]; then
@@ -344,6 +382,30 @@ if [[ "${NEED_REBUILD}" == "true" ]]; then
         pip install -r requirements.txt
     "
 
+    # Re-apply the Dockerfile torch.Stream patch after re-clone (ROCm/aiter#4817).
+    if [[ "${IMAGE_STAGE_SUFFIX}" == "-rocm724" ]]; then
+        docker exec -i ci_sglang python3 - <<'PY'
+from pathlib import Path
+p = Path("/sgl-workspace/aiter/csrc/cpp_itfs/torch_utils.py")
+s = p.read_text()
+old = """        elif isinstance(arg, torch.cuda.Stream):
+            c_args.append(ctypes.cast(arg.cuda_stream, ctypes.c_void_p))
+"""
+new = """        elif isinstance(arg, torch.Stream):
+            handle = getattr(arg, "cuda_stream", None)
+            if handle is None:
+                handle = torch.cuda.Stream(
+                    stream_id=arg.stream_id,
+                    device_index=arg.device_index,
+                    device_type=arg.device_type,
+                ).cuda_stream
+            c_args.append(ctypes.cast(handle, ctypes.c_void_p))
+"""
+if old in s:
+    p.write_text(s.replace(old, new))
+PY
+    fi
+
     if [[ "${GPU_ARCH}" == "mi35x" ]]; then
         GPU_ARCH_LIST="gfx950"
     else
@@ -351,19 +413,7 @@ if [[ "${NEED_REBUILD}" == "true" ]]; then
     fi
     echo "[CI-AITER-CHECK] GPU_ARCH_LIST=${GPU_ARCH_LIST}"
 
-    # Run the installer here rather than letting setup.py do it: setup.py
-    # swallows its errors, and the AITER_USE_SYSTEM_TRITON=1 below then keeps
-    # whatever Triton is already installed. Doing it up front fails closed.
-    if [[ "${INSTALL_AITER_TRITON}" == "true" ]]; then
-        docker exec ci_sglang bash -c "
-            set -euo pipefail
-            cd /sgl-workspace/aiter
-            test -f .github/scripts/install_triton.sh
-            bash .github/scripts/install_triton.sh
-        "
-    fi
-
-    # build AITER
+    # The image already has the Dockerfile-pinned Triton; compile against it.
     docker exec ci_sglang bash -c "
         cd /sgl-workspace/aiter && \
         AITER_USE_SYSTEM_TRITON=1 GPU_ARCHS=${GPU_ARCH_LIST} python3 setup.py develop
diff --git a/scripts/ci/amd/amd_ci_start_container.sh b/scripts/ci/amd/amd_ci_start_container.sh
index 0715e325964a..3eee603da50d 100755
--- a/scripts/ci/amd/amd_ci_start_container.sh
+++ b/scripts/ci/amd/amd_ci_start_container.sh
@@ -28,7 +28,7 @@ LOCAL_DOCKER_REGISTRY="10.44.14.109:5000"
 # Parse command line arguments
 MI30X_BASE_TAG="${DEFAULT_MI30X_BASE_TAG}"
 MI35X_BASE_TAG="${DEFAULT_MI35X_BASE_TAG}"
-CUSTOM_IMAGE=""
+CUSTOM_IMAGE="${AMD_CI_IMAGE:-}"
 BUILD_FROM_DOCKERFILE=""
 GPU_ARCH_BUILD=""
 
@@ -50,7 +50,7 @@ while [[ $# -gt 0 ]]; do
       echo "Options:"
       echo "  --mi30x-base-tag TAG       Override MI30x base image tag"
       echo "  --mi35x-base-tag TAG       Override MI35x base image tag"
-      echo "  --custom-image IMAGE       Use a specific Docker image directly"
+      echo "  --custom-image IMAGE       Use a specific Docker image directly (or set AMD_CI_IMAGE)"
       echo "  --build-from-dockerfile    Build image from docker/rocm.Dockerfile"
       echo "  --gpu-arch ARCH            GPU architecture for Dockerfile build (e.g., gfx950-rocm720)"
       echo "  --rocm-version VERSION     Override ROCm version for image lookup (e.g., rocm720)"
diff --git a/scripts/ci/amd/amd_ci_start_container_disagg.sh b/scripts/ci/amd/amd_ci_start_container_disagg.sh
index 0c6e8a1dbdac..079015076fe5 100755
--- a/scripts/ci/amd/amd_ci_start_container_disagg.sh
+++ b/scripts/ci/amd/amd_ci_start_container_disagg.sh
@@ -28,11 +28,13 @@ LOCAL_DOCKER_REGISTRY="10.44.14.109:5000"
 # Parse command line arguments
 MI30X_BASE_TAG="${DEFAULT_MI30X_BASE_TAG}"
 MI35X_BASE_TAG="${DEFAULT_MI35X_BASE_TAG}"
+CUSTOM_IMAGE="${AMD_CI_IMAGE:-}"
 
 while [[ $# -gt 0 ]]; do
   case $1 in
     --mi30x-base-tag) MI30X_BASE_TAG="$2"; shift 2;;
     --mi35x-base-tag) MI35X_BASE_TAG="$2"; shift 2;;
+    --custom-image) CUSTOM_IMAGE="$2"; shift 2;;
     --rocm-version)
       ROCM_VERSION="$2"
       MI30X_BASE_TAG="${SGLANG_VERSION}-${ROCM_VERSION}-mi30x"
@@ -40,7 +42,7 @@ while [[ $# -gt 0 ]]; do
       echo "Using ROCm version override: ${ROCM_VERSION}"
       shift 2;;
     -h|--help)
-      echo "Usage: $0 [--mi30x-base-tag TAG] [--mi35x-base-tag TAG] [--rocm-version VERSION]"
+      echo "Usage: $0 [--mi30x-base-tag TAG] [--mi35x-base-tag TAG] [--custom-image IMAGE] [--rocm-version VERSION]"
       exit 0
       ;;
     *) echo "Unknown option $1"; exit 1;;
@@ -212,20 +214,30 @@ find_latest_image() {
   esac
 }
 
-# Pull and run the latest image
-IMAGE=$(find_latest_image "${GPU_ARCH}")
-# Try the local docker registry first (avoids Docker Hub rate limits and is
-# faster on the LAN); if that fails for any reason, fall back to the
-# public registry with exponential-backoff retries. Capture stderr so the
-# real failure reason (TLS handshake, 404, connection refused, etc.) is
-# visible in the job log instead of being silently swallowed.
-if local_pull_output=$(docker pull "${LOCAL_DOCKER_REGISTRY}/${IMAGE}" 2>&1); then
-  echo "Pulled from local docker registry: ${LOCAL_DOCKER_REGISTRY}/${IMAGE}"
-  docker tag "${LOCAL_DOCKER_REGISTRY}/${IMAGE}" "${IMAGE}"
+# Determine which image to use
+if [[ -n "${CUSTOM_IMAGE}" ]]; then
+  IMAGE="${CUSTOM_IMAGE}"
+  echo "Using custom image: ${IMAGE}"
+  if [[ "${IMAGE}" == "${LOCAL_DOCKER_REGISTRY}/"* ]]; then
+    docker pull "${IMAGE}"
+  else
+    retry_with_backoff 6 docker pull "${IMAGE}"
+  fi
 else
-  echo "Local docker registry pull failed; falling back to public registry: ${IMAGE}" >&2
-  printf '%s\n' "${local_pull_output}" | sed 's/^/  [local-pull] /' >&2
-  retry_with_backoff 6 docker pull "${IMAGE}"
+  IMAGE=$(find_latest_image "${GPU_ARCH}")
+  # Try the local docker registry first (avoids Docker Hub rate limits and is
+  # faster on the LAN); if that fails for any reason, fall back to the
+  # public registry with exponential-backoff retries. Capture stderr so the
+  # real failure reason (TLS handshake, 404, connection refused, etc.) is
+  # visible in the job log instead of being silently swallowed.
+  if local_pull_output=$(docker pull "${LOCAL_DOCKER_REGISTRY}/${IMAGE}" 2>&1); then
+    echo "Pulled from local docker registry: ${LOCAL_DOCKER_REGISTRY}/${IMAGE}"
+    docker tag "${LOCAL_DOCKER_REGISTRY}/${IMAGE}" "${IMAGE}"
+  else
+    echo "Local docker registry pull failed; falling back to public registry: ${IMAGE}" >&2
+    printf '%s\n' "${local_pull_output}" | sed 's/^/  [local-pull] /' >&2
+    retry_with_backoff 6 docker pull "${IMAGE}"
+  fi
 fi
 
 CACHE_HOST=/home/runner/sglang-data