From 52c191da52390fa5508de98eddd1e3eca2dbcfb2 Mon Sep 17 00:00:00 2001
From: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
Date: Fri, 11 Sep 2026 07:58:09 +0800
Subject: [PATCH] [Deps] Retire the CUDA 12 lane (#38404)

Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>
---
 .../workflows/_docker-build-and-publish.yml   | 113 ++++--------------
 .github/workflows/patch-docker-dev.yml        |   2 +-
 .github/workflows/release-docker-dev.yml      |  14 +--
 .github/workflows/release-docker-runtime.yml  |   5 +-
 .github/workflows/release-docker.yml          |   5 +-
 .github/workflows/release-pypi-nightly.yml    |  10 +-
 .github/workflows/release-whl-deepep.yml      | 110 ++---------------
 .github/workflows/release-whl-deepgemm.yml    |  87 --------------
 .github/workflows/release-whl-kernel.yml      | 105 ----------------
 .github/workflows/retag-docker.yml            |   2 +-
 .github/workflows/trivy-scan-dev.yml          |   2 +-
 docker/Dockerfile                             |  92 +++-----------
 docker/Dockerfile.cu134                       |  81 ++-----------
 docker/sgl-deep-ep.Dockerfile                 |  20 +---
 docker/sgl-deep-gemm.Dockerfile               |   3 +-
 .../autoregressive/MiniMax/MiniMax-M3.mdx     |   2 +-
 .../autoregressive/OpenBMB/MiniCPM-V-4_6.mdx  |   3 +-
 .../ThinkingMachines/Inkling-Small.mdx        |   4 +-
 .../ThinkingMachines/Inkling.mdx              |   4 +-
 .../developer_guide/contribution_guide.mdx    |  12 +-
 docs/docs/get-started/install.mdx             |  23 +---
 .../snippets/configs/MiniMaxAI/minimax-m3.jsx |   7 +-
 .../thinkingmachines/inkling-small.jsx        |   3 +-
 .../configs/thinkingmachines/inkling.jsx      |   3 +-
 python/sglang/kernels/aot/Dockerfile          |   3 +-
 .../aot/python/sgl_kernel/load_utils.py       |   7 +-
 python/sglang/kernels/aot/rename_wheels.sh    |   8 +-
 python/sglang/kernels/jit/utils/arch.py       |   2 +-
 scripts/build_sgl_deep_gemm.sh                |   5 +-
 scripts/build_sgl_deepep.sh                   |   9 +-
 scripts/ci/cuda/ci_install_dependency.sh      | 110 +++++------------
 .../ci/utils/docker_build_metadata_args.py    |   2 +-
 scripts/rename_sgl_deep_gemm_whl.sh           |   2 +-
 scripts/update_deepep_whl_index.py            |   2 +-
 scripts/update_deepgemm_whl_index.py          |   4 +-
 scripts/update_kernel_whl_index.py            |  23 ++--
 scripts/update_nightly_whl_index.py           |   4 +-
 .../tools/test_docker_build_metadata_args.py  |  14 +--
 38 files changed, 170 insertions(+), 737 deletions(-)

diff --git a/.github/workflows/_docker-build-and-publish.yml b/.github/workflows/_docker-build-and-publish.yml
index 6a84a7b19..22f21ef70 100644
--- a/.github/workflows/_docker-build-and-publish.yml
+++ b/.github/workflows/_docker-build-and-publish.yml
@@ -1,7 +1,7 @@
 name: Build and Publish Multi-Arch Docker Images
 
-# Reusable workflow: builds CUDA 12 + CUDA 13 images for amd64 and arm64,
-# then creates multi-arch manifests with caller-specified tags.
+# Reusable workflow: builds CUDA 13 images for amd64 and arm64, then creates
+# multi-arch manifests with caller-specified tags.
 
 on:
   workflow_call:
@@ -26,7 +26,7 @@ on:
         type: string
         default: ""
       tag_config:
-        description: 'JSON array of {"cuda":"cu129|cu130","tags":["tag1","tag2"]}. Tags support {version} substitution.'
+        description: 'JSON array of {"cuda":"cu130","tags":["tag1","tag2"]}. Tags support {version} substitution.'
         required: true
         type: string
       use_environment:
@@ -50,7 +50,6 @@ jobs:
       SGL_VERSION: ${{ inputs.sgl_version }}
       IMAGE_REPO: ${{ inputs.image_repo }}
     outputs:
-      digest-cu129: ${{ steps.build-cu129.outputs.digest }}
       digest-cu130: ${{ steps.build-cu130.outputs.digest }}
     steps:
       - name: Delete huge unnecessary tools folder
@@ -69,16 +68,14 @@ jobs:
           set -euo pipefail
           BUILD_COMMIT="$(git rev-parse HEAD)"
           BUILD_URL="${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}/actions/runs/${GITHUB_RUN_ID}"
-          for CUDA_VARIANT in cu129 cu130; do
-            python3 scripts/ci/utils/docker_build_metadata_args.py \
-              --cuda "${CUDA_VARIANT}" \
-              --tag-config "${TAG_CONFIG}" \
-              --image-repo "${IMAGE_REPO}" \
-              --sgl-version "${SGL_VERSION}" \
-              --build-commit "${BUILD_COMMIT}" \
-              --build-url "${BUILD_URL}" \
-              > "/tmp/docker-metadata-${CUDA_VARIANT}.args"
-          done
+          python3 scripts/ci/utils/docker_build_metadata_args.py \
+            --cuda cu130 \
+            --tag-config "${TAG_CONFIG}" \
+            --image-repo "${IMAGE_REPO}" \
+            --sgl-version "${SGL_VERSION}" \
+            --build-commit "${BUILD_COMMIT}" \
+            --build-url "${BUILD_URL}" \
+            > /tmp/docker-metadata-cu130.args
 
       - name: Free disk space
         uses: jlumbroso/free-disk-space@main
@@ -106,34 +103,6 @@ jobs:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
 
-      - name: Build and push AMD64 image (CUDA 12)
-        id: build-cu129
-        run: |
-          VERSION_ARG=""
-          if [ -n "${SGL_VERSION}" ]; then
-            VERSION_ARG="--build-arg SGL_VERSION=${SGL_VERSION}"
-          fi
-          mapfile -t METADATA_ARGS < /tmp/docker-metadata-cu129.args
-
-          docker buildx build \
-            --target ${{ inputs.docker_target }} \
-            --platform linux/amd64 \
-            --output type=image,name=${{ inputs.image_repo }},push-by-digest=true,name-canonical=true,push=true \
-            -f docker/Dockerfile \
-            --build-arg CUDA_VERSION=12.9.2 \
-            --build-arg BUILD_TYPE=all \
-            --build-arg INSTALL_FLASHINFER_JIT_CACHE=1 \
-            "${METADATA_ARGS[@]}" \
-            ${VERSION_ARG} \
-            ${{ inputs.extra_build_args }} \
-            --metadata-file /tmp/metadata-cu129.json \
-            --no-cache \
-            .
-
-          DIGEST=$(python3 -c "import json; print(json.load(open('/tmp/metadata-cu129.json'))['containerimage.digest'])")
-          echo "Pushed digest: ${DIGEST}"
-          echo "digest=${DIGEST}" >> $GITHUB_OUTPUT
-
       - name: Build and push AMD64 image (CUDA 13)
         id: build-cu130
         run: |
@@ -171,7 +140,6 @@ jobs:
       SGL_VERSION: ${{ inputs.sgl_version }}
       IMAGE_REPO: ${{ inputs.image_repo }}
     outputs:
-      digest-cu129: ${{ steps.build-cu129.outputs.digest }}
       digest-cu130: ${{ steps.build-cu130.outputs.digest }}
     steps:
       - name: Delete huge unnecessary tools folder
@@ -190,16 +158,14 @@ jobs:
           set -euo pipefail
           BUILD_COMMIT="$(git rev-parse HEAD)"
           BUILD_URL="${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}/actions/runs/${GITHUB_RUN_ID}"
-          for CUDA_VARIANT in cu129 cu130; do
-            python3 scripts/ci/utils/docker_build_metadata_args.py \
-              --cuda "${CUDA_VARIANT}" \
-              --tag-config "${TAG_CONFIG}" \
-              --image-repo "${IMAGE_REPO}" \
-              --sgl-version "${SGL_VERSION}" \
-              --build-commit "${BUILD_COMMIT}" \
-              --build-url "${BUILD_URL}" \
-              > "/tmp/docker-metadata-${CUDA_VARIANT}.args"
-          done
+          python3 scripts/ci/utils/docker_build_metadata_args.py \
+            --cuda cu130 \
+            --tag-config "${TAG_CONFIG}" \
+            --image-repo "${IMAGE_REPO}" \
+            --sgl-version "${SGL_VERSION}" \
+            --build-commit "${BUILD_COMMIT}" \
+            --build-url "${BUILD_URL}" \
+            > /tmp/docker-metadata-cu130.args
 
       - name: Prune Docker to reclaim disk space
         run: |
@@ -216,34 +182,6 @@ jobs:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
 
-      - name: Build and push ARM64 image (CUDA 12)
-        id: build-cu129
-        run: |
-          VERSION_ARG=""
-          if [ -n "${SGL_VERSION}" ]; then
-            VERSION_ARG="--build-arg SGL_VERSION=${SGL_VERSION}"
-          fi
-          mapfile -t METADATA_ARGS < /tmp/docker-metadata-cu129.args
-
-          docker buildx build \
-            --target ${{ inputs.docker_target }} \
-            --platform linux/arm64 \
-            --output type=image,name=${{ inputs.image_repo }},push-by-digest=true,name-canonical=true,push=true \
-            -f docker/Dockerfile \
-            --build-arg CUDA_VERSION=12.9.2 \
-            --build-arg BUILD_TYPE=all \
-            --build-arg INSTALL_FLASHINFER_JIT_CACHE=1 \
-            "${METADATA_ARGS[@]}" \
-            ${VERSION_ARG} \
-            ${{ inputs.extra_build_args }} \
-            --metadata-file /tmp/metadata-cu129.json \
-            --no-cache \
-            .
-
-          DIGEST=$(python3 -c "import json; print(json.load(open('/tmp/metadata-cu129.json'))['containerimage.digest'])")
-          echo "Pushed digest: ${DIGEST}"
-          echo "digest=${DIGEST}" >> $GITHUB_OUTPUT
-
       - name: Build and push ARM64 image (CUDA 13)
         id: build-cu130
         run: |
@@ -292,22 +230,13 @@ jobs:
           TAG_CONFIG: ${{ inputs.tag_config }}
           SGL_VERSION: ${{ inputs.sgl_version }}
           IMAGE_REPO: ${{ inputs.image_repo }}
-          X86_CU129: ${{ needs.build-x86.outputs.digest-cu129 }}
           X86_CU130: ${{ needs.build-x86.outputs.digest-cu130 }}
-          ARM64_CU129: ${{ needs.build-arm64.outputs.digest-cu129 }}
           ARM64_CU130: ${{ needs.build-arm64.outputs.digest-cu130 }}
           SHORT_SHA: ${{ github.sha }}
         run: |
           echo "${TAG_CONFIG}" | jq -c '.[]' | while read -r entry; do
-            CUDA=$(echo "${entry}" | jq -r '.cuda')
-
-            if [ "${CUDA}" = "cu129" ]; then
-              X86_DIGEST="${X86_CU129}"
-              ARM64_DIGEST="${ARM64_CU129}"
-            else
-              X86_DIGEST="${X86_CU130}"
-              ARM64_DIGEST="${ARM64_CU130}"
-            fi
+            X86_DIGEST="${X86_CU130}"
+            ARM64_DIGEST="${ARM64_CU130}"
 
             TAG_ARGS=""
             for tag in $(echo "${entry}" | jq -r '.tags[]'); do
diff --git a/.github/workflows/patch-docker-dev.yml b/.github/workflows/patch-docker-dev.yml
index d46d21e17..7ecd5b674 100644
--- a/.github/workflows/patch-docker-dev.yml
+++ b/.github/workflows/patch-docker-dev.yml
@@ -8,7 +8,7 @@ on:
         required: false
         default: ""
       image_tag:
-        description: "Base image tag to patch (e.g. dev, dev-cu13, dev-cu12)"
+        description: "Base image tag to patch (e.g. dev, dev-cu13)"
         required: true
       output_tag:
         description: "Tag to publish as. Overwrites it if it already exists."
diff --git a/.github/workflows/release-docker-dev.yml b/.github/workflows/release-docker-dev.yml
index d212f0cef..bff64e962 100644
--- a/.github/workflows/release-docker-dev.yml
+++ b/.github/workflows/release-docker-dev.yml
@@ -25,7 +25,7 @@ on:
         required: false
         default: ""
       overlay_cudas:
-        description: "Comma-separated cuda variants to overlay onto (cu12, cu13). Default 'cu13'."
+        description: "Comma-separated cuda variants to overlay onto (cu13). Default 'cu13'."
         required: false
         default: "cu13"
       overlay_tag_suffix:
@@ -86,9 +86,9 @@ jobs:
           # consumers pinned to the pre-flip names.
           if [ -z "${SUFFIX}" ]; then
             # Nightly: include dated tags
-            TAG_CONFIG='[{"cuda":"cu129","tags":["dev-cu12","nightly-dev-cu12-{date}-{short_sha}"]},{"cuda":"cu130","tags":["dev","dev-cu13","nightly-dev-{date}-{short_sha}","nightly-dev-cu13-{date}-{short_sha}"]}]'
+            TAG_CONFIG='[{"cuda":"cu130","tags":["dev","dev-cu13","nightly-dev-{date}-{short_sha}","nightly-dev-cu13-{date}-{short_sha}"]}]'
           else
-            TAG_CONFIG="[{\"cuda\":\"cu129\",\"tags\":[\"dev-cu12${SUFFIX}\"]},{\"cuda\":\"cu130\",\"tags\":[\"dev${SUFFIX}\",\"dev-cu13${SUFFIX}\"]}]"
+            TAG_CONFIG="[{\"cuda\":\"cu130\",\"tags\":[\"dev${SUFFIX}\",\"dev-cu13${SUFFIX}\"]}]"
           fi
           echo "tag_config=${TAG_CONFIG}" >> $GITHUB_OUTPUT
 
@@ -97,7 +97,7 @@ jobs:
           # SUFFIX was computed above from inputs.tag / inputs.pr_number.
           if [ -n "${OVERLAY_DOCKERFILE_INPUT}" ]; then
             # Refuse to overlay without a tag/pr_number: SUFFIX would be empty,
-            # so base would point at the moving 'dev'/'dev-cu12'/'dev-cu13'
+            # so base would point at the moving 'dev'/'dev-cu13'
             # latest tags and overwrite mode would clobber them.
             if [ -z "${SUFFIX}" ]; then
               echo "overlay_dockerfile is set but neither 'tag' nor 'pr_number' was given." >&2
@@ -112,10 +112,8 @@ jobs:
               C="$(echo "${C}" | xargs)"
               if [ "${C}" = "cu13" ]; then
                 BASES='["dev'"${SUFFIX}"'","dev-cu13'"${SUFFIX}"'"]'
-              elif [ "${C}" = "cu12" ]; then
-                BASES='["dev-cu12'"${SUFFIX}"'"]'
               else
-                echo "Unknown overlay cuda variant: ${C} (expected cu12 or cu13)" >&2
+                echo "Unknown overlay cuda variant: ${C} (expected cu13)" >&2
                 exit 1
               fi
               if [ -n "${OVERLAY_SUFFIX_INPUT}" ]; then
@@ -213,7 +211,7 @@ jobs:
     if: ${{ !inputs.build_only && !inputs.tag && !inputs.pr_number }}
     uses: ./.github/workflows/_docker-cleanup-nightly.yml
     with:
-      tag_prefixes: '["nightly-dev", "nightly-dev-cu12", "nightly-dev-cu13"]'
+      tag_prefixes: '["nightly-dev", "nightly-dev-cu13"]'
       image_repo: ${{ inputs.image_repo || 'lmsysorg/sglang' }}
     secrets: inherit
 
diff --git a/.github/workflows/release-docker-runtime.yml b/.github/workflows/release-docker-runtime.yml
index 0e224bf91..eee7b1479 100644
--- a/.github/workflows/release-docker-runtime.yml
+++ b/.github/workflows/release-docker-runtime.yml
@@ -2,7 +2,7 @@ name: Release Docker Runtime Images
 #
 # Builds and publishes runtime Docker images (production-optimized, ~50% smaller):
 #   - lmsysorg/sglang:v{version}-runtime, lmsysorg/sglang:latest-runtime
-#   - lmsysorg/sglang:v{version}-cu129-runtime, lmsysorg/sglang:latest-cu129-runtime
+#   - lmsysorg/sglang:v{version}-cu130-runtime, lmsysorg/sglang:latest-cu130-runtime
 #
 on:
   push:
@@ -49,7 +49,6 @@ jobs:
       image_repo: ${{ inputs.image_repo || 'lmsysorg/sglang' }}
       tag_config: |
         [
-          {"cuda": "cu130", "tags": ["v${{ needs.resolve-version.outputs.version }}-runtime", "latest-runtime", "v${{ needs.resolve-version.outputs.version }}-cu130-runtime", "latest-cu130-runtime"]},
-          {"cuda": "cu129", "tags": ["v${{ needs.resolve-version.outputs.version }}-cu129-runtime", "latest-cu129-runtime"]}
+          {"cuda": "cu130", "tags": ["v${{ needs.resolve-version.outputs.version }}-runtime", "latest-runtime", "v${{ needs.resolve-version.outputs.version }}-cu130-runtime", "latest-cu130-runtime"]}
         ]
     secrets: inherit
diff --git a/.github/workflows/release-docker.yml b/.github/workflows/release-docker.yml
index edf21469e..fcc36a631 100644
--- a/.github/workflows/release-docker.yml
+++ b/.github/workflows/release-docker.yml
@@ -2,7 +2,7 @@ name: Release Docker Images
 #
 # Builds and publishes framework Docker images (full development environment):
 #   - lmsysorg/sglang:v{version}, lmsysorg/sglang:latest (cuda 13)
-#   - lmsysorg/sglang:v{version}-cu129, lmsysorg/sglang:latest-cu129
+#   - lmsysorg/sglang:v{version}-cu130, lmsysorg/sglang:latest-cu130
 #
 on:
   push:
@@ -49,7 +49,6 @@ jobs:
       image_repo: ${{ inputs.image_repo || 'lmsysorg/sglang' }}
       tag_config: |
         [
-          {"cuda": "cu130", "tags": ["v${{ needs.resolve-version.outputs.version }}", "latest", "v${{ needs.resolve-version.outputs.version }}-cu130", "latest-cu130"]},
-          {"cuda": "cu129", "tags": ["v${{ needs.resolve-version.outputs.version }}-cu129", "latest-cu129"]}
+          {"cuda": "cu130", "tags": ["v${{ needs.resolve-version.outputs.version }}", "latest", "v${{ needs.resolve-version.outputs.version }}-cu130", "latest-cu130"]}
         ]
     secrets: inherit
diff --git a/.github/workflows/release-pypi-nightly.yml b/.github/workflows/release-pypi-nightly.yml
index b072a8316..8a069e5de 100644
--- a/.github/workflows/release-pypi-nightly.yml
+++ b/.github/workflows/release-pypi-nightly.yml
@@ -113,13 +113,13 @@ jobs:
     environment: 'prod'
     strategy:
       fail-fast: false
-      # The wheel is CUDA-agnostic and built once — we just register the same
-      # artifact under cu129/sglang/ and cu130/sglang/ wheel indexes so users
-      # can install via either --extra-index-url. Serialize because both matrix
-      # runs clone and push to the same sgl-whl branch.
+      # The wheel is CUDA-agnostic and built once — we just register it under
+      # the cu130/sglang/ wheel index so users can install via
+      # --extra-index-url. Serialize because matrix runs clone and push to the
+      # same sgl-whl branch.
       max-parallel: 1
       matrix:
-        cuda_version: ['129', '130']
+        cuda_version: ['130']
     steps:
       - uses: actions/checkout@v4
 
diff --git a/.github/workflows/release-whl-deepep.yml b/.github/workflows/release-whl-deepep.yml
index 2a23e36d7..93b71f8ac 100644
--- a/.github/workflows/release-whl-deepep.yml
+++ b/.github/workflows/release-whl-deepep.yml
@@ -7,15 +7,6 @@ on:
         description: "Public wheel version (for example, 0.1.0 or 0.1.1rc0)"
         type: string
         required: true
-      target:
-        description: "CUDA target"
-        type: choice
-        required: false
-        default: "all"
-        options:
-          - "all"
-          - "cu129"
-          - "cu130"
       packaging-ref:
         description: "DeepEP ref containing the shared sgl_deep_ep overlay"
         type: string
@@ -30,75 +21,8 @@ env:
   TORCH_VERSION: "2.13.0"
 
 jobs:
-  build-cu129-matrix:
-    if: |
-      github.repository == 'sgl-project/sglang' &&
-      (inputs.target == 'all' || inputs.target == 'cu129')
-    strategy:
-      fail-fast: false
-      matrix:
-        python-version: ["3.10", "3.12"]
-        arch: [x86_64, aarch64]
-        include:
-          - arch: x86_64
-            runner: x64-kernel-build-node
-            source-branch: sgl-deepep-cu12-x86
-          - arch: aarch64
-            runner: arm-kernel-build-node
-            source-branch: sgl-deepep-cu12-arm
-    runs-on: ${{ matrix.runner }}
-    steps:
-      - name: Clean workspace
-        run: |
-          docker run --rm -v "${{ github.workspace }}:/workspace" alpine:3 \
-            sh -c 'rm -rf /workspace/..?* /workspace/.[!.]* /workspace/*' || true
-
-      - name: Checkout SGLang
-        uses: actions/checkout@v4
-
-      - name: Checkout DeepEP implementation
-        uses: actions/checkout@v4
-        with:
-          repository: sgl-project/DeepEP
-          ref: ${{ matrix.source-branch }}
-          path: DeepEP-source
-          submodules: recursive
-
-      - name: Checkout DeepEP packaging overlay
-        uses: actions/checkout@v4
-        with:
-          repository: sgl-project/DeepEP
-          ref: ${{ inputs['packaging-ref'] || 'sgl-deepep-packaging' }}
-          path: DeepEP-packaging
-          sparse-checkout: sgl_deep_ep
-
-      - name: Set wheel version
-        env:
-          RELEASE_VERSION: ${{ inputs.version }}
-        run: |
-          printf '%s' "${RELEASE_VERSION}" > DeepEP-packaging/sgl_deep_ep/VERSION
-
-      - name: Build wheel
-        run: |
-          chmod +x scripts/build_sgl_deepep.sh DeepEP-packaging/sgl_deep_ep/build_sgl_deep_ep.sh
-          scripts/build_sgl_deepep.sh \
-            "${{ matrix.python-version }}" \
-            "12.9" \
-            "${{ github.workspace }}/DeepEP-source" \
-            "${{ github.workspace }}/DeepEP-packaging/sgl_deep_ep" \
-            "${{ matrix.arch }}"
-
-      - name: Upload wheel
-        uses: actions/upload-artifact@v4
-        with:
-          name: deepep-wheel-cu129-py${{ matrix.python-version }}-${{ matrix.arch }}
-          path: DeepEP-source/dist/*.whl
-          if-no-files-found: error
-
   build-cu130-matrix:
-    if: |
-      github.repository == 'sgl-project/sglang' &&
-      (inputs.target == 'all' || inputs.target == 'cu130')
+    if: github.repository == 'sgl-project/sglang'
     strategy:
       fail-fast: false
       matrix:
@@ -168,19 +92,10 @@ jobs:
           if-no-files-found: error
 
   release-whl:
-    needs: [build-cu129-matrix, build-cu130-matrix]
+    needs: build-cu130-matrix
     if: |
-      always() &&
       github.repository == 'sgl-project/sglang' &&
-      (
-        (inputs.target == 'cu129' && needs.build-cu129-matrix.result == 'success') ||
-        (inputs.target == 'cu130' && needs.build-cu130-matrix.result == 'success') ||
-        (
-          inputs.target == 'all' &&
-          needs.build-cu129-matrix.result == 'success' &&
-          needs.build-cu130-matrix.result == 'success'
-        )
-      )
+      needs.build-cu130-matrix.result == 'success'
     runs-on: ubuntu-latest
     steps:
       - name: Checkout SGLang
@@ -209,14 +124,8 @@ jobs:
           git clone --branch gh-pages \
             "https://oauth2:${WHL_TOKEN}@github.com/sgl-project/whl.git" sgl-whl
 
-      - name: Update wheel indexes
-        run: |
-          if [[ "${{ inputs.target }}" == "all" || "${{ inputs.target }}" == "cu129" ]]; then
-            python3 scripts/update_deepep_whl_index.py --cuda 129
-          fi
-          if [[ "${{ inputs.target }}" == "all" || "${{ inputs.target }}" == "cu130" ]]; then
-            python3 scripts/update_deepep_whl_index.py --cuda 130
-          fi
+      - name: Update wheel index
+        run: python3 scripts/update_deepep_whl_index.py --cuda 130
 
       - name: Push wheel indexes
         env:
@@ -234,15 +143,10 @@ jobs:
           fi
 
   release-pypi:
-    needs: [build-cu129-matrix, build-cu130-matrix]
+    needs: build-cu130-matrix
     if: |
-      always() &&
       github.repository == 'sgl-project/sglang' &&
-      needs.build-cu130-matrix.result == 'success' &&
-      (
-        inputs.target == 'cu130' ||
-        (inputs.target == 'all' && needs.build-cu129-matrix.result == 'success')
-      )
+      needs.build-cu130-matrix.result == 'success'
     runs-on: ubuntu-latest
     steps:
       - name: Download PyPI wheels
diff --git a/.github/workflows/release-whl-deepgemm.yml b/.github/workflows/release-whl-deepgemm.yml
index b321bc946..ed24da92d 100644
--- a/.github/workflows/release-whl-deepgemm.yml
+++ b/.github/workflows/release-whl-deepgemm.yml
@@ -14,7 +14,6 @@ on:
         default: 'all'
         options:
           - 'all'
-          - 'cu129'
           - 'cu130'
       branch:
         description: "DeepGEMM branch to build from (default: dev)"
@@ -33,92 +32,6 @@ env:
   TORCH_VER: "2.13.0"
 
 jobs:
-  build-cu129-matrix:
-    if: |
-      github.repository == 'sgl-project/sglang' &&
-      (github.event.inputs.target == 'all' || github.event.inputs.target == 'cu129')
-    strategy:
-      matrix:
-        python-version: ["3.12"]
-        cuda-version: ["12.9"]
-        arch: [x86_64, aarch64]
-        include:
-          - arch: x86_64
-            runner: x64-kernel-build-node
-          - arch: aarch64
-            runner: arm-kernel-build-node
-    runs-on: ${{ matrix.runner }}
-    steps:
-      - name: Clean workspace (remove root-owned files from prior runs)
-        run: |
-          docker run --rm -v "${{ github.workspace }}:/workspace" alpine:3 \
-            sh -c 'rm -rf /workspace/..?* /workspace/.[!.]* /workspace/*' || true
-
-      - uses: actions/checkout@v4
-
-      - name: Checkout DeepGEMM
-        uses: actions/checkout@v4
-        with:
-          repository: sgl-project/DeepGEMM
-          ref: ${{ inputs.branch || 'dev' }}
-          path: DeepGEMM
-          submodules: recursive
-
-      - name: Set wheel version
-        run: |
-          echo -n "${{ inputs.version }}" > DeepGEMM/sgl_deep_gemm/VERSION
-          cat DeepGEMM/sgl_deep_gemm/VERSION
-
-      - name: Build wheel
-        run: |
-          chmod +x ./scripts/build_sgl_deep_gemm.sh ./scripts/rename_sgl_deep_gemm_whl.sh
-          ./scripts/build_sgl_deep_gemm.sh "${{ matrix.python-version }}" "${{ matrix.cuda-version }}" "${{ github.workspace }}/DeepGEMM" "${{ matrix.arch }}"
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          name: deepgemm-wheel-cuda${{ matrix.cuda-version }}-${{ matrix.arch }}
-          path: DeepGEMM/dist/*.whl
-
-  release-cu129:
-    needs: build-cu129-matrix
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v4
-
-      - name: Download artifacts
-        uses: actions/download-artifact@v4
-        with:
-          path: dist/
-          merge-multiple: true
-          pattern: deepgemm-wheel-cuda12.9-*
-
-      - name: Release
-        uses: softprops/action-gh-release@v2
-        with:
-          tag_name: v${{ inputs.version }}
-          repository: sgl-project/whl
-          token: ${{ secrets.GH_PAT_FOR_WHL_RELEASE }}
-          files: |
-            dist/*
-
-      - name: Clone wheel index
-        run: git clone https://oauth2:${WHL_TOKEN}@github.com/sgl-project/whl.git sgl-whl
-        env:
-          WHL_TOKEN: ${{ secrets.GH_PAT_FOR_WHL_RELEASE }}
-
-      - name: Update wheel index
-        run: python3 scripts/update_deepgemm_whl_index.py --cuda 129
-
-      - name: Push wheel index
-        run: |
-          cd sgl-whl
-          git config --local user.name "sglang-bot"
-          git config --local user.email "sglangbot@gmail.com"
-          git add -A
-          git commit -m "update sgl-deep-gemm whl index for v${{ inputs.version }}"
-          git push
-
   build-cu130-matrix:
     if: |
       github.repository == 'sgl-project/sglang' &&
diff --git a/.github/workflows/release-whl-kernel.yml b/.github/workflows/release-whl-kernel.yml
index f6e141741..560286a84 100644
--- a/.github/workflows/release-whl-kernel.yml
+++ b/.github/workflows/release-whl-kernel.yml
@@ -15,7 +15,6 @@ on:
         default: 'all'
         options:
           - 'all'
-          - 'cu129'
           - 'cu130'
           - 'rocm700'
           - 'rocm720'
@@ -35,110 +34,6 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
-  # cu130 is the PyPI-released variant; cu129 wheels are published only to the
-  # sgl-project/whl index (consumed via `pip install ...+cu129` for the legacy
-  # cuda 12.9 path), not to PyPI.
-  build-cu129-matrix:
-    if: |
-      github.repository == 'sgl-project/sglang' &&
-      (github.event_name == 'push' || github.event.inputs.target == 'all' || github.event.inputs.target == 'cu129')
-    strategy:
-      matrix:
-        python-version: ["3.10"]
-        cuda-version: ["12.9"]
-        arch: [x86_64, aarch64]
-        include:
-          - arch: x86_64
-            runner: x64-kernel-build-node
-          - arch: aarch64
-            runner: arm-kernel-build-node
-    runs-on: ${{ matrix.runner }}
-    steps:
-      # Self-hosted build nodes retain the workspace across jobs. Prior builds
-      # leave root-owned artifacts under python/sglang/kernels/aot/build/ that actions/checkout
-      # cannot remove, causing EACCES on rmdir. Wipe them via a throwaway root
-      # container before checkout recreates the workspace.
-      - name: Clean workspace (remove root-owned files from prior runs)
-        run: |
-          docker run --rm -v "${{ github.workspace }}:/workspace" alpine:3 \
-            sh -c 'rm -rf /workspace/..?* /workspace/.[!.]* /workspace/*' || true
-
-      - uses: actions/checkout@v4
-        with:
-          submodules: "recursive"
-          ref: ${{ inputs.pr_number && format('refs/pull/{0}/head', inputs.pr_number) || '' }}
-
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python-version }}
-
-      - name: Build wheels
-        run: |
-          cd python/sglang/kernels/aot
-          chmod +x ./build.sh
-          ./build.sh "${{ matrix.python-version }}" "${{ matrix.cuda-version }}" ${{ matrix.arch == 'aarch64' && 'aarch64' || '' }}
-        env:
-          BUILD_JOBS: 64
-          NVCC_THREADS: 8
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          name: wheel-python${{ matrix.python-version }}-cuda${{ matrix.cuda-version }}${{ matrix.arch == 'aarch64' && '-aarch64' || '' }}
-          path: python/sglang/kernels/aot/dist/*
-
-  release-cu129:
-    needs: build-cu129-matrix
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          ref: ${{ inputs.pr_number && format('refs/pull/{0}/head', inputs.pr_number) || '' }}
-
-      - name: Download artifacts
-        uses: actions/download-artifact@v4
-        with:
-          path: python/sglang/kernels/aot/dist/
-          merge-multiple: true
-          pattern: wheel-*
-
-      - name: Set tag name
-        id: set_tag_name
-        run: |
-          if [ -z "${{ inputs.tag_name }}" ]; then
-            TAG_NAME="v$(cat python/sglang/kernels/aot/python/sgl_kernel/version.py | cut -d'"' -f2)"
-            echo "tag_name=$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: ${{ secrets.GH_PAT_FOR_WHL_RELEASE }}
-          files: |
-            python/sglang/kernels/aot/dist/*
-
-      - name: Clone wheel index
-        run: git clone https://oauth2:${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 --cuda 129
-
-      - 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-cu130-matrix:
     if: |
       github.repository == 'sgl-project/sglang' &&
diff --git a/.github/workflows/retag-docker.yml b/.github/workflows/retag-docker.yml
index 633a275ed..aa6ab03a7 100644
--- a/.github/workflows/retag-docker.yml
+++ b/.github/workflows/retag-docker.yml
@@ -4,7 +4,7 @@ on:
   workflow_dispatch:
     inputs:
       source_tag:
-        description: "Existing image tag (e.g., v0.4.7-cu129-amd64)"
+        description: "Existing image tag (e.g., v0.4.7-cu130-amd64)"
         required: true
       target_tag:
         description: "New tag to apply (e.g., latest)"
diff --git a/.github/workflows/trivy-scan-dev.yml b/.github/workflows/trivy-scan-dev.yml
index f35476597..ab315cab6 100644
--- a/.github/workflows/trivy-scan-dev.yml
+++ b/.github/workflows/trivy-scan-dev.yml
@@ -22,7 +22,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        tag: ${{ inputs.tag && fromJSON(format('["{0}"]', inputs.tag)) || fromJSON('["dev", "dev-cu12"]') }}
+        tag: ${{ inputs.tag && fromJSON(format('["{0}"]', inputs.tag)) || fromJSON('["dev", "dev-cu13"]') }}
     steps:
       - name: Cleanup workspace (remove root-owned files from prior runs)
         run: sudo rm -rf "$GITHUB_WORKSPACE"/* || true
diff --git a/docker/Dockerfile b/docker/Dockerfile
index 49d56652f..1dc3b079b 100644
--- a/docker/Dockerfile
+++ b/docker/Dockerfile
@@ -6,7 +6,6 @@ ARG BUILD_TYPE=all
 ARG BRANCH_TYPE=remote
 ARG SGL_KERNEL_VERSION=0.4.6.post1
 ARG SGL_VERSION
-ARG SGL_DEEP_GEMM_VERSION=0.1.7
 ARG USE_LATEST_SGLANG=0
 ARG GDRCOPY_VERSION=2.5.1
 ARG SGL_NCCL_VERSION=2.30.7
@@ -185,25 +184,11 @@ RUN curl --proto '=https' --tlsv1.2 --retry 3 --retry-delay 2 -sSf https://sh.ru
 RUN --mount=type=cache,target=/root/.cache/pip \
     python3 -m pip install --upgrade pip setuptools wheel html5lib six \
     && case "$CUDA_VERSION" in \
-        12.6.3) CUINDEX=126 ;; \
-        12.9.2) CUINDEX=129 ;; \
-        13.0.3) CUINDEX=130 ;; \
+        13.0.3) ;; \
         *) echo "Unsupported CUDA version: $CUDA_VERSION" && exit 1 ;; \
     esac \
-    && if [ "$CUDA_VERSION" = "12.6.3" ]; then \
-        python3 -m pip install https://${GITHUB_ARTIFACTORY}/sgl-project/whl/releases/download/v${SGL_KERNEL_VERSION}/sglang_kernel-${SGL_KERNEL_VERSION}+cu124-cp310-abi3-manylinux2014_$(uname -m).whl --force-reinstall --no-deps \
-    ; \
-    elif [ "$CUDA_VERSION" = "12.9.2" ]; then \
-        python3 -m pip install https://${GITHUB_ARTIFACTORY}/sgl-project/whl/releases/download/v${SGL_KERNEL_VERSION}/sglang_kernel-${SGL_KERNEL_VERSION}+cu129-cp310-abi3-manylinux2014_$(uname -m).whl --force-reinstall --no-deps \
-    ; \
-    elif [ "$CUDA_VERSION" = "13.0.3" ]; then \
-        # --no-deps prevents pip from pulling torch from default PyPI
-        python3 -m pip install sglang-kernel==${SGL_KERNEL_VERSION} --force-reinstall --no-deps \
-    ; \
-    else \
-        echo "Unsupported CUDA version: $CUDA_VERSION" && exit 1 \
-    ; \
-    fi
+    # --no-deps prevents pip from pulling torch from default PyPI
+    && python3 -m pip install sglang-kernel==${SGL_KERNEL_VERSION} --force-reinstall --no-deps
 
 # Copy dep spec + Rust crate source + proto files. setuptools-rust compiles the
 # Rust extension during the stub wheel build; the crate's build.rs references
@@ -214,15 +199,11 @@ COPY rust/sglang-grpc      /tmp/sglang_deps/rust/sglang-grpc
 COPY rust/sglang-mm        /tmp/sglang_deps/rust/sglang-mm
 COPY proto                 /tmp/sglang_deps/proto
 
-# Install sglang dependencies (torch, transformers, etc.). CUDA 12 DeepEP
-# wheels live only on the SGLang index, so preinstall the local-version wheel;
-# it satisfies the public-version pyproject pin during the full dependency solve.
+# Install sglang dependencies (torch, transformers, etc.).
 # Generate constraints.txt to prevent reinstalling these deps in later stages.
 RUN --mount=type=cache,target=/root/.cache/pip \
     --mount=type=cache,target=/root/.cargo/registry \
     case "$CUDA_VERSION" in \
-        12.6.3) CUINDEX=126 ;; \
-        12.9.2) CUINDEX=129 ;; \
         13.0.3) CUINDEX=130 ;; \
         *) echo "Unsupported CUDA version: $CUDA_VERSION" && exit 1 ;; \
     esac \
@@ -232,27 +213,7 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     && echo '__version__ = "0.0.0"' > sglang/version.py \
     && touch README.md \
     && touch LICENSE \
-    && SGL_DEEP_EP_VERSION="$(sed -n 's/^[[:space:]]*"sgl-deep-ep==\([^"]*\)",/\1/p' pyproject.toml)" \
-    && test -n "${SGL_DEEP_EP_VERSION}" \
-    && if [ "${CUDA_VERSION%%.*}" = "12" ]; then \
-           python3 -m pip install \
-               "sgl-deep-ep==${SGL_DEEP_EP_VERSION}+cu129" \
-               --index-url "https://docs.sglang.ai/whl/cu129/" \
-               --no-deps; \
-       fi \
-    && if [ "${CUDA_VERSION%%.*}" = "12" ]; then \
-           sed -i 's/cuda-python>=13\.0/cuda-python>=12,<13/' pyproject.toml && \
-           sed -i 's/flashinfer_python\[cu13\]/flashinfer_python[cu12]/' pyproject.toml && \
-           sed -i 's/nvidia-cutlass-dsl\[cu13\]/nvidia-cutlass-dsl/' pyproject.toml; \
-       fi \
     && python3 -m pip install --extra-index-url https://download.pytorch.org/whl/cu${CUINDEX} ".[${BUILD_TYPE}]" \
-    && if [ "${CUDA_VERSION%%.*}" = "12" ]; then \
-           pip list --format=freeze | awk -F'==' '/-cu13(==|$)/ {print $1}' \
-               | xargs -r python3 -m pip uninstall -y && \
-           python3 -m pip install --index-url https://download.pytorch.org/whl/cu${CUINDEX} \
-               torch==2.13.0 torchvision==0.28.0 torchaudio==2.11.0 --force-reinstall; \
-           python3 -m pip install https://github.com/sgl-project/whl/releases/download/v${SGL_DEEP_GEMM_VERSION}/sgl_deep_gemm-${SGL_DEEP_GEMM_VERSION}+cu129-py3-none-manylinux2014_$(uname -m).whl --force-reinstall; \
-       fi \
     && cd /sgl-workspace \
     && rm -rf /tmp/sglang_deps \
     && pip freeze | grep -v "^sglang==" > /sgl-workspace/constraints.txt
@@ -315,8 +276,6 @@ ARG FLASHINFER_VERSION
 # Stage jit-cache/cubin artifacts into /flashinfer_jit_output for clean COPY later
 RUN --mount=type=cache,target=/root/.cache/pip \
     case "$CUDA_VERSION" in \
-        12.6.3) CUINDEX=126 ;; \
-        12.9.2) CUINDEX=129 ;; \
         13.0.3) CUINDEX=130 ;; \
         *) echo "Unsupported CUDA version: $CUDA_VERSION" && exit 1 ;; \
     esac \
@@ -503,21 +462,12 @@ RUN --mount=type=cache,target=/var/cache/apt,id=framework-apt \
 
 # Install Mooncake
 RUN --mount=type=cache,target=/root/.cache/pip \
-    CUDA_MAJOR="${CUDA_VERSION%%.*}" && \
-    if [ "$CUDA_MAJOR" -ge 13 ]; then \
-        python3 -m pip install mooncake-transfer-engine-cuda13==${MOONCAKE_VERSION}; \
-    else \
-        python3 -m pip install mooncake-transfer-engine==${MOONCAKE_VERSION}; \
-    fi
+    python3 -m pip install mooncake-transfer-engine-cuda13==${MOONCAKE_VERSION}
 
 # Install MSCCL++ Python dependencies and package (builds extension via CMake through pip)
 RUN --mount=type=cache,target=/root/.cache/pip \
     git clone --depth=1 --branch ${MSCCLPP_VERSION} https://${GITHUB_ARTIFACTORY}/microsoft/mscclpp.git /tmp/mscclpp \
     && case "${CUDA_VERSION}" in \
-           12.*) \
-               CMAKE_ARGS="-DMSCCLPP_BYPASS_GPU_CHECK=ON -DMSCCLPP_USE_CUDA=ON -DMSCCLPP_GPU_ARCHS=80,90,100,100a,103,103a" \
-               python3 -m pip install "/tmp/mscclpp[cuda12]"; \
-               ;; \
            13.*) \
                CMAKE_ARGS="-DMSCCLPP_BYPASS_GPU_CHECK=ON -DMSCCLPP_USE_CUDA=ON -DMSCCLPP_GPU_ARCHS=80,90,100,100a,103,103a" \
                python3 -m pip install "/tmp/mscclpp[cuda13]"; \
@@ -553,22 +503,15 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 # resolving dependencies and generating constraints.txt. DeepEP needs 2.30.7
 # at runtime; apply the override only after the constrained dependency solve.
 RUN --mount=type=cache,target=/root/.cache/pip \
-    if [ "${CUDA_VERSION%%.*}" = "13" ]; then \
-        python3 -m pip install --force-reinstall --no-deps \
-            "nvidia-nccl-cu13==${SGL_NCCL_VERSION}"; \
-    fi
+    python3 -m pip install --force-reinstall --no-deps \
+        "nvidia-nccl-cu13==${SGL_NCCL_VERSION}"
 
-# Per-CUDA-major package installs. The `nixl` stub package is needed (it owns
-# the `nixl` import path) but unconditionally requires nixl-cu12, so we install
-# it with --no-deps and pair it with the matching nixl-cu12 / nixl-cu13 binary
-# to avoid shipping wrong-CUDA libs on cu13 images.
-RUN --mount=type=cache,target=/root/.cache/pip if [ "${CUDA_VERSION%%.*}" = "12" ]; then \
-    python3 -m pip install nixl nixl-cu12 --no-deps ; \
-    python3 -m pip install "cuda-python>=12,<13" ; \
-elif [ "${CUDA_VERSION%%.*}" = "13" ]; then \
-    python3 -m pip install nixl nixl-cu13 --no-deps ; \
-    python3 -m pip install "cuda-python>=13,<14" ; \
-fi
+# The `nixl` stub package is needed (it owns the `nixl` import path) but
+# unconditionally requires nixl-cu12, so install it with --no-deps and pair it
+# with the nixl-cu13 binary to avoid shipping wrong-CUDA libs.
+RUN --mount=type=cache,target=/root/.cache/pip \
+    python3 -m pip install nixl nixl-cu13 --no-deps \
+    && python3 -m pip install "cuda-python>=13,<14"
 
 # Install the latest available AI Dynamo prerelease from NVIDIA's package index.
 RUN --mount=type=cache,target=/root/.cache/pip \
@@ -638,11 +581,6 @@ RUN if [ "$BRANCH_TYPE" = "local" ]; then \
 # Clean up __pycache__/tests/pyc in same RUN to avoid writing ~28k files to layer
 RUN --mount=type=cache,target=/root/.cache/pip \
     cd /sgl-workspace/sglang \
-    && if [ "${CUDA_VERSION%%.*}" = "12" ]; then \
-           sed -i 's/cuda-python>=13\.0/cuda-python>=12,<13/' python/pyproject.toml && \
-           sed -i 's/flashinfer_python\[cu13\]/flashinfer_python[cu12]/' python/pyproject.toml && \
-           sed -i 's/nvidia-cutlass-dsl\[cu13\]/nvidia-cutlass-dsl/' python/pyproject.toml; \
-       fi \
     && python3 -m pip install --no-deps -e "python[${BUILD_TYPE}]" \
     && kernels lock python \
     && ( success=0; \
@@ -659,8 +597,8 @@ RUN --mount=type=cache,target=/root/.cache/pip \
                  echo "sgl-kernel cubin download failed, retrying in 30s..." && sleep 30; \
              done; \
              # x86: if no prebuilt sgl-flash-attn3 variant matches this torch+CUDA \
-             # combo (e.g. cu129 publishes no torch>=2.10 cubin), fall back to \
-             # runtime JIT instead of failing the build, mirroring the aarch64 branch. \
+             # combo, fall back to runtime JIT instead of failing the build, \
+             # mirroring the aarch64 branch. \
              if [ "$success" != "1" ]; then \
                  echo "WARNING: no matching sgl-flash-attn3 cubin variant for this torch+CUDA; kernels will be JIT-compiled at runtime"; \
                  success=1; \
diff --git a/docker/Dockerfile.cu134 b/docker/Dockerfile.cu134
index d6d1f9a00..049437799 100644
--- a/docker/Dockerfile.cu134
+++ b/docker/Dockerfile.cu134
@@ -493,19 +493,11 @@ RUN curl --proto '=https' --tlsv1.2 --retry 3 --retry-delay 2 -sSf https://sh.ru
 RUN --mount=type=cache,target=/root/.cache/pip \
     python3 -m pip install --upgrade pip setuptools wheel html5lib six \
     && case "$CUDA_VERSION" in \
-        12.6.3) CUINDEX=126 ;; \
-        12.9.2) CUINDEX=129 ;; \
         13.0.3) CUINDEX=130 ;; \
         13.4.0) CUINDEX=134 ;; \
         *) echo "Unsupported CUDA version: $CUDA_VERSION" && exit 1 ;; \
     esac \
-    && if [ "$CUDA_VERSION" = "12.6.3" ]; then \
-        python3 -m pip install https://${GITHUB_ARTIFACTORY}/sgl-project/whl/releases/download/v${SGL_KERNEL_VERSION}/sglang_kernel-${SGL_KERNEL_VERSION}+cu124-cp310-abi3-manylinux2014_$(uname -m).whl --force-reinstall --no-deps \
-    ; \
-    elif [ "$CUDA_VERSION" = "12.9.2" ]; then \
-        python3 -m pip install https://${GITHUB_ARTIFACTORY}/sgl-project/whl/releases/download/v${SGL_KERNEL_VERSION}/sglang_kernel-${SGL_KERNEL_VERSION}+cu129-cp310-abi3-manylinux2014_$(uname -m).whl --force-reinstall --no-deps \
-    ; \
-    elif [ "$CUDA_VERSION" = "13.0.3" ]; then \
+    && if [ "$CUDA_VERSION" = "13.0.3" ]; then \
         # --no-deps prevents pip from pulling torch from default PyPI
         python3 -m pip install sglang-kernel==${SGL_KERNEL_VERSION} --force-reinstall --no-deps \
     ; \
@@ -527,15 +519,11 @@ COPY rust/sglang-grpc      /tmp/sglang_deps/rust/sglang-grpc
 COPY rust/sglang-mm        /tmp/sglang_deps/rust/sglang-mm
 COPY proto                 /tmp/sglang_deps/proto
 
-# Install sglang dependencies (torch, transformers, etc.). CUDA 12 DeepEP
-# wheels live only on the SGLang index, so preinstall the local-version wheel;
-# it satisfies the public-version pyproject pin during the full dependency solve.
+# Install sglang dependencies (torch, transformers, etc.).
 # Generate constraints.txt to prevent reinstalling these deps in later stages.
 RUN --mount=type=cache,target=/root/.cache/pip \
     --mount=type=cache,target=/root/.cargo/registry \
     case "$CUDA_VERSION" in \
-        12.6.3) CUINDEX=126 ;; \
-        12.9.2) CUINDEX=129 ;; \
         13.0.3) CUINDEX=130 ;; \
         13.4.0) CUINDEX=134 ;; \
         *) echo "Unsupported CUDA version: $CUDA_VERSION" && exit 1 ;; \
@@ -546,19 +534,6 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     && echo '__version__ = "0.0.0"' > sglang/version.py \
     && touch README.md \
     && touch LICENSE \
-    && SGL_DEEP_EP_VERSION="$(sed -n 's/^[[:space:]]*"sgl-deep-ep==\([^"]*\)",/\1/p' pyproject.toml)" \
-    && test -n "${SGL_DEEP_EP_VERSION}" \
-    && if [ "${CUDA_VERSION%%.*}" = "12" ]; then \
-           python3 -m pip install \
-               "sgl-deep-ep==${SGL_DEEP_EP_VERSION}+cu129" \
-               --index-url "https://docs.sglang.ai/whl/cu129/" \
-               --no-deps; \
-       fi \
-    && if [ "${CUDA_VERSION%%.*}" = "12" ]; then \
-           sed -i 's/cuda-python>=13\.0/cuda-python>=12,<13/' pyproject.toml && \
-           sed -i 's/flashinfer_python\[cu13\]/flashinfer_python[cu12]/' pyproject.toml && \
-           sed -i 's/nvidia-cutlass-dsl\[cu13\]/nvidia-cutlass-dsl/' pyproject.toml; \
-       fi \
     && if [ "$CUDA_VERSION" = "13.4.0" ]; then \
            sed -i "s|^  \"flashinfer_python\[cu13\]==.*|  \"flashinfer_python[cu13]==${FLASHINFER_VERSION}\",|" pyproject.toml; \
        fi \
@@ -579,17 +554,8 @@ RUN --mount=type=cache,target=/root/.cache/pip \
            TORCH_INDEX="https://download.pytorch.org/whl/cu${CUINDEX}"; \
        fi \
     && python3 -m pip install --extra-index-url "${TORCH_INDEX}" ".[${BUILD_TYPE}]" \
-    && if [ "${CUDA_VERSION%%.*}" = "13" ]; then \
-           python3 -m pip install --force-reinstall --no-deps \
-               "nvidia-nccl-cu13==${NCCL_VERSION}"; \
-       fi \
-    && if [ "${CUDA_VERSION%%.*}" = "12" ]; then \
-           pip list --format=freeze | awk -F'==' '/-cu13(==|$)/ {print $1}' \
-               | xargs -r python3 -m pip uninstall -y && \
-           python3 -m pip install --index-url https://download.pytorch.org/whl/cu${CUINDEX} \
-               torch==2.13.0 torchvision==0.28.0 torchaudio==2.11.0 --force-reinstall; \
-           python3 -m pip install https://github.com/sgl-project/whl/releases/download/v${SGL_DEEP_GEMM_VERSION}/sgl_deep_gemm-${SGL_DEEP_GEMM_VERSION}+cu129-py3-none-manylinux2014_$(uname -m).whl --force-reinstall; \
-       fi \
+    && python3 -m pip install --force-reinstall --no-deps \
+           "nvidia-nccl-cu13==${NCCL_VERSION}" \
     && if [ "$CUDA_VERSION" = "13.4.0" ]; then \
            python3 -m pip install --force-reinstall --no-deps /tmp/cu134_wheels/*.whl; \
        fi \
@@ -643,8 +609,6 @@ ARG FLASHINFER_VERSION
 # Stage jit-cache/cubin artifacts into /flashinfer_jit_output for clean COPY later
 RUN --mount=type=cache,target=/root/.cache/pip \
     case "$CUDA_VERSION" in \
-        12.6.3) CUINDEX=126 ;; \
-        12.9.2) CUINDEX=129 ;; \
         13.0.3) CUINDEX=130 ;; \
         13.4.0) CUINDEX=134 ;; \
         *) echo "Unsupported CUDA version: $CUDA_VERSION" && exit 1 ;; \
@@ -830,21 +794,12 @@ RUN --mount=type=cache,target=/var/cache/apt,id=framework-apt \
 
 # Install Mooncake
 RUN --mount=type=cache,target=/root/.cache/pip \
-    CUDA_MAJOR="${CUDA_VERSION%%.*}" && \
-    if [ "$CUDA_MAJOR" -ge 13 ]; then \
-        python3 -m pip install mooncake-transfer-engine-cuda13==${MOONCAKE_VERSION}; \
-    else \
-        python3 -m pip install mooncake-transfer-engine==${MOONCAKE_VERSION}; \
-    fi
+    python3 -m pip install mooncake-transfer-engine-cuda13==${MOONCAKE_VERSION}
 
 # Install MSCCL++ Python dependencies and package (builds extension via CMake through pip)
 RUN --mount=type=cache,target=/root/.cache/pip \
     git clone --depth=1 --branch ${MSCCLPP_VERSION} https://${GITHUB_ARTIFACTORY}/microsoft/mscclpp.git /tmp/mscclpp \
     && case "${CUDA_VERSION}" in \
-           12.*) \
-               CMAKE_ARGS="-DMSCCLPP_BYPASS_GPU_CHECK=ON -DMSCCLPP_USE_CUDA=ON -DMSCCLPP_GPU_ARCHS=80,90,100,100a,103,103a" \
-               python3 -m pip install "/tmp/mscclpp[cuda12]"; \
-               ;; \
            13.*) \
                CMAKE_ARGS="-DMSCCLPP_BYPASS_GPU_CHECK=ON -DMSCCLPP_USE_CUDA=ON -DMSCCLPP_GPU_ARCHS=80,90,100,100a,103,103a" \
                python3 -m pip install "/tmp/mscclpp[cuda13]"; \
@@ -882,17 +837,12 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     termplotlib \
     "runai-model-streamer[s3,gcs,azure]>=0.15.7"
 
-# Per-CUDA-major package installs. The `nixl` stub package is needed (it owns
-# the `nixl` import path) but unconditionally requires nixl-cu12, so we install
-# it with --no-deps and pair it with the matching nixl-cu12 / nixl-cu13 binary
-# to avoid shipping wrong-CUDA libs on cu13 images.
-RUN --mount=type=cache,target=/root/.cache/pip if [ "${CUDA_VERSION%%.*}" = "12" ]; then \
-    python3 -m pip install nixl nixl-cu12 --no-deps ; \
-    python3 -m pip install "cuda-python>=12,<13" ; \
-elif [ "${CUDA_VERSION%%.*}" = "13" ]; then \
-    python3 -m pip install nixl nixl-cu13 --no-deps ; \
-    python3 -m pip install "cuda-python>=13,<14" ; \
-fi
+# The `nixl` stub package is needed (it owns the `nixl` import path) but
+# unconditionally requires nixl-cu12, so install it with --no-deps and pair it
+# with the nixl-cu13 binary to avoid shipping wrong-CUDA libs.
+RUN --mount=type=cache,target=/root/.cache/pip \
+    python3 -m pip install nixl nixl-cu13 --no-deps \
+    && python3 -m pip install "cuda-python>=13,<14"
 
 # Install the latest available AI Dynamo prerelease from NVIDIA's package index.
 RUN --mount=type=cache,target=/root/.cache/pip \
@@ -965,11 +915,6 @@ RUN if [ "$BRANCH_TYPE" = "local" ]; then \
 # Clean up __pycache__/tests/pyc in same RUN to avoid writing ~28k files to layer
 RUN --mount=type=cache,target=/root/.cache/pip \
     cd /sgl-workspace/sglang \
-    && if [ "${CUDA_VERSION%%.*}" = "12" ]; then \
-           sed -i 's/cuda-python>=13\.0/cuda-python>=12,<13/' python/pyproject.toml && \
-           sed -i 's/flashinfer_python\[cu13\]/flashinfer_python[cu12]/' python/pyproject.toml && \
-           sed -i 's/nvidia-cutlass-dsl\[cu13\]/nvidia-cutlass-dsl/' python/pyproject.toml; \
-       fi \
     && if [ "$CUDA_VERSION" = "13.4.0" ]; then \
            sed -i "s|^  \"flashinfer_python\[cu13\]==.*|  \"flashinfer_python[cu13]==${FLASHINFER_VERSION}\",|" python/pyproject.toml; \
        fi \
@@ -989,8 +934,8 @@ RUN --mount=type=cache,target=/root/.cache/pip \
                  echo "sgl-kernel cubin download failed, retrying in 30s..." && sleep 30; \
              done; \
              # x86: if no prebuilt sgl-flash-attn3 variant matches this torch+CUDA \
-             # combo (e.g. cu129 publishes no torch>=2.10 cubin), fall back to \
-             # runtime JIT instead of failing the build, mirroring the aarch64 branch. \
+             # combo, fall back to runtime JIT instead of failing the build, \
+             # mirroring the aarch64 branch. \
              if [ "$success" != "1" ]; then \
                  echo "WARNING: no matching sgl-flash-attn3 cubin variant for this torch+CUDA; kernels will be JIT-compiled at runtime"; \
                  success=1; \
diff --git a/docker/sgl-deep-ep.Dockerfile b/docker/sgl-deep-ep.Dockerfile
index a3743073b..4ccf930f3 100644
--- a/docker/sgl-deep-ep.Dockerfile
+++ b/docker/sgl-deep-ep.Dockerfile
@@ -6,7 +6,6 @@ FROM ${BASE_IMAGE}:cuda${CUDA_VERSION}
 ARG ARCHITECTURE=x86_64
 ARG CUDA_TAG=cu130
 ARG CUDA_VERSION=13.0
-ARG GDRCOPY_VERSION=2.5.1
 ARG NCCL_VERSION=2.30.7
 ARG PYTHON_TAG=cp312-cp312
 ARG TORCH_VERSION=2.13.0
@@ -49,19 +48,6 @@ RUN set -eux; \
     ln -sf "${cuda_stub}" /usr/lib64/libcuda.so; \
     ln -sf "${cuda_stub}" "/usr/lib/${ARCHITECTURE}-linux-gnu/libcuda.so"
 
-# DeepEP v2 uses NCCL Gin on CUDA 13. Keep GDRCopy only for the CUDA 12
-# legacy NVSHMEM/IBGDA build.
-RUN set -eux; \
-    if [ "${CUDA_TAG}" = cu129 ]; then \
-        git clone --depth 1 --branch "v${GDRCOPY_VERSION}" \
-            https://github.com/NVIDIA/gdrcopy.git /opt/gdrcopy; \
-        make -C /opt/gdrcopy CUDA="${CUDA_HOME}" prefix=/usr/local lib_install; \
-        printf '%s\n' /usr/local/lib > /etc/ld.so.conf.d/gdrcopy.conf; \
-        ldconfig; \
-        test -f /usr/local/include/gdrapi.h; \
-        ldconfig -p | grep -q libgdrapi; \
-    fi
-
 RUN --mount=type=cache,id=sgl-deep-ep-pip-${CUDA_TAG}-${PYTHON_TAG}-${ARCHITECTURE},target=/root/.cache/pip \
     set -eux; \
     "${PYTHON_BIN}" -m pip uninstall -y deep-ep sgl-deep-ep || true; \
@@ -69,10 +55,8 @@ RUN --mount=type=cache,id=sgl-deep-ep-pip-${CUDA_TAG}-${PYTHON_TAG}-${ARCHITECTU
     "${PYTHON_BIN}" -m pip install --force-reinstall \
         "torch==${TORCH_VERSION}" \
         --index-url "https://download.pytorch.org/whl/${CUDA_TAG}"; \
-    if [ "${CUDA_TAG}" = cu130 ]; then \
-        "${PYTHON_BIN}" -m pip install --force-reinstall --no-deps \
-            "nvidia-nccl-cu13==${NCCL_VERSION}"; \
-    fi; \
+    "${PYTHON_BIN}" -m pip install --force-reinstall --no-deps \
+        "nvidia-nccl-cu13==${NCCL_VERSION}"; \
     "${PYTHON_BIN}" -m pip install \
         "auditwheel>=6.0" \
         build \
diff --git a/docker/sgl-deep-gemm.Dockerfile b/docker/sgl-deep-gemm.Dockerfile
index 9fc653014..c76b2a67a 100644
--- a/docker/sgl-deep-gemm.Dockerfile
+++ b/docker/sgl-deep-gemm.Dockerfile
@@ -27,8 +27,7 @@ RUN --mount=type=cache,id=sgl-deep-gemm-pip,target=/root/.cache/pip \
     set -eux; \
     case "${CUDA_VERSION}" in \
       13.0) CU_TAG=cu130 ;; \
-      12.9) CU_TAG=cu129 ;; \
-      *)    CU_TAG=cu130 ;; \
+      *) echo "Unsupported CUDA version: ${CUDA_VERSION}" && exit 1 ;; \
     esac; \
     ${PYTHON_ROOT_PATH}/bin/pip install torch==${TORCH_VER} --index-url https://${PYTORCH_MIRROR}/whl/${CU_TAG}; \
     ${PYTHON_ROOT_PATH}/bin/pip install --index-url ${PIP_DEFAULT_INDEX} \
diff --git a/docs/cookbook/autoregressive/MiniMax/MiniMax-M3.mdx b/docs/cookbook/autoregressive/MiniMax/MiniMax-M3.mdx
index 8326f3dc8..d24cb14c1 100644
--- a/docs/cookbook/autoregressive/MiniMax/MiniMax-M3.mdx
+++ b/docs/cookbook/autoregressive/MiniMax/MiniMax-M3.mdx
@@ -39,7 +39,7 @@ Then run the **Python** output of the command panel below in that environment. T
 docker pull lmsysorg/sglang:dev-cu13-minimax-m3
 ```
 
-The command panel below fills in the right tag per platform: `dev-cu13-minimax-m3` (CUDA 13 — B300, GB200, GB300), `dev-cu12-minimax-m3` (CUDA 12 — Hopper H200), or `dev-minimax-m3` (default). On AMD Instinct it uses the matching ROCm image (MI300X/MI325X → `aigmkt/minimax-m3-sglang-rocm700-mi30x`, MI350X/MI355X → `aigmkt/minimax-m3-sglang-rocm720-mi35x`). For how to launch the image, see [Install → Method 3: Using Docker](../../../docs/get-started/install#method-3-using-docker), substituting the inner `sglang serve ...` with what the command generator produces.
+The command panel below fills in the right tag per platform: `dev-cu13-minimax-m3` (CUDA 13 — B300, GB200, GB300), `dev-cu12-minimax-m3` (CUDA 12 — Hopper H200; final build, the CUDA 12 lane is retired), or `dev-minimax-m3` (default). On AMD Instinct it uses the matching ROCm image (MI300X/MI325X → `aigmkt/minimax-m3-sglang-rocm700-mi30x`, MI350X/MI355X → `aigmkt/minimax-m3-sglang-rocm720-mi35x`). For how to launch the image, see [Install → Method 3: Using Docker](../../../docs/get-started/install#method-3-using-docker), substituting the inner `sglang serve ...` with what the command generator produces.
 
 <Note>
 These M3 dev images now **bundle MiniMax's MSA sparse-attention kernel** (`fmha_sm100`), so Blackwell users get the recommended fast path automatically — no manual install needed (see **§2.1**). On a custom image without it, the same recipe still serves on the built-in Triton sparse path.
diff --git a/docs/cookbook/autoregressive/OpenBMB/MiniCPM-V-4_6.mdx b/docs/cookbook/autoregressive/OpenBMB/MiniCPM-V-4_6.mdx
index 04ec72563..55a788b49 100644
--- a/docs/cookbook/autoregressive/OpenBMB/MiniCPM-V-4_6.mdx
+++ b/docs/cookbook/autoregressive/OpenBMB/MiniCPM-V-4_6.mdx
@@ -32,7 +32,8 @@ Pull the nightly Docker image (rolling tag, tracks `main`):
 # CUDA 13 (Hopper / Blackwell, default)
 docker pull lmsysorg/sglang:dev
 
-# CUDA 12 (Ampere or older drivers)
+# CUDA 12 (Ampere or older drivers). Final build: the CUDA 12 lane is retired,
+# so this tag no longer tracks main.
 docker pull lmsysorg/sglang:dev-cu12
 ```
 
diff --git a/docs/cookbook/autoregressive/ThinkingMachines/Inkling-Small.mdx b/docs/cookbook/autoregressive/ThinkingMachines/Inkling-Small.mdx
index 5bc7305aa..ce9315173 100644
--- a/docs/cookbook/autoregressive/ThinkingMachines/Inkling-Small.mdx
+++ b/docs/cookbook/autoregressive/ThinkingMachines/Inkling-Small.mdx
@@ -35,7 +35,7 @@ There are two multi-arch (amd64 / arm64) CUDA builds plus a ROCm build; pick the
 
 ```bash Command
 docker pull lmsysorg/sglang:dev-inkling-dspark      # CUDA 13
-docker pull lmsysorg/sglang:dev-cu12-inkling-dspark # CUDA 12
+docker pull lmsysorg/sglang:dev-cu12-inkling-dspark # CUDA 12 (final build; lane retired)
 docker pull lmsysorg/sglang:dev-inkling-small-dgx-spark # DGX Spark (GB10 / SM121)
 docker pull lmsysorg/sglang-rocm:dev-rocm720-mi35x-inkling-dspark  # AMD MI350X / MI355X
 ```
@@ -309,4 +309,4 @@ To try it, select the **Long Context** strategy in the Deploy panel above for an
 
 The **DSpark** deploy strategy is the second speculative-decoding path for Inkling-Small. Unlike **MTP**, which drives Inkling-Small's own multi-layer draft head, DSpark runs a **separate draft checkpoint** — `RadixArk/Inkling-Small-DSpark` — served unquantized alongside the NVFP4 target.
 
-DSpark support ships in the images listed in §1 (`dev-inkling-dspark` for CUDA 13, `dev-cu12-inkling-dspark` for CUDA 12), so no separate build is needed. Verified end-to-end on B200 (TP=8, NVFP4).
+DSpark support ships in the images listed in §1 (`dev-inkling-dspark` for CUDA 13, `dev-cu12-inkling-dspark` for CUDA 12 — a final build, since that lane is retired), so no separate build is needed. Verified end-to-end on B200 (TP=8, NVFP4).
diff --git a/docs/cookbook/autoregressive/ThinkingMachines/Inkling.mdx b/docs/cookbook/autoregressive/ThinkingMachines/Inkling.mdx
index b7a74b979..4ac2d511b 100644
--- a/docs/cookbook/autoregressive/ThinkingMachines/Inkling.mdx
+++ b/docs/cookbook/autoregressive/ThinkingMachines/Inkling.mdx
@@ -35,7 +35,7 @@ There are two multi-arch (amd64 / arm64) CUDA builds plus a ROCm build; pick the
 
 ```bash Command
 docker pull lmsysorg/sglang:dev-inkling-dspark      # CUDA 13
-docker pull lmsysorg/sglang:dev-cu12-inkling-dspark # CUDA 12
+docker pull lmsysorg/sglang:dev-cu12-inkling-dspark # CUDA 12 (final build; lane retired)
 docker pull lmsysorg/sglang-rocm:dev-rocm720-mi35x-inkling-dspark  # AMD MI350X / MI355X
 ```
 
@@ -306,4 +306,4 @@ To try it, select the **Long Context** strategy in the Deploy panel above for an
 
 The **DSpark** deploy strategy is the second speculative-decoding path for Inkling. Unlike **MTP**, which drives Inkling's own multi-layer draft head, DSpark runs a **separate draft checkpoint** — `RadixArk/Inkling-DSpark-Preview` — served unquantized alongside the NVFP4 target.
 
-DSpark support ships in the images listed in §1 (`dev-inkling-dspark` for CUDA 13, `dev-cu12-inkling-dspark` for CUDA 12), so no separate build is needed. Verified end-to-end on B200 (TP=8, NVFP4).
+DSpark support ships in the images listed in §1 (`dev-inkling-dspark` for CUDA 13, `dev-cu12-inkling-dspark` for CUDA 12 — a final build, since that lane is retired), so no separate build is needed. Verified end-to-end on B200 (TP=8, NVFP4).
diff --git a/docs/docs/developer_guide/contribution_guide.mdx b/docs/docs/developer_guide/contribution_guide.mdx
index 09005fa17..dc1ef9f67 100644
--- a/docs/docs/developer_guide/contribution_guide.mdx
+++ b/docs/docs/developer_guide/contribution_guide.mdx
@@ -206,9 +206,9 @@ Develop SGLang's customized DeepGEMM package on the [`dev` branch of `sgl-projec
 
 ### Update sgl-deep-ep
 
-Develop `sgl-deep-ep` in [`sgl-project/DeepEP`](https://github.com/sgl-project/DeepEP). Use the implementation branch for the target platform: `sgl-deepep` for CUDA 13 on x86_64 or aarch64, `sgl-deepep-cu12-x86` for CUDA 12.9 on x86_64, or `sgl-deepep-cu12-arm` for CUDA 12.9 on aarch64. Merge packaging changes into `sgl-deepep-packaging`. The [`sgl-deep-ep` README](https://github.com/sgl-project/DeepEP/blob/sgl-deepep-packaging/sgl_deep_ep/README.md) describes the platform prerequisites and release matrix.
+Develop `sgl-deep-ep` in [`sgl-project/DeepEP`](https://github.com/sgl-project/DeepEP). Use `sgl-deepep`, the implementation branch for CUDA 13 on x86_64 or aarch64. Merge packaging changes into `sgl-deepep-packaging`. The [`sgl-deep-ep` README](https://github.com/sgl-project/DeepEP/blob/sgl-deepep-packaging/sgl_deep_ep/README.md) describes the platform prerequisites and release matrix.
 
-To validate locally, check out the selected implementation branch as `DeepEP-source` and the packaging branch as `DeepEP-packaging`. Install the required build dependencies first; CUDA 12.9 builds also require GDRCopy. The following CUDA 13 example builds a wheel for the host architecture, installs that exact wheel, and verifies that its guarded package import succeeds:
+To validate locally, check out the selected implementation branch as `DeepEP-source` and the packaging branch as `DeepEP-packaging`. Install the required build dependencies first. The following example builds a wheel for the host architecture, installs that exact wheel, and verifies that its guarded package import succeeds:
 
 ```bash
 DEEPEP_OUTPUT_DIR="$(mktemp -d)"
@@ -223,17 +223,13 @@ python3 -m pip install --force-reinstall --no-deps \
 python3 -c "import deep_ep; print(deep_ep.__file__)"
 ```
 
-Use `12.9` instead of `13.0` for a CUDA 12.9 build. The import check validates packaging and binary loading, but not communication correctness. On a configured multi-GPU host, also run the test appropriate for the implementation branch:
+The import check validates packaging and binary loading, but not communication correctness. On a configured multi-GPU host, also run:
 
 ```bash
-# CUDA 13 implementation branch
 python3 DeepEP-source/tests/elastic/test_ep.py --num-processes 8
-
-# CUDA 12.9 implementation branches
-python3 DeepEP-source/tests/test_intranode.py --num-processes 8
 ```
 
-Adjust `--num-processes` to the available GPUs and run the internode or low-latency tests when those transports changed. After local validation, ask the SGLang team to run the [sgl-deep-ep release workflow](https://github.com/sgl-project/sglang/actions/workflows/release-whl-deepep.yml) with the new version, CUDA target, and packaging ref. After verifying the published wheels for the supported Python versions and architectures, update the `sgl-deep-ep` pin in `python/pyproject.toml` before landing dependent SGLang changes.
+Adjust `--num-processes` to the available GPUs and run the internode or low-latency tests when those transports changed. After local validation, ask the SGLang team to run the [sgl-deep-ep release workflow](https://github.com/sgl-project/sglang/actions/workflows/release-whl-deepep.yml) with the new version and packaging ref. After verifying the published wheels for the supported Python versions and architectures, update the `sgl-deep-ep` pin in `python/pyproject.toml` before landing dependent SGLang changes.
 
 ## Tips for newcomers
 
diff --git a/docs/docs/get-started/install.mdx b/docs/docs/get-started/install.mdx
index 669bcf482..b7d715566 100644
--- a/docs/docs/get-started/install.mdx
+++ b/docs/docs/get-started/install.mdx
@@ -29,15 +29,11 @@ uv pip install --prerelease=allow sglang
 Some of SGLang's dependencies only publish pre-releases on PyPI, so without `--prerelease=allow` uv older than 0.12.0 silently installs SGLang 0.5.9. On [uv 0.12.0](https://github.com/astral-sh/uv/releases/tag/0.12.0) and newer the flag is a harmless no-op.
 </Note>
 
-The major version of Cuda is 13 by default. To install sglang under Cuda 12 with pip or uv, please try the following commands:
-```bash Command
-pip install --upgrade pip
-pip install uv
-uv pip install --prerelease=allow sglang
-uv pip install --force-reinstall torch==2.13.0 torchaudio==2.11.0 torchvision --index-url https://download.pytorch.org/whl/cu129
-uv pip install --force-reinstall sglang-kernel --index-url https://docs.sglang.ai/whl/cu129/
-uv pip install --force-reinstall sgl-deep-gemm --index-url https://docs.sglang.ai/whl/cu129/ --no-deps
-```
+<Note>
+SGLang requires CUDA 13. The CUDA 12 (`cu129`) wheels and images are retired,
+because PyTorch 2.14 publishes no CUDA 12.9 builds. SGLang 0.5.19 is the last
+release with a CUDA 12 lane.
+</Note>
 
 ### Nightly builds
 
@@ -49,13 +45,6 @@ pip install uv
 uv pip install --prerelease=allow --index-strategy unsafe-best-match --extra-index-url https://docs.sglang.ai/whl/cu130/ sglang
 ```
 
-To install a nightly build under Cuda 12, swap the index to `cu129`:
-```bash Command
-pip install --upgrade pip
-pip install uv
-uv pip install --prerelease=allow --index-strategy unsafe-best-match --extra-index-url https://docs.sglang.ai/whl/cu129/ sglang
-```
-
 ### Quick fixes to common problems
 - If you encounter `OSError: CUDA_HOME environment variable is not set`. Please set it to your CUDA install root with either of the following solutions:
   1. Use `export CUDA_HOME=/usr/local/cuda-<your-cuda-version>` to set the `CUDA_HOME` environment variable.
@@ -113,7 +102,7 @@ docker run --gpus all \
 You can also find the nightly docker images [here](https://hub.docker.com/r/lmsysorg/sglang/tags?name=nightly).
 
 Notes:
-- SGLang is shipped with CUDA 13 environment by default. To run SGLang on CUDA 12 environment, please use images with `-cu12` or `-cu129` suffix, such as `lmsysorg/sglang:latest-cu129` or `lmsysorg/sglang:dev-cu12`.
+- SGLang images ship a CUDA 13 environment. The `-cu12` / `-cu129` images are no longer published; `lmsysorg/sglang:v0.5.19-cu129` is the last CUDA 12 tag.
 
 ## Method 4: Using Kubernetes
 
diff --git a/docs/src/snippets/configs/MiniMaxAI/minimax-m3.jsx b/docs/src/snippets/configs/MiniMaxAI/minimax-m3.jsx
index 333c4ff1d..09fb65dd4 100644
--- a/docs/src/snippets/configs/MiniMaxAI/minimax-m3.jsx
+++ b/docs/src/snippets/configs/MiniMaxAI/minimax-m3.jsx
@@ -93,9 +93,10 @@ sgl-eval run mmmu_pro \\
 
   dockerImages: {
     // M3-specific dev images (multi-arch amd64+arm64). cu13 carries the sm_103
-    // (B300/GB300) + Grace arm64 builds; cu12 is the Hopper/CUDA-12 build;
-    // dev-minimax-m3 is the rolling default. M3 model support is not yet in a
-    // tagged release, so :latest cannot serve it.
+    // (B300/GB300) + Grace arm64 builds; cu12 is the final Hopper/CUDA-12 build
+    // (that lane is retired, so it no longer rebuilds); dev-minimax-m3 is the
+    // rolling default. M3 model support is not yet in a tagged release, so
+    // :latest cannot serve it.
     b200: "lmsysorg/sglang:dev-minimax-m3",
     b300: "lmsysorg/sglang:dev-cu13-minimax-m3",
     gb200: "lmsysorg/sglang:dev-cu13-minimax-m3",
diff --git a/docs/src/snippets/configs/thinkingmachines/inkling-small.jsx b/docs/src/snippets/configs/thinkingmachines/inkling-small.jsx
index 40b019389..eb3e7d8ae 100644
--- a/docs/src/snippets/configs/thinkingmachines/inkling-small.jsx
+++ b/docs/src/snippets/configs/thinkingmachines/inkling-small.jsx
@@ -70,7 +70,8 @@ export const config = {
 -d '{ "model": "{{MODEL_NAME}}", "messages": [{"role":"user","content":"Hello"}] }'`,
 
   // NVIDIA: two multi-arch CUDA builds (dev-inkling-dspark for CUDA 13,
-  // dev-cu12-inkling-dspark for CUDA 12) — pick by your CUDA version, not by GPU.
+  // dev-cu12-inkling-dspark for CUDA 12, a final build since that lane is
+  // retired) — pick by your CUDA version, not by GPU.
   // Panel defaults to cu13. AMD: dev-rocm720-mi35x-inkling-dspark (sglang-rocm repo).
   // DGX Spark uses a dedicated arm64 CUDA 13 image with NCCL 2.30.7.
   // All tiers ship from the same images, DSpark included (except DGX Spark).
diff --git a/docs/src/snippets/configs/thinkingmachines/inkling.jsx b/docs/src/snippets/configs/thinkingmachines/inkling.jsx
index 05de1befa..9287736eb 100644
--- a/docs/src/snippets/configs/thinkingmachines/inkling.jsx
+++ b/docs/src/snippets/configs/thinkingmachines/inkling.jsx
@@ -73,7 +73,8 @@ export const config = {
 -d '{ "model": "{{MODEL_NAME}}", "messages": [{"role":"user","content":"Hello"}] }'`,
 
   // NVIDIA: two multi-arch CUDA builds (dev-inkling-dspark for CUDA 13,
-  // dev-cu12-inkling-dspark for CUDA 12) — pick by your CUDA version, not by GPU.
+  // dev-cu12-inkling-dspark for CUDA 12, a final build since that lane is
+  // retired) — pick by your CUDA version, not by GPU.
   // Panel defaults to cu13. AMD: dev-rocm720-mi35x-inkling-dspark (sglang-rocm repo).
   // All tiers ship from the same images, DSpark included.
   dockerImages: {
diff --git a/python/sglang/kernels/aot/Dockerfile b/python/sglang/kernels/aot/Dockerfile
index f8f2853c3..d1821c15c 100644
--- a/python/sglang/kernels/aot/Dockerfile
+++ b/python/sglang/kernels/aot/Dockerfile
@@ -98,8 +98,7 @@ RUN --mount=type=cache,id=sgl-kernel-pip,target=/root/.cache/pip \
     set -eux; \
     case "${CUDA_VERSION}" in \
       13.0) TORCH_VER=2.13.0; CU_TAG=cu130 ;; \
-      12.9) TORCH_VER=2.13.0; CU_TAG=cu129 ;; \
-      *)    TORCH_VER=2.13.0; CU_TAG=cu126 ;; \
+      *) echo "Unsupported CUDA version: ${CUDA_VERSION}" && exit 1 ;; \
     esac; \
     ${PYTHON_ROOT_PATH}/bin/pip install torch==${TORCH_VER} --index-url ${PYTORCH_INDEX_BASE}/${CU_TAG}; \
     ${PYTHON_ROOT_PATH}/bin/pip install ninja setuptools==75.0.0 wheel==0.41.0 numpy uv scikit-build-core --index-url ${PIP_DEFAULT_INDEX}
diff --git a/python/sglang/kernels/aot/python/sgl_kernel/load_utils.py b/python/sglang/kernels/aot/python/sgl_kernel/load_utils.py
index f55cc1ab0..6b2239529 100644
--- a/python/sglang/kernels/aot/python/sgl_kernel/load_utils.py
+++ b/python/sglang/kernels/aot/python/sgl_kernel/load_utils.py
@@ -167,12 +167,7 @@ def _load_architecture_specific_ops():
 
     # All attempts failed
     cuda_version = torch.version.cuda
-    if cuda_version and cuda_version.startswith("12"):
-        install_hint = (
-            "pip install sglang-kernel --index-url https://docs.sglang.ai/whl/cu129/"
-        )
-    else:
-        install_hint = "pip install --upgrade sglang-kernel"
+    install_hint = "pip install --upgrade sglang-kernel"
 
     error_msg = f"""
 [sgl_kernel] CRITICAL: Could not load any common_ops library!
diff --git a/python/sglang/kernels/aot/rename_wheels.sh b/python/sglang/kernels/aot/rename_wheels.sh
index 76a513679..d5772b34d 100755
--- a/python/sglang/kernels/aot/rename_wheels.sh
+++ b/python/sglang/kernels/aot/rename_wheels.sh
@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
-# Align CUDA wheel filenames (+cu124/+cu129/+cu130) with internal METADATA Version and
+# Align CUDA wheel filenames (+cu130) with internal METADATA Version and
 # WHEEL tags after build (fixes pip "inconsistent version" when only the .whl name changed).
 # Unpack → patch WHEEL/METADATA → wheel pack (RECORD regenerated; no hand-editing).
 set -ex
@@ -7,11 +7,7 @@ set -ex
 WHEEL_DIR="dist"
 
 detect_cuda_suffix() {
-    if ls /usr/local/ 2>/dev/null | grep -q "12.4"; then
-        echo "+cu124"
-    elif ls /usr/local/ 2>/dev/null | grep -q "12.9"; then
-        echo "+cu129"
-    elif ls /usr/local/ 2>/dev/null | grep -q "13.0"; then
+    if ls /usr/local/ 2>/dev/null | grep -q "13.0"; then
         echo "+cu130"
     else
         echo ""
diff --git a/python/sglang/kernels/jit/utils/arch.py b/python/sglang/kernels/jit/utils/arch.py
index 9c68aacff..ade8d4927 100644
--- a/python/sglang/kernels/jit/utils/arch.py
+++ b/python/sglang/kernels/jit/utils/arch.py
@@ -42,7 +42,7 @@ def _jit_cuda_version() -> tuple[int, ...]:
     """CUDA version of the nvcc that JIT builds actually run.
 
     The target has to match the compiler, not the toolkit PyTorch was built
-    against: a cu129 wheel on a CUDA 12.8 toolkit would otherwise select
+    against: a cu130 wheel on a CUDA 12.8 toolkit would otherwise select
     `sm_120f`, which nvcc 12.8 rejects. Resolve nvcc the way tvm-ffi does
     (`CUDA_HOME` / `CUDA_PATH`, then `$PATH`, then `/usr/local/cuda`) and fall
     back to `torch.version.cuda` when it cannot be probed.
diff --git a/scripts/build_sgl_deep_gemm.sh b/scripts/build_sgl_deep_gemm.sh
index a042cfe0a..834d2efdb 100755
--- a/scripts/build_sgl_deep_gemm.sh
+++ b/scripts/build_sgl_deep_gemm.sh
@@ -3,12 +3,12 @@
 #
 # Usage: build_sgl_deep_gemm.sh <PYTHON_VERSION> <CUDA_VERSION> <DEEPGEMM_SRC> [ARCH]
 #   PYTHON_VERSION: e.g. 3.10
-#   CUDA_VERSION:   e.g. 12.9 or 13.0
+#   CUDA_VERSION:   e.g. 13.0
 #   DEEPGEMM_SRC:   path to a checkout of sgl-project/DeepGEMM
 #   ARCH:           x86_64 (default) or aarch64
 #
 # Writes:
-#   <DEEPGEMM_SRC>/dist/      — wheel(s) tagged +cu129 / +cu130 and manylinux
+#   <DEEPGEMM_SRC>/dist/      — wheel(s) tagged +cu130 and manylinux
 #   <DEEPGEMM_SRC>/dist-pypi/ — cu130 only: same wheel(s) with +cu130 stripped
 #                              (PyPI rejects local-version segments)
 set -ex
@@ -25,7 +25,6 @@ ARCH="${4:-$(uname -i)}"
 
 case "${CUDA_VERSION}" in
   13.0) CU_TAG=cu130 ;;
-  12.9) CU_TAG=cu129 ;;
   *)
     echo "Unsupported CUDA_VERSION: ${CUDA_VERSION}" >&2
     exit 1
diff --git a/scripts/build_sgl_deepep.sh b/scripts/build_sgl_deepep.sh
index b1fc31c57..eb76f14a9 100755
--- a/scripts/build_sgl_deepep.sh
+++ b/scripts/build_sgl_deepep.sh
@@ -4,8 +4,8 @@
 # Usage:
 #   build_sgl_deepep.sh <python-version> <cuda-version> <deepep-source> <packaging-overlay> [architecture]
 #
-# Writes CUDA-tagged wheels to <deepep-source>/dist. CUDA 13 builds also write
-# PyPI-ready wheels without the local CUDA version to <deepep-source>/dist-pypi.
+# Writes CUDA-tagged wheels to <deepep-source>/dist, plus PyPI-ready wheels
+# without the local CUDA version in <deepep-source>/dist-pypi.
 
 set -euo pipefail
 
@@ -14,7 +14,7 @@ usage() {
 Usage: build_sgl_deepep.sh <python-version> <cuda-version> <deepep-source> <packaging-overlay> [architecture]
 
   python-version:     3.10, 3.11, 3.12, or 3.13
-  cuda-version:       12.9 or 13.0
+  cuda-version:       13.0
   deepep-source:      checkout of the selected DeepEP implementation branch
   packaging-overlay: path to the shared DeepEP sgl_deep_ep directory
   architecture:       x86_64 or aarch64 (defaults to the current machine)
@@ -45,9 +45,6 @@ case "${PYTHON_VERSION}" in
 esac
 
 case "${CUDA_VERSION}" in
-    12.9)
-        CUDA_TAG=cu129
-        ;;
     13.0)
         CUDA_TAG=cu130
         ;;
diff --git a/scripts/ci/cuda/ci_install_dependency.sh b/scripts/ci/cuda/ci_install_dependency.sh
index c1afa9ed5..d69f3a253 100755
--- a/scripts/ci/cuda/ci_install_dependency.sh
+++ b/scripts/ci/cuda/ci_install_dependency.sh
@@ -2,7 +2,7 @@
 # Install dependencies for CUDA CI jobs.
 #
 # CU_VERSION (default: cu130) controls PyTorch index URL, FlashInfer JIT cache
-# index, and nvrtc variant selection.
+# index, and the sglang wheel index. CUDA 13 only.
 set -euxo pipefail
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
@@ -32,10 +32,14 @@ mark_step_done() {
 
 configure_environment() {
     # CU_VERSION controls PyTorch index URL, FlashInfer JIT cache index, and
-    # nvrtc variant selection (cu12 vs cu13).
+    # the sglang wheel index. Only CUDA 13 lanes exist: PyTorch 2.14 publishes
+    # no CUDA 12 wheels for the cu129 index the retired cu12 lane used.
     CU_VERSION="${CU_VERSION:-cu130}"
     CU_STRIP="${CU_VERSION#cu}"
-    CU_MAJOR="${CU_STRIP:0:2}"
+    case "${CU_STRIP}" in
+        13*) ;;
+        *) echo "FATAL: unsupported CU_VERSION=${CU_VERSION}; only CUDA 13 is supported"; exit 1 ;;
+    esac
 
     OPTIONAL_DEPS="${1:-}"
 
@@ -366,11 +370,6 @@ remove_stale_cuda12_nvidia_wheels() {
     local -a NVIDIA_WHEELS_TO_RESTORE=()
     local -a STALE_CUDA12_NVIDIA_WHEELS=()
 
-    if [ "$CU_MAJOR" != "13" ]; then
-        mark_step_done "${FUNCNAME[0]}"
-        return
-    fi
-
     mapfile -t INSTALLED_NVIDIA_WHEELS < <(
         python3 -m pip list --format=freeze | sed -n '/^nvidia-.*==/p'
     )
@@ -462,30 +461,6 @@ install_pytorch_stack() {
     mark_step_done "${FUNCNAME[0]}"
 }
 
-install_cuda12_deepep_wheel() {
-    if [ "$CU_MAJOR" = "13" ]; then
-        echo "CUDA 13 uses the public sgl-deep-ep wheel declared in python/pyproject.toml"
-        mark_step_done "${FUNCNAME[0]}"
-        return
-    fi
-
-    local version
-    version=$(grep -Po -m1 '"sgl-deep-ep==\K[^"]+' python/pyproject.toml || true)
-    if [ -z "$version" ]; then
-        echo "ERROR: python/pyproject.toml must pin sgl-deep-ep"
-        exit 1
-    fi
-
-    # CUDA 12 wheels intentionally live only on the SGLang wheel index. Their
-    # local version satisfies the public-version pyproject pin, so the later
-    # editable SGLang install keeps this CUDA-matched wheel.
-    $PIP_CMD install "sgl-deep-ep==${version}+${CU_VERSION}" \
-        --index-url "https://docs.sglang.ai/whl/${CU_VERSION}/" \
-        --force-reinstall --no-deps $PIP_INSTALL_SUFFIX
-
-    mark_step_done "${FUNCNAME[0]}"
-}
-
 require_prebuilt_rust_exts() {
     # Stages whose download succeeded set this to none. Runs before
     # setup_pip_toolchain uninstalls sglang, so clearing it here still reaches
@@ -551,14 +526,10 @@ install_sglang() {
 }
 
 install_nccl() {
-    if [ "$CU_MAJOR" = "13" ]; then
-        # PyTorch pins 2.29.7, so this override must run after every command
-        # that resolves Python dependencies (including lmms-eval).
-        $PIP_CMD install "nvidia-nccl-cu13==2.30.7" \
-            --force-reinstall --no-deps $PIP_INSTALL_SUFFIX
-    else
-        echo "CUDA ${CU_MAJOR} does not require the NCCL Gin wheel"
-    fi
+    # PyTorch pins 2.29.7, so this override must run after every command
+    # that resolves Python dependencies (including lmms-eval).
+    $PIP_CMD install "nvidia-nccl-cu13==2.30.7" \
+        --force-reinstall --no-deps $PIP_INSTALL_SUFFIX
 
     mark_step_done "${FUNCNAME[0]}"
 }
@@ -628,8 +599,8 @@ install_sglang_kernel() {
 
     if [ "${CUSTOM_BUILD_SGL_KERNEL:-}" != "true" ]; then
         # The PyPI default wheel tracks one CUDA version (currently cu130); other
-        # runners (e.g. h20 / cu129) need the +${CU_VERSION}-tagged wheel from the
-        # sglang index, linked against the right libnvrtc.
+        # runners need the +${CU_VERSION}-tagged wheel from the sglang index,
+        # linked against the right libnvrtc.
         SGL_KERNEL_WANTED="${SGL_KERNEL_VERSION_FROM_SRT}+${CU_VERSION}"
         if installed_wheel_ok sglang-kernel "${SGL_KERNEL_WANTED}" reject-local; then
             echo "sglang-kernel==${SGL_KERNEL_WANTED} already installed, keeping it"
@@ -640,18 +611,11 @@ install_sglang_kernel() {
         echo "CUSTOM_BUILD_SGL_KERNEL=true: keeping freshly built sgl-kernel wheel."
     fi
     SGL_DEEP_GEMM_VERSION=$(grep -Po -m1 '(?<=sgl-deep-gemm==)[0-9A-Za-z\.\-]+' python/pyproject.toml)
-    if [ "$CU_MAJOR" = "13" ]; then
-        SGL_DEEP_GEMM_WANTED="${SGL_DEEP_GEMM_VERSION}"
-    else
-        SGL_DEEP_GEMM_WANTED="${SGL_DEEP_GEMM_VERSION}+cu129"
-    fi
     # No reject-local: nothing builds sgl-deep-gemm locally.
-    if installed_wheel_ok sgl-deep-gemm "${SGL_DEEP_GEMM_WANTED}"; then
-        echo "sgl-deep-gemm==${SGL_DEEP_GEMM_WANTED} already installed, keeping it"
-    elif [ "$CU_MAJOR" = "13" ]; then
-        $PIP_CMD install "sgl-deep-gemm==${SGL_DEEP_GEMM_VERSION}" --force-reinstall $PIP_INSTALL_SUFFIX
+    if installed_wheel_ok sgl-deep-gemm "${SGL_DEEP_GEMM_VERSION}"; then
+        echo "sgl-deep-gemm==${SGL_DEEP_GEMM_VERSION} already installed, keeping it"
     else
-        $PIP_CMD install "https://github.com/sgl-project/whl/releases/download/v${SGL_DEEP_GEMM_VERSION}/sgl_deep_gemm-${SGL_DEEP_GEMM_VERSION}+cu129-py3-none-manylinux2014_$(uname -m).whl" --force-reinstall $PIP_INSTALL_SUFFIX
+        $PIP_CMD install "sgl-deep-gemm==${SGL_DEEP_GEMM_VERSION}" --force-reinstall $PIP_INSTALL_SUFFIX
     fi
 
     mark_step_done "${FUNCNAME[0]}"
@@ -739,17 +703,10 @@ stabilize_flashinfer_jit_paths() {
 install_extra_deps() {
     MOONCAKE_VERSION="0.3.13"
     NIXL_VERSION="1.3.0"
-    if [ "$CU_MAJOR" = "13" ]; then
-        MOONCAKE_PKG="mooncake-transfer-engine-cuda13==${MOONCAKE_VERSION}"
-        MOONCAKE_STALE_PKG="mooncake-transfer-engine"
-        NIXL_BIN_NAME="nixl-cu13"
-        EXTRA_NVIDIA_SPECS="nvidia-cuda-nvrtc"
-    else
-        MOONCAKE_PKG="mooncake-transfer-engine==${MOONCAKE_VERSION}"
-        MOONCAKE_STALE_PKG="mooncake-transfer-engine-cuda13"
-        NIXL_BIN_NAME="nixl-cu12"
-        EXTRA_NVIDIA_SPECS="nvidia-cuda-nvrtc-cu12"
-    fi
+    MOONCAKE_PKG="mooncake-transfer-engine-cuda13==${MOONCAKE_VERSION}"
+    MOONCAKE_STALE_PKG="mooncake-transfer-engine"
+    NIXL_BIN_NAME="nixl-cu13"
+    EXTRA_NVIDIA_SPECS="nvidia-cuda-nvrtc"
     # Both variants own the same mooncake/ package files and bin/ scripts
     # (mooncake_master, etc.). Uninstalling the stale variant deletes shared
     # files that the live variant's RECORD still references, so we force a
@@ -834,20 +791,18 @@ verify_imports() {
 import ctypes
 import importlib.metadata
 import os
-import sys
 
-if sys.argv[1] == "13":
-    if importlib.metadata.version("nvidia-nccl-cu13") != "2.30.7":
-        raise SystemExit("nvidia-nccl-cu13 was changed after the final CI override")
-    nccl = ctypes.CDLL("libnccl.so.2")
-    nccl_version = ctypes.c_int()
-    status = nccl.ncclGetVersion(ctypes.byref(nccl_version))
-    if status != 0 or nccl_version.value != 23007:
-        raise SystemExit(
-            f"expected NCCL runtime 2.30.7, got status={status}, "
-            f"raw_version={nccl_version.value}"
-        )
-    print("NCCL package and runtime versions are 2.30.7")
+if importlib.metadata.version("nvidia-nccl-cu13") != "2.30.7":
+    raise SystemExit("nvidia-nccl-cu13 was changed after the final CI override")
+nccl = ctypes.CDLL("libnccl.so.2")
+nccl_version = ctypes.c_int()
+status = nccl.ncclGetVersion(ctypes.byref(nccl_version))
+if status != 0 or nccl_version.value != 23007:
+    raise SystemExit(
+        f"expected NCCL runtime 2.30.7, got status={status}, "
+        f"raw_version={nccl_version.value}"
+    )
+print("NCCL package and runtime versions are 2.30.7")
 
 import torch
 print(torch.version.cuda)
@@ -882,7 +837,7 @@ for mod in ("server", "grpc", "multimodal"):
     except Exception as exc:
         raise SystemExit(f"{name} is present but does not load: {exc!r}")
     print(f"{name} loads")
-' "$CU_MAJOR"
+'
 
     mark_step_done "${FUNCNAME[0]}"
 }
@@ -904,7 +859,6 @@ main() {
     remove_stale_cuda12_nvidia_wheels
     uninstall_stale_flashinfer
     install_pytorch_stack
-    install_cuda12_deepep_wheel
     setup_cargo_cache
     install_sglang
     release_cargo_cache_lock
diff --git a/scripts/ci/utils/docker_build_metadata_args.py b/scripts/ci/utils/docker_build_metadata_args.py
index 79a41a656..ea83d6449 100644
--- a/scripts/ci/utils/docker_build_metadata_args.py
+++ b/scripts/ci/utils/docker_build_metadata_args.py
@@ -3,7 +3,7 @@ import datetime
 import json
 import sys
 
-MOVING_TAGS = {"dev", "dev-cu12", "dev-cu13", "latest"}
+MOVING_TAGS = {"dev", "dev-cu13", "latest"}
 
 
 def render_tag_template(tag: str, version: str, date: str, short_sha: str) -> str:
diff --git a/scripts/rename_sgl_deep_gemm_whl.sh b/scripts/rename_sgl_deep_gemm_whl.sh
index 7a91c1842..570675abf 100755
--- a/scripts/rename_sgl_deep_gemm_whl.sh
+++ b/scripts/rename_sgl_deep_gemm_whl.sh
@@ -8,7 +8,7 @@
 #
 # Usage: rename_wheels.sh <WHEEL_DIR> <CU_TAG> <ARCH>
 #   WHEEL_DIR: directory containing the *.whl file (e.g. DeepGEMM/dist)
-#   CU_TAG:    cu129 | cu130
+#   CU_TAG:    cu130
 #   ARCH:      x86_64 | aarch64
 set -ex
 
diff --git a/scripts/update_deepep_whl_index.py b/scripts/update_deepep_whl_index.py
index fd9b09720..c0c9a7c7d 100644
--- a/scripts/update_deepep_whl_index.py
+++ b/scripts/update_deepep_whl_index.py
@@ -5,7 +5,7 @@ import hashlib
 import pathlib
 import re
 
-SUPPORTED_CUDA_VERSIONS = ("129", "130")
+SUPPORTED_CUDA_VERSIONS = ("130",)
 WHEEL_PATTERN = re.compile(
     r"^sgl_deep_ep-(?P<version>[0-9][^-]*)-[^-]+-[^-]+-[^-]+\.whl$"
 )
diff --git a/scripts/update_deepgemm_whl_index.py b/scripts/update_deepgemm_whl_index.py
index 9be622a9f..f02dede82 100644
--- a/scripts/update_deepgemm_whl_index.py
+++ b/scripts/update_deepgemm_whl_index.py
@@ -1,14 +1,14 @@
 # Generates a PEP 503 simple index for sgl-deep-gemm wheels under
 # sgl-whl/cu<version>/sgl-deep-gemm/index.html. Mirrors the layout used by
 # update_kernel_whl_index.py so consumers can `pip install
-# sgl-deep-gemm --extra-index-url https://...whl/cu129`.
+# sgl-deep-gemm --extra-index-url https://...whl/cu130`.
 
 import argparse
 import hashlib
 import pathlib
 import re
 
-SUPPORTED_CUDA_VERSIONS = ["129", "130"]
+SUPPORTED_CUDA_VERSIONS = ["130"]
 
 
 def update_wheel_index(cuda_version, wheel_dir):
diff --git a/scripts/update_kernel_whl_index.py b/scripts/update_kernel_whl_index.py
index 3c89e70a7..46845479b 100644
--- a/scripts/update_kernel_whl_index.py
+++ b/scripts/update_kernel_whl_index.py
@@ -5,9 +5,9 @@ import hashlib
 import pathlib
 import re
 
-# All the CUDA versions that the wheels will cover
-SUPPORTED_CUDA_VERSIONS = ["129", "130"]
 DEFAULT_CUDA_VERSION = "130"
+# Local version a CUDA wheel carries, e.g. sglang_kernel-0.4.6.post1+cu130-...whl
+CUDA_LOCAL_VERSION_PATTERN = re.compile(r"\+cu(\d+)")
 
 
 def check_wheel_cuda_version(path_name, target_cuda_version):
@@ -19,16 +19,17 @@ def check_wheel_cuda_version(path_name, target_cuda_version):
     ):
         return False
 
-    # For other CUDA versions, the wheel path name will contain the cuda version suffix, e.g. sglang_kernel-0.4.0+cu130-cp310-abi3-manylinux2014_x86_64.whl
-    if target_cuda_version != DEFAULT_CUDA_VERSION:
-        return target_cuda_version in path_name
+    # Match on the wheel's own +cuNNN tag rather than a list of known versions,
+    # so a wheel built for a CUDA version this script has never heard of is
+    # rejected instead of landing in the target index.
+    match = CUDA_LOCAL_VERSION_PATTERN.search(path_name)
+    if match is not None:
+        return match.group(1) == target_cuda_version
 
-    # For the default CUDA version, the wheel path name will not contain any cuda version suffix, e.g. sglang_kernel-0.4.0-cp310-abi3-manylinux2014_x86_64.whl
-    # So we need to check if the wheel path name contains any other cuda version suffix
-    for cuda_version in SUPPORTED_CUDA_VERSIONS:
-        if cuda_version != DEFAULT_CUDA_VERSION and cuda_version in path_name:
-            return False
-    return True
+    # An untagged wheel is the default-CUDA build, e.g.
+    # sglang_kernel-0.4.0-cp310-abi3-manylinux2014_x86_64.whl (PyPI rejects
+    # local versions, so that upload strips the tag).
+    return target_cuda_version == DEFAULT_CUDA_VERSION
 
 
 def update_wheel_index(cuda_version=DEFAULT_CUDA_VERSION, rocm_version=None):
diff --git a/scripts/update_nightly_whl_index.py b/scripts/update_nightly_whl_index.py
index f79fda6a3..05d556690 100755
--- a/scripts/update_nightly_whl_index.py
+++ b/scripts/update_nightly_whl_index.py
@@ -35,7 +35,7 @@ def update_wheel_index(
     Args:
         commit_hash: Short git commit hash (e.g., 'c5f1e86')
         nightly_version: Full nightly version string (e.g., '0.5.6.post1.dev7716+gc5f1e86')
-        cuda_version: CUDA version string (e.g., '129' or '130')
+        cuda_version: CUDA version string (e.g., '130')
         build_date: Build date in YYYY-MM-DD format (e.g., '2025-12-13')
     """
     dist_dir = pathlib.Path("dist")
@@ -176,7 +176,7 @@ def main():
         "--cuda-version",
         type=str,
         default="130",
-        help="CUDA version (e.g., '129' or '130'). Defaults to '130'.",
+        help="CUDA version (e.g., '130'). Defaults to '130'.",
     )
     parser.add_argument(
         "--build-date",
diff --git a/test/registered/unit/tools/test_docker_build_metadata_args.py b/test/registered/unit/tools/test_docker_build_metadata_args.py
index 7aca2444b..92811a5ef 100644
--- a/test/registered/unit/tools/test_docker_build_metadata_args.py
+++ b/test/registered/unit/tools/test_docker_build_metadata_args.py
@@ -76,9 +76,9 @@ class TestDockerBuildMetadataArgs(unittest.TestCase):
 
     def test_release_metadata_prefers_versioned_tag(self):
         args = self.run_helper(
-            cuda="cu129",
+            cuda="cu134",
             tag_config=[
-                {"cuda": "cu129", "tags": ["v{version}", "latest"]},
+                {"cuda": "cu134", "tags": ["v{version}", "latest"]},
                 {"cuda": "cu130", "tags": ["v{version}-cu130", "latest-cu130"]},
             ],
         )
@@ -99,7 +99,7 @@ class TestDockerBuildMetadataArgs(unittest.TestCase):
             cuda="cu130",
             image_repo="lmsysorg/sglang-staging",
             tag_config=[
-                {"cuda": "cu129", "tags": ["v{version}-runtime", "latest-runtime"]},
+                {"cuda": "cu134", "tags": ["v{version}-runtime", "latest-runtime"]},
                 {
                     "cuda": "cu130",
                     "tags": ["v{version}-cu130-runtime", "latest-cu130-runtime"],
@@ -114,11 +114,11 @@ class TestDockerBuildMetadataArgs(unittest.TestCase):
 
     def test_dev_nightly_metadata_prefers_unique_tag_from_checked_out_commit(self):
         args = self.run_helper(
-            cuda="cu129",
+            cuda="cu134",
             version="",
             build_commit="1234567890abcdef",
             tag_config=[
-                {"cuda": "cu129", "tags": ["dev", "nightly-dev-{date}-{short_sha}"]},
+                {"cuda": "cu134", "tags": ["dev", "nightly-dev-{date}-{short_sha}"]},
                 {
                     "cuda": "cu130",
                     "tags": ["dev-cu13", "nightly-dev-cu13-{date}-{short_sha}"],
@@ -140,7 +140,7 @@ class TestDockerBuildMetadataArgs(unittest.TestCase):
             cuda="cu130",
             version="",
             tag_config=[
-                {"cuda": "cu129", "tags": ["dev-my-test"]},
+                {"cuda": "cu134", "tags": ["dev-my-test"]},
                 {"cuda": "cu130", "tags": ["dev-cu13-my-test"]},
             ],
         )
@@ -153,7 +153,7 @@ class TestDockerBuildMetadataArgs(unittest.TestCase):
     def test_missing_cuda_entry_fails(self):
         with self.assertRaisesRegex(ValueError, "cu130"):
             self.helper.select_tag(
-                json.dumps([{"cuda": "cu129", "tags": ["v{version}"]}]),
+                json.dumps([{"cuda": "cu134", "tags": ["v{version}"]}]),
                 "cu130",
                 "0.6.0",
                 "20260429",