diff --git a/.github/workflows/nightly-test-npu.yml b/.github/workflows/nightly-test-npu.yml
index 44071afc7..9b4253a76 100644
--- a/.github/workflows/nightly-test-npu.yml
+++ b/.github/workflows/nightly-test-npu.yml
@@ -63,7 +63,7 @@ jobs:
           fi
 
           if [ -z "${{ inputs.image_a3 }}" ]; then
-            echo "image_a3=swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-a3-ubuntu22.04-py3.11" >> $GITHUB_OUTPUT
+            echo "image_a3=swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:9.0.0-a3-ubuntu22.04-py3.11" >> $GITHUB_OUTPUT
           else
             echo "image_a3=${{ inputs.image_a3 }}" >> $GITHUB_OUTPUT
           fi
@@ -128,7 +128,7 @@ jobs:
         run: |
           pip install sglang_router
           hf download lmms-lab/MMMU --repo-type dataset
-          pip install sentence_transformers torchaudio==2.8.0
+          pip install sentence_transformers torchaudio==2.10.0
           pip install protobuf==6.31.1 zss pre-commit wandb>=0.16.0 tenacity==8.3.0 loguru openpyxl latex2sympy2 zstandard transformers-stream-generator tqdm-multiprocess pycocoevalcap
           pip install yt-dlp sentencepiece==0.1.99 nltk av ftfy sqlitedict==2.1.0 sacrebleu>=1.5.0 pytablewriter black==24.1.0 isort==5.13.2 peft>=0.2.0 accelerate>=0.29.1
           pip install jsonlines httpx==0.25.0 evaluate>=0.4.0 datasets==2.16.1 numexpr xgrammar==0.2.0 numpy==1.26.4 dotenv
@@ -194,7 +194,7 @@ jobs:
         run: |
           pip install sglang_router
           hf download lmms-lab/MMMU --repo-type dataset
-          pip install sentence_transformers torchaudio==2.8.0
+          pip install sentence_transformers torchaudio==2.10.0
           pip install protobuf==6.31.1 zss pre-commit wandb>=0.16.0 tenacity==8.3.0 loguru openpyxl latex2sympy2 zstandard transformers-stream-generator tqdm-multiprocess pycocoevalcap
           pip install yt-dlp sentencepiece==0.1.99 nltk av ftfy sqlitedict==2.1.0 sacrebleu>=1.5.0 pytablewriter black==24.1.0 isort==5.13.2 peft>=0.2.0 accelerate>=0.29.1
           pip install jsonlines httpx==0.25.0 evaluate>=0.4.0 datasets==2.16.1 numexpr xgrammar==0.2.0 numpy==1.26.4 dotenv
@@ -261,7 +261,7 @@ jobs:
         run: |
           pip install sglang_router
           hf download lmms-lab/MMMU --repo-type dataset
-          pip install sentence_transformers torchaudio==2.8.0
+          pip install sentence_transformers torchaudio==2.10.0
           pip install protobuf==6.31.1 zss pre-commit wandb>=0.16.0 tenacity==8.3.0 loguru openpyxl latex2sympy2 zstandard transformers-stream-generator tqdm-multiprocess pycocoevalcap
           pip install yt-dlp sentencepiece==0.1.99 nltk av ftfy sqlitedict==2.1.0 sacrebleu>=1.5.0 pytablewriter black==24.1.0 isort==5.13.2 peft>=0.2.0 accelerate>=0.29.1
           pip install jsonlines httpx==0.25.0 evaluate>=0.4.0 datasets==2.16.1 numexpr xgrammar==0.2.0 numpy==1.26.4 dotenv
@@ -328,7 +328,7 @@ jobs:
         run: |
           pip install sglang_router
           hf download lmms-lab/MMMU --repo-type dataset
-          pip install sentence_transformers torchaudio==2.8.0
+          pip install sentence_transformers torchaudio==2.10.0
           pip install protobuf==6.31.1 zss pre-commit wandb>=0.16.0 tenacity==8.3.0 loguru openpyxl latex2sympy2 zstandard transformers-stream-generator tqdm-multiprocess pycocoevalcap
           pip install yt-dlp sentencepiece==0.1.99 nltk av ftfy sqlitedict==2.1.0 sacrebleu>=1.5.0 pytablewriter black==24.1.0 isort==5.13.2 peft>=0.2.0 accelerate>=0.29.1
           pip install jsonlines httpx==0.25.0 evaluate>=0.4.0 datasets==2.16.1 numexpr xgrammar==0.2.0 numpy==1.26.4 dotenv
@@ -395,7 +395,7 @@ jobs:
         run: |
           pip install sglang_router
           hf download lmms-lab/MMMU --repo-type dataset
-          pip install sentence_transformers torchaudio==2.8.0
+          pip install sentence_transformers torchaudio==2.10.0
           pip install protobuf==6.31.1 zss pre-commit wandb>=0.16.0 tenacity==8.3.0 loguru openpyxl latex2sympy2 zstandard transformers-stream-generator tqdm-multiprocess pycocoevalcap
           pip install yt-dlp sentencepiece==0.1.99 nltk av ftfy sqlitedict==2.1.0 sacrebleu>=1.5.0 pytablewriter black==24.1.0 isort==5.13.2 peft>=0.2.0 accelerate>=0.29.1
           pip install jsonlines httpx==0.25.0 evaluate>=0.4.0 datasets==2.16.1 numexpr xgrammar==0.2.0 numpy==1.26.4 dotenv
diff --git a/.github/workflows/pr-test-npu.yml b/.github/workflows/pr-test-npu.yml
index 0a1b29608..2edd382ec 100644
--- a/.github/workflows/pr-test-npu.yml
+++ b/.github/workflows/pr-test-npu.yml
@@ -77,8 +77,21 @@ jobs:
     uses: ./.github/workflows/pr-gate.yml
     secrets: inherit
 
+  set-image-config:
+    runs-on: ubuntu-latest
+    outputs:
+      CANN_image_a3: ${{ steps.set-vars.outputs.CANN_image_a3 }}
+      CANN_image_910b: ${{ steps.set-vars.outputs.CANN_image_910b }}
+    steps:
+      # When triggered by PR, no inputs parameters are used. The latest community code is tested by default.
+      - name: Set image config
+        id: set-vars
+        run: |
+          echo "CANN_image_a3=swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:9.0.0-a3-ubuntu22.04-py3.11" >> $GITHUB_OUTPUT
+          echo "CANN_image_910b=swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:9.0.0-910b-ubuntu22.04-py3.11" >> $GITHUB_OUTPUT
+
   stage-b-test-1-npu-a2:
-    needs: [check-changes, pr-gate]
+    needs: [check-changes, pr-gate, set-image-config]
     if: needs.check-changes.outputs.main_package == 'true'
     runs-on: linux-aarch64-a2-1
     strategy:
@@ -86,7 +99,7 @@ jobs:
       matrix:
         part: [ 0, 1 ]
     container:
-      image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-910b-ubuntu22.04-py3.11
+      image: ${{ needs.set-image-config.outputs.CANN_image_910b }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -131,7 +144,7 @@ jobs:
           python3 run_suite.py --hw npu --suite stage-b-test-1-npu-a2 --auto-partition-id ${{ matrix.part }} --auto-partition-size 2
 
   stage-b-test-2-npu-a2:
-    needs: [check-changes, pr-gate]
+    needs: [check-changes, pr-gate, set-image-config]
     if: needs.check-changes.outputs.main_package == 'true'
     runs-on: linux-aarch64-a2-2
     strategy:
@@ -139,7 +152,7 @@ jobs:
       matrix:
         part: [0, 1]
     container:
-      image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-910b-ubuntu22.04-py3.11
+      image: $${{ needs.set-image-config.outputs.CANN_image_910b }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -184,11 +197,11 @@ jobs:
           python3 run_suite.py --hw npu --suite stage-b-test-2-npu-a2 --auto-partition-id ${{ matrix.part }} --auto-partition-size 2
 
   stage-b-test-4-npu-a3:
-    needs: [check-changes, pr-gate]
+    needs: [check-changes, pr-gate, set-image-config]
     if: needs.check-changes.outputs.main_package == 'true'
     runs-on: linux-aarch64-a3-4
     container:
-      image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-a3-ubuntu22.04-py3.11
+      image: ${{ needs.set-image-config.outputs.CANN_image_a3 }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -234,11 +247,11 @@ jobs:
 
 
   stage-b-test-16-npu-a3:
-    needs: [check-changes, pr-gate]
+    needs: [check-changes, pr-gate, set-image-config]
     if: needs.check-changes.outputs.main_package == 'true'
     runs-on: linux-aarch64-a3-16
     container:
-      image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-a3-ubuntu22.04-py3.11
+      image: ${{ needs.set-image-config.outputs.CANN_image_a3 }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -283,11 +296,11 @@ jobs:
           python3 run_suite.py --hw npu --suite stage-b-test-16-npu-a3 --timeout-per-file 3600
 
   multimodal-gen-test-1-npu-a3:
-    needs: [check-changes, pr-gate]
+    needs: [check-changes, pr-gate, set-image-config]
     if: needs.check-changes.outputs.multimodal_gen == 'true'
     runs-on: linux-aarch64-a3-2
     container:
-      image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.3.rc2-a3-ubuntu22.04-py3.11
+      image: ${{ needs.set-image-config.outputs.CANN_image_a3 }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -310,7 +323,7 @@ jobs:
           pip config set global.index-url http://${CACHING_URL}/pypi/simple
           pip config set global.trusted-host "${CACHING_URL}"
 
-          bash scripts/ci/npu/npu_ci_install_dependency.sh a3 diffusion
+          bash scripts/ci/npu/npu_ci_install_dependency.sh a3
           # copy required file from our daily cache
           cp ~/.cache/modelscope/hub/datasets/otavia/ShareGPT_Vicuna_unfiltered/ShareGPT_V3_unfiltered_cleaned_split.json /tmp
           # copy gsm8k dataset
@@ -327,7 +340,6 @@ jobs:
           STREAMS_PER_DEVICE: 32
           SGLANG_DIFFUSION_ARTIFACT_DIR: ${{ github.workspace }}/diffusion-failures
         run: |
-          export PATH="/usr/local/Ascend/8.3.RC1/compiler/bishengir/bin:${PATH}"
           cd python
           python3 sglang/multimodal_gen/test/run_suite_npu.py --suite 1-npu
 
@@ -341,11 +353,11 @@ jobs:
           retention-days: 7
 
   multimodal-gen-test-2-npu-a3:
-    needs: [check-changes, pr-gate]
+    needs: [check-changes, pr-gate, set-image-config]
     if: needs.check-changes.outputs.multimodal_gen == 'true'
     runs-on: linux-aarch64-a3-16
     container:
-      image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.3.rc2-a3-ubuntu22.04-py3.11
+      image: ${{ needs.set-image-config.outputs.CANN_image_a3 }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -368,7 +380,7 @@ jobs:
           pip config set global.index-url http://${CACHING_URL}/pypi/simple
           pip config set global.trusted-host "${CACHING_URL}"
 
-          bash scripts/ci/npu/npu_ci_install_dependency.sh a3 diffusion
+          bash scripts/ci/npu/npu_ci_install_dependency.sh a3
           # copy required file from our daily cache
           cp ~/.cache/modelscope/hub/datasets/otavia/ShareGPT_Vicuna_unfiltered/ShareGPT_V3_unfiltered_cleaned_split.json /tmp
           # copy gsm8k dataset
@@ -385,7 +397,6 @@ jobs:
           STREAMS_PER_DEVICE: 32
           SGLANG_DIFFUSION_ARTIFACT_DIR: ${{ github.workspace }}/diffusion-failures
         run: |
-          export PATH="/usr/local/Ascend/8.3.RC1/compiler/bishengir/bin:${PATH}"
           cd python
           python3 sglang/multimodal_gen/test/run_suite_npu.py --suite 2-npu
 
@@ -399,11 +410,11 @@ jobs:
           retention-days: 7
 
   multimodal-gen-test-8-npu-a3:
-    needs: [check-changes, pr-gate]
+    needs: [check-changes, pr-gate, set-image-config]
     if: needs.check-changes.outputs.multimodal_gen == 'true'
     runs-on: linux-aarch64-a3-8
     container:
-      image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-a3-ubuntu22.04-py3.11
+      image: ${{ needs.set-image-config.outputs.CANN_image_a3 }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -426,7 +437,7 @@ jobs:
           pip config set global.index-url http://${CACHING_URL}/pypi/simple
           pip config set global.trusted-host "${CACHING_URL}"
 
-          bash scripts/ci/npu/npu_ci_install_dependency.sh a3 diffusion
+          bash scripts/ci/npu/npu_ci_install_dependency.sh a3
           # copy required file from our daily cache
           cp ~/.cache/modelscope/hub/datasets/otavia/ShareGPT_Vicuna_unfiltered/ShareGPT_V3_unfiltered_cleaned_split.json /tmp
           # copy gsm8k dataset
diff --git a/.github/workflows/release-docker-npu-nightly.yml b/.github/workflows/release-docker-npu-nightly.yml
index c0dfc0ad6..39ade16a3 100644
--- a/.github/workflows/release-docker-npu-nightly.yml
+++ b/.github/workflows/release-docker-npu-nightly.yml
@@ -19,7 +19,7 @@ jobs:
     runs-on: ubuntu-22.04-arm
     strategy:
       matrix:
-        cann_version: ["8.5.0"]
+        cann_version: ["9.0.0"]
         device_type: ["910b", "a3"]
     steps:
       - name: Checkout repository
@@ -80,6 +80,6 @@ jobs:
           push: ${{ github.repository == 'sgl-project/sglang' && github.event_name != 'pull_request' }}
           provenance: false
           build-args: |
-            SGLANG_KERNEL_NPU_TAG=2026.05.01
+            SGLANG_KERNEL_NPU_TAG=2026.05.01.post3
             CANN_VERSION=${{ matrix.cann_version }}
             DEVICE_TYPE=${{ matrix.device_type }}
diff --git a/.github/workflows/release-docker-npu.yml b/.github/workflows/release-docker-npu.yml
index 03019407d..30c3a06b3 100644
--- a/.github/workflows/release-docker-npu.yml
+++ b/.github/workflows/release-docker-npu.yml
@@ -14,7 +14,7 @@ jobs:
     runs-on: ubuntu-22.04-arm
     strategy:
       matrix:
-        cann_version: ["8.5.0"]
+        cann_version: ["9.0.0"]
         device_type: ["910b", "a3"]
     steps:
       - name: Checkout repository
@@ -87,7 +87,7 @@ jobs:
           push: ${{ github.repository == 'sgl-project/sglang' && github.event_name != 'pull_request' }}
           provenance: false
           build-args: |
-            SGLANG_KERNEL_NPU_TAG=2026.05.01
+            SGLANG_KERNEL_NPU_TAG=2026.05.01.post3
             CANN_VERSION=${{ matrix.cann_version }}
             DEVICE_TYPE=${{ matrix.device_type }}
             SGLANG_TAG=${{ steps.version.outputs.version }}
diff --git a/docker/npu.Dockerfile b/docker/npu.Dockerfile
index 44a8cbaf3..c0dfc55f9 100644
--- a/docker/npu.Dockerfile
+++ b/docker/npu.Dockerfile
@@ -11,11 +11,13 @@ ARG CANN_VERSION
 ARG DEVICE_TYPE
 ARG PIP_INDEX_URL="https://pypi.org/simple/"
 ARG APTMIRROR=""
-ARG PYTORCH_VERSION="2.8.0"
-ARG TORCHVISION_VERSION="0.23.0"
-ARG TORCHAUDIO_VERSION="2.8.0"
-ARG PTA_URL_ARM64="https://gitcode.com/Ascend/pytorch/releases/download/v7.3.0-pytorch2.8.0/torch_npu-2.8.0.post2-cp311-cp311-manylinux_2_28_aarch64.whl"
-ARG PTA_URL_AMD64="https://gitcode.com/Ascend/pytorch/releases/download/v7.3.0-pytorch2.8.0/torch_npu-2.8.0.post2-cp311-cp311-manylinux_2_28_x86_64.whl"
+ARG PYTORCH_VERSION="2.10.0"
+ARG TORCHVISION_VERSION="0.25.0"
+ARG TORCHAUDIO_VERSION="2.10.0"
+ARG PTA_URL_ARM64="https://gitcode.com/Ascend/pytorch/releases/download/v26.0.0-pytorch2.10.0/torch_npu-2.10.0-cp311-cp311-manylinux_2_28_aarch64.whl"
+ARG PTA_URL_AMD64="https://gitcode.com/Ascend/pytorch/releases/download/v26.0.0-pytorch2.10.0/torch_npu-2.10.0-cp311-cp311-manylinux_2_28_x86_64.whl"
+ARG TRITON_URL_ARM64="https://gitcode.com/Ascend/triton-ascend/releases/download/v3.2.1/triton_ascend-3.2.1-cp311-cp311-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl"
+ARG TRITON_URL_AMD64="https://gitcode.com/Ascend/triton-ascend/releases/download/v3.2.1/triton_ascend-3.2.1-cp311-cp311-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl"
 ARG SGLANG_TAG=main
 ARG ASCEND_CANN_PATH=/usr/local/Ascend/ascend-toolkit
 ARG SGLANG_KERNEL_NPU_TAG=main
@@ -26,9 +28,11 @@ ARG DEVICE_TYPE
 RUN if [ "$TARGETARCH" = "amd64" ]; then \
       echo "Using x86_64 dependencies"; \
       echo "PTA_URL=$PTA_URL_AMD64" >> /etc/environment_new; \
+      echo "TRITON_URL=$TRITON_URL_AMD64" >> /etc/environment_new; \
     elif [ "$TARGETARCH" = "arm64" ]; then \
       echo "Using aarch64 dependencies"; \
       echo "PTA_URL=$PTA_URL_ARM64" >> /etc/environment_new; \
+      echo "TRITON_URL=$TRITON_URL_ARM64" >> /etc/environment_new; \
     else \
       echo "Unsupported TARGETARCH: $TARGETARCH"; exit 1; \
     fi
@@ -72,7 +76,13 @@ ENV LC_ALL=en_US.UTF-8
 
 
 ### Install MemFabric
-RUN ${PIP_INSTALL} memfabric-hybrid==1.0.5
+RUN ${PIP_INSTALL} memfabric-hybrid==1.0.8
+
+### Install zbal
+RUN if [ "$TARGETARCH" = "arm64" ]; then \
+      ${PIP_INSTALL} memfabric-zbal==1.1.1; \
+    fi
+
 ### Install SGLang Model Gateway
 RUN ${PIP_INSTALL} sglang-router
 
@@ -84,7 +94,9 @@ RUN . /etc/environment_new && \
 
 
 ## Install triton-ascend
-RUN (${PIP_INSTALL} pybind11 triton-ascend)
+RUN . /etc/environment_new && \
+    (${PIP_INSTALL} pybind11) && \
+    (${PIP_INSTALL} ${TRITON_URL})
 
 # Install SGLang (editable mode to preserve source and git history)
 RUN git clone https://github.com/sgl-project/sglang --branch $SGLANG_TAG /sgl-workspace/sglang && \
@@ -96,8 +108,8 @@ RUN git clone https://github.com/sgl-project/sglang --branch $SGLANG_TAG /sgl-wo
 RUN ${PIP_INSTALL} wheel==0.45.1 pybind11 pyyaml decorator scipy attrs psutil \
     && mkdir sgl-kernel-npu \
     && cd sgl-kernel-npu \
-    && wget https://github.com/sgl-project/sgl-kernel-npu/releases/download/${SGLANG_KERNEL_NPU_TAG}/sgl-kernel-npu-${SGLANG_KERNEL_NPU_TAG}-torch2.8.0-py311-cann${CANN_VERSION}-${DEVICE_TYPE}-$(arch).zip \
-    && unzip sgl-kernel-npu-${SGLANG_KERNEL_NPU_TAG}-torch2.8.0-py311-cann${CANN_VERSION}-${DEVICE_TYPE}-$(arch).zip \
+    && wget https://github.com/sgl-project/sgl-kernel-npu/releases/download/${SGLANG_KERNEL_NPU_TAG}/sgl-kernel-npu-${SGLANG_KERNEL_NPU_TAG}-torch2.10.0-py311-cann${CANN_VERSION}-${DEVICE_TYPE}-$(arch).zip \
+    && unzip sgl-kernel-npu-${SGLANG_KERNEL_NPU_TAG}-torch2.10.0-py311-cann${CANN_VERSION}-${DEVICE_TYPE}-$(arch).zip \
     && ${PIP_INSTALL} deep_ep*.whl sgl_kernel_npu*.whl \
     && cd .. && rm -rf sgl-kernel-npu \
     && cd "$(python3 -m pip show deep-ep | awk '/^Location:/ {print $2}')" && ln -sf deep_ep/deep_ep_cpp*.so
diff --git a/scripts/ci/npu/npu_ci_install_dependency.sh b/scripts/ci/npu/npu_ci_install_dependency.sh
index 83434b9ce..12d4f1521 100755
--- a/scripts/ci/npu/npu_ci_install_dependency.sh
+++ b/scripts/ci/npu/npu_ci_install_dependency.sh
@@ -4,7 +4,6 @@ set -euo pipefail
 PIP_INSTALL="python3 -m pip install --no-cache-dir"
 UV_PIP_INSTALL="uv pip install "
 DEVICE_TYPE=$1
-OPTIONAL_DEPS="${2:-}"
 
 
 # Install the required dependencies in CI.
@@ -43,36 +42,30 @@ ${UV_PIP_INSTALL} wheel==0.45.1 pybind11 pyyaml decorator scipy attrs psutil
 
 
 ### Install MemFabric
-${UV_PIP_INSTALL} memfabric-hybrid==1.0.5
+${UV_PIP_INSTALL} memfabric-hybrid==1.0.8
 
 
 ### Install PyTorch and PTA
-if [ -n "$OPTIONAL_DEPS" ]; then
-    PYTORCH_VERSION="2.10.0"
-    TORCHVISION_VERSION="0.25.0"
-    ${UV_PIP_INSTALL} torch==${PYTORCH_VERSION} torchvision==${TORCHVISION_VERSION} --index-url ${TORCH_CACHE_URL:="https://download.pytorch.org/whl/cpu"} --extra-index-url ${PYPI_CACHE_URL:="https://pypi.org/simple/"}
-    PTA_URL="https://gitcode.com/Ascend/pytorch/releases/download/7.3.0.alpha002/torch_npu-2.10.0rc2-cp311-cp311-manylinux_2_28_aarch64.whl"
-    # GitCode does not allow UV downloads.
-    ${PIP_INSTALL} ${PTA_URL}
-else
-    PYTORCH_VERSION="2.8.0"
-    TORCHVISION_VERSION="0.23.0"
-    TORCHAUDIO_VERSION="2.8.0"
-    ${UV_PIP_INSTALL} torch==${PYTORCH_VERSION} torchvision==${TORCHVISION_VERSION} torchaudio==${TORCHAUDIO_VERSION} --index-url ${TORCH_CACHE_URL:="https://download.pytorch.org/whl/cpu"} --extra-index-url ${PYPI_CACHE_URL:="https://pypi.org/simple/"}
-    PTA_URL="https://gitcode.com/Ascend/pytorch/releases/download/v7.3.0-pytorch2.8.0/torch_npu-2.8.0.post2-cp311-cp311-manylinux_2_28_aarch64.whl"
-    ${PIP_INSTALL} ${PTA_URL}
-fi
+PYTORCH_VERSION="2.10.0"
+TORCHVISION_VERSION="0.25.0"
+TORCHAUDIO_VERSION="2.10.0"
+${UV_PIP_INSTALL} torch==${PYTORCH_VERSION} torchvision==${TORCHVISION_VERSION} torchaudio==${TORCHAUDIO_VERSION} --index-url ${TORCH_CACHE_URL:="https://download.pytorch.org/whl/cpu"} --extra-index-url ${PYPI_CACHE_URL:="https://pypi.org/simple/"}
+PTA_URL="https://gitcode.com/Ascend/pytorch/releases/download/v26.0.0-pytorch2.10.0/torch_npu-2.10.0-cp311-cp311-manylinux_2_28_aarch64.whl"
+# GitCode does not allow UV downloads.
+${PIP_INSTALL} ${PTA_URL}
 
+### Install zbal
+${UV_PIP_INSTALL} memfabric-zbal==1.1.1
 
 ### Install Triton-Ascend
-${UV_PIP_INSTALL} triton-ascend
+${PIP_INSTALL} "https://gitcode.com/Ascend/triton-ascend/releases/download/v3.2.1/triton_ascend-3.2.1-cp311-cp311-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl"
 
 
 ### Install sgl-kernel-npu
-SGLANG_KERNEL_NPU_TAG="2026.05.01"
+SGLANG_KERNEL_NPU_TAG="2026.05.01.post3"
 mkdir sgl-kernel-npu
-(cd sgl-kernel-npu && wget "${GITHUB_PROXY_URL:=""}https://github.com/sgl-project/sgl-kernel-npu/releases/download/${SGLANG_KERNEL_NPU_TAG}/sgl-kernel-npu-${SGLANG_KERNEL_NPU_TAG}-torch${PYTORCH_VERSION}-py311-cann8.5.0-${DEVICE_TYPE}-$(arch).zip" \
-&& unzip ./sgl-kernel-npu-${SGLANG_KERNEL_NPU_TAG}-torch${PYTORCH_VERSION}-py311-cann8.5.0-${DEVICE_TYPE}-$(arch).zip \
+(cd sgl-kernel-npu && wget "${GITHUB_PROXY_URL:=""}https://github.com/sgl-project/sgl-kernel-npu/releases/download/${SGLANG_KERNEL_NPU_TAG}/sgl-kernel-npu-${SGLANG_KERNEL_NPU_TAG}-torch${PYTORCH_VERSION}-py311-cann9.0.0-${DEVICE_TYPE}-$(arch).zip" \
+&& unzip ./sgl-kernel-npu-${SGLANG_KERNEL_NPU_TAG}-torch${PYTORCH_VERSION}-py311-cann9.0.0-${DEVICE_TYPE}-$(arch).zip \
 && ${UV_PIP_INSTALL} ./deep_ep*.whl ./sgl_kernel_npu*.whl \
 && (cd "$(python3 -m pip show deep-ep | grep -E '^Location:' | awk '{print $2}')" && ln -s deep_ep/deep_ep_cpp*.so))