diff --git a/.github/workflows/build-and-publish.yml b/.github/workflows/build-and-publish.yml
new file mode 100644
index 0000000..aa09692
--- /dev/null
+++ b/.github/workflows/build-and-publish.yml
@@ -0,0 +1,98 @@
+name: build-and-publish
+
+on:
+  workflow_dispatch:
+    inputs:
+      tag:
+        description: "Image tag to publish (e.g. latest)"
+        required: true
+        default: "latest"
+
+env:
+  IMAGE_REPO: kyuz0/vllm-therock-gfx1151
+  DOCKER_BUILDKIT: "1"
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Put Docker on /mnt
+        run: |
+          set -eux
+          echo '{ "data-root": "/mnt/docker" }' | sudo tee /etc/docker/daemon.json
+          sudo systemctl stop docker
+          sudo rm -rf /var/lib/docker || true
+          sudo mkdir -p /mnt/docker
+          sudo systemctl start docker
+          docker info | grep "Docker Root Dir"
+
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Free disk space
+        shell: bash
+        run: |
+          set -euxo pipefail
+          echo "Disk BEFORE:"; df -h
+          sudo rm -rf /usr/local/lib/android /usr/share/dotnet /opt/ghc || true
+          sudo rm -rf /opt/hostedtoolcache/CodeQL /opt/hostedtoolcache/go || true
+          docker system prune -af || true
+          docker builder prune -af || true
+          sudo apt-get clean
+          sudo rm -rf /var/lib/apt/lists/*
+          sudo rm -rf /opt/hostedtoolcache
+          echo "Disk AFTER:"; df -h
+
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v3
+
+      - name: BuildKit GC config (8GB cap)
+        run: |
+          cat > /tmp/buildkitd.toml <<'EOF'
+          [worker.oci]
+            gc = true
+            gckeepstorage = 8000   # MB
+          EOF
+
+      - name: Set up Buildx (with GC)
+        uses: docker/setup-buildx-action@v3
+        with:
+          buildkitd-flags: --config /tmp/buildkitd.toml
+
+      - name: Log in to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: docker.io/${{ env.IMAGE_REPO }}
+          tags: |
+            type=raw,value=${{ github.event.inputs.tag }}
+            type=sha
+            type=raw,value={{date 'YYYYMMDD-HHmmss'}}
+          labels: |
+            org.opencontainers.image.source=https://github.com/${{ github.repository }}
+            org.opencontainers.image.revision=${{ github.sha }}
+
+      - name: Build and push
+        id: build
+        uses: docker/build-push-action@v6
+        with:
+          context: .
+          file: ./Dockerfile
+          platforms: linux/amd64
+          push: true
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
+          provenance: false
+          sbom: false
+          no-cache: true
+
+      - name: Prune buildx cache
+        if: always()
+        run: |
+          docker buildx prune -af --verbose --min-free-space 4gb || true
diff --git a/Dockerfile b/Dockerfile
new file mode 100644
index 0000000..fcf1d20
--- /dev/null
+++ b/Dockerfile
@@ -0,0 +1,129 @@
+FROM registry.fedoraproject.org/fedora:43
+
+# 1. System Base & Build Tools
+# Added 'gperftools-libs' for tcmalloc (fixes double-free)
+RUN dnf -y install --setopt=install_weak_deps=False --nodocs \
+      python3.13 python3.13-devel git rsync libatomic bash ca-certificates curl \
+      gcc gcc-c++ binutils make ffmpeg-free \
+      cmake ninja-build aria2c tar xz vim nano \
+      libdrm-devel zlib-devel openssl-devel \
+      numactl-devel gperftools-libs \
+  && dnf clean all && rm -rf /var/cache/dnf/*
+
+# 2. Install "TheRock" ROCm SDK (Tarball Method)
+WORKDIR /tmp
+ARG ROCM_MAJOR_VER=7
+ARG GFX=gfx1151
+RUN set -euo pipefail; \
+    BASE="https://therock-nightly-tarball.s3.amazonaws.com"; \
+    PREFIX="therock-dist-linux-${GFX}-${ROCM_MAJOR_VER}"; \
+    KEY="$(curl -s "${BASE}?list-type=2&prefix=${PREFIX}" \
+      | tr '<' '\n' \
+      | grep -o "therock-dist-linux-${GFX}-${ROCM_MAJOR_VER}\..*\.tar\.gz" \
+      | sort -V | tail -n1)"; \
+    echo "Downloading Latest Tarball: ${KEY}"; \
+    aria2c -x 16 -s 16 -j 16 --file-allocation=none "${BASE}/${KEY}" -o therock.tar.gz; \
+    mkdir -p /opt/rocm; \
+    tar xzf therock.tar.gz -C /opt/rocm --strip-components=1; \
+    rm therock.tar.gz
+
+# 3. Configure Global ROCm Environment
+# We add LD_PRELOAD for tcmalloc here to fix the shutdown crash
+RUN export ROCM_PATH=/opt/rocm && \
+    BITCODE_PATH=$(find /opt/rocm -type d -name bitcode -print -quit) && \
+    printf '%s\n' \
+      "export ROCM_PATH=/opt/rocm" \
+      "export HIP_PLATFORM=amd" \
+      "export HIP_PATH=/opt/rocm" \
+      "export HIP_CLANG_PATH=/opt/rocm/llvm/bin" \
+      "export HIP_DEVICE_LIB_PATH=$BITCODE_PATH" \
+      "export PATH=$ROCM_PATH/bin:$ROCM_PATH/llvm/bin:\$PATH" \
+      "export LD_LIBRARY_PATH=$ROCM_PATH/lib:$ROCM_PATH/lib64:$ROCM_PATH/llvm/lib:\$LD_LIBRARY_PATH" \
+      "export ROCBLAS_USE_HIPBLASLT=1" \
+      "export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1" \
+      "export VLLM_TARGET_DEVICE=rocm" \
+      "export HIP_FORCE_DEV_KERNARG=1" \
+      "export RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES=1" \
+      "export LD_PRELOAD=/usr/lib64/libtcmalloc_minimal.so.4" \
+      > /etc/profile.d/rocm-sdk.sh && \
+    chmod 0644 /etc/profile.d/rocm-sdk.sh
+
+# 4. Python Venv Setup
+RUN /usr/bin/python3.13 -m venv /opt/venv
+ENV VIRTUAL_ENV=/opt/venv
+ENV PATH=/opt/venv/bin:$PATH
+ENV PIP_NO_CACHE_DIR=1
+RUN printf 'source /opt/venv/bin/activate\n' > /etc/profile.d/venv.sh
+RUN python -m pip install --upgrade pip wheel packaging "setuptools<80.0.0"
+
+# 5. Install PyTorch (TheRock Nightly)
+RUN python -m pip install \
+  --index-url https://rocm.nightlies.amd.com/v2-staging/gfx1151/ \
+  --pre torch torchaudio torchvision
+
+WORKDIR /opt
+
+# 6. Clone vLLM
+RUN git clone https://github.com/vllm-project/vllm.git /opt/vllm
+WORKDIR /opt/vllm
+
+# --- PATCHING ---
+RUN echo "import sys, re" > patch_strix.py && \
+    echo "from pathlib import Path" >> patch_strix.py && \
+    # Patch 1: __init__.py
+    echo "p = Path('vllm/platforms/__init__.py')" >> patch_strix.py && \
+    echo "txt = p.read_text()" >> patch_strix.py && \
+    echo "txt = txt.replace('import amdsmi', '# import amdsmi')" >> patch_strix.py && \
+    echo "txt = re.sub(r'is_rocm = .*', 'is_rocm = True', txt)" >> patch_strix.py && \
+    echo "txt = re.sub(r'if len\(amdsmi\.amdsmi_get_processor_handles\(\)\) > 0:', 'if True:', txt)" >> patch_strix.py && \
+    echo "txt = txt.replace('amdsmi.amdsmi_init()', 'pass')" >> patch_strix.py && \
+    echo "txt = txt.replace('amdsmi.amdsmi_shut_down()', 'pass')" >> patch_strix.py && \
+    echo "p.write_text(txt)" >> patch_strix.py && \
+    # Patch 2: rocm.py
+    echo "p = Path('vllm/platforms/rocm.py')" >> patch_strix.py && \
+    echo "txt = p.read_text()" >> patch_strix.py && \
+    echo "header = 'import sys\nfrom unittest.mock import MagicMock\nsys.modules[\"amdsmi\"] = MagicMock()\n'" >> patch_strix.py && \
+    echo "txt = header + txt" >> patch_strix.py && \
+    echo "txt = re.sub(r'device_type = .*', 'device_type = \"cuda\"', txt)" >> patch_strix.py && \
+    echo "txt = re.sub(r'device_name = .*', 'device_name = \"cuda\"', txt)" >> patch_strix.py && \
+    echo "txt += '\n    def get_device_name(self, device_id: int = 0) -> str:\n        return \"AMD-gfx1151\"\n'" >> patch_strix.py && \
+    echo "p.write_text(txt)" >> patch_strix.py && \
+    echo "print('Successfully patched vLLM for Strix Halo')" >> patch_strix.py && \
+    python patch_strix.py && \
+    sed -i 's/gfx1200;gfx1201/gfx1151;gfx1200;gfx1201/' CMakeLists.txt
+
+# 7. Build vLLM (Wheel Method) with CLANG Host Compiler
+RUN python -m pip install --upgrade cmake ninja packaging wheel numpy "setuptools-scm>=8" "setuptools<80.0.0" scikit-build-core pybind11
+ENV ROCM_HOME="/opt/rocm"
+ENV HIP_PATH="/opt/rocm"
+ENV VLLM_TARGET_DEVICE="rocm"
+ENV PYTORCH_ROCM_ARCH="gfx1151"
+ENV MAX_JOBS="4"
+
+# --- CRITICAL FIX FOR SEGFAULT ---
+# We force the Host Compiler (CC/CXX) to be the ROCm Clang, not Fedora GCC.
+# This aligns the ABI of the compiled vLLM extensions with PyTorch.
+ENV CC="/opt/rocm/llvm/bin/clang"
+ENV CXX="/opt/rocm/llvm/bin/clang++"
+
+RUN export HIP_DEVICE_LIB_PATH=$(find /opt/rocm -type d -name bitcode -print -quit) && \
+    echo "Compiling with Bitcode: $HIP_DEVICE_LIB_PATH" && \
+    export CMAKE_ARGS="-DROCM_PATH=/opt/rocm -DHIP_PATH=/opt/rocm" && \
+    python -m pip wheel --no-build-isolation --no-deps -w /tmp/dist -v . && \
+    python -m pip install /tmp/dist/*.whl
+
+# 8. Final Cleanup & Runtime
+WORKDIR /opt
+RUN chmod -R a+rwX /opt && \
+    find /opt/venv -type f -name "*.so" -exec strip -s {} + 2>/dev/null || true && \
+    find /opt/venv -type d -name "__pycache__" -prune -exec rm -rf {} + && \
+    rm -rf /root/.cache/pip || true && \
+    dnf clean all && rm -rf /var/cache/dnf/*
+
+COPY scripts/01-rocm-env-for-triton.sh /etc/profile.d/01-rocm-env-for-triton.sh
+COPY scripts/99-toolbox-banner.sh /etc/profile.d/99-toolbox-banner.sh
+COPY scripts/zz-venv-last.sh /etc/profile.d/zz-venv-last.sh
+RUN chmod 0644 /etc/profile.d/*.sh
+RUN printf 'ulimit -S -c 0\n' > /etc/profile.d/90-nocoredump.sh && chmod 0644 /etc/profile.d/90-nocoredump.sh
+
+CMD ["/bin/bash"]
diff --git a/Dockerfile.vllm-therock-gfx1151-aotriton b/Dockerfile.vllm-therock-gfx1151-aotriton
deleted file mode 100644
index 5df88e5..0000000
--- a/Dockerfile.vllm-therock-gfx1151-aotriton
+++ /dev/null
@@ -1,142 +0,0 @@
-FROM kyuz0/pytorch-therock-gfx1151-aotriton-builder:latest AS vllm-builder
-
-# Clone vLLM repository (shallow clone)
-RUN git clone --depth 1 https://github.com/vllm-project/vllm.git
-
-# Install vLLM build dependencies and build vLLM
-RUN source .venv/bin/activate && \
-    cd vllm && \
-    uv pip install ninja cmake wheel pybind11 && \
-    uv pip install --upgrade numba scipy huggingface-hub[cli] "numpy<2" && \
-    python use_existing_torch.py && \
-    sed -i '/amdsmi==/d' requirements/rocm-build.txt && \
-    sed -i '/pytorch-triton-rocm/d' requirements/rocm-build.txt && \
-    sed -i '/triton==/d' requirements/rocm-build.txt && \
-    uv pip install -r requirements/rocm-build.txt
-
-# Apply gfx1151 fixes
-RUN cd vllm && \
-    sed -i 's/gfx1200;gfx1201/gfx1151;gfx1200;gfx1201/' CMakeLists.txt && \
-    sed -i '/torch == 2.8.0,/d' pyproject.toml && \
-    sed -i 's/import torch/try:\n    import torch\n    from torch.utils.cpp_extension import CUDA_HOME, ROCM_HOME\n    TORCH_AVAILABLE = True\nexcept ImportError:\n    torch = None\n    CUDA_HOME = None\n    ROCM_HOME = None\n    TORCH_AVAILABLE = False/' setup.py && \
-    sed -i 's/from torch.utils.cpp_extension import CUDA_HOME, ROCM_HOME/# Moved to try block above/' setup.py && \
-    sed -i 's/torch.version.cuda is None/TORCH_AVAILABLE and torch.version.cuda is None/' setup.py && \
-    sed -i 's/has_cuda = torch.version.cuda is not None/has_cuda = TORCH_AVAILABLE and torch.version.cuda is not None/' setup.py && \
-    sed -i 's/torch.version.hip is not None/TORCH_AVAILABLE and torch.version.hip is not None/' setup.py && \
-    sed -i 's/rocm_version = get_rocm_version() or torch.version.hip/rocm_version = get_rocm_version() or (torch.version.hip if TORCH_AVAILABLE else None)/' setup.py && \
-    sed -i 's/cuda_major, cuda_minor = torch.version.cuda.split(".")/cuda_major, cuda_minor = torch.version.cuda.split(".") if TORCH_AVAILABLE else ("0", "0")/' setup.py
-
-# Fix ROCm platform detection
-RUN cd vllm && \
-    git checkout HEAD -- vllm/platforms/__init__.py && \
-    sed -i '/def rocm_platform_plugin/,/return "vllm.platforms.rocm.RocmPlatform" if is_rocm else None/s/is_rocm = False/is_rocm = False/' vllm/platforms/__init__.py && \
-    sed -i '/def rocm_platform_plugin/,/return "vllm.platforms.rocm.RocmPlatform" if is_rocm else None/s/logger.debug("Checking if ROCm platform is available.")/logger.debug("Checking if ROCm platform is available.")\n    \n    # Skip amdsmi check due to segfault issues - default to ROCm for AMD systems/' vllm/platforms/__init__.py && \
-    sed -i '/def rocm_platform_plugin/,/return "vllm.platforms.rocm.RocmPlatform" if is_rocm else None/s/try:\n        import amdsmi/try:\n        import torch/' vllm/platforms/__init__.py && \
-    sed -i '/def rocm_platform_plugin/,/return "vllm.platforms.rocm.RocmPlatform" if is_rocm else None/s/amdsmi.amdsmi_init()/# amdsmi disabled - using torch detection/' vllm/platforms/__init__.py && \
-    sed -i '/def rocm_platform_plugin/,/return "vllm.platforms.rocm.RocmPlatform" if is_rocm else None/s/try:\n            if len(amdsmi.amdsmi_get_processor_handles()) > 0:/if hasattr(torch, '\''version'\'') and hasattr(torch.version, '\''hip'\'') and torch.version.hip is not None:/' vllm/platforms/__init__.py && \
-    sed -i '/def rocm_platform_plugin/,/return "vllm.platforms.rocm.RocmPlatform" if is_rocm else None/s/is_rocm = True\n                logger.debug("Confirmed ROCm platform is available.")/is_rocm = True\n            logger.debug("ROCm platform detected via torch.version.hip")/' vllm/platforms/__init__.py && \
-    sed -i '/def rocm_platform_plugin/,/return "vllm.platforms.rocm.RocmPlatform" if is_rocm else None/s/else:\n                logger.debug("ROCm platform is not available because"\n                             " no GPU is found.")/else:\n            # Fallback: assume ROCm if we'\''re not CUDA and not other platforms\n            logger.debug("Defaulting to ROCm platform (amdsmi disabled due to segfault)")\n            is_rocm = True/' vllm/platforms/__init__.py && \
-    sed -i '/def rocm_platform_plugin/,/return "vllm.platforms.rocm.RocmPlatform" if is_rocm else None/s/finally:\n            amdsmi.amdsmi_shut_down()/finally:\n            # amdsmi disabled\n            pass/' vllm/platforms/__init__.py && \
-    sed -i '/def rocm_platform_plugin/,/return "vllm.platforms.rocm.RocmPlatform" if is_rocm else None/s/logger.debug("ROCm platform is not available because: %s", str(e))/logger.debug("ROCm platform check failed: %s", str(e))\n        # Still default to ROCm as fallback\n        is_rocm = True/' vllm/platforms/__init__.py
-
-# Also patch vllm/platforms/rocm.py to avoid amdsmi at runtime
-RUN cd vllm && python - <<'PY'
-from pathlib import Path
-p = Path("vllm/platforms/rocm.py")
-s = p.read_text()
-
-# Add amdsmi stubs if missing
-if "def amdsmi_init" not in s:
-    s += """
-
-# --- vllm-therock patch: tolerate missing 'amdsmi' ---
-try:
-    amdsmi
-except Exception:
-    amdsmi = None
-
-if 'amdsmi_init' not in globals():
-    def amdsmi_init(): return None
-    def amdsmi_shut_down(): return None
-"""
-
-# Override get_device_name to avoid amdsmi
-if "vllm_therock_rocm_get_device_name" not in s:
-    s += r"""
-
-def vllm_therock_rocm_get_device_name(self, device_id: int = 0):
-    try:
-        import torch
-        return torch.cuda.get_device_name(device_id)
-    except Exception:
-        return "AMD-gfx1151"
-
-try:
-    RocmPlatform.get_device_name = vllm_therock_rocm_get_device_name
-except Exception:
-    pass
-"""
-p.write_text(s)
-print("Patched", p)
-PY
-
-
-# Build vLLM
-RUN source .venv/bin/activate && \
-    cd vllm && \
-    uv pip uninstall amdsmi || echo "amdsmi not installed" && \
-    printf '#!/bin/bash\necho "gfx1151"\n' > /usr/local/bin/amdgpu-arch && \
-    chmod +x /usr/local/bin/amdgpu-arch && \
-    printf '#!/bin/bash\necho "gfx1151"\n' > /usr/bin/amdgpu-arch && \
-    chmod +x /usr/bin/amdgpu-arch && \
-    printf '#!/bin/bash\necho "gfx1151"\n' > /bin/amdgpu-arch && \
-    chmod +x /bin/amdgpu-arch && \
-    export PYTORCH_ROCM_ARCH="gfx1151" && \
-    /torch-therock/.venv/bin/python -c "import torch; print('torch==' + torch.__version__)" > /tmp/constraints.txt && \
-    /torch-therock/.venv/bin/python -c "import triton; print('pytorch-triton-rocm==' + getattr(triton, '__version__', 'unknown'))" >> /tmp/constraints.txt || echo "# triton version not found" >> /tmp/constraints.txt && \
-    TORCH_CMAKE_PATH=$(/torch-therock/.venv/bin/python -c "import torch; print(torch.utils.cmake_prefix_path)") && \
-    VLLM_TARGET_DEVICE=rocm CMAKE_PREFIX_PATH="$TORCH_CMAKE_PATH" Torch_DIR="$TORCH_CMAKE_PATH/Torch" CMAKE_ARGS="-DGPU_TARGETS=gfx1151 -DHIP_TARGETS=gfx1151 -DAMDGPU_TARGETS=gfx1151" /torch-therock/.venv/bin/pip install . --no-build-isolation --constraint /tmp/constraints.txt
-
-# Runtime stage
-FROM archlinux:latest
-
-# Install runtime dependencies + compilation tools
-RUN pacman -Syu --noconfirm && \
-    pacman -S --noconfirm ca-certificates gcc make cmake ninja git && \
-    pacman -Scc --noconfirm && \
-    git clone --depth 1 https://github.com/pyenv/pyenv.git /opt/pyenv && \
-    export PYENV_ROOT=/opt/pyenv && \
-    export PATH=$PYENV_ROOT/bin:$PATH && \
-    eval "$(pyenv init -)" && \
-    pyenv install 3.12.9 && \
-    pyenv global 3.12.9
-
-# Copy complete environment from builder
-COPY --from=vllm-builder /opt/pyenv /opt/pyenv
-COPY --from=vllm-builder /torch-therock/.venv /torch-therock/.venv
-COPY --from=vllm-builder /torch-therock/*.sh /torch-therock/
-
-# Set environment
-ENV PYENV_ROOT=/opt/pyenv
-ENV PYENV_VERSION=3.12.9
-ENV PATH="/opt/pyenv/versions/3.12.9/bin:/torch-therock/.venv/bin:$PATH"
-ENV PYTORCH_ROCM_ARCH=gfx1151
-
-WORKDIR /torch-therock
-
-# Test installation
-RUN /torch-therock/.venv/bin/python -c "import torch; print('PyTorch version:', torch.__version__)" && \
-    /torch-therock/.venv/bin/python -c "import vllm; print('vLLM version:', vllm.__version__)"
-
-# Toolbx compatibility - fix permissions and add environment setup
-RUN chmod -R a+rwX /torch-therock
-
-# Copy toolbx scripts
-COPY scripts/vllm-env.sh /etc/profile.d/vllm-env.sh
-COPY scripts/vllm-banner.sh /etc/profile.d/vllm-banner.sh
-RUN chmod 644 /etc/profile.d/vllm-env.sh /etc/profile.d/vllm-banner.sh
-
-COPY scripts/start-vllm.sh /usr/local/bin/start-vllm
-RUN chmod 755 /usr/local/bin/start-vllm
-
-CMD ["bash", "-c", "source .venv/bin/activate && bash"]
\ No newline at end of file
diff --git a/scripts/01-rocm-env-for-triton.sh b/scripts/01-rocm-env-for-triton.sh
new file mode 100644
index 0000000..3ad1166
--- /dev/null
+++ b/scripts/01-rocm-env-for-triton.sh
@@ -0,0 +1,3 @@
+# Required for Strix Halo / RDNA3.5 on vLLM
+export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
+export VLLM_TARGET_DEVICE=rocm
\ No newline at end of file
diff --git a/scripts/vllm-banner.sh b/scripts/99-toolbox-banner.sh
similarity index 100%
rename from scripts/vllm-banner.sh
rename to scripts/99-toolbox-banner.sh
diff --git a/scripts/vllm-env.sh b/scripts/vllm-env.sh
deleted file mode 100644
index 943b87d..0000000
--- a/scripts/vllm-env.sh
+++ /dev/null
@@ -1,30 +0,0 @@
-#!/usr/bin/env bash
-# Auto-activate vLLM environment for toolbx
-
-# Activate PyTorch + vLLM environment
-source /torch-therock/.venv/bin/activate
-
-# ROCm and performance environment variables
-export PYTORCH_ROCM_ARCH=gfx1151
-export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
-export VLLM_USE_TRITON_FLASH_ATTN=0
-export TORCH_COMPILE_DEBUG=1
-export VLLM_COMPILE_LEVEL=3
-
-# Detect and export ROCm toolchain paths
-eval "$(
-python3 - <<'PY'
-try:
-    import pathlib, _rocm_sdk_core as r
-    base = pathlib.Path(r.__file__).parent / "lib" / "llvm" / "bin"
-    lib  = pathlib.Path(r.__file__).parent / "lib"
-    print(f'export TRITON_HIP_LLD_PATH="{base / "ld.lld"}"')
-    print(f'export TRITON_HIP_CLANG_PATH="{base / "clang++"}"')
-    print(f'export LD_LIBRARY_PATH="{lib}:$LD_LIBRARY_PATH"')
-except ImportError:
-    pass
-PY
-)" 2>/dev/null || true
-
-# Enable flash attention
-export FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE
\ No newline at end of file
diff --git a/scripts/zz-venv-last.sh b/scripts/zz-venv-last.sh
new file mode 100644
index 0000000..eee3228
--- /dev/null
+++ b/scripts/zz-venv-last.sh
@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+# Ensure /opt/venv/bin is first even if ~/.local/bin or ~/.cargo/bin prepend themselves via user dotfiles.
+
+_venv_path_fix() {
+  # remove any existing /opt/venv/bin entries, then prepend one
+  local newpath
+  newpath="$(printf '%s' "$PATH" | awk -v RS=: -v ORS=: '$0!="/opt/venv/bin"{print}')"
+  PATH="/opt/venv/bin:${newpath%:}"
+}
+
+# run once after shell init; don't duplicate
+case "$PROMPT_COMMAND" in
+  *_venv_path_fix*) : ;;
+  *) PROMPT_COMMAND="_venv_path_fix${PROMPT_COMMAND:+;$PROMPT_COMMAND}" ;;
+esac
+