diff --git a/.tekton/universal-image-py312-cuda128-torch280-pull-request.yaml b/.tekton/universal-image-py312-cuda128-torch280-pull-request.yaml
index 2fcbfadf7..a391f3762 100644
--- a/.tekton/universal-image-py312-cuda128-torch280-pull-request.yaml
+++ b/.tekton/universal-image-py312-cuda128-torch280-pull-request.yaml
@@ -18,6 +18,8 @@ metadata:
   name: universal-image-py312-cuda128-torch280-on-pull-request
   namespace: open-data-hub-tenant
 spec:
+  timeouts:
+    pipeline: 9h  
   params:
   - name: git-url
     value: '{{source_url}}'
diff --git a/.tekton/universal-image-py312-cuda128-torch280-push.yaml b/.tekton/universal-image-py312-cuda128-torch280-push.yaml
index 7e471edc1..a788a2c16 100644
--- a/.tekton/universal-image-py312-cuda128-torch280-push.yaml
+++ b/.tekton/universal-image-py312-cuda128-torch280-push.yaml
@@ -17,6 +17,8 @@ metadata:
   name: universal-image-py312-cuda128-torch280-on-push
   namespace: open-data-hub-tenant
 spec:
+  timeouts:
+    pipeline: 9h  
   params:
   - name: git-url
     value: '{{source_url}}'
diff --git a/images/universal/training/py312-cuda128-torch280/Dockerfile b/images/universal/training/py312-cuda128-torch280/Dockerfile
index ca8d30b49..603c80bb3 100644
--- a/images/universal/training/py312-cuda128-torch280/Dockerfile
+++ b/images/universal/training/py312-cuda128-torch280/Dockerfile
@@ -19,7 +19,8 @@ LABEL name="universal:py312-cuda128-torch280" \
       io.k8s.display-name="Universal CUDA 12.8 Python 3.12 (Workbench + Runtime)" \
       io.k8s.description="Universal image: Jupyter workbench by default; runtime when command provided. Includes RDMA/IB libs, Torch 2.8.0 cu128, FlashAttention 2.8.3."
 
-COPY ./images/universal/training/py312-cuda128-torch280/LICENSE.md /licenses/cuda-license.md
+## TODO: Add license file
+# COPY LICENSE.md /licenses/cuda-license.md
 
 # For OS installs we need elevated privileges; base may default to 1001
 USER 0
@@ -108,10 +109,10 @@ RUN pip install --retries 5 --timeout 300 --no-cache-dir numpy==2.3.3
 
 # Install build backend for VCS package and the SDK itself (no build isolation so backend is visible)
 RUN pip install --retries 5 --timeout 300 --no-cache-dir hatchling hatch-vcs
-RUN pip install --retries 5 --timeout 300 --no-cache-dir --no-build-isolation "git+https://github.com/briangallagher/sdk@training-hub"
+RUN pip install --retries 5 --timeout 300 --no-cache-dir --no-build-isolation "git+https://github.com/briangallagher/sdk@add-training-hub"
 
 # Provide ninja via pip (RHEL/UBI repo ninja-build may be unavailable)
-RUN pip install --retries 5 --timeout 300 --no-cache-dir ninja
+# RUN pip install --retries 5 --timeout 300 --no-cache-dir ninja
 
 # Install remaining runtime packages (resolved from default PyPI), including FlashAttention
 # Note: We intentionally do not use a Pipfile/lock here to avoid mixing resolvers with the base (uv lock),
@@ -119,26 +120,29 @@ RUN pip install --retries 5 --timeout 300 --no-cache-dir ninja
 RUN pip install --retries 5 --timeout 300 --no-cache-dir \
     flash-attn==2.8.3 --no-build-isolation \
     accelerate==1.10.0 \
-    transformers==4.55.2 \
+    transformers==4.57.1 \
     peft==0.17.0 \
     tqdm==4.67.1 \
     datasets==4.0.0 \
     pydantic>=2.11.7 \
     aiofiles==24.1.0 \
+    deprecated==1.2.18 \
+    typer==0.19.2 \
     "protobuf>=5.28.0,<6.0.0" \
     "simpleeval>=0.9.13,<1.0" \
     safetensors==0.6.2 \
     packaging==25.0 \
     pyyaml==6.0.2 \
     py-cpuinfo==9.0.0 \
-    numba==0.61.2 \
+    numba==0.62.1 \
     rich==14.1.0 \
     tensorboard==2.19.0 \
-    bitsandbytes>=0.45.3 \
-    liger-kernel==0.5.10 \
+    bitsandbytes==0.48.1 \
+    liger-kernel==0.6.2 \
+    einops==0.8.1 \
+    kernels==0.10.3 \
     "sentencepiece>=0.1.99,<0.3" \
-    tokenizers==0.21.4 \
-    training-hub==0.2.0 \
+    tokenizers==0.22.1 \
     trl==0.21.0 \
     deepspeed>=0.14.3 \
     async-timeout==4.0.3 \
@@ -147,9 +151,24 @@ RUN pip install --retries 5 --timeout 300 --no-cache-dir \
     huggingface-hub==0.34.4 \
     mlflow==3.4.0 \
     psutil==7.0.0 \
+    training-hub==0.3.0 \
+    instructlab-training==0.12.1 \
+    rhai-innovation-mini-trainer==0.3.0 \
  && chmod -R g+w /opt/app-root/lib/python3.12/site-packages \
  && fix-permissions /opt/app-root -P
 
+
+# Build helpers to compile PyTorch extensions
+RUN pip install -U pip setuptools wheel ninja cmake
+
+# Optional: set GPU archs if you hit arch issues
+# ENV TORCH_CUDA_ARCH_LIST="8.0;8.6;8.9;9.0"
+
+# Deterministic 2-step: sub-dep first, then parent without deps
+RUN pip install --no-build-isolation --no-cache-dir causal-conv1d==1.5.3.post1 && \
+    pip install --no-build-isolation --no-cache-dir mamba-ssm==2.2.6.post3 --no-deps && \
+    fix-permissions /opt/app-root -P
+
 # Provide a POSIX entrypoint wrapper to choose behavior based on invocation
 COPY --chmod=0755 ./images/universal/training/py312-cuda128-torch280/entrypoint-universal.sh /usr/local/bin/entrypoint-universal.sh
 
diff --git a/images/universal/training/py312-cuda128-torch280/Dockerfile.debug b/images/universal/training/py312-cuda128-torch280/Dockerfile.debug
new file mode 100644
index 000000000..8ba44a97f
--- /dev/null
+++ b/images/universal/training/py312-cuda128-torch280/Dockerfile.debug
@@ -0,0 +1,324 @@
+# Universal image Dockerfile
+#
+# Base image:
+# - Minimal Jupyter CUDA workbench with CUDA 12.8 and Python 3.12
+# - Provides JupyterLab, Elyra integration, addons, and default ENTRYPOINT start-notebook.sh
+# - Source: quay.io/opendatahub/workbench-images:cuda-jupyter-minimal-ubi9-python-3.12-2025a_20250903
+#
+# Design intent:
+# - Preserve workbench behavior by default (no args → start-notebook.sh)
+# - Add runtime capabilities on top (Python ML/training stack, RDMA/IB packages)
+# - Avoid duplicating dependencies provided by the base image
+# - Allow headless runtime mode when a command is provided (args → exec that command)
+
+FROM quay.io/opendatahub/workbench-images:cuda-jupyter-minimal-ubi9-python-3.12-2025a_20250903
+
+LABEL name="universal:py312-cuda128-torch280" \
+      summary="Universal CUDA 12.8 Python 3.12 image with PyTorch 2.8.0" \
+      description="Universal image combining minimal Jupyter workbench and runtime ML stack (CUDA 12.8, PyTorch 2.8.0, FlashAttention 2.8.3) on UBI9" \
+      io.k8s.display-name="Universal CUDA 12.8 Python 3.12 (Workbench + Runtime)" \
+      io.k8s.description="Universal image: Jupyter workbench by default; runtime when command provided. Includes RDMA/IB libs, Torch 2.8.0 cu128, FlashAttention 2.8.3."
+
+COPY ./images/universal/training/py312-cuda128-torch280/LICENSE.md /licenses/cuda-license.md
+
+# For OS installs we need elevated privileges; base may default to 1001
+USER 0
+WORKDIR /opt/app-root/bin
+
+# Keep NVIDIA driver capability constraints consistent with runtime image behavior
+ENV NVIDIA_VISIBLE_DEVICES=all \
+    NVIDIA_DRIVER_CAPABILITIES=compute,utility \
+    CUDA_VERSION=12.8 \
+    PIP_DEFAULT_TIMEOUT=600 \
+    PIP_DISABLE_PIP_VERSION_CHECK=1 \
+    PIP_UPGRADE_STRATEGY=only-if-needed
+
+# Follow runtime: enable CUDA and Mellanox OFED repositories for RDMA/IB packages.
+# Note: The base image already includes CUDA 12.8 runtime; we only add missing components (e.g., RDMA libs).
+RUN dnf config-manager \
+    --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel9/x86_64/cuda-rhel9.repo \
+ && dnf config-manager \
+    --add-repo https://linux.mellanox.com/public/repo/mlnx_ofed/latest/rhel9.5/mellanox_mlnx_ofed.repo \
+ && dnf install -y --disablerepo="*" --enablerepo="cuda-rhel9-x86_64,mlnx_ofed_24.10-1.1.4.0_base,ubi-9-appstream-rpms,ubi-9-baseos-rpms" \
+    libibverbs-utils \
+    infiniband-diags \
+    libibumad3 \
+    librdmacm \
+    librdmacm-utils \
+    rdma-core \
+    mlnx-tools \
+ && dnf clean all \
+ && rm -rf /var/cache/dnf/*
+
+# Install CUDA NVCC and build toolchain required to build FlashAttention from source
+# NOTE: Use command-line CUDA packages to avoid Nsight GUI deps (X11 libs) not available in UBI
+RUN dnf install -y --disablerepo="*" --enablerepo="cuda-rhel9-x86_64,ubi-9-appstream-rpms,ubi-9-baseos-rpms" \
+    cuda-command-line-tools-12-8 \
+    cuda-cudart-devel-12-8 \
+    cuda-libraries-devel-12-8 \
+    cuda-compat-12-8 \
+    cuda-compiler-12-8 \
+    cuda-nvcc-12-8-12.8.93-1 \
+    gcc \
+    gcc-c++ \
+    make \
+    python3-devel \
+    cmake \
+    git \
+ && echo "/usr/local/nvidia/lib" >> /etc/ld.so.conf.d/nvidia.conf \
+ && echo "/usr/local/nvidia/lib64" >> /etc/ld.so.conf.d/nvidia.conf \
+ && dnf clean all \
+ && rm -rf /var/cache/dnf/*
+
+# Ensure CUDA_HOME points to the toolkit and nvcc is discoverable, then sanity check nvcc
+ARG CUDA_ARCH_LIST=9.0
+ENV CUDA_HOME=/usr/local/cuda \
+    PATH=/usr/local/nvidia/bin:${CUDA_HOME}/bin:${PATH} \
+    LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64:$CUDA_HOME/lib64:$CUDA_HOME/extras/CUPTI/lib64:$LD_LIBRARY_PATH \
+    TORCH_CUDA_ARCH_LIST="${CUDA_ARCH_LIST}"
+
+# Extra verbosity and progress for CUDA extension builds (Ninja progress, Torch verbose)
+ENV USE_NINJA=1 \
+    TORCH_CUDA_VERBOSE_BUILD=1 \
+    NINJA_STATUS="[%f/%t %es] "
+# NOTE: Optional build-time CUDA checks (remove if not needed for faster builds)
+# Verify CUDA toolkit present and nvcc available
+RUN /usr/local/cuda/bin/nvcc -V
+# Verify key CUDA libs are discoverable
+RUN ldconfig -p | grep -E 'libcudart|libcublas|libcudnn' || (echo "[fail-fast] CUDA libs not found in ldconfig" >&2; exit 1)
+RUN mkdir -p /opt/app-root/.ccache && chown -R 1001:0 /opt/app-root/.ccache
+
+# Quick preflight: verify torch wheel and flash-attn index are reachable to fail fast before large downloads
+ARG TORCH_WHEEL_FILE=https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp312-cp312-manylinux_2_28_x86_64.whl
+RUN curl -IfsS --connect-timeout 10 --max-time 20 "$TORCH_WHEEL_FILE" > /dev/null || (echo "[fail-fast] Torch cu128 wheel not reachable: $TORCH_WHEEL_FILE" >&2; exit 1)
+RUN curl -IfsS --connect-timeout 10 --max-time 20 https://pypi.org/simple/flash-attn/ > /dev/null || (echo "[fail-fast] PyPI flash-attn index not reachable" >&2; exit 1)
+#
+# Additional diagnostics to help debug build env before Python installs
+RUN python -m pip debug --verbose || true
+RUN gcc --version | head -n1 || true
+RUN g++ --version | head -n1 || true
+RUN /usr/local/cuda/bin/nvcc -V | head -n1 || true
+
+# Switch back to the non-root user for Python environment changes
+USER 1001
+
+WORKDIR /opt/app-root/src
+
+# Ensure user installs land in app-root and are discoverable by Python
+ENV PYTHONUSERBASE=/opt/app-root \
+    PYTHONNOUSERSITE=0
+
+# Speed up repeated native builds using ccache (if present)
+ENV CCACHE_DIR=/opt/app-root/.ccache \
+    CCACHE_MAXSIZE=5G \
+    CCACHE_COMPRESS=1
+
+# Add runtime Python dependencies on top of the minimal Jupyter stack.
+# We intentionally avoid re-installing minimal-provided packages (e.g., jupyterlab) to prevent downgrades.
+# Torch/cu128 must match CUDA 12.8. FlashAttention is mandatory and currently supported on amd64.
+ARG TARGETARCH
+# Enforce amd64 for FlashAttention wheel availability
+RUN if [ "$TARGETARCH" != "amd64" ]; then echo "FlashAttention is mandatory and requires amd64 prebuilt wheels. Build with --platform linux/amd64." >&2; exit 1; fi
+
+# Install torch from the PyTorch CUDA index separately to avoid affecting other packages' index resolution
+RUN python -m pip install --retries 5 --timeout 300 --no-cache-dir torch==2.8.0 --index-url https://download.pytorch.org/whl/cu128
+
+# Diagnostics: show interpreter, sys.path, and user site locations
+RUN python -c "import sys,site,os; print('exe:',sys.executable); print('sys.path:',sys.path); print('userbase:',site.getuserbase()); print('usersite:',site.getusersitepackages()); print('PYTHONNOUSERSITE=',os.environ.get('PYTHONNOUSERSITE'))"
+RUN python -m pip show torch || true && python -c "import importlib.util; print('torch_spec:', importlib.util.find_spec('torch'))"
+
+# NOTE: Optional build-time check (remove if not needed): verify torch build has CUDA enabled
+RUN python - <<'PY'
+import torch, sys
+print("[check] torch", torch.__version__, "cuda build:", torch.version.cuda)
+sys.exit(0 if torch.backends.cuda.is_built() else 1)
+PY
+
+# Install numpy ahead of building extensions that expect it
+RUN python -m pip install --retries 5 --timeout 300 --no-cache-dir numpy==2.3.4
+
+# Install build backend for VCS package and the SDK itself (no build isolation so backend is visible)
+RUN python -m pip install --retries 5 --timeout 300 --no-cache-dir hatchling hatch-vcs
+RUN python -m pip install --retries 5 --timeout 300 --no-cache-dir --no-build-isolation "git+https://github.com/briangallagher/sdk@training-hub"
+#
+# NOTE: kubeflow pulls kubernetes==34.x which requires urllib3<2.4, but
+# training-hub requires urllib3>=2.4. There is no kubernetes>=35 on PyPI yet.
+# We intentionally keep urllib3>=2.4 for training-hub and accept the mismatch.
+# To avoid resolution failure, we do NOT try to force-upgrade kubernetes here.
+# RUN python -m pip install --retries 5 --timeout 180 --no-cache-dir \
+#     "kubernetes>=35.0.0" "urllib3>=2.4,<3"
+
+# Ensure modern build tooling for extensions
+RUN python -m pip install --retries 5 --timeout 300 --no-cache-dir -U pip setuptools wheel ninja cmake
+#
+# Fail-fast: ensure binary wheels exist for packages that are expensive to build
+RUN mkdir -p /tmp/wheels && \
+    python -m pip download --retries 5 --timeout 120 --only-binary=:all: --no-deps -d /tmp/wheels \
+      numba==0.62.1 bitsandbytes==0.48.1 || \
+    (echo "[fail-fast] Missing binary wheel for numba or bitsandbytes on this platform (refusing to build from source)." >&2; exit 1)
+
+# Install remaining runtime packages (resolved from default PyPI), including FlashAttention
+# Note: We intentionally do not use a Pipfile/lock here to avoid mixing resolvers with the base (uv lock),
+#       to control CUDA/FA install order and indexes, and to reduce lock churn across arches/ABI-specific wheels.
+RUN python -m pip install --retries 5 --timeout 300 --no-cache-dir --prefer-binary --only-binary=numba,bitsandbytes --upgrade-strategy only-if-needed \
+    flash-attn==2.8.3 --no-build-isolation \
+    accelerate==1.10.0 \
+    transformers==4.57.1 \
+    peft==0.17.0 \
+    tqdm==4.67.1 \
+    datasets==4.0.0 \
+    pydantic>=2.11.7 \
+    aiofiles==24.1.0 \
+    deprecated==1.2.18 \
+    typer==0.19.2 \
+    "protobuf>=5.28.0,<6.0.0" \
+    "simpleeval>=0.9.13,<1.0" \
+    safetensors==0.6.2 \
+    packaging==25.0 \
+    pyyaml==6.0.2 \
+    py-cpuinfo==9.0.0 \
+    numba==0.62.1 \
+    rich==14.1.0 \
+    tensorboard==2.19.0 \
+    bitsandbytes==0.48.1 \
+    liger-kernel==0.6.2 \
+    einops==0.8.1 \
+    kernels==0.10.3 \
+    "sentencepiece>=0.1.99,<0.3" \
+    tokenizers==0.22.1 \
+    trl==0.21.0 \
+    deepspeed>=0.14.3 \
+    async-timeout==4.0.3 \
+    aiohttp==3.12.15 \
+    hf-xet==1.1.8 \
+    huggingface-hub==0.34.4 \
+    mlflow==3.4.0 \
+    psutil==7.0.0 \
+    training-hub==0.3.0 \
+    instructlab-training==0.12.1 \
+    rhai-innovation-mini-trainer==0.3.0 \
+ && chmod -R g+w /opt/app-root/lib/python3.12/site-packages \
+ && fix-permissions /opt/app-root -P
+#
+# WARNING: Skipping `pip check` due to known kubernetes(urllib3<2.4) vs
+# training-hub(urllib3>=2.4) requirement mismatch. Re-enable once upstream
+# loosens kubernetes urllib3 bounds or kubeflow no longer pins 34.x.
+# RUN python -m pip check || (python -m pip freeze; exit 1)
+#
+# Numba diagnostics (helps catch llvmlite/LLVM/NumPy mismatches quickly)
+RUN python -m numba -s || (echo "[diagnostics] numba sysinfo failed" >&2; exit 1)
+
+# Provide CUDA user-space libraries via pip, aligning with runtime for extension builds
+RUN python -m pip install --retries 5 --timeout 300 --no-cache-dir \
+    nvidia-nccl-cu12==2.27.3 \
+    nvidia-cublas-cu12==12.8.4.1 \
+    nvidia-cuda-cupti-cu12==12.8.90 \
+    nvidia-cuda-nvrtc-cu12==12.8.93 \
+    nvidia-cuda-runtime-cu12==12.8.90 \
+    nvidia-cudnn-cu12==9.10.2.21 \
+    nvidia-cufft-cu12==11.3.3.83 \
+    nvidia-cufile-cu12==1.13.1.3 \
+    nvidia-curand-cu12==10.3.9.90 \
+    nvidia-cusolver-cu12==11.7.3.90 \
+    nvidia-cusparse-cu12==12.5.8.93 \
+    nvidia-cusparselt-cu12==0.7.1 \
+    nvidia-nvjitlink-cu12==12.8.93 \
+    nvidia-nvtx-cu12==12.8.90 \
+ && fix-permissions /opt/app-root -P
+
+# Ensure cuDNN from pip is discoverable during source builds
+ENV LD_LIBRARY_PATH="/opt/app-root/lib/python3.12/site-packages/nvidia/cudnn/lib:${LD_LIBRARY_PATH}"
+
+# Deterministic 2-step with logging; optional ccache + parallelism for speed
+# 1) Build and log wheel; 2) Install from local wheel
+RUN set -e; \
+    if command -v ccache >/dev/null 2>&1; then \
+      export CC="ccache gcc" CXX="ccache g++" CUDAHOSTCXX="ccache g++"; \
+      CC_LAUNCH='-DCMAKE_CUDA_COMPILER_LAUNCHER=ccache -DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccache'; \
+    else \
+      export CC="gcc" CXX="g++" CUDAHOSTCXX="g++"; \
+      CC_LAUNCH=''; \
+    fi; \
+    ARCH_NO_DOT="$(echo "${CUDA_ARCH_LIST}" | tr -d '.')" ; \
+    export TORCH_CUDA_ARCH_LIST="${CUDA_ARCH_LIST}" ; \
+    export PYTORCH_NVCC_FLAGS="${PYTORCH_NVCC_FLAGS:-} -gencode arch=compute_${ARCH_NO_DOT},code=sm_${ARCH_NO_DOT}" ; \
+    MAX_JOBS="$(nproc)" TORCH_CUDA_ARCH_LIST="${CUDA_ARCH_LIST}" PYTORCH_NVCC_FLAGS="${PYTORCH_NVCC_FLAGS}" python -m pip install --no-cache-dir --no-build-isolation -vv --log /tmp/pip-causal.log causal-conv1d==1.5.3.post1 && \
+    CMAKE_GENERATOR=Ninja CMAKE_BUILD_PARALLEL_LEVEL="$(nproc)" MAX_JOBS="$(nproc)" CMAKE_ARGS="-DCMAKE_VERBOSE_MAKEFILE=OFF -DCMAKE_BUILD_TYPE=Release ${CC_LAUNCH} -DCMAKE_CUDA_ARCHITECTURES=${ARCH_NO_DOT}" \
+    echo "[diag] nproc=$(nproc)"; df -h /tmp || true; cmake --version || true; ninja --version || true; which nvcc && nvcc -V || true; which g++ && g++ --version | head -n1 || true; \
+    export FORCE_CUDA=1 TORCH_EXTENSIONS_VERBOSE=0; \
+    printf '%s\n' \
+      '#!/usr/bin/env bash' \
+      'set -euo pipefail' \
+      'build_log=/tmp/pip-mamba-ssm.log' \
+      'mkdir -p /tmp/wheels' \
+      ': > "$build_log"' \
+      'echo "[diag] starting wheel build at: $(date)"' \
+      'TIME_LIMIT_SECS=1200' \
+      'ARCH_NO_DOT="$(echo "${CUDA_ARCH_LIST:-9.0}" | tr -d ".")"' \
+      'export TORCH_CUDA_ARCH_LIST="${CUDA_ARCH_LIST:-9.0}"' \
+      'export PYTORCH_NVCC_FLAGS="${PYTORCH_NVCC_FLAGS:-} -gencode arch=compute_${ARCH_NO_DOT},code=sm_${ARCH_NO_DOT}"' \
+      'if command -v ccache >/dev/null 2>&1; then' \
+      '  export CC="ccache gcc" CXX="ccache g++" CUDAHOSTCXX="ccache g++"' \
+      '  CC_LAUNCH="-DCMAKE_CUDA_COMPILER_LAUNCHER=ccache -DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccache"' \
+      'else' \
+      '  export CC="gcc" CXX="g++" CUDAHOSTCXX="g++"' \
+      '  CC_LAUNCH=""' \
+      'fi' \
+      'export CMAKE_GENERATOR=Ninja CMAKE_BUILD_PARALLEL_LEVEL="$(nproc)" MAX_JOBS="$(nproc)"' \
+      'export CMAKE_ARGS="-DCMAKE_VERBOSE_MAKEFILE=OFF -DCMAKE_BUILD_TYPE=Release ${CC_LAUNCH} -DCMAKE_CUDA_ARCHITECTURES=${ARCH_NO_DOT}"' \
+      'export FORCE_CUDA=1 TORCH_EXTENSIONS_VERBOSE=0' \
+      'rm -rf "$HOME/.cache/torch_extensions" /tmp/wheels/* || true' \
+      'PYTHON_WHEEL_CMD='\''python -m pip wheel --no-cache-dir --no-build-isolation -vv'\''' \
+      '$PYTHON_WHEEL_CMD \' \
+      '  --log "$build_log" \' \
+      '  mamba-ssm==2.2.6.post3 --no-deps -w /tmp/wheels &' \
+      'pid=$!' \
+      'echo "[diag] following pip log at $build_log" >&2' \
+      'stdbuf -oL -eL tail -F "$build_log" >&2 & tail_pid=$!' \
+      'start=$(date +%s)' \
+      'while kill -0 "$pid" 2>/dev/null; do' \
+      '  now=$(date +%s)' \
+      '  elapsed=$(( now - start ))' \
+      '  minute=$(( elapsed / 60 ))' \
+      '  echo "[heartbeat] $(date): elapsed=${elapsed}s (minute=${minute})"' \
+      '  echo "[heartbeat] $(date): elapsed=${elapsed}s (minute=${minute})" >&2' \
+      '  echo "[heartbeat] ps snapshot:" >&2' \
+      '  ps -o pid,pgid,stat,etime,pcpu,pmem,comm,args -p "$pid" || true' \
+      '  echo "[heartbeat] compiled objects so far: $(find /tmp -type f -name "*.o" | wc -l || true)"' \
+      '  echo "[heartbeat] compiled objects so far: $(find /tmp -type f -name "*.o" | wc -l || true)" >&2' \
+      '  df -h /tmp || true' \
+      '  df -h /tmp >&2 || true' \
+      '  if [ "$elapsed" -ge "$TIME_LIMIT_SECS" ]; then' \
+      '    echo "[timeout] wheel build exceeded $(( TIME_LIMIT_SECS / 60 )) minutes; killing...";' \
+      '    pgid=$(ps -o pgid= "$pid" 2>/dev/null | tr -d " " || true);' \
+      '    if [ -n "$pgid" ]; then' \
+      '      echo "[timeout] killing process group ${pgid}";' \
+      '      kill -TERM -"${pgid}" || true; sleep 10; kill -KILL -"${pgid}" || true;' \
+      '    fi' \
+      '    kill "$pid" || true; sleep 10; kill -9 "$pid" || true;' \
+      '    echo "--- pip mamba-ssm log (tail) ---"; tail -n 1000 "$build_log" || true;' \
+      '    exit 1;' \
+      '  fi' \
+      '  sleep 15' \
+      'done' \
+      'wait "$pid"; status=$?' \
+      'kill "$tail_pid" 2>/dev/null || true' \
+      'if [ "$status" -ne 0 ]; then' \
+      '  echo "[fail] wheel build exited with status $status";' \
+      '  echo "--- pip mamba-ssm log (tail) ---"; tail -n 1000 "$build_log" || true;' \
+      '  exit "$status";' \
+      'fi' \
+      > /tmp/build_mamba_wheel.sh && \
+    chmod +x /tmp/build_mamba_wheel.sh && \
+    timeout -k 120s 22m bash /tmp/build_mamba_wheel.sh || (echo '--- pip mamba-ssm log (tail) ---'; tail -n 1000 /tmp/pip-mamba-ssm.log; exit 1) && \
+    python -m pip install --no-cache-dir /tmp/wheels/*.whl && \
+    rm -rf /tmp/wheels && \
+    fix-permissions /opt/app-root -P
+
+# Provide a POSIX entrypoint wrapper to choose behavior based on invocation
+COPY --chmod=0755 ./images/universal/training/py312-cuda128-torch280/entrypoint-universal.sh /usr/local/bin/entrypoint-universal.sh
+
+# Set ENTRYPOINT to the wrapper so that providing a command runs headless.
+# Default CMD maintains workbench behavior (no args → start-notebook.sh)
+ENTRYPOINT ["/usr/local/bin/entrypoint-universal.sh"]
+CMD ["start-notebook.sh"]
diff --git a/images/universal/training/py312-cuda128-torch280/README.md b/images/universal/training/py312-cuda128-torch280/README.md
index 194af76b4..713abedb6 100644
--- a/images/universal/training/py312-cuda128-torch280/README.md
+++ b/images/universal/training/py312-cuda128-torch280/README.md
@@ -4,6 +4,7 @@ CUDA enabled container image for Training Workbench and Training Runtime in Open
 
 It includes the following layers:
 * UBI 9
+* Minimal Workbench
 * Python 3.12
 * CUDA 12.8
 * PyTorch 2.8.0