add timeouts

Signed-off-by: Brian Gallagher <[email protected]>

Author: briangallagher

.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}}'

.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}}'

images/universal/training/py312-cuda128-torch280/Dockerfile

@@ -30,7 +30,8 @@ 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_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).
@@ -54,22 +55,32 @@ RUN dnf config-manager \
 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
 ENV CUDA_HOME=/usr/local/cuda \
-    PATH=/usr/local/cuda/bin:$PATH \
-    LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH \
-    TORCH_CUDA_ARCH_LIST="8.0;8.6;8.9;9.0"
+    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 
+    # TODO: revisit this, going with default list for now.
+    # TORCH_CUDA_ARCH_LIST="8.0;8.6;8.9;9.0"
 
+# 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
@@ -80,12 +91,22 @@ RUN ldconfig -p | grep -E 'libcudart|libcublas|libcudnn' || (echo "[fail-fast] C
 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
+
 # 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.
@@ -94,7 +115,11 @@ ARG TARGETARCH
 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 pip install --retries 5 --timeout 300 --no-cache-dir torch==2.8.0 --index-url https://download.pytorch.org/whl/cu128
+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'
@@ -104,19 +129,32 @@ sys.exit(0 if torch.backends.cuda.is_built() else 1)
 PY
 
 # Install numpy ahead of building extensions that expect it
-RUN pip install --retries 5 --timeout 300 --no-cache-dir numpy==2.3.3
+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 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 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"
 
-# Provide ninja via pip (RHEL/UBI repo ninja-build may be unavailable)
-RUN pip install --retries 5 --timeout 300 --no-cache-dir ninja
+# 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 pip install --retries 5 --timeout 300 --no-cache-dir \
+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 \
@@ -125,6 +163,8 @@ RUN pip install --retries 5 --timeout 300 --no-cache-dir \
     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 \
@@ -140,9 +180,6 @@ RUN pip install --retries 5 --timeout 300 --no-cache-dir \
     kernels==0.10.3 \
     "sentencepiece>=0.1.99,<0.3" \
     tokenizers==0.22.1 \
-    instructlab-training==0.12.1 \
-    rhai-innovation-mini-trainer==0.3.0 \
-    training-hub==0.3.0 \
     trl==0.21.0 \
     deepspeed>=0.14.3 \
     async-timeout==4.0.3 \
@@ -151,12 +188,97 @@ 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
+#
+# 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: sub-dep first, then parent without deps (align with runtime)
-RUN pip install --retries 5 --timeout 300 --no-cache-dir --no-build-isolation causal-conv1d==1.5.3.post1 && \
-    pip install --retries 5 --timeout 300 --no-cache-dir --no-build-isolation mamba-ssm==2.2.6.post3 --no-deps && \
+# Deterministic 2-step with verbose logging for debugging builds
+# 1) Build and log wheel via heredoc script (avoids fragile inline quoting); 2) Install from local wheel
+RUN 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=ON -DCMAKE_BUILD_TYPE=Release -DCMAKE_CUDA_FLAGS:STRING="--ptxas-options=-v -v" -DCMAKE_CXX_FLAGS:STRING="-v"' \
+    echo "[diag] nproc=$(nproc)"; df -h /tmp || true; cmake --version || true; ninja --version || true; which nvcc && nvcc -V || true; which g++ && g++ --version | head -n1 || true; \
+    export PYTORCH_NVCC_FLAGS="-v -lineinfo" FORCE_CUDA=1 TORCH_EXTENSIONS_VERBOSE=1; \
+    printf '%s\n' \
+      '#!/usr/bin/env bash' \
+      'set -euo pipefail' \
+      'exec > >(stdbuf -oL cat) 2> >(stdbuf -oL cat >&2)' \
+      'build_log=/tmp/pip-mamba-ssm.log' \
+      'mkdir -p /tmp/wheels' \
+      ': > "$build_log"' \
+      'echo "[diag] starting wheel build at: $(date)"' \
+      'TIME_LIMIT_SECS=1200' \
+      'PYTHON_WHEEL_CMD='\''python -m pip wheel --no-cache-dir --no-build-isolation -vv'\''' \
+      '$PYTHON_WHEEL_CMD \' \
+      '  --log "$build_log" \' \
+      '  mamba-ssm==2.2.6.post3 --no-deps -w /tmp/wheels &' \
+      'pid=$!' \
+      '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=${minute})" >&2' \
+      '  echo "[heartbeat] ps snapshot:" >&2' \
+      '  ps -o pid,pgid,stat,etime,pcpu,pmem,comm,args -p "$pid" || true' \
+      '  echo "[heartbeat] recent log:" >&2' \
+      '  tail -n 80 "$build_log" >&2 || true' \
+      '  echo "[heartbeat] compiled objects so far: $(find /tmp -type f -name "*.o" | wc -l || true)" >&2' \
+      '  df -h /tmp >&2 || true' \
+      '  if [ "$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 30' \
+      'done' \
+      'wait "$pid"; status=$?' \
+      '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

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