[PATCH] Tensorflow Training 2.18 CVE Patch

dlc_developer_config.toml

@@ -37,12 +37,12 @@ deep_canary_mode = false
 [build]
 # Add in frameworks you would like to build. By default, builds are disabled unless you specify building an image.
 # available frameworks - ["base", "vllm", "autogluon", "huggingface_tensorflow", "huggingface_pytorch", "huggingface_tensorflow_trcomp", "huggingface_pytorch_trcomp", "pytorch_trcomp", "tensorflow", "pytorch", "stabilityai_pytorch"]
-build_frameworks = []
+build_frameworks = ["tensorflow"]
 
 
 # By default we build both training and inference containers. Set true/false values to determine which to build.
 build_training = true
-build_inference = true
+build_inference = false
 
 # Set do_build to "false" to skip builds and test the latest image built by this PR
 # Note: at least one build is required to set do_build to "false"
@@ -120,7 +120,7 @@ use_scheduler = false
 
 # Standard Framework Training
 dlc-pr-pytorch-training = ""
-dlc-pr-tensorflow-2-training = ""
+dlc-pr-tensorflow-2-training = "tensorflow/training/buildspec-2-18-sm.yml"
 dlc-pr-autogluon-training = ""
 
 # ARM64 Training

tensorflow/training/buildspec-2-18-ec2.yml

@@ -5,7 +5,7 @@ framework: &FRAMEWORK tensorflow
 version: &VERSION 2.18.0
 short_version: &SHORT_VERSION "2.18"
 arch_type: x86
-autopatch_build: "True"
+# autopatch_build: "True"
 
 repository_info:
   training_repository: &TRAINING_REPOSITORY

tensorflow/training/buildspec-2-18-sm.yml

@@ -5,7 +5,7 @@ framework: &FRAMEWORK tensorflow
 version: &VERSION 2.18.0
 short_version: &SHORT_VERSION "2.18"
 arch_type: x86
-autopatch_build: "True"
+# autopatch_build: "True"
 
 repository_info:
   training_repository: &TRAINING_REPOSITORY

tensorflow/training/docker/2.18/py3/Dockerfile.cpu

@@ -272,15 +272,16 @@ RUN $PYTHON -m pip install --no-cache-dir -U \
     seaborn \
     shap
 
+
 RUN $PYTHON -m pip install --no-cache-dir -U \
-    "sagemaker<3" \
-    sagemaker-experiments==0.* \
-    sagemaker-tensorflow-training \
-    sagemaker-training \
+    "sagemaker>=2.172.0,<3" \
+    "sagemaker-experiments==0.1.7" \
+    "sagemaker-tensorflow-training==20.5.0" \
+    "sagemaker-training>=4.3.0,<=4.7.4" \
     "sagemaker-studio-analytics-extension<1" \
     "sparkmagic<1" \
     "sagemaker-studio-sparkmagic-lib<1" \
-    smclarify
+    "smclarify"
 
 # Remove python kernel installed by sparkmagic
 RUN /usr/local/bin/jupyter-kernelspec remove -f python3

tensorflow/training/docker/2.18/py3/Dockerfile.ec2.cpu.py_scan_allowlist.json

@@ -0,0 +1,3 @@
+{
+    "77740": "[Package: protobuf] Affected versions of this package are vulnerable to a potential Denial of Service (DoS) attack due to unbounded recursion when parsing untrusted Protocol Buffers data. The pure-Python implementation fails to enforce recursion depth limits when processing recursive groups, recursive messages, or a series of SGROUP tags, leading to stack overflow conditions that can crash the application by exceeding Python's recursion limit."
+}

tensorflow/training/docker/2.18/py3/Dockerfile.sagemaker.cpu.py_scan_allowlist.json

@@ -0,0 +1,3 @@
+{
+    "77740": "[Package: protobuf] Affected versions of this package are vulnerable to a potential Denial of Service (DoS) attack due to unbounded recursion when parsing untrusted Protocol Buffers data. The pure-Python implementation fails to enforce recursion depth limits when processing recursive groups, recursive messages, or a series of SGROUP tags, leading to stack overflow conditions that can crash the application by exceeding Python's recursion limit."
+}

tensorflow/training/docker/2.18/py3/cu125/Dockerfile.ec2.gpu.py_scan_allowlist.json

@@ -0,0 +1,3 @@
+{
+    "77740": "[Package: protobuf] Affected versions of this package are vulnerable to a potential Denial of Service (DoS) attack due to unbounded recursion when parsing untrusted Protocol Buffers data. The pure-Python implementation fails to enforce recursion depth limits when processing recursive groups, recursive messages, or a series of SGROUP tags, leading to stack overflow conditions that can crash the application by exceeding Python's recursion limit."
+}

tensorflow/training/docker/2.18/py3/cu125/Dockerfile.gpu

@@ -246,6 +246,20 @@ RUN mkdir /tmp/efa-ofi-nccl \
  && make install \
  && rm -rf /tmp/efa-ofi-nccl
 
+# patch nvjpeg
+RUN mkdir -p /tmp/nvjpeg \
+&& cd /tmp/nvjpeg \
+&& wget https://developer.download.nvidia.com/compute/cuda/redist/libnvjpeg/linux-x86_64/libnvjpeg-linux-x86_64-12.4.0.76-archive.tar.xz \
+&& tar -xvf libnvjpeg-linux-x86_64-12.4.0.76-archive.tar.xz \
+&& rm -rf /usr/local/cuda/targets/x86_64-linux/lib/libnvjpeg* \
+&& rm -rf /usr/local/cuda/targets/x86_64-linux/include/nvjpeg.h \
+&& cp libnvjpeg-linux-x86_64-12.4.0.76-archive/lib/libnvjpeg* /usr/local/cuda/targets/x86_64-linux/lib/ \
+&& cp libnvjpeg-linux-x86_64-12.4.0.76-archive/include/* /usr/local/cuda/targets/x86_64-linux/include/ \
+&& rm -rf /tmp/nvjpeg \
+# patch cuobjdump and nvdisasm
+&& rm -rf /usr/local/cuda/bin/cuobjdump* \
+&& rm -rf /usr/local/cuda/bin/nvdisasm* 
+
 # Allow OpenSSH to talk to containers without asking for confirmation
 RUN cat /etc/ssh/ssh_config | grep -v StrictHostKeyChecking > /etc/ssh/ssh_config.new \
    && echo "    StrictHostKeyChecking no" >> /etc/ssh/ssh_config.new \
@@ -363,14 +377,14 @@ RUN $PYTHON -m pip install --no-cache-dir -U \
     shap
 
 RUN $PYTHON -m pip install --no-cache-dir -U \
-    "sagemaker<3" \
-    sagemaker-experiments==0.* \
-    sagemaker-tensorflow-training \
-    sagemaker-training \
+    "sagemaker>=2.172.0,<3" \
+    "sagemaker-experiments==0.1.7" \
+    "sagemaker-tensorflow-training==20.5.0" \
+    "sagemaker-training>=4.3.0,<=4.7.4" \
     "sagemaker-studio-analytics-extension<1" \
     "sparkmagic<1" \
     "sagemaker-studio-sparkmagic-lib<1" \
-    smclarify
+    "smclarify"
 
 # install boost
 # tensorflow is compiled with --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=1"

tensorflow/training/docker/2.18/py3/cu125/Dockerfile.sagemaker.gpu.py_scan_allowlist.json

@@ -0,0 +1,3 @@
+{
+    "77740": "[Package: protobuf] Affected versions of this package are vulnerable to a potential Denial of Service (DoS) attack due to unbounded recursion when parsing untrusted Protocol Buffers data. The pure-Python implementation fails to enforce recursion depth limits when processing recursive groups, recursive messages, or a series of SGROUP tags, leading to stack overflow conditions that can crash the application by exceeding Python's recursion limit."
+}