Remove unused binaries (#4949)

* Remove unused binaries

* test TF

* test base

* test vllm

* retest TF

* test PT arm64

* Restore ['dlc_developer_config.toml']

dlc_developer_config.toml:
('Restore to '
 'https://raw.githubusercontent.com/aws/deep-learning-containers/master/dlc_developer_config.toml')

Author: sallyseok

miscellaneous_scripts/dockerfile_patch_script.sh

@@ -43,10 +43,9 @@ fi
 chmod +x $PATCHING_INFO_PATH/patch-details/install_script_language.sh && \
 $PATCHING_INFO_PATH/patch-details/install_script_language.sh
 
-# Upgrade sagemaker-training package to latest
-# For PT 2.7 sagemaker has dependency on protobuf 3.20.3 and sagemaker-training 4.8.3
-if pip show sagemaker-training; then
-    pip install "sagemaker-training>4.7.4,<=4.8.3" --upgrade
+# Upgrade sagemaker-training 
+if [[ $LATEST_RELEASED_IMAGE_URI =~ ^763104351884\.dkr\.ecr\.us-west-2\.amazonaws\.com/pytorch-training:2\.[4-6](.+)sagemaker ]]; then
+    pip install "sagemaker-training>4.7.4,<5" --upgrade
 fi
 
 # For PT inference sagemaker images, replace torchserve-entrypoint.py with the latest one
@@ -69,6 +68,12 @@ if [[ $LATEST_RELEASED_IMAGE_URI =~ ^763104351884\.dkr\.ecr\.us-west-2\.amazonaw
     chmod +x /usr/local/bin/start_cuda_compat.sh
 fi
 
+# For all GPU images, remove cuobjdump and nvdisasm
+if [[ $LATEST_RELEASED_IMAGE_URI =~ ^763104351884\.dkr\.ecr\.us-west-2\.amazonaws\.com/(pytorch|tensorflow)(.+)gpu(.+) ]]; then
+    rm -rf /usr/local/cuda/bin/cuobjdump*
+    rm -rf /usr/local/cuda/bin/nvdisasm*
+fi
+
 pip cache purge
 
 ## Update GPG key in case Nginx exists

scripts/install_cuda.sh

@@ -12,6 +12,7 @@ function prune_cuda {
     # - Debugging tools (compute-sanitizer, debugger)
     # - Profiling tools (Nsight Compute, Nsight Systems)
     # - Legacy tools (Visual Profiler)
+    # - ELF file processing components
     # This keeps only the essential runtime libraries, headers and development tools
     rm -rf /usr/local/cuda/compute-sanitizer/docs \
         /usr/local/cuda/nsight-compute-****.*.*/docs \
@@ -24,7 +25,9 @@ function prune_cuda {
         /usr/local/cuda/compute-sanitizer \
         /usr/local/cuda/extras/Debugger \
         /usr/local/cuda/nsight-compute-****.*.* \
-        /usr/local/cuda/nsight-systems-****.*.*
+        /usr/local/cuda/nsight-systems-****.*.* \
+        /usr/local/cuda/bin/cuobjdump* \
+        /usr/local/cuda/bin/nvdisasm*
     rm -rf /usr/local/cuda/doc
     rm -rf /usr/local/cuda/samples
     rm -rf /usr/local/cuda/share/doc

vllm/x86_64/gpu/Dockerfile

@@ -42,6 +42,9 @@ RUN chmod +x /usr/local/bin/deep_learning_container.py && \
     ${HOME_DIR}/oss_compliance/generate_oss_compliance.sh ${HOME_DIR} ${PYTHON} && \
     # create symlink for python
     ln -s /usr/bin/python3 /usr/bin/python && \
+    # remove cuobjdump and nvdisasm
+    rm -rf /usr/local/cuda/bin/cuobjdump* && \
+    rm -rf /usr/local/cuda/bin/nvdisasm* && \
     # clean up
     rm -rf ${HOME_DIR}/oss_compliance* && \
     rm -rf /tmp/tmp* && \