diff --git a/.buildkite/pipelines/validate_pytorch_allowlist.yml.sh b/.buildkite/pipelines/validate_pytorch_allowlist.yml.sh
index 936352d55..53a65d8c7 100755
--- a/.buildkite/pipelines/validate_pytorch_allowlist.yml.sh
+++ b/.buildkite/pipelines/validate_pytorch_allowlist.yml.sh
@@ -8,10 +8,10 @@
 # compliance with the Elastic License 2.0 and the foregoing additional
 # limitation.
 
-# Use the same Docker image as the build steps — it has Python 3.12 and
-# the source-built torch package, giving exact version parity with the
-# libtorch that pytorch_inference links against.
-VALIDATION_IMAGE="${DOCKER_IMAGE:-docker.elastic.co/ml-dev/ml-linux-build:34}"
+# Always validate against the published PyTorch Linux dependency image (same tag as
+# Linux compile agents: torch + MKL under /usr/local/gcc133 per dev-tools/docker/pytorch_linux_image).
+# Optional override for experiments: PYTORCH_ALLOWLIST_VALIDATION_IMAGE.
+VALIDATION_IMAGE="${PYTORCH_ALLOWLIST_VALIDATION_IMAGE:-docker.elastic.co/ml-dev/ml-linux-dependency-build:pytorch_latest}"
 
 cat <<EOL
 steps:
diff --git a/dev-tools/docker/pytorch_linux_image/Dockerfile b/dev-tools/docker/pytorch_linux_image/Dockerfile
index 80e1f0801..5429c5624 100644
--- a/dev-tools/docker/pytorch_linux_image/Dockerfile
+++ b/dev-tools/docker/pytorch_linux_image/Dockerfile
@@ -36,12 +36,29 @@ ENV LDFLAGS_FOR_TARGET="-Wl,-z,relro -Wl,-z,now"
 
 ARG build_dir=/usr/src
 
-# Update paths to use the compiler in C++17 mode. /usr/local/bin is on PATH in this stage only
-# so the PyTorch compile step can invoke sccache without shipping it inside /usr/local/gcc133.
+# Update paths to use the compiler in C++17 mode. Keep /usr/local/bin on PATH in this stage so
+# the sccache binary (installed above) is visible during the PyTorch RUN steps; sccache stays
+# under /usr/local/bin (not under /usr/local/gcc133) so it is not copied into the final image with
+# the toolchain tree.
 ENV LD_LIBRARY_PATH=/usr/local/gcc133/lib64:/usr/local/gcc133/lib:/usr/lib:/lib
 ENV PATH=/usr/local/gcc133/bin:/usr/local/bin:/usr/bin:/bin:/usr/sbin:/sbin
 ENV CXX="g++ -std=gnu++17"
 
+# Install Intel MKL into gcc133 (same pattern as dev-tools/docker/linux_image/Dockerfile) so
+# libtorch_cpu.so resolves MKL at runtime in the final image after COPY --from=builder.
+RUN \
+  echo -e '[oneAPI]\n\
+name=Intel oneAPI repository\n\
+baseurl=https://yum.repos.intel.com/oneapi\n\
+enabled=1\n\
+gpgcheck=1\n\
+repo_gpgcheck=1\n\
+gpgkey=https://yum.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB\n' > /etc/yum.repos.d/oneAPI.repo && \
+  dnf install -y intel-oneapi-mkl-devel-2024.0 && \
+  (cd /opt/intel/oneapi/mkl/2024.0 && tar cf - lib) | (cd /usr/local/gcc133 && tar xvf -) && \
+  dnf clean all && \
+  rm -rf /var/cache/dnf /opt/intel/oneapi/mkl/2024.0/doc /tmp/*
+
 # Clone PyTorch and build LibTorch
 # PYTORCH_BUILD_VERSION is only set for tagged branches (e.g. v2.7.1);
 # for main/viable/strict PyTorch derives the version from version.txt.
@@ -100,6 +117,10 @@ RUN --mount=type=secret,id=gcs_key \
 
 FROM rockylinux:8
 COPY --from=builder /usr/local/gcc133 /usr/local/gcc133
+# Match linux_image final stage: MKL + libtorch under gcc133; needed for `import torch`
+# when running without the full compile-time shell (e.g. allowlist validation).
+ENV LD_LIBRARY_PATH=/usr/local/gcc133/lib64:/usr/local/gcc133/lib:/usr/lib:/lib
+ENV PATH=/usr/local/gcc133/bin:/usr/local/bin:/usr/bin:/bin:/usr/sbin:/sbin
 # Python 3.12 + torch site-packages for allowlist validation
 COPY --from=builder /usr/local/bin/python3.12 /usr/local/bin/python3.12
 COPY --from=builder /usr/local/bin/pip3.12 /usr/local/bin/pip3.12