diff --git a/.github/config/image/tensorflow-2.20-inference-sagemaker-cpu.yml b/.github/config/image/tensorflow-2.20-inference-sagemaker-cpu.yml
new file mode 100644
index 000000000000..0ab46d6128b1
--- /dev/null
+++ b/.github/config/image/tensorflow-2.20-inference-sagemaker-cpu.yml
@@ -0,0 +1,22 @@
+image:
+  name: "tensorflow-sagemaker-cpu"
+  description: "TensorFlow Serving 2.20 CPU inference for SageMaker"
+common:
+  framework: "tensorflow"
+  framework_version: "2.20.0"
+  job_type: "inference"
+  python_version: "py312"
+  os_version: "amzn2023"
+  customer_type: "sagemaker"
+  platform: "sagemaker"
+  arch_type: "x86"
+  prod_image: "tensorflow-inference:2.20-cpu-amzn2023-sagemaker"
+  device_type: "cpu"
+  contributor: "None"
+release:
+  release: true
+  force_release: false
+  public_registry: true
+  private_registry: true
+  enable_soci: true
+  environment: production
diff --git a/.github/config/image/tensorflow-2.20-inference-sagemaker-cuda.yml b/.github/config/image/tensorflow-2.20-inference-sagemaker-cuda.yml
new file mode 100644
index 000000000000..1a29dee0202f
--- /dev/null
+++ b/.github/config/image/tensorflow-2.20-inference-sagemaker-cuda.yml
@@ -0,0 +1,23 @@
+image:
+  name: "tensorflow-sagemaker-cuda"
+  description: "TensorFlow Serving 2.20 CUDA inference for SageMaker"
+common:
+  framework: "tensorflow"
+  framework_version: "2.20.0"
+  job_type: "inference"
+  python_version: "py312"
+  cuda_version: "cu129"
+  os_version: "amzn2023"
+  customer_type: "sagemaker"
+  platform: "sagemaker"
+  arch_type: "x86"
+  prod_image: "tensorflow-inference:2.20-cu129-amzn2023-sagemaker"
+  device_type: "gpu"
+  contributor: "None"
+release:
+  release: true
+  force_release: false
+  public_registry: true
+  private_registry: true
+  enable_soci: true
+  environment: production
diff --git a/.github/workflows/pr-tensorflow-inference-sagemaker-cpu.yml b/.github/workflows/pr-tensorflow-inference-sagemaker-cpu.yml
new file mode 100644
index 000000000000..bdbc5ecb4fbf
--- /dev/null
+++ b/.github/workflows/pr-tensorflow-inference-sagemaker-cpu.yml
@@ -0,0 +1,357 @@
+name: PR - TensorFlow Inference SageMaker CPU
+
+# Mirrors .github/workflows/pr-tensorflow-sagemaker-cpu.yml from the
+# tensorflow-2.21-currency branch (TF 2.21 training PR #6107). Inference deltas:
+#   - path triggers narrowed to docker/tensorflow/inference/** (the training
+#     workflow's globs would also fire on inference changes — flagged as a
+#     latent issue in the training workflow; not fixed here)
+#   - image config glob: tensorflow-*-inference-sagemaker-cpu.yml
+#   - version detection regex matches docker/tensorflow/inference/<v>/ paths
+#   - ECR tag uses tensorflow-inference (mirrors master TF 2.19 inference repo
+#     naming) and TF_SERVING_VERSION from versions-cpu.env (2.20.0) — note this
+#     is the TFS binary version, not the framework version
+#   - --build-args swapped: drop OPEN_MPI/TF_VERSION (training-only),
+#     add TF_SERVING_VERSION/TFS_SHORT_VERSION/NGINX_VERSION/NJS_VERSION
+#   - drop unit-test (no test/tensorflow/unit/ inference suite exists; Phase 5
+#     scope. TODO marker below if added later)
+#   - sagemaker-test points at test/tensorflow/integration/inference/ which
+#     Phase 5 will populate
+
+on:
+  pull_request:
+    branches: [main]
+    types: [opened, reopened, synchronize]
+    paths:
+      - ".github/config/image/tensorflow-*-inference-sagemaker-cpu.yml"
+      - ".github/workflows/pr-tensorflow-inference-sagemaker-cpu.yml"
+      - "docker/tensorflow/inference/*/Dockerfile.cpu"
+      - "docker/tensorflow/inference/*/cpu/**"
+      - "docker/tensorflow/inference/*/versions-cpu.env"
+      - "scripts/common/**"
+      - "scripts/tensorflow/inference/**"
+      - "scripts/telemetry/**"
+      - "test/tensorflow/integration/inference/**"
+      - "test/sanity/**"
+      - "test/telemetry/**"
+      - "!docs/**"
+
+permissions:
+  contents: read
+  pull-requests: read
+
+env:
+  FORCE_COLOR: "1"
+  # TFS version (binary), not framework version. See versions-cpu.env header.
+  LATEST_TENSORFLOW_INFERENCE_VERSION: "2.20"
+
+jobs:
+  # ============================================================
+  # Gate: permission check on base branch
+  # ============================================================
+  gatekeeper:
+    runs-on: ubuntu-latest
+    concurrency:
+      group: ${{ github.workflow }}-gate-${{ github.event.pull_request.number }}
+      cancel-in-progress: true
+    steps:
+      - name: Checkout base branch (safe)
+        uses: actions/checkout@v5
+        with:
+          ref: ${{ github.event.pull_request.base.sha }}
+          fetch-depth: 1
+
+      - name: Run permission gate (from base)
+        uses: ./.github/actions/pr-permission-gate
+
+  # ============================================================
+  # Detect changed TensorFlow inference versions + build per-version configs
+  # matrix + run pre-commit + path-based change detection
+  # ============================================================
+  check-changes:
+    needs: [gatekeeper]
+    if: success()
+    runs-on: ubuntu-latest
+    concurrency:
+      group: ${{ github.workflow }}-check-${{ github.event.pull_request.number }}
+      cancel-in-progress: true
+    outputs:
+      versions: ${{ steps.versions.outputs.versions }}
+      configs: ${{ steps.versions.outputs.configs }}
+      build-change: ${{ steps.changes.outputs.build-change }}
+      sanity-test-change: ${{ steps.changes.outputs.sanity-test-change }}
+      telemetry-test-change: ${{ steps.changes.outputs.telemetry-test-change }}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v5
+
+      - name: Setup python
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.12"
+
+      - name: Run pre-commit
+        uses: pre-commit/action@v3.0.1
+        with:
+          extra_args: --all-files
+
+      - name: Install yq
+        run: |
+          if ! command -v yq &> /dev/null; then
+            sudo wget -qO /usr/local/bin/yq https://github.com/mikefarah/yq/releases/latest/download/yq_linux_amd64
+            sudo chmod +x /usr/local/bin/yq
+          fi
+
+      - name: Detect TensorFlow inference versions and build configs matrix
+        id: versions
+        run: |
+          # Match either the docker path (docker/tensorflow/inference/2.20/...)
+          # or the config-file naming (tensorflow-2.20-inference-...).
+          VERSIONS=$(git diff --name-only origin/main...HEAD \
+            | grep -oP '(?:docker/tensorflow/inference/|tensorflow-)\K[0-9]+\.[0-9]+' \
+            | sort -u)
+          if [ -z "$VERSIONS" ]; then
+            VERSIONS="$LATEST_TENSORFLOW_INFERENCE_VERSION"
+          fi
+          JSON=$(echo "$VERSIONS" | jq -R -s -c 'split("\n") | map(select(length > 0))')
+          echo "versions=${JSON}" >> $GITHUB_OUTPUT
+          echo "Detected versions: ${JSON}"
+
+          # Build a configs matrix: each entry carries all metadata fields
+          CONFIGS="[]"
+          for V in $VERSIONS; do
+            CONFIG_FILE=".github/config/image/tensorflow-${V}-inference-sagemaker-cpu.yml"
+            if [ -f "$CONFIG_FILE" ]; then
+              CONFIGS=$(echo "$CONFIGS" | jq -c \
+                --arg v "$V" \
+                --arg fw "$(yq '.common.framework' $CONFIG_FILE)" \
+                --arg fwv "$(yq '.common.framework_version' $CONFIG_FILE)" \
+                --arg py "$(yq '.common.python_version' $CONFIG_FILE)" \
+                --arg cuda "$(yq '.common.cuda_version // ""' $CONFIG_FILE)" \
+                --arg os "$(yq '.common.os_version' $CONFIG_FILE)" \
+                --arg ct "$(yq '.common.job_type' $CONFIG_FILE)" \
+                --arg dt "$(yq '.common.device_type // "cpu"' $CONFIG_FILE)" \
+                --arg at "$(yq '.common.arch_type // "x86"' $CONFIG_FILE)" \
+                --arg contrib "$(yq '.common.contributor // "None"' $CONFIG_FILE)" \
+                --arg cust "$(yq '.common.customer_type // ""' $CONFIG_FILE)" \
+                --arg prod "$(yq '.common.prod_image' $CONFIG_FILE)" \
+                '. + [{"version": $v, "framework": $fw, "framework_version": $fwv, "python_version": $py, "cuda_version": $cuda, "os_version": $os, "container_type": $ct, "device_type": $dt, "arch_type": $at, "contributor": $contrib, "customer_type": $cust, "prod_image": $prod}]')
+            fi
+          done
+          echo "configs=${CONFIGS}" >> $GITHUB_OUTPUT
+          echo "Configs matrix: ${CONFIGS}"
+
+      - name: Detect file changes
+        id: changes
+        uses: dorny/paths-filter@v4
+        with:
+          filters: |
+            build-change:
+              - ".github/config/image/tensorflow-*-inference-sagemaker-cpu.yml"
+              - "docker/tensorflow/inference/*/Dockerfile.cpu"
+              - "docker/tensorflow/inference/*/cpu/**"
+              - "docker/tensorflow/inference/*/versions-cpu.env"
+              - "scripts/common/setup_oss_compliance.sh"
+              - "scripts/tensorflow/inference/**"
+              - "scripts/telemetry/bash_telemetry.sh.template"
+            sanity-test-change:
+              - "test/sanity/**"
+            telemetry-test-change:
+              - "test/telemetry/**"
+
+  # ============================================================
+  # Build CPU SageMaker inference images (matrix over detected versions)
+  # ============================================================
+  build-image:
+    needs: [check-changes]
+    if: needs.check-changes.outputs.build-change == 'true'
+    runs-on:
+      - codebuild-runner-${{ github.run_id }}-${{ github.run_attempt }}
+        fleet:default-runner
+        buildspec-override:true
+    strategy:
+      matrix:
+        include: ${{ fromJson(needs.check-changes.outputs.configs) }}
+      fail-fast: false
+    concurrency:
+      group: ${{ github.workflow }}-build-${{ matrix.version }}-${{ github.event.pull_request.number }}
+      cancel-in-progress: true
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v5
+
+      - name: Setup buildkitd
+        run: .github/scripts/buildkitd.sh
+
+      - name: ECR login
+        uses: ./.github/actions/ecr-authenticate
+        with:
+          aws-account-id: ${{ vars.CI_AWS_ACCOUNT_ID }}
+          aws-region: ${{ vars.AWS_REGION }}
+
+      - name: Build sagemaker inference image
+        id: build-sagemaker
+        run: |
+          VERSION="${{ matrix.version }}"
+          source docker/tensorflow/inference/${VERSION}/versions-cpu.env
+          # ECR tag uses TF Serving binary version (TF_SERVING_VERSION from
+          # versions-cpu.env) — mirrors master TF 2.19 inference repo naming.
+          CI_IMAGE_URI="${{ vars.CI_AWS_ACCOUNT_ID }}.dkr.ecr.${{ vars.AWS_REGION }}.amazonaws.com/ci:tensorflow-inference-${TF_SERVING_VERSION}-cpu-py312-sagemaker-${{ matrix.version }}-pr-${{ github.event.pull_request.number }}"
+
+          # Derive label values to match check_labels.py expectations
+          FRAMEWORK_LABEL=$(echo "${{ matrix.framework }}" | tr '_' '-')
+          FWK_VER_LABEL=$(echo "${{ matrix.framework_version }}" | tr '.' '-')
+          OS_LABEL=$(echo "${{ matrix.os_version }}" | tr '.' '-')
+
+          docker buildx build --progress plain \
+            --build-arg FRAMEWORK=${{ matrix.framework }} \
+            --build-arg PYTHON_VERSION=${PYTHON_VERSION} \
+            --build-arg TF_SERVING_VERSION=${TF_SERVING_VERSION} \
+            --build-arg TFS_SHORT_VERSION=${VERSION} \
+            --build-arg NGINX_VERSION=${NGINX_VERSION} \
+            --build-arg NJS_VERSION=${NJS_VERSION} \
+            --build-arg DLC_MAJOR_VERSION=${DLC_MAJOR_VERSION} \
+            --build-arg DLC_MINOR_VERSION=${DLC_MINOR_VERSION} \
+            --label "com.amazonaws.ml.engines.sagemaker.dlc.framework.${FRAMEWORK_LABEL}.${FWK_VER_LABEL}=true" \
+            --label "com.amazonaws.ml.engines.sagemaker.dlc.device.cpu=true" \
+            --label "com.amazonaws.ml.engines.sagemaker.dlc.job.${{ matrix.container_type }}=true" \
+            --label "com.amazonaws.ml.engines.sagemaker.dlc.arch.${{ matrix.arch_type }}=true" \
+            --label "com.amazonaws.ml.engines.sagemaker.dlc.os.${OS_LABEL}=true" \
+            --label "com.amazonaws.ml.engines.sagemaker.dlc.python.${{ matrix.python_version }}=true" \
+            --label "com.amazonaws.ml.engines.sagemaker.dlc.contributor.${{ matrix.contributor }}=true" \
+            --cache-to=type=inline \
+            --cache-from=type=registry,ref=${CI_IMAGE_URI} \
+            --tag ${CI_IMAGE_URI} \
+            --push \
+            --target sagemaker \
+            -f docker/tensorflow/inference/${VERSION}/Dockerfile.cpu .
+
+          echo "image-uri=${CI_IMAGE_URI}" >> $GITHUB_OUTPUT
+
+  # ============================================================
+  # Sanity tests
+  # ============================================================
+  sanity-test:
+    needs: [check-changes, build-image]
+    if: |
+      always() && !failure() && !cancelled() &&
+      (needs.check-changes.outputs.build-change == 'true' || needs.check-changes.outputs.sanity-test-change == 'true')
+    strategy:
+      matrix:
+        include: ${{ fromJson(needs.check-changes.outputs.configs) }}
+      fail-fast: false
+    uses: ./.github/workflows/reusable-sanity-tests.yml
+    with:
+      image-uri: ${{ needs.build-image.result == 'success' && format('{0}.dkr.ecr.{1}.amazonaws.com/ci:tensorflow-inference-{2}-cpu-py312-sagemaker-{3}-pr-{4}', vars.CI_AWS_ACCOUNT_ID, vars.AWS_REGION, matrix.framework_version, matrix.version, github.event.pull_request.number) || format('{0}.dkr.ecr.{1}.amazonaws.com/{2}', vars.PROD_AWS_ACCOUNT_ID, vars.AWS_REGION, matrix.prod_image) }}
+      aws-account-id: ${{ needs.build-image.result == 'success' && vars.CI_AWS_ACCOUNT_ID || vars.PROD_AWS_ACCOUNT_ID }}
+      aws-region: ${{ vars.AWS_REGION }}
+      framework: ${{ matrix.framework }}
+      framework-version: ${{ matrix.framework_version }}
+      python-version: ${{ matrix.python_version }}
+      cuda-version: ${{ matrix.cuda_version }}
+      os-version: ${{ matrix.os_version }}
+      customer-type: ${{ matrix.customer_type }}
+      arch-type: ${{ matrix.arch_type }}
+      device-type: ${{ matrix.device_type }}
+      contributor: ${{ matrix.contributor }}
+      container-type: ${{ matrix.container_type }}
+
+  # ============================================================
+  # Security tests
+  # ============================================================
+  security-test:
+    needs: [check-changes, build-image]
+    if: success()
+    strategy:
+      matrix:
+        include: ${{ fromJson(needs.check-changes.outputs.configs) }}
+      fail-fast: false
+    uses: ./.github/workflows/reusable-security-tests.yml
+    with:
+      image-uri: ${{ needs.build-image.result == 'success' && format('{0}.dkr.ecr.{1}.amazonaws.com/ci:tensorflow-inference-{2}-cpu-py312-sagemaker-{3}-pr-{4}', vars.CI_AWS_ACCOUNT_ID, vars.AWS_REGION, matrix.framework_version, matrix.version, github.event.pull_request.number) || format('{0}.dkr.ecr.{1}.amazonaws.com/{2}', vars.PROD_AWS_ACCOUNT_ID, vars.AWS_REGION, matrix.prod_image) }}
+      aws-account-id: ${{ needs.build-image.result == 'success' && vars.CI_AWS_ACCOUNT_ID || vars.PROD_AWS_ACCOUNT_ID }}
+      aws-region: ${{ vars.AWS_REGION }}
+      framework: ${{ matrix.framework }}
+      framework-version: ${{ matrix.framework_version }}
+
+  # ============================================================
+  # Telemetry tests
+  # ============================================================
+  telemetry-test:
+    needs: [check-changes, build-image]
+    if: |
+      always() && !failure() && !cancelled() &&
+      (needs.check-changes.outputs.build-change == 'true' || needs.check-changes.outputs.telemetry-test-change == 'true')
+    strategy:
+      matrix:
+        include: ${{ fromJson(needs.check-changes.outputs.configs) }}
+      fail-fast: false
+    concurrency:
+      group: ${{ github.workflow }}-telemetry-test-${{ matrix.version }}-${{ github.event.pull_request.number }}
+      cancel-in-progress: false
+    uses: ./.github/workflows/reusable-telemetry-tests.yml
+    with:
+      image-uri: ${{ needs.build-image.result == 'success' && format('{0}.dkr.ecr.{1}.amazonaws.com/ci:tensorflow-inference-{2}-cpu-py312-sagemaker-{3}-pr-{4}', vars.CI_AWS_ACCOUNT_ID, vars.AWS_REGION, matrix.framework_version, matrix.version, github.event.pull_request.number) || format('{0}.dkr.ecr.{1}.amazonaws.com/{2}', vars.PROD_AWS_ACCOUNT_ID, vars.AWS_REGION, matrix.prod_image) }}
+      aws-account-id: ${{ needs.build-image.result == 'success' && vars.CI_AWS_ACCOUNT_ID || vars.PROD_AWS_ACCOUNT_ID }}
+      aws-region: ${{ vars.AWS_REGION }}
+      framework: ${{ matrix.framework }}
+      framework-version: ${{ matrix.framework_version }}
+      container-type: ${{ matrix.container_type }}
+
+  # TODO: unit-test job — training has one targeting test/tensorflow/unit/.
+  #       Inference equivalent does not exist yet. Add in Phase 5 if a unit
+  #       suite is created.
+
+  # TODO: MME-specific test — image config sets
+  #       com.amazonaws.sagemaker.capabilities.multi-models=true. Phase 5 should
+  #       add an MME endpoint smoke test once the integration test scaffolding
+  #       is in place.
+
+  # ============================================================
+  # SageMaker integration tests (launch real CPU SM endpoints)
+  # ============================================================
+  # Phase 5 will populate test/tensorflow/integration/inference/. Until then this
+  # job will run pytest on an empty dir (pytest exits 5 = no tests collected; we
+  # treat that as success here so the workflow can land before Phase 5).
+  sagemaker-test:
+    needs: [check-changes, build-image, sanity-test, security-test]
+    if: success()
+    runs-on:
+      - codebuild-runner-${{ github.run_id }}-${{ github.run_attempt }}
+        fleet:default-runner
+        buildspec-override:true
+    strategy:
+      matrix:
+        include: ${{ fromJson(needs.check-changes.outputs.configs) }}
+      fail-fast: false
+    concurrency:
+      group: ${{ github.workflow }}-sagemaker-${{ matrix.version }}-${{ github.event.pull_request.number }}
+      cancel-in-progress: true
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v5
+
+      - name: ECR login
+        uses: ./.github/actions/ecr-authenticate
+        with:
+          aws-account-id: ${{ vars.CI_AWS_ACCOUNT_ID }}
+          aws-region: ${{ vars.AWS_REGION }}
+
+      - name: Install test dependencies
+        run: |
+          pip install -r test/tensorflow/integration/inference/requirements.txt
+
+      - name: Run SageMaker inference tests
+        env:
+          PYTHONPATH: ${{ github.workspace }}/test
+          TEST_IMAGE_URI: ${{ format('{0}.dkr.ecr.{1}.amazonaws.com/ci:tensorflow-inference-{2}-cpu-py312-sagemaker-{3}-pr-{4}', vars.CI_AWS_ACCOUNT_ID, vars.AWS_REGION, matrix.framework_version, matrix.version, github.event.pull_request.number) }}
+          SM_ROLE_ARN: arn:aws:iam::${{ vars.CI_AWS_ACCOUNT_ID }}:role/SageMakerRole
+        run: |
+          # pytest exit 5 (no tests collected) is acceptable until Phase 5 adds
+          # tests under test/tensorflow/integration/inference/.
+          set +e
+          pytest test/tensorflow/integration/inference/ -v
+          rc=$?
+          set -e
+          if [ "$rc" != "0" ] && [ "$rc" != "5" ]; then
+            exit $rc
+          fi
diff --git a/.github/workflows/pr-tensorflow-inference-sagemaker-cuda.yml b/.github/workflows/pr-tensorflow-inference-sagemaker-cuda.yml
new file mode 100644
index 000000000000..c387d8f1701d
--- /dev/null
+++ b/.github/workflows/pr-tensorflow-inference-sagemaker-cuda.yml
@@ -0,0 +1,363 @@
+name: PR - TensorFlow Inference SageMaker CUDA
+
+# Mirrors .github/workflows/pr-tensorflow-sagemaker-cuda.yml from the
+# tensorflow-2.21-currency branch (TF 2.21 training PR #6107). Inference deltas:
+#   - path triggers narrowed to docker/tensorflow/inference/** (the training
+#     workflow's globs would also fire on inference changes — flagged as a
+#     latent issue in the training workflow; not fixed here)
+#   - image config glob: tensorflow-*-inference-sagemaker-cuda.yml
+#   - version detection regex matches docker/tensorflow/inference/<v>/ paths
+#   - ECR tag uses tensorflow-inference (mirrors master TF 2.19 inference repo
+#     naming) and TF_SERVING_VERSION from versions-cuda.env (2.20.0) — note this
+#     is the TFS binary version, not the framework version
+#   - --build-args swapped: drop NCCL/EFA/OpenMPI/TF_VERSION (training-only),
+#     add TF_SERVING_VERSION/TFS_SHORT_VERSION/NGINX_VERSION/NJS_VERSION
+#   - drop single-gpu-test (no in-container GPU smoke for an inference image —
+#     sanity already covers serving smoke; full SM endpoint test runs on real GPU)
+#   - drop unit-test (no test/tensorflow/unit/ inference suite exists; Phase 5
+#     scope. TODO marker below if added later)
+#   - sagemaker-test points at test/tensorflow/integration/inference/ which
+#     Phase 5 will populate
+
+on:
+  pull_request:
+    branches: [main]
+    types: [opened, reopened, synchronize]
+    paths:
+      - ".github/config/image/tensorflow-*-inference-sagemaker-cuda.yml"
+      - ".github/workflows/pr-tensorflow-inference-sagemaker-cuda.yml"
+      - "docker/tensorflow/inference/*/Dockerfile.cuda"
+      - "docker/tensorflow/inference/*/cuda/**"
+      - "docker/tensorflow/inference/*/versions-cuda.env"
+      - "scripts/common/**"
+      - "scripts/tensorflow/inference/**"
+      - "scripts/telemetry/**"
+      - "test/tensorflow/integration/inference/**"
+      - "test/sanity/**"
+      - "test/telemetry/**"
+      - "!docs/**"
+
+permissions:
+  contents: read
+  pull-requests: read
+
+env:
+  FORCE_COLOR: "1"
+  # TFS version (binary), not framework version. See versions-cuda.env header.
+  LATEST_TENSORFLOW_INFERENCE_VERSION: "2.20"
+
+jobs:
+  # ============================================================
+  # Gate: permission check on base branch
+  # ============================================================
+  gatekeeper:
+    runs-on: ubuntu-latest
+    concurrency:
+      group: ${{ github.workflow }}-gate-${{ github.event.pull_request.number }}
+      cancel-in-progress: true
+    steps:
+      - name: Checkout base branch (safe)
+        uses: actions/checkout@v5
+        with:
+          ref: ${{ github.event.pull_request.base.sha }}
+          fetch-depth: 1
+
+      - name: Run permission gate (from base)
+        uses: ./.github/actions/pr-permission-gate
+
+  # ============================================================
+  # Detect changed TensorFlow inference versions + build per-version configs
+  # matrix + run pre-commit + path-based change detection
+  # ============================================================
+  check-changes:
+    needs: [gatekeeper]
+    if: success()
+    runs-on: ubuntu-latest
+    concurrency:
+      group: ${{ github.workflow }}-check-${{ github.event.pull_request.number }}
+      cancel-in-progress: true
+    outputs:
+      versions: ${{ steps.versions.outputs.versions }}
+      configs: ${{ steps.versions.outputs.configs }}
+      build-change: ${{ steps.changes.outputs.build-change }}
+      sanity-test-change: ${{ steps.changes.outputs.sanity-test-change }}
+      telemetry-test-change: ${{ steps.changes.outputs.telemetry-test-change }}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v5
+
+      - name: Setup python
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.12"
+
+      - name: Run pre-commit
+        uses: pre-commit/action@v3.0.1
+        with:
+          extra_args: --all-files
+
+      - name: Install yq
+        run: |
+          if ! command -v yq &> /dev/null; then
+            sudo wget -qO /usr/local/bin/yq https://github.com/mikefarah/yq/releases/latest/download/yq_linux_amd64
+            sudo chmod +x /usr/local/bin/yq
+          fi
+
+      - name: Detect TensorFlow inference versions and build configs matrix
+        id: versions
+        run: |
+          # Match either the docker path (docker/tensorflow/inference/2.20/...)
+          # or the config-file naming (tensorflow-2.20-inference-...).
+          VERSIONS=$(git diff --name-only origin/main...HEAD \
+            | grep -oP '(?:docker/tensorflow/inference/|tensorflow-)\K[0-9]+\.[0-9]+' \
+            | sort -u)
+          if [ -z "$VERSIONS" ]; then
+            VERSIONS="$LATEST_TENSORFLOW_INFERENCE_VERSION"
+          fi
+          JSON=$(echo "$VERSIONS" | jq -R -s -c 'split("\n") | map(select(length > 0))')
+          echo "versions=${JSON}" >> $GITHUB_OUTPUT
+          echo "Detected versions: ${JSON}"
+
+          # Build a configs matrix: each entry carries all metadata fields
+          CONFIGS="[]"
+          for V in $VERSIONS; do
+            CONFIG_FILE=".github/config/image/tensorflow-${V}-inference-sagemaker-cuda.yml"
+            if [ -f "$CONFIG_FILE" ]; then
+              CONFIGS=$(echo "$CONFIGS" | jq -c \
+                --arg v "$V" \
+                --arg fw "$(yq '.common.framework' $CONFIG_FILE)" \
+                --arg fwv "$(yq '.common.framework_version' $CONFIG_FILE)" \
+                --arg py "$(yq '.common.python_version' $CONFIG_FILE)" \
+                --arg cuda "$(yq '.common.cuda_version' $CONFIG_FILE)" \
+                --arg os "$(yq '.common.os_version' $CONFIG_FILE)" \
+                --arg ct "$(yq '.common.job_type' $CONFIG_FILE)" \
+                --arg dt "$(yq '.common.device_type // "gpu"' $CONFIG_FILE)" \
+                --arg at "$(yq '.common.arch_type // "x86"' $CONFIG_FILE)" \
+                --arg contrib "$(yq '.common.contributor // "None"' $CONFIG_FILE)" \
+                --arg cust "$(yq '.common.customer_type // ""' $CONFIG_FILE)" \
+                --arg prod "$(yq '.common.prod_image' $CONFIG_FILE)" \
+                '. + [{"version": $v, "framework": $fw, "framework_version": $fwv, "python_version": $py, "cuda_version": $cuda, "os_version": $os, "container_type": $ct, "device_type": $dt, "arch_type": $at, "contributor": $contrib, "customer_type": $cust, "prod_image": $prod}]')
+            fi
+          done
+          echo "configs=${CONFIGS}" >> $GITHUB_OUTPUT
+          echo "Configs matrix: ${CONFIGS}"
+
+      - name: Detect file changes
+        id: changes
+        uses: dorny/paths-filter@v4
+        with:
+          filters: |
+            build-change:
+              - ".github/config/image/tensorflow-*-inference-sagemaker-cuda.yml"
+              - "docker/tensorflow/inference/*/Dockerfile.cuda"
+              - "docker/tensorflow/inference/*/cuda/**"
+              - "docker/tensorflow/inference/*/versions-cuda.env"
+              - "scripts/common/setup_oss_compliance.sh"
+              - "scripts/common/start_cuda_compat.sh"
+              - "scripts/tensorflow/inference/**"
+              - "scripts/telemetry/bash_telemetry.sh.template"
+            sanity-test-change:
+              - "test/sanity/**"
+            telemetry-test-change:
+              - "test/telemetry/**"
+
+  # ============================================================
+  # Build SageMaker inference images (matrix over detected versions)
+  # ============================================================
+  build-image:
+    needs: [check-changes]
+    if: needs.check-changes.outputs.build-change == 'true'
+    runs-on:
+      - codebuild-runner-${{ github.run_id }}-${{ github.run_attempt }}
+        fleet:x86-build-runner
+        buildspec-override:true
+    strategy:
+      matrix:
+        include: ${{ fromJson(needs.check-changes.outputs.configs) }}
+      fail-fast: false
+    concurrency:
+      group: ${{ github.workflow }}-build-${{ matrix.version }}-${{ github.event.pull_request.number }}
+      cancel-in-progress: true
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v5
+
+      - name: Setup buildkitd
+        run: .github/scripts/buildkitd.sh
+
+      - name: ECR login
+        uses: ./.github/actions/ecr-authenticate
+        with:
+          aws-account-id: ${{ vars.CI_AWS_ACCOUNT_ID }}
+          aws-region: ${{ vars.AWS_REGION }}
+
+      - name: Build sagemaker inference image
+        id: build-sagemaker
+        run: |
+          VERSION="${{ matrix.version }}"
+          source docker/tensorflow/inference/${VERSION}/versions-cuda.env
+          # ECR tag uses TF Serving binary version (TF_SERVING_VERSION from
+          # versions-cuda.env) — mirrors master TF 2.19 inference repo naming.
+          CI_IMAGE_URI="${{ vars.CI_AWS_ACCOUNT_ID }}.dkr.ecr.${{ vars.AWS_REGION }}.amazonaws.com/ci:tensorflow-inference-${TF_SERVING_VERSION}-gpu-py312-cu129-sagemaker-${{ matrix.version }}-pr-${{ github.event.pull_request.number }}"
+
+          # Derive label values to match check_labels.py expectations
+          FRAMEWORK_LABEL=$(echo "${{ matrix.framework }}" | tr '_' '-')
+          FWK_VER_LABEL=$(echo "${{ matrix.framework_version }}" | tr '.' '-')
+          CUDA_LABEL="${{ matrix.cuda_version }}"
+          OS_LABEL=$(echo "${{ matrix.os_version }}" | tr '.' '-')
+
+          docker buildx build --progress plain \
+            --build-arg FRAMEWORK=${{ matrix.framework }} \
+            --build-arg CUDA_VERSION=${CUDA_VERSION} \
+            --build-arg PYTHON_VERSION=${PYTHON_VERSION} \
+            --build-arg TF_SERVING_VERSION=${TF_SERVING_VERSION} \
+            --build-arg TFS_SHORT_VERSION=${VERSION} \
+            --build-arg NGINX_VERSION=${NGINX_VERSION} \
+            --build-arg NJS_VERSION=${NJS_VERSION} \
+            --build-arg DLC_MAJOR_VERSION=${DLC_MAJOR_VERSION} \
+            --build-arg DLC_MINOR_VERSION=${DLC_MINOR_VERSION} \
+            --label "com.amazonaws.ml.engines.sagemaker.dlc.framework.${FRAMEWORK_LABEL}.${FWK_VER_LABEL}=true" \
+            --label "com.amazonaws.ml.engines.sagemaker.dlc.device.gpu.${CUDA_LABEL}=true" \
+            --label "com.amazonaws.ml.engines.sagemaker.dlc.job.${{ matrix.container_type }}=true" \
+            --label "com.amazonaws.ml.engines.sagemaker.dlc.arch.${{ matrix.arch_type }}=true" \
+            --label "com.amazonaws.ml.engines.sagemaker.dlc.os.${OS_LABEL}=true" \
+            --label "com.amazonaws.ml.engines.sagemaker.dlc.python.${{ matrix.python_version }}=true" \
+            --label "com.amazonaws.ml.engines.sagemaker.dlc.contributor.${{ matrix.contributor }}=true" \
+            --cache-to=type=inline \
+            --cache-from=type=registry,ref=${CI_IMAGE_URI} \
+            --tag ${CI_IMAGE_URI} \
+            --push \
+            --target sagemaker \
+            -f docker/tensorflow/inference/${VERSION}/Dockerfile.cuda .
+
+          echo "image-uri=${CI_IMAGE_URI}" >> $GITHUB_OUTPUT
+
+  # ============================================================
+  # Sanity tests (labels, filesystem, OSS compliance)
+  # ============================================================
+  sanity-test:
+    needs: [check-changes, build-image]
+    if: |
+      always() && !failure() && !cancelled() &&
+      (needs.check-changes.outputs.build-change == 'true' || needs.check-changes.outputs.sanity-test-change == 'true')
+    strategy:
+      matrix:
+        include: ${{ fromJson(needs.check-changes.outputs.configs) }}
+      fail-fast: false
+    uses: ./.github/workflows/reusable-sanity-tests.yml
+    with:
+      image-uri: ${{ needs.build-image.result == 'success' && format('{0}.dkr.ecr.{1}.amazonaws.com/ci:tensorflow-inference-{2}-gpu-py312-cu129-sagemaker-{3}-pr-{4}', vars.CI_AWS_ACCOUNT_ID, vars.AWS_REGION, matrix.framework_version, matrix.version, github.event.pull_request.number) || format('{0}.dkr.ecr.{1}.amazonaws.com/{2}', vars.PROD_AWS_ACCOUNT_ID, vars.AWS_REGION, matrix.prod_image) }}
+      aws-account-id: ${{ needs.build-image.result == 'success' && vars.CI_AWS_ACCOUNT_ID || vars.PROD_AWS_ACCOUNT_ID }}
+      aws-region: ${{ vars.AWS_REGION }}
+      framework: ${{ matrix.framework }}
+      framework-version: ${{ matrix.framework_version }}
+      python-version: ${{ matrix.python_version }}
+      cuda-version: ${{ matrix.cuda_version }}
+      os-version: ${{ matrix.os_version }}
+      customer-type: ${{ matrix.customer_type }}
+      arch-type: ${{ matrix.arch_type }}
+      device-type: ${{ matrix.device_type }}
+      contributor: ${{ matrix.contributor }}
+      container-type: ${{ matrix.container_type }}
+
+  # ============================================================
+  # Security tests (ECR scan, CVE allowlist)
+  # ============================================================
+  security-test:
+    needs: [check-changes, build-image]
+    if: success()
+    strategy:
+      matrix:
+        include: ${{ fromJson(needs.check-changes.outputs.configs) }}
+      fail-fast: false
+    uses: ./.github/workflows/reusable-security-tests.yml
+    with:
+      image-uri: ${{ needs.build-image.result == 'success' && format('{0}.dkr.ecr.{1}.amazonaws.com/ci:tensorflow-inference-{2}-gpu-py312-cu129-sagemaker-{3}-pr-{4}', vars.CI_AWS_ACCOUNT_ID, vars.AWS_REGION, matrix.framework_version, matrix.version, github.event.pull_request.number) || format('{0}.dkr.ecr.{1}.amazonaws.com/{2}', vars.PROD_AWS_ACCOUNT_ID, vars.AWS_REGION, matrix.prod_image) }}
+      aws-account-id: ${{ needs.build-image.result == 'success' && vars.CI_AWS_ACCOUNT_ID || vars.PROD_AWS_ACCOUNT_ID }}
+      aws-region: ${{ vars.AWS_REGION }}
+      framework: ${{ matrix.framework }}
+      framework-version: ${{ matrix.framework_version }}
+
+  # ============================================================
+  # Telemetry tests (opt-out, environment variables)
+  # ============================================================
+  telemetry-test:
+    needs: [check-changes, build-image]
+    if: |
+      always() && !failure() && !cancelled() &&
+      (needs.check-changes.outputs.build-change == 'true' || needs.check-changes.outputs.telemetry-test-change == 'true')
+    strategy:
+      matrix:
+        include: ${{ fromJson(needs.check-changes.outputs.configs) }}
+      fail-fast: false
+    concurrency:
+      group: ${{ github.workflow }}-telemetry-test-${{ matrix.version }}-${{ github.event.pull_request.number }}
+      cancel-in-progress: false
+    uses: ./.github/workflows/reusable-telemetry-tests.yml
+    with:
+      image-uri: ${{ needs.build-image.result == 'success' && format('{0}.dkr.ecr.{1}.amazonaws.com/ci:tensorflow-inference-{2}-gpu-py312-cu129-sagemaker-{3}-pr-{4}', vars.CI_AWS_ACCOUNT_ID, vars.AWS_REGION, matrix.framework_version, matrix.version, github.event.pull_request.number) || format('{0}.dkr.ecr.{1}.amazonaws.com/{2}', vars.PROD_AWS_ACCOUNT_ID, vars.AWS_REGION, matrix.prod_image) }}
+      aws-account-id: ${{ needs.build-image.result == 'success' && vars.CI_AWS_ACCOUNT_ID || vars.PROD_AWS_ACCOUNT_ID }}
+      aws-region: ${{ vars.AWS_REGION }}
+      framework: ${{ matrix.framework }}
+      framework-version: ${{ matrix.framework_version }}
+      container-type: ${{ matrix.container_type }}
+
+  # TODO: unit-test job — training has one targeting test/tensorflow/unit/.
+  #       Inference equivalent (e.g., test/tensorflow/integration/inference/unit/
+  #       exercising tfs_utils / multi_model_utils in-container) does not exist
+  #       yet. Add in Phase 5 if a unit suite is created.
+
+  # TODO: MME-specific test — image config sets
+  #       com.amazonaws.sagemaker.capabilities.multi-models=true. Phase 5 should
+  #       add an MME endpoint smoke test (multi-model load via tensorflowServing.js)
+  #       once the integration test scaffolding is in place.
+
+  # ============================================================
+  # SageMaker integration tests (launch real SM endpoints on GPU)
+  # ============================================================
+  # Phase 5 will populate test/tensorflow/integration/inference/. Until then this
+  # job will run pytest on an empty dir (pytest exits 5 = no tests collected; we
+  # treat that as success here so the workflow can land before Phase 5).
+  sagemaker-test:
+    needs: [check-changes, build-image, sanity-test, security-test]
+    if: success()
+    runs-on:
+      - codebuild-runner-${{ github.run_id }}-${{ github.run_attempt }}
+        fleet:default-runner
+        buildspec-override:true
+    strategy:
+      matrix:
+        include: ${{ fromJson(needs.check-changes.outputs.configs) }}
+      fail-fast: false
+    concurrency:
+      group: ${{ github.workflow }}-sagemaker-${{ matrix.version }}-${{ github.event.pull_request.number }}
+      cancel-in-progress: true
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v5
+
+      - name: ECR login
+        uses: ./.github/actions/ecr-authenticate
+        with:
+          aws-account-id: ${{ vars.CI_AWS_ACCOUNT_ID }}
+          aws-region: ${{ vars.AWS_REGION }}
+
+      - name: Install test dependencies
+        run: |
+          pip install -r test/tensorflow/integration/inference/requirements.txt
+
+      - name: Run SageMaker inference tests
+        env:
+          PYTHONPATH: ${{ github.workspace }}/test
+          TEST_IMAGE_URI: ${{ format('{0}.dkr.ecr.{1}.amazonaws.com/ci:tensorflow-inference-{2}-gpu-py312-cu129-sagemaker-{3}-pr-{4}', vars.CI_AWS_ACCOUNT_ID, vars.AWS_REGION, matrix.framework_version, matrix.version, github.event.pull_request.number) }}
+          SM_ROLE_ARN: arn:aws:iam::${{ vars.CI_AWS_ACCOUNT_ID }}:role/SageMakerRole
+        run: |
+          # pytest exit 5 (no tests collected) is acceptable until Phase 5 adds
+          # tests under test/tensorflow/integration/inference/.
+          set +e
+          pytest test/tensorflow/integration/inference/ -v
+          rc=$?
+          set -e
+          if [ "$rc" != "0" ] && [ "$rc" != "5" ]; then
+            exit $rc
+          fi
diff --git a/docker/tensorflow/inference/2.20/Dockerfile.cpu b/docker/tensorflow/inference/2.20/Dockerfile.cpu
new file mode 100644
index 000000000000..356bf3e81e3f
--- /dev/null
+++ b/docker/tensorflow/inference/2.20/Dockerfile.cpu
@@ -0,0 +1,225 @@
+# ============================================================================
+# TensorFlow Serving 2.20 Inference DLC — Amazon Linux 2023 (CPU)
+# Multi-stage build:
+#   builder-base ──┬── builder-oss   (OSS license generation — isolated)
+#                  └── builder-njs   (compile nginx + njs dynamic module)
+#   runtime-base ──── sagemaker      (SageMaker inference, MME-capable)
+#
+# All version defaults mirror docker/tensorflow/inference/2.20/versions-cpu.env.
+# Workflows source versions-cpu.env and pass ARGs via --build-arg — single
+# source of truth.
+#
+# Same structure as Dockerfile.cuda; deltas:
+#   - amazonlinux:2023 base (no CUDA).
+#   - tensorflow/serving:${TF_SERVING_VERSION}-devel (no -gpu) for binary copy.
+#   - tensorflow-serving-api (no -gpu suffix) installed via --no-deps.
+# ============================================================================
+
+# ── Global ARGs (available to all stages) ───────────────────────────────────
+ARG DLC_MAJOR_VERSION=1
+ARG DLC_MINOR_VERSION=0
+ARG PYTHON_VERSION=3.12
+ARG TF_SERVING_VERSION=2.20.0
+ARG TFS_SHORT_VERSION=2.20
+ARG NGINX_VERSION=1.30.2
+ARG NJS_VERSION=0.9.9
+
+
+# ── Stage: build_image (TFS upstream binary source) ──────────────────────────
+FROM tensorflow/serving:${TF_SERVING_VERSION}-devel AS build_image
+
+
+# ── Stage: builder-base (Python venv + lockfile deps) ───────────────────────
+FROM amazonlinux:2023 AS builder-base
+ARG PYTHON_VERSION
+
+RUN dnf install -y --allowerasing \
+  python${PYTHON_VERSION} python${PYTHON_VERSION}-devel python${PYTHON_VERSION}-pip \
+  gcc gcc-c++ make cmake git openssl-devel ninja-build \
+  tar xz curl wget \
+  && dnf clean all
+
+COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
+
+ENV UV_PROJECT_ENVIRONMENT="/opt/venv"
+RUN python${PYTHON_VERSION} -m venv /opt/venv
+ENV PATH="/opt/venv/bin:${PATH}"
+
+COPY docker/tensorflow/inference/2.20/cpu/pyproject.toml docker/tensorflow/inference/2.20/cpu/uv.lock /tmp/build/
+WORKDIR /tmp/build
+RUN --mount=type=cache,target=/root/.cache/uv uv sync --frozen --no-dev --no-install-project --inexact
+
+
+# ── Stage: builder-oss (generates license files in isolation) ───────────────
+FROM amazonlinux:2023 AS builder-oss
+ARG PYTHON_VERSION
+RUN dnf install -y --allowerasing python${PYTHON_VERSION} curl && dnf clean all
+COPY --from=builder-base /opt/venv /opt/venv
+COPY scripts/common/setup_oss_compliance.sh /tmp/setup_oss_compliance.sh
+RUN PATH="/opt/venv/bin:${PATH}" bash /tmp/setup_oss_compliance.sh python${PYTHON_VERSION} \
+  && touch /root/THIRD_PARTY_SOURCE_CODE_URLS
+
+
+# ── Stage: builder-njs (compile nginx + njs dynamic module from source) ─────
+# Output: /out/modules/ngx_http_js_module.so
+# Identical to the cuda variant's builder-njs — njs build is CPU-only and
+# produces an architecture-matched .so that loads into the AL2023 nginx.
+FROM amazonlinux:2023 AS builder-njs
+ARG NGINX_VERSION
+ARG NJS_VERSION
+
+RUN dnf install -y --allowerasing \
+  gcc gcc-c++ make \
+  pcre-devel pcre2-devel zlib-devel openssl-devel libxml2-devel libxslt-devel \
+  curl tar gzip xz \
+  && dnf clean all
+
+WORKDIR /tmp/njs-build
+
+RUN curl -fsSLO "https://nginx.org/download/nginx-${NGINX_VERSION}.tar.gz" \
+  && tar xzf nginx-${NGINX_VERSION}.tar.gz \
+  && curl -fsSL "https://github.com/nginx/njs/archive/refs/tags/${NJS_VERSION}.tar.gz" -o njs-${NJS_VERSION}.tar.gz \
+  && tar xzf njs-${NJS_VERSION}.tar.gz
+
+RUN cd nginx-${NGINX_VERSION} \
+  && ./configure \
+    --with-compat \
+    --add-dynamic-module=/tmp/njs-build/njs-${NJS_VERSION}/nginx \
+  && make -j$(nproc) modules \
+  && mkdir -p /out/modules \
+  && cp objs/ngx_http_js_module.so /out/modules/
+
+
+# ── Stage: runtime-base (shared base for output stages) ─────────────────────
+FROM amazonlinux:2023 AS runtime-base
+ARG PYTHON_VERSION
+ARG TF_SERVING_VERSION
+ARG TFS_SHORT_VERSION
+ARG DLC_MAJOR_VERSION
+ARG DLC_MINOR_VERSION
+
+# Labels live on runtime-base so all output stages inherit them.
+LABEL maintainer="Amazon AI"
+LABEL dlc_major_version="${DLC_MAJOR_VERSION}"
+LABEL dlc_minor_version="${DLC_MINOR_VERSION}"
+LABEL framework="tensorflow"
+LABEL framework_version="${TF_SERVING_VERSION}"
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+  PYTHONUNBUFFERED=1 \
+  PYTHONIOENCODING=UTF-8 \
+  LANG=C.UTF-8 \
+  LC_ALL=C.UTF-8 \
+  DLC_CONTAINER_TYPE=inference \
+  MODEL_BASE_PATH=/models \
+  MODEL_NAME=model
+
+# Runtime system deps via dnf — equivalent of master TF 2.19 CPU apt-get block.
+RUN dnf install -y --allowerasing \
+  python${PYTHON_VERSION} python${PYTHON_VERSION}-pip python${PYTHON_VERSION}-devel \
+  nginx \
+  gcc gcc-c++ make git \
+  tar gzip xz which findutils util-linux \
+  libpng-devel freetype-devel zlib-devel \
+  openssl unzip jq curl wget \
+  && dnf clean all
+
+# Copy venv from builder-base
+COPY --from=builder-base /opt/venv /opt/venv
+
+# Copy njs dynamic modules
+COPY --from=builder-njs /out/modules/ngx_http_js_module.so /usr/lib64/nginx/modules/
+
+# TF Serving binary
+COPY --from=build_image /usr/local/bin/tensorflow_model_server /usr/local/bin/tensorflow_model_server
+
+# python symlink — some TF tooling expects /usr/local/bin/python.
+RUN ln -sf $(which python${PYTHON_VERSION}) /usr/local/bin/python \
+  && ln -sf $(which pip3) /usr/local/bin/pip
+
+ENV PATH="/opt/venv/bin:${PATH}"
+ENV LD_LIBRARY_PATH="/usr/local/lib:${LD_LIBRARY_PATH}"
+
+# Models dir
+RUN mkdir -p ${MODEL_BASE_PATH}
+
+# License file (S3 bucket pre-provisioned for tensorflow-2.20)
+RUN curl -fsSLo /license.txt "https://aws-dlc-licenses.s3.amazonaws.com/tensorflow-${TFS_SHORT_VERSION}/license.txt" \
+ || echo "tensorflow-${TFS_SHORT_VERSION}/license.txt not yet provisioned in aws-dlc-licenses bucket; placeholder emitted" >/license.txt
+
+WORKDIR /
+
+
+# ── Stage: sagemaker (SageMaker inference output, MME-capable) ──────────────
+FROM runtime-base AS sagemaker
+ARG TF_SERVING_VERSION
+ARG TFS_SHORT_VERSION
+ARG PYTHON_VERSION
+
+# SageMaker inference labels — accept-bind-to-port allows pipeline use of
+# SAGEMAKER_BIND_TO_PORT; multi-models=true enables MME loading.
+LABEL com.amazonaws.sagemaker.capabilities.accept-bind-to-port=true
+LABEL com.amazonaws.sagemaker.capabilities.multi-models=true
+
+ENV SAGEMAKER_TFS_VERSION="${TFS_SHORT_VERSION}"
+ENV PATH="$PATH:/sagemaker"
+
+# SageMaker BYOC paths
+RUN mkdir -p /opt/ml/input/data /opt/ml/model /opt/ml/output /opt/ml/code
+
+# tensorflow-serving-api — installed inline with --no-deps; see locked
+# decision Q1 in cpu/pyproject.toml header.
+RUN /opt/venv/bin/uv pip install --no-deps --no-cache "tensorflow-serving-api==${TF_SERVING_VERSION}" 2>/dev/null \
+  || /opt/venv/bin/pip install --no-deps --no-cache-dir "tensorflow-serving-api==${TF_SERVING_VERSION}"
+
+# SageMaker handler artifacts (TFS toolkit ported in Phase 3 from master TF 2.19
+# build_artifacts/sagemaker/: serve/serve.py, python_service.py, tfs_utils.py,
+# multi_model_utils.py, tensorflowServing.js, nginx.conf.template).
+COPY scripts/tensorflow/inference/sagemaker /sagemaker
+
+# Telemetry
+COPY scripts/telemetry/deep_learning_container.py /usr/local/bin/deep_learning_container.py
+COPY scripts/telemetry/bash_telemetry.sh.template /tmp/bash_telemetry.sh.template
+ARG FRAMEWORK="tensorflow"
+ARG CONTAINER_TYPE="inference"
+RUN chmod +x /usr/local/bin/deep_learning_container.py \
+  && sed -e "s/{{FRAMEWORK}}/${FRAMEWORK}/g" \
+    -e "s/{{FRAMEWORK_VERSION}}/${TF_SERVING_VERSION}/g" \
+    -e "s/{{CONTAINER_TYPE}}/${CONTAINER_TYPE}/g" \
+    /tmp/bash_telemetry.sh.template >/usr/local/bin/bash_telemetry.sh \
+  && chmod +x /usr/local/bin/bash_telemetry.sh \
+  && rm /tmp/bash_telemetry.sh.template
+
+# Security patch — run after all installers so every OS package is covered.
+RUN dnf upgrade -y --security --releasever latest \
+  && dnf upgrade -y libcurl libcurl-minimal --refresh \
+  && dnf clean all
+
+# Telemetry bashrc hook — MUST be after `dnf upgrade --security`.
+RUN echo 'source /usr/local/bin/bash_telemetry.sh' >>/etc/bashrc \
+  && echo 'source /usr/local/bin/bash_telemetry.sh' >>/root/.bashrc
+
+# OSS compliance
+COPY --from=builder-oss /root/THIRD_PARTY_SOURCE_CODE_URLS /root/THIRD_PARTY_SOURCE_CODE_URLS
+COPY --from=builder-oss /root/PYTHON_PACKAGES_LICENSES /root/PYTHON_PACKAGES_LICENSES
+COPY --from=builder-oss /root/LINUX_PACKAGES_LICENSES /root/LINUX_PACKAGES_LICENSES
+COPY --from=builder-oss /root/BUILD_FROM_SOURCE_PACKAGES_LICENCES /root/BUILD_FROM_SOURCE_PACKAGES_LICENCES
+COPY --from=builder-oss /usr/local/bin/testOSSCompliance /usr/local/bin/testOSSCompliance
+
+# SM entrypoint — start_cuda_compat.sh is a safe no-op on CPU (the compat .so
+# does not exist), so we ship it for uniformity with the GPU image's entrypoint
+# contract. PR #6107 training Dockerfile.cpu makes the same simplification.
+COPY scripts/tensorflow/inference/dockerd_entrypoint.sh /usr/local/bin/dockerd_entrypoint.sh
+COPY scripts/tensorflow/inference/tf_serving_entrypoint.sh /usr/local/bin/tf_serving_entrypoint.sh
+COPY scripts/common/start_cuda_compat.sh /usr/local/bin/start_cuda_compat.sh
+RUN chmod +x /usr/local/bin/dockerd_entrypoint.sh \
+  /usr/local/bin/tf_serving_entrypoint.sh \
+  /usr/local/bin/start_cuda_compat.sh
+
+RUN rm -rf /tmp/* /root/.cache
+
+# Expose ports — TF Serving native gRPC (8500) and REST (8501).
+EXPOSE 8500 8501
+
+ENTRYPOINT ["bash", "-m", "/usr/local/bin/dockerd_entrypoint.sh"]
+CMD ["/usr/local/bin/tf_serving_entrypoint.sh"]
\ No newline at end of file
diff --git a/docker/tensorflow/inference/2.20/Dockerfile.cuda b/docker/tensorflow/inference/2.20/Dockerfile.cuda
new file mode 100644
index 000000000000..ec6f732edd85
--- /dev/null
+++ b/docker/tensorflow/inference/2.20/Dockerfile.cuda
@@ -0,0 +1,263 @@
+# ============================================================================
+# TensorFlow Serving 2.20 Inference DLC — Amazon Linux 2023 (CUDA 12.9.1)
+# Multi-stage build:
+#   builder-base ──┬── builder-oss   (OSS license generation — isolated)
+#                  └── builder-njs   (compile nginx + njs dynamic module)
+#   runtime-base ──── sagemaker      (SageMaker inference, MME-capable)
+#
+# All version defaults mirror docker/tensorflow/inference/2.20/versions-cuda.env.
+# Workflows source versions-cuda.env and pass ARGs via --build-arg — single
+# source of truth.
+#
+# This image is SageMaker-only (no `runtime` stage). It is patterned after
+# PR #6107 (TF 2.21 training v2) for stage layout, with two key deltas:
+#   1. No EFA / NCCL / OpenMPI — inference does not run multi-node MPI.
+#   2. NEW `builder-njs` stage that compiles nginx-mod-njs from source because
+#      AL2023's core repo does not ship nginx-module-njs (master TF 2.19 used
+#      Ubuntu's nginx.org apt repo, which does not exist for AL2023).
+# ============================================================================
+
+# ── Global ARGs (available to all stages) ───────────────────────────────────
+ARG DLC_MAJOR_VERSION=1
+ARG DLC_MINOR_VERSION=0
+ARG CUDA_VERSION=12.9.1
+ARG PYTHON_VERSION=3.12
+ARG TF_SERVING_VERSION=2.20.0
+ARG TFS_SHORT_VERSION=2.20
+ARG NGINX_VERSION=1.30.2
+ARG NJS_VERSION=0.9.9
+
+
+# ── Stage: build_image (TFS upstream binary source) ──────────────────────────
+# Pinned by tag; provides /usr/local/bin/tensorflow_model_server compiled by
+# Google against CUDA 12.x. CUDA minor-version forward compatibility lets the
+# 12.x-built binary run on the CUDA 12.9.1 runtime base below.
+FROM tensorflow/serving:${TF_SERVING_VERSION}-devel-gpu AS build_image
+
+
+# ── Stage: builder-base (Python venv + lockfile deps) ───────────────────────
+FROM nvidia/cuda:${CUDA_VERSION}-devel-amzn2023 AS builder-base
+ARG PYTHON_VERSION
+
+RUN dnf install -y --allowerasing \
+  python${PYTHON_VERSION} python${PYTHON_VERSION}-devel python${PYTHON_VERSION}-pip \
+  gcc gcc-c++ make cmake git openssl-devel ninja-build \
+  tar xz curl \
+  && dnf clean all
+
+COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
+
+ENV UV_PROJECT_ENVIRONMENT="/opt/venv"
+RUN python${PYTHON_VERSION} -m venv /opt/venv
+ENV PATH="/opt/venv/bin:${PATH}"
+
+COPY docker/tensorflow/inference/2.20/cuda/pyproject.toml docker/tensorflow/inference/2.20/cuda/uv.lock /tmp/build/
+WORKDIR /tmp/build
+RUN --mount=type=cache,target=/root/.cache/uv uv sync --frozen --no-dev --no-install-project --inexact
+
+
+# ── Stage: builder-oss (generates license files in isolation) ───────────────
+FROM nvidia/cuda:${CUDA_VERSION}-runtime-amzn2023 AS builder-oss
+ARG PYTHON_VERSION
+RUN dnf install -y --allowerasing python${PYTHON_VERSION} curl && dnf clean all
+COPY --from=builder-base /opt/venv /opt/venv
+COPY scripts/common/setup_oss_compliance.sh /tmp/setup_oss_compliance.sh
+RUN PATH="/opt/venv/bin:${PATH}" bash /tmp/setup_oss_compliance.sh python${PYTHON_VERSION} \
+  && touch /root/THIRD_PARTY_SOURCE_CODE_URLS
+
+
+# ── Stage: builder-njs (compile nginx + njs dynamic module from source) ─────
+# Output: /out/modules/ngx_http_js_module.so
+# We use the CUDA -devel base for build toolchain consistency with builder-base;
+# this stage produces a CPU-only .so artifact that is identical for cuda and cpu
+# variants (kept in this Dockerfile as well as Dockerfile.cpu for self-contained
+# multi-stage builds).
+FROM nvidia/cuda:${CUDA_VERSION}-devel-amzn2023 AS builder-njs
+ARG NGINX_VERSION
+ARG NJS_VERSION
+
+RUN dnf install -y --allowerasing \
+  gcc gcc-c++ make \
+  pcre-devel pcre2-devel zlib-devel openssl-devel libxml2-devel libxslt-devel \
+  curl tar gzip xz \
+  && dnf clean all
+
+WORKDIR /tmp/njs-build
+
+# Fetch nginx + njs sources, both with checksum-validated release tarballs
+# from the upstream maintainer.
+RUN curl -fsSLO "https://nginx.org/download/nginx-${NGINX_VERSION}.tar.gz" \
+  && tar xzf nginx-${NGINX_VERSION}.tar.gz \
+  && curl -fsSL "https://github.com/nginx/njs/archive/refs/tags/${NJS_VERSION}.tar.gz" -o njs-${NJS_VERSION}.tar.gz \
+  && tar xzf njs-${NJS_VERSION}.tar.gz
+
+# Configure nginx with the njs dynamic module, then build only the modules.
+# --with-compat is REQUIRED so the resulting .so loads into a stock nginx
+# binary (we install nginx via dnf in runtime-base; the .so must be ABI-compat).
+RUN cd nginx-${NGINX_VERSION} \
+  && ./configure \
+    --with-compat \
+    --add-dynamic-module=/tmp/njs-build/njs-${NJS_VERSION}/nginx \
+  && make -j$(nproc) modules \
+  && mkdir -p /out/modules \
+  && cp objs/ngx_http_js_module.so /out/modules/
+
+
+# ── Stage: runtime-base (shared base for output stages) ─────────────────────
+FROM nvidia/cuda:${CUDA_VERSION}-runtime-amzn2023 AS runtime-base
+ARG CUDA_VERSION
+ARG PYTHON_VERSION
+ARG TF_SERVING_VERSION
+ARG TFS_SHORT_VERSION
+ARG DLC_MAJOR_VERSION
+ARG DLC_MINOR_VERSION
+
+# Labels live on runtime-base so all output stages inherit them.
+LABEL maintainer="Amazon AI"
+LABEL dlc_major_version="${DLC_MAJOR_VERSION}"
+LABEL dlc_minor_version="${DLC_MINOR_VERSION}"
+LABEL framework="tensorflow"
+LABEL framework_version="${TF_SERVING_VERSION}"
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+  PYTHONUNBUFFERED=1 \
+  PYTHONIOENCODING=UTF-8 \
+  LANG=C.UTF-8 \
+  LC_ALL=C.UTF-8 \
+  DLC_CONTAINER_TYPE=inference \
+  CUDA_HOME=/usr/local/cuda \
+  MODEL_BASE_PATH=/models \
+  MODEL_NAME=model
+
+# Runtime system deps via dnf — equivalent of master TF 2.19 GPU apt-get block,
+# minus the training-specific packages (NCCL/EFA/OpenMPI). nginx is installed
+# from AL2023 core (1.24.x); the njs .so we built in builder-njs is dropped
+# into /usr/lib64/nginx/modules. tar/gzip/which/findutils/util-linux are
+# baseline shell utilities the SM handler scripts and TFS launcher rely on.
+RUN dnf install -y --allowerasing \
+  python${PYTHON_VERSION} python${PYTHON_VERSION}-pip python${PYTHON_VERSION}-devel \
+  nginx \
+  gcc gcc-c++ make git \
+  tar gzip xz which findutils util-linux \
+  libpng-devel freetype-devel zlib-devel \
+  openssl unzip jq curl wget \
+  && dnf clean all
+
+# Copy venv from builder-base
+COPY --from=builder-base /opt/venv /opt/venv
+
+# Copy njs dynamic module(s) into the system nginx modules dir. The package
+# nginx (1.24.x) on AL2023 honors `load_module modules/ngx_http_js_module.so`
+# in its conf, since we built with --with-compat against matching nginx headers.
+COPY --from=builder-njs /out/modules/ngx_http_js_module.so /usr/lib64/nginx/modules/
+
+# TF Serving binary — copied from the official upstream devel image.
+COPY --from=build_image /usr/local/bin/tensorflow_model_server /usr/local/bin/tensorflow_model_server
+
+# python symlink — some TF tooling expects /usr/local/bin/python.
+RUN ln -sf $(which python${PYTHON_VERSION}) /usr/local/bin/python \
+  && ln -sf $(which pip3) /usr/local/bin/pip
+
+ENV PATH="/opt/venv/bin:/usr/local/cuda/bin:${PATH}"
+ENV LD_LIBRARY_PATH="/usr/local/cuda/lib64:/usr/local/lib:${LD_LIBRARY_PATH}"
+
+# Models dir
+RUN mkdir -p ${MODEL_BASE_PATH}
+
+# License file (S3 bucket pre-provisioned for tensorflow-2.20)
+RUN curl -fsSLo /license.txt "https://aws-dlc-licenses.s3.amazonaws.com/tensorflow-${TFS_SHORT_VERSION}/license.txt" \
+ || echo "tensorflow-${TFS_SHORT_VERSION}/license.txt not yet provisioned in aws-dlc-licenses bucket; placeholder emitted" >/license.txt
+
+WORKDIR /
+
+
+# ── Stage: sagemaker (SageMaker inference output, MME-capable) ──────────────
+FROM runtime-base AS sagemaker
+ARG TF_SERVING_VERSION
+ARG TFS_SHORT_VERSION
+ARG PYTHON_VERSION
+# (Labels are inherited from runtime-base — do not redeclare maintainer/framework.)
+
+# SageMaker inference labels — accept-bind-to-port allows pipeline use of
+# SAGEMAKER_BIND_TO_PORT; multi-models=true enables MME loading via
+# tensorflowServing.js (handler scripts ported in Phase 3).
+LABEL com.amazonaws.sagemaker.capabilities.accept-bind-to-port=true
+LABEL com.amazonaws.sagemaker.capabilities.multi-models=true
+LABEL com.amazonaws.sagemaker.inference.cuda.verified_versions=12.9
+
+ENV SAGEMAKER_TFS_VERSION="${TFS_SHORT_VERSION}"
+ENV PATH="$PATH:/sagemaker"
+
+# SageMaker BYOC paths
+RUN mkdir -p /opt/ml/input/data /opt/ml/model /opt/ml/output /opt/ml/code
+
+# tensorflow-serving-api-gpu — installed inline with --no-deps so the ~600 MB
+# `tensorflow` framework wheel is NOT pulled into an inference image. See
+# locked decision Q1 in cuda/pyproject.toml header. The transitive runtime
+# needs (numpy / protobuf / grpcio) are already in the venv from builder-base.
+RUN /opt/venv/bin/uv pip install --no-deps --no-cache "tensorflow-serving-api-gpu==${TF_SERVING_VERSION}" 2>/dev/null \
+  || /opt/venv/bin/pip install --no-deps --no-cache-dir "tensorflow-serving-api-gpu==${TF_SERVING_VERSION}"
+
+# SageMaker handler artifacts (TFS toolkit ported in Phase 3 from master TF 2.19
+# build_artifacts/sagemaker/: serve/serve.py, python_service.py, tfs_utils.py,
+# multi_model_utils.py, tensorflowServing.js, nginx.conf.template).
+COPY scripts/tensorflow/inference/sagemaker /sagemaker
+
+# Telemetry (PT/TF v2 cross-framework pattern)
+COPY scripts/telemetry/deep_learning_container.py /usr/local/bin/deep_learning_container.py
+COPY scripts/telemetry/bash_telemetry.sh.template /tmp/bash_telemetry.sh.template
+ARG FRAMEWORK="tensorflow"
+ARG CONTAINER_TYPE="inference"
+RUN chmod +x /usr/local/bin/deep_learning_container.py \
+  && sed -e "s/{{FRAMEWORK}}/${FRAMEWORK}/g" \
+    -e "s/{{FRAMEWORK_VERSION}}/${TF_SERVING_VERSION}/g" \
+    -e "s/{{CONTAINER_TYPE}}/${CONTAINER_TYPE}/g" \
+    /tmp/bash_telemetry.sh.template >/usr/local/bin/bash_telemetry.sh \
+  && chmod +x /usr/local/bin/bash_telemetry.sh \
+  && rm /tmp/bash_telemetry.sh.template
+
+# Security patch — run after all installers so every OS package is covered.
+# Force libcurl refresh — AL2023's --security filter sometimes misses libcurl
+# CVE patches that ship as regular updates rather than security advisories.
+RUN dnf upgrade -y --security --releasever latest \
+  && dnf upgrade -y cuda-compat-* \
+  && dnf upgrade -y libcurl libcurl-minimal --refresh \
+  && dnf clean all
+
+# Telemetry bashrc hook — MUST be after `dnf upgrade --security` because dnf
+# may replace /etc/bashrc during the upgrade, silently wiping out any `source`
+# line added earlier (PT main pattern).
+RUN echo 'source /usr/local/bin/bash_telemetry.sh' >>/etc/bashrc \
+  && echo 'source /usr/local/bin/bash_telemetry.sh' >>/root/.bashrc
+
+# OSS compliance (copy artifacts from builder-oss)
+COPY --from=builder-oss /root/THIRD_PARTY_SOURCE_CODE_URLS /root/THIRD_PARTY_SOURCE_CODE_URLS
+COPY --from=builder-oss /root/PYTHON_PACKAGES_LICENSES /root/PYTHON_PACKAGES_LICENSES
+COPY --from=builder-oss /root/LINUX_PACKAGES_LICENSES /root/LINUX_PACKAGES_LICENSES
+COPY --from=builder-oss /root/BUILD_FROM_SOURCE_PACKAGES_LICENCES /root/BUILD_FROM_SOURCE_PACKAGES_LICENCES
+COPY --from=builder-oss /usr/local/bin/testOSSCompliance /usr/local/bin/testOSSCompliance
+
+# SM entrypoint — dockerd_entrypoint.sh + tf_serving_entrypoint.sh ported in
+# Phase 3 from master TF 2.19 build_artifacts/. ENTRYPOINT below invokes
+# dockerd_entrypoint.sh which (1) runs telemetry, (2) sources start_cuda_compat
+# when the installed tensorflow-serving-api wheel is the -gpu variant, then
+# (3) `eval`s "$@" — so SM's `docker run … serve` resolves /sagemaker/serve
+# (PATH includes /sagemaker), and a bare `docker run` falls through to the
+# default CMD (tf_serving_entrypoint.sh — non-SM smoke path).
+COPY scripts/tensorflow/inference/dockerd_entrypoint.sh /usr/local/bin/dockerd_entrypoint.sh
+COPY scripts/tensorflow/inference/tf_serving_entrypoint.sh /usr/local/bin/tf_serving_entrypoint.sh
+COPY scripts/common/start_cuda_compat.sh /usr/local/bin/start_cuda_compat.sh
+RUN chmod +x /usr/local/bin/dockerd_entrypoint.sh \
+  /usr/local/bin/tf_serving_entrypoint.sh \
+  /usr/local/bin/start_cuda_compat.sh
+
+RUN rm -rf /tmp/* /root/.cache
+
+# Expose ports — TF Serving native gRPC (8500) and REST (8501).
+# Master TF 2.19 SM image exposes the same pair; the SM frontend nginx (port
+# 8080 from the SM contract) is set up at runtime by the handler scripts and
+# does not need an EXPOSE here.
+EXPOSE 8500 8501
+
+ENTRYPOINT ["bash", "-m", "/usr/local/bin/dockerd_entrypoint.sh"]
+CMD ["/usr/local/bin/tf_serving_entrypoint.sh"]
\ No newline at end of file
diff --git a/docker/tensorflow/inference/2.20/cpu/pyproject.toml b/docker/tensorflow/inference/2.20/cpu/pyproject.toml
new file mode 100644
index 000000000000..ea7afd5a9472
--- /dev/null
+++ b/docker/tensorflow/inference/2.20/cpu/pyproject.toml
@@ -0,0 +1,67 @@
+[project]
+name = "tensorflow-serving-inference-dlc-cpu"
+version = "2.20.0"
+requires-python = ">=3.12,<3.13"
+# ──────────────────────────────────────────────────────────────────────────────
+# TF Serving 2.20 CPU AL2023 inference DLC — runtime-base dependencies
+#
+# The TF Serving binary itself is NOT installed via pip — it is COPYed out of
+# tensorflow/serving:2.20.0-devel in the Dockerfile. The `tensorflow-serving-api`
+# wheel is the matched gRPC stub (used by python_service.py to call the local
+# serving binary on port 8500).
+#
+# IMPORTANT — locked decision Q1 (handoff):
+#   `tensorflow-serving-api==2.20.0` is INTENTIONALLY OMITTED from this
+#   `dependencies` list. The PyPI wheel declares a hard `tensorflow` runtime
+#   dep (≈600 MB framework wheel) which we do not want in an inference image.
+#   uv's resolver does not support per-package `--no-deps`, so the wheel is
+#   instead installed inline in the Dockerfile via:
+#       uv pip install --no-deps tensorflow-serving-api==2.20.0
+#   The transitive runtime needs of tfs-api (numpy, protobuf, grpcio) are
+#   declared explicitly below so a coherent venv is locked. OSS license
+#   tracking still picks up tfs-api at compliance-scan time because the
+#   wheel is physically present in the image.
+#
+# Source for this dep list: master TF 2.19 inference Dockerfile.cpu
+# (RUN pip install blocks), translated to uv/pyproject form per training
+# PR #6107 pattern.
+# ──────────────────────────────────────────────────────────────────────────────
+dependencies = [
+    # ── tfs-api transitive runtime needs (explicit pins; see header note) ─────
+    "numpy==1.26.4",
+    "protobuf>=3.20,<7",
+    "grpcio>=1.0",
+
+    # ── AWS / network / common utilities (carried over from master 2.19) ─────
+    "awscli<2",
+    "boto3",
+    "botocore",
+    "requests",
+    "packaging",
+    "gevent",
+
+    # ── Falcon-based SM handler stack ─────────────────────────────────────────
+    # Both pins user-locked per handoff §16.
+    "falcon==3.1.0",
+    "gunicorn>=22.0.0",
+
+    # ── Compat / security pins inherited from training PR #6107 ───────────────
+    "cython<3",
+    "aiohttp>=3.14.0",
+    "urllib3",
+]
+
+[project.optional-dependencies]
+sagemaker = [
+    # See cuda/pyproject.toml for rationale on each entry; CPU mirrors CUDA
+    # for SM-stage deps. `sagemaker-tensorflow-serving-container` is NOT on
+    # PyPI; the `sagemaker` SDK is intentionally omitted to avoid pulling
+    # torch + nvidia-* into an inference image. Handler scripts use boto3
+    # directly (matches master TF 2.19 inference behavior).
+    "pandas",
+    "scikit-learn",
+    "cloudpickle",
+]
+
+[tool.uv]
+environments = ["sys_platform == 'linux' and platform_machine == 'x86_64'"]
diff --git a/docker/tensorflow/inference/2.20/cpu/uv.lock b/docker/tensorflow/inference/2.20/cpu/uv.lock
new file mode 100644
index 000000000000..5325aa6d44ed
--- /dev/null
+++ b/docker/tensorflow/inference/2.20/cpu/uv.lock
@@ -0,0 +1,550 @@
+version = 1
+revision = 3
+requires-python = "==3.12.*"
+resolution-markers = [
+    "platform_machine == 'x86_64' and sys_platform == 'linux'",
+]
+supported-markers = [
+    "platform_machine == 'x86_64' and sys_platform == 'linux'",
+]
+
+[[package]]
+name = "aiohappyeyeballs"
+version = "2.6.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/33/c6/61a2d7b7572279226bb2e7f61d7a19ca7c90da0329c93fa0d560cbf288d8/aiohappyeyeballs-2.6.2.tar.gz", hash = "sha256:e202810ee718bd01fc6ef49e8ea53d023d5cb6b581076d7925aa499fa55dbe64", size = 22591, upload-time = "2026-05-20T15:12:24.631Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5f/fc/a7bf5b6e4e617b45f90f2d9d2a68519c249c81dd4fc2658c7a2a61c4f4b7/aiohappyeyeballs-2.6.2-py3-none-any.whl", hash = "sha256:4708045e2d7a6c6bdf8aafa8ed39649eaf926a4543b54560659129e3365953c4", size = 15062, upload-time = "2026-05-20T15:12:23.328Z" },
+]
+
+[[package]]
+name = "aiohttp"
+version = "3.14.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "aiohappyeyeballs", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "aiosignal", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "attrs", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "frozenlist", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "multidict", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "propcache", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "typing-extensions", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "yarl", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/82/78/8ea7308cac6934de8c74a14f3d5f65d1c89287426688be79538d0e5c013d/aiohttp-3.14.1.tar.gz", hash = "sha256:307f2cff90a764d329e77040603fa032db89c5c24fdad50c4c15334cba744035", size = 7955794, upload-time = "2026-06-07T21:09:35.529Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/20/9c/d445818389df371f56d141d881153ba23183c4735a03f7356ffb43f7757d/aiohttp-3.14.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:3e6fc1a85fa7194a1a7d19f44e8609180f4a8eb5fa4c7ed8b4355f080fad235c", size = 1790278, upload-time = "2026-06-07T21:06:54.049Z" },
+    { url = "https://files.pythonhosted.org/packages/70/0a/e0075ce9ca0279ee1d4f0c0b85f54fea02ebc83c3007651a72bece658fec/aiohttp-3.14.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:6f71173be42d3241d428f760122febb748de0623f44308a6f120d0dd9ec572e3", size = 1767580, upload-time = "2026-06-07T21:07:07.873Z" },
+]
+
+[[package]]
+name = "aiosignal"
+version = "1.4.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "frozenlist", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "typing-extensions", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/61/62/06741b579156360248d1ec624842ad0edf697050bbaf7c3e46394e106ad1/aiosignal-1.4.0.tar.gz", hash = "sha256:f47eecd9468083c2029cc99945502cb7708b082c232f9aca65da147157b251c7", size = 25007, upload-time = "2025-07-03T22:54:43.528Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/fb/76/641ae371508676492379f16e2fa48f4e2c11741bd63c48be4b12a6b09cba/aiosignal-1.4.0-py3-none-any.whl", hash = "sha256:053243f8b92b990551949e63930a839ff0cf0b0ebbe0597b0f3fb19e1a0fe82e", size = 7490, upload-time = "2025-07-03T22:54:42.156Z" },
+]
+
+[[package]]
+name = "attrs"
+version = "26.1.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/9a/8e/82a0fe20a541c03148528be8cac2408564a6c9a0cc7e9171802bc1d26985/attrs-26.1.0.tar.gz", hash = "sha256:d03ceb89cb322a8fd706d4fb91940737b6642aa36998fe130a9bc96c985eff32", size = 952055, upload-time = "2026-03-19T14:22:25.026Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/64/b4/17d4b0b2a2dc85a6df63d1157e028ed19f90d4cd97c36717afef2bc2f395/attrs-26.1.0-py3-none-any.whl", hash = "sha256:c647aa4a12dfbad9333ca4e71fe62ddc36f4e63b2d260a37a8b83d2f043ac309", size = 67548, upload-time = "2026-03-19T14:22:23.645Z" },
+]
+
+[[package]]
+name = "awscli"
+version = "1.45.28"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "botocore", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "colorama", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "docutils", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "pyyaml", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "rsa", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "s3transfer", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/0e/32/68ff637efdba82e49fd12ddc068fc19edd7f6b4b3ee5bb5976379daaeb25/awscli-1.45.28.tar.gz", hash = "sha256:50e788ed5dc17eb916fd1b9c74d581e4b6a86147c0e60b8b2d25b60ceb32d458", size = 1896334, upload-time = "2026-06-11T19:28:58.014Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/32/41/d9a71dfafd7670e7a79fa4a0d9f073ebbecd33ff4d498f0715d39e4a4464/awscli-1.45.28-py3-none-any.whl", hash = "sha256:783a45041e364c1be9640f730cbcc18313767c592f555477cf0911b1c9511455", size = 4648370, upload-time = "2026-06-11T19:28:54.876Z" },
+]
+
+[[package]]
+name = "boto3"
+version = "1.43.28"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "botocore", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "jmespath", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "s3transfer", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/4e/f2/a976b2a81d8dc7ff675f4b614367a185727061130184a28da0f53f446b97/boto3-1.43.28.tar.gz", hash = "sha256:8391fdcc4d8e1d4e0bf96575a7e5610964a4d401dafa4dccb0a5bade8dd3fbb0", size = 113202, upload-time = "2026-06-11T19:29:01.464Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c3/a5/47db150ea6380f11569b87d3ad064e3c929e5abe227a549d472fab6f5f3a/boto3-1.43.28-py3-none-any.whl", hash = "sha256:4fe6df2163aea02b561eca0d685e2f41a059d71f03721a3e79c3b522e79a3b56", size = 140536, upload-time = "2026-06-11T19:29:00.143Z" },
+]
+
+[[package]]
+name = "botocore"
+version = "1.43.28"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "jmespath", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "python-dateutil", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "urllib3", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/02/dc/1b01808003f88f8a328732c979f20cb0456791048b4440fc4abcae08c1a0/botocore-1.43.28.tar.gz", hash = "sha256:9bbad501a68e4ffdbeff76a382507f5d7827abc316f34a218ab76f5293e6c78d", size = 15503514, upload-time = "2026-06-11T19:28:50.989Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/fa/8c/14916c353ce8a29d14cf6308c2bef842bbb25dde6defc620e26e28063331/botocore-1.43.28-py3-none-any.whl", hash = "sha256:8147adea89b4c9324e842cd8c01ea1a0e17c92cb6ebeaa8cb774f821cb5a7629", size = 15188401, upload-time = "2026-06-11T19:28:47.244Z" },
+]
+
+[[package]]
+name = "certifi"
+version = "2026.5.20"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/f3/ce/ee2ecad540810a79593028e88299baeae54d346cc7a0d94b6199988b89b1/certifi-2026.5.20.tar.gz", hash = "sha256:69dea482ab64caa7b9f6aba1c6bf48bb6a5448d1c0f1b17ab42ad8c763a5344d", size = 135422, upload-time = "2026-05-20T11:46:50.073Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/59/8c/57e832b7af6d7c5abe66eb3fbe3a3a32f4d11ea23a1aa7131371035be991/certifi-2026.5.20-py3-none-any.whl", hash = "sha256:3c52e209ba0a4ad7aebe60436a4ab349c39e1e602e8c134221e546902ad25897", size = 134134, upload-time = "2026-05-20T11:46:48.578Z" },
+]
+
+[[package]]
+name = "charset-normalizer"
+version = "3.4.7"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/e7/a1/67fe25fac3c7642725500a3f6cfe5821ad557c3abb11c9d20d12c7008d3e/charset_normalizer-3.4.7.tar.gz", hash = "sha256:ae89db9e5f98a11a4bf50407d4363e7b09b31e55bc117b4f7d80aab97ba009e5", size = 144271, upload-time = "2026-04-02T09:28:39.342Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/4b/f8/d0118a2f5f23b02cd166fa385c60f9b0d4f9194f574e2b31cef350ad7223/charset_normalizer-3.4.7-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:5649fd1c7bade02f320a462fdefd0b4bd3ce036065836d4f42e0de958038e116", size = 216589, upload-time = "2026-04-02T09:26:29.239Z" },
+    { url = "https://files.pythonhosted.org/packages/8a/2d/a572df5c9204ab7688ec1edc895a73ebded3b023bb07364710b05dd1c9be/charset_normalizer-3.4.7-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:bb6d88045545b26da47aa879dd4a89a71d1dce0f0e549b1abcb31dfe4a8eac49", size = 218468, upload-time = "2026-04-02T09:26:40.17Z" },
+    { url = "https://files.pythonhosted.org/packages/db/8f/61959034484a4a7c527811f4721e75d02d653a35afb0b6054474d8185d4c/charset_normalizer-3.4.7-py3-none-any.whl", hash = "sha256:3dce51d0f5e7951f8bb4900c257dad282f49190fdbebecd4ba99bcc41fef404d", size = 61958, upload-time = "2026-04-02T09:28:37.794Z" },
+]
+
+[[package]]
+name = "cloudpickle"
+version = "3.1.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/27/fb/576f067976d320f5f0114a8d9fa1215425441bb35627b1993e5afd8111e5/cloudpickle-3.1.2.tar.gz", hash = "sha256:7fda9eb655c9c230dab534f1983763de5835249750e85fbcef43aaa30a9a2414", size = 22330, upload-time = "2025-11-03T09:25:26.604Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/88/39/799be3f2f0f38cc727ee3b4f1445fe6d5e4133064ec2e4115069418a5bb6/cloudpickle-3.1.2-py3-none-any.whl", hash = "sha256:9acb47f6afd73f60dc1df93bb801b472f05ff42fa6c84167d25cb206be1fbf4a", size = 22228, upload-time = "2025-11-03T09:25:25.534Z" },
+]
+
+[[package]]
+name = "colorama"
+version = "0.4.6"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/d8/53/6f443c9a4a8358a93a6792e2acffb9d9d5cb0a5cfd8802644b7b1c9a02e4/colorama-0.4.6.tar.gz", hash = "sha256:08695f5cb7ed6e0531a20572697297273c47b8cae5a63ffc6d6ed5c201be6e44", size = 27697, upload-time = "2022-10-25T02:36:22.414Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/d1/d6/3965ed04c63042e047cb6a3e6ed1a63a35087b6a609aa3a15ed8ac56c221/colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6", size = 25335, upload-time = "2022-10-25T02:36:20.889Z" },
+]
+
+[[package]]
+name = "cython"
+version = "0.29.37"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/2a/97/8cc3fe7c6de4796921236a64d00ca8a95565772e57f0d3caae68d880b592/Cython-0.29.37.tar.gz", hash = "sha256:f813d4a6dd94adee5d4ff266191d1d95bf6d4164a4facc535422c021b2504cfb", size = 2099621, upload-time = "2023-12-19T09:22:39.397Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c9/aa/99a0eac01136c0c75feb3210d107c49f93d49d5cb97f19e99318b9ecefdd/Cython-0.29.37-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e8af5975ecfae254d8c0051204fca995dda8f93cf9f0bbf7571e3cda2b0cef4d", size = 1951525, upload-time = "2023-12-19T09:57:00.464Z" },
+    { url = "https://files.pythonhosted.org/packages/14/5f/f5efba4409474892d650ba7e57349afa5d4c41d06f5ba3356c32891c75a6/Cython-0.29.37-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:29415d8eb2fdc1ea518ca4810c50a2d062b387d4c9fbcfb3352346e93db22c6d", size = 1950936, upload-time = "2023-12-19T09:24:07.949Z" },
+    { url = "https://files.pythonhosted.org/packages/7e/26/9d8de10005fedb1eceabe713348d43bae1dbab1786042ca0751a2e2b0f8c/Cython-0.29.37-py2.py3-none-any.whl", hash = "sha256:95f1d6a83ef2729e67b3fa7318c829ce5b07ac64c084cd6af11c228e0364662c", size = 989503, upload-time = "2023-12-19T09:22:32.861Z" },
+]
+
+[[package]]
+name = "docutils"
+version = "0.19"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/6b/5c/330ea8d383eb2ce973df34d1239b3b21e91cd8c865d21ff82902d952f91f/docutils-0.19.tar.gz", hash = "sha256:33995a6753c30b7f577febfc2c50411fec6aac7f7ffeb7c4cfe5991072dcf9e6", size = 2056383, upload-time = "2022-07-05T20:17:31.045Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/93/69/e391bd51bc08ed9141ecd899a0ddb61ab6465309f1eb470905c0c8868081/docutils-0.19-py3-none-any.whl", hash = "sha256:5e1de4d849fee02c63b040a4a3fd567f4ab104defd8a5511fbbc24a8a017efbc", size = 570472, upload-time = "2022-07-05T20:17:26.388Z" },
+]
+
+[[package]]
+name = "falcon"
+version = "3.1.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/36/53/4fd90c6c841bc2e4be29ab92c65e5406df9096c421f138bef9d95d43afc9/falcon-3.1.0.tar.gz", hash = "sha256:f2760bd18c16393a6fb5e55f371f67921edb72febe693a82b3c5e82195d087b7", size = 645631, upload-time = "2022-03-25T16:13:19.137Z" }
+
+[[package]]
+name = "frozenlist"
+version = "1.8.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/2d/f5/c831fac6cc817d26fd54c7eaccd04ef7e0288806943f7cc5bbf69f3ac1f0/frozenlist-1.8.0.tar.gz", hash = "sha256:3ede829ed8d842f6cd48fc7081d7a41001a56f1f38603f9d49bf3020d59a31ad", size = 45875, upload-time = "2025-10-06T05:38:17.865Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/6a/bd/d91c5e39f490a49df14320f4e8c80161cfcce09f1e2cde1edd16a551abb3/frozenlist-1.8.0-cp312-cp312-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:494a5952b1c597ba44e0e78113a7266e656b9794eec897b19ead706bd7074383", size = 242411, upload-time = "2025-10-06T05:36:09.801Z" },
+    { url = "https://files.pythonhosted.org/packages/8f/96/007e5944694d66123183845a106547a15944fbbb7154788cbf7272789536/frozenlist-1.8.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:776f352e8329135506a1d6bf16ac3f87bc25b28e765949282dcc627af36123aa", size = 239096, upload-time = "2025-10-06T05:36:22.129Z" },
+    { url = "https://files.pythonhosted.org/packages/9a/9a/e35b4a917281c0b8419d4207f4334c8e8c5dbf4f3f5f9ada73958d937dcc/frozenlist-1.8.0-py3-none-any.whl", hash = "sha256:0c18a16eab41e82c295618a77502e17b195883241c563b00f0aa5106fc4eaa0d", size = 13409, upload-time = "2025-10-06T05:38:16.721Z" },
+]
+
+[[package]]
+name = "gevent"
+version = "26.5.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "greenlet", marker = "platform_machine == 'x86_64' and platform_python_implementation == 'CPython' and sys_platform == 'linux'" },
+    { name = "zope-event", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "zope-interface", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/c4/cb/98aa3a299e2fc4a2372b5d124863e02965b64579ffc29fe54d0641e65b2f/gevent-26.5.0.tar.gz", hash = "sha256:1655eb04c1e20d71b2aa4a3c7528162dd58ff6cc46a037af1f01f534c80fefba", size = 6712354, upload-time = "2026-05-20T21:22:45.132Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/4e/69/1559b1f6b5107a9118fccd300240879bd581b6d87b03d568d0d155ea702c/gevent-26.5.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:bccff69c462e3650a0fd1d4e9cfc8b6effe15f3e9b1cad20a7bb5ce14b057efd", size = 2114915, upload-time = "2026-05-20T20:35:25.041Z" },
+    { url = "https://files.pythonhosted.org/packages/f9/3c/2fe77ee6e3d381b3c50c0b7d6c4c08c08b8ff5e8c0d9dd51a3b426d61eec/gevent-26.5.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:0bf57df54f1c66273bf3601c2a1e41b12138fe848933718369663bc54f177ca2", size = 2140806, upload-time = "2026-05-20T20:43:28.895Z" },
+]
+
+[[package]]
+name = "greenlet"
+version = "3.5.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/6d/6e/802acd792aebb2256fbbee8cacf2727faaeb6f240ac11008f09eae4414bc/greenlet-3.5.1.tar.gz", hash = "sha256:5a56aeb7d5d9cc4b3a735efb5095bd4b4f6f0e4f93e5ca876d0e2315137b7829", size = 197356, upload-time = "2026-05-20T15:05:03.917Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/40/69/b91cda0647df839483201545913514c2827ebea5e5ccdf931842763bc127/greenlet-3.5.1-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:add5217d68b31130f0beca584d7fef4878327d2e31642b66618a14eef312b63b", size = 611358, upload-time = "2026-05-20T13:14:26.37Z" },
+    { url = "https://files.pythonhosted.org/packages/65/2c/18cece62045e74598c3c393f70dce4a63f56222015ba29a5d4eeb04f764c/greenlet-3.5.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:c5551170cf4f5ff5623e9af81323751979fee2c731e2287b61f73cd27257b823", size = 1635625, upload-time = "2026-05-20T13:14:34.027Z" },
+]
+
+[[package]]
+name = "grpcio"
+version = "1.81.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "typing-extensions", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/b0/b5/1ff353970a87eda4c98251e34d2dfd214abd4982dc89119c9252a2a482d2/grpcio-1.81.1.tar.gz", hash = "sha256:6fa10a767143a5e82e8eaab53918af0cd8909a57a27f8cb2288b80a613ac671b", size = 13026582, upload-time = "2026-06-11T12:46:51.673Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/eb/8c/dea020b6d91508cd84463917a63149ec196ee7db505d032ae43fcb3303b9/grpcio-1.81.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:766bc7c9a9c340342f4c864ccbda8e78111e4751f13b895812b9c148fb79e9d0", size = 6809167, upload-time = "2026-06-11T12:45:32.52Z" },
+    { url = "https://files.pythonhosted.org/packages/25/7a/71437c7f3596e5246155c515852795a85a1a8d228190212432b13b97a95d/grpcio-1.81.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:8ea1936c26b99999b27479853039a7f34713f56c49375ad52b38535ec93a796c", size = 7849660, upload-time = "2026-06-11T12:45:40.627Z" },
+]
+
+[[package]]
+name = "gunicorn"
+version = "26.0.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "packaging", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/6d/b7/a4a3f632f823e432ce6bc65f62961b7980c898c77f075a2f7118cb3846fe/gunicorn-26.0.0.tar.gz", hash = "sha256:ca9346f85e3a4aeeb64d491045c16b9a35647abd37ea15efe53080eb8b090baf", size = 727286, upload-time = "2026-05-05T06:38:25.529Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e6/40/9c2384fc2be4ad25dd4a49decd5ad9ea5a3639814c11bd40ab77cb9f0a14/gunicorn-26.0.0-py3-none-any.whl", hash = "sha256:40233d26a5f0d1872916188c276e21641155111c2853f0c2cd55260aec0d24fc", size = 212009, upload-time = "2026-05-05T06:38:23.007Z" },
+]
+
+[[package]]
+name = "idna"
+version = "3.18"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/cd/63/9496c57188a2ee585e0f1db071d75089a11e98aa86eb99d9d7618fc1edce/idna-3.18.tar.gz", hash = "sha256:ffb385a7e039654cef1ab9ef32c6fafe283c0c0467bba1d9029738ce4a14a848", size = 196711, upload-time = "2026-06-02T14:34:07.794Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/1e/5e/d4e9f1a599fb8e573b7b87160658329fbf28d19eac2718f51fc3def3aa5a/idna-3.18-py3-none-any.whl", hash = "sha256:7f952cbe720b688055e3f87de14f5c3e5fdaa8bc3928985c4077ca689de849a2", size = 65455, upload-time = "2026-06-02T14:34:06.319Z" },
+]
+
+[[package]]
+name = "jmespath"
+version = "1.1.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/d3/59/322338183ecda247fb5d1763a6cbe46eff7222eaeebafd9fa65d4bf5cb11/jmespath-1.1.0.tar.gz", hash = "sha256:472c87d80f36026ae83c6ddd0f1d05d4e510134ed462851fd5f754c8c3cbb88d", size = 27377, upload-time = "2026-01-22T16:35:26.279Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/14/2f/967ba146e6d58cf6a652da73885f52fc68001525b4197effc174321d70b4/jmespath-1.1.0-py3-none-any.whl", hash = "sha256:a5663118de4908c91729bea0acadca56526eb2698e83de10cd116ae0f4e97c64", size = 20419, upload-time = "2026-01-22T16:35:24.919Z" },
+]
+
+[[package]]
+name = "joblib"
+version = "1.5.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/41/f2/d34e8b3a08a9cc79a50b2208a93dce981fe615b64d5a4d4abee421d898df/joblib-1.5.3.tar.gz", hash = "sha256:8561a3269e6801106863fd0d6d84bb737be9e7631e33aaed3fb9ce5953688da3", size = 331603, upload-time = "2025-12-15T08:41:46.427Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/7b/91/984aca2ec129e2757d1e4e3c81c3fcda9d0f85b74670a094cc443d9ee949/joblib-1.5.3-py3-none-any.whl", hash = "sha256:5fc3c5039fc5ca8c0276333a188bbd59d6b7ab37fe6632daa76bc7f9ec18e713", size = 309071, upload-time = "2025-12-15T08:41:44.973Z" },
+]
+
+[[package]]
+name = "multidict"
+version = "6.7.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/1a/c2/c2d94cbe6ac1753f3fc980da97b3d930efe1da3af3c9f5125354436c073d/multidict-6.7.1.tar.gz", hash = "sha256:ec6652a1bee61c53a3e5776b6049172c53b6aaba34f18c9ad04f82712bac623d", size = 102010, upload-time = "2026-01-26T02:46:45.979Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f3/8d/5e5be3ced1d12966fefb5c4ea3b2a5b480afcea36406559442c6e31d4a48/multidict-6.7.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:bfde23ef6ed9db7eaee6c37dcec08524cb43903c60b285b172b6c094711b3961", size = 256322, upload-time = "2026-01-26T02:44:03.56Z" },
+    { url = "https://files.pythonhosted.org/packages/2b/17/ebb9644da78c4ab36403739e0e6e0e30ebb135b9caf3440825001a0bddcb/multidict-6.7.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:fc5907494fccf3e7d3f94f95c91d6336b092b5fc83811720fae5e2765890dfba", size = 251377, upload-time = "2026-01-26T02:44:12.042Z" },
+    { url = "https://files.pythonhosted.org/packages/81/08/7036c080d7117f28a4af526d794aab6a84463126db031b007717c1a6676e/multidict-6.7.1-py3-none-any.whl", hash = "sha256:55d97cc6dae627efa6a6e548885712d4864b81110ac76fa4e534c03819fa4a56", size = 12319, upload-time = "2026-01-26T02:46:44.004Z" },
+]
+
+[[package]]
+name = "narwhals"
+version = "2.22.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/62/3c/c4ef2164a71c1a63d7f1ae411c4082c5fa872405106db60a4b7114989ad7/narwhals-2.22.1.tar.gz", hash = "sha256:d62920805a0a43b7ff8b54b0c0d3142d796f8a9301836ada37e573d6a33cbcd9", size = 647493, upload-time = "2026-06-05T12:34:34.051Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/48/ca/36339329c4604adbcc99c899b7eb1ce1a555c499b6a6860757dc9bfed36d/narwhals-2.22.1-py3-none-any.whl", hash = "sha256:60567d774edf77db53906f89d9fbd164e66e56d66d388e1e6990f17ac33cfb53", size = 454815, upload-time = "2026-06-05T12:34:32.289Z" },
+]
+
+[[package]]
+name = "numpy"
+version = "1.26.4"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/65/6e/09db70a523a96d25e115e71cc56a6f9031e7b8cd166c1ac8438307c14058/numpy-1.26.4.tar.gz", hash = "sha256:2a02aba9ed12e4ac4eb3ea9421c420301a0c6460d9830d74a9df87efa4912010", size = 15786129, upload-time = "2024-02-06T00:26:44.495Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/0f/50/de23fde84e45f5c4fda2488c759b69990fd4512387a8632860f3ac9cd225/numpy-1.26.4-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:675d61ffbfa78604709862923189bad94014bef562cc35cf61d3a07bba02a7ed", size = 17950613, upload-time = "2024-02-05T23:56:56.054Z" },
+    { url = "https://files.pythonhosted.org/packages/76/8c/2ba3902e1a0fc1c74962ea9bb33a534bb05984ad7ff9515bf8d07527cadd/numpy-1.26.4-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:1dda2e7b4ec9dd512f84935c5f126c8bd8b9f2fc001e9f54af255e8c5f16b0e0", size = 17786643, upload-time = "2024-02-05T23:57:56.585Z" },
+]
+
+[[package]]
+name = "packaging"
+version = "26.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/d7/f1/e7a6dd94a8d4a5626c03e4e99c87f241ba9e350cd9e6d75123f992427270/packaging-26.2.tar.gz", hash = "sha256:ff452ff5a3e828ce110190feff1178bb1f2ea2281fa2075aadb987c2fb221661", size = 228134, upload-time = "2026-04-24T20:15:23.917Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/df/b2/87e62e8c3e2f4b32e5fe99e0b86d576da1312593b39f47d8ceef365e95ed/packaging-26.2-py3-none-any.whl", hash = "sha256:5fc45236b9446107ff2415ce77c807cee2862cb6fac22b8a73826d0693b0980e", size = 100195, upload-time = "2026-04-24T20:15:22.081Z" },
+]
+
+[[package]]
+name = "pandas"
+version = "3.0.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "python-dateutil", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/f8/87/4341c6252d1c47b08768c3d25ac487362bf403f0313ddae4a2a26c9b1b4c/pandas-3.0.3.tar.gz", hash = "sha256:696a4a00a2a2a35d4e5deb3fc946641b96c944f02230e4f76137fe35d806c4fc", size = 4651414, upload-time = "2026-05-11T18:54:29.21Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/65/b6/09b01cdbc15224e2850365192d17b7bdebb8bdbd8780ed221fcdf0d9a515/pandas-3.0.3-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6dc0b3fd2169c9157deed50b4d519553a3655c8c6a96027136d654592be973a9", size = 10894600, upload-time = "2026-05-11T18:52:45.02Z" },
+    { url = "https://files.pythonhosted.org/packages/f8/45/830bb57f533a4604b355e07edcb8ea18cf88b5f94e5fca92f27052d7c597/pandas-3.0.3-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:f8894dc474d648fe7b6ff0ca9b0bd73950d19952bc1a6534540762c5d79d305c", size = 11950889, upload-time = "2026-05-11T18:52:50.905Z" },
+]
+
+[[package]]
+name = "propcache"
+version = "0.5.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/ec/44/c87281c333769159c50594f22610f77398a47ccbfbbf23074e744e86f87c/propcache-0.5.2.tar.gz", hash = "sha256:01c4fc7480cd0598bb4b57022df55b9ca296da7fc5a8760bd8451a7e63a7d427", size = 50208, upload-time = "2026-05-08T21:02:12.199Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/14/15/5574111ae50dd6e879456888c0eadd4c5a869959775854e18e18a6b345f3/propcache-0.5.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6f328175a2cde1f0ff2c4ed8ce968b9dcfb55f3a7153f39e2957ed994da13476", size = 61639, upload-time = "2026-05-08T21:00:19.692Z" },
+    { url = "https://files.pythonhosted.org/packages/12/fd/77fe5936d8c3086ca9048f7f415f122ed82e53884a9ec193646b42deef06/propcache-0.5.2-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:c80f4ba3e8f00189165999a742ee526ebeccedf6c3f7beb0c7df821e9772435a", size = 62514, upload-time = "2026-05-08T21:00:30.098Z" },
+    { url = "https://files.pythonhosted.org/packages/3a/ed/1cdcab6ba3d6ab7feca11fc14f0eeea80755bb53ef4e892079f31b10a25f/propcache-0.5.2-py3-none-any.whl", hash = "sha256:be1ddfcbb376e3de5d2e2db1d58d6d67463e6b4f9f040c000de8e300295465fe", size = 14036, upload-time = "2026-05-08T21:02:10.673Z" },
+]
+
+[[package]]
+name = "protobuf"
+version = "6.33.6"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/66/70/e908e9c5e52ef7c3a6c7902c9dfbb34c7e29c25d2f81ade3856445fd5c94/protobuf-6.33.6.tar.gz", hash = "sha256:a6768d25248312c297558af96a9f9c929e8c4cee0659cb07e780731095f38135", size = 444531, upload-time = "2026-03-18T19:05:00.988Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/16/92/d1e32e3e0d894fe00b15ce28ad4944ab692713f2e7f0a99787405e43533a/protobuf-6.33.6-cp39-abi3-manylinux2014_x86_64.whl", hash = "sha256:e9db7e292e0ab79dd108d7f1a94fe31601ce1ee3f7b79e0692043423020b0593", size = 323436, upload-time = "2026-03-18T19:04:55.768Z" },
+    { url = "https://files.pythonhosted.org/packages/c4/72/02445137af02769918a93807b2b7890047c32bfb9f90371cbc12688819eb/protobuf-6.33.6-py3-none-any.whl", hash = "sha256:77179e006c476e69bf8e8ce866640091ec42e1beb80b213c3900006ecfba6901", size = 170656, upload-time = "2026-03-18T19:04:59.826Z" },
+]
+
+[[package]]
+name = "pyasn1"
+version = "0.6.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/5c/5f/6583902b6f79b399c9c40674ac384fd9cd77805f9e6205075f828ef11fb2/pyasn1-0.6.3.tar.gz", hash = "sha256:697a8ecd6d98891189184ca1fa05d1bb00e2f84b5977c481452050549c8a72cf", size = 148685, upload-time = "2026-03-17T01:06:53.382Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5d/a0/7d793dce3fa811fe047d6ae2431c672364b462850c6235ae306c0efd025f/pyasn1-0.6.3-py3-none-any.whl", hash = "sha256:a80184d120f0864a52a073acc6fc642847d0be408e7c7252f31390c0f4eadcde", size = 83997, upload-time = "2026-03-17T01:06:52.036Z" },
+]
+
+[[package]]
+name = "python-dateutil"
+version = "2.9.0.post0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "six", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/66/c0/0c8b6ad9f17a802ee498c46e004a0eb49bc148f2fd230864601a86dcf6db/python-dateutil-2.9.0.post0.tar.gz", hash = "sha256:37dd54208da7e1cd875388217d5e00ebd4179249f90fb72437e91a35459a0ad3", size = 342432, upload-time = "2024-03-01T18:36:20.211Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ec/57/56b9bcc3c9c6a792fcbaf139543cee77261f3651ca9da0c93f5c1221264b/python_dateutil-2.9.0.post0-py2.py3-none-any.whl", hash = "sha256:a8b2bc7bffae282281c8140a97d3aa9c14da0b136dfe83f850eea9a5f7470427", size = 229892, upload-time = "2024-03-01T18:36:18.57Z" },
+]
+
+[[package]]
+name = "pyyaml"
+version = "6.0.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/05/8e/961c0007c59b8dd7729d542c61a4d537767a59645b82a0b521206e1e25c2/pyyaml-6.0.3.tar.gz", hash = "sha256:d76623373421df22fb4cf8817020cbb7ef15c725b9d5e45f17e189bfc384190f", size = 130960, upload-time = "2025-09-25T21:33:16.546Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/8b/9d/b3589d3877982d4f2329302ef98a8026e7f4443c765c46cfecc8858c6b4b/pyyaml-6.0.3-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ba1cc08a7ccde2d2ec775841541641e4548226580ab850948cbfda66a1befcdc", size = 807870, upload-time = "2025-09-25T21:32:16.431Z" },
+    { url = "https://files.pythonhosted.org/packages/be/8e/98435a21d1d4b46590d5459a22d88128103f8da4c2d4cb8f14f2a96504e1/pyyaml-6.0.3-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:41715c910c881bc081f1e8872880d3c650acf13dfa8214bad49ed4cede7c34ea", size = 790181, upload-time = "2025-09-25T21:32:18.834Z" },
+]
+
+[[package]]
+name = "requests"
+version = "2.34.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "certifi", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "charset-normalizer", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "idna", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "urllib3", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/ac/c3/e2a2b89f2d3e2179abd6d00ebd70bff6273f37fb3e0cc209f48b39d00cbf/requests-2.34.2.tar.gz", hash = "sha256:f288924cae4e29463698d6d60bc6a4da69c89185ad1e0bcc4104f584e960b9ed", size = 142856, upload-time = "2026-05-14T19:25:27.735Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a0/f4/c67b0b3f1b9245e8d266f0f112c500d50e5b4e83cb6f3b71b6528104182a/requests-2.34.2-py3-none-any.whl", hash = "sha256:2a0d60c172f83ac6ab31e4554906c0f3b3588d37b5cb939b1c061f4907e278e0", size = 73075, upload-time = "2026-05-14T19:25:26.443Z" },
+]
+
+[[package]]
+name = "rsa"
+version = "4.7.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "pyasn1", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/db/b5/475c45a58650b0580421746504b680cd2db4e81bc941e94ca53785250269/rsa-4.7.2.tar.gz", hash = "sha256:9d689e6ca1b3038bc82bf8d23e944b6b6037bc02301a574935b2dd946e0353b9", size = 39711, upload-time = "2021-02-24T10:55:05.846Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e9/93/0c0f002031f18b53af7a6166103c02b9c0667be528944137cc954ec921b3/rsa-4.7.2-py3-none-any.whl", hash = "sha256:78f9a9bf4e7be0c5ded4583326e7461e3a3c5aae24073648b4bdfa797d78c9d2", size = 34505, upload-time = "2021-02-24T10:55:03.55Z" },
+]
+
+[[package]]
+name = "s3transfer"
+version = "0.18.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "botocore", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/e0/1f/12417f7f493fc45e1f9fd5d4a9b6c125cf8d2cf3f8ddbdfab3e76406e9d6/s3transfer-0.18.0.tar.gz", hash = "sha256:3760b8b7ec1315da54048b2d626276732bee4300d054d492d4e1d43e20d4ecbd", size = 160560, upload-time = "2026-05-28T19:39:09.124Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/2b/58/a58fc997655386daa2e25784e30c288aa3e3819e401f77029ee4899fb55a/s3transfer-0.18.0-py3-none-any.whl", hash = "sha256:239c13b09e65ad0346e1be7348b8a202dcad44ac7ea7c6eb858fc881dce739b6", size = 88572, upload-time = "2026-05-28T19:39:07.999Z" },
+]
+
+[[package]]
+name = "scikit-learn"
+version = "1.9.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "joblib", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "narwhals", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "numpy", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "scipy", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "threadpoolctl", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/fa/6f/37092bdb25f712817231799fc5674d8e704066a8a70c1d2d40517e18b4ab/scikit_learn-1.9.0.tar.gz", hash = "sha256:8833266989d3a5110178a9fae30783675460724d0e1efb13b14901d2c660c557", size = 7750767, upload-time = "2026-06-02T11:54:32.706Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/6c/c2/63fdda36c56437eeb44aaf9493c8bcd62ce230ab1598924fc626ffbfa943/scikit_learn-1.9.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:056c92bb67ad4c28463c2f2653d9701449201e7e7a9e94e321be0f71c4fef2b8", size = 9132097, upload-time = "2026-06-02T11:53:33.456Z" },
+]
+
+[[package]]
+name = "scipy"
+version = "1.17.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/7a/97/5a3609c4f8d58b039179648e62dd220f89864f56f7357f5d4f45c29eb2cc/scipy-1.17.1.tar.gz", hash = "sha256:95d8e012d8cb8816c226aef832200b1d45109ed4464303e997c5b13122b297c0", size = 30573822, upload-time = "2026-02-23T00:26:24.851Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/01/8e/1e35281b8ab6d5d72ebe9911edcdffa3f36b04ed9d51dec6dd140396e220/scipy-1.17.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:02ae3b274fde71c5e92ac4d54bc06c42d80e399fec704383dcd99b301df37458", size = 35235890, upload-time = "2026-02-23T00:18:49.188Z" },
+    { url = "https://files.pythonhosted.org/packages/65/94/7698add8f276dbab7a9de9fb6b0e02fc13ee61d51c7c3f85ac28b65e1239/scipy-1.17.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:f590cd684941912d10becc07325a3eeb77886fe981415660d9265c4c418d0bea", size = 37625856, upload-time = "2026-02-23T00:19:00.307Z" },
+]
+
+[[package]]
+name = "six"
+version = "1.17.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/94/e7/b2c673351809dca68a0e064b6af791aa332cf192da575fd474ed7d6f16a2/six-1.17.0.tar.gz", hash = "sha256:ff70335d468e7eb6ec65b95b99d3a2836546063f63acc5171de367e834932a81", size = 34031, upload-time = "2024-12-04T17:35:28.174Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/b7/ce/149a00dd41f10bc29e5921b496af8b574d8413afcd5e30dfa0ed46c2cc5e/six-1.17.0-py2.py3-none-any.whl", hash = "sha256:4721f391ed90541fddacab5acf947aa0d3dc7d27b2e1e8eda2be8970586c3274", size = 11050, upload-time = "2024-12-04T17:35:26.475Z" },
+]
+
+[[package]]
+name = "tensorflow-serving-inference-dlc-cpu"
+version = "2.20.0"
+source = { virtual = "." }
+dependencies = [
+    { name = "aiohttp", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "awscli", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "boto3", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "botocore", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "cython", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "falcon", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "gevent", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "grpcio", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "gunicorn", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "numpy", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "packaging", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "protobuf", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "requests", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "urllib3", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+
+[package.optional-dependencies]
+sagemaker = [
+    { name = "cloudpickle", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "pandas", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "scikit-learn", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+
+[package.metadata]
+requires-dist = [
+    { name = "aiohttp", specifier = ">=3.14.0" },
+    { name = "awscli", specifier = "<2" },
+    { name = "boto3" },
+    { name = "botocore" },
+    { name = "cloudpickle", marker = "extra == 'sagemaker'" },
+    { name = "cython", specifier = "<3" },
+    { name = "falcon", specifier = "==3.1.0" },
+    { name = "gevent" },
+    { name = "grpcio", specifier = ">=1.0" },
+    { name = "gunicorn", specifier = ">=22.0.0" },
+    { name = "numpy", specifier = "==1.26.4" },
+    { name = "packaging" },
+    { name = "pandas", marker = "extra == 'sagemaker'" },
+    { name = "protobuf", specifier = ">=3.20,<7" },
+    { name = "requests" },
+    { name = "scikit-learn", marker = "extra == 'sagemaker'" },
+    { name = "urllib3" },
+]
+provides-extras = ["sagemaker"]
+
+[[package]]
+name = "threadpoolctl"
+version = "3.6.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/b7/4d/08c89e34946fce2aec4fbb45c9016efd5f4d7f24af8e5d93296e935631d8/threadpoolctl-3.6.0.tar.gz", hash = "sha256:8ab8b4aa3491d812b623328249fab5302a68d2d71745c8a4c719a2fcaba9f44e", size = 21274, upload-time = "2025-03-13T13:49:23.031Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/32/d5/f9a850d79b0851d1d4ef6456097579a9005b31fea68726a4ae5f2d82ddd9/threadpoolctl-3.6.0-py3-none-any.whl", hash = "sha256:43a0b8fd5a2928500110039e43a5eed8480b918967083ea48dc3ab9f13c4a7fb", size = 18638, upload-time = "2025-03-13T13:49:21.846Z" },
+]
+
+[[package]]
+name = "typing-extensions"
+version = "4.15.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/72/94/1a15dd82efb362ac84269196e94cf00f187f7ed21c242792a923cdb1c61f/typing_extensions-4.15.0.tar.gz", hash = "sha256:0cea48d173cc12fa28ecabc3b837ea3cf6f38c6d1136f85cbaaf598984861466", size = 109391, upload-time = "2025-08-25T13:49:26.313Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/18/67/36e9267722cc04a6b9f15c7f3441c2363321a3ea07da7ae0c0707beb2a9c/typing_extensions-4.15.0-py3-none-any.whl", hash = "sha256:f0fa19c6845758ab08074a0cfa8b7aecb71c999ca73d62883bc25cc018c4e548", size = 44614, upload-time = "2025-08-25T13:49:24.86Z" },
+]
+
+[[package]]
+name = "urllib3"
+version = "2.7.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/53/0c/06f8b233b8fd13b9e5ee11424ef85419ba0d8ba0b3138bf360be2ff56953/urllib3-2.7.0.tar.gz", hash = "sha256:231e0ec3b63ceb14667c67be60f2f2c40a518cb38b03af60abc813da26505f4c", size = 433602, upload-time = "2026-05-07T16:13:18.596Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/7f/3e/5db95bcf282c52709639744ca2a8b149baccf648e39c8cc87553df9eae0c/urllib3-2.7.0-py3-none-any.whl", hash = "sha256:9fb4c81ebbb1ce9531cce37674bbc6f1360472bc18ca9a553ede278ef7276897", size = 131087, upload-time = "2026-05-07T16:13:17.151Z" },
+]
+
+[[package]]
+name = "yarl"
+version = "1.24.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "idna", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "multidict", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "propcache", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/79/12/1e8f37460ea0f7eb59c221fdaf0ed75e7ac43e97f8093b9c6f411df50a78/yarl-1.24.2.tar.gz", hash = "sha256:9ac374123c6fd7abf64d1fec93962b0bd4ee2c19751755a762a72dd96c0378f8", size = 210798, upload-time = "2026-05-19T21:31:05.599Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/7a/da/323a01c349bd5fb01bb6652e314d9bb218cee630a736bdb810ad50e4013f/yarl-1.24.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7e7ebcdef69dec6c6451e616f32b622a6d4a2e92b445c992f7c8e5274a6bbc4c", size = 105612, upload-time = "2026-05-19T21:29:04.247Z" },
+    { url = "https://files.pythonhosted.org/packages/48/2d/1c8d89c7c5f9cad9fb2902445d94e2ab1d7aa35de029afbb8ae95c42d00f/yarl-1.24.2-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:e30dd55825dc554ec5b66a94953b8eda8745926514c5089dfcacecb9c99b5bd1", size = 105719, upload-time = "2026-05-19T21:29:18.367Z" },
+    { url = "https://files.pythonhosted.org/packages/fd/4d/4b880086bd0d3e034d25647be1d830afc3e3f610e98c4ab3490af6b1b6d5/yarl-1.24.2-py3-none-any.whl", hash = "sha256:2783d9226db8797636cd6896e4de81feed252d1db72265686c9558d97a4d94b9", size = 53576, upload-time = "2026-05-19T21:31:03.909Z" },
+]
+
+[[package]]
+name = "zope-event"
+version = "6.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/93/41/faa10af34d48d9cd6fa0249a1162943ad84a9590bd1a06939981e6640416/zope_event-6.2.tar.gz", hash = "sha256:b97d5d6327067ee6b9dfcbdf606ade9ade70991e19c162e808ea39e5fcf0f8d3", size = 18958, upload-time = "2026-04-28T06:24:10.578Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/9e/33/848922889e946d4befc415c219fe516af75c49555d8e736e183bfd30db42/zope_event-6.2-py3-none-any.whl", hash = "sha256:5e755153ac4faf64c10a4b6dd3307680166a3edf65b38df22df592610f8fa874", size = 6525, upload-time = "2026-04-28T06:24:09.176Z" },
+]
+
+[[package]]
+name = "zope-interface"
+version = "8.5"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/08/dc/50550cfcbb2ea3cbca5f1d7ed05c8aa840f831a0f2d63aec0a953f7c590e/zope_interface-8.5.tar.gz", hash = "sha256:7a3ba1c5877f0f3e3906b02ddf793abed2becc2948116414ce0e1dd820b68d6d", size = 257957, upload-time = "2026-05-26T06:50:14.574Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c5/66/1036543d6a66bc04c19df3cf650f3ad938a002ab0a443c24e23e8de5e8b9/zope_interface-8.5-cp312-cp312-manylinux1_x86_64.manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:5e970dabea777a24b0b0bbf9dae3ab75ce8b2d8e948edf4875627034b21f3560", size = 270689, upload-time = "2026-05-26T06:49:31.767Z" },
+]
diff --git a/docker/tensorflow/inference/2.20/cuda/pyproject.toml b/docker/tensorflow/inference/2.20/cuda/pyproject.toml
new file mode 100644
index 000000000000..5e27349df5f1
--- /dev/null
+++ b/docker/tensorflow/inference/2.20/cuda/pyproject.toml
@@ -0,0 +1,90 @@
+[project]
+name = "tensorflow-serving-inference-dlc"
+version = "2.20.0"
+requires-python = ">=3.12,<3.13"
+# ──────────────────────────────────────────────────────────────────────────────
+# TF Serving 2.20 CUDA AL2023 inference DLC — runtime-base dependencies
+#
+# The TF Serving binary itself is NOT installed via pip — it is COPYed out of
+# tensorflow/serving:2.20.0-devel-gpu in the Dockerfile. The
+# `tensorflow-serving-api-gpu` wheel is the matched gRPC stub (used by
+# python_service.py to call the local serving binary on port 8500).
+#
+# IMPORTANT — locked decision Q1 (handoff):
+#   `tensorflow-serving-api-gpu==2.20.0` is INTENTIONALLY OMITTED from this
+#   `dependencies` list. The PyPI wheel declares a hard `tensorflow` runtime
+#   dep (≈600 MB framework wheel) which we do not want in an inference image.
+#   uv's resolver does not support per-package `--no-deps`, so the wheel is
+#   instead installed inline in the Dockerfile via:
+#       uv pip install --no-deps tensorflow-serving-api-gpu==2.20.0
+#   The transitive runtime needs of tfs-api (numpy, protobuf, grpcio) are
+#   declared explicitly below so a coherent venv is locked. OSS license
+#   tracking still picks up tfs-api at compliance-scan time because the
+#   wheel is physically present in the image.
+#
+# Source for this dep list: master TF 2.19 inference Dockerfile.gpu
+# (RUN pip install blocks at lines ~155 and SM-stage line ~265), translated
+# to uv/pyproject form per training PR #6107 pattern.
+# ──────────────────────────────────────────────────────────────────────────────
+dependencies = [
+    # ── tfs-api transitive runtime needs (explicit pins; see header note) ─────
+    # numpy 1.26.4 — TF SavedModel/numpy ABI compat. Must match the version
+    # the TF 2.21 training image uses to export SavedModels (handoff §16).
+    "numpy==1.26.4",
+    # protobuf — tfs-api uses protobuf for the gRPC stubs. Master 2.19 pinned
+    # protobuf==6.33.4; we relax to >=3.20,<7 here so the resolver can pick the
+    # latest line that satisfies the rest of the venv. Floor 3.20 matches the
+    # tfs-api wheel's metadata; ceiling 7 prevents future major-version churn.
+    "protobuf>=3.20,<7",
+    # grpcio — gRPC client used by python_service.py. Latest stable line.
+    "grpcio>=1.0",
+
+    # ── AWS / network / common utilities (carried over from master 2.19) ─────
+    "awscli<2",
+    "boto3",
+    "botocore",
+    "requests",
+    "packaging",
+    "gevent",
+
+    # ── Falcon-based SM handler stack ─────────────────────────────────────────
+    # Both pins user-locked per handoff §16. Master GPU used falcon==3.0.1, but
+    # we standardize on 3.1.0 (matches master CPU and is forward-compatible).
+    "falcon==3.1.0",
+    "gunicorn>=22.0.0",
+
+    # ── Compat / security pins inherited from training PR #6107 ───────────────
+    # cython<3 — TF/numpy ABI compat; same constraint as training image.
+    "cython<3",
+    # aiohttp>=3.14.0 — security pin (CVE-2026-34993, CVE-2026-47265). Pulled
+    # transitively by aiobotocore / boto3 ecosystem. Direct pin forces resolver
+    # to pick the patched line.
+    "aiohttp>=3.14.0",
+    "urllib3",
+]
+
+[project.optional-dependencies]
+sagemaker = [
+    # NOTE: `sagemaker-tensorflow-serving-container` is NOT on PyPI (verified
+    # via `uv lock` failure on 2026-06-11; handoff §14.4 flagged this risk).
+    # Master TF 2.19 inference references it but never pip-installs it from
+    # PyPI — it is vendored as the local handler scripts under
+    # docker/build_artifacts/sagemaker/, which we port to scripts/tensorflow/
+    # inference/sagemaker/ in Phase 3.
+    #
+    # Master TF 2.19 inference SM stage does NOT install the `sagemaker` SDK
+    # (verified by re-reading tensorflow/inference/docker/2.19/py3/cu122/Dockerfile.gpu
+    # lines 247-275). The handler scripts (python_service.py, multi_model_utils.py)
+    # use boto3 directly. Removing the SDK from this extra avoids transitively
+    # pulling torch + ~20 nvidia-cu13 packages we have no use for in an
+    # inference image.
+    #
+    # Lightweight data tooling — kept so customer inference handlers can use
+    # them without an extra pip install:
+    "pandas",
+    "scikit-learn",
+    "cloudpickle",
+]
+
+[tool.uv]
+environments = ["sys_platform == 'linux' and platform_machine == 'x86_64'"]
diff --git a/docker/tensorflow/inference/2.20/cuda/uv.lock b/docker/tensorflow/inference/2.20/cuda/uv.lock
new file mode 100644
index 000000000000..f63d2c8f949f
--- /dev/null
+++ b/docker/tensorflow/inference/2.20/cuda/uv.lock
@@ -0,0 +1,550 @@
+version = 1
+revision = 3
+requires-python = "==3.12.*"
+resolution-markers = [
+    "platform_machine == 'x86_64' and sys_platform == 'linux'",
+]
+supported-markers = [
+    "platform_machine == 'x86_64' and sys_platform == 'linux'",
+]
+
+[[package]]
+name = "aiohappyeyeballs"
+version = "2.6.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/33/c6/61a2d7b7572279226bb2e7f61d7a19ca7c90da0329c93fa0d560cbf288d8/aiohappyeyeballs-2.6.2.tar.gz", hash = "sha256:e202810ee718bd01fc6ef49e8ea53d023d5cb6b581076d7925aa499fa55dbe64", size = 22591, upload-time = "2026-05-20T15:12:24.631Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5f/fc/a7bf5b6e4e617b45f90f2d9d2a68519c249c81dd4fc2658c7a2a61c4f4b7/aiohappyeyeballs-2.6.2-py3-none-any.whl", hash = "sha256:4708045e2d7a6c6bdf8aafa8ed39649eaf926a4543b54560659129e3365953c4", size = 15062, upload-time = "2026-05-20T15:12:23.328Z" },
+]
+
+[[package]]
+name = "aiohttp"
+version = "3.14.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "aiohappyeyeballs", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "aiosignal", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "attrs", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "frozenlist", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "multidict", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "propcache", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "typing-extensions", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "yarl", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/82/78/8ea7308cac6934de8c74a14f3d5f65d1c89287426688be79538d0e5c013d/aiohttp-3.14.1.tar.gz", hash = "sha256:307f2cff90a764d329e77040603fa032db89c5c24fdad50c4c15334cba744035", size = 7955794, upload-time = "2026-06-07T21:09:35.529Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/20/9c/d445818389df371f56d141d881153ba23183c4735a03f7356ffb43f7757d/aiohttp-3.14.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:3e6fc1a85fa7194a1a7d19f44e8609180f4a8eb5fa4c7ed8b4355f080fad235c", size = 1790278, upload-time = "2026-06-07T21:06:54.049Z" },
+    { url = "https://files.pythonhosted.org/packages/70/0a/e0075ce9ca0279ee1d4f0c0b85f54fea02ebc83c3007651a72bece658fec/aiohttp-3.14.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:6f71173be42d3241d428f760122febb748de0623f44308a6f120d0dd9ec572e3", size = 1767580, upload-time = "2026-06-07T21:07:07.873Z" },
+]
+
+[[package]]
+name = "aiosignal"
+version = "1.4.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "frozenlist", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "typing-extensions", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/61/62/06741b579156360248d1ec624842ad0edf697050bbaf7c3e46394e106ad1/aiosignal-1.4.0.tar.gz", hash = "sha256:f47eecd9468083c2029cc99945502cb7708b082c232f9aca65da147157b251c7", size = 25007, upload-time = "2025-07-03T22:54:43.528Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/fb/76/641ae371508676492379f16e2fa48f4e2c11741bd63c48be4b12a6b09cba/aiosignal-1.4.0-py3-none-any.whl", hash = "sha256:053243f8b92b990551949e63930a839ff0cf0b0ebbe0597b0f3fb19e1a0fe82e", size = 7490, upload-time = "2025-07-03T22:54:42.156Z" },
+]
+
+[[package]]
+name = "attrs"
+version = "26.1.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/9a/8e/82a0fe20a541c03148528be8cac2408564a6c9a0cc7e9171802bc1d26985/attrs-26.1.0.tar.gz", hash = "sha256:d03ceb89cb322a8fd706d4fb91940737b6642aa36998fe130a9bc96c985eff32", size = 952055, upload-time = "2026-03-19T14:22:25.026Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/64/b4/17d4b0b2a2dc85a6df63d1157e028ed19f90d4cd97c36717afef2bc2f395/attrs-26.1.0-py3-none-any.whl", hash = "sha256:c647aa4a12dfbad9333ca4e71fe62ddc36f4e63b2d260a37a8b83d2f043ac309", size = 67548, upload-time = "2026-03-19T14:22:23.645Z" },
+]
+
+[[package]]
+name = "awscli"
+version = "1.45.28"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "botocore", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "colorama", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "docutils", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "pyyaml", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "rsa", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "s3transfer", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/0e/32/68ff637efdba82e49fd12ddc068fc19edd7f6b4b3ee5bb5976379daaeb25/awscli-1.45.28.tar.gz", hash = "sha256:50e788ed5dc17eb916fd1b9c74d581e4b6a86147c0e60b8b2d25b60ceb32d458", size = 1896334, upload-time = "2026-06-11T19:28:58.014Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/32/41/d9a71dfafd7670e7a79fa4a0d9f073ebbecd33ff4d498f0715d39e4a4464/awscli-1.45.28-py3-none-any.whl", hash = "sha256:783a45041e364c1be9640f730cbcc18313767c592f555477cf0911b1c9511455", size = 4648370, upload-time = "2026-06-11T19:28:54.876Z" },
+]
+
+[[package]]
+name = "boto3"
+version = "1.43.28"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "botocore", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "jmespath", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "s3transfer", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/4e/f2/a976b2a81d8dc7ff675f4b614367a185727061130184a28da0f53f446b97/boto3-1.43.28.tar.gz", hash = "sha256:8391fdcc4d8e1d4e0bf96575a7e5610964a4d401dafa4dccb0a5bade8dd3fbb0", size = 113202, upload-time = "2026-06-11T19:29:01.464Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c3/a5/47db150ea6380f11569b87d3ad064e3c929e5abe227a549d472fab6f5f3a/boto3-1.43.28-py3-none-any.whl", hash = "sha256:4fe6df2163aea02b561eca0d685e2f41a059d71f03721a3e79c3b522e79a3b56", size = 140536, upload-time = "2026-06-11T19:29:00.143Z" },
+]
+
+[[package]]
+name = "botocore"
+version = "1.43.28"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "jmespath", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "python-dateutil", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "urllib3", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/02/dc/1b01808003f88f8a328732c979f20cb0456791048b4440fc4abcae08c1a0/botocore-1.43.28.tar.gz", hash = "sha256:9bbad501a68e4ffdbeff76a382507f5d7827abc316f34a218ab76f5293e6c78d", size = 15503514, upload-time = "2026-06-11T19:28:50.989Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/fa/8c/14916c353ce8a29d14cf6308c2bef842bbb25dde6defc620e26e28063331/botocore-1.43.28-py3-none-any.whl", hash = "sha256:8147adea89b4c9324e842cd8c01ea1a0e17c92cb6ebeaa8cb774f821cb5a7629", size = 15188401, upload-time = "2026-06-11T19:28:47.244Z" },
+]
+
+[[package]]
+name = "certifi"
+version = "2026.5.20"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/f3/ce/ee2ecad540810a79593028e88299baeae54d346cc7a0d94b6199988b89b1/certifi-2026.5.20.tar.gz", hash = "sha256:69dea482ab64caa7b9f6aba1c6bf48bb6a5448d1c0f1b17ab42ad8c763a5344d", size = 135422, upload-time = "2026-05-20T11:46:50.073Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/59/8c/57e832b7af6d7c5abe66eb3fbe3a3a32f4d11ea23a1aa7131371035be991/certifi-2026.5.20-py3-none-any.whl", hash = "sha256:3c52e209ba0a4ad7aebe60436a4ab349c39e1e602e8c134221e546902ad25897", size = 134134, upload-time = "2026-05-20T11:46:48.578Z" },
+]
+
+[[package]]
+name = "charset-normalizer"
+version = "3.4.7"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/e7/a1/67fe25fac3c7642725500a3f6cfe5821ad557c3abb11c9d20d12c7008d3e/charset_normalizer-3.4.7.tar.gz", hash = "sha256:ae89db9e5f98a11a4bf50407d4363e7b09b31e55bc117b4f7d80aab97ba009e5", size = 144271, upload-time = "2026-04-02T09:28:39.342Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/4b/f8/d0118a2f5f23b02cd166fa385c60f9b0d4f9194f574e2b31cef350ad7223/charset_normalizer-3.4.7-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:5649fd1c7bade02f320a462fdefd0b4bd3ce036065836d4f42e0de958038e116", size = 216589, upload-time = "2026-04-02T09:26:29.239Z" },
+    { url = "https://files.pythonhosted.org/packages/8a/2d/a572df5c9204ab7688ec1edc895a73ebded3b023bb07364710b05dd1c9be/charset_normalizer-3.4.7-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:bb6d88045545b26da47aa879dd4a89a71d1dce0f0e549b1abcb31dfe4a8eac49", size = 218468, upload-time = "2026-04-02T09:26:40.17Z" },
+    { url = "https://files.pythonhosted.org/packages/db/8f/61959034484a4a7c527811f4721e75d02d653a35afb0b6054474d8185d4c/charset_normalizer-3.4.7-py3-none-any.whl", hash = "sha256:3dce51d0f5e7951f8bb4900c257dad282f49190fdbebecd4ba99bcc41fef404d", size = 61958, upload-time = "2026-04-02T09:28:37.794Z" },
+]
+
+[[package]]
+name = "cloudpickle"
+version = "3.1.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/27/fb/576f067976d320f5f0114a8d9fa1215425441bb35627b1993e5afd8111e5/cloudpickle-3.1.2.tar.gz", hash = "sha256:7fda9eb655c9c230dab534f1983763de5835249750e85fbcef43aaa30a9a2414", size = 22330, upload-time = "2025-11-03T09:25:26.604Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/88/39/799be3f2f0f38cc727ee3b4f1445fe6d5e4133064ec2e4115069418a5bb6/cloudpickle-3.1.2-py3-none-any.whl", hash = "sha256:9acb47f6afd73f60dc1df93bb801b472f05ff42fa6c84167d25cb206be1fbf4a", size = 22228, upload-time = "2025-11-03T09:25:25.534Z" },
+]
+
+[[package]]
+name = "colorama"
+version = "0.4.6"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/d8/53/6f443c9a4a8358a93a6792e2acffb9d9d5cb0a5cfd8802644b7b1c9a02e4/colorama-0.4.6.tar.gz", hash = "sha256:08695f5cb7ed6e0531a20572697297273c47b8cae5a63ffc6d6ed5c201be6e44", size = 27697, upload-time = "2022-10-25T02:36:22.414Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/d1/d6/3965ed04c63042e047cb6a3e6ed1a63a35087b6a609aa3a15ed8ac56c221/colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6", size = 25335, upload-time = "2022-10-25T02:36:20.889Z" },
+]
+
+[[package]]
+name = "cython"
+version = "0.29.37"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/2a/97/8cc3fe7c6de4796921236a64d00ca8a95565772e57f0d3caae68d880b592/Cython-0.29.37.tar.gz", hash = "sha256:f813d4a6dd94adee5d4ff266191d1d95bf6d4164a4facc535422c021b2504cfb", size = 2099621, upload-time = "2023-12-19T09:22:39.397Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c9/aa/99a0eac01136c0c75feb3210d107c49f93d49d5cb97f19e99318b9ecefdd/Cython-0.29.37-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e8af5975ecfae254d8c0051204fca995dda8f93cf9f0bbf7571e3cda2b0cef4d", size = 1951525, upload-time = "2023-12-19T09:57:00.464Z" },
+    { url = "https://files.pythonhosted.org/packages/14/5f/f5efba4409474892d650ba7e57349afa5d4c41d06f5ba3356c32891c75a6/Cython-0.29.37-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:29415d8eb2fdc1ea518ca4810c50a2d062b387d4c9fbcfb3352346e93db22c6d", size = 1950936, upload-time = "2023-12-19T09:24:07.949Z" },
+    { url = "https://files.pythonhosted.org/packages/7e/26/9d8de10005fedb1eceabe713348d43bae1dbab1786042ca0751a2e2b0f8c/Cython-0.29.37-py2.py3-none-any.whl", hash = "sha256:95f1d6a83ef2729e67b3fa7318c829ce5b07ac64c084cd6af11c228e0364662c", size = 989503, upload-time = "2023-12-19T09:22:32.861Z" },
+]
+
+[[package]]
+name = "docutils"
+version = "0.19"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/6b/5c/330ea8d383eb2ce973df34d1239b3b21e91cd8c865d21ff82902d952f91f/docutils-0.19.tar.gz", hash = "sha256:33995a6753c30b7f577febfc2c50411fec6aac7f7ffeb7c4cfe5991072dcf9e6", size = 2056383, upload-time = "2022-07-05T20:17:31.045Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/93/69/e391bd51bc08ed9141ecd899a0ddb61ab6465309f1eb470905c0c8868081/docutils-0.19-py3-none-any.whl", hash = "sha256:5e1de4d849fee02c63b040a4a3fd567f4ab104defd8a5511fbbc24a8a017efbc", size = 570472, upload-time = "2022-07-05T20:17:26.388Z" },
+]
+
+[[package]]
+name = "falcon"
+version = "3.1.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/36/53/4fd90c6c841bc2e4be29ab92c65e5406df9096c421f138bef9d95d43afc9/falcon-3.1.0.tar.gz", hash = "sha256:f2760bd18c16393a6fb5e55f371f67921edb72febe693a82b3c5e82195d087b7", size = 645631, upload-time = "2022-03-25T16:13:19.137Z" }
+
+[[package]]
+name = "frozenlist"
+version = "1.8.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/2d/f5/c831fac6cc817d26fd54c7eaccd04ef7e0288806943f7cc5bbf69f3ac1f0/frozenlist-1.8.0.tar.gz", hash = "sha256:3ede829ed8d842f6cd48fc7081d7a41001a56f1f38603f9d49bf3020d59a31ad", size = 45875, upload-time = "2025-10-06T05:38:17.865Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/6a/bd/d91c5e39f490a49df14320f4e8c80161cfcce09f1e2cde1edd16a551abb3/frozenlist-1.8.0-cp312-cp312-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:494a5952b1c597ba44e0e78113a7266e656b9794eec897b19ead706bd7074383", size = 242411, upload-time = "2025-10-06T05:36:09.801Z" },
+    { url = "https://files.pythonhosted.org/packages/8f/96/007e5944694d66123183845a106547a15944fbbb7154788cbf7272789536/frozenlist-1.8.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:776f352e8329135506a1d6bf16ac3f87bc25b28e765949282dcc627af36123aa", size = 239096, upload-time = "2025-10-06T05:36:22.129Z" },
+    { url = "https://files.pythonhosted.org/packages/9a/9a/e35b4a917281c0b8419d4207f4334c8e8c5dbf4f3f5f9ada73958d937dcc/frozenlist-1.8.0-py3-none-any.whl", hash = "sha256:0c18a16eab41e82c295618a77502e17b195883241c563b00f0aa5106fc4eaa0d", size = 13409, upload-time = "2025-10-06T05:38:16.721Z" },
+]
+
+[[package]]
+name = "gevent"
+version = "26.5.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "greenlet", marker = "platform_machine == 'x86_64' and platform_python_implementation == 'CPython' and sys_platform == 'linux'" },
+    { name = "zope-event", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "zope-interface", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/c4/cb/98aa3a299e2fc4a2372b5d124863e02965b64579ffc29fe54d0641e65b2f/gevent-26.5.0.tar.gz", hash = "sha256:1655eb04c1e20d71b2aa4a3c7528162dd58ff6cc46a037af1f01f534c80fefba", size = 6712354, upload-time = "2026-05-20T21:22:45.132Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/4e/69/1559b1f6b5107a9118fccd300240879bd581b6d87b03d568d0d155ea702c/gevent-26.5.0-cp312-cp312-manylinux_2_28_x86_64.whl", hash = "sha256:bccff69c462e3650a0fd1d4e9cfc8b6effe15f3e9b1cad20a7bb5ce14b057efd", size = 2114915, upload-time = "2026-05-20T20:35:25.041Z" },
+    { url = "https://files.pythonhosted.org/packages/f9/3c/2fe77ee6e3d381b3c50c0b7d6c4c08c08b8ff5e8c0d9dd51a3b426d61eec/gevent-26.5.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:0bf57df54f1c66273bf3601c2a1e41b12138fe848933718369663bc54f177ca2", size = 2140806, upload-time = "2026-05-20T20:43:28.895Z" },
+]
+
+[[package]]
+name = "greenlet"
+version = "3.5.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/6d/6e/802acd792aebb2256fbbee8cacf2727faaeb6f240ac11008f09eae4414bc/greenlet-3.5.1.tar.gz", hash = "sha256:5a56aeb7d5d9cc4b3a735efb5095bd4b4f6f0e4f93e5ca876d0e2315137b7829", size = 197356, upload-time = "2026-05-20T15:05:03.917Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/40/69/b91cda0647df839483201545913514c2827ebea5e5ccdf931842763bc127/greenlet-3.5.1-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:add5217d68b31130f0beca584d7fef4878327d2e31642b66618a14eef312b63b", size = 611358, upload-time = "2026-05-20T13:14:26.37Z" },
+    { url = "https://files.pythonhosted.org/packages/65/2c/18cece62045e74598c3c393f70dce4a63f56222015ba29a5d4eeb04f764c/greenlet-3.5.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:c5551170cf4f5ff5623e9af81323751979fee2c731e2287b61f73cd27257b823", size = 1635625, upload-time = "2026-05-20T13:14:34.027Z" },
+]
+
+[[package]]
+name = "grpcio"
+version = "1.81.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "typing-extensions", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/b0/b5/1ff353970a87eda4c98251e34d2dfd214abd4982dc89119c9252a2a482d2/grpcio-1.81.1.tar.gz", hash = "sha256:6fa10a767143a5e82e8eaab53918af0cd8909a57a27f8cb2288b80a613ac671b", size = 13026582, upload-time = "2026-06-11T12:46:51.673Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/eb/8c/dea020b6d91508cd84463917a63149ec196ee7db505d032ae43fcb3303b9/grpcio-1.81.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:766bc7c9a9c340342f4c864ccbda8e78111e4751f13b895812b9c148fb79e9d0", size = 6809167, upload-time = "2026-06-11T12:45:32.52Z" },
+    { url = "https://files.pythonhosted.org/packages/25/7a/71437c7f3596e5246155c515852795a85a1a8d228190212432b13b97a95d/grpcio-1.81.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:8ea1936c26b99999b27479853039a7f34713f56c49375ad52b38535ec93a796c", size = 7849660, upload-time = "2026-06-11T12:45:40.627Z" },
+]
+
+[[package]]
+name = "gunicorn"
+version = "26.0.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "packaging", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/6d/b7/a4a3f632f823e432ce6bc65f62961b7980c898c77f075a2f7118cb3846fe/gunicorn-26.0.0.tar.gz", hash = "sha256:ca9346f85e3a4aeeb64d491045c16b9a35647abd37ea15efe53080eb8b090baf", size = 727286, upload-time = "2026-05-05T06:38:25.529Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e6/40/9c2384fc2be4ad25dd4a49decd5ad9ea5a3639814c11bd40ab77cb9f0a14/gunicorn-26.0.0-py3-none-any.whl", hash = "sha256:40233d26a5f0d1872916188c276e21641155111c2853f0c2cd55260aec0d24fc", size = 212009, upload-time = "2026-05-05T06:38:23.007Z" },
+]
+
+[[package]]
+name = "idna"
+version = "3.18"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/cd/63/9496c57188a2ee585e0f1db071d75089a11e98aa86eb99d9d7618fc1edce/idna-3.18.tar.gz", hash = "sha256:ffb385a7e039654cef1ab9ef32c6fafe283c0c0467bba1d9029738ce4a14a848", size = 196711, upload-time = "2026-06-02T14:34:07.794Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/1e/5e/d4e9f1a599fb8e573b7b87160658329fbf28d19eac2718f51fc3def3aa5a/idna-3.18-py3-none-any.whl", hash = "sha256:7f952cbe720b688055e3f87de14f5c3e5fdaa8bc3928985c4077ca689de849a2", size = 65455, upload-time = "2026-06-02T14:34:06.319Z" },
+]
+
+[[package]]
+name = "jmespath"
+version = "1.1.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/d3/59/322338183ecda247fb5d1763a6cbe46eff7222eaeebafd9fa65d4bf5cb11/jmespath-1.1.0.tar.gz", hash = "sha256:472c87d80f36026ae83c6ddd0f1d05d4e510134ed462851fd5f754c8c3cbb88d", size = 27377, upload-time = "2026-01-22T16:35:26.279Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/14/2f/967ba146e6d58cf6a652da73885f52fc68001525b4197effc174321d70b4/jmespath-1.1.0-py3-none-any.whl", hash = "sha256:a5663118de4908c91729bea0acadca56526eb2698e83de10cd116ae0f4e97c64", size = 20419, upload-time = "2026-01-22T16:35:24.919Z" },
+]
+
+[[package]]
+name = "joblib"
+version = "1.5.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/41/f2/d34e8b3a08a9cc79a50b2208a93dce981fe615b64d5a4d4abee421d898df/joblib-1.5.3.tar.gz", hash = "sha256:8561a3269e6801106863fd0d6d84bb737be9e7631e33aaed3fb9ce5953688da3", size = 331603, upload-time = "2025-12-15T08:41:46.427Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/7b/91/984aca2ec129e2757d1e4e3c81c3fcda9d0f85b74670a094cc443d9ee949/joblib-1.5.3-py3-none-any.whl", hash = "sha256:5fc3c5039fc5ca8c0276333a188bbd59d6b7ab37fe6632daa76bc7f9ec18e713", size = 309071, upload-time = "2025-12-15T08:41:44.973Z" },
+]
+
+[[package]]
+name = "multidict"
+version = "6.7.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/1a/c2/c2d94cbe6ac1753f3fc980da97b3d930efe1da3af3c9f5125354436c073d/multidict-6.7.1.tar.gz", hash = "sha256:ec6652a1bee61c53a3e5776b6049172c53b6aaba34f18c9ad04f82712bac623d", size = 102010, upload-time = "2026-01-26T02:46:45.979Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/f3/8d/5e5be3ced1d12966fefb5c4ea3b2a5b480afcea36406559442c6e31d4a48/multidict-6.7.1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:bfde23ef6ed9db7eaee6c37dcec08524cb43903c60b285b172b6c094711b3961", size = 256322, upload-time = "2026-01-26T02:44:03.56Z" },
+    { url = "https://files.pythonhosted.org/packages/2b/17/ebb9644da78c4ab36403739e0e6e0e30ebb135b9caf3440825001a0bddcb/multidict-6.7.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:fc5907494fccf3e7d3f94f95c91d6336b092b5fc83811720fae5e2765890dfba", size = 251377, upload-time = "2026-01-26T02:44:12.042Z" },
+    { url = "https://files.pythonhosted.org/packages/81/08/7036c080d7117f28a4af526d794aab6a84463126db031b007717c1a6676e/multidict-6.7.1-py3-none-any.whl", hash = "sha256:55d97cc6dae627efa6a6e548885712d4864b81110ac76fa4e534c03819fa4a56", size = 12319, upload-time = "2026-01-26T02:46:44.004Z" },
+]
+
+[[package]]
+name = "narwhals"
+version = "2.22.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/62/3c/c4ef2164a71c1a63d7f1ae411c4082c5fa872405106db60a4b7114989ad7/narwhals-2.22.1.tar.gz", hash = "sha256:d62920805a0a43b7ff8b54b0c0d3142d796f8a9301836ada37e573d6a33cbcd9", size = 647493, upload-time = "2026-06-05T12:34:34.051Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/48/ca/36339329c4604adbcc99c899b7eb1ce1a555c499b6a6860757dc9bfed36d/narwhals-2.22.1-py3-none-any.whl", hash = "sha256:60567d774edf77db53906f89d9fbd164e66e56d66d388e1e6990f17ac33cfb53", size = 454815, upload-time = "2026-06-05T12:34:32.289Z" },
+]
+
+[[package]]
+name = "numpy"
+version = "1.26.4"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/65/6e/09db70a523a96d25e115e71cc56a6f9031e7b8cd166c1ac8438307c14058/numpy-1.26.4.tar.gz", hash = "sha256:2a02aba9ed12e4ac4eb3ea9421c420301a0c6460d9830d74a9df87efa4912010", size = 15786129, upload-time = "2024-02-06T00:26:44.495Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/0f/50/de23fde84e45f5c4fda2488c759b69990fd4512387a8632860f3ac9cd225/numpy-1.26.4-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:675d61ffbfa78604709862923189bad94014bef562cc35cf61d3a07bba02a7ed", size = 17950613, upload-time = "2024-02-05T23:56:56.054Z" },
+    { url = "https://files.pythonhosted.org/packages/76/8c/2ba3902e1a0fc1c74962ea9bb33a534bb05984ad7ff9515bf8d07527cadd/numpy-1.26.4-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:1dda2e7b4ec9dd512f84935c5f126c8bd8b9f2fc001e9f54af255e8c5f16b0e0", size = 17786643, upload-time = "2024-02-05T23:57:56.585Z" },
+]
+
+[[package]]
+name = "packaging"
+version = "26.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/d7/f1/e7a6dd94a8d4a5626c03e4e99c87f241ba9e350cd9e6d75123f992427270/packaging-26.2.tar.gz", hash = "sha256:ff452ff5a3e828ce110190feff1178bb1f2ea2281fa2075aadb987c2fb221661", size = 228134, upload-time = "2026-04-24T20:15:23.917Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/df/b2/87e62e8c3e2f4b32e5fe99e0b86d576da1312593b39f47d8ceef365e95ed/packaging-26.2-py3-none-any.whl", hash = "sha256:5fc45236b9446107ff2415ce77c807cee2862cb6fac22b8a73826d0693b0980e", size = 100195, upload-time = "2026-04-24T20:15:22.081Z" },
+]
+
+[[package]]
+name = "pandas"
+version = "3.0.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "python-dateutil", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/f8/87/4341c6252d1c47b08768c3d25ac487362bf403f0313ddae4a2a26c9b1b4c/pandas-3.0.3.tar.gz", hash = "sha256:696a4a00a2a2a35d4e5deb3fc946641b96c944f02230e4f76137fe35d806c4fc", size = 4651414, upload-time = "2026-05-11T18:54:29.21Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/65/b6/09b01cdbc15224e2850365192d17b7bdebb8bdbd8780ed221fcdf0d9a515/pandas-3.0.3-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6dc0b3fd2169c9157deed50b4d519553a3655c8c6a96027136d654592be973a9", size = 10894600, upload-time = "2026-05-11T18:52:45.02Z" },
+    { url = "https://files.pythonhosted.org/packages/f8/45/830bb57f533a4604b355e07edcb8ea18cf88b5f94e5fca92f27052d7c597/pandas-3.0.3-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:f8894dc474d648fe7b6ff0ca9b0bd73950d19952bc1a6534540762c5d79d305c", size = 11950889, upload-time = "2026-05-11T18:52:50.905Z" },
+]
+
+[[package]]
+name = "propcache"
+version = "0.5.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/ec/44/c87281c333769159c50594f22610f77398a47ccbfbbf23074e744e86f87c/propcache-0.5.2.tar.gz", hash = "sha256:01c4fc7480cd0598bb4b57022df55b9ca296da7fc5a8760bd8451a7e63a7d427", size = 50208, upload-time = "2026-05-08T21:02:12.199Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/14/15/5574111ae50dd6e879456888c0eadd4c5a869959775854e18e18a6b345f3/propcache-0.5.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6f328175a2cde1f0ff2c4ed8ce968b9dcfb55f3a7153f39e2957ed994da13476", size = 61639, upload-time = "2026-05-08T21:00:19.692Z" },
+    { url = "https://files.pythonhosted.org/packages/12/fd/77fe5936d8c3086ca9048f7f415f122ed82e53884a9ec193646b42deef06/propcache-0.5.2-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:c80f4ba3e8f00189165999a742ee526ebeccedf6c3f7beb0c7df821e9772435a", size = 62514, upload-time = "2026-05-08T21:00:30.098Z" },
+    { url = "https://files.pythonhosted.org/packages/3a/ed/1cdcab6ba3d6ab7feca11fc14f0eeea80755bb53ef4e892079f31b10a25f/propcache-0.5.2-py3-none-any.whl", hash = "sha256:be1ddfcbb376e3de5d2e2db1d58d6d67463e6b4f9f040c000de8e300295465fe", size = 14036, upload-time = "2026-05-08T21:02:10.673Z" },
+]
+
+[[package]]
+name = "protobuf"
+version = "6.33.6"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/66/70/e908e9c5e52ef7c3a6c7902c9dfbb34c7e29c25d2f81ade3856445fd5c94/protobuf-6.33.6.tar.gz", hash = "sha256:a6768d25248312c297558af96a9f9c929e8c4cee0659cb07e780731095f38135", size = 444531, upload-time = "2026-03-18T19:05:00.988Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/16/92/d1e32e3e0d894fe00b15ce28ad4944ab692713f2e7f0a99787405e43533a/protobuf-6.33.6-cp39-abi3-manylinux2014_x86_64.whl", hash = "sha256:e9db7e292e0ab79dd108d7f1a94fe31601ce1ee3f7b79e0692043423020b0593", size = 323436, upload-time = "2026-03-18T19:04:55.768Z" },
+    { url = "https://files.pythonhosted.org/packages/c4/72/02445137af02769918a93807b2b7890047c32bfb9f90371cbc12688819eb/protobuf-6.33.6-py3-none-any.whl", hash = "sha256:77179e006c476e69bf8e8ce866640091ec42e1beb80b213c3900006ecfba6901", size = 170656, upload-time = "2026-03-18T19:04:59.826Z" },
+]
+
+[[package]]
+name = "pyasn1"
+version = "0.6.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/5c/5f/6583902b6f79b399c9c40674ac384fd9cd77805f9e6205075f828ef11fb2/pyasn1-0.6.3.tar.gz", hash = "sha256:697a8ecd6d98891189184ca1fa05d1bb00e2f84b5977c481452050549c8a72cf", size = 148685, upload-time = "2026-03-17T01:06:53.382Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/5d/a0/7d793dce3fa811fe047d6ae2431c672364b462850c6235ae306c0efd025f/pyasn1-0.6.3-py3-none-any.whl", hash = "sha256:a80184d120f0864a52a073acc6fc642847d0be408e7c7252f31390c0f4eadcde", size = 83997, upload-time = "2026-03-17T01:06:52.036Z" },
+]
+
+[[package]]
+name = "python-dateutil"
+version = "2.9.0.post0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "six", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/66/c0/0c8b6ad9f17a802ee498c46e004a0eb49bc148f2fd230864601a86dcf6db/python-dateutil-2.9.0.post0.tar.gz", hash = "sha256:37dd54208da7e1cd875388217d5e00ebd4179249f90fb72437e91a35459a0ad3", size = 342432, upload-time = "2024-03-01T18:36:20.211Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/ec/57/56b9bcc3c9c6a792fcbaf139543cee77261f3651ca9da0c93f5c1221264b/python_dateutil-2.9.0.post0-py2.py3-none-any.whl", hash = "sha256:a8b2bc7bffae282281c8140a97d3aa9c14da0b136dfe83f850eea9a5f7470427", size = 229892, upload-time = "2024-03-01T18:36:18.57Z" },
+]
+
+[[package]]
+name = "pyyaml"
+version = "6.0.3"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/05/8e/961c0007c59b8dd7729d542c61a4d537767a59645b82a0b521206e1e25c2/pyyaml-6.0.3.tar.gz", hash = "sha256:d76623373421df22fb4cf8817020cbb7ef15c725b9d5e45f17e189bfc384190f", size = 130960, upload-time = "2025-09-25T21:33:16.546Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/8b/9d/b3589d3877982d4f2329302ef98a8026e7f4443c765c46cfecc8858c6b4b/pyyaml-6.0.3-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:ba1cc08a7ccde2d2ec775841541641e4548226580ab850948cbfda66a1befcdc", size = 807870, upload-time = "2025-09-25T21:32:16.431Z" },
+    { url = "https://files.pythonhosted.org/packages/be/8e/98435a21d1d4b46590d5459a22d88128103f8da4c2d4cb8f14f2a96504e1/pyyaml-6.0.3-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:41715c910c881bc081f1e8872880d3c650acf13dfa8214bad49ed4cede7c34ea", size = 790181, upload-time = "2025-09-25T21:32:18.834Z" },
+]
+
+[[package]]
+name = "requests"
+version = "2.34.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "certifi", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "charset-normalizer", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "idna", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "urllib3", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/ac/c3/e2a2b89f2d3e2179abd6d00ebd70bff6273f37fb3e0cc209f48b39d00cbf/requests-2.34.2.tar.gz", hash = "sha256:f288924cae4e29463698d6d60bc6a4da69c89185ad1e0bcc4104f584e960b9ed", size = 142856, upload-time = "2026-05-14T19:25:27.735Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/a0/f4/c67b0b3f1b9245e8d266f0f112c500d50e5b4e83cb6f3b71b6528104182a/requests-2.34.2-py3-none-any.whl", hash = "sha256:2a0d60c172f83ac6ab31e4554906c0f3b3588d37b5cb939b1c061f4907e278e0", size = 73075, upload-time = "2026-05-14T19:25:26.443Z" },
+]
+
+[[package]]
+name = "rsa"
+version = "4.7.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "pyasn1", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/db/b5/475c45a58650b0580421746504b680cd2db4e81bc941e94ca53785250269/rsa-4.7.2.tar.gz", hash = "sha256:9d689e6ca1b3038bc82bf8d23e944b6b6037bc02301a574935b2dd946e0353b9", size = 39711, upload-time = "2021-02-24T10:55:05.846Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/e9/93/0c0f002031f18b53af7a6166103c02b9c0667be528944137cc954ec921b3/rsa-4.7.2-py3-none-any.whl", hash = "sha256:78f9a9bf4e7be0c5ded4583326e7461e3a3c5aae24073648b4bdfa797d78c9d2", size = 34505, upload-time = "2021-02-24T10:55:03.55Z" },
+]
+
+[[package]]
+name = "s3transfer"
+version = "0.18.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "botocore", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/e0/1f/12417f7f493fc45e1f9fd5d4a9b6c125cf8d2cf3f8ddbdfab3e76406e9d6/s3transfer-0.18.0.tar.gz", hash = "sha256:3760b8b7ec1315da54048b2d626276732bee4300d054d492d4e1d43e20d4ecbd", size = 160560, upload-time = "2026-05-28T19:39:09.124Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/2b/58/a58fc997655386daa2e25784e30c288aa3e3819e401f77029ee4899fb55a/s3transfer-0.18.0-py3-none-any.whl", hash = "sha256:239c13b09e65ad0346e1be7348b8a202dcad44ac7ea7c6eb858fc881dce739b6", size = 88572, upload-time = "2026-05-28T19:39:07.999Z" },
+]
+
+[[package]]
+name = "scikit-learn"
+version = "1.9.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "joblib", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "narwhals", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "numpy", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "scipy", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "threadpoolctl", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/fa/6f/37092bdb25f712817231799fc5674d8e704066a8a70c1d2d40517e18b4ab/scikit_learn-1.9.0.tar.gz", hash = "sha256:8833266989d3a5110178a9fae30783675460724d0e1efb13b14901d2c660c557", size = 7750767, upload-time = "2026-06-02T11:54:32.706Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/6c/c2/63fdda36c56437eeb44aaf9493c8bcd62ce230ab1598924fc626ffbfa943/scikit_learn-1.9.0-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:056c92bb67ad4c28463c2f2653d9701449201e7e7a9e94e321be0f71c4fef2b8", size = 9132097, upload-time = "2026-06-02T11:53:33.456Z" },
+]
+
+[[package]]
+name = "scipy"
+version = "1.17.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "numpy", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/7a/97/5a3609c4f8d58b039179648e62dd220f89864f56f7357f5d4f45c29eb2cc/scipy-1.17.1.tar.gz", hash = "sha256:95d8e012d8cb8816c226aef832200b1d45109ed4464303e997c5b13122b297c0", size = 30573822, upload-time = "2026-02-23T00:26:24.851Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/01/8e/1e35281b8ab6d5d72ebe9911edcdffa3f36b04ed9d51dec6dd140396e220/scipy-1.17.1-cp312-cp312-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:02ae3b274fde71c5e92ac4d54bc06c42d80e399fec704383dcd99b301df37458", size = 35235890, upload-time = "2026-02-23T00:18:49.188Z" },
+    { url = "https://files.pythonhosted.org/packages/65/94/7698add8f276dbab7a9de9fb6b0e02fc13ee61d51c7c3f85ac28b65e1239/scipy-1.17.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:f590cd684941912d10becc07325a3eeb77886fe981415660d9265c4c418d0bea", size = 37625856, upload-time = "2026-02-23T00:19:00.307Z" },
+]
+
+[[package]]
+name = "six"
+version = "1.17.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/94/e7/b2c673351809dca68a0e064b6af791aa332cf192da575fd474ed7d6f16a2/six-1.17.0.tar.gz", hash = "sha256:ff70335d468e7eb6ec65b95b99d3a2836546063f63acc5171de367e834932a81", size = 34031, upload-time = "2024-12-04T17:35:28.174Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/b7/ce/149a00dd41f10bc29e5921b496af8b574d8413afcd5e30dfa0ed46c2cc5e/six-1.17.0-py2.py3-none-any.whl", hash = "sha256:4721f391ed90541fddacab5acf947aa0d3dc7d27b2e1e8eda2be8970586c3274", size = 11050, upload-time = "2024-12-04T17:35:26.475Z" },
+]
+
+[[package]]
+name = "tensorflow-serving-inference-dlc"
+version = "2.20.0"
+source = { virtual = "." }
+dependencies = [
+    { name = "aiohttp", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "awscli", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "boto3", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "botocore", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "cython", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "falcon", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "gevent", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "grpcio", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "gunicorn", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "numpy", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "packaging", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "protobuf", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "requests", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "urllib3", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+
+[package.optional-dependencies]
+sagemaker = [
+    { name = "cloudpickle", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "pandas", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "scikit-learn", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+
+[package.metadata]
+requires-dist = [
+    { name = "aiohttp", specifier = ">=3.14.0" },
+    { name = "awscli", specifier = "<2" },
+    { name = "boto3" },
+    { name = "botocore" },
+    { name = "cloudpickle", marker = "extra == 'sagemaker'" },
+    { name = "cython", specifier = "<3" },
+    { name = "falcon", specifier = "==3.1.0" },
+    { name = "gevent" },
+    { name = "grpcio", specifier = ">=1.0" },
+    { name = "gunicorn", specifier = ">=22.0.0" },
+    { name = "numpy", specifier = "==1.26.4" },
+    { name = "packaging" },
+    { name = "pandas", marker = "extra == 'sagemaker'" },
+    { name = "protobuf", specifier = ">=3.20,<7" },
+    { name = "requests" },
+    { name = "scikit-learn", marker = "extra == 'sagemaker'" },
+    { name = "urllib3" },
+]
+provides-extras = ["sagemaker"]
+
+[[package]]
+name = "threadpoolctl"
+version = "3.6.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/b7/4d/08c89e34946fce2aec4fbb45c9016efd5f4d7f24af8e5d93296e935631d8/threadpoolctl-3.6.0.tar.gz", hash = "sha256:8ab8b4aa3491d812b623328249fab5302a68d2d71745c8a4c719a2fcaba9f44e", size = 21274, upload-time = "2025-03-13T13:49:23.031Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/32/d5/f9a850d79b0851d1d4ef6456097579a9005b31fea68726a4ae5f2d82ddd9/threadpoolctl-3.6.0-py3-none-any.whl", hash = "sha256:43a0b8fd5a2928500110039e43a5eed8480b918967083ea48dc3ab9f13c4a7fb", size = 18638, upload-time = "2025-03-13T13:49:21.846Z" },
+]
+
+[[package]]
+name = "typing-extensions"
+version = "4.15.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/72/94/1a15dd82efb362ac84269196e94cf00f187f7ed21c242792a923cdb1c61f/typing_extensions-4.15.0.tar.gz", hash = "sha256:0cea48d173cc12fa28ecabc3b837ea3cf6f38c6d1136f85cbaaf598984861466", size = 109391, upload-time = "2025-08-25T13:49:26.313Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/18/67/36e9267722cc04a6b9f15c7f3441c2363321a3ea07da7ae0c0707beb2a9c/typing_extensions-4.15.0-py3-none-any.whl", hash = "sha256:f0fa19c6845758ab08074a0cfa8b7aecb71c999ca73d62883bc25cc018c4e548", size = 44614, upload-time = "2025-08-25T13:49:24.86Z" },
+]
+
+[[package]]
+name = "urllib3"
+version = "2.7.0"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/53/0c/06f8b233b8fd13b9e5ee11424ef85419ba0d8ba0b3138bf360be2ff56953/urllib3-2.7.0.tar.gz", hash = "sha256:231e0ec3b63ceb14667c67be60f2f2c40a518cb38b03af60abc813da26505f4c", size = 433602, upload-time = "2026-05-07T16:13:18.596Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/7f/3e/5db95bcf282c52709639744ca2a8b149baccf648e39c8cc87553df9eae0c/urllib3-2.7.0-py3-none-any.whl", hash = "sha256:9fb4c81ebbb1ce9531cce37674bbc6f1360472bc18ca9a553ede278ef7276897", size = 131087, upload-time = "2026-05-07T16:13:17.151Z" },
+]
+
+[[package]]
+name = "yarl"
+version = "1.24.2"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "idna", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "multidict", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "propcache", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/79/12/1e8f37460ea0f7eb59c221fdaf0ed75e7ac43e97f8093b9c6f411df50a78/yarl-1.24.2.tar.gz", hash = "sha256:9ac374123c6fd7abf64d1fec93962b0bd4ee2c19751755a762a72dd96c0378f8", size = 210798, upload-time = "2026-05-19T21:31:05.599Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/7a/da/323a01c349bd5fb01bb6652e314d9bb218cee630a736bdb810ad50e4013f/yarl-1.24.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7e7ebcdef69dec6c6451e616f32b622a6d4a2e92b445c992f7c8e5274a6bbc4c", size = 105612, upload-time = "2026-05-19T21:29:04.247Z" },
+    { url = "https://files.pythonhosted.org/packages/48/2d/1c8d89c7c5f9cad9fb2902445d94e2ab1d7aa35de029afbb8ae95c42d00f/yarl-1.24.2-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:e30dd55825dc554ec5b66a94953b8eda8745926514c5089dfcacecb9c99b5bd1", size = 105719, upload-time = "2026-05-19T21:29:18.367Z" },
+    { url = "https://files.pythonhosted.org/packages/fd/4d/4b880086bd0d3e034d25647be1d830afc3e3f610e98c4ab3490af6b1b6d5/yarl-1.24.2-py3-none-any.whl", hash = "sha256:2783d9226db8797636cd6896e4de81feed252d1db72265686c9558d97a4d94b9", size = 53576, upload-time = "2026-05-19T21:31:03.909Z" },
+]
+
+[[package]]
+name = "zope-event"
+version = "6.2"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/93/41/faa10af34d48d9cd6fa0249a1162943ad84a9590bd1a06939981e6640416/zope_event-6.2.tar.gz", hash = "sha256:b97d5d6327067ee6b9dfcbdf606ade9ade70991e19c162e808ea39e5fcf0f8d3", size = 18958, upload-time = "2026-04-28T06:24:10.578Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/9e/33/848922889e946d4befc415c219fe516af75c49555d8e736e183bfd30db42/zope_event-6.2-py3-none-any.whl", hash = "sha256:5e755153ac4faf64c10a4b6dd3307680166a3edf65b38df22df592610f8fa874", size = 6525, upload-time = "2026-04-28T06:24:09.176Z" },
+]
+
+[[package]]
+name = "zope-interface"
+version = "8.5"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/08/dc/50550cfcbb2ea3cbca5f1d7ed05c8aa840f831a0f2d63aec0a953f7c590e/zope_interface-8.5.tar.gz", hash = "sha256:7a3ba1c5877f0f3e3906b02ddf793abed2becc2948116414ce0e1dd820b68d6d", size = 257957, upload-time = "2026-05-26T06:50:14.574Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c5/66/1036543d6a66bc04c19df3cf650f3ad938a002ab0a443c24e23e8de5e8b9/zope_interface-8.5-cp312-cp312-manylinux1_x86_64.manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:5e970dabea777a24b0b0bbf9dae3ab75ce8b2d8e948edf4875627034b21f3560", size = 270689, upload-time = "2026-05-26T06:49:31.767Z" },
+]
diff --git a/docker/tensorflow/inference/2.20/versions-cpu.env b/docker/tensorflow/inference/2.20/versions-cpu.env
new file mode 100644
index 000000000000..dde4c749e6e6
--- /dev/null
+++ b/docker/tensorflow/inference/2.20/versions-cpu.env
@@ -0,0 +1,32 @@
+# versions-cpu.env — Pinned versions for TensorFlow Serving 2.20 CPU AL2023 inference DLC
+# Source this file: source versions-cpu.env
+
+# ── Image metadata ──────────────────────────────────────────────
+export DLC_MAJOR_VERSION="1"
+export DLC_MINOR_VERSION="0"
+export BUILD_DATE="$(date +%Y%m%d)"
+
+# ── Python ──────────────────────────────────────────────────────
+export PYTHON_VERSION="3.12"
+
+# ── TensorFlow Serving ─────────────────────────────────────────
+# The inference image ships the tensorflow_model_server binary copied out of
+# tensorflow/serving:2.20.0-devel. PyPI tensorflow-serving-api==2.20.0 is the
+# matched gRPC stub (installed inline in the Dockerfile via --no-deps; see locked
+# decision Q1 in cpu/pyproject.toml header).
+# Note: this is TF Serving's version, NOT the framework version (2.21). TF Serving's
+# release cadence trails the framework — 2.20 is the latest published serving release.
+export TF_SERVING_VERSION="2.20.0"
+
+# Git commit pin on the tensorflow/serving repo (branch r2.20). Used by
+# `setup.sources.sh` and OSS-compliance manifest references. Sourced from
+# `git ls-remote https://github.com/tensorflow/serving refs/heads/r2.20` at the
+# start of Phase 2 (2026-06-11).
+export TF_SERVING_VERSION_GIT_COMMIT="bc7e9d2b9a419294c4526e293c2df8727b0d3116"
+
+# ── nginx + njs (compiled in builder-njs stage) ────────────────
+# AL2023 base repo ships nginx 1.24.x without nginx-module-njs; we therefore
+# build njs as a dynamic module from source against a matched nginx source tree.
+# Same versions as the CUDA variant — the njs build does not depend on CUDA.
+export NGINX_VERSION="1.30.2"
+export NJS_VERSION="0.9.9"
diff --git a/docker/tensorflow/inference/2.20/versions-cuda.env b/docker/tensorflow/inference/2.20/versions-cuda.env
new file mode 100644
index 000000000000..ee5d7f7cbd0f
--- /dev/null
+++ b/docker/tensorflow/inference/2.20/versions-cuda.env
@@ -0,0 +1,46 @@
+# versions-cuda.env — Pinned versions for TensorFlow Serving 2.20 CUDA AL2023 inference DLC
+# Source this file: source versions-cuda.env
+
+# ── Image metadata ──────────────────────────────────────────────
+export DLC_MAJOR_VERSION="1"
+export DLC_MINOR_VERSION="0"
+export BUILD_DATE="$(date +%Y%m%d)"
+
+# ── CUDA ───────────────────────────────────────────────────────
+# CUDA 12.9.1 — matches the TF 2.21 training image base. tensorflow/serving:2.20.0-devel-gpu
+# is compiled against CUDA 12.x (upstream-pinned by Google's TFS build matrix); we rely on
+# CUDA minor-version forward compatibility to run the serving binary on CUDA 12.9.1 runtime.
+# This is the same compat chain the training image uses for the TF 2.21 wheel built against
+# CUDA 12.5 — verified empirically there.
+export CUDA_VERSION="12.9.1"
+
+# ── Python ──────────────────────────────────────────────────────
+export PYTHON_VERSION="3.12"
+
+# ── TensorFlow Serving ─────────────────────────────────────────
+# The inference image ships the tensorflow_model_server binary copied out of
+# tensorflow/serving:2.20.0-devel-gpu. PyPI tensorflow-serving-api-gpu==2.20.0 is the
+# matched gRPC stub (installed inline in the Dockerfile via --no-deps; see locked
+# decision Q1 in cuda/pyproject.toml header).
+# Note: this is TF Serving's version, NOT the framework version (2.21). TF Serving's
+# release cadence trails the framework — 2.20 is the latest published serving release.
+export TF_SERVING_VERSION="2.20.0"
+
+# Git commit pin on the tensorflow/serving repo (branch r2.20). Used by
+# `setup.sources.sh` for any apt source bootstrapping that the upstream TFS build
+# scripts perform during the binary copy or by reference inside the OSS-compliance
+# manifest. Sourced from `git ls-remote https://github.com/tensorflow/serving refs/heads/r2.20`
+# at the start of Phase 2 (2026-06-11).
+export TF_SERVING_VERSION_GIT_COMMIT="bc7e9d2b9a419294c4526e293c2df8727b0d3116"
+
+# ── nginx + njs (compiled in builder-njs stage) ────────────────
+# AL2023 base repo ships nginx 1.24.x without nginx-module-njs; we therefore
+# build njs as a dynamic module from source against a matched nginx source tree
+# (mainline 1.30.x, current stable line as of 2026-06).
+#
+# nginx 1.30.2 — current mainline stable; ships CVE-2026-9256 fix
+# (heap buffer overflow in ngx_http_rewrite_module).
+# njs 0.9.9 — latest stable; ships CVE-2026-8711 fix (heap buffer overflow
+# in js_fetch_proxy directive).
+export NGINX_VERSION="1.30.2"
+export NJS_VERSION="0.9.9"
diff --git a/scripts/tensorflow/inference/dockerd_entrypoint.sh b/scripts/tensorflow/inference/dockerd_entrypoint.sh
new file mode 100755
index 000000000000..341ba60575a3
--- /dev/null
+++ b/scripts/tensorflow/inference/dockerd_entrypoint.sh
@@ -0,0 +1,12 @@
+#!/usr/bin/env bash
+
+# Execute telemetry script if it exists, suppress errors
+bash /usr/local/bin/bash_telemetry.sh >/dev/null 2>&1 || true
+
+TF_SERVING_PACKAGE=$(pip list | grep tensorflow-serving | cut -d ' ' -f 1)
+
+if [[ ${TF_SERVING_PACKAGE} == *"gpu"* ]]; then
+  bash /usr/local/bin/start_cuda_compat.sh
+fi
+
+eval '"$@"'
diff --git a/scripts/tensorflow/inference/sagemaker/__init__.py b/scripts/tensorflow/inference/sagemaker/__init__.py
new file mode 100644
index 000000000000..04fbf5d9a144
--- /dev/null
+++ b/scripts/tensorflow/inference/sagemaker/__init__.py
@@ -0,0 +1,12 @@
+# Copyright 2019-2020 Amazon.com, Inc. or its affiliates. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"). You
+# may not use this file except in compliance with the License. A copy of
+# the License is located at
+#
+#     http://aws.amazon.com/apache2.0/
+#
+# or in the "license" file accompanying this file. This file is
+# distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF
+# ANY KIND, either express or implied. See the License for the specific
+# language governing permissions and limitations under the License.
diff --git a/scripts/tensorflow/inference/sagemaker/multi_model_utils.py b/scripts/tensorflow/inference/sagemaker/multi_model_utils.py
new file mode 100644
index 000000000000..79bf7a88464c
--- /dev/null
+++ b/scripts/tensorflow/inference/sagemaker/multi_model_utils.py
@@ -0,0 +1,53 @@
+# Copyright 2019-2020 Amazon.com, Inc. or its affiliates. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"). You
+# may not use this file except in compliance with the License. A copy of
+# the License is located at
+#
+#     http://aws.amazon.com/apache2.0/
+#
+# or in the "license" file accompanying this file. This file is
+# distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF
+# ANY KIND, either express or implied. See the License for the specific
+# language governing permissions and limitations under the License.
+import fcntl
+import signal
+import time
+from contextlib import contextmanager
+
+MODEL_CONFIG_FILE = "/sagemaker/model-config.cfg"
+DEFAULT_LOCK_FILE = "/sagemaker/lock-file.lock"
+
+
+@contextmanager
+def lock(path=DEFAULT_LOCK_FILE):
+    f = open(path, "w", encoding="utf8")
+    fd = f.fileno()
+    fcntl.lockf(fd, fcntl.LOCK_EX)
+
+    try:
+        yield
+    finally:
+        time.sleep(1)
+        fcntl.lockf(fd, fcntl.LOCK_UN)
+
+
+@contextmanager
+def timeout(seconds=60):
+    def _raise_timeout_error(signum, frame):
+        raise Exception(408, "Timed out after {} seconds".format(seconds))
+
+    try:
+        signal.signal(signal.SIGALRM, _raise_timeout_error)
+        signal.alarm(seconds)
+        yield
+    finally:
+        signal.alarm(0)
+
+
+class MultiModelException(Exception):
+    def __init__(self, code, msg, pid):
+        Exception.__init__(self, code, msg)
+        self.pid = pid
+        self.code = code
+        self.msg = msg
diff --git a/scripts/tensorflow/inference/sagemaker/nginx.conf.template b/scripts/tensorflow/inference/sagemaker/nginx.conf.template
new file mode 100644
index 000000000000..a975069e57d2
--- /dev/null
+++ b/scripts/tensorflow/inference/sagemaker/nginx.conf.template
@@ -0,0 +1,66 @@
+load_module modules/ngx_http_js_module.so;
+
+worker_processes auto;
+daemon off;
+pid /tmp/nginx.pid;
+error_log  /dev/stderr %NGINX_LOG_LEVEL%;
+
+worker_rlimit_nofile 4096;
+
+events {
+  worker_connections 2048;
+}
+
+http {
+  include /etc/nginx/mime.types;
+  default_type application/json;
+  access_log /dev/stdout combined;
+  js_import tensorflowServing.js;
+
+  proxy_read_timeout %PROXY_READ_TIMEOUT%;
+
+  upstream tfs_upstream {
+    %TFS_UPSTREAM%;
+  }
+
+  upstream gunicorn_upstream {
+    server unix:/tmp/gunicorn.sock fail_timeout=1;
+  }
+
+  server {
+    listen %NGINX_HTTP_PORT% deferred;
+    client_max_body_size 0;
+    client_body_buffer_size 100m;
+    subrequest_output_buffer_size 100m;
+
+    set $tfs_version %TFS_VERSION%;
+    set $default_tfs_model %TFS_DEFAULT_MODEL_NAME%;
+
+    location /tfs {
+        rewrite ^/tfs/(.*) /$1  break;
+        proxy_redirect off;
+        proxy_pass_request_headers off;
+        proxy_set_header Content-Type 'application/json';
+        proxy_set_header Accept 'application/json';
+        proxy_pass http://tfs_upstream;
+    }
+
+    location /ping {
+        %FORWARD_PING_REQUESTS%;
+    }
+
+    location /invocations {
+        %FORWARD_INVOCATION_REQUESTS%;
+    }
+
+    location /models {
+        proxy_pass http://gunicorn_upstream/models;
+    }
+
+    location / {
+        return 404 '{"error": "Not Found"}';
+    }
+
+    keepalive_timeout 3;
+  }
+}
diff --git a/scripts/tensorflow/inference/sagemaker/python_service.py b/scripts/tensorflow/inference/sagemaker/python_service.py
new file mode 100644
index 000000000000..fb7fd708da11
--- /dev/null
+++ b/scripts/tensorflow/inference/sagemaker/python_service.py
@@ -0,0 +1,687 @@
+# Copyright 2019-2020 Amazon.com, Inc. or its affiliates. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"). You
+# may not use this file except in compliance with the License. A copy of
+# the License is located at
+#
+#     http://aws.amazon.com/apache2.0/
+#
+# or in the "license" file accompanying this file. This file is
+# distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF
+# ANY KIND, either express or implied. See the License for the specific
+# language governing permissions and limitations under the License.
+
+# monkey patching to ensure that all I/O operations are properly made asynchronous.
+import gevent.monkey
+
+gevent.monkey.patch_all()
+
+import argparse  # noqa: E402
+import copy  # noqa: E402
+import importlib.util  # noqa: E402
+import json  # noqa: E402
+import logging  # noqa: E402
+import os  # noqa: E402
+import pickle  # noqa: E402
+import random  # noqa: E402
+import shutil  # noqa: E402
+import signal  # noqa: E402
+import subprocess  # noqa: E402
+import sys  # noqa: E402
+
+import falcon  # noqa: E402
+import grpc  # noqa: E402
+import requests  # noqa: E402
+import tfs_utils  # noqa: E402
+from multi_model_utils import MultiModelException, lock  # noqa: E402
+
+SAGEMAKER_MULTI_MODEL_ENABLED = os.environ.get("SAGEMAKER_MULTI_MODEL", "false").lower() == "true"
+INFERENCE_SCRIPT_PATH = (
+    "/opt/ml/code/inference.py"
+    if SAGEMAKER_MULTI_MODEL_ENABLED
+    else "/opt/ml/model/code/inference.py"
+)
+
+SAGEMAKER_BATCHING_ENABLED = os.environ.get("SAGEMAKER_TFS_ENABLE_BATCHING", "false").lower()
+MODEL_CONFIG_FILE_PATH = "/sagemaker/model-config.cfg"
+TFS_GRPC_PORTS = os.environ.get("TFS_GRPC_PORTS")
+TFS_REST_PORTS = os.environ.get("TFS_REST_PORTS")
+SAGEMAKER_TFS_PORT_RANGE = os.environ.get("SAGEMAKER_SAFE_PORT_RANGE")
+TFS_INSTANCE_COUNT = int(os.environ.get("SAGEMAKER_TFS_INSTANCE_COUNT", "1"))
+
+logging.basicConfig(
+    format="%(process)d %(asctime)s %(levelname)-8s %(message)s", force=True, level=logging.INFO
+)
+log = logging.getLogger(__name__)
+
+CUSTOM_ATTRIBUTES_HEADER = "X-Amzn-SageMaker-Custom-Attributes"
+MME_TFS_INSTANCE_STATUS_FILE = "/sagemaker/tfs_instance.pickle"
+
+
+def default_handler(data, context):
+    """A default inference request handler that directly send post request to TFS rest port with
+    un-processed data and return un-processed response
+    :param data: input data
+    :param context: context instance that contains tfs_rest_uri
+    :return: inference response from TFS model server
+    """
+    data = data.read().decode("utf-8")
+    if not isinstance(data, str):
+        data = json.loads(data)
+    response = requests.post(context.rest_uri, data=data)
+    return response.content, context.accept_header
+
+
+class TfsInstanceStatus:
+    def __init__(self, rest_port: str, grpc_port: str, pid: int):
+        self.rest_port = rest_port
+        self.grpc_port = grpc_port
+        self.pid = pid
+
+    def __repr__(self):
+        return f"TFS Instance Status (rest_port : {self.rest_port}, grpc_port: {self.grpc_port}, pid: {self.pid}))"
+
+
+class PythonServiceResource:
+    def __init__(self):
+        if SAGEMAKER_MULTI_MODEL_ENABLED:
+            self._mme_tfs_instances_status: dict[str, [TfsInstanceStatus]] = {}
+            self._tfs_ports = self._parse_sagemaker_port_range_mme(SAGEMAKER_TFS_PORT_RANGE)
+            self._tfs_available_ports = self._parse_sagemaker_port_range_mme(
+                SAGEMAKER_TFS_PORT_RANGE
+            )
+            # If Multi-Model mode is enabled, dependencies/handlers will be imported
+            # during the _handle_load_model_post()
+            self.model_handlers = {}
+        else:
+            self._tfs_grpc_ports = self._parse_concat_ports(TFS_GRPC_PORTS)
+            self._tfs_rest_ports = self._parse_concat_ports(TFS_REST_PORTS)
+
+            self._channels = {}
+            for grpc_port in self._tfs_grpc_ports:
+                # Initialize grpc channel here so gunicorn worker could have mapping
+                # between each grpc port and channel
+                self._setup_channel(grpc_port)
+
+        self._default_handlers_enabled = False
+        if os.path.exists(INFERENCE_SCRIPT_PATH):
+            # Single-Model Mode & Multi-Model Mode both use one inference.py
+            self._handler, self._input_handler, self._output_handler = self._import_handlers()
+            self._handlers = self._make_handler(
+                self._handler, self._input_handler, self._output_handler
+            )
+        else:
+            self._handlers = default_handler
+            self._default_handlers_enabled = True
+
+        self._tfs_enable_batching = SAGEMAKER_BATCHING_ENABLED == "true"
+        self._tfs_default_model_name = os.environ.get("TFS_DEFAULT_MODEL_NAME", "None")
+        self._tfs_inter_op_parallelism = os.environ.get("SAGEMAKER_TFS_INTER_OP_PARALLELISM", 0)
+        self._tfs_intra_op_parallelism = os.environ.get("SAGEMAKER_TFS_INTRA_OP_PARALLELISM", 0)
+        self._tfs_instance_count = int(os.environ.get("SAGEMAKER_TFS_INSTANCE_COUNT", 1))
+        self._gunicorn_workers = int(os.environ.get("SAGEMAKER_GUNICORN_WORKERS", 1))
+        self._tfs_wait_time_seconds = int(
+            os.environ.get("SAGEMAKER_TFS_WAIT_TIME_SECONDS", 55 // self._tfs_instance_count)
+        )
+
+    def on_post(self, req, res, model_name=None):
+        if model_name or "invocations" in req.uri:
+            self._handle_invocation_post(req, res, model_name)
+        else:
+            data = json.loads(req.stream.read().decode("utf-8"))
+            self._handle_load_model_post(res, data)
+
+    def _parse_concat_ports(self, concat_ports):
+        return concat_ports.split(",")
+
+    def _pick_port(self, ports):
+        return random.choice(ports)
+
+    def _parse_sagemaker_port_range_mme(self, port_range):
+        lower, upper = port_range.split("-")
+        lower = int(lower)
+        upper = lower + int((int(upper) - lower) * 0.9)  # only utilizing 90% of the ports
+        rest_port = lower
+        grpc_port = (lower + upper) // 2
+        tfs_ports = {
+            "rest_port": [port for port in range(rest_port, grpc_port)],
+            "grpc_port": [port for port in range(grpc_port, upper)],
+        }
+        return tfs_ports
+
+    def _ports_available(self):
+        rest_ports = self._tfs_available_ports["rest_port"]
+        grpc_ports = self._tfs_available_ports["grpc_port"]
+        return len(rest_ports) > 0 and len(grpc_ports) > 0
+
+    def _update_ports_available(self):
+        self._tfs_available_ports = copy.deepcopy(self._tfs_ports)
+        for _, tf_status_list in self._mme_tfs_instances_status.items():
+            for tf_status in tf_status_list:
+                if tf_status.rest_port in self._tfs_available_ports["rest_port"]:
+                    self._tfs_available_ports["rest_port"].remove(tf_status.rest_port)
+                if tf_status.grpc_port in self._tfs_available_ports["grpc_port"]:
+                    self._tfs_available_ports["grpc_port"].remove(tf_status.grpc_port)
+        log.info(f"available ports : {self._tfs_available_ports}")
+
+    def _load_model(self, model_name, base_path, rest_port, grpc_port, model_index):
+        if self.validate_model_dir(base_path):
+            try:
+                self._import_custom_modules(model_name)
+                tfs_config = tfs_utils.create_tfs_config_individual_model(model_name, base_path)
+                tfs_config_file = "/sagemaker/tfs-config/{}/{}/model-config.cfg".format(
+                    model_name, model_index
+                )
+                log.info("tensorflow serving model config: \n%s\n", tfs_config)
+                os.makedirs(os.path.dirname(tfs_config_file))
+                with open(tfs_config_file, "w", encoding="utf8") as f:
+                    f.write(tfs_config)
+
+                batching_config_file = "/sagemaker/batching/{}/{}/batching-config.cfg".format(
+                    model_name, model_index
+                )
+                if self._tfs_enable_batching:
+                    tfs_utils.create_batching_config(batching_config_file)
+
+                cmd = tfs_utils.tfs_command(
+                    grpc_port,
+                    rest_port,
+                    tfs_config_file,
+                    self._tfs_enable_batching,
+                    batching_config_file,
+                    tfs_intra_op_parallelism=self._tfs_intra_op_parallelism,
+                    tfs_inter_op_parallelism=self._tfs_inter_op_parallelism,
+                )
+                log.info("MME starts tensorflow serving with command: {}".format(cmd))
+                p = subprocess.Popen(cmd.split())
+
+                tfs_utils.wait_for_model(rest_port, model_name, self._tfs_wait_time_seconds, p.pid)
+
+                log.info("started tensorflow serving (pid: %d)", p.pid)
+
+                return {
+                    "status": falcon.HTTP_200,
+                    "body": json.dumps(
+                        {
+                            "success": "Successfully loaded model {}, "
+                            "listening on rest port {} "
+                            "and grpc port {}.".format(model_name, rest_port, grpc_port)
+                        },
+                    ),
+                    "pid": p.pid,
+                }
+            except MultiModelException as multi_model_exception:
+                if multi_model_exception.code == 409:
+                    return {
+                        "status": falcon.HTTP_409,
+                        "body": multi_model_exception.msg,
+                        "pid": multi_model_exception.pid,
+                    }
+                elif multi_model_exception.code == 408:
+                    cpu_memory_usage = tfs_utils.get_cpu_memory_util()
+                    log.info(f"cpu memory usage {cpu_memory_usage}")
+                    if cpu_memory_usage > 70:
+                        return {
+                            "status": falcon.HTTP_507,
+                            "body": "Memory exhausted: not enough memory to start TFS instance",
+                            "pid": multi_model_exception.pid,
+                        }
+                    return {
+                        "status": falcon.HTTP_408,
+                        "body": multi_model_exception.msg,
+                        "pid": multi_model_exception.pid,
+                    }
+                else:
+                    return {
+                        "status": falcon.HTTP_500,
+                        "body": multi_model_exception.msg,
+                        "pid": multi_model_exception.pid,
+                    }
+            except FileExistsError as e:
+                return {
+                    "status": falcon.HTTP_409,
+                    "body": json.dumps(
+                        {"error": "Model {} is already loaded. {}".format(model_name, str(e))}
+                    ),
+                }
+            except OSError as os_error:
+                log.error(f"failed to load model with exception {os_error}")
+                if os_error.errno == 12:
+                    return {
+                        "status": falcon.HTTP_507,
+                        "body": "Memory exhausted: not enough memory to start TFS instance",
+                    }
+                else:
+                    return {
+                        "status": falcon.HTTP_500,
+                        "body": os_error.strerror,
+                    }
+        else:
+            return {
+                "status": falcon.HTTP_404,
+                "body": json.dumps(
+                    {
+                        "error": "Could not find valid base path {} for servable {}".format(
+                            base_path, model_name
+                        )
+                    }
+                ),
+            }
+
+    def _handle_load_model_post(self, res, data):  # noqa: C901
+        with lock():
+            model_name = data["model_name"]
+            base_path = data["url"]
+
+            # sync sync_local_mme_instance_status & update available ports
+            self._sync_local_mme_instance_status()
+            self._update_ports_available()
+            self._sync_model_handlers()
+
+            # model is already loaded
+            if model_name in self._mme_tfs_instances_status:
+                res.status = falcon.HTTP_409
+                res.body = json.dumps({"error": "Model {} is already loaded.".format(model_name)})
+                return
+
+            is_load_successful = True
+            response = {}
+            for i in range(self._tfs_instance_count):
+                # check if there are available ports
+                if not self._ports_available():
+                    is_load_successful = False
+                    response["status"] = falcon.HTTP_507
+                    response["body"] = json.dumps(
+                        {"error": "Memory exhausted: no available ports to load the model."}
+                    )
+                    break
+                tfs_rest_port = self._tfs_available_ports["rest_port"].pop()
+                tfs_grpc_port = self._tfs_available_ports["grpc_port"].pop()
+
+                response = self._load_model(model_name, base_path, tfs_rest_port, tfs_grpc_port, i)
+
+                if "pid" in response:
+                    self._mme_tfs_instances_status.setdefault(model_name, []).append(
+                        TfsInstanceStatus(tfs_rest_port, tfs_grpc_port, response["pid"])
+                    )
+
+                if response["status"] != falcon.HTTP_200:
+                    log.info(f"Failed to load model : {model_name}")
+                    is_load_successful = False
+                    break
+
+            if not is_load_successful:
+                log.info(f"Failed to load model : {model_name}, Starting to cleanup...")
+                self._delete_model(model_name)
+                self._remove_model_config(model_name)
+            else:
+                self._upload_mme_instance_status()
+
+            res.status = response["status"]
+            res.body = response["body"]
+
+    def _import_custom_modules(self, model_name):
+        inference_script_path = "/opt/ml/models/{}/model/code/inference.py".format(model_name)
+        python_lib_path = "/opt/ml/models/{}/model/code/lib".format(model_name)
+        if os.path.exists(python_lib_path):
+            log.info(
+                "Add Python code library for the model {} found at path {}.".format(
+                    model_name, python_lib_path
+                )
+            )
+            sys.path.append(python_lib_path)
+        else:
+            log.info(
+                "Python code library for the model {} not found at path {}.".format(
+                    model_name, python_lib_path
+                )
+            )
+        if os.path.exists(inference_script_path):
+            log.info(
+                "Importing handlers from model-specific inference script for the model {} found at path {}.".format(
+                    model_name, inference_script_path
+                )
+            )
+            handler, input_handler, output_handler = self._import_handlers(inference_script_path)
+            model_handlers = self._make_handler(handler, input_handler, output_handler)
+            self.model_handlers[model_name] = model_handlers
+        else:
+            log.info(
+                "Model-specific inference script for the model {} not found at path {}.".format(
+                    model_name, inference_script_path
+                )
+            )
+
+    def _handle_invocation_post(self, req, res, model_name=None):
+        if SAGEMAKER_MULTI_MODEL_ENABLED:
+            if model_name:
+                if self._gunicorn_workers > 1:
+                    if model_name not in self._mme_tfs_instances_status or not self._check_pid(
+                        self._mme_tfs_instances_status[model_name][0].pid
+                    ):
+                        with lock():
+                            self._sync_local_mme_instance_status()
+                        self._sync_model_handlers()
+
+                if model_name not in self._mme_tfs_instances_status:
+                    res.status = falcon.HTTP_404
+                    res.body = json.dumps(
+                        {"error": "Model {} is not loaded yet.".format(model_name)}
+                    )
+                    return
+                else:
+                    log.info("model name: {}".format(model_name))
+                    rest_ports = [
+                        status.rest_port for status in self._mme_tfs_instances_status[model_name]
+                    ]
+                    rest_port = self._pick_port(rest_ports)
+                    log.info("rest port: {}".format(str(rest_port)))
+                    grpc_ports = [
+                        status.grpc_port for status in self._mme_tfs_instances_status[model_name]
+                    ]
+                    grpc_port = grpc_ports[rest_ports.index(rest_port)]
+                    log.info("grpc port: {}".format(str(grpc_port)))
+                    data, context = tfs_utils.parse_request(
+                        req,
+                        rest_port,
+                        grpc_port,
+                        self._tfs_default_model_name,
+                        model_name=model_name,
+                    )
+            else:
+                res.status = falcon.HTTP_400
+                res.body = json.dumps({"error": "Invocation request does not contain model name."})
+                return
+        else:
+            # Randomly pick port used for routing incoming request.
+            grpc_port = self._pick_port(self._tfs_grpc_ports)
+            rest_port = self._pick_port(self._tfs_rest_ports)
+            data, context = tfs_utils.parse_request(
+                req,
+                rest_port,
+                grpc_port,
+                self._tfs_default_model_name,
+                channel=self._channels[grpc_port],
+            )
+
+        try:
+            res.status = falcon.HTTP_200
+            handlers = self._handlers
+            if SAGEMAKER_MULTI_MODEL_ENABLED and model_name in self.model_handlers:
+                log.info(
+                    "Model-specific inference script for the model {} exists, importing handlers.".format(
+                        model_name
+                    )
+                )
+                handlers = self.model_handlers[model_name]
+            elif not self._default_handlers_enabled:
+                log.info(
+                    "Universal inference script exists at path {}, importing handlers.".format(
+                        INFERENCE_SCRIPT_PATH
+                    )
+                )
+            else:
+                log.info(
+                    "Model-specific inference script and universal inference script both do not exist, using default handlers."
+                )
+            res.body, res.content_type = handlers(data, context)
+        except Exception as e:  # pylint: disable=broad-except
+            log.exception("exception handling request: {}".format(e))
+            res.status = falcon.HTTP_500
+            res.body = json.dumps({"error": str(e)}).encode("utf-8")  # pylint: disable=E1101
+
+    def _setup_channel(self, grpc_port):
+        if grpc_port not in self._channels:
+            log.info("Creating grpc channel for port: %s", grpc_port)
+            self._channels[grpc_port] = grpc.insecure_channel("localhost:{}".format(grpc_port))
+
+    def _import_handlers(self, inference_script=INFERENCE_SCRIPT_PATH):
+        spec = importlib.util.spec_from_file_location("inference", inference_script)
+        inference = importlib.util.module_from_spec(spec)
+        spec.loader.exec_module(inference)
+
+        _custom_handler, _custom_input_handler, _custom_output_handler = None, None, None
+        if hasattr(inference, "handler"):
+            _custom_handler = inference.handler
+        elif hasattr(inference, "input_handler") and hasattr(inference, "output_handler"):
+            _custom_input_handler = inference.input_handler
+            _custom_output_handler = inference.output_handler
+        else:
+            raise NotImplementedError("Handlers are not implemented correctly in user script.")
+
+        return _custom_handler, _custom_input_handler, _custom_output_handler
+
+    def _make_handler(self, custom_handler, custom_input_handler, custom_output_handler):
+        if custom_handler:
+            return custom_handler
+
+        def handler(data, context):
+            processed_input = custom_input_handler(data, context)
+            response = requests.post(context.rest_uri, data=processed_input)
+            return custom_output_handler(response, context)
+
+        return handler
+
+    def on_get(self, req, res, model_name=None):  # pylint: disable=W0613
+        with lock():
+            self._sync_local_mme_instance_status()
+            if model_name is None:
+                models_info = {}
+                uri = "http://localhost:{}/v1/models/{}"
+                for model, tfs_instance_status in self._mme_tfs_instances_status.items():
+                    try:
+                        info = json.loads(
+                            requests.get(
+                                uri.format(tfs_instance_status[0].rest_port, model)
+                            ).content
+                        )
+                        models_info[model] = info
+                    except ValueError as e:
+                        log.exception("exception handling request: {}".format(e))
+                        res.status = falcon.HTTP_500
+                        res.body = json.dumps({"error": str(e)}).encode("utf-8")
+                res.status = falcon.HTTP_200
+                res.body = json.dumps(models_info)
+            else:
+                if model_name not in self._mme_tfs_instances_status:
+                    res.status = falcon.HTTP_404
+                    res.body = json.dumps(
+                        {"error": "Model {} is loaded yet.".format(model_name)}
+                    ).encode("utf-8")
+                else:
+                    port = self._mme_tfs_instances_status[model_name].rest_port
+                    uri = "http://localhost:{}/v1/models/{}".format(port, model_name)
+                    try:
+                        info = requests.get(uri)
+                        res.status = falcon.HTTP_200
+                        res.body = json.dumps({"model": info}).encode("utf-8")
+                    except ValueError as e:
+                        log.exception("exception handling GET models request.")
+                        res.status = falcon.HTTP_500
+                        res.body = json.dumps({"error": str(e)}).encode("utf-8")
+
+    def on_delete(self, req, res, model_name):  # pylint: disable=W0613
+        with lock():
+            self._sync_local_mme_instance_status()
+            if model_name not in self._mme_tfs_instances_status:
+                res.status = falcon.HTTP_404
+                res.body = json.dumps({"error": "Model {} is not loaded yet".format(model_name)})
+            else:
+                try:
+                    self._delete_model(model_name)
+                    self._remove_model_config(model_name)
+                    del self._mme_tfs_instances_status[model_name]
+                    self._upload_mme_instance_status()
+                    res.status = falcon.HTTP_200
+                    res.body = json.dumps(
+                        {"success": "Successfully unloaded model {}.".format(model_name)}
+                    )
+                except OSError as error:
+                    res.status = falcon.HTTP_500
+                    res.body = json.dumps({"error": str(error)}).encode("utf-8")
+
+    def _delete_model(self, model_name):
+        if model_name not in self._mme_tfs_instances_status:
+            return
+        for tfs_status in self._mme_tfs_instances_status[model_name]:
+            os.kill(tfs_status.pid, signal.SIGKILL)
+
+    def _remove_model_config(self, model_name):
+        shutil.rmtree("/sagemaker/tfs-config/{}".format(model_name), ignore_errors=True)
+        shutil.rmtree("/sagemaker/batching/{}".format(model_name), ignore_errors=True)
+
+    def validate_model_dir(self, model_path):
+        # model base path doesn't exits
+        if not os.path.exists(model_path):
+            return False
+        versions = []
+        for _, dirs, _ in os.walk(model_path):
+            for dirname in dirs:
+                if dirname.isdigit():
+                    versions.append(dirname)
+        return self.validate_model_versions(versions)
+
+    def validate_model_versions(self, versions):
+        if not versions:
+            return False
+        for v in versions:
+            if v.isdigit():
+                # TensorFlow model server will succeed with any versions found
+                # even if there are directories that's not a valid model version,
+                # the loading will succeed.
+                return True
+        return False
+
+    def _upload_mme_instance_status(self):
+        log.info(
+            "uploaded mme instance status file with content: {}".format(
+                self._mme_tfs_instances_status
+            )
+        )
+        with open(MME_TFS_INSTANCE_STATUS_FILE, "wb") as handle:
+            pickle.dump(self._mme_tfs_instances_status, handle, protocol=pickle.HIGHEST_PROTOCOL)
+
+    def _sync_local_mme_instance_status(self):
+        if not os.path.exists(MME_TFS_INSTANCE_STATUS_FILE):
+            log.info("mme instance status file does not found.")
+            return
+        with open(MME_TFS_INSTANCE_STATUS_FILE, "rb") as handle:
+            self._mme_tfs_instances_status = pickle.load(handle)
+        log.info(
+            "updated local mme instance status with content: {}".format(
+                self._mme_tfs_instances_status
+            )
+        )
+
+    def _sync_model_handlers(self):
+        for model_name, _ in self._mme_tfs_instances_status.items():
+            if model_name not in self.model_handlers:
+                self._import_custom_modules(model_name)
+
+    def _check_pid(self, pid):
+        """Check For the existence of a unix pid."""
+        try:
+            os.kill(pid, 0)
+        except OSError:
+            return False
+        else:
+            return True
+
+
+class PingResource:
+    def on_get(self, req, res):  # pylint: disable=W0613
+        res.status = falcon.HTTP_200
+
+
+class ServiceResources:
+    def __init__(self):
+        self._enable_model_manager = SAGEMAKER_MULTI_MODEL_ENABLED
+        self._python_service_resource = PythonServiceResource()
+        self._ping_resource = PingResource()
+
+    def add_routes(self, application):
+        application.add_route("/ping", self._ping_resource)
+        application.add_route("/invocations", self._python_service_resource)
+
+        if self._enable_model_manager:
+            application.add_route("/models", self._python_service_resource)
+            application.add_route("/models/{model_name}", self._python_service_resource)
+            application.add_route("/models/{model_name}/invoke", self._python_service_resource)
+
+
+app = falcon.API()
+resources = ServiceResources()
+resources.add_routes(app)
+
+if __name__ == "__main__":
+    # Define the command-line arguments
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "-b", "--bind", type=str, required=True, help="Specify a server socket to bind."
+    )
+    parser.add_argument(
+        "-k",
+        "--worker-class",
+        type=str,
+        required=True,
+        choices=["sync", "eventlet", "gevent", "tornado", "gthread", "sync"],
+        help="The type of worker process to run",
+    )
+    parser.add_argument("-c", "--chdir", type=str, required=True, help="Change root dir")
+    parser.add_argument(
+        "-w",
+        "--workers",
+        type=int,
+        required=True,
+        help="The number of worker processes. This number should generally be between 2-4 workers per core in the server.",
+    )
+    parser.add_argument("-t", "--threads", type=int, required=True, help="The number of threads")
+    parser.add_argument("-l", "--log-level", type=str, required=True)
+    parser.add_argument("-o", "--timeout", type=int, required=True, help="Gunicorn timeout")
+
+    # Parse the command-line arguments
+    args = parser.parse_args()
+
+    # Create gunicorn options
+    options = {
+        "bind": args.bind,
+        "worker_class": args.worker_class,
+        "chdir": args.chdir,
+        "workers": args.workers,
+        "threads": args.threads,
+        "loglevel": args.log_level,
+        "timeout": args.timeout,
+        "raw_env": [
+            f"TFS_GRPC_PORTS={TFS_GRPC_PORTS}",
+            f"TFS_REST_PORTS={TFS_REST_PORTS}",
+            f"SAGEMAKER_MULTI_MODEL={os.environ.get('SAGEMAKER_MULTI_MODEL')}",
+            f"SAGEMAKER_SAFE_PORT_RANGE={SAGEMAKER_TFS_PORT_RANGE}",
+            f"SAGEMAKER_TFS_WAIT_TIME_SECONDS={os.environ.get('SAGEMAKER_TFS_WAIT_TIME_SECONDS')}",
+            f"SAGEMAKER_TFS_INTER_OP_PARALLELISM={os.environ.get('SAGEMAKER_TFS_INTER_OP_PARALLELISM', 0)}",
+            f"SAGEMAKER_TFS_INTRA_OP_PARALLELISM={os.environ.get('SAGEMAKER_TFS_INTRA_OP_PARALLELISM', 0)}",
+            f"SAGEMAKER_TFS_INSTANCE_COUNT={os.environ.get('SAGEMAKER_TFS_INSTANCE_COUNT', '1')}",
+            f"SAGEMAKER_GUNICORN_WORKERS={os.environ.get('SAGEMAKER_GUNICORN_WORKERS', '1')}",
+        ],
+    }
+
+    from gunicorn.app.base import BaseApplication
+
+    class StandaloneApplication(BaseApplication):
+        def __init__(self, app, options=None):
+            self.options = options or {}
+            self.application = app
+            super().__init__()
+
+        def load_config(self):
+            config = {
+                key: value
+                for key, value in self.options.items()
+                if key in self.cfg.settings and value is not None
+            }
+            for key, value in config.items():
+                self.cfg.set(key.lower(), value)
+
+        def load(self):
+            return self.application
+
+    StandaloneApplication(app, options).run()
diff --git a/scripts/tensorflow/inference/sagemaker/serve b/scripts/tensorflow/inference/sagemaker/serve
new file mode 100755
index 000000000000..9fac6a93ab41
--- /dev/null
+++ b/scripts/tensorflow/inference/sagemaker/serve
@@ -0,0 +1,3 @@
+#!/bin/bash
+
+python3 /sagemaker/serve.py
diff --git a/scripts/tensorflow/inference/sagemaker/serve.py b/scripts/tensorflow/inference/sagemaker/serve.py
new file mode 100644
index 000000000000..577095c8a8b4
--- /dev/null
+++ b/scripts/tensorflow/inference/sagemaker/serve.py
@@ -0,0 +1,522 @@
+# Copyright 2018-2020 Amazon.com, Inc. or its affiliates. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"). You
+# may not use this file except in compliance with the License. A copy of
+# the License is located at
+#
+#     http://aws.amazon.com/apache2.0/
+#
+# or in the "license" file accompanying this file. This file is
+# distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF
+# ANY KIND, either express or implied. See the License for the specific
+# language governing permissions and limitations under the License.
+import logging
+import os
+import re
+import signal
+import subprocess
+from contextlib import contextmanager
+
+import boto3
+import tfs_utils
+
+logging.basicConfig(
+    format="%(process)d %(asctime)s %(levelname)-8s %(message)s", force=True, level=logging.INFO
+)
+log = logging.getLogger(__name__)
+
+JS_PING = "js_content tensorflowServing.ping"
+JS_INVOCATIONS = "js_content tensorflowServing.invocations"
+GUNICORN_PING = "proxy_pass http://gunicorn_upstream/ping"
+GUNICORN_INVOCATIONS = "proxy_pass http://gunicorn_upstream/invocations"
+CODE_DIR = (
+    "/opt/ml/code"
+    if os.environ.get("SAGEMAKER_MULTI_MODEL", "False").lower() == "true"
+    else "/opt/ml/model/code"
+)
+PYTHON_LIB_PATH = os.path.join(CODE_DIR, "lib")
+REQUIREMENTS_PATH = os.path.join(CODE_DIR, "requirements.txt")
+INFERENCE_PATH = os.path.join(CODE_DIR, "inference.py")
+
+
+class ServiceManager(object):
+    def __init__(self):
+        self._state = "initializing"
+        self._nginx = None
+        self._tfs = []
+        self._gunicorn = None
+        self._gunicorn_command = None
+        self._gunicorn_env = None
+        self._enable_python_service = False
+        self._tfs_version = os.environ.get("SAGEMAKER_TFS_VERSION", "1.13")
+        self._nginx_http_port = os.environ.get("SAGEMAKER_BIND_TO_PORT", "8080")
+        self._nginx_loglevel = os.environ.get("SAGEMAKER_TFS_NGINX_LOGLEVEL", "error")
+        self._tfs_default_model_name = os.environ.get("SAGEMAKER_TFS_DEFAULT_MODEL_NAME", "None")
+        self._sagemaker_port_range = os.environ.get("SAGEMAKER_SAFE_PORT_RANGE", None)
+        self._gunicorn_workers = os.environ.get("SAGEMAKER_GUNICORN_WORKERS", 1)
+        self._gunicorn_threads = os.environ.get("SAGEMAKER_GUNICORN_THREADS", 1)
+        self._gunicorn_loglevel = os.environ.get("SAGEMAKER_GUNICORN_LOGLEVEL", "info")
+        self._tfs_config_path = "/sagemaker/model-config.cfg"
+        self._tfs_batching_config_path = "/sagemaker/batching-config.cfg"
+
+        _enable_batching = os.environ.get("SAGEMAKER_TFS_ENABLE_BATCHING", "false").lower()
+        _enable_multi_model_endpoint = os.environ.get("SAGEMAKER_MULTI_MODEL", "false").lower()
+        # Use this to specify memory that is needed to initialize CUDA/cuDNN and other GPU libraries
+        self._tfs_gpu_margin = float(os.environ.get("SAGEMAKER_TFS_FRACTIONAL_GPU_MEM_MARGIN", 0.2))
+        self._tfs_instance_count = int(os.environ.get("SAGEMAKER_TFS_INSTANCE_COUNT", 1))
+        self._tfs_wait_time_seconds = int(
+            os.environ.get("SAGEMAKER_TFS_WAIT_TIME_SECONDS", 55 // self._tfs_instance_count)
+        )
+        self._tfs_inter_op_parallelism = os.environ.get("SAGEMAKER_TFS_INTER_OP_PARALLELISM", 0)
+        self._tfs_intra_op_parallelism = os.environ.get("SAGEMAKER_TFS_INTRA_OP_PARALLELISM", 0)
+        self._gunicorn_worker_class = os.environ.get("SAGEMAKER_GUNICORN_WORKER_CLASS", "gevent")
+        self._gunicorn_timeout_seconds = int(
+            os.environ.get("SAGEMAKER_GUNICORN_TIMEOUT_SECONDS", 30)
+        )
+        self._nginx_proxy_read_timeout_seconds = int(
+            os.environ.get("SAGEMAKER_NGINX_PROXY_READ_TIMEOUT_SECONDS", 60)
+        )
+
+        # Nginx proxy read timeout should not be less than the GUnicorn timeout. If it is, this
+        # can result in upstream time out errors.
+        if self._gunicorn_timeout_seconds > self._nginx_proxy_read_timeout_seconds:
+            log.info(
+                "GUnicorn timeout was higher than Nginx proxy read timeout."
+                " Setting Nginx proxy read timeout from {} seconds to {} seconds"
+                " to match GUnicorn timeout.".format(
+                    self._nginx_proxy_read_timeout_seconds, self._gunicorn_timeout_seconds
+                )
+            )
+            self._nginx_proxy_read_timeout_seconds = self._gunicorn_timeout_seconds
+
+        if os.environ.get("OMP_NUM_THREADS") is None:
+            os.environ["OMP_NUM_THREADS"] = "1"
+
+        if _enable_multi_model_endpoint not in ["true", "false"]:
+            raise ValueError("SAGEMAKER_MULTI_MODEL must be 'true' or 'false'")
+        self._tfs_enable_multi_model_endpoint = _enable_multi_model_endpoint == "true"
+
+        self._need_python_service()
+        log.info("PYTHON SERVICE: {}".format(str(self._enable_python_service)))
+
+        if _enable_batching not in ["true", "false"]:
+            raise ValueError("SAGEMAKER_TFS_ENABLE_BATCHING must be 'true' or 'false'")
+        self._tfs_enable_batching = _enable_batching == "true"
+
+        if _enable_multi_model_endpoint not in ["true", "false"]:
+            raise ValueError("SAGEMAKER_MULTI_MODEL must be 'true' or 'false'")
+        self._tfs_enable_multi_model_endpoint = _enable_multi_model_endpoint == "true"
+
+        self._use_gunicorn = self._enable_python_service or self._tfs_enable_multi_model_endpoint
+
+        if self._sagemaker_port_range is not None:
+            parts = self._sagemaker_port_range.split("-")
+            low = int(parts[0])
+            hi = int(parts[1])
+            self._tfs_grpc_ports = []
+            self._tfs_rest_ports = []
+            if low + 2 * self._tfs_instance_count > hi:
+                raise ValueError(
+                    "not enough ports available in SAGEMAKER_SAFE_PORT_RANGE ({})".format(
+                        self._sagemaker_port_range
+                    )
+                )
+            # select non-overlapping grpc and rest ports based on tfs instance count
+            for i in range(self._tfs_instance_count):
+                self._tfs_grpc_ports.append(str(low + 2 * i))
+                self._tfs_rest_ports.append(str(low + 2 * i + 1))
+            # concat selected ports respectively in order to pass them to python service
+            self._tfs_grpc_concat_ports = self._concat_ports(self._tfs_grpc_ports)
+            self._tfs_rest_concat_ports = self._concat_ports(self._tfs_rest_ports)
+        else:
+            # just use the standard default ports
+            self._tfs_grpc_ports = ["9000"]
+            self._tfs_rest_ports = ["8501"]
+            # provide single concat port here for default case
+            self._tfs_grpc_concat_ports = "9000"
+            self._tfs_rest_concat_ports = "8501"
+
+        # set environment variable for python service
+        os.environ["TFS_GRPC_PORTS"] = self._tfs_grpc_concat_ports
+        os.environ["TFS_REST_PORTS"] = self._tfs_rest_concat_ports
+
+    def _need_python_service(self):
+        if (
+            os.path.exists(INFERENCE_PATH)
+            or os.path.exists(REQUIREMENTS_PATH)
+            or os.path.exists(PYTHON_LIB_PATH)
+        ):
+            self._enable_python_service = True
+        if os.environ.get("SAGEMAKER_MULTI_MODEL_UNIVERSAL_BUCKET") and os.environ.get(
+            "SAGEMAKER_MULTI_MODEL_UNIVERSAL_PREFIX"
+        ):
+            self._enable_python_service = True
+
+    def _concat_ports(self, ports):
+        str_ports = [str(port) for port in ports]
+        concat_str_ports = ",".join(str_ports)
+        return concat_str_ports
+
+    def _create_tfs_config(self):
+        models = tfs_utils.find_models()
+
+        if not models:
+            raise ValueError("no SavedModel bundles found!")
+
+        if self._tfs_default_model_name == "None":
+            default_model = os.path.basename(models[0])
+            if default_model:
+                self._tfs_default_model_name = default_model
+                log.info("using default model name: {}".format(self._tfs_default_model_name))
+            else:
+                log.info("no default model detected")
+
+        # config (may) include duplicate 'config' keys, so we can't just dump a dict
+        config = "model_config_list: {\n"
+        for m in models:
+            config += "  config: {\n"
+            config += "    name: '{}'\n".format(os.path.basename(m))
+            config += "    base_path: '{}'\n".format(m)
+            config += "    model_platform: 'tensorflow'\n"
+
+            config += "    model_version_policy: {\n"
+            config += "      specific: {\n"
+            for version in tfs_utils.find_model_versions(m):
+                config += "        versions: {}\n".format(version)
+            config += "      }\n"
+            config += "    }\n"
+
+            config += "  }\n"
+        config += "}\n"
+
+        log.info("tensorflow serving model config: \n%s\n", config)
+
+        with open(self._tfs_config_path, "w", encoding="utf8") as f:
+            f.write(config)
+
+    def _setup_gunicorn(self):
+        python_path_content = []
+        python_path_option = ""
+
+        bucket = os.environ.get("SAGEMAKER_MULTI_MODEL_UNIVERSAL_BUCKET", None)
+        prefix = os.environ.get("SAGEMAKER_MULTI_MODEL_UNIVERSAL_PREFIX", None)
+
+        if not os.path.exists(CODE_DIR) and bucket and prefix:
+            self._download_scripts(bucket, prefix)
+
+        if self._enable_python_service:
+            lib_path_exists = os.path.exists(PYTHON_LIB_PATH)
+            requirements_exists = os.path.exists(REQUIREMENTS_PATH)
+            python_path_content = ["/opt/ml/model/code"]
+            python_path_option = "--pythonpath "  # noqa: F841
+
+            if lib_path_exists:
+                python_path_content.append(PYTHON_LIB_PATH)
+
+            if requirements_exists:
+                if lib_path_exists:
+                    log.warning(
+                        "loading modules in '{}', ignoring requirements.txt".format(PYTHON_LIB_PATH)
+                    )
+                else:
+                    log.info("installing packages from requirements.txt...")
+                    pip_install_cmd = "pip3 install -r {}".format(REQUIREMENTS_PATH)
+                    try:
+                        subprocess.check_call(pip_install_cmd.split())
+                    except subprocess.CalledProcessError:
+                        log.error("failed to install required packages, exiting.")
+                        self._stop()
+                        raise ChildProcessError("failed to install required packages.")
+
+        gunicorn_command = (
+            "python3 /sagemaker/python_service.py -b unix:/tmp/gunicorn.sock -k {} --chdir /sagemaker "
+            "--workers {} --threads {} --log-level {} --timeout {} "
+        ).format(
+            self._gunicorn_worker_class,
+            self._gunicorn_workers,
+            self._gunicorn_threads,
+            self._gunicorn_loglevel,
+            self._gunicorn_timeout_seconds,
+        )
+
+        log.info("gunicorn command: {}".format(gunicorn_command))
+        self._gunicorn_command = gunicorn_command
+        gunicorn_env = {
+            "TFS_GRPC_PORTS": self._tfs_grpc_concat_ports,
+            "TFS_REST_PORTS": self._tfs_rest_concat_ports,
+            "SAGEMAKER_MULTI_MODEL": str(self._tfs_enable_multi_model_endpoint),
+            "SAGEMAKER_TFS_WAIT_TIME_SECONDS": str(self._tfs_wait_time_seconds),
+            "SAGEMAKER_TFS_INTER_OP_PARALLELISM": str(self._tfs_inter_op_parallelism),
+            "SAGEMAKER_TFS_INTRA_OP_PARALLELISM": str(self._tfs_intra_op_parallelism),
+            "SAGEMAKER_TFS_INSTANCE_COUNT": str(self._tfs_instance_count),
+            "PYTHONPATH": ":".join(python_path_content),
+            "SAGEMAKER_GUNICORN_WORKERS": str(self._gunicorn_workers),
+        }
+        if self._sagemaker_port_range is not None:
+            gunicorn_env["SAGEMAKER_SAFE_PORT_RANGE"] = self._sagemaker_port_range
+        log.info(f"gunicorn env: {gunicorn_env}")
+        self._gunicorn_env = gunicorn_env
+
+    def _download_scripts(self, bucket, prefix):
+        log.info("checking boto session region ...")
+        boto_session = boto3.session.Session()
+        boto_region = boto_session.region_name
+        if boto_region in ("us-iso-east-1", "us-gov-west-1"):
+            raise ValueError("Universal scripts is not supported in us-iso-east-1 or us-gov-west-1")
+
+        log.info("downloading universal scripts ...")
+        client = boto3.client("s3")
+        resource = boto3.resource("s3")
+        # download files
+        paginator = client.get_paginator("list_objects")
+        for result in paginator.paginate(Bucket=bucket, Delimiter="/", Prefix=prefix):
+            for file in result.get("Contents", []):
+                destination = os.path.join(CODE_DIR, file.get("Key").split("/")[-1])
+                if not os.path.exists(os.path.dirname(destination)):
+                    os.makedirs(os.path.dirname(destination))
+                resource.meta.client.download_file(bucket, file.get("Key"), destination)
+
+    def _create_nginx_tfs_upstream(self):
+        indentation = "    "
+        tfs_upstream = ""
+        for port in self._tfs_rest_ports:
+            tfs_upstream += "{}server localhost:{};\n".format(indentation, port)
+        tfs_upstream = tfs_upstream[len(indentation) : -2]
+
+        return tfs_upstream
+
+    def _create_nginx_config(self):
+        template = self._read_nginx_template()
+        pattern = re.compile(r"%(\w+)%")
+
+        template_values = {
+            "TFS_VERSION": self._tfs_version,
+            "TFS_UPSTREAM": self._create_nginx_tfs_upstream(),
+            "TFS_DEFAULT_MODEL_NAME": self._tfs_default_model_name,
+            "NGINX_HTTP_PORT": self._nginx_http_port,
+            "NGINX_LOG_LEVEL": self._nginx_loglevel,
+            "FORWARD_PING_REQUESTS": GUNICORN_PING if self._use_gunicorn else JS_PING,
+            "FORWARD_INVOCATION_REQUESTS": (
+                GUNICORN_INVOCATIONS if self._use_gunicorn else JS_INVOCATIONS
+            ),
+            "PROXY_READ_TIMEOUT": str(self._nginx_proxy_read_timeout_seconds),
+        }
+
+        config = pattern.sub(lambda x: template_values[x.group(1)], template)
+        log.info("nginx config: \n%s\n", config)
+
+        with open("/sagemaker/nginx.conf", "w", encoding="utf8") as f:
+            f.write(config)
+
+    def _read_nginx_template(self):
+        with open("/sagemaker/nginx.conf.template", "r", encoding="utf8") as f:
+            template = f.read()
+            if not template:
+                raise ValueError("failed to read nginx.conf.template")
+
+            return template
+
+    def _enable_per_process_gpu_memory_fraction(self):
+        nvidia_smi_exist = os.path.exists("/usr/bin/nvidia-smi")
+        if self._tfs_instance_count > 1 and nvidia_smi_exist:
+            return True
+
+        return False
+
+    def _get_number_of_gpu_on_host(self):
+        nvidia_smi_exist = os.path.exists("/usr/bin/nvidia-smi")
+        if nvidia_smi_exist:
+            return len(
+                subprocess.check_output(["nvidia-smi", "-L"]).decode("utf-8").strip().split("\n")
+            )
+        return 0
+
+    def _calculate_per_process_gpu_memory_fraction(self):
+        return round((1 - self._tfs_gpu_margin) / float(self._tfs_instance_count), 4)
+
+    def _start_tfs(self):
+        self._log_version("tensorflow_model_server --version", "tensorflow version info:")
+
+        for i in range(self._tfs_instance_count):
+            p = self._start_single_tfs(i)
+            self._tfs.append(p)
+
+    def _start_gunicorn(self):
+        self._log_version("gunicorn --version", "gunicorn version info:")
+        env = os.environ.copy()
+        env["TFS_DEFAULT_MODEL_NAME"] = self._tfs_default_model_name
+        env.update(self._gunicorn_env)
+        p = subprocess.Popen(self._gunicorn_command.split(), env=env)
+        log.info("started gunicorn (pid: %d)", p.pid)
+        self._gunicorn = p
+
+    def _start_nginx(self):
+        self._log_version("/usr/sbin/nginx -V", "nginx version info:")
+        p = subprocess.Popen("/usr/sbin/nginx -c /sagemaker/nginx.conf".split())
+        log.info("started nginx (pid: %d)", p.pid)
+        self._nginx = p
+
+    def _log_version(self, command, message):
+        try:
+            output = (
+                subprocess.check_output(command.split(), stderr=subprocess.STDOUT)
+                .decode("utf-8", "backslashreplace")
+                .strip()
+            )
+            log.info("{}\n{}".format(message, output))
+        except subprocess.CalledProcessError:
+            log.warning("failed to run command: %s", command)
+
+    def _stop(self, *args):  # pylint: disable=W0613
+        self._state = "stopping"
+        log.info("stopping services")
+        try:
+            os.kill(self._nginx.pid, signal.SIGQUIT)
+        except OSError:
+            pass
+        try:
+            if self._gunicorn:
+                os.kill(self._gunicorn.pid, signal.SIGTERM)
+        except OSError:
+            pass
+        try:
+            for tfs in self._tfs:
+                os.kill(tfs.pid, signal.SIGTERM)
+        except OSError:
+            pass
+
+        self._state = "stopped"
+        log.info("stopped")
+
+    def _wait_for_gunicorn(self):
+        while True:
+            if os.path.exists("/tmp/gunicorn.sock"):
+                log.info("gunicorn server is ready!")
+                return
+
+    def _wait_for_tfs(self):
+        for i in range(self._tfs_instance_count):
+            tfs_utils.wait_for_model(
+                self._tfs_rest_ports[i], self._tfs_default_model_name, self._tfs_wait_time_seconds
+            )
+
+    @contextmanager
+    def _timeout(self, seconds):
+        def _raise_timeout_error(signum, frame):
+            raise TimeoutError("time out after {} seconds".format(seconds))
+
+        try:
+            signal.signal(signal.SIGALRM, _raise_timeout_error)
+            signal.alarm(seconds)
+            yield
+        finally:
+            signal.alarm(0)
+
+    def _is_tfs_process(self, pid):
+        for p in self._tfs:
+            if p.pid == pid:
+                return True
+        return False
+
+    def _find_tfs_process(self, pid):
+        for index, p in enumerate(self._tfs):
+            if p.pid == pid:
+                return index
+        return None
+
+    def _restart_single_tfs(self, pid):
+        instance_id = self._find_tfs_process(pid)
+        if instance_id is None:
+            raise ValueError("Cannot find tfs with pid: {};".format(pid))
+        p = self._start_single_tfs(instance_id)
+        self._tfs[instance_id] = p
+
+    def _start_single_tfs(self, instance_id):
+        cmd = tfs_utils.tfs_command(
+            self._tfs_grpc_ports[instance_id],
+            self._tfs_rest_ports[instance_id],
+            self._tfs_config_path,
+            self._tfs_enable_batching,
+            self._tfs_batching_config_path,
+            tfs_intra_op_parallelism=self._tfs_intra_op_parallelism,
+            tfs_inter_op_parallelism=self._tfs_inter_op_parallelism,
+            tfs_enable_gpu_memory_fraction=self._enable_per_process_gpu_memory_fraction(),
+            tfs_gpu_memory_fraction=self._calculate_per_process_gpu_memory_fraction(),
+        )
+        log.info("tensorflow serving command: {}".format(cmd))
+
+        num_gpus = self._get_number_of_gpu_on_host()
+        if num_gpus > 1:
+            # utilizing multi-gpu
+            worker_env = os.environ.copy()
+            worker_env["CUDA_VISIBLE_DEVICES"] = str(instance_id % num_gpus)
+            p = subprocess.Popen(cmd.split(), env=worker_env)
+            log.info(
+                "started tensorflow serving (pid: {}) on GPU: {}".format(
+                    p.pid, instance_id % num_gpus
+                )
+            )
+        else:
+            # cpu and single gpu
+            p = subprocess.Popen(cmd.split())
+            log.info("started tensorflow serving (pid: {})".format(p.pid))
+
+        return p
+
+    def _monitor(self):
+        while True:
+            pid, status = os.wait()
+
+            if self._state != "started":
+                break
+
+            if pid == self._nginx.pid:
+                log.warning("unexpected nginx exit (status: {}). restarting.".format(status))
+                self._start_nginx()
+
+            elif self._is_tfs_process(pid):
+                log.warning(
+                    "unexpected tensorflow serving exit (status: {}). restarting.".format(status)
+                )
+                try:
+                    self._restart_single_tfs(pid)
+                except (ValueError, OSError) as error:
+                    log.error("Failed to restart tensorflow serving. {}".format(error))
+
+            elif self._gunicorn and pid == self._gunicorn.pid:
+                log.warning("unexpected gunicorn exit (status: {}). restarting.".format(status))
+                self._start_gunicorn()
+
+    def start(self):
+        log.info("starting services")
+        self._state = "starting"
+        signal.signal(signal.SIGTERM, self._stop)
+
+        if self._tfs_enable_batching:
+            log.info("batching is enabled")
+            tfs_utils.create_batching_config(self._tfs_batching_config_path)
+
+        if self._tfs_enable_multi_model_endpoint:
+            log.info("multi-model endpoint is enabled, TFS model servers will be started later")
+        else:
+            self._create_tfs_config()
+            self._start_tfs()
+            self._wait_for_tfs()
+
+        self._create_nginx_config()
+
+        if self._use_gunicorn:
+            self._setup_gunicorn()
+            self._start_gunicorn()
+            # make sure gunicorn is up
+            with self._timeout(seconds=self._gunicorn_timeout_seconds):
+                self._wait_for_gunicorn()
+
+        self._start_nginx()
+        self._state = "started"
+        self._monitor()
+        self._stop()
+
+
+if __name__ == "__main__":
+    ServiceManager().start()
diff --git a/scripts/tensorflow/inference/sagemaker/tensorflowServing.js b/scripts/tensorflow/inference/sagemaker/tensorflowServing.js
new file mode 100644
index 000000000000..65732ac8b386
--- /dev/null
+++ b/scripts/tensorflow/inference/sagemaker/tensorflowServing.js
@@ -0,0 +1,239 @@
+var tfs_base_uri = '/tfs/v1/models/'
+var custom_attributes_header = 'X-Amzn-SageMaker-Custom-Attributes'
+
+function invocations(r) {
+    var ct = r.headersIn['Content-Type']
+
+    if ('application/json' == ct || 'application/jsonlines' == ct || 'application/jsons' == ct) {
+        json_request(r)
+    } else if ('text/csv' == ct) {
+        csv_request(r)
+    } else {
+        return_error(r, 415, 'Unsupported Media Type: ' + (ct || 'Unknown'))
+    }
+}
+
+function ping(r) {
+    var uri = make_tfs_uri(r, false)
+
+    function callback (reply) {
+        if (reply.status == 200 && reply.responseText.includes('"AVAILABLE"')) {
+            r.return(200)
+        } else {
+            r.error('failed ping' + reply.responseText)
+            r.return(502)
+        }
+    }
+
+    r.subrequest(uri, callback)
+}
+
+function ping_without_model(r) {
+    // hack for TF 1.11 and MME
+    // for TF 1.11, send an arbitrary fixed request to the default model.
+    // if response is 400, the model is ok (but input was bad), so return 200
+    // for MME, the default model name is None and does not exist
+    // also return 200 in unlikely case our request was really valid
+
+    var uri = make_tfs_uri(r, true)
+    var options = {
+        method: 'POST',
+        body: '{"instances": "invalid"}'
+    }
+
+    function callback (reply) {
+        if (reply.status == 200 || reply.status == 400 ||
+        reply.responseText.includes('Servable not found for request: Latest(None)')) {
+            r.return(200)
+        } else {
+            r.error('failed ping' + reply.responseText)
+            r.return(502)
+        }
+    }
+
+    r.subrequest(uri, options, callback)
+}
+
+function return_error(r, code, message) {
+    if (message) {
+        r.return(code, '{"error": "' + message + '"}')
+    } else {
+        r.return(code)
+    }
+}
+
+function tfs_json_request(r, json) {
+    var uri = make_tfs_uri(r, true)
+    var options = {
+        method: 'POST',
+        body: json
+    }
+
+    var accept = r.headersIn.Accept
+    function callback (reply) {
+        var body = reply.responseText
+        if (reply.status == 400) {
+            // "fix" broken json escaping in \'instances\' message
+            body = body.replace("\\'instances\\'", "'instances'")
+        }
+
+        if (accept != undefined) {
+            var content_types = accept.trim().replace(" ", "").split(",")
+            if (content_types.includes('application/jsonlines') || content_types.includes('application/json')) {
+                body = body.replace(/\n/g, '')
+                r.headersOut['Content-Type'] = content_types[0]
+            }
+        }
+        r.return(reply.status, body)
+    }
+
+    r.subrequest(uri, options, callback)
+
+}
+
+function make_tfs_uri(r, with_method) {
+    var attributes = parse_custom_attributes(r)
+
+    var uri = tfs_base_uri + attributes['tfs-model-name']
+    if ('tfs-model-version' in attributes) {
+        uri += '/versions/' + attributes['tfs-model-version']
+    }
+
+    if (with_method) {
+        uri += ':' + (attributes['tfs-method'] || 'predict')
+    }
+
+    return uri
+}
+
+function parse_custom_attributes(r) {
+    var attributes = {}
+    var kv_pattern = /tfs-[a-z\-]+=[^,]+/g
+    var header = r.headersIn[custom_attributes_header]
+    if (header) {
+        var matches = header.match(kv_pattern)
+        if (matches) {
+            for (var i = 0; i < matches.length; i++) {
+                var kv = matches[i].split('=')
+                if (kv.length === 2) {
+                    attributes[kv[0]] = kv[1]
+                }
+            }
+        }
+    }
+
+    // for MME invocations, tfs-model-name is in the uri, or use default_tfs_model
+    if (!attributes['tfs-model-name']) {
+        var uri_pattern = /\/models\/[^,]+\/invoke/g
+        var model_name = r.uri.match(uri_pattern)
+        if (model_name[0]) {
+            model_name = r.uri.replace('/models/', '').replace('/invoke', '')
+            attributes['tfs-model-name'] = model_name
+        } else {
+            attributes['tfs-model-name'] = r.variables.default_tfs_model
+        }
+    }
+
+    return attributes
+}
+
+function json_request(r) {
+    var data = r.requestText
+
+    if (is_tfs_json(data)) {
+        tfs_json_request(r, data)
+    } else if (is_json_lines(data)) {
+        json_lines_request(r, data)
+    } else {
+        generic_json_request(r, data)
+    }
+}
+
+function is_tfs_json(data) {
+    return /"(instances|inputs|examples)"\s*:/.test(data)
+}
+
+function is_json_lines(data) {
+    // objects separated only by (optional) whitespace means jsons/json-lines
+    return /[}\]]\s*[\[{]/.test(data)
+}
+
+function generic_json_request(r, data) {
+    if (! /^\s*\[\s*\[/.test(data)) {
+        data = '[' + data + ']'
+    }
+
+    var json = '{"instances":' + data + '}'
+    tfs_json_request(r, json)
+}
+
+function json_lines_request(r, data) {
+    var lines = data.trim().split(/\r?\n/)
+    var builder = []
+    builder.push('{"instances":')
+    if (lines.length != 1) {
+        builder.push('[')
+    }
+
+    for (var i = 0; i < lines.length; i++) {
+        var line = lines[i].trim()
+        if (line) {
+            var instance = (i == 0) ? '' : ','
+            instance += line
+            builder.push(instance)
+        }
+    }
+
+    builder.push(lines.length == 1 ? '}' : ']}')
+    tfs_json_request(r, builder.join(''))
+}
+
+function csv_request(r) {
+    var data = r.requestText
+    // look for initial quote or numeric-only data in 1st field
+    var needs_quotes = data.search(/^\s*("|[\d.Ee+\-]+.*)/) != 0
+    var lines = data.trim().split(/\r?\n/)
+    var builder = []
+    builder.push('{"instances":[')
+
+    for (var i = 0; i < lines.length; i++) {
+        var line = lines[i].trim()
+        if (line) {
+            var line_builder = []
+            // Only wrap line in brackets if there are multiple columns.
+            // If there's only one column and it has a string with a comma,
+            // the input will be wrapped in an extra set of brackets.
+            var has_multiple_columns = line.search(',') != -1
+
+            if (has_multiple_columns) {
+                line_builder.push('[')
+            }
+
+            if (needs_quotes) {
+                line_builder.push('"')
+                line_builder.push(line.replace('"', '\\"').replace(',', '","'))
+                line_builder.push('"')
+            } else {
+                line_builder.push(line)
+            }
+
+            if (has_multiple_columns) {
+                line_builder.push(']')
+            }
+
+            var json_line = line_builder.join('')
+            builder.push(json_line)
+
+            if (i != lines.length - 1)
+                builder.push(',')
+        }
+    }
+
+    builder.push(']}')
+    tfs_json_request(r, builder.join(''))
+}
+
+export default {invocations, ping, ping_without_model, return_error,
+    tfs_json_request, make_tfs_uri, parse_custom_attributes,
+    json_request, is_tfs_json, is_json_lines, generic_json_request,
+    json_lines_request, csv_request};
diff --git a/scripts/tensorflow/inference/sagemaker/tfs_utils.py b/scripts/tensorflow/inference/sagemaker/tfs_utils.py
new file mode 100644
index 000000000000..22493fd2b1a7
--- /dev/null
+++ b/scripts/tensorflow/inference/sagemaker/tfs_utils.py
@@ -0,0 +1,336 @@
+# Copyright 2020 Amazon.com, Inc. or its affiliates. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"). You
+# may not use this file except in compliance with the License. A copy of
+# the License is located at
+#
+#     http://aws.amazon.com/apache2.0/
+#
+# or in the "license" file accompanying this file. This file is
+# distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF
+# ANY KIND, either express or implied. See the License for the specific
+# language governing permissions and limitations under the License.
+
+import json
+import logging
+import multiprocessing
+import os
+import re
+import time
+from collections import namedtuple
+
+import requests
+from multi_model_utils import MultiModelException
+from urllib3.exceptions import MaxRetryError, NewConnectionError
+from urllib3.util.retry import Retry
+
+logging.basicConfig(level=logging.INFO)
+log = logging.getLogger(__name__)
+
+DEFAULT_CONTENT_TYPE = "application/json"
+DEFAULT_ACCEPT_HEADER = "application/json"
+CUSTOM_ATTRIBUTES_HEADER = "X-Amzn-SageMaker-Custom-Attributes"
+
+Context = namedtuple(
+    "Context",
+    "model_name, model_version, method, rest_uri, grpc_port, channel, "
+    "custom_attributes, request_content_type, accept_header, content_length",
+)
+
+
+def parse_request(req, rest_port, grpc_port, default_model_name, model_name=None, channel=None):
+    tfs_attributes = parse_tfs_custom_attributes(req)
+    tfs_uri = make_tfs_uri(rest_port, tfs_attributes, default_model_name, model_name)
+
+    if not model_name:
+        model_name = tfs_attributes.get("tfs-model-name")
+
+    context = Context(
+        model_name,
+        tfs_attributes.get("tfs-model-version"),
+        tfs_attributes.get("tfs-method"),
+        tfs_uri,
+        grpc_port,
+        channel,
+        req.get_header(CUSTOM_ATTRIBUTES_HEADER),
+        req.get_header("Content-Type") or DEFAULT_CONTENT_TYPE,
+        req.get_header("Accept") or DEFAULT_ACCEPT_HEADER,
+        req.content_length,
+    )
+
+    data = req.stream
+    return data, context
+
+
+def make_tfs_uri(port, attributes, default_model_name, model_name=None):
+    log.info("sagemaker tfs attributes: \n{}".format(attributes))
+
+    tfs_model_name = model_name or attributes.get("tfs-model-name", default_model_name)
+    tfs_model_version = attributes.get("tfs-model-version")
+    tfs_method = attributes.get("tfs-method", "predict")
+
+    uri = "http://localhost:{}/v1/models/{}".format(port, tfs_model_name)
+    if tfs_model_version:
+        uri += "/versions/" + tfs_model_version
+    uri += ":" + tfs_method
+    return uri
+
+
+def parse_tfs_custom_attributes(req):
+    attributes = {}
+    header = req.get_header(CUSTOM_ATTRIBUTES_HEADER)
+    if header:
+        matches = re.findall(r"(tfs-[a-z\-]+=[^,]+)", header)
+        attributes = dict(attribute.split("=") for attribute in matches)
+    return attributes
+
+
+def create_tfs_config_individual_model(model_name, base_path):
+    config = "model_config_list: {\n"
+    config += "  config: {\n"
+    config += "    name: '{}'\n".format(model_name)
+    config += "    base_path: '{}'\n".format(base_path)
+    config += "    model_platform: 'tensorflow'\n"
+
+    config += "    model_version_policy: {\n"
+    config += "      specific: {\n"
+    for version in find_model_versions(base_path):
+        config += "        versions: {}\n".format(version)
+    config += "      }\n"
+    config += "    }\n"
+
+    config += "  }\n"
+    config += "}\n"
+    return config
+
+
+def tfs_command(
+    tfs_grpc_port,
+    tfs_rest_port,
+    tfs_config_path,
+    tfs_enable_batching,
+    tfs_batching_config_file,
+    tfs_intra_op_parallelism=None,
+    tfs_inter_op_parallelism=None,
+    tfs_enable_gpu_memory_fraction=False,
+    tfs_gpu_memory_fraction=None,
+):
+    cmd = (
+        "tensorflow_model_server "
+        "--port={} "
+        "--rest_api_port={} "
+        "--model_config_file={} "
+        "--max_num_load_retries=0 {} {} {} {}".format(
+            tfs_grpc_port,
+            tfs_rest_port,
+            tfs_config_path,
+            get_tfs_batching_args(tfs_enable_batching, tfs_batching_config_file),
+            get_tensorflow_intra_op_parallelism_args(tfs_intra_op_parallelism),
+            get_tensorflow_inter_op_parallelism_args(tfs_inter_op_parallelism),
+            get_tfs_gpu_mem_args(tfs_enable_gpu_memory_fraction, tfs_gpu_memory_fraction),
+        )
+    )
+    return cmd
+
+
+def find_models():
+    base_path = "/opt/ml/model"
+    models = []
+    for f in _find_saved_model_files(base_path):
+        parts = f.split("/")
+        if len(parts) >= 6 and re.match(r"^\d+$", parts[-2]):
+            model_path = "/".join(parts[0:-2])
+            if model_path not in models:
+                models.append(model_path)
+    return models
+
+
+def find_model_versions(model_path):
+    """Remove leading zeros from the version number, returns list of versions"""
+    return [
+        version[:-1].lstrip("0") + version[-1]
+        for version in os.listdir(model_path)
+        if version.isnumeric()
+    ]
+
+
+def _find_saved_model_files(path):
+    for e in os.scandir(path):
+        if e.is_dir():
+            yield from _find_saved_model_files(os.path.join(path, e.name))
+        else:
+            if e.name == "saved_model.pb":
+                yield os.path.join(path, e.name)
+
+
+def get_tfs_batching_args(enable_batching, tfs_batching_config):
+    if enable_batching:
+        return "--enable_batching=true --batching_parameters_file={}".format(tfs_batching_config)
+    else:
+        return ""
+
+
+def get_tensorflow_intra_op_parallelism_args(tfs_intra_op_parallelism):
+    if tfs_intra_op_parallelism:
+        return "--tensorflow_intra_op_parallelism={}".format(tfs_intra_op_parallelism)
+    else:
+        return ""
+
+
+def get_tensorflow_inter_op_parallelism_args(tfs_inter_op_parallelism):
+    if tfs_inter_op_parallelism:
+        return "--tensorflow_inter_op_parallelism={}".format(tfs_inter_op_parallelism)
+    else:
+        return ""
+
+
+def get_tfs_gpu_mem_args(enable_gpu_memory_fraction, gpu_memory_fraction):
+    if enable_gpu_memory_fraction and gpu_memory_fraction:
+        return "--per_process_gpu_memory_fraction={}".format(gpu_memory_fraction)
+    else:
+        return ""
+
+
+def create_batching_config(batching_config_file):
+    class _BatchingParameter:
+        def __init__(self, key, env_var, value, defaulted_message):
+            self.key = key
+            self.env_var = env_var
+            self.value = value
+            self.defaulted_message = defaulted_message
+
+    cpu_count = multiprocessing.cpu_count()
+    batching_parameters = [
+        _BatchingParameter(
+            "max_batch_size",
+            "SAGEMAKER_TFS_MAX_BATCH_SIZE",
+            8,
+            "max_batch_size defaulted to {}. Set {} to override default. "
+            "Tuning this parameter may yield better performance.",
+        ),
+        _BatchingParameter(
+            "batch_timeout_micros",
+            "SAGEMAKER_TFS_BATCH_TIMEOUT_MICROS",
+            1000,
+            "batch_timeout_micros defaulted to {}. Set {} to override "
+            "default. Tuning this parameter may yield better performance.",
+        ),
+        _BatchingParameter(
+            "num_batch_threads",
+            "SAGEMAKER_TFS_NUM_BATCH_THREADS",
+            cpu_count,
+            "num_batch_threads defaulted to {},the number of CPUs. Set {} to override default.",
+        ),
+        _BatchingParameter(
+            "max_enqueued_batches",
+            "SAGEMAKER_TFS_MAX_ENQUEUED_BATCHES",
+            # Batch limits number of concurrent requests, which limits number
+            # of enqueued batches, so this can be set high for Batch
+            100000000 if "SAGEMAKER_BATCH" in os.environ else cpu_count,
+            "max_enqueued_batches defaulted to {}. Set {} to override default. "
+            "Tuning this parameter may be necessary to tune out-of-memory "
+            "errors occur.",
+        ),
+    ]
+
+    warning_message = ""
+    for batching_parameter in batching_parameters:
+        if batching_parameter.env_var in os.environ:
+            batching_parameter.value = os.environ[batching_parameter.env_var]
+        else:
+            warning_message += batching_parameter.defaulted_message.format(
+                batching_parameter.value, batching_parameter.env_var
+            )
+            warning_message += "\n"
+    if warning_message:
+        log.warning(warning_message)
+
+    config = ""
+    for batching_parameter in batching_parameters:
+        config += "%s { value: %s }\n" % (batching_parameter.key, batching_parameter.value)
+
+    log.info("batching config: \n%s\n", config)
+    with open(batching_config_file, "w", encoding="utf8") as f:
+        f.write(config)
+
+
+def wait_for_model(rest_port, model_name, timeout_seconds, pid=None):
+    """
+    Notice:
+    The calculation for retry count based on timeout_seconds might introduce a small delta (0.1s) for each retry
+    which might cause total timeout longer than timeout_seconds
+    """
+    tfs_url = "http://localhost:{}/v1/models/{}".format(rest_port, model_name)
+    start = time.time()
+    try:
+        session = requests.Session()
+        backoff_factor = 0.1
+        # sleep = {backoff factor} * (2 ^ ({number of retries so far} - 1))
+        retry_count = retry_from_timeout(timeout_seconds, backoff_factor)
+        retries = Retry(total=retry_count, backoff_factor=backoff_factor)
+        session.mount("http://", requests.adapters.HTTPAdapter(max_retries=retries))
+        log.info(
+            "Trying to connect with model server: {} with timeout : {} and retry : {}".format(
+                tfs_url, timeout_seconds, retry_count
+            )
+        )
+        response = session.get(tfs_url, timeout=0.1)
+        log.info(
+            f"tfs response status_code: {response.status_code} with content : {json.loads(response.content)}"
+        )
+        end = time.time()
+        if response.status_code == 200:
+            if is_model_ready(response):
+                return
+            elif wait_for_model_ready(tfs_url, timeout_seconds - int(end - start)):
+                return
+
+        raise MultiModelException(408, "Timed out after {} seconds".format(timeout_seconds), pid)
+    except (
+        ConnectionRefusedError,
+        NewConnectionError,
+        MaxRetryError,
+        requests.exceptions.ConnectionError,
+    ):
+        raise MultiModelException(408, "Timed out after {} seconds".format(timeout_seconds), pid)
+
+
+def is_model_ready(response):
+    versions = json.loads(response.content)["model_version_status"]
+    if all(version["state"] == "AVAILABLE" for version in versions):
+        return True
+    return False
+
+
+def wait_for_model_ready(url, timeout_seconds):
+    try:
+        while timeout_seconds > 0:
+            response = requests.get(url, timeout=0.1)
+            log.info(
+                f"wait_for_model_ready response status_code : {response.status_code} "
+                f"response : {json.loads(response.content)} timeout in : {timeout_seconds}s"
+            )
+            if response.status_code != 200:
+                return False
+            if is_model_ready(response):
+                return True
+            timeout_seconds -= 1
+        return False
+    except requests.exceptions.RequestException:
+        return False
+
+
+def retry_from_timeout(timeout_seconds, backoff_factor):
+    retry_count = 1
+    retry_time = 0
+    while retry_time < timeout_seconds:
+        retry_count += 1
+        retry_time += backoff_factor * (2 ** (retry_count + 1))
+    return retry_count
+
+
+def get_cpu_memory_util():
+    total_memory, used_memory, free_memory = map(
+        int, os.popen("free -t -m").readlines()[-1].split()[1:]
+    )
+    return round((used_memory / total_memory) * 100, 2)
diff --git a/scripts/tensorflow/inference/tf_serving_entrypoint.sh b/scripts/tensorflow/inference/tf_serving_entrypoint.sh
new file mode 100755
index 000000000000..60d67f9a7515
--- /dev/null
+++ b/scripts/tensorflow/inference/tf_serving_entrypoint.sh
@@ -0,0 +1,6 @@
+#!/bin/bash
+
+# Execute telemetry script if it exists, suppress errors
+bash /usr/local/bin/bash_telemetry.sh >/dev/null 2>&1 || true
+
+/usr/local/bin/tensorflow_model_server --port=8500 --rest_api_port=8501 --model_name=${MODEL_NAME} --model_base_path=${MODEL_BASE_PATH}/${MODEL_NAME} "$@"
diff --git a/test/security/data/ecr_scan_allowlist/tensorflow/tensorflow-2.20.0.json b/test/security/data/ecr_scan_allowlist/tensorflow/tensorflow-2.20.0.json
new file mode 100644
index 000000000000..c179c6356f21
--- /dev/null
+++ b/test/security/data/ecr_scan_allowlist/tensorflow/tensorflow-2.20.0.json
@@ -0,0 +1,12 @@
+[
+    {
+        "vulnerability_id": "CVE-2025-23339",
+        "reason": "cuda-toolkit-config-common 12.9.79 — fix requires CUDA 13 major bump (>=0:13.0.48-1) which cascades through TFS binary compatibility, cuDNN, and the parallel TF 2.21 training PR. CUDA 12.9 is pinned to match the AL2023 base image; CUDA 13 currency tracked separately.",
+        "review_by": "2026-09-13"
+    },
+    {
+        "vulnerability_id": "CVE-2025-23308",
+        "reason": "cuda-toolkit-config-common 12.9.79 — fix requires CUDA 13 major bump (>=0:13.0.48-1) which cascades through TFS binary compatibility, cuDNN, and the parallel TF 2.21 training PR. CUDA 12.9 is pinned to match the AL2023 base image; CUDA 13 currency tracked separately.",
+        "review_by": "2026-09-13"
+    }
+]
diff --git a/test/tensorflow/integration/inference/.gitkeep b/test/tensorflow/integration/inference/.gitkeep
new file mode 100644
index 000000000000..e69de29bb2d1
diff --git a/test/tensorflow/integration/inference/__init__.py b/test/tensorflow/integration/inference/__init__.py
new file mode 100644
index 000000000000..e69de29bb2d1
diff --git a/test/tensorflow/integration/inference/conftest.py b/test/tensorflow/integration/inference/conftest.py
new file mode 100644
index 000000000000..506170347040
--- /dev/null
+++ b/test/tensorflow/integration/inference/conftest.py
@@ -0,0 +1,137 @@
+"""Pytest fixtures for TF 2.20 inference integration tests on SageMaker.
+
+Uses the SageMaker Python SDK v3 (``sagemaker>=3.0.0``) — the v2 Estimator /
+Model / Predictor classes were removed in v3 in favor of the unified
+``ModelBuilder`` and the ``sagemaker-core`` resource layer
+(``Endpoint``, ``EndpointConfig``, ``Model``, ``ContainerDefinition``,
+``ProductionVariant``). For these DLC tests we already have a custom
+``image_uri`` and a pre-built ``model.tar.gz``, so the simplest v3 path is
+the resource layer directly: ``Model.create -> EndpointConfig.create ->
+Endpoint.create -> endpoint.invoke()``. ``ModelBuilder`` is the right choice
+when the SDK should auto-detect the framework / container / packaging — for
+us, those are all fixed by the test fixture inputs.
+
+Fixtures intentionally defer all AWS calls until test-execution time so that
+``pytest --collect-only`` works in environments without AWS credentials.
+"""
+
+from __future__ import annotations
+
+import os
+import time
+from uuid import uuid4
+
+import pytest
+
+
+@pytest.fixture(scope="session")
+def aws_region() -> str:
+    """AWS region for SageMaker operations. Defaults to us-west-2."""
+    return os.environ.get("AWS_REGION", "us-west-2")
+
+
+@pytest.fixture(scope="session")
+def sagemaker_role_arn() -> str:
+    """SageMaker execution role ARN. Skips the test if not set."""
+    arn = os.environ.get("SAGEMAKER_ROLE_ARN")
+    if not arn:
+        pytest.skip("SAGEMAKER_ROLE_ARN not set")
+    return arn
+
+
+@pytest.fixture(scope="session")
+def inference_image_uri() -> str:
+    """ECR URI for the TF 2.20 inference image under test. Skips if not set."""
+    uri = os.environ.get("INFERENCE_IMAGE_URI")
+    if not uri:
+        pytest.skip("INFERENCE_IMAGE_URI not set")
+    return uri
+
+
+@pytest.fixture(scope="session")
+def boto_session(aws_region: str):
+    """A boto3 session bound to the configured region.
+
+    Used purely as a transport for ``sagemaker.core.helper.session_helper.Session``
+    and for the underlying ``s3`` client when uploading model artifacts; no
+    SageMaker control-plane calls go through it directly.
+    """
+    import boto3
+
+    return boto3.Session(region_name=aws_region)
+
+
+@pytest.fixture(scope="session")
+def sagemaker_session(boto_session):
+    """A SageMaker SDK v3 session.
+
+    ``sagemaker.core.helper.session_helper.Session`` is the v3 replacement for
+    the v2 ``sagemaker.Session``. We use it for ``default_bucket()`` and
+    ``upload_data()``; resource-layer ``create()`` calls accept it via the
+    ``session=`` kwarg.
+    """
+    from sagemaker.core.helper.session_helper import Session
+
+    return Session(boto_session=boto_session)
+
+
+@pytest.fixture
+def unique_name():
+    """Returns a callable producing collision-resistant resource names.
+
+    Usage:
+        name = unique_name("tf220-single")
+    """
+
+    def _make(prefix: str) -> str:
+        return f"{prefix}-{int(time.time())}-{uuid4().hex[:6]}"
+
+    return _make
+
+
+@pytest.fixture
+def cleanup_endpoint(boto_session):
+    """Yield-style fixture that tears down endpoint, endpoint config, and model.
+
+    Uses the v3 ``sagemaker-core`` resource layer (``Endpoint.get(...).delete()``,
+    etc.) rather than raw boto3 SDK calls, so cleanup code matches the deploy
+    code in the tests. The ``session=`` kwarg on resource ``get`` / ``create``
+    methods accepts a raw ``boto3.session.Session`` (see
+    ``sagemaker.core.utils.utils.SageMakerClient``); pass ``boto_session``
+    rather than the helper ``Session``.
+
+    Usage:
+        def test_x(cleanup_endpoint, ...):
+            cleanup_endpoint(endpoint_name, model_name=model_name)
+            # ... deploy + predict ...
+    """
+    registered: list[dict] = []
+
+    def _register(endpoint_name: str, model_name: str | None = None) -> None:
+        registered.append({"endpoint_name": endpoint_name, "model_name": model_name})
+
+    yield _register
+
+    # Import lazily so collection works without the SDK installed.
+    from sagemaker.core.resources import Endpoint, EndpointConfig, Model
+
+    for item in registered:
+        endpoint_name = item["endpoint_name"]
+        model_name = item["model_name"]
+
+        # Endpoint config name == endpoint name in our deploy flow below.
+        for resource_cls, get_kwargs in (
+            (Endpoint, {"endpoint_name": endpoint_name}),
+            (EndpointConfig, {"endpoint_config_name": endpoint_name}),
+        ):
+            try:
+                resource_cls.get(session=boto_session, **get_kwargs).delete()
+            except Exception:
+                # Best-effort teardown: swallow NotFound / already-deleted.
+                pass
+
+        if model_name:
+            try:
+                Model.get(model_name=model_name, session=boto_session).delete()
+            except Exception:
+                pass
diff --git a/test/tensorflow/integration/inference/requirements.txt b/test/tensorflow/integration/inference/requirements.txt
new file mode 100644
index 000000000000..f60dca57bf56
--- /dev/null
+++ b/test/tensorflow/integration/inference/requirements.txt
@@ -0,0 +1,3 @@
+boto3
+pytest
+sagemaker>=3.0.0
diff --git a/test/tensorflow/integration/inference/resources/__init__.py b/test/tensorflow/integration/inference/resources/__init__.py
new file mode 100644
index 000000000000..e69de29bb2d1
diff --git a/test/tensorflow/integration/inference/resources/build_sample_model.py b/test/tensorflow/integration/inference/resources/build_sample_model.py
new file mode 100644
index 000000000000..9d6020160ad9
--- /dev/null
+++ b/test/tensorflow/integration/inference/resources/build_sample_model.py
@@ -0,0 +1,60 @@
+"""Build a tiny TensorFlow SavedModel and tar it for SageMaker deployment.
+
+The model performs ``y = x * multiplier`` for a runtime-supplied multiplier.
+Tarballs are written to a caller-provided directory; nothing is checked in.
+"""
+
+from __future__ import annotations
+
+import os
+import tarfile
+import tempfile
+from pathlib import Path
+
+
+def build_sample_model(
+    output_dir: str | os.PathLike | None = None,
+    multiplier: float = 2.0,
+    model_name: str = "model",
+    tar_filename: str = "model.tar.gz",
+) -> str:
+    """Build a SavedModel that multiplies its input by ``multiplier``.
+
+    The SavedModel is laid out under ``<output_dir>/export/Servo/1`` so that
+    SageMaker's TensorFlow Serving handler can discover the version directory.
+    Returns the absolute path to the produced ``model.tar.gz``.
+    """
+    import tensorflow as tf
+
+    output_dir = Path(output_dir) if output_dir else Path(tempfile.mkdtemp(prefix="tf220-sample-"))
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    # SageMaker TFS expects: model.tar.gz -> <model_name>/<version>/saved_model.pb
+    saved_model_dir = output_dir / model_name / "1"
+    saved_model_dir.mkdir(parents=True, exist_ok=True)
+
+    multiplier_const = tf.constant(multiplier, dtype=tf.float32)
+
+    class MultiplierModel(tf.Module):
+        @tf.function(input_signature=[tf.TensorSpec(shape=[None, None], dtype=tf.float32)])
+        def __call__(self, x):
+            return {"output": x * multiplier_const}
+
+    model = MultiplierModel()
+    tf.saved_model.save(
+        model,
+        str(saved_model_dir),
+        signatures={"serving_default": model.__call__},
+    )
+
+    tar_path = output_dir / tar_filename
+    with tarfile.open(tar_path, "w:gz") as tar:
+        # Archive the model_name dir as the top-level entry inside the tarball.
+        tar.add(str(output_dir / model_name), arcname=model_name)
+
+    return str(tar_path)
+
+
+if __name__ == "__main__":
+    path = build_sample_model()
+    print(path)
diff --git a/test/tensorflow/integration/inference/test_multi_model_endpoint.py b/test/tensorflow/integration/inference/test_multi_model_endpoint.py
new file mode 100644
index 000000000000..ebea37e42d58
--- /dev/null
+++ b/test/tensorflow/integration/inference/test_multi_model_endpoint.py
@@ -0,0 +1,143 @@
+"""Multi-model endpoint (MME) integration test for TF 2.20 inference DLC.
+
+Builds two tiny SavedModels (``y = 2x`` and ``y = 3x``), uploads both to a
+shared S3 prefix, deploys a SageMaker MME backed by the v2 inference image,
+and asserts that ``target_model`` routes invocations correctly.
+
+Uses the SageMaker Python SDK v3 ``sagemaker-core`` resource layer — v3
+removed ``sagemaker.multidatamodel.MultiDataModel``. The native MME wire
+contract (``ContainerDefinition.mode = "MultiModel"``,
+``model_data_url = s3://bucket/prefix/``, plus the
+``X-Amzn-SageMaker-Target-Model`` header on invoke) is unchanged, so we
+express it directly: ``Model.create`` with ``mode="MultiModel"`` and an S3
+prefix in ``model_data_url``, then ``endpoint.invoke(target_model=...)``
+which sets the runtime header for us.
+"""
+
+from __future__ import annotations
+
+import json
+import tempfile
+from pathlib import Path
+
+import pytest
+
+from .resources.build_sample_model import build_sample_model
+
+INSTANCE_TYPE = "ml.c5.xlarge"
+
+
+def _values_from_predictions(predictions) -> list:
+    """Pull the numeric output list out of either signature-keyed or raw rows."""
+    assert predictions and isinstance(predictions, list)
+    first = predictions[0]
+    if isinstance(first, dict) and "output" in first:
+        return first["output"]
+    return first
+
+
+def test_mme_two_models(
+    boto_session,
+    sagemaker_session,
+    sagemaker_role_arn,
+    inference_image_uri,
+    unique_name,
+    cleanup_endpoint,
+):
+    from sagemaker.core.resources import (
+        ContainerDefinition,
+        Endpoint,
+        EndpointConfig,
+        Model,
+        ProductionVariant,
+    )
+
+    with tempfile.TemporaryDirectory(prefix="tf220-mme-") as workdir:
+        workdir_path = Path(workdir)
+
+        # Build two models with different multipliers, each in its own subdir
+        # so build_sample_model doesn't collide on the SavedModel layout.
+        model1_dir = workdir_path / "m1"
+        model2_dir = workdir_path / "m2"
+        model1_tar = build_sample_model(
+            output_dir=model1_dir, multiplier=2.0, model_name="model", tar_filename="model1.tar.gz"
+        )
+        model2_tar = build_sample_model(
+            output_dir=model2_dir, multiplier=3.0, model_name="model", tar_filename="model2.tar.gz"
+        )
+
+        bucket = sagemaker_session.default_bucket()
+        run_id = unique_name("mme")
+        s3_key_prefix = f"tf220-inference-tests/mme-models/{run_id}"
+
+        # Upload each tarball under the shared MME prefix so the runtime can
+        # resolve target_model relative to the same S3 location.
+        sagemaker_session.upload_data(path=model1_tar, bucket=bucket, key_prefix=s3_key_prefix)
+        sagemaker_session.upload_data(path=model2_tar, bucket=bucket, key_prefix=s3_key_prefix)
+        s3_model_prefix = f"s3://{bucket}/{s3_key_prefix}/"
+
+        endpoint_name = unique_name("tf220-mme")
+        model_name = unique_name("tf220-mme-model")
+        cleanup_endpoint(endpoint_name, model_name=model_name)
+
+        # 1. Create a multi-model SageMaker Model. The MME contract is
+        #    expressed at the container definition level: mode="MultiModel"
+        #    plus an S3 *prefix* (not a single tar) in model_data_url.
+        Model.create(
+            model_name=model_name,
+            primary_container=ContainerDefinition(
+                image=inference_image_uri,
+                mode="MultiModel",
+                model_data_url=s3_model_prefix,
+            ),
+            execution_role_arn=sagemaker_role_arn,
+            session=boto_session,
+        )
+
+        # 2. Endpoint config + endpoint — same shape as single-model.
+        EndpointConfig.create(
+            endpoint_config_name=endpoint_name,
+            production_variants=[
+                ProductionVariant(
+                    variant_name="AllTraffic",
+                    model_name=model_name,
+                    initial_instance_count=1,
+                    instance_type=INSTANCE_TYPE,
+                ),
+            ],
+            session=boto_session,
+        )
+
+        endpoint = Endpoint.create(
+            endpoint_name=endpoint_name,
+            endpoint_config_name=endpoint_name,
+            session=boto_session,
+        )
+        endpoint.wait_for_status("InService")
+
+        payload = json.dumps({"instances": [[1.0, 2.0, 3.0]]})
+
+        # 3. Invoke each model by name. ``target_model`` maps to the
+        #    X-Amzn-SageMaker-Target-Model header that selects the tarball
+        #    within the MME's S3 prefix.
+        resp1 = endpoint.invoke(
+            body=payload,
+            content_type="application/json",
+            accept="application/json",
+            target_model="model1.tar.gz",
+        )
+        body1 = json.loads(resp1.body.read().decode("utf-8"))
+        assert "predictions" in body1, f"model1 response missing predictions: {body1!r}"
+        values1 = _values_from_predictions(body1["predictions"])
+        assert values1 == pytest.approx([2.0, 4.0, 6.0]), f"model1 got {values1!r}"
+
+        resp2 = endpoint.invoke(
+            body=payload,
+            content_type="application/json",
+            accept="application/json",
+            target_model="model2.tar.gz",
+        )
+        body2 = json.loads(resp2.body.read().decode("utf-8"))
+        assert "predictions" in body2, f"model2 response missing predictions: {body2!r}"
+        values2 = _values_from_predictions(body2["predictions"])
+        assert values2 == pytest.approx([3.0, 6.0, 9.0]), f"model2 got {values2!r}"
diff --git a/test/tensorflow/integration/inference/test_single_model_endpoint.py b/test/tensorflow/integration/inference/test_single_model_endpoint.py
new file mode 100644
index 000000000000..c632377a4d8a
--- /dev/null
+++ b/test/tensorflow/integration/inference/test_single_model_endpoint.py
@@ -0,0 +1,123 @@
+"""Single-model endpoint integration test for TF 2.20 inference DLC.
+
+Builds a tiny ``y = 2x`` SavedModel, deploys it to a single-instance SageMaker
+endpoint backed by the v2 inference image under test, and asserts the
+predicted values.
+
+Uses the SageMaker Python SDK v3 ``sagemaker-core`` resource layer
+(``Endpoint``, ``EndpointConfig``, ``Model``, ``ContainerDefinition``,
+``ProductionVariant``) — the v2 ``TensorFlowModel`` / ``Predictor`` classes
+were removed in v3. ``ModelBuilder`` is the v3 entry point for
+auto-detected deployments, but for DLC tests we already supply the
+``image_uri`` and a pre-built ``model.tar.gz``, so we go straight to the
+resource layer (the same surface ``ModelBuilder`` calls underneath).
+"""
+
+from __future__ import annotations
+
+import json
+import tempfile
+from pathlib import Path
+
+import pytest
+
+from .resources.build_sample_model import build_sample_model
+
+INSTANCE_TYPE = "ml.c5.xlarge"
+
+
+def test_single_model_predict(
+    boto_session,
+    sagemaker_session,
+    sagemaker_role_arn,
+    inference_image_uri,
+    unique_name,
+    cleanup_endpoint,
+):
+    from sagemaker.core.resources import (
+        ContainerDefinition,
+        Endpoint,
+        EndpointConfig,
+        Model,
+        ProductionVariant,
+    )
+
+    with tempfile.TemporaryDirectory(prefix="tf220-single-") as workdir:
+        tar_path = build_sample_model(
+            output_dir=workdir,
+            multiplier=2.0,
+            model_name="model",
+        )
+
+        # Upload the tarball via the v3 helper Session — same default-bucket /
+        # upload_data ergonomics as v2.
+        bucket = sagemaker_session.default_bucket()
+        key_prefix = f"tf220-inference-tests/{Path(tar_path).stem}-{unique_name('single')}"
+        model_data = sagemaker_session.upload_data(
+            path=tar_path,
+            bucket=bucket,
+            key_prefix=key_prefix,
+        )
+
+        endpoint_name = unique_name("tf220-single")
+        model_name = unique_name("tf220-single-model")
+        cleanup_endpoint(endpoint_name, model_name=model_name)
+
+        # 1. Create the SageMaker Model — points at our DLC image and the
+        #    uploaded SavedModel tar.gz.
+        Model.create(
+            model_name=model_name,
+            primary_container=ContainerDefinition(
+                image=inference_image_uri,
+                model_data_url=model_data,
+            ),
+            execution_role_arn=sagemaker_role_arn,
+            session=boto_session,
+        )
+
+        # 2. Create the EndpointConfig with a single ProductionVariant.
+        EndpointConfig.create(
+            endpoint_config_name=endpoint_name,
+            production_variants=[
+                ProductionVariant(
+                    variant_name="AllTraffic",
+                    model_name=model_name,
+                    initial_instance_count=1,
+                    instance_type=INSTANCE_TYPE,
+                ),
+            ],
+            session=boto_session,
+        )
+
+        # 3. Create the Endpoint and wait for it to come InService.
+        endpoint = Endpoint.create(
+            endpoint_name=endpoint_name,
+            endpoint_config_name=endpoint_name,
+            session=boto_session,
+        )
+        endpoint.wait_for_status("InService")
+
+        # 4. Invoke. ``Endpoint.invoke`` returns an InvokeEndpointOutput whose
+        #    ``body`` is a streaming bytes-like object.
+        payload = json.dumps({"instances": [[1.0, 2.0, 3.0]]})
+        result = endpoint.invoke(
+            body=payload,
+            content_type="application/json",
+            accept="application/json",
+        )
+        response = json.loads(result.body.read().decode("utf-8"))
+
+        assert "predictions" in response, f"missing predictions key in {response!r}"
+        predictions = response["predictions"]
+        assert predictions and isinstance(predictions, list)
+
+        # Output signature is {"output": x * 2.0} -> TFS surfaces the tensor
+        # under the signature output key when there is a single named tensor;
+        # some TFS versions instead return the raw list. Handle both.
+        first = predictions[0]
+        if isinstance(first, dict) and "output" in first:
+            values = first["output"]
+        else:
+            values = first
+
+        assert values == pytest.approx([2.0, 4.0, 6.0]), f"got {values!r}"
diff --git a/test/tensorflow/pytest.ini b/test/tensorflow/pytest.ini
new file mode 100644
index 000000000000..eea2c180278f
--- /dev/null
+++ b/test/tensorflow/pytest.ini
@@ -0,0 +1 @@
+[pytest]