diff --git a/.github/workflows/ci.yaml b/.github/workflows/ci.yaml
index 143406673a8..3647f00e9cf 100644
--- a/.github/workflows/ci.yaml
+++ b/.github/workflows/ci.yaml
@@ -78,7 +78,7 @@ jobs:
           make docker-print TARGET=docker-gpu
           make docker-print TARGET=docker-auditor
           make docker-print TARGET=docker-python-base
-          make docker-print TARGET=gpu-wheels
+          make docker-print TARGET=nmp-automodel-gpu-wheels
           make docker-print TARGET=nmp-automodel
           make docker-print TARGET=nmp-unsloth
 
diff --git a/docker-bake.hcl b/docker-bake.hcl
index 8ebbc057728..edec3cd3dcd 100644
--- a/docker-bake.hcl
+++ b/docker-bake.hcl
@@ -18,7 +18,7 @@ variable "IMAGE_REGISTRY" {
   default = "my-registry"
 }
 
-# Registry where pinned base images are published (nmp-python-base, nmp-customizer-tasks-base, etc.)
+# Registry where pinned base images are published.
 # In CI this matches CI_REGISTRY_IMAGE; locally override if needed.
 variable "BASE_REGISTRY" {
   default = "my-registry"
@@ -36,10 +36,6 @@ variable "PYTHON_DEV_BASE_TARGET" {
   default = ""
 }
 
-variable "CUSTOMIZER_TASKS_BASE_CONTEXT" {
-  default = ""
-}
-
 variable "AUTOMODEL_BASE_CONTEXT" {
   default = ""
 }
@@ -48,10 +44,6 @@ variable "USE_LOCAL_WHEELS" {
   default = ""
 }
 
-variable "FAISS_WHEEL_CONTEXT" {
-  default = ""
-}
-
 variable "CAUSAL_CONV1D_WHEEL_CONTEXT" {
   default = ""
 }
@@ -86,11 +78,6 @@ variable "BASE_TAG_AUTOMODEL" {
   default = "dbdc61fdb117f48bc9569b4274afe836cf0de030"
 }
 
-# Pin for nmp-customizer-tasks-base (built by nmp-customizer-tasks-base-builder)
-variable "BASE_TAG_CUSTOMIZER_TASKS" {
-  default = "dbdc61fdb117f48bc9569b4274afe836cf0de030"
-}
-
 # The tag for base images if needed
 variable "WHEELS_TAG" {
   default = "3fd6986ff173b598446ffac06d9be3f84b482495"
@@ -124,14 +111,6 @@ variable "SAFE_SYNTHESIZER_CONTAINER_VARIANT" {
   default = "cu129"
 }
 
-variable "FAISS_VERSION" {
-  default = "v1.9.0"
-}
-
-variable "FAISS_CUDA_ARCHS" {
-  default = "75;80;86;90"
-}
-
 # Versions for the mamba wheel builder.
 variable "MAMBA_22_COMMIT" {
   default = "6b32be06d026e170b3fdaf3ae6282c5a6ff57b06"
@@ -145,11 +124,6 @@ variable "CAUSAL_CONV1D_VERSION" {
   default = "v1.5.3"
 }
 
-function "get_faiss_wheel_image" {
-  params = []
-  result = "${WHEELS_REGISTRY}/faiss-gpu-wheel:${WHEELS_TAG}"
-}
-
 function "get_causal_conv1d_wheel_image" {
   params = []
   result = "${WHEELS_REGISTRY}/causal-conv1d-wheel:${WHEELS_TAG}"
@@ -182,21 +156,11 @@ function "python_dev_base_target" {
   result = notequal(PYTHON_DEV_BASE_TARGET, "") ? PYTHON_DEV_BASE_TARGET : notequal(USE_PREBUILT_BASES, "") ? "nmp-python-dev-base" : "nmp-python-dev-base-builder"
 }
 
-function "customizer_tasks_base_context" {
-  params = []
-  result = notequal(CUSTOMIZER_TASKS_BASE_CONTEXT, "") ? CUSTOMIZER_TASKS_BASE_CONTEXT : notequal(USE_PREBUILT_BASES, "") ? "docker-image://${BASE_REGISTRY}/nmp-customizer-tasks-base:${BASE_TAG_CUSTOMIZER_TASKS}" : "target:nmp-customizer-tasks-base-builder"
-}
-
 function "automodel_base_context" {
   params = []
   result = notequal(AUTOMODEL_BASE_CONTEXT, "") ? AUTOMODEL_BASE_CONTEXT : notequal(USE_PREBUILT_BASES, "") ? "docker-image://${BASE_REGISTRY}/nmp-automodel-base:${BASE_TAG_AUTOMODEL}" : "target:nmp-automodel-base-builder"
 }
 
-function "faiss_wheel_context" {
-  params = []
-  result = notequal(FAISS_WHEEL_CONTEXT, "") ? FAISS_WHEEL_CONTEXT : notequal(USE_LOCAL_WHEELS, "") ? "target:faiss-gpu-wheel" : "docker-image://${get_faiss_wheel_image()}"
-}
-
 function "causal_conv1d_wheel_context" {
   params = []
   result = notequal(CAUSAL_CONV1D_WHEEL_CONTEXT, "") ? CAUSAL_CONV1D_WHEEL_CONTEXT : notequal(USE_LOCAL_WHEELS, "") ? "target:causal-conv1d-wheel" : "docker-image://${get_causal_conv1d_wheel_image()}"
@@ -259,15 +223,6 @@ group "docker-auditor" {
   ]
 }
 
-# Legacy groups for backward compatibility
-group "docker-amd64-only" {
-  targets = [
-    "customizer-rl-docker",
-    "customizer-tasks-docker",
-    "docker-gpu"
-  ]
-}
-
 group "all-multi-platform" {
   targets = [
     "docker-multi-platform",
@@ -277,7 +232,6 @@ group "all-multi-platform" {
 group "docker-multi-platform" {
   targets = [
     "docker-cpu",
-    "customizer-megatron-docker",
     "auditor-tasks-docker",
   ]
 }
@@ -298,7 +252,7 @@ group "all-arm64" {
 group "all-amd64" {
   targets = [
     "docker-multi-platform",
-    "docker-amd64-only",
+    "docker-gpu",
   ]
 }
 
@@ -306,7 +260,6 @@ group "docker" {
   targets = [
     "docker-cpu",
     "docker-gpu",
-    "docker-customizer",
     "docker-auditor",
   ]
 }
@@ -332,14 +285,6 @@ group "docker-gpu" {
   ]
 }
 
-group "gpu-wheels" {
-  targets = [
-    "faiss-gpu-wheel",
-    "causal-conv1d-wheel",
-    "mamba-ssm-wheel",
-  ]
-}
-
 group "nmp-automodel-gpu-wheels" {
   targets = [
     "causal-conv1d-wheel",
@@ -363,17 +308,6 @@ group "nmp-unsloth" {
   ]
 }
 
-# Final Customizer Training Images - WARNING: No not build base CUDA images using shared buildkit
-# Only build the final image.  All base image builds should happen using arch specific runners.
-group "docker-customizer" {
-  targets = [
-    "customizer-rl-docker",
-    "customizer-rl-smoke-test",
-    "customizer-tasks-docker",
-    "customizer-tasks-smoke-test",
-  ]
-}
-
 # Base images for consolidated containers
 target "nmp-python-base" {
   target     = python_base_target()
@@ -561,63 +495,6 @@ target "nmp-cpu-tasks-docker" {
   platforms  = get_platforms()
 }
 
-# Safe Synthesizer Tasks - GPU-intensive Safe Synthesizer task execution
-
-# NMP Customizer Tasks - GPU Tasks for Customizer (transformers>=5.0.0, mamba-ssm 2.3.0)
-
-# Builder for the Customizer Tasks base image
-target "nmp-customizer-tasks-base-builder" {
-  target     = "nmp-customizer-tasks-base-builder"
-  context    = "."
-  dockerfile = "docker/Dockerfile.nmp-customizer-tasks"
-  cache-to   = maybe_registry_cache_to("nmp-customizer-tasks-base")
-  cache-from = maybe_registry_cache_from("nmp-customizer-tasks-base")
-  tags       = base_tags("nmp-customizer-tasks-base")
-  output     = image_output()
-  contexts = {
-    causal-conv1d-wheel-image = causal_conv1d_wheel_context()
-    faiss-wheel-image         = faiss_wheel_context()
-    mamba-ssm-wheel-image     = mamba_ssm_wheel_context()
-    nmp-gpu-base              = "target:nmp-gpu-base"
-    nmp-workspace             = "target:nmp-workspace"
-  }
-  platforms  = ["linux/amd64"]
-}
-
-# The base image for Customizer Tasks
-target "nmp-customizer-tasks-base" {
-  target     = "nmp-customizer-tasks-base-builder"
-  context    = "."
-  contexts = {
-    causal-conv1d-wheel-image = causal_conv1d_wheel_context()
-    faiss-wheel-image         = faiss_wheel_context()
-    mamba-ssm-wheel-image     = mamba_ssm_wheel_context()
-    nmp-gpu-base              = "target:nmp-gpu-base"
-    nmp-workspace             = "target:nmp-workspace"
-  }
-  dockerfile = "docker/Dockerfile.nmp-customizer-tasks"
-  platforms  = get_platforms()
-}
-
-
-# FAISS GPU Wheel Builder
-# Clones FAISS from GitHub and produces a Python wheel with CUDA support.
-# The wheel lives at /faiss/wheels/*.whl inside the image.
-target "faiss-gpu-wheel" {
-  target     = "faiss-gpu-wheel"
-  context    = "."
-  dockerfile = "docker/base/Dockerfile.faiss-gpu-wheel"
-  cache-to   = maybe_registry_cache_to("faiss-gpu-wheel")
-  cache-from = maybe_registry_cache_from("faiss-gpu-wheel")
-  tags       = wheel_tags("faiss-gpu-wheel")
-  output     = image_output()
-  args = {
-    FAISS_VERSION = FAISS_VERSION
-    CUDA_ARCHS    = FAISS_CUDA_ARCHS
-  }
-  platforms = get_platforms()
-}
-
 # Mamba Wheel Builders
 # Builds Python wheels for mamba-ssm and causal-conv1d in parallel.
 # Both only ship source distributions on PyPI; this pre-builds them for
@@ -804,133 +681,6 @@ target "buildkit-test" {
   platforms       = get_platforms()
 }
 
-# Customizer
-
-# Customizer base targets (used as context injection)
-target "customizer-megatron-base" {
-  target     = "customizer-megatron-base"
-  context    = "."
-  dockerfile = "docker/Dockerfile.nmp-customizer"
-  platforms  = get_platforms()
-}
-
-target "customizer-rl-shared-base" {
-  target     = "customizer-rl-shared-base"
-  context    = "."
-  dockerfile = "docker/Dockerfile.nmp-customizer"
-  # nemo-rl only for amd64 at the moment
-  platforms  = ["linux/amd64"]
-}
-
-target "customizer-combined-base" {
-  target     = "customizer-combined-base"
-  context    = "."
-  dockerfile = "docker/Dockerfile.nmp-customizer"
-  platforms  = ["linux/amd64"]
-}
-
-# Customizer Base Build Targets
-
-# Customizer Final Images - base injected via contexts
-target "customizer-combined-docker" {
-  target     = "customizer-runtime"
-  context    = "."
-  dockerfile = "docker/Dockerfile.nmp-customizer"
-  contexts = {
-    nmp-workspace   = "target:nmp-workspace"
-    customizer-base = "target:customizer-combined-base"
-  }
-  cache-to   = maybe_registry_cache_to("customizer-combined")
-  cache-from = maybe_registry_cache_from("customizer-combined")
-  tags       = sha_and_maybe_latest_tags("customizer-combined")
-  output     = image_output()
-  platforms  = ["linux/amd64"]
-}
-
-target "customizer-megatron-docker" {
-  target     = "customizer-runtime"
-  context    = "."
-  dockerfile = "docker/Dockerfile.nmp-customizer"
-  contexts = {
-    nmp-workspace   = "target:nmp-workspace"
-    customizer-base = "target:customizer-megatron-base"
-  }
-  cache-to   = maybe_registry_cache_to("customizer-megatron")
-  cache-from = maybe_registry_cache_from("customizer-megatron")
-  tags       = sha_and_maybe_latest_tags("customizer-megatron")
-  output     = image_output()
-  platforms  = get_platforms()
-}
-
-target "customizer-tasks-smoke-test" {
-  target     = "smoke-test"
-  context    = "."
-  dockerfile = "docker/Dockerfile.nmp-customizer-tasks"
-  contexts = {
-    nmp-customizer-tasks-base = customizer_tasks_base_context()
-    causal-conv1d-wheel-image = causal_conv1d_wheel_context()
-    faiss-wheel-image         = faiss_wheel_context()
-    mamba-ssm-wheel-image     = mamba_ssm_wheel_context()
-    nmp-gpu-base              = "target:nmp-gpu-base"
-    nmp-workspace             = "target:nmp-workspace"
-  }
-  cache-from = maybe_registry_cache_from("customizer-tasks")
-  output     = ["type=cacheonly"]
-  platforms  = ["linux/amd64"]
-}
-
-target "customizer-tasks-docker" {
-  target     = "runtime"
-  context    = "."
-  dockerfile = "docker/Dockerfile.nmp-customizer-tasks"
-  contexts = {
-    nmp-customizer-tasks-base = customizer_tasks_base_context()
-    causal-conv1d-wheel-image = causal_conv1d_wheel_context()
-    faiss-wheel-image         = faiss_wheel_context()
-    mamba-ssm-wheel-image     = mamba_ssm_wheel_context()
-    nmp-gpu-base              = "target:nmp-gpu-base"
-    nmp-workspace             = "target:nmp-workspace"
-  }
-  cache-to   = maybe_registry_cache_to("customizer-tasks")
-  cache-from = maybe_registry_cache_from("customizer-tasks")
-  tags       = sha_and_maybe_latest_tags("customizer-tasks")
-  output     = image_output()
-  platforms  = ["linux/amd64"]
-}
-
-target "customizer-rl-docker" {
-  target     = "customizer-runtime"
-  context    = "."
-  dockerfile = "docker/Dockerfile.nmp-customizer"
-  contexts = {
-    nmp-workspace   = "target:nmp-workspace"
-    customizer-base = "target:customizer-rl-shared-base"
-  }
-  cache-to   = maybe_registry_cache_to("customizer-rl")
-  cache-from = maybe_registry_cache_from("customizer-rl")
-  tags       = sha_and_maybe_latest_tags("customizer-rl")
-  output     = image_output()
-  # only amd64 for now, until the next release of nemo-rl
-  platforms  = ["linux/amd64"]
-}
-
-target "customizer-rl-smoke-test" {
-  target     = "smoke-test"
-  context    = "."
-  dockerfile = "docker/Dockerfile.nmp-customizer"
-  contexts = {
-    nmp-workspace   = "target:nmp-workspace"
-    customizer-base = "target:customizer-rl-shared-base"
-  }
-  args = {
-    SMOKE_MARKER = "smoke_customizer_rl"
-  }
-  cache-from = maybe_registry_cache_from("customizer-rl")
-  output     = ["type=cacheonly"]
-  platforms  = ["linux/amd64"]
-}
-
-
 # Automodel and Unsloth
 
 target "automodel-platform-workspace" {
diff --git a/docker/Dockerfile.nmp-customizer b/docker/Dockerfile.nmp-customizer
deleted file mode 100644
index 4aec1fdca11..00000000000
--- a/docker/Dockerfile.nmp-customizer
+++ /dev/null
@@ -1,213 +0,0 @@
-#######
-# Legacy Customizer runtime Dockerfile.
-#
-# This file keeps the non-Automodel customizer images available while the team
-# decides whether to retire them:
-# - combined - includes nemo-rl plus Automodel/Megatron extras for comparison
-# - rl       - includes nemo-rl v0.4.0
-# - megatron - derived from nvidia/nemo:25.11.01
-#
-# The old customizer-automodel image and base builder are intentionally not
-# defined here.
-#######
-
-ARG SMOKE_MARKER=smoke_customizer_rl
-
-# BASE IMAGES
-
-#customizer-rl-shared-base:
-FROM nvcr.io/nvidia/nemo-rl:v0.4.0 AS customizer-rl-shared-base
-
-# setup virtual environment properly
-ENV VIRTUAL_ENV=/opt/nemo_rl_venv \
-    NO_OVERRIDE_REQUIREMENTS_PATH=/app/services/customizer/src/nmp/customizer/tasks/training/backends/nemo_rl/no_override_requirements.txt
-
-# nvcr.io/nvidia/nemo-rl:v0.4.0's VENV is built with root and the .cache and .local are unavailable...
-RUN chmod 755 /root /root/.cache /root/.cache/uv /root/.local /root/.local/share /root/.local/share/uv
-
-# CVE fix: dirmngr (CVE-2025-68973): upgrade to 2.2.27-3ubuntu2.5+
-# CVE-2025-68973 (High, out-of-SLA): Upgrade gnupg packages.
-# --only-upgrade: upgrades installed packages only; skips absent ones silently.
-# Remove unused nsight cli which is out of date
-RUN apt-get update && apt-get install -y --only-upgrade \
-    gnupg gnupg2 gpg gpg-agent gpgconf gpgsm \
-    dirmngr gnupg-utils gpgv \
-    && dpkg-query -f '${binary:Package}\n' -W 'nsight*' | xargs -r apt-get purge -y \
-    && apt-get autoremove \
-    && apt-get clean && rm -rf /var/lib/apt/lists/*
-
-# Fix CVE security vulnerabilities in pre-installed packages.
-# NOTE: These overrides are done here in the Dockerfile rather than in pyproject.toml
-# because the vulnerable packages come pre-installed in the NGC base image venv
-# (nvcr.io/nvidia/nemo-rl:v0.4.0), not from our uv workspace. The uv workspace's
-# override-dependencies only affect packages resolved by our own services; they have
-# no effect on packages already baked into the base image's virtual environment.
-# --no-deps is essential here for two reasons:
-#   1. The nemo-rl base image ships several packages with corrupted dist-info
-#      (missing METADATA files). Dependency resolution reads all installed
-#      package metadata, so it fails before installing anything. --no-deps
-#      skips resolution entirely.
-#   2. Without --no-deps, uv would pull transitive dependencies and potentially
-#      replace NGC-optimized builds of torch, numpy, etc. with generic PyPI
-#      wheels, breaking GPU functionality. --no-deps installs only the listed
-#      packages, leaving all other installed packages untouched.
-RUN uv pip install --upgrade setuptools && \
-    uv pip install --no-deps --no-cache \
-    'anyio>=4.10.0' \
-    "urllib3>=2.6.3" \
-    "starlette>=0.49.1" \
-    "protobuf>=5.29.6,<6.0.dev0" \
-    "cryptography>=46.0.5" \
-    "python-multipart>=0.0.20" \
-    "aiohttp>=3.13.3" \
-    "cbor2>=5.9.0" \
-    "pyasn1>=0.6.3" \
-    "xgrammar>=0.1.32" \
-    "Pillow>=12.1.1" \
-    "wheel>=0.46.2" \
-    "nltk>=3.9.3" \
-    'mlflow>=3.8.0rc0' \
-    'filelock>=3.25.0'
-
-# ray CVE fix (GHSA-6wgj-66m2-xxp2, GHSA-gx77-xgc2-4888, GHSA-q279-jhrf-cc6v):
-# ray 2.52+ restructured internal modules (ray._private to ray._common). A simple
-# overlay leaves remnant files from the old version. We must:
-#   1. Purge the old ray package directory and dist-info entirely
-#   2. Remove pre-built worker venvs (/opt/ray_venvs) so NeMo-RL rebuilds them
-#      with the new ray version at runtime
-#   3. Install the new ray wheel cleanly
-RUN set -e && \
-    rm -rf $VIRTUAL_ENV/lib/python3.12/site-packages/ray \
-           $VIRTUAL_ENV/lib/python3.12/site-packages/ray-*.dist-info && \
-    uv pip install --no-deps --no-cache 'ray>=2.52.0' && \
-    for venv_dir in /opt/ray_venvs/*/; do \
-        [ -d "$venv_dir/lib" ] || continue; \
-        rm -rf "$venv_dir"/lib/python*/site-packages/ray \
-               "$venv_dir"/lib/python*/site-packages/ray-*.dist-info; \
-        uv pip install --no-deps --no-cache --python "$venv_dir/bin/python" 'ray>=2.52.0'; \
-    done && \
-    python3 -c "import ray; print(f'ray {ray.__version__} at {ray.__file__}'); from ray._common.ray_constants import LOGGING_ROTATE_BYTES; print('ray._common OK')"
-
-# vllm CVE fix (GHSA-4r2x-xpjr-7cvv, GHSA-3f6c-7fw2-ppm4, GHSA-7972-pg2x-xr59,
-# GHSA-8fr4-5q9j-m8gm, GHSA-pmqf-x6x8-p7qw, GHSA-qh4c-xf7m-gxfc,
-# GHSA-wr9h-g72x-mwhm, GHSA-79j6-g2m3-jgfw, GHSA-rxc4-3w6r-4v47):
-# vllm is not used by the customizer's current training paths (DPO, SFT, LoRA).
-# It ships in the nemo-rl base image for generation/rollout (GRPO) which is not
-# yet implemented in v2. Removing it eliminates the CVEs without an upgrade.
-RUN rm -rf $VIRTUAL_ENV/lib/python3.12/site-packages/vllm \
-           $VIRTUAL_ENV/lib/python3.12/site-packages/vllm-*.dist-info && \
-    for venv_dir in /opt/ray_venvs/*/; do \
-        [ -d "$venv_dir/lib" ] || continue; \
-        rm -rf "$venv_dir"/lib/python*/site-packages/vllm \
-               "$venv_dir"/lib/python*/site-packages/vllm-*.dist-info; \
-    done
-
-#customizer-combined-base:
-FROM customizer-rl-shared-base AS customizer-combined-base
-# add megatron and automodel, not installed by default
-RUN --mount=type=cache,target=/root/.cache/uv \
-    cd /opt/nemo-rl && UV_HTTP_TIMEOUT=120 uv sync --frozen --extra automodel --extra mcore
-
-#customizer-megatron-base
-FROM nvcr.io/nvidia/nemo:25.11.01 AS customizer-megatron-base
-# nemo contains most recent version of megatron
-ENV VIRTUAL_ENV=/opt/venv
-
-# CVE fix: dirmngr (CVE-2025-68973): upgrade to 2.2.27-3ubuntu2.5+
-# CVE-2025-68973 (High, out-of-SLA): Upgrade gnupg packages.
-# --only-upgrade: upgrades installed packages only; skips absent ones silently.
-# Remove unused nsight cli which is out of date
-# Remove vllm which is not used (remediation)
-RUN apt-get update && apt-get install -y --only-upgrade \
-    gnupg gnupg2 gpg gpg-agent gpgconf gpgsm \
-    dirmngr gnupg-utils gpgv \
-    && dpkg-query -f '${binary:Package}\n' -W 'nsight*' | xargs -r apt-get purge -y \
-    && apt-get autoremove \
-    && apt-get clean && rm -rf /var/lib/apt/lists/* && \
-    rm -rf /usr/local/lib/python*/site-packages/vllm \
-           /usr/local/lib/python*/site-packages/vllm-*.dist-info
-
-# Fix CVE security vulnerabilities in pre-installed packages.
-# NOTE: These overrides are done here in the Dockerfile rather than in pyproject.toml
-# because the vulnerable packages come pre-installed in the NGC base image venv
-# (nvcr.io/nvidia/nemo:25.11.01), not from our uv workspace. The uv workspace's
-# override-dependencies only affect packages resolved by our own services; they have
-# no effect on packages already baked into the base image's virtual environment.
-# nemo-toolkit is intentionally NOT upgraded here: it is tightly coupled to the
-# NGC PyTorch build (nvcr.io/nvidia/nemo:25.11.01 uses a custom PyTorch 2.9.0a0
-# with CUDA 13.0 patches). Upgrading could silently replace the NGC build with a
-# PyPI wheel lacking CUDA 13.0 support and break distributed training.
-# Track via GHSA-9379-mwvr-7wxx/GHSA-hvjw-vp7g-39h5 (nemo-toolkit).
-RUN uv pip install --no-cache --python $VIRTUAL_ENV/bin/python \
-    "urllib3>=2.6.3" \
-    "black>=26.3.1" \
-    "cbor2>=5.9.0" \
-    "pyasn1>=0.6.3" \
-    "aiohttp>=3.13.3" \
-    "xgrammar>=0.1.32" \
-    "Pillow>=12.1.1" \
-    "nbconvert>=7.17.0" \
-    'nemo-toolkit>=2.6.1' \
-    'tornado>=6.5.5' \
-    "onnx>=1.21.0"
-# onnx: CVE-2026-28500 has no upstream patch yet; track for future upgrade
-
-## FINAL RUNTIME IMAGE
-FROM customizer-base AS customizer-runtime
-
-# Default environment variables
-ENV GUNICORN_WORKERS=1 \
-    PORT=8000 \
-    HAYSTACK_TELEMETRY_ENABLED=False \
-    HF_HUB_ENABLE_HF_TRANSFER=1 \
-    OTEL_PYTHON_EXCLUDED_URLS="health"
-
-ARG USERNAME=ubuntu
-ARG USER_UID=1000
-ARG USER_GID=$USER_UID
-COPY --chown=${USER_UID}:${USER_GID} --from=nmp-workspace / /app
-ENV USERNAME=${USERNAME} \
-    HOME=/home/${USERNAME}
-
-# Set up user home directory
-RUN mkdir -p /home/$USERNAME/.cache \
-    && ln -s /home/ubuntu /home/customizer \
-    && chown -R $USERNAME /home/$USERNAME/ /home/customizer
-
-# Copy built venv and configure runtime
-# If NO_OVERRIDE_REQUIREMENTS_PATH is set, add --overrides $NO_OVERRIDE_REQUIREMENTS_PATH to the uv pip install command
-# NO_OVERRIDE_REQUIREMENTS_PATH is a requirements.txt formatted file that lists packages that should not be overridden
-# when installing packages into the target $VIRTUAL_ENV
-RUN uv pip install --python $VIRTUAL_ENV/bin/python --no-cache \
-    ${NO_OVERRIDE_REQUIREMENTS_PATH:+--overrides $NO_OVERRIDE_REQUIREMENTS_PATH} \
-    -e /app/services/customizer
-
-# CVE fix: python-multipart and black for all customizer images.
-RUN uv pip install --python $VIRTUAL_ENV/bin/python --upgrade setuptools && \
-    uv pip install --python $VIRTUAL_ENV/bin/python --no-cache \
-    "python-multipart>=0.0.22" \
-    "black>=26.3.1"
-
-COPY docker/scripts/cve-cleanup.sh /bin/
-RUN bash /bin/cve-cleanup.sh
-
-ENV LD_LIBRARY_PATH=/opt/amazon-efa-ofi/ofi/lib:/opt/amazon-efa-ofi/efa/lib:/opt/amazon/aws-ofi-nccl/lib:${LD_LIBRARY_PATH}
-
-USER ${USER_UID}:${USER_GID}
-
-# =============================================================================
-# Smoke test: verify critical imports resolve correctly.
-# Catches two failure classes without requiring GPU hardware:
-#   - ModuleNotFoundError  : package missing from the image
-#   - ImportError          : CUDA extension .so has an undefined symbol
-#                            (ABI mismatch). Surfaces at .so load time before
-#                            any GPU device is needed.
-# SMOKE_MARKER selects which test suite runs.
-# Built as part of the docker-customizer bake group but never pushed.
-# =============================================================================
-FROM customizer-runtime AS smoke-test
-ARG SMOKE_MARKER
-USER 0
-COPY tests/smoke_gpu/ /smoke_test/
-RUN uv pip install --python $VIRTUAL_ENV/bin/python --no-cache --reinstall pytest && \
-    $VIRTUAL_ENV/bin/pytest /smoke_test/ -m ${SMOKE_MARKER} -v
diff --git a/docker/Dockerfile.nmp-customizer-tasks b/docker/Dockerfile.nmp-customizer-tasks
deleted file mode 100644
index 95ca16cc7f1..00000000000
--- a/docker/Dockerfile.nmp-customizer-tasks
+++ /dev/null
@@ -1,141 +0,0 @@
-# syntax=docker/dockerfile:1
-# NMP Customizer Tasks - GPU Tasks for Customizer
-# Requirements:
-# - nemo-platform
-# - python 3.11
-# - mamba-ssm 2.3.0 (compatible with nemotron v3)
-# - causal-conv1d
-
-# =============================================================================
-# Base Image
-# =============================================================================
-# Supplied by bake. Local builds use the base builder target; CI can override
-# the context to a pinned docker-image:// reference.
-FROM nmp-customizer-tasks-base AS customizer-tasks-base
-
-# =============================================================================
-# FAISS wheel source (ARM64 only)
-# =============================================================================
-# faiss-gpu-cu12 has no ARM64 wheel on PyPI. The bake file supplies this
-# context from a wheel image or from the local faiss-gpu-wheel target.
-FROM faiss-wheel-image AS faiss-wheel-src
-
-# =============================================================================
-# Wheel sources (causal-conv1d + mamba-ssm)
-# =============================================================================
-# Pre-built wheels from the gpu-wheels pipeline.
-# Both packages only ship source distributions on PyPI and require nvcc to
-# compile. Building them inline is slow; instead consume pre-built wheels
-# published by the gpu-wheels CI pipeline.
-FROM causal-conv1d-wheel-image AS causal-conv1d-wheel-src
-FROM mamba-ssm-wheel-image AS mamba-ssm-wheel-src
-
-# =============================================================================
-# Builder: Install all dependencies
-# =============================================================================
-FROM nmp-gpu-base AS nmp-customizer-tasks-base-builder
-COPY --from=nmp-workspace . .
-
-# ensuring we build using the same base as the runtime image, and with the currently configured dependencies
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv sync --frozen --group gpu-tasks --extra cu128 --no-editable
-
-# Install locally-built FAISS wheel on ARM64.
-# faiss-gpu-cu12 is excluded by platform marker on aarch64; the equivalent
-# FAISS GPU wheel is provided by the faiss-wheel-src context.
-RUN --mount=type=cache,target=/root/.cache/uv \
-    --mount=from=faiss-wheel-src,target=/tmp/faiss-wheel-src \
-    if ls /tmp/faiss-wheel-src/wheels/*.whl >/dev/null 2>&1; then \
-    uv pip install --no-deps /tmp/faiss-wheel-src/wheels/*.whl; \
-    fi
-
-# Install pre-built CUDA extension wheels (causal-conv1d, mamba-ssm)
-# Wheels are built by the gpu-wheels pipeline. Update WHEELS_TAG in docker-bake.hcl
-# when new wheels are published.
-# Install mamba_ssm with --no-deps to avoid replacing the pinned torch/CUDA stack,
-# then add its lightweight Python runtime dependency explicitly.
-RUN --mount=type=cache,target=/root/.cache/uv \
-    --mount=from=causal-conv1d-wheel-src,target=/tmp/causal-conv1d-wheel-src,readonly \
-    --mount=from=mamba-ssm-wheel-src,target=/tmp/mamba-ssm-wheel-src,readonly \
-    uv pip install --no-deps \
-    /tmp/causal-conv1d-wheel-src/wheels/causal_conv1d-*cp311*.whl \
-    /tmp/mamba-ssm-wheel-src/wheels/mamba_ssm-2.3.0-cp311*.whl && \
-    uv pip install --no-deps "einops>=0.8,<0.9"
-
-# =============================================================================
-# Runtime: Copy packages in separate layers
-# =============================================================================
-FROM nmp-gpu-base AS runtime
-WORKDIR /app
-
-# Layer 1: NVIDIA CUDA libraries (~4.3GB)
-COPY --from=customizer-tasks-base /app/.venv/lib/python3.11/site-packages/nvidia \
-    /app/.venv/lib/python3.11/site-packages/nvidia
-
-# Layer 2: PyTorch (~2GB)
-COPY --from=customizer-tasks-base /app/.venv/lib/python3.11/site-packages/torch \
-    /app/.venv/lib/python3.11/site-packages/torch
-COPY --from=customizer-tasks-base /app/.venv/lib/python3.11/site-packages/torchgen \
-    /app/.venv/lib/python3.11/site-packages/torchgen
-COPY --from=customizer-tasks-base /app/.venv/lib/python3.11/site-packages/functorch \
-    /app/.venv/lib/python3.11/site-packages/functorch
-COPY --from=customizer-tasks-base /app/.venv/lib/python3.11/site-packages/torch-*.dist-info \
-    /app/.venv/lib/python3.11/site-packages/
-
-# Layer 3: Inference engines (~3.4GB)
-COPY --from=customizer-tasks-base /app/.venv/lib/python3.11/site-packages/flashinfer_cubin \
-    /app/.venv/lib/python3.11/site-packages/flashinfer_cubin
-COPY --from=customizer-tasks-base /app/.venv/lib/python3.11/site-packages/triton \
-    /app/.venv/lib/python3.11/site-packages/triton
-
-#TODO needs to be built for arm64
-COPY --from=customizer-tasks-base /app/.venv/lib/python3.11/site-packages/xformers \
-    /app/.venv/lib/python3.11/site-packages/xformers
-
-# Layer 4: CUDA Extensions (causal-conv1d, mamba-ssm)
-COPY --from=nmp-customizer-tasks-base-builder /app/.venv/lib/python3.11/site-packages/causal_conv1d \
-    /app/.venv/lib/python3.11/site-packages/causal_conv1d
-COPY --from=nmp-customizer-tasks-base-builder /app/.venv/lib/python3.11/site-packages/mamba_ssm \
-    /app/.venv/lib/python3.11/site-packages/mamba_ssm
-
-# Layer 5: Everything else (venv structure, bin/, dist-info, other packages)
-RUN --mount=from=nmp-customizer-tasks-base-builder,source=/app/.venv,target=/mnt/venv,readonly \
-    cd /mnt/venv && \
-    tar --exclude='lib/python3.11/site-packages/nvidia' \
-    --exclude='lib/python3.11/site-packages/torch' \
-    --exclude='lib/python3.11/site-packages/torchgen' \
-    --exclude='lib/python3.11/site-packages/functorch' \
-    --exclude='lib/python3.11/site-packages/torch-*.dist-info' \
-    --exclude='lib/python3.11/site-packages/flashinfer_cubin' \
-    --exclude='lib/python3.11/site-packages/vllm' \
-    --exclude='lib/python3.11/site-packages/triton' \
-    --exclude='lib/python3.11/site-packages/xformers' \
-    -cf - . | tar -xf - -C /app/.venv
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install "transformers>=5.0.0,<5.1.0"
-
-# setuptools: GHSA-58pv-8j8x-9vj2
-RUN uv pip install --upgrade setuptools
-
-COPY docker/scripts/cve-cleanup.sh /bin/
-RUN bash /bin/cve-cleanup.sh
-
-ENV PATH="/app/.venv/bin:$PATH"
-ENTRYPOINT ["nemo-platform"]
-# Override CMD with "run" in orchestration (e.g., Helm, docker-compose)
-CMD ["--help"]
-
-# =============================================================================
-# Smoke test: verify all customizer task imports resolve correctly.
-# Catches two failure classes without requiring GPU hardware:
-#   - ModuleNotFoundError  : package missing from the image
-#   - ImportError          : CUDA extension .so has an undefined symbol
-#                            (ABI mismatch). Surfaces at .so load time before
-#                            any GPU device is needed.
-# Built as part of the docker-customizer bake group but never pushed.
-# =============================================================================
-FROM runtime AS smoke-test
-COPY tests/smoke_gpu/ /smoke_test/
-RUN uv pip install --no-cache --reinstall pytest && \
-    pytest /smoke_test/ -m smoke_customizer_tasks -v
diff --git a/docker/README.md b/docker/README.md
index 11c1bda906b..203f62cc651 100644
--- a/docker/README.md
+++ b/docker/README.md
@@ -8,5 +8,5 @@ Examples:
 ```bash
 docker buildx bake --print docker-cpu
 docker buildx bake --print nmp-automodel
-docker buildx bake --print docker-customizer
+docker buildx bake --print nmp-unsloth
 ```
diff --git a/docker/base/Dockerfile.faiss-gpu-wheel b/docker/base/Dockerfile.faiss-gpu-wheel
deleted file mode 100644
index 2a3e1e8b380..00000000000
--- a/docker/base/Dockerfile.faiss-gpu-wheel
+++ /dev/null
@@ -1,95 +0,0 @@
-# syntax=docker/dockerfile:1
-#######
-# FAISS GPU Wheel Builder
-#
-# Builds the FAISS Python wheel with CUDA support.
-# The resulting image stores the wheel at /wheels/*.whl.
-#
-# Downstream images can install the wheel with:
-#   COPY --from=faiss-gpu-wheel /wheels/*.whl /tmp/wheels/
-#   RUN pip install /tmp/wheels/*.whl
-#
-# Build args:
-#   FAISS_VERSION  - git tag/branch/SHA to clone  (default: v1.9.0)
-#   CUDA_ARCHS     - semicolon-separated SM targets (default: 75;80;86;90)
-#   CUDA_VERSION   - CUDA devel image version       (default: 12.8.1)
-#
-#   CUDA_ARCHS reference:
-#     75 = Turing  (T4, RTX 2080)
-#     80 = Ampere  (A100)
-#     86 = Ampere  (A10G, RTX 3080/3090)
-#     87 = Ampere  (Jetson AGX Orin)
-#     89 = Ada     (RTX 4090)
-#     90 = Hopper  (H100)
-#######
-
-ARG CUDA_VERSION=12.8.1
-
-FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS faiss-gpu-wheel-builder
-
-ARG FAISS_VERSION=v1.9.0
-ARG CUDA_ARCHS="75;80;86;90"
-
-ENV DEBIAN_FRONTEND=noninteractive
-
-# Install build dependencies.
-# python3.11 + python3.11-dev are available in ubuntu 22.04 universe.
-RUN apt-get update && apt-get install -y --no-install-recommends \
-    build-essential \
-    ca-certificates \
-    git \
-    libgflags-dev \
-    libopenblas-dev \
-    python3.11 \
-    python3.11-dev \
-    python3.11-venv \
-    swig \
-    && rm -rf /var/lib/apt/lists/*
-
-# Install uv and create a Python 3.11 venv with build-time deps.
-# Ubuntu/Debian disables ensurepip for the system Python; uv manages its
-# own Python toolchain and doesn't depend on pip or ensurepip.
-# cmake is installed via PyPI to get a version >= 3.24 (Ubuntu 22.04 ships 3.22).
-COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
-
-RUN uv venv --python 3.11 /opt/venv \
-    && uv pip install --python /opt/venv/bin/python3 \
-        "cmake>=3.24" ninja numpy packaging setuptools wheel
-
-ENV PATH="/opt/venv/bin:$PATH"
-
-# Clone FAISS at the requested version.
-RUN git clone --depth 1 --branch ${FAISS_VERSION} \
-    https://github.com/facebookresearch/faiss.git /faiss/src
-
-# Configure.
-RUN cmake -B /faiss/build \
-    -DFAISS_ENABLE_GPU=ON \
-    -DFAISS_ENABLE_RAFT=OFF \
-    -DFAISS_ENABLE_PYTHON=ON \
-    -DFAISS_ENABLE_MKL=OFF \
-    -DFAISS_ENABLE_EXTRAS=OFF \
-    -DBUILD_TESTING=OFF \
-    -DBLA_VENDOR=OpenBLAS \
-    -DCMAKE_BUILD_TYPE=Release \
-    -DCMAKE_CUDA_ARCHITECTURES="${CUDA_ARCHS}" \
-    -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/g++ \
-    -DPython_EXECUTABLE=/opt/venv/bin/python3 \
-    /faiss/src
-
-# Build CPU library, GPU library, and Python bindings.
-RUN cmake --build /faiss/build --target faiss faiss_gpu swigfaiss \
-    --parallel $(nproc)
-
-# Package into a wheel.
-RUN cd /faiss/build/faiss/python \
-    && python setup.py bdist_wheel \
-    && mkdir -p /wheels \
-    && cp dist/*.whl /wheels/
-
-RUN ls -lh /wheels/
-
-CMD ["ls", "-lh", "/wheels/"]
-
-FROM scratch AS faiss-gpu-wheel
-COPY --from=faiss-gpu-wheel-builder /wheels /wheels
diff --git a/docker/base/Dockerfile.mamba-wheel b/docker/base/Dockerfile.mamba-wheel
index 06a0754c74b..8c9e7198d14 100644
--- a/docker/base/Dockerfile.mamba-wheel
+++ b/docker/base/Dockerfile.mamba-wheel
@@ -9,11 +9,11 @@
 # Both only ship source distributions on PyPI and require nvcc to compile.
 # The two builds are independent stages so BuildKit runs them in parallel.
 # Each image stores its wheel at /wheels/*.whl.
-# Build via the gpu-wheels or nmp-automodel-gpu-wheels bake group.
+# Build via the nmp-automodel-gpu-wheels bake group.
 #
 # Build args:
 #   CAUSAL_CONV1D_VERSION  - git tag to build              (default: v1.5.3)
-#   MAMBA_22_COMMIT           - git SHA or tag to build
+#   MAMBA_22_COMMIT        - git SHA or tag to build
 #   CUDA_VERSION           - CUDA devel image version      (default: 12.8.1)
 #   TORCH_CUDA_ARCH_LIST   - semicolon-separated SM targets (default: "8.0;8.6;9.0")
 #######
@@ -97,7 +97,7 @@ COPY --from=ghcr.io/astral-sh/uv:0.9.14 /uv /usr/local/bin/uv
 
 
 # =============================================================================
-# causal-conv1d wheel — Python 3.11 (for safe-synthesizer-tasks, nmp-customizer-tasks, nmp-automodel-tasks, and nmp-automodel-tasks)
+# causal-conv1d wheel - Python 3.11
 # =============================================================================
 FROM mamba-wheel-base AS causal-conv1d-wheel-builder
 
@@ -139,7 +139,7 @@ RUN mkdir -p /wheels && \
     rm -rf /src/causal-conv1d
 
 # The final causal-conv1d-wheel image contains:
-#   - causal_conv1d-*-cp311-*.whl  (for Python 3.11 consumers: safe-synthesizer-tasks, nmp-customizer-tasks, nmp-automodel-tasks)
+#   - causal_conv1d-*-cp311-*.whl  (for Python 3.11 consumers)
 #   - causal_conv1d-*-cp312-*.whl  (for Python 3.12 consumers)
 # Consumers must pin the Python tag glob (e.g. causal_conv1d-*cp311*.whl) to select the right one.
 FROM scratch AS causal-conv1d-wheel
@@ -181,22 +181,6 @@ RUN mkdir -p /wheels && \
     uv build --wheel --no-build-isolation --out-dir=/wheels . && \
     rm -rf /src/mamba
 
-# =============================================================================
-# mamba-ssm 2.3.0 wheel — Python 3.11 (for nmp-customizer-tasks)
-# =============================================================================
-FROM mamba-wheel-base AS mamba-ssm-23-wheel-builder
-
-ARG MAMBA_23_COMMIT=v2.3.0
-
-RUN mkdir -p /wheels && \
-    git clone https://github.com/state-spaces/mamba.git /src/mamba && \
-    cd /src/mamba && \
-    git checkout ${MAMBA_23_COMMIT} && \
-    sed -i "/triton/d" setup.py && \
-    sed -i "/triton/d" pyproject.toml && \
-    uv build --wheel --no-build-isolation --out-dir=/wheels . && \
-    rm -rf /src/mamba
-
 # =============================================================================
 # mamba-ssm 2.3.0 wheel — Python 3.12 (for Python 3.12 consumers)
 # =============================================================================
@@ -231,15 +215,14 @@ RUN mkdir -p /wheels && \
     rm -rf /src/mamba
 
 
-# The final mamba-ssm-wheel image contains four versions:
-#   - mamba_ssm-2.2.5-cp311-*.whl  (from MAMBA_22_COMMIT=6b32be06, for safe-synthesizer-tasks / Python 3.11)
-#   - mamba_ssm-2.2.5-cp312-*.whl  (from MAMBA_22_COMMIT=6b32be06, for Python 3.12 consumers, e.g. automodel)
-#   - mamba_ssm-2.3.0-cp311-*.whl  (from v2.3.0, for nmp-customizer-tasks / Python 3.11)
+# The final mamba-ssm-wheel image contains:
+#   - mamba_ssm-2.2.5-cp311-*.whl  (from MAMBA_22_COMMIT=6b32be06, for Python 3.11 consumers)
+#   - mamba_ssm-2.2.5-cp312-*.whl  (from MAMBA_22_COMMIT=6b32be06, for Python 3.12 consumers)
 #   - mamba_ssm-2.3.0-cp312-*.whl  (from v2.3.0, for Python 3.12 consumers)
+#   - cu13.1.1/mamba_ssm-2.3.0-cp312-*.whl  (from v2.3.0, for Automodel on CUDA 13.1.1)
 # Consumers must pin both version AND Python tag glob to select the correct wheel.
 FROM scratch AS mamba-ssm-wheel
 COPY --from=mamba-ssm-wheel-builder /wheels /wheels
 COPY --from=mamba-ssm-25-wheel-builder-py312 /wheels /wheels
-COPY --from=mamba-ssm-23-wheel-builder /wheels /wheels
 COPY --from=mamba-ssm-23-wheel-builder-py312 /wheels /wheels
 COPY --from=mamba-ssm-23-wheel-builder-py312-cu13.1.1 /wheels /wheels/cu13.1.1
diff --git a/docker/base/Dockerfile.nmp-gpu-base-py312 b/docker/base/Dockerfile.nmp-gpu-base-py312
index c6ef88a91ff..ae223278af2 100644
--- a/docker/base/Dockerfile.nmp-gpu-base-py312
+++ b/docker/base/Dockerfile.nmp-gpu-base-py312
@@ -2,7 +2,6 @@
 #######
 # NMP GPU Base Image (Python 3.12)
 # Base for GPU task containers that require Python 3.12
-# (e.g. nmp-customizer-tasks)
 #######
 
 ARG CUDA_VERSION=12.8.1