diff --git a/Dockerfile b/Dockerfile
index 8d0805b..e8e8bcb 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -223,7 +223,8 @@ RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
 # Copy artifacts from Builder Stage
 COPY --from=builder /workspace/wheels /workspace/wheels
 RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
-    uv pip install /workspace/wheels/*.whl
+    uv pip install /workspace/wheels/*.whl && \
+    rm -rf /workspace/wheels
 
 # Setup Env for Runtime
 ENV TORCH_CUDA_ARCH_LIST=12.1a
@@ -239,3 +240,18 @@ RUN chmod +x $VLLM_BASE_DIR/run-cluster-node.sh
 # Final extra deps
 RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
     uv pip install ray[default] fastsafetensors "apache-tvm-ffi<0.2"
+
+# Cleanup
+# RUN uv pip uninstall absl-py apex argon2-cffi \
+#     argon2-cffi-bindings arrow asttokens astunparse async-lru audioread babel beautifulsoup4 \
+#     black bleach comm contourpy cycler datasets debugpy decorator defusedxml dllist dm-tree \
+#     execnet executing expecttest fastjsonschema fonttools fqdn gast hypothesis \
+#     ipykernel ipython ipython_pygments_lexers isoduration isort jedi joblib jupyter-events \
+#     jupyter-lsp jupyter_client jupyter_core jupyter_server jupyter_server_terminals jupyterlab \
+#     jupyterlab_code_formatter jupyterlab_code_formatter jupyterlab_pygments jupyterlab_server \
+#     jupyterlab_tensorboard_pro jupytext kiwisolver matplotlib matplotlib-inline matplotlib-inline \
+#     mistune ml_dtypes mock nbclient nbconvert nbformat nest-asyncio notebook notebook_shim \
+#     opt_einsum optree outlines_core overrides pandas pandocfilters parso pexpect polygraphy pooch \
+#     pyarrow pycocotools pytest-flakefinder pytest-rerunfailures pytest-shard pytest-xdist \
+#     scikit-learn scipy Send2Trash soundfile soupsieve soxr spin stack-data \
+#     wcwidth webcolors xdoctest Werkzeug
\ No newline at end of file
diff --git a/Dockerfile.mxfp4 b/Dockerfile.mxfp4
index 11af230..1902e32 100644
--- a/Dockerfile.mxfp4
+++ b/Dockerfile.mxfp4
@@ -258,7 +258,8 @@ RUN mkdir -p tiktoken_encodings && \
 # No need to copy source code, as it's already in the site-packages
 COPY --from=builder /workspace/wheels /workspace/wheels
 RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
-    uv pip install /workspace/wheels/*.whl
+    uv pip install /workspace/wheels/*.whl && \
+    rm -rf /workspace/wheels
 
 # Setup Env for Runtime
 ENV TORCH_CUDA_ARCH_LIST="12.0;12.1"
@@ -274,3 +275,4 @@ RUN chmod +x $VLLM_BASE_DIR/run-cluster-node.sh
 # Final extra deps
 RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
     uv pip install ray[default] fastsafetensors
+
diff --git a/Dockerfile.wheels b/Dockerfile.wheels
index 3e42d47..8311661 100644
--- a/Dockerfile.wheels
+++ b/Dockerfile.wheels
@@ -1,6 +1,9 @@
 # syntax=docker/dockerfile:1.6
 
-FROM nvcr.io/nvidia/pytorch:26.01-py3
+# =========================================================
+# STAGE 1: Base Image (Installs Dependencies)
+# =========================================================
+FROM nvcr.io/nvidia/pytorch:26.01-py3 AS base
 
 ENV DEBIAN_FRONTEND=noninteractive
 ENV PIP_BREAK_SYSTEM_PACKAGES=1
@@ -42,7 +45,7 @@ COPY fastsafetensors.patch .
 
 # Install fastsafetensors
 RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
-    uv pip install -U fastsafetensors
+    uv pip install -U fastsafetensors ray[default] "apache-tvm-ffi<0.2"
 
 # --- VLLM SOURCE CACHE BUSTER ---
 # Change THIS argument to force a fresh git clone and rebuild of vLLM
@@ -92,7 +95,17 @@ ENV TIKTOKEN_ENCODINGS_BASE=$VLLM_BASE_DIR/tiktoken_encodings
 COPY run-cluster-node.sh $VLLM_BASE_DIR/
 RUN chmod +x $VLLM_BASE_DIR/run-cluster-node.sh
 
-# Final extra deps
-RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
-    uv pip install ray[default] fastsafetensors "apache-tvm-ffi<0.2"
-
+# Cleanup unneeded packages to reduce image size
+# RUN uv pip uninstall absl-py apex argon2-cffi \
+#     argon2-cffi-bindings arrow asttokens astunparse async-lru audioread babel beautifulsoup4 \
+#     black bleach comm contourpy cycler datasets debugpy decorator defusedxml dllist dm-tree \
+#     execnet executing expecttest fastjsonschema fonttools fqdn gast hypothesis \
+#     ipykernel ipython ipython_pygments_lexers isoduration isort jedi joblib jupyter-events \
+#     jupyter-lsp jupyter_client jupyter_core jupyter_server jupyter_server_terminals jupyterlab \
+#     jupyterlab_code_formatter jupyterlab_code_formatter jupyterlab_pygments jupyterlab_server \
+#     jupyterlab_tensorboard_pro jupytext kiwisolver matplotlib matplotlib-inline matplotlib-inline \
+#     mistune ml_dtypes mock nbclient nbconvert nbformat nest-asyncio notebook notebook_shim \
+#     opt_einsum optree outlines_core overrides pandas pandocfilters parso pexpect polygraphy pooch \
+#     pyarrow pycocotools pytest-flakefinder pytest-rerunfailures pytest-shard pytest-xdist \
+#     scikit-learn scipy Send2Trash soundfile soupsieve soxr spin stack-data \
+#     wcwidth webcolors xdoctest Werkzeug
\ No newline at end of file