diff --git a/.github/workflows/llm-finetuner.yml b/.github/workflows/llm-finetuner.yml
new file mode 100644
index 0000000..6e42056
--- /dev/null
+++ b/.github/workflows/llm-finetuner.yml
@@ -0,0 +1,15 @@
+on:
+  workflow_dispatch:
+    inputs:
+      commit:
+        description: 'Commit to build'
+        required: true
+
+
+jobs:
+  build:
+    uses: ./.github/workflows/build.yml
+    with: 
+      image-name: llm-finetuner
+      folder: llm-finetuner
+      build-args: "COMMIT=${{ inputs.commit }}"
\ No newline at end of file
diff --git a/llm-finetuner/Dockerfile b/llm-finetuner/Dockerfile
new file mode 100644
index 0000000..c41708d
--- /dev/null
+++ b/llm-finetuner/Dockerfile
@@ -0,0 +1,82 @@
+# syntax=docker/dockerfile:1.2
+
+ARG BASE_IMAGE=ghcr.io/coreweave/ml-containers/torch:afecfe9-base-cuda11.8.0-torch2.0.0-vision0.15.1
+
+FROM alpine/git:2.36.3 as downloader
+WORKDIR /git
+
+# Note: avoid using a branch name as the commit to prevent erroneous caching.
+ARG COMMIT
+
+# Download only the required files from the repository
+RUN git clone --filter=blob:none --depth 1 --no-single-branch --no-checkout \
+      https://github.com/coreweave/kubernetes-cloud && \
+    cd kubernetes-cloud && \
+    git sparse-checkout init && \
+    git config --worktree core.sparseCheckoutCone false && \
+    git config --worktree advice.detachedHead false && \
+    git sparse-checkout set /finetuner-workflow/finetuner && \
+    git checkout "${COMMIT}" && \
+    rm -rf .git && \
+    [ -d finetuner-workflow/finetuner ] && \
+    ln -s $(realpath finetuner-workflow/finetuner) /src
+
+WORKDIR /src
+
+# Dependencies requiring NVCC are built ahead of time in a separate stage
+# so that the ~2 GiB dev library installations don't have to be included
+# in the final finetuner image.
+FROM ${BASE_IMAGE} as builder
+RUN apt-get install -y --no-install-recommends \
+        cuda-nvcc-11-8 cuda-nvml-dev-11-8 libcurand-dev-11-8 \
+        libcublas-dev-11-8 libcusparse-dev-11-8 \
+        libcusolver-dev-11-8 cuda-nvprof-11-8 \
+        cuda-profiler-api-11-8 \
+        ninja-build \
+        # gcc-10/g++-10/lld do not need to be installed here, but they improve the build.
+        # gfortran-10 is just for compiler_wrapper.f95.
+        gcc-10 g++-10 gfortran-10 lld && \
+    apt-get clean && \
+    update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 10 && \
+    update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 10 && \
+    update-alternatives --install \
+      /usr/bin/gfortran gfortran /usr/bin/gfortran-10 10 && \
+    update-alternatives --install /usr/bin/ld ld /usr/bin/ld.lld 1
+
+RUN mkdir /wheels /build
+WORKDIR /build
+COPY compiler_wrapper.f95 .
+COPY --from=downloader /src/requirements-precompilable.txt .
+RUN gfortran -O3 ./compiler_wrapper.f95 -o ./compiler && \
+    python3 -m pip install -U --no-cache-dir \
+      packaging setuptools wheel pip && \
+    # Only DS_BUILT_UTILS and DS_BUILD_CPU_ADAM are necessary for the finetuner
+    # See: https://www.deepspeed.ai/tutorials/advanced-install
+    DS_BUILD_UTILS=1 DS_BUILD_CPU_ADAM=1 \
+      # DeepSpeed forces -march=native into the compiler options,
+      # making the result dependent on the processor architecture
+      # used on the builder machine.
+      # The compiler wrapper normalizes -march=native to -march=skylake
+      # along with a couple other transformations before invoking GCC.
+      CC=$(realpath -e ./compiler) \
+      python3 -m pip wheel -w /wheels \
+      --no-cache-dir --no-build-isolation --no-deps \
+      -r requirements-precompilable.txt && \
+    rm ./*
+
+WORKDIR /wheels
+
+FROM ${BASE_IMAGE}
+RUN mkdir /app
+WORKDIR /app
+RUN --mount=type=bind,from=builder,source=/wheels,target=. \
+    pip3 install --no-cache-dir ./*.whl
+COPY --from=downloader /src/requirements.txt .
+COPY --from=downloader /src/requirements-precompilable.txt .
+RUN pip3 install --no-cache-dir -r requirements.txt
+COPY --from=downloader /src/ds_config.json .
+COPY --from=downloader /src/finetuner.py .
+COPY --from=downloader /src/evaluator.py .
+COPY --from=downloader /src/inference.py .
+COPY --from=downloader /src/utils.py .
+CMD [ "/usr/bin/python3", "finetuner.py" ]
diff --git a/llm-finetuner/compiler_wrapper.f95 b/llm-finetuner/compiler_wrapper.f95
new file mode 100644
index 0000000..f8c13bd
--- /dev/null
+++ b/llm-finetuner/compiler_wrapper.f95
@@ -0,0 +1,76 @@
+PROGRAM compiler_wrapper
+    ! Wraps GCC invocations,
+    ! replacing -D__AVX512__ and -D__SCALAR__ preprocessor definitions
+    ! with -D__AVX256__, and -march=native with -march=skylake,
+    ! for better reproducibility and compatibility.
+    IMPLICIT NONE
+    INTEGER :: i, exitcode = 0, full_length = 0, truncated = 0
+    CHARACTER(len=:), ALLOCATABLE :: arg, command
+    ALLOCATE(CHARACTER(len=128) :: arg)
+    command = "gcc"
+
+    DO i = 1, COMMAND_ARGUMENT_COUNT()
+        DO
+            CALL GET_COMMAND_ARGUMENT(i, arg, full_length, truncated)
+            IF (truncated == 0) THEN
+                EXIT
+            ELSE IF (truncated == -1) THEN
+                DEALLOCATE(arg)
+                ALLOCATE(CHARACTER(len=full_length) :: arg)
+            ELSE
+                CALL EXIT(95)
+            END IF
+        END DO
+        IF (arg == "-march=native") THEN
+            command = command // " '-march=skylake'"
+        ELSE IF (arg == "-D__AVX512__" .OR. arg == "-D__SCALAR__") THEN
+            command = command // " '-D__AVX256__'"
+        ELSE
+            command = command // shell_escaped(arg)
+        END IF
+    END DO
+    CALL SYSTEM(command, exitcode)
+    IF (exitcode > 255) THEN
+        exitcode = MAX(IAND(exitcode, 255), 1)
+    END IF
+    CALL EXIT(exitcode)
+
+
+    CONTAINS
+        FUNCTION shell_escaped(str) RESULT(out)
+            ! Turns [str] into [ 'str'] and replaces all
+            ! internal ['] characters with ['"'"']
+            IMPLICIT NONE
+            CHARACTER(len=*), INTENT(IN) :: str
+            CHARACTER(len=:), ALLOCATABLE :: out
+            INTEGER :: old_i, out_i, old_len, out_len
+
+            old_len = LEN_TRIM(str)
+            ! Figure out the new length to allocate by scanning `str`.
+            ! This always needs to add at least [ '] at the beginning
+            ! and ['] at the end, so the length increases by at least 3.
+            out_len = old_len + 3
+            DO old_i = 1, old_len
+                IF (str(old_i:old_i) == "'") THEN
+                    out_len = out_len + 4
+                END IF
+            END DO
+            ALLOCATE(CHARACTER(len=out_len) :: out)
+
+            ! Copy over the string, performing necessary escapes.
+            out(1:2) = " '"
+            out_i = 3
+            DO old_i = 1, old_len
+                IF (str(old_i:old_i) == "'") THEN
+                    ! Escape internal single-quotes
+                    out(out_i:out_i + 4) = '''"''"'''
+                    out_i = out_i + 5
+                ELSE
+                    ! No escaping needed
+                    out(out_i:out_i) = str(old_i:old_i)
+                    out_i = out_i + 1
+                END IF
+            END DO
+            out(out_i:out_i) = "'"
+        END FUNCTION
+END PROGRAM