serverless/Dockerfile

# Reproducible Dockerfile equivalent of build.sh (which assembles the same
# image daemonlessly with crane, since RunPod pods have no Docker daemon).
#
# Build from the HF-repo root (context needs serverless/ and apply_patches.sh),
# passing your HF token as a BuildKit secret so it is never baked into a layer:
#
#   DOCKER_BUILDKIT=1 docker build \
#     -f serverless/Dockerfile \
#     --secret id=hf_token,env=HF_TOKEN \
#     --build-arg ENDPOINT=qwen-edit-turbo \
#     -t plx1029/comfyui-serverless:qwen-edit-turbo-v7 .
#
# Layer order mirrors build.sh: custom nodes + models first (heavy, rarely
# change), serverless code last (tiny, changes often) — so code iterations
# never re-push model layers.

FROM plx1029/comfyui-qwen:v6

ARG ENDPOINT=qwen-edit-turbo
ARG CUSTOM_SCRIPTS_COMMIT=main
ARG QWENEDITUTILS_COMMIT=main

# ---- custom nodes (rgthree already ships in the base) ----
RUN git clone https://github.com/pythongosssss/ComfyUI-Custom-Scripts.git \
        /workspace/ComfyUI/custom_nodes/ComfyUI-Custom-Scripts && \
    git -C /workspace/ComfyUI/custom_nodes/ComfyUI-Custom-Scripts checkout ${CUSTOM_SCRIPTS_COMMIT} && \
    git clone https://github.com/lrzjason/Comfyui-QwenEditUtils.git \
        /workspace/ComfyUI/custom_nodes/Comfyui-QwenEditUtils && \
    git -C /workspace/ComfyUI/custom_nodes/Comfyui-QwenEditUtils checkout ${QWENEDITUTILS_COMMIT}

# ---- models: one RUN (= one layer) per file, list must match endpoint.json ----
COPY serverless/endpoints/${ENDPOINT}/endpoint.json /tmp/endpoint.json
RUN pip install -q -U huggingface_hub
# hf download reads the token from the mounted secret; one layer per model
RUN --mount=type=secret,id=hf_token HF_TOKEN=$(cat /run/secrets/hf_token) \
    hf download aleph65/ComfyUI models/diffusion_models/qwen_image_edit_2511_bf16.safetensors --local-dir /workspace/ComfyUI
RUN --mount=type=secret,id=hf_token HF_TOKEN=$(cat /run/secrets/hf_token) \
    hf download aleph65/ComfyUI models/text_encoders/qwen_2.5_vl_7b.safetensors --local-dir /workspace/ComfyUI
RUN --mount=type=secret,id=hf_token HF_TOKEN=$(cat /run/secrets/hf_token) \
    hf download aleph65/ComfyUI models/vae/qwen_image_vae.safetensors --local-dir /workspace/ComfyUI
RUN --mount=type=secret,id=hf_token HF_TOKEN=$(cat /run/secrets/hf_token) \
    hf download aleph65/ComfyUI models/loras/Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors --local-dir /workspace/ComfyUI
RUN --mount=type=secret,id=hf_token HF_TOKEN=$(cat /run/secrets/hf_token) \
    hf download aleph65/ComfyUI models/loras/Qwen-Image-Edit-2511-Lightning-8steps-V1.0-bf16.safetensors --local-dir /workspace/ComfyUI
RUN --mount=type=secret,id=hf_token HF_TOKEN=$(cat /run/secrets/hf_token) \
    hf download aleph65/ComfyUI models/loras/Qwen_LoRA_Skin_Fix_v2.safetensors --local-dir /workspace/ComfyUI
RUN --mount=type=secret,id=hf_token HF_TOKEN=$(cat /run/secrets/hf_token) \
    hf download aleph65/ComfyUI models/loras/Qwen4Play_v2.safetensors --local-dir /workspace/ComfyUI

# ---- serverless code (top layer: cheap to rebuild) ----
COPY apply_patches.sh /workspace/apply_patches.sh
RUN bash /workspace/apply_patches.sh
COPY serverless/ /opt/serverless/src/
RUN rm -rf /opt/serverless/src/Dockerfile /opt/serverless/src/build.sh && \
    find /opt/serverless/src/endpoints -mindepth 1 -maxdepth 1 ! -name "${ENDPOINT}" -exec rm -rf {} + && \
    pip install -q --target /opt/serverless/lib runpod && \
    chmod +x /opt/serverless/src/start.sh

CMD ["/opt/serverless/src/start.sh"]