Spaces:

msradam
/

riprap

Running

riprap / Dockerfile

deploy(l4): self-contained Riprap mirror

3dbff85 6 months ago

4.16 kB

# Riprap — Hugging Face Spaces deployment for the personal Space
# (msradam/riprap-nyc) on L4 hardware.
#
# Differences from the canonical Dockerfile:
#
#   1. L4 has 24 GB VRAM (vs 16 GB on T4 small), so we co-host the
#      riprap-models service inside the same container instead of
#      proxying to the AMD MI300X droplet. No external dependency.
#
#   2. We bake granite4.1:8b at *build* time. The build sandbox could
#      not previously fit Granite + EO toolchain together; this Dockerfile
#      keeps the EO install at runtime (entrypoint.l4.sh) and frees the
#      sandbox budget for the 8B pull.
#
#   3. CUDA + ROCm-free torch — the inline riprap-models service uses
#      the cu124 wheels installed via requirements.txt + the additional
#      delta in services/riprap-models/requirements.txt.
#
# DO NOT push this image to the lablab Space — that one stays pointed
# at the MI300X droplet for AMD-judging continuity.

FROM nvidia/cuda:12.4.1-cudnn-runtime-ubuntu22.04 AS base

ENV DEBIAN_FRONTEND=noninteractive
RUN apt-get update && apt-get install -y --no-install-recommends \
        python3 python3-pip python3-venv python-is-python3 \
        curl ca-certificates zstd procps git \
        gdal-bin libgdal-dev libgeos-dev libproj-dev \
        libgl1 libglib2.0-0 \
    && rm -rf /var/lib/apt/lists/*

RUN useradd -m -u 1000 user
ENV HOME=/home/user \
    PATH=/home/user/.local/bin:/usr/local/bin:/usr/bin:/bin \
    PYTHONUNBUFFERED=1 \
    HF_HOME=/home/user/.cache/huggingface \
    OLLAMA_HOST=127.0.0.1:11434 \
    OLLAMA_NUM_PARALLEL=1 \
    OLLAMA_KEEP_ALIVE=24h \
    OLLAMA_MAX_LOADED_MODELS=2 \
    OLLAMA_FLASH_ATTENTION=1 \
    OLLAMA_KV_CACHE_TYPE=q8_0 \
    OLLAMA_DEBUG=1 \
    OLLAMA_MODELS=/home/user/.ollama/models \
    RIPRAP_OLLAMA_3B_TAG=granite4.1:8b \
    RIPRAP_LLM_PRIMARY=ollama \
    RIPRAP_LLM_BASE_URL=http://127.0.0.1:11434/v1 \
    RIPRAP_ML_BACKEND=remote \
    RIPRAP_ML_BASE_URL=http://127.0.0.1:7861

RUN curl -fsSL https://ollama.com/install.sh | sh

WORKDIR /home/user/app

# Web app deps (torch cu124 lands via sentence-transformers / etc.).
COPY --chown=user:user requirements.txt ./
RUN pip install --no-cache-dir --upgrade pip && \
    pip install --no-cache-dir -r requirements.txt

# riprap-models delta deps. Use the existing requirements.txt at the
# *service* level, but skip requirements-full.txt — its ROCm-frozen
# torch pin would clobber the cu124 wheels installed above.
COPY --chown=user:user services/riprap-models/requirements.txt /tmp/req-models.txt
RUN pip install --no-cache-dir -r /tmp/req-models.txt

# Bake torchvision (CUDA 12.4 wheel) and peft at build time. The
# canonical entrypoint.sh runtime-installs torchvision via the EO
# toolchain path because the canonical CPU Space's build sandbox is
# too tight; L4 builds have more room, and a properly matched
# torchvision avoids the `torchvision::nms does not exist` runtime
# error the canonical setup hits. peft is required by the riprap-
# models service for the TerraMind LoRA inference path.
RUN pip install --no-cache-dir \
        --index-url https://download.pytorch.org/whl/cu124 \
        torchvision \
    && pip install --no-cache-dir peft==0.18.1

# Bake Granite 4.1 weights into the image (EO toolchain is installed
# at runtime — see entrypoint.l4.sh — to keep the build sandbox under
# its disk threshold).
RUN mkdir -p $OLLAMA_MODELS && \
    ollama serve & \
    OPID=$! && \
    for i in $(seq 1 30); do curl -sf http://127.0.0.1:11434/ > /dev/null && break; sleep 1; done && \
    ollama pull granite4.1:8b && \
    kill $OPID 2>/dev/null || true && \
    sleep 2

# App code, fixtures, and inline model service.
COPY --chown=user:user app/ ./app/
COPY --chown=user:user web/ ./web/
COPY --chown=user:user scripts/ ./scripts/
COPY --chown=user:user data/ ./data/
COPY --chown=user:user corpus/ ./corpus/
COPY --chown=user:user services/riprap-models/main.py ./riprap_models.py
COPY --chown=user:user agent.py riprap.py ./
COPY --chown=user:user entrypoint.sh ./entrypoint.sh
RUN chmod +x ./entrypoint.sh

RUN chown -R user:user /home/user
USER user

EXPOSE 7860
CMD ["./entrypoint.sh"]

	# Riprap — Hugging Face Spaces deployment for the personal Space
	# (msradam/riprap-nyc) on L4 hardware.
	#
	# Differences from the canonical Dockerfile:
	#
	# 1. L4 has 24 GB VRAM (vs 16 GB on T4 small), so we co-host the
	# riprap-models service inside the same container instead of
	# proxying to the AMD MI300X droplet. No external dependency.
	#
	# 2. We bake granite4.1:8b at build time. The build sandbox could
	# not previously fit Granite + EO toolchain together; this Dockerfile
	# keeps the EO install at runtime (entrypoint.l4.sh) and frees the
	# sandbox budget for the 8B pull.
	#
	# 3. CUDA + ROCm-free torch — the inline riprap-models service uses
	# the cu124 wheels installed via requirements.txt + the additional
	# delta in services/riprap-models/requirements.txt.
	#
	# DO NOT push this image to the lablab Space — that one stays pointed
	# at the MI300X droplet for AMD-judging continuity.

	FROM nvidia/cuda:12.4.1-cudnn-runtime-ubuntu22.04 AS base

	ENV DEBIAN_FRONTEND=noninteractive
	RUN apt-get update && apt-get install -y --no-install-recommends \
	python3 python3-pip python3-venv python-is-python3 \
	curl ca-certificates zstd procps git \
	gdal-bin libgdal-dev libgeos-dev libproj-dev \
	libgl1 libglib2.0-0 \
	&& rm -rf /var/lib/apt/lists/*

	RUN useradd -m -u 1000 user
	ENV HOME=/home/user \
	PATH=/home/user/.local/bin:/usr/local/bin:/usr/bin:/bin \
	PYTHONUNBUFFERED=1 \
	HF_HOME=/home/user/.cache/huggingface \
	OLLAMA_HOST=127.0.0.1:11434 \
	OLLAMA_NUM_PARALLEL=1 \
	OLLAMA_KEEP_ALIVE=24h \
	OLLAMA_MAX_LOADED_MODELS=2 \
	OLLAMA_FLASH_ATTENTION=1 \
	OLLAMA_KV_CACHE_TYPE=q8_0 \
	OLLAMA_DEBUG=1 \
	OLLAMA_MODELS=/home/user/.ollama/models \
	RIPRAP_OLLAMA_3B_TAG=granite4.1:8b \
	RIPRAP_LLM_PRIMARY=ollama \
	RIPRAP_LLM_BASE_URL=http://127.0.0.1:11434/v1 \
	RIPRAP_ML_BACKEND=remote \
	RIPRAP_ML_BASE_URL=http://127.0.0.1:7861

	RUN curl -fsSL https://ollama.com/install.sh \| sh

	WORKDIR /home/user/app

	# Web app deps (torch cu124 lands via sentence-transformers / etc.).
	COPY --chown=user:user requirements.txt ./
	RUN pip install --no-cache-dir --upgrade pip && \
	pip install --no-cache-dir -r requirements.txt

	# riprap-models delta deps. Use the existing requirements.txt at the
	# service level, but skip requirements-full.txt — its ROCm-frozen
	# torch pin would clobber the cu124 wheels installed above.
	COPY --chown=user:user services/riprap-models/requirements.txt /tmp/req-models.txt
	RUN pip install --no-cache-dir -r /tmp/req-models.txt

	# Bake torchvision (CUDA 12.4 wheel) and peft at build time. The
	# canonical entrypoint.sh runtime-installs torchvision via the EO
	# toolchain path because the canonical CPU Space's build sandbox is
	# too tight; L4 builds have more room, and a properly matched
	# torchvision avoids the `torchvision::nms does not exist` runtime
	# error the canonical setup hits. peft is required by the riprap-
	# models service for the TerraMind LoRA inference path.
	RUN pip install --no-cache-dir \
	--index-url https://download.pytorch.org/whl/cu124 \
	torchvision \
	&& pip install --no-cache-dir peft==0.18.1

	# Bake Granite 4.1 weights into the image (EO toolchain is installed
	# at runtime — see entrypoint.l4.sh — to keep the build sandbox under
	# its disk threshold).
	RUN mkdir -p $OLLAMA_MODELS && \
	ollama serve & \
	OPID=$! && \
	for i in $(seq 1 30); do curl -sf http://127.0.0.1:11434/ > /dev/null && break; sleep 1; done && \
	ollama pull granite4.1:8b && \
	kill $OPID 2>/dev/null \|\| true && \
	sleep 2

	# App code, fixtures, and inline model service.
	COPY --chown=user:user app/ ./app/
	COPY --chown=user:user web/ ./web/
	COPY --chown=user:user scripts/ ./scripts/
	COPY --chown=user:user data/ ./data/
	COPY --chown=user:user corpus/ ./corpus/
	COPY --chown=user:user services/riprap-models/main.py ./riprap_models.py
	COPY --chown=user:user agent.py riprap.py ./
	COPY --chown=user:user entrypoint.sh ./entrypoint.sh
	RUN chmod +x ./entrypoint.sh

	RUN chown -R user:user /home/user
	USER user

	EXPOSE 7860
	CMD ["./entrypoint.sh"]