Spaces:

Shwy1001
/

BioPatch

Sleeping

BioPatch / Dockerfile

Jerry

BioPatch backend (clean state, history rewritten)

10795db 4 months ago

4.96 kB

# syntax=docker/dockerfile:1.7
# ---------------------------------------------------------------------------
# BioPatch backend — Hugging Face Spaces + Fly.io compatible.
#
# IMPORTANT — LLM backend on this image:
#   Default: JARVIS_LLM_BACKEND=local  → llama-cpp-python in-process
#   Default model: BioMistral-7B Q4_K_M (~4.4 GB weights, ~6 GB RAM with
#   8K context). Tuned for the HF Spaces cpu-upgrade tier (8 vCPU /
#   32 GB RAM). On the free tier this OOMs at load time — fall back to
#   a smaller model by overriding JARVIS_LLM_REPO / JARVIS_LLM_FILE /
#   JARVIS_LLM_CHAT_FORMAT in the Space Variables.
#
#   Why local and not HF Inference on a Space:
#   HF Spaces sandbox blocks outbound DNS to api-inference.huggingface.co
#   (anti-loop policy), so the hosted Inference API is unreachable from
#   inside a Space. Local llama-cpp works.
#
#   Set JARVIS_LLM_BACKEND=hf if you deploy this image OUTSIDE a Space
#   (e.g. Fly.io) and prefer the hosted API; you'll also need HF_TOKEN.
#
# Persistent storage:
#   When the Space has persistent storage enabled, /data is mounted as
#   a writable volume. The env vars below default to /data paths so the
#   SQLite DB, wound photos and the GGUF model cache survive restarts.
#   Without persistent storage, override these to /app paths via the
#   Space Variables.
#
# Thread safety: every BLAS / OpenMP / sklearn parallel pool is capped
# at 1 thread + the RF model is forced to n_jobs=1 in code. This
# eliminates the GGML_ASSERT / 'Failed joining thread' cascade that
# crashed the container under simulator load (joblib pool corrupted
# llama's GGML workspace).
# ---------------------------------------------------------------------------

ARG PYTHON_VERSION=3.12
FROM python:${PYTHON_VERSION}-slim AS base

ENV PYTHONDONTWRITEBYTECODE=1 \
    PYTHONUNBUFFERED=1 \
    PIP_NO_CACHE_DIR=1 \
    PIP_DISABLE_PIP_VERSION_CHECK=1 \
    PIP_PREFER_BINARY=1

RUN apt-get update && apt-get install -y --no-install-recommends \
    build-essential \
    libopenblas-dev \
    ca-certificates \
    curl \
    && rm -rf /var/lib/apt/lists/*

# HF Spaces runs as UID 1000.
RUN useradd -m -u 1000 user

WORKDIR /app

# Core deps first.
COPY backend/requirements.txt /app/requirements.txt
RUN pip install -r requirements.txt

# LLM stack — pull pre-built CPU wheels from abetlen's dedicated index.
# llama-cpp-python intentionally publishes only source dists to PyPI;
# the binary wheels (built per Python version × platform) live at
# https://abetlen.github.io/llama-cpp-python/whl/cpu. With --only-binary
# pip refuses to compile from source, so if the wheel is genuinely
# missing for our Python it errors out in seconds with a clear message
# instead of OOMing 8 minutes into a C++ compile on the HF free tier.
COPY backend/requirements-llm.txt /app/requirements-llm.txt
RUN pip install --only-binary=:all: \
      --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu \
      -r requirements-llm.txt

COPY backend /app

RUN python -m scripts.seed_clinical_data || true

# Thread caps — keep sklearn / OpenBLAS / OpenMP / MKL single-threaded.
# Combined with the n_jobs=1 forced on the loaded RandomForest model,
# this guarantees the joblib threadpool never spawns and never collides
# with llama-cpp's GGML workspace.
ENV OMP_NUM_THREADS=1 \
    OPENBLAS_NUM_THREADS=1 \
    MKL_NUM_THREADS=1 \
    BLIS_NUM_THREADS=1 \
    NUMEXPR_MAX_THREADS=1 \
    JARVIS_LLM_BACKEND=local \
    JARVIS_LLM_REPO=MaziyarPanahi/BioMistral-7B-GGUF \
    JARVIS_LLM_FILE=BioMistral-7B.Q4_K_M.gguf \
    JARVIS_LLM_CHAT_FORMAT=mistral-instruct \
    JARVIS_LLM_CACHE=/data/models \
    JARVIS_LLM_CTX=4096 \
    JARVIS_LLM_THREADS=8 \
    JARVIS_HF_MODEL=meta-llama/Llama-3.2-3B-Instruct \
    DATABASE_URL=sqlite:////data/biopatch.db \
    WOUND_PHOTO_DIR=/data/wound_photos \
    HF_HOME=/data/.cache/huggingface \
    PORT=7860 \
    ENABLE_SIMULATOR=true \
    SIMULATOR_INTERVAL_SECONDS=15.0 \
    AUTO_SEED=true

# Pre-create fallback dirs under /app for hosts WITHOUT persistent /data
# (Fly.io, local docker run, etc.). On a HF Space with persistent storage
# enabled, /data is mounted at runtime and the env vars above override
# these paths.
RUN mkdir -p /app/ai/artifacts/models /app/data/wound_photos /app/.cache/huggingface \
    && chown -R user:user /app

# Stay as root for the runtime. HF Spaces persistent /data is mounted
# with FUSE-style perms that silently reject chown from inside the
# container, so dropping to uid 1000 leaves the volume unwritable
# (sqlite OperationalError: unable to open database file). Running
# uvicorn as root sidesteps the perm check; the HF sandbox confines
# the container regardless. The threading caps + RF n_jobs=1 already
# neutralise the previous GGML / joblib collision, so non-root is no
# longer required for stability.

EXPOSE 8000 7860

CMD ["sh", "-c", "uvicorn main:app --host 0.0.0.0 --port ${PORT:-8000} --workers 1"]

	# syntax=docker/dockerfile:1.7
	# ---------------------------------------------------------------------------
	# BioPatch backend — Hugging Face Spaces + Fly.io compatible.
	#
	# IMPORTANT — LLM backend on this image:
	# Default: JARVIS_LLM_BACKEND=local → llama-cpp-python in-process
	# Default model: BioMistral-7B Q4_K_M (~4.4 GB weights, ~6 GB RAM with
	# 8K context). Tuned for the HF Spaces cpu-upgrade tier (8 vCPU /
	# 32 GB RAM). On the free tier this OOMs at load time — fall back to
	# a smaller model by overriding JARVIS_LLM_REPO / JARVIS_LLM_FILE /
	# JARVIS_LLM_CHAT_FORMAT in the Space Variables.
	#
	# Why local and not HF Inference on a Space:
	# HF Spaces sandbox blocks outbound DNS to api-inference.huggingface.co
	# (anti-loop policy), so the hosted Inference API is unreachable from
	# inside a Space. Local llama-cpp works.
	#
	# Set JARVIS_LLM_BACKEND=hf if you deploy this image OUTSIDE a Space
	# (e.g. Fly.io) and prefer the hosted API; you'll also need HF_TOKEN.
	#
	# Persistent storage:
	# When the Space has persistent storage enabled, /data is mounted as
	# a writable volume. The env vars below default to /data paths so the
	# SQLite DB, wound photos and the GGUF model cache survive restarts.
	# Without persistent storage, override these to /app paths via the
	# Space Variables.
	#
	# Thread safety: every BLAS / OpenMP / sklearn parallel pool is capped
	# at 1 thread + the RF model is forced to n_jobs=1 in code. This
	# eliminates the GGML_ASSERT / 'Failed joining thread' cascade that
	# crashed the container under simulator load (joblib pool corrupted
	# llama's GGML workspace).
	# ---------------------------------------------------------------------------

	ARG PYTHON_VERSION=3.12
	FROM python:${PYTHON_VERSION}-slim AS base

	ENV PYTHONDONTWRITEBYTECODE=1 \
	PYTHONUNBUFFERED=1 \
	PIP_NO_CACHE_DIR=1 \
	PIP_DISABLE_PIP_VERSION_CHECK=1 \
	PIP_PREFER_BINARY=1

	RUN apt-get update && apt-get install -y --no-install-recommends \
	build-essential \
	libopenblas-dev \
	ca-certificates \
	curl \
	&& rm -rf /var/lib/apt/lists/*

	# HF Spaces runs as UID 1000.
	RUN useradd -m -u 1000 user

	WORKDIR /app

	# Core deps first.
	COPY backend/requirements.txt /app/requirements.txt
	RUN pip install -r requirements.txt

	# LLM stack — pull pre-built CPU wheels from abetlen's dedicated index.
	# llama-cpp-python intentionally publishes only source dists to PyPI;
	# the binary wheels (built per Python version × platform) live at
	# https://abetlen.github.io/llama-cpp-python/whl/cpu. With --only-binary
	# pip refuses to compile from source, so if the wheel is genuinely
	# missing for our Python it errors out in seconds with a clear message
	# instead of OOMing 8 minutes into a C++ compile on the HF free tier.
	COPY backend/requirements-llm.txt /app/requirements-llm.txt
	RUN pip install --only-binary=:all: \
	--extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu \
	-r requirements-llm.txt

	COPY backend /app

	RUN python -m scripts.seed_clinical_data \|\| true

	# Thread caps — keep sklearn / OpenBLAS / OpenMP / MKL single-threaded.
	# Combined with the n_jobs=1 forced on the loaded RandomForest model,
	# this guarantees the joblib threadpool never spawns and never collides
	# with llama-cpp's GGML workspace.
	ENV OMP_NUM_THREADS=1 \
	OPENBLAS_NUM_THREADS=1 \
	MKL_NUM_THREADS=1 \
	BLIS_NUM_THREADS=1 \
	NUMEXPR_MAX_THREADS=1 \
	JARVIS_LLM_BACKEND=local \
	JARVIS_LLM_REPO=MaziyarPanahi/BioMistral-7B-GGUF \
	JARVIS_LLM_FILE=BioMistral-7B.Q4_K_M.gguf \
	JARVIS_LLM_CHAT_FORMAT=mistral-instruct \
	JARVIS_LLM_CACHE=/data/models \
	JARVIS_LLM_CTX=4096 \
	JARVIS_LLM_THREADS=8 \
	JARVIS_HF_MODEL=meta-llama/Llama-3.2-3B-Instruct \
	DATABASE_URL=sqlite:////data/biopatch.db \
	WOUND_PHOTO_DIR=/data/wound_photos \
	HF_HOME=/data/.cache/huggingface \
	PORT=7860 \
	ENABLE_SIMULATOR=true \
	SIMULATOR_INTERVAL_SECONDS=15.0 \
	AUTO_SEED=true

	# Pre-create fallback dirs under /app for hosts WITHOUT persistent /data
	# (Fly.io, local docker run, etc.). On a HF Space with persistent storage
	# enabled, /data is mounted at runtime and the env vars above override
	# these paths.
	RUN mkdir -p /app/ai/artifacts/models /app/data/wound_photos /app/.cache/huggingface \
	&& chown -R user:user /app

	# Stay as root for the runtime. HF Spaces persistent /data is mounted
	# with FUSE-style perms that silently reject chown from inside the
	# container, so dropping to uid 1000 leaves the volume unwritable
	# (sqlite OperationalError: unable to open database file). Running
	# uvicorn as root sidesteps the perm check; the HF sandbox confines
	# the container regardless. The threading caps + RF n_jobs=1 already
	# neutralise the previous GGML / joblib collision, so non-root is no
	# longer required for stability.

	EXPOSE 8000 7860

	CMD ["sh", "-c", "uvicorn main:app --host 0.0.0.0 --port ${PORT:-8000} --workers 1"]