Initial marker-api benchmark setup: Containerfile, app.py, convert_both.py, compare_conversions.py

2026-06-05 13:26:25 +02:00
commit 460fc0003b
33 changed files with 58290 additions and 0 deletions
@@ -0,0 +1,68 @@
+# ===== containerfile (podman & docker) for marker-api on powermac  =====
+#
+# Targets 2x AMD Vega 20 (GFX906) with ROCm using gfx906-patched PyTorch.
+# deepseek-ocr and GPT services are on the same host at :8082 / :8080.
+
+FROM mixa3607/pytorch-gfx906:v2.7.1-rocm-6.3.3
+
+# install system deps & remove system-installed python3-blinker to avoid pip upgrade conflicts
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        curl ca-certificates tini procps git gcc g++ zlib1g-dev libjpeg-dev \
+        libpango-1.0-0 libharfbuzz0b libpangoft2-1.0-0 && \
+    apt-get remove -y python3-blinker || true && \
+    rm -rf /var/lib/apt/lists/*
+
+# ---- runtime ----
+RUN groupadd -r marker && (groupadd -r render || true) && (groupadd -r video || true) && useradd -r -g marker -G render,video marker && \
+    mkdir -p /app/conversion_results /app/uploads /app/.cache && \
+    chown -R marker:marker /app
+
+WORKDIR /app
+
+# ---- deps ----
+COPY marker/ /app/marker/
+
+# We use --break-system-packages and make sure we do NOT use --ignore-installed so that pre-installed
+# gfx906 patched torch, torchvision, and torchaudio are preserved and NOT overwritten by PyPI releases.
+RUN pip install --no-cache-dir --break-system-packages -e "/app/marker[full]" && \
+    pip install --no-cache-dir --break-system-packages flask gunicorn
+
+# ---- font: download GoNotoCurrent-Regular.ttf at build time ----
+RUN mkdir -p /app/marker/static/fonts && \
+    curl -sL -o /app/marker/static/fonts/GoNotoCurrent-Regular.ttf \
+    "https://models.datalab.to/artifacts/GoNotoCurrent-Regular.ttf" && \
+    chown marker:marker /app/marker/static/fonts/GoNotoCurrent-Regular.ttf
+
+# ---- final image ----
+COPY app.py          /app/
+COPY gunicorn.conf.py /app/
+COPY entrypoint.sh    /app/entrypoint.sh
+
+RUN chmod +x /app/entrypoint.sh
+
+ENV \
+    HOME=/app \
+    PYTHONUNBUFFERED=1 \
+    PYTHONDONTWRITEBYTECODE=1 \
+    MARKER_OUTPUT_DIR=/app/conversion_results \
+    UPLOAD_DIR=/app/uploads \
+    OLLAMA_HOST=http://localhost:8082 \
+    DEESEEK_OCR_MODEL=deepseek-ocr \
+    AMD_COMPUTE=true \
+    TORCH_DEVICE= \
+    MODEL_DTYPE=float32 \
+    PORT=8000 \
+    HOST=0.0.0.0 \
+    LLM_SERVICE=marker.services.openai.OpenAIService \
+    USE_LLM=false \
+    OPENAI_BASE_URL=http://localhost:8082/v1 \
+    OPENAI_MODEL= \
+    GUNICORN_WORKERS=1
+
+USER marker
+
+EXPOSE 8000
+
+ENTRYPOINT ["/app/entrypoint.sh"]
+CMD ["gunicorn", "-c", "gunicorn.conf.py", "app:app_instance"]