clone Dockerfile from LevArtesa/grpo-humanizer-training

Dockerfile

@@ -0,0 +1,39 @@
+FROM pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel
+
+# Create user for HF Spaces (runs as uid 1000)
+RUN useradd -m -u 1000 user || true
+ENV HOME=/home/user
+ENV PATH=/home/user/.local/bin:$PATH
+
+# Set working directory
+WORKDIR /app
+
+# Copy requirements and install
+# requirements.txt already pins python-docx>=1.1 and hypothesis>=6.100
+# which are required by Dataset v2 tooling (document_cutter, PBT tests).
+COPY requirements.txt .
+RUN pip install --no-cache-dir --upgrade pip && \
+    pip install --no-cache-dir -r requirements.txt
+
+# Copy all files (training/* plus scripts/* when the build context is the
+# repo root). entrypoint.sh lands at /app/entrypoint.sh.
+COPY . .
+
+# Make the dispatch entrypoint executable. The script selects between
+# serve / train / train_v2 / filter_dataset_v2 / evaluate based on the
+# ENTRYPOINT_MODE environment variable (see training/entrypoint.sh for
+# details). Default is "serve" to preserve the existing HF Space
+# inference behaviour.
+RUN chmod +x /app/entrypoint.sh
+
+# Fix permissions
+RUN chown -R 1000:1000 /app /home/user
+USER 1000
+
+# Set cache dirs to writable locations
+ENV HF_HOME=/home/user/.cache/huggingface
+ENV TORCH_HOME=/home/user/.cache/torch
+
+# Default to the inference server. Override by setting
+# ENTRYPOINT_MODE=train_v2 (or another supported value) on the HF Space.
+CMD ["/app/entrypoint.sh"]