CUDA + cudnn + PyTorch 2.5.1

Dockerfile

@@ -1,29 +1,34 @@
-# Read the doc: https://huggingface.co/docs/hub/spaces-sdks-docker
-# you will also find guides on how best to write your Dockerfile
+# CUDA + cudnn + PyTorch 2.5.1 (CUDA 12.4)
+FROM pytorch/pytorch:2.5.1-cuda12.4-cudnn9-devel
 
-FROM python:3.11
-RUN apt-get update && apt-get install -y openssh-client sshpass
+# Sistema y build tools
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    git openssh-client sshpass build-essential cmake ninja-build kmod curl ca-certificates && \
+    rm -rf /var/lib/apt/lists/*
 
-RUN useradd -m -u 1000 user
-USER user
-ENV PATH="/home/user/.local/bin:$PATH"
-
-WORKDIR /app
+# CUDA env
+ENV CUDA_HOME=/usr/local/cuda
+ENV PATH=$CUDA_HOME/bin:$PATH
+ENV LD_LIBRARY_PATH=/usr/local/cuda/lib64:${LD_LIBRARY_PATH}
 
-COPY --chown=user . /app
-RUN pip install -e .
+# Evitar reinstall de torch/torchvision ya presentes y asegurar numpy
+RUN pip install --no-cache-dir -U pip && pip install --no-cache-dir numpy
 
-# Instala las dependencias de construcción necesarias
-RUN pip install numpy torch
+# Código app
+WORKDIR /app
+COPY . /app
 
-# Establece la variable de entorno CUDA_HOME
-ENV CUDA_HOME=/usr/local/cuda
+# Instalar el proyecto (usa pyproject del repo)
+RUN pip install --no-cache-dir -e .
 
-# Clona el fork de vLLM de StarVector
-RUN git clone https://github.com/starvector/vllm.git /app/vllm
+# Sustituir vLLM oficial por el fork de StarVector
+RUN git clone https://github.com/starvector/vllm.git /app/vllm && \
+    pip install --no-cache-dir -e /app/vllm
 
-# Instala el fork de vLLM
-RUN cd /app/vllm && pip install -e .
+# Crear usuario no root después de instalar dependencias
+RUN useradd -m -u 1000 user && chown -R user:user /app
+USER user
 
+# Permisos y arranque
 RUN chmod +x ./start.sh
-CMD ["./start.sh"]
+CMD ["./start.sh"]