Update Dockerfile

Dockerfile

@@ -1,21 +1,36 @@
-# Use official vLLM image which has everything pre-configured 
-FROM vllm/vllm-openai:latest
+# Base image with CUDA 12.1 and Ubuntu 20.04
+FROM nvidia/cuda:12.1.1-runtime-ubuntu20.04
+ENV PIP_NO_CACHE_DIR=1
+ENV DEBIAN_FRONTEND=noninteractive
+ENV TZ=Etc/UTC
+ENV TRANSFORMERS_CACHE=/tmp/hf-cache
+ENV HF_HOME=/tmp/hf-home
+ENV HF_HUB_CACHE=/tmp/hf-cache
+
+# Install system dependencies
+RUN apt-get update && apt-get install -y \
+    git python3 python3-pip ffmpeg libsndfile1 tzdata && \
+    rm -rf /var/lib/apt/lists/*
+
+# Install Python and upgrade pip and setuptools
+RUN ln -s /usr/bin/python3 /usr/bin/python && \
+    pip3 install --upgrade pip && \
+    pip3 install --upgrade setuptools wheel
+
+# Install required dependencies
+RUN pip install tokenizers==0.19.1 hf_transfer
+
+# Install vLLM with audio support from nightly wheels (this should work better)
+RUN pip install -U "vllm[audio]" --extra-index-url https://wheels.vllm.ai/nightly
 
 # Set environment variables
 ENV HF_HUB_ENABLE_HF_TRANSFER=1
 ENV TRANSFORMERS_CACHE=/data/.cache/huggingface/transformers
-ENV HF_HOME=/tmp/hf-home
-ENV TORCHINDUCTOR_CACHE_DIR=/tmp/torch_cache
-ENV USER=user
-ENV HOME=/tmp
-
-# Install additional dependencies if needed
-RUN pip install hf_transfer
 
 # Expose port
 EXPOSE 8000
 
-# Serve Voxtral with proper parameters
+# Serve with updated parameters
 CMD ["vllm", "serve", "mistralai/Voxtral-Mini-3B-2507", \
      "--tokenizer-mode", "mistral", \
      "--config-format", "mistral", \