Update Dockerfile

Dockerfile

@@ -1,4 +1,4 @@
-FROM nvidia/cuda:11.8.0-cudnn8-runtime-ubuntu22.04
+FROM nvidia/cuda:12.1.1-cudnn8-runtime-ubuntu22.04
 
 ENV DEBIAN_FRONTEND=noninteractive
 
@@ -23,10 +23,8 @@ WORKDIR /app
 
 # ---- HEAVY INSTALLS (cached unless dependencies change) ----
 
-# PyTorch (CUDA 11.8) - need >= 2.4 for SNAC's weight_norm
-RUN pip install --no-cache-dir \
-    torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 \
-    --index-url https://download.pytorch.org/whl/cu118
+# vLLM 0.7.3 (installs compatible PyTorch 2.5.1+cu121 automatically)
+RUN pip install --no-cache-dir vllm==0.7.3
 
 # Clone FLOAT repo
 RUN git clone https://github.com/deepbrainai-research/float.git /app/float_repo
@@ -49,8 +47,7 @@ RUN python -c "from huggingface_hub import snapshot_download; \
     snapshot_download('hubertsiuzdak/snac_24khz')" \
     || echo 'WARNING: SNAC pre-download failed'
 
-# NOTE: Both Orpheus models (pretrained + finetuned) are GATED
-# They download at runtime using HF_TOKEN secret
+# NOTE: Both Orpheus models are GATED - download at runtime with HF_TOKEN
 # Accept terms at: https://huggingface.co/canopylabs/orpheus-3b-0.1-pretrained
 # And optionally: https://huggingface.co/canopylabs/orpheus-3b-0.1-ft