Update Dockerfile

Dockerfile

@@ -19,14 +19,9 @@ WORKDIR $HOME/app
 COPY --chown=user . $HOME/app
 
 RUN python3 -m pip install aphrodite-engine
-RUN wget -O model.gguf https://huggingface.co/MarsupialAI/Yi-34B-200K-RPMerge_GGUF/resolve/main/Yi-34B-200K-RPMerge_Q4km.gguf?download=true
-RUN wget -O config.json "https://huggingface.co/brucethemoose/Yi-34B-200K-RPMerge/resolve/main/config.json?download=true"
-RUN wget -O special_tokens_map.json "https://huggingface.co/brucethemoose/Yi-34B-200K-RPMerge/resolve/main/special_tokens_map.json?download=true"
-RUN wget -O tokenizer.json "https://huggingface.co/brucethemoose/Yi-34B-200K-RPMerge/resolve/main/tokenizer.json?download=true"
-RUN wget -O tokenizer_config.json "https://huggingface.co/brucethemoose/Yi-34B-200K-RPMerge/resolve/main/tokenizer_config.json?download=true"
-RUN wget -O gguf_to_torch.py https://raw.githubusercontent.com/PygmalionAI/aphrodite-engine/main/examples/gguf_to_torch.py
-RUN python3 -m pip install protobuf==3.20.*
-RUN python3 gguf_to_torch.py --input $HOME/app/model.gguf --output $HOME/app/
+RUN pip install huggingface-hub hf-transfer
+ENV HF_HUB_ENABLE_HF_TRANSFER=1
+RUN huggingface-cli download brucethemoose/Yi-34B-200K-RPMerge --local-dir $HOME/goliath-gptq --local-dir-use-symlinks False --cache-dir $HOME/cache
 EXPOSE 7860
 # Should expose on all network interfaces so 0.0.0.0 not only localhost
-CMD /bin/bash -c "/bin/python3 -m aphrodite.endpoints.openai.api_server -q gguf --dtype half -tp 4 -gmu 0.95 --kv-cache-dtype fp8_e5m2 --max-num-seqs 15 --served-model-name "MarsupialAI/Yi-34B-200K-RPMerge_GGUF" --enforce-eager --port 7860 --host 0.0.0.0 --model $HOME/app/"
+CMD /bin/bash -c "/bin/python3 -m aphrodite.endpoints.openai.api_server $ENGINE_ARGS --port 7860 --host 0.0.0.0 --model ~/goliath-gptq"