Spaces:
Runtime error
Runtime error
Commit
•
4ac8560
1
Parent(s):
2b7e75a
Update Dockerfile
Browse files- Dockerfile +0 -1
Dockerfile
CHANGED
|
@@ -45,7 +45,6 @@ RUN huggingface-cli download LoneStriker/TinyLlama-1.1B-32k-Instruct-8.0bpw-h8-e
|
|
| 45 |
EXPOSE 7860
|
| 46 |
RUN pip install aioprometheus
|
| 47 |
ENV PYTHONPATH=$HOME/aphrodite-engine
|
| 48 |
-
RUN cd $HOME/aphrodite-engine && python3 setup.py build_ext --inplace
|
| 49 |
|
| 50 |
# Command to run the API server from the cloned directory
|
| 51 |
CMD ["/bin/bash", "-c", "cd $HOME/aphrodite-engine/aphrodite/endpoints/kobold && /bin/python3 api_server.py -q exl2 --dtype auto -gmu 0.95 --kv-cache-dtype fp8_e5m2 --max-num-seqs 15 --served-model-name \"BagelMIsteryTour-v2-8x7B-AWQ\" --enforce-eager -tp 4 --port 7860 --host 0.0.0.0 --model ~/goliath-gptq"]
|
|
|
|
| 45 |
EXPOSE 7860
|
| 46 |
RUN pip install aioprometheus
|
| 47 |
ENV PYTHONPATH=$HOME/aphrodite-engine
|
|
|
|
| 48 |
|
| 49 |
# Command to run the API server from the cloned directory
|
| 50 |
CMD ["/bin/bash", "-c", "cd $HOME/aphrodite-engine/aphrodite/endpoints/kobold && /bin/python3 api_server.py -q exl2 --dtype auto -gmu 0.95 --kv-cache-dtype fp8_e5m2 --max-num-seqs 15 --served-model-name \"BagelMIsteryTour-v2-8x7B-AWQ\" --enforce-eager -tp 4 --port 7860 --host 0.0.0.0 --model ~/goliath-gptq"]
|