Spaces:
Runtime error
Runtime error
Commit
•
d0da2f1
1
Parent(s):
1fe6da2
Update Dockerfile
Browse files- Dockerfile +3 -0
Dockerfile
CHANGED
|
@@ -44,5 +44,8 @@ RUN huggingface-cli download LoneStriker/TinyLlama-1.1B-32k-Instruct-8.0bpw-h8-e
|
|
| 44 |
# Expose the port the API server will listen on
|
| 45 |
EXPOSE 7860
|
| 46 |
RUN pip install aioprometheus
|
|
|
|
|
|
|
|
|
|
| 47 |
# Command to run the API server from the cloned directory
|
| 48 |
CMD ["/bin/bash", "-c", "cd $HOME/aphrodite-engine/aphrodite/endpoints/openai && /bin/python3 api_server.py -q exl2 --dtype auto -gmu 0.95 --kv-cache-dtype fp8_e5m2 --max-num-seqs 15 --served-model-name \"BagelMIsteryTour-v2-8x7B-AWQ\" --enforce-eager -tp 4 --port 7860 --host 0.0.0.0 --model ~/goliath-gptq"]
|
|
|
|
| 44 |
# Expose the port the API server will listen on
|
| 45 |
EXPOSE 7860
|
| 46 |
RUN pip install aioprometheus
|
| 47 |
+
ENV PYTHONPATH=$HOME/aphrodite-engine
|
| 48 |
+
|
| 49 |
+
|
| 50 |
# Command to run the API server from the cloned directory
|
| 51 |
CMD ["/bin/bash", "-c", "cd $HOME/aphrodite-engine/aphrodite/endpoints/openai && /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"]
|