Spaces:
Sleeping
Sleeping
MAX_BATCH_TOKENS=32768, MAX_INPUT_LENGTH=8192
Browse files- Dockerfile +3 -2
Dockerfile
CHANGED
|
@@ -3,11 +3,12 @@ FROM ghcr.io/huggingface/text-embeddings-inference:cpu-1.9
|
|
| 3 |
ENV MODEL_ID=Qwen/Qwen3-Embedding-0.6B
|
| 4 |
ENV PORT=7860
|
| 5 |
ENV MAX_CONCURRENT_REQUESTS=8
|
| 6 |
-
ENV MAX_BATCH_TOKENS=
|
|
|
|
| 7 |
ENV MAX_CLIENT_BATCH_SIZE=32
|
| 8 |
ENV DTYPE=float16
|
| 9 |
ENV MAX_INPUT_LENGTH=8192
|
| 10 |
|
| 11 |
EXPOSE 7860
|
| 12 |
|
| 13 |
-
ENTRYPOINT ["text-embeddings-router"]
|
|
|
|
| 3 |
ENV MODEL_ID=Qwen/Qwen3-Embedding-0.6B
|
| 4 |
ENV PORT=7860
|
| 5 |
ENV MAX_CONCURRENT_REQUESTS=8
|
| 6 |
+
ENV MAX_BATCH_TOKENS=32768
|
| 7 |
+
ENV MAX_BATCH_REQUESTS=16
|
| 8 |
ENV MAX_CLIENT_BATCH_SIZE=32
|
| 9 |
ENV DTYPE=float16
|
| 10 |
ENV MAX_INPUT_LENGTH=8192
|
| 11 |
|
| 12 |
EXPOSE 7860
|
| 13 |
|
| 14 |
+
ENTRYPOINT ["text-embeddings-router"]
|