svc-gradiotei-fm-template

Paused

chansung commited on Mar 21

Commit

73f0ed3

•

1 Parent(s): 6892c67

Update entrypoint.sh.template

Files changed (1) hide show

entrypoint.sh.template CHANGED Viewed

@@ -3,14 +3,14 @@
 if [[ "$QUANTIZATION" == "false" ]]; then
   text-generation-launcher --model-id $MODEL_NAME \
     --num-shard 1 --port 8080 --trust-remote-code \
-    --max-concurrent-request $MAX_CONCURRENT_REQUESTS \
     --max-top-n-tokens $MAX_INPUT_LENGTH \
     --max-input-length $MAX_TOTAL_TOKENS \
     &
 else
   text-generation-launcher --model-id $MODEL_NAME \
     --num-shard 1 --port 8080 --trust-remote-code \
-    --max-concurrent-request $MAX_CONCURRENT_REQUESTS \
     --max-top-n-tokens $MAX_INPUT_LENGTH \
     --max-input-length $MAX_TOTAL_TOKENS \
     --quantize $QUANTIZATION \

 if [[ "$QUANTIZATION" == "false" ]]; then
   text-generation-launcher --model-id $MODEL_NAME \
     --num-shard 1 --port 8080 --trust-remote-code \
+    --max-concurrent-requests $MAX_CONCURRENT_REQUESTS \
     --max-top-n-tokens $MAX_INPUT_LENGTH \
     --max-input-length $MAX_TOTAL_TOKENS \
     &
 else
   text-generation-launcher --model-id $MODEL_NAME \
     --num-shard 1 --port 8080 --trust-remote-code \
+    --max-concurrent-requests $MAX_CONCURRENT_REQUESTS \
     --max-top-n-tokens $MAX_INPUT_LENGTH \
     --max-input-length $MAX_TOTAL_TOKENS \
     --quantize $QUANTIZATION \