feat(paddleocr-vl): add PaddleOCR-VL GPU Dockerfile, pin vllm, update CPU image deps, and improve entrypoint and tests

image_support_files/paddleocr-vl-entrypoint.sh

@@ -11,12 +11,16 @@ HOST="${HOST:-0.0.0.0}"
 PORT="${PORT:-8000}"
 MAX_BATCHED_TOKENS="${MAX_BATCHED_TOKENS:-16384}"
 GPU_MEMORY_UTILIZATION="${GPU_MEMORY_UTILIZATION:-0.9}"
+MAX_MODEL_LEN="${MAX_MODEL_LEN:-8192}"
+ENFORCE_EAGER="${ENFORCE_EAGER:-false}"
 
 echo "Model: ${MODEL_NAME}"
 echo "Host: ${HOST}"
 echo "Port: ${PORT}"
 echo "Max batched tokens: ${MAX_BATCHED_TOKENS}"
 echo "GPU memory utilization: ${GPU_MEMORY_UTILIZATION}"
+echo "Max model length: ${MAX_MODEL_LEN}"
+echo "Enforce eager: ${ENFORCE_EAGER}"
 echo ""
 
 # Check GPU availability
@@ -31,13 +35,25 @@ fi
 echo "Starting vLLM server..."
 echo "==================================="
 
-# Start vLLM server with PaddleOCR-VL
-exec vllm serve "${MODEL_NAME}" \
-    --trust-remote-code \
-    --host "${HOST}" \
-    --port "${PORT}" \
-    --max-num-batched-tokens "${MAX_BATCHED_TOKENS}" \
-    --gpu-memory-utilization "${GPU_MEMORY_UTILIZATION}" \
-    --no-enable-prefix-caching \
-    --mm-processor-cache-gb 0 \
+# Build vLLM command
+VLLM_ARGS=(
+    serve "${MODEL_NAME}"
+    --trust-remote-code
+    --host "${HOST}"
+    --port "${PORT}"
+    --max-num-batched-tokens "${MAX_BATCHED_TOKENS}"
+    --gpu-memory-utilization "${GPU_MEMORY_UTILIZATION}"
+    --max-model-len "${MAX_MODEL_LEN}"
+    --no-enable-prefix-caching
+    --mm-processor-cache-gb 0
     --served-model-name "paddleocr-vl"
+    --limit-mm-per-prompt '{"image": 1}'
+)
+
+# Add enforce-eager if enabled (disables CUDA graphs, saves memory)
+if [ "${ENFORCE_EAGER}" = "true" ]; then
+    VLLM_ARGS+=(--enforce-eager)
+fi
+
+# Start vLLM server with PaddleOCR-VL
+exec vllm "${VLLM_ARGS[@]}"