added comments after using skill

Author: Eric Zhong

.github/config/model-tests/vllm-multimodal-benchmark-tests.yml

@@ -28,10 +28,12 @@ benchmark:
       test_script: "vllm_embedding_benchmark_test.sh"
       min_rps: 3
 
+    # LightOnOCR-2-1B: 1B VLM for document OCR (LightOnOcrForConditionalGeneration).
+    # bf16 ~2GB, fits L40S easily. Thresholds are provisional — calibrate after first run.
     - name: "lighton-ocr-2-1b"
       s3_model: "lightonocr-2-1b.tar.gz"
       runner_label: "gpu-l40s-1gpu-runners"
-      extra_args: "--dtype bfloat16 --max-model-len 16384"
+      extra_args: "--dtype bfloat16 --max-model-len 4096"
       min_throughput: 30
       min_rps: 1
       benchmark_profiles: "baseline,high_concurrency"