inference: add CLI overrides for fps/total_pixels and vLLM memory knobs

Author: paularamo

scripts/inference.py

@@ -148,6 +148,20 @@ def main():
         type=str,
         help="Output directory for debugging",
     )
+    #Adding the next two arguments to improve performance on the GPU
+    parser.add_argument(
+        "--gpu-memory-utilization",
+        type=float,
+        default=0.60,
+        help="Target fraction of GPU memory vLLM can use for model + KV cache",
+    )
+    parser.add_argument(
+        "--max-model-len",
+        type=int,
+        default=1024,
+        help="Maximum sequence length for sizing KV cache",
+    )
+
     args = parser.parse_args()
 
     images: list[str] = args.images or []
@@ -204,8 +218,11 @@ def main():
         revision=args.revision,
         limit_mm_per_prompt={"image": len(images), "video": len(videos)},
         enforce_eager=True,
+        gpu_memory_utilization=args.gpu_memory_utilization,
+        max_model_len=args.max_model_len,
     )
 
+
     # Process inputs
     processor: transformers.Qwen2_5_VLProcessor = (
         transformers.AutoProcessor.from_pretrained(args.model)
@@ -239,14 +256,22 @@ def main():
         "mm_processor_kwargs": video_kwargs,
     }
     outputs = llm.generate([llm_inputs], sampling_params=sampling_params)
+
     print(SEPARATOR)
+    full_texts = []
     for output in outputs[0].outputs:
         output_text = output.text
+        full_texts.append(output_text)
         print("Assistant:")
         print(textwrap.indent(output_text.rstrip(), "  "))
     print(SEPARATOR)
 
-    result, _ = extract_tagged_text(output_text)
+    # Debug: show raw length so we know if it’s really short
+    print(f"[DEBUG] Total outputs: {len(full_texts)}")
+    print(f"[DEBUG] Last output length: {len(full_texts[-1]) if full_texts else 0}")
+
+    result, _ = extract_tagged_text(full_texts[-1])
+
     if args.verbose and result:
         pprint_dict(result, "Result")