fix(infer): add repetition_penalty and CLI ngram parameters to prevent text looping

README.md

@@ -289,9 +289,15 @@ Useful options:
 ```shell
 --model_dir baidu/Unlimited-OCR   # Local path or Hugging Face model ID
 --gpu 0                           # CUDA_VISIBLE_DEVICES value
+--repetition_penalty 1.1          # Repetition penalty factor to prevent text looping (>1.0 to enable)
+--no_repeat_ngram_size 35         # N-gram size to ban repeated sequences (0 to disable)
+--ngram_window 128                # Sliding window size for n-gram repetition blocking
 --server_log ./log/sglang_server.log
 ```
 
+> **Preventing Text Repetition Loops:**
+> On certain dense or structured document pages, short text phrases may repeat endlessly. Pass `--repetition_penalty 1.1` in `infer.py` or set `repetition_penalty=1.1` in `model.generate()` to prevent repetition loops. You can also tune `no_repeat_ngram_size` (e.g. `15`-`35`) to detect shorter repeating units.
+
 For OmniDocBench evaluation, you need to perform the following post-processing.
 ```python
 def remove_det(raw: str) -> str:

infer.py

@@ -32,6 +32,7 @@
 CONTEXT_LENGTH = 32768
 NO_REPEAT_NGRAM_SIZE = 35
 NGRAM_WINDOW = 128
+REPETITION_PENALTY = 1.0
 REQUEST_TIMEOUT = 1200
 MAX_RETRIES = 5
 NO_REPEAT_NGRAM_PROCESSOR_STR = None
@@ -194,11 +195,15 @@ def infer_one(image_path: str, output_file: str | None, args, idx: int) -> dict:
         "stream": True,
         "images_config": {"image_mode": args.image_mode},
     }
-    if NO_REPEAT_NGRAM_SIZE > 0 and NGRAM_WINDOW > 0:
+    if getattr(args, "repetition_penalty", REPETITION_PENALTY) > 1.0:
+        payload["repetition_penalty"] = args.repetition_penalty
+    ngram_size = getattr(args, "no_repeat_ngram_size", NO_REPEAT_NGRAM_SIZE)
+    window_size = getattr(args, "ngram_window", NGRAM_WINDOW)
+    if ngram_size > 0 and window_size > 0:
         payload["custom_logit_processor"] = get_ngram_processor_str()
         payload["custom_params"] = {
-            "ngram_size": NO_REPEAT_NGRAM_SIZE,
-            "window_size": NGRAM_WINDOW,
+            "ngram_size": ngram_size,
+            "window_size": window_size,
         }
 
     name = os.path.basename(image_path)
@@ -312,6 +317,9 @@ def parse_args():
     parser.add_argument("--gpu", default="0")
     parser.add_argument("--model_dir", default="baidu/Unlimited-OCR")
     parser.add_argument("--image_mode", choices=("gundam", "base"), default="gundam")
+    parser.add_argument("--no_repeat_ngram_size", type=int, default=35, help="N-gram size to ban repeated sequences (0 to disable)")
+    parser.add_argument("--ngram_window", type=int, default=128, help="Sliding window size for n-gram repetition blocking")
+    parser.add_argument("--repetition_penalty", type=float, default=1.0, help="Repetition penalty factor to prevent text looping (>1.0 to enable)")
     parser.add_argument("--server_log", default="./log/sglang_server.log")
     return parser.parse_args()
 

tests/test_infer_repetition.py

@@ -0,0 +1,66 @@
+"""
+Unit test for repetition penalty and ngram parameters in infer.py (Fixes #55).
+"""
+
+import sys
+import unittest
+from unittest.mock import MagicMock, patch
+
+import infer
+
+
+class TestInferRepetitionControl(unittest.TestCase):
+    def test_parse_args_defaults(self):
+        with patch.object(sys, "argv", ["infer.py"]):
+            args = infer.parse_args()
+            self.assertEqual(args.repetition_penalty, 1.0)
+            self.assertEqual(args.no_repeat_ngram_size, 35)
+            self.assertEqual(args.ngram_window, 128)
+
+    def test_parse_args_custom_repetition(self):
+        test_argv = [
+            "infer.py",
+            "--repetition_penalty",
+            "1.1",
+            "--no_repeat_ngram_size",
+            "15",
+            "--ngram_window",
+            "256",
+        ]
+        with patch.object(sys, "argv", test_argv):
+            args = infer.parse_args()
+            self.assertEqual(args.repetition_penalty, 1.1)
+            self.assertEqual(args.no_repeat_ngram_size, 15)
+            self.assertEqual(args.ngram_window, 256)
+
+    @patch("infer.requests.post")
+    @patch("infer.get_ngram_processor_str", return_value="MockNGramProcessor")
+    @patch("infer.build_content", return_value=[{"type": "text", "text": "test"}])
+    def test_infer_one_payload_building(self, mock_build_content, mock_get_processor, mock_post):
+        mock_response = MagicMock()
+        mock_response.status_code = 200
+        mock_response.iter_lines.return_value = [b"data: [DONE]"]
+        mock_post.return_value = mock_response
+
+        args = MagicMock()
+        args.image_mode = "gundam"
+        args.repetition_penalty = 1.1
+        args.no_repeat_ngram_size = 20
+        args.ngram_window = 64
+
+        infer.infer_one("dummy.png", None, args, 1)
+
+        self.assertTrue(mock_post.called)
+        _, kwargs = mock_post.call_args
+        import json
+
+        payload = json.loads(kwargs["data"])
+
+        self.assertEqual(payload.get("repetition_penalty"), 1.1)
+        self.assertEqual(payload.get("custom_logit_processor"), "MockNGramProcessor")
+        self.assertEqual(payload["custom_params"]["ngram_size"], 20)
+        self.assertEqual(payload["custom_params"]["window_size"], 64)
+
+
+if __name__ == "__main__":
+    unittest.main()