Fix glm-ocr.py max_tokens exceeding max_model_len

GLM-OCR SDK sets max_tokens=16384 but our max_model_len is 8192.
vLLM silently caps generation, making the high default misleading.
Align default to 8192 and document the SDK discrepancy.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

glm-ocr.py

@@ -214,7 +214,7 @@ def main(
     image_column: str = "image",
     batch_size: int = 16,
     max_model_len: int = 8192,
-    max_tokens: int = 16384,
+    max_tokens: int = 8192,
     temperature: float = 0.01,
     top_p: float = 0.00001,
     repetition_penalty: float = 1.1,
@@ -278,6 +278,7 @@ def main(
     # glmocr/config.py PageLoaderConfig: temperature=0.01, top_p=0.00001,
     # top_k=1, repetition_penalty=1.1, max_tokens=16384
     # generation_config.json on HF also sets do_sample=false (greedy)
+    # Note: SDK uses max_tokens=16384 but vLLM caps at max_model_len (8192)
     sampling_params = SamplingParams(
         temperature=temperature,
         top_p=top_p,
@@ -454,8 +455,8 @@ Examples:
     parser.add_argument(
         "--max-tokens",
         type=int,
-        default=16384,
-        help="Maximum tokens to generate (default: 16384)",
+        default=8192,
+        help="Maximum tokens to generate (default: 8192, capped by max-model-len)",
     )
     parser.add_argument(
         "--temperature",