Update app.py

app.py

@@ -14,14 +14,15 @@ import yaml
 log = logging.getLogger("glmocr_selfhosted_space")
 logging.basicConfig(level=logging.INFO)
 
-MODEL_ID = os.getenv("GLMOCR_MODEL_ID", "zai-org/GLM-OCR")
-OCR_API_PORT = int(os.getenv("OCR_API_PORT", "8080"))
+MODEL_ID = "zai-org/GLM-OCR"
+OCR_API_PORT = 8080
 OCR_API_BASE = f"http://127.0.0.1:{OCR_API_PORT}"
-# GLM-OCR self-hosted requests can ask up to 8192 output tokens.
-# Keep server context length aligned unless explicitly overridden by env.
-VLLM_MAX_MODEL_LEN = os.getenv("VLLM_MAX_MODEL_LEN", "8192")
-VLLM_GPU_MEMORY_UTIL = os.getenv("VLLM_GPU_MEMORY_UTIL", "0.75")
-VLLM_EXTRA_ARGS = os.getenv("VLLM_EXTRA_ARGS", "").strip()
+# Fixed runtime settings (no environment variables required).
+VLLM_MAX_MODEL_LEN = "8192"
+VLLM_GPU_MEMORY_UTIL = "0.75"
+VLLM_EXTRA_ARGS = ""
+# Keep output budget below context length to leave room for prompt tokens.
+GLMOCR_MAX_OUTPUT_TOKENS = 2048
 
 _parser = None
 _vllm_proc = None
@@ -88,6 +89,8 @@ def _configure_glmocr_to_keep_header_footer() -> None:
     ocr_api["api_path"] = "/v1/chat/completions"
     ocr_api["api_mode"] = "openai"
     ocr_api["verify_ssl"] = False
+    page_loader = pipeline.setdefault("page_loader", {})
+    page_loader["max_tokens"] = GLMOCR_MAX_OUTPUT_TOKENS
 
     layout = pipeline.setdefault("layout", {})
     label_task_mapping = layout.setdefault("label_task_mapping", {})