From 30aa84f0571ba4417e2f9ed67a4ba637b41e96ba Mon Sep 17 00:00:00 2001
From: xiawei <xiawei@4pdadmin.com>
Date: Fri, 14 Nov 2025 15:36:52 +0800
Subject: [PATCH] DESC: fix infer

---
 modeling_deepseekocr.py | 21 ++++++++++-----------
 1 file changed, 10 insertions(+), 11 deletions(-)

diff --git a/modeling_deepseekocr.py b/modeling_deepseekocr.py
index 05ebf94..b39f210 100644
--- a/modeling_deepseekocr.py
+++ b/modeling_deepseekocr.py
@@ -944,17 +944,6 @@ class DeepseekOCRForCausalLM(DeepseekV2ForCausalLM):
                         no_repeat_ngram_size = 35,
                         use_cache = True
                         )
-                
-
-        if '<image>' in conversation[0]['content'] and eval_mode:
-                outputs = tokenizer.decode(output_ids[0, input_ids.unsqueeze(0).cuda().shape[1]:])
-                stop_str = '<｜end▁of▁sentence｜>'
-                if outputs.endswith(stop_str):
-                    outputs = outputs[:-len(stop_str)]
-                # re_match
-                outputs = outputs.strip()
-
-                return outputs
         
         if '<image>' in conversation[0]['content'] and test_compress:
             outputs = tokenizer.decode(output_ids[0, input_ids.unsqueeze(0).cuda().shape[1]:])
@@ -1035,3 +1024,13 @@ class DeepseekOCRForCausalLM(DeepseekV2ForCausalLM):
                 plt.close()
 
             result.save(f"{output_path}/result_with_boxes.jpg")
+
+        if '<image>' in conversation[0]['content'] and eval_mode:
+            outputs = tokenizer.decode(output_ids[0, input_ids.unsqueeze(0).cuda().shape[1]:])
+            stop_str = '<｜end▁of▁sentence｜>'
+            if outputs.endswith(stop_str):
+                outputs = outputs[:-len(stop_str)]
+            # re_match
+            outputs = outputs.strip()
+
+            return outputs