[Engine][DataProcessor] fix decode token (PaddlePaddle#7102)

Author: zhuangzhuang12

fastdeploy/engine/common_engine.py

@@ -1935,6 +1935,11 @@ def _decode_token(self, token_ids, req_id, is_end):
                 token_ids = cum_tokens[prefix_offset:read_offset]
             else:
                 token_ids = []
+
+            if is_end and delta_text == "" and len(cum_tokens) > 0:
+                read_offset = self.data_processor.decode_status[req_id][1]
+                token_ids = cum_tokens[read_offset:]
+
             if is_end:
                 del self.data_processor.decode_status[req_id]
         return delta_text, token_ids

tests/engine/test_decode_token.py

@@ -0,0 +1,161 @@
+"""
+Author:
+Date: 2026-03-31 10:40:18
+LastEditors:
+LastEditTime: 2026-04-01 11:00:47
+FilePath: /fastdeploy/test_decode_token.py
+"""
+
+"""
+# Copyright (c) 2025  PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+
+import os
+import sys
+import unittest
+from unittest.mock import MagicMock, patch
+
+sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", ".."))
+
+from fastdeploy.engine.common_engine import EngineService
+from fastdeploy.utils import envs
+
+
+def _make_mock_ids2tokens(decode_status, undecoded_tokens=None):
+    """
+    Create a mock ids2tokens that simulates incremental decoding behavior.
+
+    Simulates the non-HF path of DataProcessor.ids2tokens:
+    - decode_status[task_id] = [prefix_offset, read_offset, cumulative_token_ids, cumulative_text]
+    - Returns (delta_text, previous_token_ids, previous_texts)
+
+    Args:
+        decode_status: shared dict for tracking decode state
+        undecoded_tokens: set of token IDs that produce no visible text (delta_text=""),
+                          simulating tokens that cannot be decoded incrementally.
+                          read_offset will NOT advance for these tokens.
+    """
+    undecoded_tokens = undecoded_tokens or set()
+
+    # Simple token->char mapping for testing
+    token_map = {
+        1000: "你",
+        1001: "好",
+    }
+
+    def mock_ids2tokens(token_ids, task_id):
+        if task_id not in decode_status:
+            decode_status[task_id] = [0, 0, [], ""]
+
+        previous_token_ids = list(decode_status[task_id][2])
+        previous_texts = decode_status[task_id][3]
+
+        # Append new tokens to cumulative list
+        decode_status[task_id][2] += token_ids
+
+        # Check if all new tokens are "undecoded" (produce no visible text)
+        all_undecoded = all(tid in undecoded_tokens for tid in token_ids) if token_ids else True
+
+        if all_undecoded and token_ids:
+            # These tokens can't be decoded yet - don't advance read_offset
+            delta_text = ""
+        else:
+            # Normal decoding
+            delta_text = ""
+            for tid in token_ids:
+                delta_text += token_map.get(tid, f"<{tid}>")
+
+            if token_ids:
+                # Only advance offsets when there are actual tokens
+                cum_len = len(decode_status[task_id][2])
+                decode_status[task_id][0] = max(0, cum_len - 1)  # prefix_offset
+                decode_status[task_id][1] = cum_len  # read_offset
+                decode_status[task_id][3] += delta_text
+
+        return delta_text, previous_token_ids, previous_texts
+
+    return mock_ids2tokens
+
+
+class TestDecodeToken(unittest.TestCase):
+    """Test case for _decode_token method with mocked tokenizer"""
+
+    def setUp(self):
+        self.req_id = "test_req_123"
+        self._setup_engine()
+
+    def _setup_engine(self, undecoded_tokens=None):
+        self.decode_status = {}
+
+        self.data_processor = MagicMock()
+        self.data_processor.decode_status = self.decode_status
+        self.data_processor.ids2tokens = _make_mock_ids2tokens(self.decode_status, undecoded_tokens)
+
+        self.engine = MagicMock(spec=EngineService)
+        self.engine.data_processor = self.data_processor
+        self.engine._decode_token = EngineService._decode_token.__get__(self.engine, EngineService)
+
+        # Common init for decode_status
+        self.decode_status[self.req_id] = [0, 0, [], ""]
+
+    def _assert_cleaned_up(self):
+        self.assertNotIn(self.req_id, self.data_processor.decode_status)
+
+    def test_empty_end(self):
+        """Empty token_ids with is_end=True should return empty and cleanup"""
+        with patch.object(envs, "FD_ENABLE_RETURN_TEXT", True):
+            delta_text, returned_tokens = self.engine._decode_token([], self.req_id, is_end=True)
+            self.assertEqual(delta_text, "")
+            self.assertEqual(returned_tokens, [])
+            self._assert_cleaned_up()
+
+    def test_incremental_decoding_and_cleanup(self):
+        """Tokens added in multiple steps should decode correctly and cleanup at end"""
+        with patch.object(envs, "FD_ENABLE_RETURN_TEXT", True):
+            for token_id in [1000, 1001]:  # "你", "好"
+                delta_text, _ = self.engine._decode_token([token_id], self.req_id, is_end=False)
+                self.assertTrue(len(delta_text) > 0)
+
+            delta_text, _ = self.engine._decode_token([], self.req_id, is_end=True)
+            self._assert_cleaned_up()
+
+    def test_undecoded_tokens_on_end(self):
+        """Test that tokens which produce no visible text during streaming
+        are force-decoded when is_end=True"""
+        # Re-setup with 109584 as an undecoded token (produces no delta_text during streaming)
+        self._setup_engine(undecoded_tokens={109584})
+        self.decode_status[self.req_id] = [0, 0, [], ""]
+
+        with patch.object(envs, "FD_ENABLE_RETURN_TEXT", True), patch.dict(os.environ, {"DEBUG_DECODE": "1"}):
+            all_delta = ""
+
+            delta_text, _ = self.engine._decode_token([109584], self.req_id, is_end=False)
+            all_delta += delta_text
+
+            # Now end the stream - force decode should recover any remaining text
+            delta_end, _ = self.engine._decode_token([109584], self.req_id, is_end=False)
+            all_delta += delta_end
+            delta_end, _ = self.engine._decode_token([109584], self.req_id, is_end=False)
+            all_delta += delta_end
+            delta_end, token_ids = self.engine._decode_token([], self.req_id, is_end=True)
+            all_delta += delta_end
+
+            # The full text must be recovered either during streaming or at end
+            self.assertEqual(token_ids, [109584, 109584, 109584])
+            self._assert_cleaned_up()
+
+
+if __name__ == "__main__":
+    unittest.main()