[Bugfix][Frontend] Detokenize echoed prompts after truncation (#61876)

Signed-off-by: Cy <15118312+cycancode@users.noreply.github.com>
Co-authored-by: Cy <15118312+cycancode@users.noreply.github.com>

tests/renderers/test_completions.py

@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+import asyncio
 import io
 from collections.abc import Sequence
 from dataclasses import dataclass
@@ -376,6 +377,32 @@ class TestRenderPrompt:
         assert results[0]["prompt_token_ids"] == tokens
         assert results[0]["prompt"] == "[1, 2, 3, 4]"
 
+    @pytest.mark.parametrize(
+        ("truncation_side", "expected"),
+        [("left", [106, 107, 108, 109]), ("right", [100, 101, 102, 103])],
+    )
+    def test_detokenization_matches_truncated_tokens(self, truncation_side, expected):
+        """Echoed prompt text must match the truncated tokens sent to the engine."""
+        renderer = _build_renderer(MockModelConfig())
+        params = TokenizeParams(
+            max_total_tokens=100,
+            truncate_prompt_tokens=4,
+            truncation_side=truncation_side,
+            needs_detokenization=True,
+        )
+
+        def render():
+            return renderer.render_prompts(
+                _preprocess_prompt(renderer.model_config, list(range(100, 110)))
+            )
+
+        results = renderer.tokenize_prompts(render(), params)
+        async_results = asyncio.run(renderer.tokenize_prompts_async(render(), params))
+
+        for result in (*results, *async_results):
+            assert result["prompt_token_ids"] == expected
+            assert result["prompt"] == str(expected)
+
     def test_explicit_side_tokenizer_unbounded(self):
         renderer = _build_renderer(MockModelConfig())
 

vllm/renderers/base.py

@@ -687,13 +687,17 @@ class BaseRenderer(ABC, Generic[_T]):
             prompt = self._tokenize_prompt(prompt, params)
             prompt = self._apply_prompt_char_offset(prompt, prompt_char_offset)
 
+        prompt = params.apply_post_tokenization(self.tokenizer, prompt)  # type: ignore[arg-type]
+
+        # Detokenize after truncation so the text matches the tokens sent
+        # to the engine.
         if params.needs_detokenization and "prompt" not in prompt:
             if "prompt_token_ids" not in prompt:
                 raise RuntimeError("Cannot run detokenization on embeddings")
 
             prompt = self._detokenize_prompt(prompt)  # type: ignore[arg-type]
 
-        return params.apply_post_tokenization(self.tokenizer, prompt)  # type: ignore[arg-type]
+        return prompt
 
     @overload
     async def _tokenize_singleton_prompt_async(
@@ -727,13 +731,17 @@ class BaseRenderer(ABC, Generic[_T]):
             prompt = await self._tokenize_prompt_async(prompt, params)
             prompt = self._apply_prompt_char_offset(prompt, prompt_char_offset)
 
+        prompt = params.apply_post_tokenization(self.tokenizer, prompt)  # type: ignore[arg-type]
+
+        # Detokenize after truncation so the text matches the tokens sent
+        # to the engine.
         if params.needs_detokenization and "prompt" not in prompt:
             if "prompt_token_ids" not in prompt:
                 raise RuntimeError("Cannot run detokenization on embeddings")
 
             prompt = await self._detokenize_prompt_async(prompt)  # type: ignore[arg-type]
 
-        return params.apply_post_tokenization(self.tokenizer, prompt)  # type: ignore[arg-type]
+        return prompt
 
     def _tokenize_enc_dec_prompt(
         self,