[Bugfix] Allow --skip-tokenizer-init with echo and return_token_ids (vllm-project#26238)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Signed-off-by: xuebwang-amd <xuebwang@amd.com>

Author: DarkLight1337

tests/entrypoints/openai/test_token_in_token_out.py

@@ -54,7 +54,7 @@ async def test_token_in_token_out_and_logprobs(server):
             prompt=token_ids,
             max_tokens=20,
             temperature=0,
-            echo=False,
+            echo=True,
             extra_body={
                 "return_token_ids": True,
             },

vllm/entrypoints/openai/serving_completion.py

@@ -691,5 +691,6 @@ def _build_render_config(
             truncate_prompt_tokens=request.truncate_prompt_tokens,
             add_special_tokens=request.add_special_tokens,
             cache_salt=request.cache_salt,
-            needs_detokenization=bool(request.echo),
+            needs_detokenization=bool(request.echo
+                                      and not request.return_token_ids),
         )