fix: add missing bos_token to example templates

[toslunar]

This PR updates chat template examples to align with the latest (>=0.4.3) /v1/chat/completions spec on BOS.

Some chat templates don't contain bos_token, especially when it's written before v0.4.3. According to the excellent investigation #9519, the chat serving does not add BOS again, while awkwardness remains in the offline chat interface (to be fixed).

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[DarkLight1337]

Thanks for fixing, can you add some unit tests to ensure that the BOS token isn't added twice in this case?

[toslunar]

Added a test.

Is it OK to place it to a new directory (tests/examples_tests) because it seems the first time to test files in examples? I named it examples_tests because it shouldn't be confused with data directories like tests/data and tests/prompts.

[DarkLight1337]

I would just name the directory tests/examples. Can you add the tests to CI by editing this file?

[DarkLight1337]

Please update the failing tests to consider the BOS token.

[toslunar]

Oops, this is unexpected to me. I'll fix the enumeration of Jinja files.

https://buildkite.com/vllm/fastcheck/builds/10539#0193fc47-aa1b-4af4-8c53-a27c0af8deb1

[2024-12-25T06:56:26Z] ============================= test session starts ==============================
[2024-12-25T06:56:26Z] platform linux -- Python 3.12.8, pytest-8.3.3, pluggy-1.5.0 -- /usr/bin/python3
[2024-12-25T06:56:26Z] cachedir: .pytest_cache
[2024-12-25T06:56:26Z] rootdir: /vllm-workspace
[2024-12-25T06:56:26Z] configfile: pyproject.toml
[2024-12-25T06:56:26Z] plugins: anyio-4.6.2.post1, buildkite-test-collector-0.1.9, asyncio-0.24.0, forked-1.6.0, rerunfailures-14.0, shard-0.1.2
[2024-12-25T06:56:26Z] asyncio: mode=Mode.STRICT, default_loop_scope=None
collected 2 items                                                              
[2024-12-25T06:56:26Z] Running 2 items in this shard: tests/examples/test_jinja.py::test_bos[1-path0], tests/examples/test_jinja.py::test_bos[3-path0]
[2024-12-25T06:56:26Z]
[2024-12-25T06:56:26Z] examples/test_jinja.py::test_bos[1-path0] SKIPPED (got empty paramet...)
[2024-12-25T06:56:26Z] examples/test_jinja.py::test_bos[3-path0] SKIPPED (got empty paramet...)
[2024-12-25T06:56:26Z]
[2024-12-25T06:56:26Z] ============================== 2 skipped in 0.03s ==============================

[DarkLight1337]

I think you also need to update the entrypoints tests to consider the BOS token.

[DarkLight1337]

Actually now that I think more about it... wouldn't it be better to instead update LLM.chat to automatically add BOS tokens? That way our chat templates would remain consistent with HF.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @toslunar.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[toslunar]

I suppose this PR could be incremental and it only resolves the inconsistency among example templates.

Actually now that I think more about it... wouldn't it be better to instead update LLM.chat to automatically add BOS tokens? That way our chat templates would remain consistent with HF.

Sorry. I didn't get the concrete suggestion. Do you mean making LLM.chat add a BOS token if there's not (for backward compatibility)?

• non-chat: HF's tokenizer __call__ defaults to add_special_tokens=True.
• chat: HF's apply_chat_template forces add_special_tokens=False. I use tokenize=False when I need BOS.