[Bugfix] Allow prefill of assistant response when using mistral_common

[sasha0552]

The mistral_common tokenizer supports prefilling of an assistant message, when prefix=True is used in the last message (role="assistant" only).

However, continue_final_message=False makes no sense for the mistral_common tokenizer because ChatCompletionRequest cannot have a last message with role="assistant" without prefix=True.

Also, AFAIK, OpenAI uses continue_final_message=True-like behavior (continues the last assistant message), so I think this behavior should be replicated in vLLM (potentially replacing add_generation_prompt/continue_final_message in chat completion endpoint?), but that's beyond the scope of this PR. (I'm only fixing the error that vLLM gives when an application sends a message chain in which the last message is an assistant.)

(Also, warnings about add_generation_prompt/continue_final_message when using mistral_common are annoying and should only be logged once (or don't logged at all), but that too is beyond the scope of this PR).

[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]

@patrickvonplaten can you verify whether this is reasonable? I am not familiar with using Mistral tokenizer.

Just a minor nit:

[patrickvonplaten]

Nice catch! If the last message is an assistant message we should indeed set prefix=True