[Bugfix] Handle <|tool_call|> token in granite tool parser

[tjohnson31415]

Small tweak to the granite tool parser to support handle the <|tool_call|> token that can be generated, eg.

<|start of role|>user<|end of role|>What is temperature in Boston?<|end of text|>
<|start of role|>assistant<|end of role|><|tool call|>[{“name”: “get temp”, ...}]<|end of text|>
<|start of role|>tool response<|end of role|>{“temp”: 20.5, “unit”: “C”}<|end of text|>

REF: Table 3 on page 14 in https://github.com/ibm-granite/granite-3.0-language-models/blob/main/paper.pdf

Supporting this matters if the request has skip_special_tokens=False. Currently, the parser will not parse out the tool call when the output starts with <|tool_call|>.

I have also noticed interesting behavior between the granite model and llama w.r.t handling the tool call special token. With meta-llama/Llama-3.1-8B-Instruct, including skip_special_tokens=True on the vLLM request does not skip generation of the <|python_tag|> special token.

cc: @maxdebayser

[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!