[Frontend] Add segments to OpenAI Requests

[ruediste]

I extended the format of the completions prompt:

{
  ...
  prompt: [
        { text: "<|fim_prefix|>", split_special_tokens: false },
        { text: prefix, split_special_tokens: true },
        { text: "<|fim_suffix|>", split_special_tokens: false },
        { text: suffix, split_special_tokens: true },
        { text: "<|fim_middle|>", split_special_tokens: false },
      ]
  ...
}

This allows fine control over the generated tokens, protection against prompt injection, and low latency since all of this can be achieved with a single request.

Sample Usage using the javascript openai client:

const response = await openai.completions.create({
      model: config.model,
      prompt: [
        { text: "<|fim_prefix|>", split_special_tokens: false },
        { text: prefix, split_special_tokens: true },
        { text: "<|fim_suffix|>", split_special_tokens: false },
        { text: suffix, split_special_tokens: true },
        { text: "<|fim_middle|>", split_special_tokens: false },
      ] as {
        text: string;
        split_special_tokens: boolean;
      }[] as any,
      stream: true,
      max_tokens: 25,
      temperature: 0,
    });
``1`

[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.

🚀

[mgoin]

Hi @ruediste this interface is a new idea to me and I don't see it in the OpenAI API, is there another server or API that implements this? It would be helpful to have a full e2e example of how to use this feature to demonstrate its intended use.

[ruediste]

A full e2e test is somewhat difficult, as it would in my case include a VS Code extension. But I'm happy to give an example:

Imaging a inline autocomplete extension. The user is editing some code:

// print `Hello World` to the console
console.<FIM>

You would get a prefix of

// print `Hello World` to the console
console.

and the suffix

The normal completions request:

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{ 
    "model": "coder", 
    "prompt": "<|fim_prefix|>// print `Hello World` to the console\nconsole.<|fim_suffix|>\n<|fim_middle|>"
    }'

You get the intended (Qwen 2.5 Coder 0.5B) response: log(\"Hello World!\")

Now imagine the following code:

// print `<|endoftext|>` to the console
console.<FIM>

Request:

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{ 
    "model": "coder", 
    "prompt": "<|fim_prefix|>// print `<|endoftext|>` to the console<|fim_suffix|>\n<|fim_middle|>"
    }'

the response is \nvar name = `Akash`\nconsole.log(name);

With the new API:

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{ 
    "model": "coder", 
    "prompt": [
    	{"text": "<|fim_prefix|>", "split_special_tokens": false},`
  	{"text": "// print `<|endoftext|>` to the console\nconsole.", "split_special_tokens": true},
    	{"text": "<|fim_suffix|>", "split_special_tokens": false},
 	{"text": "\n", "split_special_tokens": true},
    	{"text": "<|fim_middle|>", "split_special_tokens": false}
    ]}'

The response is as expected: log('<|endoftext|>')