Feature/vllm/input embedding completion api

[Nan2018]

adds support for passing prompt_embeds as b64 encoded bytes to the completions api.

Start the server with

VLLM_USE_V1=0 vllm serve HuggingFaceH4/zephyr-7b-beta

query example:

url = "http://localhost:8000/v1/completions"

prompt_embeds = []
for input_embeds in inputs_embeds: # inputs_embeds is list of 2d tensors of shape (seq_len, embed_dim)
    buff = io.BytesIO()
    torch.save(input_embeds.detach().cpu(), buff)
    prompt_embeds.append(b64encode(buff.getvalue()).decode("utf-8"))

resps = requests.post(
    url,
    json={
        "model": request_model,
        "prompt_embeds": prompt_embeds,
    },
)

# or with openai client

completions = client.completions.create(
    model=request_model,
    prompt="",
    extra_body={"prompt_embeds": prompt_embeds},
)

Note

this does not work with lora or prompt adapters

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

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 either: Add ready label to the PR or enable auto-merge.

🚀

[CandiedCode]

Now that #15428 has landed, this will be great to use that functionality 🙏🏽 . @DarkLight1337 I'm curious on what your thoughts are on the issues @Nan2018 is having.