Implement chat_vllm()

[hadley]

Fixes #140

@cboettig can you please take a look and let me know if this meets your needs?

[cboettig]

@hadley thanks much and apologies for the delay. I can confirm this is working now. I ran into issues with gorilla, maybe you did too, but was able to test by running other tool-use-enabled models in vllm locally.

e.g. using Groq/Llama-3-Groq-8B-Tool-Use:

#library(elmer)
devtools::load_all()

# vllm is installable as a python module, could have used reticulate
system("vllm serve --dtype=half --enable-auto-tool-choice --tool-call-parser llama3_json  Groq/Llama-3-Groq-8B-Tool-Use")


chat <- chat_vllm(
  base_url = "http://localhost:8000/v1",
  model = "Groq/Llama-3-Groq-8B-Tool-Use",
  api_key = ""
)

tool_rnorm <- tool(
  rnorm,
  "Drawn numbers from a random normal distribution",
  n = type_integer("The number of observations. Must be a positive integer."),
  mean = type_number("The mean value of the distribution."),
  sd = type_number("The standard deviation of the distribution. Must be a non-negative number.")
)
# Then register it
chat$register_tool(tool_rnorm)

# Then ask a question that needs it.
chat$chat("
  Give me five numbers from a random normal distribution of mean 5 and sd 1.
")

Groq/Llama-3-Groq-8B-Tool-Use is available of ollama too so should be easy to compare (though sadly no mpb support yet on vllm but fine if you have an nvidia card handy).

Thanks so much for adding this! Looking forward to testing out in the classroom!

[cboettig]

Maybe this needs to be extended to structured data too? In calling extract_data() with vllm I'm seeing the errors and it looks like strict is still in there, but not sure if it's related. lemme know if that's worth a separate issue or is an easy extension to this.

[hadley]

Oh yeah, I bet I forgot about that.