mirror of
https://github.com/discourse/discourse-ai.git
synced 2025-03-08 18:29:32 +00:00
4f1a3effe0
Both endpoints provide OpenAI-compatible servers. The only difference is that Vllm doesn't support passing tools as a separate parameter. Even if the tool param is supported, it ultimately relies on the model's ability to handle native functions, which is not the case with the models we have today. As a part of this change, we are dropping support for StableBeluga/Llama2 models. They don't have a chat_template, meaning the new API can translate them. These changes let us remove some of our existing dialects and are a first step in our plan to support any LLM by defining them as data-driven concepts. I rewrote the "translate" method to use a template method and extracted the tool support strategies into its classes to simplify the code. Finally, these changes bring support for Ollama when running in dev mode. It only works with Mistral for now, but it will change soon..
97 lines
2.8 KiB
Ruby
97 lines
2.8 KiB
Ruby
# frozen_string_literal: true
|
|
|
|
module DiscourseAi
|
|
module Completions
|
|
module Endpoints
|
|
class Vllm < Base
|
|
class << self
|
|
def can_contact?(endpoint_name, model_name)
|
|
endpoint_name == "vllm" &&
|
|
%w[mistralai/Mixtral-8x7B-Instruct-v0.1 mistralai/Mistral-7B-Instruct-v0.2].include?(
|
|
model_name,
|
|
)
|
|
end
|
|
|
|
def dependant_setting_names
|
|
%w[ai_vllm_endpoint_srv ai_vllm_endpoint]
|
|
end
|
|
|
|
def correctly_configured?(_model_name)
|
|
SiteSetting.ai_vllm_endpoint_srv.present? || SiteSetting.ai_vllm_endpoint.present?
|
|
end
|
|
|
|
def endpoint_name(model_name)
|
|
"vLLM - #{model_name}"
|
|
end
|
|
end
|
|
|
|
def normalize_model_params(model_params)
|
|
model_params = model_params.dup
|
|
|
|
# max_tokens, temperature are already supported
|
|
if model_params[:stop_sequences]
|
|
model_params[:stop] = model_params.delete(:stop_sequences)
|
|
end
|
|
|
|
model_params
|
|
end
|
|
|
|
def default_options
|
|
{ max_tokens: 2000, model: model }
|
|
end
|
|
|
|
def provider_id
|
|
AiApiAuditLog::Provider::Vllm
|
|
end
|
|
|
|
private
|
|
|
|
def model_uri
|
|
service = DiscourseAi::Utils::DnsSrv.lookup(SiteSetting.ai_vllm_endpoint_srv)
|
|
if service.present?
|
|
api_endpoint = "https://#{service.target}:#{service.port}/v1/chat/completions"
|
|
else
|
|
api_endpoint = "#{SiteSetting.ai_vllm_endpoint}/v1/chat/completions"
|
|
end
|
|
@uri ||= URI(api_endpoint)
|
|
end
|
|
|
|
def prepare_payload(prompt, model_params, _dialect)
|
|
default_options
|
|
.merge(model_params)
|
|
.merge(messages: prompt)
|
|
.tap { |payload| payload[:stream] = true if @streaming_mode }
|
|
end
|
|
|
|
def prepare_request(payload)
|
|
headers = { "Referer" => Discourse.base_url, "Content-Type" => "application/json" }
|
|
|
|
headers["X-API-KEY"] = SiteSetting.ai_vllm_api_key if SiteSetting.ai_vllm_api_key.present?
|
|
|
|
Net::HTTP::Post.new(model_uri, headers).tap { |r| r.body = payload }
|
|
end
|
|
|
|
def partials_from(decoded_chunk)
|
|
decoded_chunk
|
|
.split("\n")
|
|
.map do |line|
|
|
data = line.split("data: ", 2)[1]
|
|
data == "[DONE]" ? nil : data
|
|
end
|
|
.compact
|
|
end
|
|
|
|
def extract_completion_from(response_raw)
|
|
parsed = JSON.parse(response_raw, symbolize_names: true).dig(:choices, 0)
|
|
# half a line sent here
|
|
return if !parsed
|
|
|
|
response_h = @streaming_mode ? parsed.dig(:delta) : parsed.dig(:message)
|
|
|
|
response_h.dig(:content)
|
|
end
|
|
end
|
|
end
|
|
end
|
|
end
|