2023-12-26 12:49:55 -05:00
|
|
|
# frozen_string_literal: true
|
|
|
|
|
|
|
|
module DiscourseAi
|
|
|
|
module Completions
|
|
|
|
module Endpoints
|
|
|
|
class Vllm < Base
|
2024-07-30 12:44:57 -04:00
|
|
|
def self.can_contact?(model_provider)
|
|
|
|
model_provider == "vllm"
|
2023-12-26 12:49:55 -05:00
|
|
|
end
|
|
|
|
|
2024-01-04 07:53:47 -05:00
|
|
|
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
|
|
|
|
|
2023-12-26 12:49:55 -05:00
|
|
|
def default_options
|
2024-07-30 12:44:57 -04:00
|
|
|
{ max_tokens: 2000, model: llm_model.name }
|
2023-12-26 12:49:55 -05:00
|
|
|
end
|
|
|
|
|
|
|
|
def provider_id
|
|
|
|
AiApiAuditLog::Provider::Vllm
|
|
|
|
end
|
|
|
|
|
|
|
|
private
|
|
|
|
|
|
|
|
def model_uri
|
2024-07-30 12:44:57 -04:00
|
|
|
if llm_model.url.to_s.starts_with?("srv://")
|
2024-08-02 08:52:33 -04:00
|
|
|
service = DiscourseAi::Utils::DnsSrv.lookup(llm_model.url.sub("srv://", ""))
|
2024-05-07 09:02:16 -04:00
|
|
|
api_endpoint = "https://#{service.target}:#{service.port}/v1/chat/completions"
|
2023-12-26 12:49:55 -05:00
|
|
|
else
|
2024-07-30 12:44:57 -04:00
|
|
|
api_endpoint = llm_model.url
|
2023-12-26 12:49:55 -05:00
|
|
|
end
|
2024-07-30 12:44:57 -04:00
|
|
|
|
2023-12-26 12:49:55 -05:00
|
|
|
@uri ||= URI(api_endpoint)
|
|
|
|
end
|
|
|
|
|
2024-08-02 08:52:33 -04:00
|
|
|
def prepare_payload(prompt, model_params, dialect)
|
|
|
|
payload = default_options.merge(model_params).merge(messages: prompt)
|
|
|
|
payload[:stream] = true if @streaming_mode
|
|
|
|
|
|
|
|
payload
|
2023-12-26 12:49:55 -05:00
|
|
|
end
|
|
|
|
|
|
|
|
def prepare_request(payload)
|
|
|
|
headers = { "Referer" => Discourse.base_url, "Content-Type" => "application/json" }
|
2024-01-10 17:23:07 -05:00
|
|
|
|
2024-05-16 08:50:22 -04:00
|
|
|
api_key = llm_model&.api_key || SiteSetting.ai_vllm_api_key
|
|
|
|
headers["X-API-KEY"] = api_key if api_key.present?
|
2024-01-10 17:23:07 -05:00
|
|
|
|
2023-12-26 12:49:55 -05:00
|
|
|
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
|
2024-05-07 09:02:16 -04:00
|
|
|
|
|
|
|
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
|
2023-12-26 12:49:55 -05:00
|
|
|
end
|
|
|
|
end
|
|
|
|
end
|
|
|
|
end
|