mirror of
https://github.com/discourse/discourse-ai.git
synced 2025-02-16 00:14:48 +00:00
0634b85a81
* UX: Validations to Llm-backed features (except AI Bot) This change is part of an ongoing effort to prevent enabling a broken feature due to lack of configuration. We also want to explicit which provider we are going to use. For example, Claude models are available through AWS Bedrock and Anthropic, but the configuration differs. Validations are: * You must choose a model before enabling the feature. * You must turn off the feature before setting the model to blank. * You must configure each model settings before being able to select it. * Add provider name to summarization options * vLLM can technically support same models as HF * Check we can talk to the selected model * Check for Bedrock instead of anthropic as a site could have both creds setup
113 lines
3.1 KiB
Ruby
113 lines
3.1 KiB
Ruby
# frozen_string_literal: true
|
|
|
|
module DiscourseAi
|
|
module Completions
|
|
module Endpoints
|
|
class HuggingFace < Base
|
|
class << self
|
|
def can_contact?(endpoint_name, model_name)
|
|
return false unless endpoint_name == "hugging_face"
|
|
|
|
%w[
|
|
StableBeluga2
|
|
Upstage-Llama-2-*-instruct-v2
|
|
Llama2-*-chat-hf
|
|
Llama2-chat-hf
|
|
mistralai/Mixtral-8x7B-Instruct-v0.1
|
|
mistralai/Mistral-7B-Instruct-v0.2
|
|
].include?(model_name)
|
|
end
|
|
|
|
def dependant_setting_names
|
|
%w[ai_hugging_face_api_url]
|
|
end
|
|
|
|
def correctly_configured?(_model_name)
|
|
SiteSetting.ai_hugging_face_api_url.present?
|
|
end
|
|
|
|
def endpoint_name(model_name)
|
|
"Hugging Face - #{model_name}"
|
|
end
|
|
end
|
|
|
|
def default_options
|
|
{ parameters: { repetition_penalty: 1.1, temperature: 0.7, return_full_text: false } }
|
|
end
|
|
|
|
def normalize_model_params(model_params)
|
|
model_params = model_params.dup
|
|
|
|
if model_params[:stop_sequences]
|
|
model_params[:stop] = model_params.delete(:stop_sequences)
|
|
end
|
|
|
|
if model_params[:max_tokens]
|
|
model_params[:max_new_tokens] = model_params.delete(:max_tokens)
|
|
end
|
|
|
|
model_params
|
|
end
|
|
|
|
def provider_id
|
|
AiApiAuditLog::Provider::HuggingFaceTextGeneration
|
|
end
|
|
|
|
private
|
|
|
|
def model_uri
|
|
URI(SiteSetting.ai_hugging_face_api_url)
|
|
end
|
|
|
|
def prepare_payload(prompt, model_params, _dialect)
|
|
default_options
|
|
.merge(inputs: prompt)
|
|
.tap do |payload|
|
|
payload[:parameters].merge!(model_params)
|
|
|
|
token_limit = SiteSetting.ai_hugging_face_token_limit || 4_000
|
|
|
|
payload[:parameters][:max_new_tokens] = token_limit - prompt_size(prompt)
|
|
|
|
payload[:stream] = true if @streaming_mode
|
|
end
|
|
end
|
|
|
|
def prepare_request(payload)
|
|
headers =
|
|
{ "Content-Type" => "application/json" }.tap do |h|
|
|
if SiteSetting.ai_hugging_face_api_key.present?
|
|
h["Authorization"] = "Bearer #{SiteSetting.ai_hugging_face_api_key}"
|
|
end
|
|
end
|
|
|
|
Net::HTTP::Post.new(model_uri, headers).tap { |r| r.body = payload }
|
|
end
|
|
|
|
def extract_completion_from(response_raw)
|
|
parsed = JSON.parse(response_raw, symbolize_names: true)
|
|
|
|
if @streaming_mode
|
|
# Last chunk contains full response, which we already yielded.
|
|
return if parsed.dig(:token, :special)
|
|
|
|
parsed.dig(:token, :text).to_s
|
|
else
|
|
parsed[0][:generated_text].to_s
|
|
end
|
|
end
|
|
|
|
def partials_from(decoded_chunk)
|
|
decoded_chunk
|
|
.split("\n")
|
|
.map do |line|
|
|
data = line.split("data:", 2)[1]
|
|
data&.squish == "[DONE]" ? nil : data
|
|
end
|
|
.compact
|
|
end
|
|
end
|
|
end
|
|
end
|
|
end
|