LLM.ts

import { type StreamCallbackResponse } from "../callbacks/CallbackManager.js";

import type { LLMOptions } from "portkey-ai";
import { getCallbackManager } from "../internal/settings/CallbackManager.js";
import { BaseLLM } from "./base.js";
import type { PortkeySession } from "./portkey.js";
import { getPortkeySession } from "./portkey.js";
import { ReplicateSession } from "./replicate_ai.js";
import type {
  ChatMessage,
  ChatResponse,
  ChatResponseChunk,
  LLMChatParamsNonStreaming,
  LLMChatParamsStreaming,
  LLMMetadata,
  MessageType,
} from "./types.js";
import { extractText, wrapLLMEvent } from "./utils.js";

export const ALL_AVAILABLE_LLAMADEUCE_MODELS = {
  "Llama-2-70b-chat-old": {
    contextWindow: 4096,
    replicateApi:
      "replicate/llama70b-v2-chat:e951f18578850b652510200860fc4ea62b3b16fac280f83ff32282f87bbd2e48",
    //^ Previous 70b model. This is also actually 4 bit, although not exllama.
  },
  "Llama-2-70b-chat-4bit": {
    contextWindow: 4096,
    replicateApi:
      "meta/llama-2-70b-chat:02e509c789964a7ea8736978a43525956ef40397be9033abf9fd2badfe68c9e3",
    //^ Model is based off of exllama 4bit.
  },
  "Llama-2-13b-chat-old": {
    contextWindow: 4096,
    replicateApi:
      "a16z-infra/llama13b-v2-chat:df7690f1994d94e96ad9d568eac121aecf50684a0b0963b25a41cc40061269e5",
  },
  //^ Last known good 13b non-quantized model. In future versions they add the SYS and INST tags themselves
  "Llama-2-13b-chat-4bit": {
    contextWindow: 4096,
    replicateApi:
      "meta/llama-2-13b-chat:f4e2de70d66816a838a89eeeb621910adffb0dd0baba3976c96980970978018d",
  },
  "Llama-2-7b-chat-old": {
    contextWindow: 4096,
    replicateApi:
      "a16z-infra/llama7b-v2-chat:4f0a4744c7295c024a1de15e1a63c880d3da035fa1f49bfd344fe076074c8eea",
    //^ Last (somewhat) known good 7b non-quantized model. In future versions they add the SYS and INST
    // tags themselves
    // https://github.com/replicate/cog-llama-template/commit/fa5ce83912cf82fc2b9c01a4e9dc9bff6f2ef137
    // Problem is that they fix the max_new_tokens issue in the same commit. :-(
  },
  "Llama-2-7b-chat-4bit": {
    contextWindow: 4096,
    replicateApi:
      "meta/llama-2-7b-chat:13c3cdee13ee059ab779f0291d29054dab00a47dad8261375654de5540165fb0",
  },
};

export enum DeuceChatStrategy {
  A16Z = "a16z",
  META = "meta",
  METAWBOS = "metawbos",
  //^ This is not exactly right because SentencePiece puts the BOS and EOS token IDs in after tokenization
  // Unfortunately any string only API won't support these properly.
  REPLICATE4BIT = "replicate4bit",
  //^ To satisfy Replicate's 4 bit models' requirements where they also insert some INST tags
  REPLICATE4BITWNEWLINES = "replicate4bitwnewlines",
  //^ Replicate's documentation recommends using newlines: https://replicate.com/blog/how-to-prompt-llama
}

/**
 * Llama2 LLM implementation
 */
export class LlamaDeuce extends BaseLLM {
  model: keyof typeof ALL_AVAILABLE_LLAMADEUCE_MODELS;
  chatStrategy: DeuceChatStrategy;
  temperature: number;
  topP: number;
  maxTokens?: number;
  replicateSession: ReplicateSession;

  constructor(init?: Partial<LlamaDeuce>) {
    super();
    this.model = init?.model ?? "Llama-2-70b-chat-4bit";
    this.chatStrategy =
      init?.chatStrategy ??
      (this.model.endsWith("4bit")
        ? DeuceChatStrategy.REPLICATE4BITWNEWLINES // With the newer Replicate models they do the system message themselves.
        : DeuceChatStrategy.METAWBOS); // With BOS and EOS seems to work best, although they all have problems past a certain point
    this.temperature = init?.temperature ?? 0.1; // minimum temperature is 0.01 for Replicate endpoint
    this.topP = init?.topP ?? 1;
    this.maxTokens =
      init?.maxTokens ??
      ALL_AVAILABLE_LLAMADEUCE_MODELS[this.model].contextWindow; // For Replicate, the default is 500 tokens which is too low.
    this.replicateSession = init?.replicateSession ?? new ReplicateSession();
  }

  get metadata() {
    return {
      model: this.model,
      temperature: this.temperature,
      topP: this.topP,
      maxTokens: this.maxTokens,
      contextWindow: ALL_AVAILABLE_LLAMADEUCE_MODELS[this.model].contextWindow,
      tokenizer: undefined,
    };
  }

  mapMessagesToPrompt(messages: ChatMessage[]) {
    if (this.chatStrategy === DeuceChatStrategy.A16Z) {
      return this.mapMessagesToPromptA16Z(messages);
    } else if (this.chatStrategy === DeuceChatStrategy.META) {
      return this.mapMessagesToPromptMeta(messages);
    } else if (this.chatStrategy === DeuceChatStrategy.METAWBOS) {
      return this.mapMessagesToPromptMeta(messages, { withBos: true });
    } else if (this.chatStrategy === DeuceChatStrategy.REPLICATE4BIT) {
      return this.mapMessagesToPromptMeta(messages, {
        replicate4Bit: true,
        withNewlines: true,
      });
    } else if (this.chatStrategy === DeuceChatStrategy.REPLICATE4BITWNEWLINES) {
      return this.mapMessagesToPromptMeta(messages, {
        replicate4Bit: true,
        withNewlines: true,
      });
    } else {
      return this.mapMessagesToPromptMeta(messages);
    }
  }

  mapMessagesToPromptA16Z(messages: ChatMessage[]) {
    return {
      prompt:
        messages.reduce((acc, message) => {
          return (
            (acc && `${acc}\n\n`) +
            `${this.mapMessageTypeA16Z(message.role)}${message.content}`
          );
        }, "") + "\n\nAssistant:",
      //^ Here we're differing from A16Z by omitting the space. Generally spaces at the end of prompts decrease performance due to tokenization
      systemPrompt: undefined,
    };
  }

  mapMessageTypeA16Z(messageType: MessageType): string {
    switch (messageType) {
      case "user":
        return "User: ";
      case "assistant":
        return "Assistant: ";
      case "system":
        return "";
      default:
        throw new Error("Unsupported LlamaDeuce message type");
    }
  }

  mapMessagesToPromptMeta(
    messages: ChatMessage[],
    opts?: {
      withBos?: boolean;
      replicate4Bit?: boolean;
      withNewlines?: boolean;
    },
  ) {
    const {
      withBos = false,
      replicate4Bit = false,
      withNewlines = false,
    } = opts ?? {};
    const DEFAULT_SYSTEM_PROMPT = `You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.

If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.`;

    const B_SYS = "<<SYS>>\n";
    const E_SYS = "\n<</SYS>>\n\n";
    const B_INST = "[INST]";
    const E_INST = "[/INST]";
    const BOS = "<s>";
    const EOS = "</s>";

    if (messages.length === 0) {
      return { prompt: "", systemPrompt: undefined };
    }

    messages = [...messages]; // so we can use shift without mutating the original array

    let systemPrompt = undefined;
    if (messages[0].role === "system") {
      const systemMessage = messages.shift()!;

      if (replicate4Bit) {
        systemPrompt = systemMessage.content;
      } else {
        const systemStr = `${B_SYS}${systemMessage.content}${E_SYS}`;

        // TS Bug: https://github.com/microsoft/TypeScript/issues/9998
        // @ts-ignore
        if (messages[0].role !== "user") {
          throw new Error(
            "LlamaDeuce: if there is a system message, the second message must be a user message.",
          );
        }

        const userContent = messages[0].content;

        messages[0].content = `${systemStr}${userContent}`;
      }
    } else {
      if (!replicate4Bit) {
        messages[0].content = `${B_SYS}${DEFAULT_SYSTEM_PROMPT}${E_SYS}${messages[0].content}`;
      }
    }

    return {
      prompt: messages.reduce((acc, message, index) => {
        const content = extractText(message.content);
        if (index % 2 === 0) {
          return (
            `${acc}${withBos ? BOS : ""}${B_INST} ${content.trim()} ${E_INST}` +
            (withNewlines ? "\n" : "")
          );
        } else {
          return (
            `${acc} ${content.trim()}` +
            (withNewlines ? "\n" : " ") +
            (withBos ? EOS : "")
          ); // Yes, the EOS comes after the space. This is not a mistake.
        }
      }, ""),
      systemPrompt,
    };
  }

  chat(
    params: LLMChatParamsStreaming,
  ): Promise<AsyncIterable<ChatResponseChunk>>;
  chat(params: LLMChatParamsNonStreaming): Promise<ChatResponse>;
  @wrapLLMEvent
  async chat(
    params: LLMChatParamsNonStreaming | LLMChatParamsStreaming,
  ): Promise<ChatResponse | AsyncIterable<ChatResponseChunk>> {
    const { messages, stream } = params;
    const api = ALL_AVAILABLE_LLAMADEUCE_MODELS[this.model]
      .replicateApi as `${string}/${string}:${string}`;

    const { prompt, systemPrompt } = this.mapMessagesToPrompt(messages);

    const replicateOptions: any = {
      input: {
        prompt,
        system_prompt: systemPrompt,
        temperature: this.temperature,
        top_p: this.topP,
      },
    };

    if (this.model.endsWith("4bit")) {
      replicateOptions.input.max_new_tokens = this.maxTokens;
    } else {
      replicateOptions.input.max_length = this.maxTokens;
    }

    //TODO: Add streaming for this
    if (stream) {
      throw new Error("Streaming not supported for LlamaDeuce");
    }

    //Non-streaming
    const response = await this.replicateSession.replicate.run(
      api,
      replicateOptions,
    );
    return {
      message: {
        content: (response as Array<string>).join("").trimStart(),
        //^ We need to do this because Replicate returns a list of strings (for streaming functionality which is not exposed by the run function)
        role: "assistant",
      },
    };
  }
}

export class Portkey extends BaseLLM {
  apiKey?: string = undefined;
  baseURL?: string = undefined;
  mode?: string = undefined;
  llms?: [LLMOptions] | null = undefined;
  session: PortkeySession;

  constructor(init?: Partial<Portkey>) {
    super();
    this.apiKey = init?.apiKey;
    this.baseURL = init?.baseURL;
    this.mode = init?.mode;
    this.llms = init?.llms;
    this.session = getPortkeySession({
      apiKey: this.apiKey,
      baseURL: this.baseURL,
      llms: this.llms,
      mode: this.mode,
    });
  }

  get metadata(): LLMMetadata {
    throw new Error("metadata not implemented for Portkey");
  }

  chat(
    params: LLMChatParamsStreaming,
  ): Promise<AsyncIterable<ChatResponseChunk>>;
  chat(params: LLMChatParamsNonStreaming): Promise<ChatResponse>;
  @wrapLLMEvent
  async chat(
    params: LLMChatParamsNonStreaming | LLMChatParamsStreaming,
  ): Promise<ChatResponse | AsyncIterable<ChatResponseChunk>> {
    const { messages, stream, additionalChatOptions } = params;
    if (stream) {
      return this.streamChat(messages, additionalChatOptions);
    } else {
      const bodyParams = additionalChatOptions || {};
      const response = await this.session.portkey.chatCompletions.create({
        messages: messages.map((message) => ({
          content: extractText(message.content),
          role: message.role,
        })),
        ...bodyParams,
      });

      const content = response.choices[0].message?.content ?? "";
      const role = response.choices[0].message?.role || "assistant";
      return { message: { content, role: role as MessageType } };
    }
  }

  async *streamChat(
    messages: ChatMessage[],
    params?: Record<string, any>,
  ): AsyncIterable<ChatResponseChunk> {
    const chunkStream = await this.session.portkey.chatCompletions.create({
      messages: messages.map((message) => ({
        content: extractText(message.content),
        role: message.role,
      })),
      ...params,
      stream: true,
    });

    //Indices
    let idx_counter: number = 0;
    for await (const part of chunkStream) {
      //Increment
      part.choices[0].index = idx_counter;
      const is_done: boolean =
        part.choices[0].finish_reason === "stop" ? true : false;
      //onLLMStream Callback

      const stream_callback: StreamCallbackResponse = {
        index: idx_counter,
        isDone: is_done,
        // token: part,
      };
      getCallbackManager().dispatchEvent("stream", stream_callback);

      idx_counter++;

      yield { delta: part.choices[0].delta?.content ?? "" };
    }
    return;
  }
}