build/server/chunks/LlamaCppService-fS5DnT85.js

import { existsSync, mkdirSync, writeFileSync } from 'fs';
import { resolve } from 'path';
import { d as private_env } from './shared-server-49TKSBDM.js';
import dns from 'node:dns';

dns.setDefaultResultOrder("ipv4first");
class LlamaCppService {
  url = "";
  llmParams;
  constructor(params) {
    this.url = params.url;
    this.llmParams = params;
  }
  async health() {
    try {
      const r = await fetch(`${this.url}/health`, {
        method: "GET",
        headers: {
          "Content-Type": "application/json"
        }
      });
      const data = await r.json();
      if (data.status === "ok" || data.status === "error" || data.status === "loading model") {
        return data.status;
      }
    } catch (error) {
      console.log(JSON.parse(JSON.stringify(error)));
    }
    return "unavailable";
  }
  async tokenize(prompt, abortController) {
    const response = await fetch(`${this.url}/tokenize`, {
      method: "POST",
      headers: {
        "Content-Type": "application/json"
      },
      body: JSON.stringify({
        "content": prompt
      }),
      signal: abortController.signal
    });
    if (response.ok) {
      const data = await response.json();
      if (data.tokens) {
        return data.tokens;
      }
    } else if (response.status === 404) {
      console.log("Tokenization endpoint not found (404).");
    } else {
      console.log(`llama.cpp: Failed to tokenize: ${await response.text()}`);
    }
    return null;
  }
  async detokenize(tokens, abortController) {
    const response = await fetch(`${this.url}/detokenize`, {
      method: "POST",
      headers: {
        "Content-Type": "application/json"
      },
      body: JSON.stringify({
        "tokens": tokens
      }),
      signal: abortController.signal
    });
    if (response.ok) {
      const data = await response.json();
      if (data.content) {
        return data.content.trim();
      }
    } else if (response.status === 404) {
      console.log("Detokenization endpoint not found (404).");
    } else {
      console.log(`Failed to detokenize: ${await response.text()}`);
    }
    return null;
  }
  createRequest(prompt, llmParams) {
    const actualPrompt = this.applyLlmTemplateToPrompt(prompt);
    const request = {
      "stream": true,
      "stop": llmParams.predict_params?.stop || ["</s>", "bot:", "user:"],
      "prompt": actualPrompt
    };
    request["n_predict"] = llmParams.predict_params?.n_predict || -1;
    request["temperature"] = llmParams.predict_params?.temperature || 0;
    if (llmParams.predict_params?.top_k != null) {
      request["top_k"] = Number(llmParams.predict_params.top_k);
    }
    if (llmParams.predict_params?.top_p != null) {
      request["top_p"] = Number(llmParams.predict_params.top_p);
    }
    if (llmParams.predict_params?.min_p != null) {
      request["min_p"] = Number(llmParams.predict_params.min_p);
    }
    if (llmParams.predict_params?.seed != null) {
      request["seed"] = Number(llmParams.predict_params.seed);
    }
    if (llmParams.predict_params?.n_keep != null) {
      request["n_keep"] = Number(llmParams.predict_params.n_keep);
    }
    if (llmParams.predict_params?.cache_prompt != null) {
      request["cache_prompt"] = Boolean(llmParams.predict_params.cache_prompt);
    }
    if (llmParams.predict_params?.repeat_penalty != null) {
      request["repeat_penalty"] = Boolean(llmParams.predict_params.repeat_penalty);
    }
    if (llmParams.predict_params?.repeat_last_n != null) {
      request["repeat_last_n"] = Boolean(llmParams.predict_params.repeat_last_n);
    }
    if (llmParams.predict_params?.frequency_penalty != null) {
      request["frequency_penalty"] = Boolean(llmParams.predict_params.frequency_penalty);
    }
    if (llmParams.predict_params?.presence_penalty != null) {
      request["presence_penalty"] = Boolean(llmParams.predict_params.presence_penalty);
    }
    return request;
  }
  applyLlmTemplateToPrompt(prompt) {
    let actualPrompt = prompt;
    if (this.llmParams.template != void 0) {
      actualPrompt = this.llmParams.template.replace("{{PROMPT}}", actualPrompt);
    }
    return actualPrompt;
  }
  async trimTokenizedText(text, userRequest, { abortController }) {
    let tokens = await this.tokenize(text, abortController);
    let systemPromptTokenCount = 0;
    if (this.llmParams.predict_params?.system_prompt) {
      systemPromptTokenCount = (await this.tokenize(this.llmParams.predict_params?.system_prompt, abortController))?.length || 0;
    }
    const originalTokenCount = tokens?.length || -1;
    const auxTokens = await this.tokenize(this.applyLlmTemplateToPrompt(userRequest), abortController);
    const maxLength = Number(this.llmParams.context) - Number(this.llmParams.predict_params?.n_predict) - (auxTokens?.length ?? 0) - systemPromptTokenCount;
    if (tokens !== null) {
      tokens = tokens.slice(0, maxLength);
      const detokenizedPrompt = await this.detokenize(tokens, abortController);
      if (detokenizedPrompt !== null) {
        text = detokenizedPrompt;
      } else {
        text = text.substring(0, maxLength);
      }
    } else {
      text = text.substring(0, maxLength);
    }
    return { result: text, originalTokenCount, slicedTokenCount: tokens?.length || -1 };
  }
  predict({ abortController }) {
    return async ({ prompt, llmParams }) => {
      const request = this.createRequest(prompt, llmParams);
      console.log(`Predict request. Url: ${this.url}`);
      console.log(`Prompt: ${request.prompt.substring(0, 30)}`);
      let r;
      while (true) {
        r = await fetch(`${this.url}/completion`, {
          method: "POST",
          headers: {
            "Content-Type": "application/json",
            "Accept": "text/event-stream"
          },
          body: JSON.stringify(request),
          signal: abortController.signal
        });
        if (r.status === 404) {
          if (!private_env.LLM_API_404_RETRY_INTERVAL) {
            break;
          }
          console.log(`Received 404, retrying after ${private_env.LLM_API_404_RETRY_INTERVAL} seconds...`);
          await new Promise((resolve2) => setTimeout(resolve2, Number(private_env.LLM_API_404_RETRY_INTERVAL) * 1e3));
        } else {
          break;
        }
      }
      if (!r.ok) {
        throw new Error(`Failed to generate text: ${await r.text()}`);
      }
      const encoder = new TextDecoderStream();
      const reader = await r.body?.pipeThrough(encoder).getReader();
      let t = this;
      return async function* () {
        let stop = false;
        let generatedText = "";
        let tokenId = 0;
        while (!stop) {
          const out = await reader?.read() ?? { done: false, value: void 0 };
          if (out.done) {
            stop = true;
            reader?.cancel();
            t.createLogFile(JSON.stringify({
              "request": prompt,
              "response": generatedText
            }), "llm-service");
            return;
          }
          if (!out.value) {
            stop = true;
            reader?.cancel();
            return;
          }
          let tokenValue = "";
          if (out.value.startsWith("data: ")) {
            try {
              const data = JSON.parse(out.value.slice(6));
              tokenValue = data.content;
            } catch (e) {
            }
          } else if (private_env.LLM_API_VERSION == "v1") {
            tokenValue = out.value;
          }
          generatedText += tokenValue.replace("</s>", "") ?? "";
          yield {
            token: {
              id: tokenId++,
              text: tokenValue.replace("</s>", "") ?? "",
              logprob: 0,
              special: false
            },
            generated_text: null,
            details: null
          };
        }
      }();
    };
  }
  conversation({ abortController }) {
    return async ({ history, llmParams }) => {
      let prompt = history[history.length - 1][0];
      const request = this.createRequest(prompt, llmParams);
      let tokens = await this.tokenize(prompt, abortController);
      if (tokens !== null) {
        tokens = tokens.slice(0, Number(this.llmParams.context) - Number(this.llmParams.predict_params?.n_predict));
        const detokenizedPrompt = await this.detokenize(tokens, abortController);
        if (detokenizedPrompt !== null) {
          prompt = detokenizedPrompt;
        } else {
          prompt = prompt.substring(0, Number(this.llmParams.context) - Number(this.llmParams.predict_params?.n_predict));
        }
      } else {
        prompt = prompt.substring(0, Number(this.llmParams.context) - Number(this.llmParams.predict_params?.n_predict));
      }
      const r = await fetch(`${this.url}/completion`, {
        method: "POST",
        headers: {
          "Content-Type": "application/json",
          "Accept": "text/event-stream"
        },
        body: JSON.stringify(request),
        signal: abortController.signal
      });
      if (!r.ok) {
        throw new Error(`Failed to generate text: ${await r.text()}`);
      }
      const encoder = new TextDecoderStream();
      const reader = r.body?.pipeThrough(encoder).getReader();
      return async function* () {
        let stop = false;
        let generatedText = "";
        let tokenId = 0;
        while (!stop) {
          const out = await reader?.read() ?? { done: false, value: void 0 };
          if (out.done) {
            reader?.cancel();
            return;
          }
          if (!out.value) {
            reader?.cancel();
            return;
          }
          let tokenValue = "";
          if (out.value.startsWith("data: ")) {
            try {
              let data2 = JSON.parse(out.value.slice(6));
              tokenValue = data2.content;
            } catch (e) {
            }
          } else if (private_env.LLM_API_VERSION == "v1") {
            tokenValue = out.value;
          }
          generatedText += tokenValue.replace("</s>", "") ?? "";
          yield {
            token: {
              id: tokenId++,
              text: tokenValue.replace("</s>", "") ?? "",
              logprob: 0,
              special: false
            },
            generated_text: null,
            details: null
          };
        }
      }();
    };
  }
  formatPrompt(history) {
    let prompt = "";
    for (const [userPrompt, botResponse] of history) {
      prompt += ` [INST] ${userPrompt} [/INST]`;
      if (botResponse) {
        prompt += `${botResponse}`;
      }
    }
    return prompt;
  }
  createLogFile(text, namePrefix = "") {
    if (!private_env.LOGS_ROOT_FOLDER) {
      return;
    }
    try {
      const logsDirectory = resolve(private_env.LOGS_ROOT_FOLDER + "/llama");
      if (!existsSync(logsDirectory)) {
        mkdirSync(logsDirectory, {
          recursive: true
        });
      }
      const timestamp = (/* @__PURE__ */ new Date()).toISOString().replace(/[:.]/g, "");
      const logFilePath = resolve(logsDirectory, `${namePrefix}${timestamp}.json`);
      writeFileSync(logFilePath, text);
      console.log(`Log file created: ${logFilePath}`);
    } catch (e) {
      console.log(`Failed to create log file in llama service`);
      console.log(e);
    }
  }
}

export { LlamaCppService as L };
//# sourceMappingURL=LlamaCppService-fS5DnT85.js.map