feat: pull model from jan hub

Author: louis-jan

cortex-js/src/infrastructure/commanders/usecases/models.cli.usecases.ts

@@ -99,10 +99,9 @@ export class ModelsCliUsecases {
   }
 
   async pullModel(modelId: string) {
-    if (modelId.includes('/')) {
+    if (modelId.includes('/') || modelId.includes(':')) {
       await this.pullHuggingFaceModel(modelId);
     }
-
     const bar = new SingleBar({}, Presets.shades_classic);
     bar.start(100, 0);
     const callback = (progress: number) => {
@@ -112,20 +111,35 @@ export class ModelsCliUsecases {
   }
 
   private async pullHuggingFaceModel(modelId: string) {
-    const data = await this.fetchHuggingFaceRepoData(modelId);
-    const { quantization } = await this.inquirerService.inquirer.prompt({
-      type: 'list',
-      name: 'quantization',
-      message: 'Select quantization',
-      choices: data.siblings
-        .map((e) => e.quantization)
-        .filter((e) => e != null),
-    });
-
-    const sibling = data.siblings
-      .filter((e) => !!e.quantization)
-      .find((e: any) => e.quantization === quantization);
+    let data: HuggingFaceRepoData;
+    if (modelId.includes('/'))
+      data = await this.fetchHuggingFaceRepoData(modelId);
+    else data = await this.fetchJanRepoData(modelId);
+
+    let sibling;
+
+    const listChoices = data.siblings
+      .filter((e) => e.quantization != null)
+      .map((e) => {
+        return {
+          name: e.quantization,
+          value: e.quantization,
+        };
+      });
 
+    if (listChoices.length > 1) {
+      const { quantization } = await this.inquirerService.inquirer.prompt({
+        type: 'list',
+        name: 'quantization',
+        message: 'Select quantization',
+        choices: listChoices,
+      });
+      sibling = data.siblings
+        .filter((e) => !!e.quantization)
+        .find((e: any) => e.quantization === quantization);
+    } else {
+      sibling = data.siblings.find((e) => e.rfilename.includes('.gguf'));
+    }
     if (!sibling) throw 'No expected quantization found';
 
     let stopWord = '';
@@ -141,9 +155,7 @@ export class ModelsCliUsecases {
 
       // @ts-expect-error "tokenizer.ggml.tokens"
       stopWord = metadata['tokenizer.ggml.tokens'][index] ?? '';
-    } catch (err) {
-      console.log('Failed to get stop word: ', err);
-    }
+    } catch (err) {}
 
     const stopWords: string[] = [];
     if (stopWord.length > 0) {
@@ -163,6 +175,7 @@ export class ModelsCliUsecases {
       description: '',
       settings: {
         prompt_template: promptTemplate,
+        llama_model_path: sibling.rfilename,
       },
       parameters: {
         stop: stopWords,
@@ -209,6 +222,59 @@ export class ModelsCliUsecases {
     }
   }
 
+  private async fetchJanRepoData(modelId: string) {
+    const repo = modelId.split(':')[0];
+    const tree = modelId.split(':')[1];
+    const url = this.toHuggingFaceUrl(`janhq/${repo}`, tree);
+    const res = await fetch(url);
+    const response:
+      | {
+          path: string;
+          size: number;
+        }[]
+      | { error: string } = await res.json();
+
+    if ('error' in response && response.error != null) {
+      throw new Error(response.error);
+    }
+
+    const data: HuggingFaceRepoData = {
+      siblings: Array.isArray(response)
+        ? response.map((e) => {
+            return {
+              rfilename: e.path,
+              downloadUrl: `https://huggingface.co/janhq/${repo}/resolve/${tree}/${e.path}`,
+              fileSize: e.size ?? 0,
+            };
+          })
+        : [],
+      tags: ['gguf'],
+      id: modelId,
+      modelId: modelId,
+      author: 'janhq',
+      sha: '',
+      downloads: 0,
+      lastModified: '',
+      private: false,
+      disabled: false,
+      gated: false,
+      pipeline_tag: 'text-generation',
+      cardData: {},
+      createdAt: '',
+    };
+
+    AllQuantizations.forEach((quantization) => {
+      data.siblings.forEach((sibling: any) => {
+        if (!sibling.quantization && sibling.rfilename.includes(quantization)) {
+          sibling.quantization = quantization;
+        }
+      });
+    });
+
+    data.modelUrl = url;
+    return data;
+  }
+
   private async fetchHuggingFaceRepoData(repoId: string) {
     const sanitizedUrl = this.toHuggingFaceUrl(repoId);
 
@@ -245,24 +311,7 @@ export class ModelsCliUsecases {
     return data;
   }
 
-  private toHuggingFaceUrl(repoId: string): string {
-    try {
-      const url = new URL(`https://huggingface.co/${repoId}`);
-      if (url.host !== 'huggingface.co') {
-        throw `Invalid Hugging Face repo URL: ${repoId}`;
-      }
-
-      const paths = url.pathname.split('/').filter((e) => e.trim().length > 0);
-      if (paths.length < 2) {
-        throw `Invalid Hugging Face repo URL: ${repoId}`;
-      }
-
-      return `${url.origin}/api/models/${paths[0]}/${paths[1]}`;
-    } catch (err) {
-      if (repoId.startsWith('https')) {
-        throw new Error(`Cannot parse url: ${repoId}`);
-      }
-      throw err;
-    }
+  private toHuggingFaceUrl(repoId: string, tree?: string): string {
+    return `https://huggingface.co/api/models/${repoId}${tree ? `/tree/${tree}` : ''}`;
   }
 }

cortex-js/src/infrastructure/commanders/utils/normalize-model-id.ts

@@ -0,0 +1,3 @@
+export const normalizeModelId = (modelId: string): string => {
+  return modelId.replace(':', '%3A');
+};

cortex-js/src/infrastructure/providers/cortex/cortex.provider.ts

@@ -6,6 +6,7 @@ import { Model, ModelSettingParams } from '@/domain/models/model.interface';
 import { HttpService } from '@nestjs/axios';
 import { defaultCortexCppHost, defaultCortexCppPort } from 'constant';
 import { readdirSync } from 'node:fs';
+import { normalizeModelId } from '@/infrastructure/commanders/utils/normalize-model-id';
 
 /**
  * A class that implements the InferenceExtension interface from the @janhq/core package.
@@ -32,7 +33,10 @@ export default class CortexProvider extends OAIEngineExtension {
   ): Promise<void> {
     const modelsContainerDir = this.modelDir();
 
-    const modelFolderFullPath = join(modelsContainerDir, model.id);
+    const modelFolderFullPath = join(
+      modelsContainerDir,
+      normalizeModelId(model.id),
+    );
     const ggufFiles = readdirSync(modelFolderFullPath).filter((file) => {
       return file.endsWith('.gguf');
     });

cortex-js/src/usecases/cortex/cortex.usecases.ts

@@ -32,7 +32,7 @@ export class CortexUsecases {
     );
 
     if (!existsSync(cortexCppPath)) {
-      throw new Error('Cortex binary not found');
+      throw new Error('The engine is not available, please run "cortex init".');
     }
 
     // go up one level to get the binary folder, have to also work on windows

cortex-js/src/usecases/models/models.usecases.ts

@@ -23,6 +23,7 @@ import { ExtensionRepository } from '@/domain/repositories/extension.interface';
 import { EngineExtension } from '@/domain/abstracts/engine.abstract';
 import { HttpService } from '@nestjs/axios';
 import { ModelSettingParamsDto } from '@/infrastructure/dtos/models/model-setting-params.dto';
+import { normalizeModelId } from '@/infrastructure/commanders/utils/normalize-model-id';
 
 @Injectable()
 export class ModelsUsecases {
@@ -106,7 +107,7 @@ export class ModelsUsecases {
       return;
     }
 
-    const modelFolder = join(modelsContainerDir, id);
+    const modelFolder = join(modelsContainerDir, normalizeModelId(id));
 
     return this.modelRepository
       .delete(id)
@@ -205,7 +206,7 @@ export class ModelsUsecases {
       mkdirSync(modelsContainerDir, { recursive: true });
     }
 
-    const modelFolder = join(modelsContainerDir, model.id);
+    const modelFolder = join(modelsContainerDir, normalizeModelId(model.id));
     await promises.mkdir(modelFolder, { recursive: true });
     const destination = join(modelFolder, fileName);