Support/anyscale finetuning

CHANGELOG.md

@@ -0,0 +1,10 @@
+# Changelog
+
+[5/2] Support for Together AI and Bedrock models. Parity with the [Python release](https://github.com/Tanuki/tanuki.py).
+
+[25/1] Initial Typescript Release in line with the [Python release](https://github.com/Tanuki/tanuki.py).
+
+[27/11] Renamed MonkeyPatch to Tanuki, support for [embeddings](https://github.com/monkeypatch/tanuki.py/blob/update_docs/docs/embeddings_support.md) and [function configurability](https://github.com/monkeypatch/tanuki.py/blob/update_docs/docs/function_configurability.md) is released!
+* Use embeddings to integrate Tanuki with downstream RAG implementations using OpenAI Ada-2 model.
+*  Function configurability allows to configure Tanuki function executions to ignore certain implemented aspects (finetuning, data-storage communications) for improved latency and serverless integrations.
+

README.md

@@ -4,13 +4,6 @@ Build LLM-powered apps that get cheaper and faster over time.
 
 ---
 
-## Release
-[25/1] Initial Typescript Release in line with the [Python release](https://github.com/Tanuki/tanuki.py).
-
-[27/11] Renamed MonkeyPatch to Tanuki, support for [embeddings](https://github.com/monkeypatch/tanuki.py/blob/update_docs/docs/embeddings_support.md) and [function configurability](https://github.com/monkeypatch/tanuki.py/blob/update_docs/docs/function_configurability.md) is released!
-* Use embeddings to integrate Tanuki with downstream RAG implementations using OpenAI Ada-2 model.
-*  Function configurability allows to configure Tanuki function executions to ignore certain implemented aspects (finetuning, data-storage communications) for improved latency and serverless integrations.
-
 Join us on [Discord](https://discord.gg/uUzX5DYctk)
 
 ## Contents

docs/anyscale.md

@@ -0,0 +1,43 @@
+# Anyscale models
+
+Tanuki now supports the finetuning for all models accessible by the Anyscale API as student models. Currently out of the box we support the following hosted models for finetuning
+* Llama-2-7b-chat-hf
+* Llama-2-13b-chat-hf
+* Llama-2-70b-chat-hf
+* Mistral-7B-Instruct-v0.1
+
+
+Anyscale models use the OpenAI package so there is no need to install any extra packages. 
+
+To specify custom student models, a configuration flag for the student_model needs to be sent to the `patch` constructor as shown below at the examples section. 
+
+If no student_model is specified, OpenAIs gpt-3.5-turbo-1106 is used as the default student model.
+
+## Setup
+
+Set your Anyscale API key using:
+
+```
+export ANYSCALE_API_KEY=...
+```
+
+## Examples
+
+### Using the Llama-2-7b-chat-hf as the student model
+```python
+@tanuki.patch(student_model = "Llama-2-7b-chat-hf")
+def example_function(input: TypedInput) -> TypedOutput:
+    """(Optional) Include the description of how your function will be used."""
+
+@tanuki.align
+def test_example_function():
+
+    assert example_function(example_typed_input) == example_typed_output
+
+```
+
+To use the other supported student models, the following text handler should be sent in to the student_model attribute at the `@tanuki.patch` decorator
+* To use meta-llama/Llama-2-7b-chat-hf as a student model, student_model = "Llama-2-7b-chat-hf"
+* To use meta-llama/Llama-2-13b-chat-hf as a student model, student_model = "Llama-2-13b-chat-hf"
+* To use meta-llama/Llama-2-70b-chat-hf as a student model, student_model = "Llama-2-70b-chat-hf"
+* To use mistralai/Mistral-7B-Instruct-v0.1 as a student model, student_model = "Mistral-7B-Instruct-v0.1"

package.json

@@ -1,6 +1,6 @@
 {
   "name": "tanuki.ts",
-  "version": "0.1.4-rc-1",
+  "version": "0.2.0",
   "description": "TypeScript client for building LLM-powered applications",
   "main": "./lib/index.js",
   "type": "module",

src/APIManager.ts

@@ -3,19 +3,34 @@ import {
   LLAMA_BEDROCK_PROVIDER,
   TITAN_BEDROCK_PROVIDER,
   TOGETHER_AI_PROVIDER,
+  ANYSCALE_PROVIDER,
 } from './constants';
 import { FinetuneJob } from './models/finetuneJob';
 import { Embedding } from './models/embedding';
 import { BaseModelConfig } from './languageModels/llmConfigs/baseModelConfig';
+import { OpenAIConfig } from './languageModels/llmConfigs/openAIConfig';
+import Buffer from 'buffer';
 
 export interface Finetunable {
-  listFinetuned: (limit: number) => Promise<FinetuneJob[]>;
-  getFinetuned: (jobId: string) => Promise<FinetuneJob>;
-  finetune: (fileBuffer: Buffer, suffix: string) => Promise<FinetuneJob>;
+  listFinetuned: (
+    modelConfig: OpenAIConfig,
+    limit: number,
+    ...args: any[]
+  ) => Promise<FinetuneJob[]>;
+  getFinetuned: (
+    jobId: string,
+    modelConfig: OpenAIConfig
+  ) => Promise<FinetuneJob>;
+  finetune: (
+    fileBuffer: Buffer,
+    suffix: string,
+    modelConfig: OpenAIConfig
+  ) => Promise<FinetuneJob>;
 }
+
 export interface Inferable {
   generate: (
-    model: BaseModelConfig | string,
+    model: BaseModelConfig,
     systemMessage: string,
     prompt: any,
     kwargs: any
@@ -24,7 +39,7 @@ export interface Inferable {
 export interface Embeddable {
   embed: (
     texts: string[],
-    model: BaseModelConfig | string,
+    model: BaseModelConfig,
     kwargs: any
   ) => Promise<Embedding<any>[]>;
 }
@@ -50,9 +65,26 @@ class APIManager {
   }
 
   private async addApiProvider(provider: string): Promise<void> {
-    if (provider === OPENAI_PROVIDER) {
+    if (provider === ANYSCALE_PROVIDER) {
+      const { AnyscaleAPI } = await import('./languageModels/anyscaleAPI');
+      try {
+        this.apiProviders[provider] = new AnyscaleAPI();
+      } catch (e) {
+        throw new Error(
+          `You need to install the openai package to use the Anyscale api provider. 
+                      Please install it with \`pip install openai\``
+        );
+      }
+    } else if (provider === OPENAI_PROVIDER) {
       const { OpenAIAPI } = await import('./languageModels/openAIAPI');
-      this.apiProviders[provider] = new OpenAIAPI();
+      try {
+        this.apiProviders[provider] = new OpenAIAPI();
+      } catch (e) {
+        throw new Error(
+          `You need to install the openai package to use the OpenAI api provider. 
+                        Please install it with \`pip install openai\``
+        );
+      }
     } else if (provider === LLAMA_BEDROCK_PROVIDER) {
       const { LLamaBedrockAPI } = await import(
         './languageModels/llamaBedrockAPI'

src/constants.ts

@@ -3,6 +3,7 @@ import { ClaudeConfig } from './languageModels/llmConfigs/claudeConfig';
 import { LlamaBedrockConfig } from './languageModels/llmConfigs/llamaConfig';
 import { TitanBedrockConfig } from './languageModels/llmConfigs/titanConfig';
 import { TogetherAIConfig } from './languageModels/llmConfigs/togetherAIConfig';
+import { AnyscaleConfig } from './languageModels/llmConfigs/anyscaleConfig';
 export const EXAMPLE_ELEMENT_LIMIT = 1000;
 
 // These represent the file extensions for the symbolic patch and alignment datasets
@@ -36,6 +37,7 @@ export const ENVVAR = 'TANUKI_LOG_DIR';
 
 // default models
 export const DEFAULT_TEACHER_MODEL_NAMES = ['gpt-4', 'gpt-4-32k'];
+
 export const DEFAULT_DISTILLED_MODEL_NAME = 'gpt-3.5-turbo-1106';
 export const DEFAULT_EMBEDDING_MODEL_NAME = 'ada-002';
 
@@ -45,6 +47,8 @@ export const BEDROCK_PROVIDER = 'bedrock';
 export const LLAMA_BEDROCK_PROVIDER = 'llama_bedrock';
 export const TITAN_BEDROCK_PROVIDER = 'aws_titan_bedrock';
 export const TOGETHER_AI_PROVIDER = 'together_ai';
+export const ANYSCALE_PROVIDER = 'anyscale';
+
 // model type strings
 export const TEACHER_MODEL = 'teacher';
 export const DISTILLED_MODEL = 'distillation';
@@ -123,6 +127,27 @@ export const DEFAULT_STUDENT_MODELS = {
   'gpt-3.5-turbo-1106': new OpenAIConfig({
     modelName: '',
     contextLength: 14000,
+    baseModelForSft: 'gpt-3.5-turbo-1106',
+  }),
+  'Llama-2-7b-chat-hf': new AnyscaleConfig({
+    modelName: '',
+    contextLength: 3000,
+    baseModelForSft: 'meta-llama/Llama-2-7b-chat-hf',
+  }),
+  'Llama-2-13b-chat-hf': new AnyscaleConfig({
+    modelName: '',
+    contextLength: 3000,
+    baseModelForSft: 'meta-llama/Llama-2-13b-chat-hf',
+  }),
+  'Llama-2-70b-chat-hf': new AnyscaleConfig({
+    modelName: '',
+    contextLength: 3000,
+    baseModelForSft: 'meta-llama/Llama-2-70b-chat-hf',
+  }),
+  'Mistral-7B-Instruct-v0.1': new AnyscaleConfig({
+    modelName: '',
+    contextLength: 3000,
+    baseModelForSft: 'mistralai/Mistral-7B-Instruct-v0.1',
   }),
 };