Merge pull request #53 from matlab-deep-learning/remote-ollama

Add support for remote ollama

+llms/+internal/callOllamaChatAPI.m

@@ -37,9 +37,13 @@
     nvp.Seed
     nvp.TimeOut
     nvp.StreamFun
+    nvp.Endpoint
 end
 
-URL = "http://localhost:11434/api/chat";
+URL = nvp.Endpoint + "/api/chat";
+if ~startsWith(URL,"http")
+    URL = "http://" + URL;
+end
 
 % The JSON for StopSequences must have an array, and cannot say "stop": "foo".
 % The easiest way to ensure that is to never pass in a scalar …

.github/workflows/ci.yml

@@ -14,19 +14,23 @@ jobs:
         run: |
             # Run the background, there is no way to daemonise at the moment
             ollama serve &
+            # Run a second server to test different endpoint
+            OLLAMA_HOST=127.0.0.1:11435 OLLAMA_MODELS=/tmp/ollama/models ollama serve &
 
             # A short pause is required before the HTTP port is opened
             sleep 5
 
             # This endpoint blocks until ready
             time curl -i http://localhost:11434
+            time curl -i http://localhost:11435
 
             # For debugging, record Ollama version
             ollama --version
 
-      - name: Pull mistral model
+      - name: Pull models
         run: |
             ollama pull mistral
+            OLLAMA_HOST=127.0.0.1:11435 ollama pull qwen2:0.5b
       - name: Set up MATLAB
         uses: matlab-actions/setup-matlab@v2
         with:
@@ -39,6 +43,7 @@ jobs:
             AZURE_OPENAI_DEPLOYMENT: ${{ secrets.AZURE_DEPLOYMENT }}
             AZURE_OPENAI_ENDPOINT: ${{ secrets.AZURE_ENDPOINT }}
             AZURE_OPENAI_API_KEY: ${{ secrets.AZURE_KEY }}
+            SECOND_OLLAMA_ENDPOINT: 127.0.0.1:11435
         uses: matlab-actions/run-tests@v2
         with:
           test-results-junit: test-results/results.xml

.gitignore

@@ -6,3 +6,5 @@ papers_to_read.csv
 data/*
 examples/data/*
 ._*
+.nfs*
+.DS_Store

README.md

@@ -2,7 +2,7 @@
 
 [![Open in MATLAB Online](https://www.mathworks.com/images/responsive/global/open-in-matlab-online.svg)](https://matlab.mathworks.com/open/github/v1?repo=matlab-deep-learning/llms-with-matlab) [![View Large Language Models (LLMs) with MATLAB on File Exchange](https://www.mathworks.com/matlabcentral/images/matlab-file-exchange.svg)](https://www.mathworks.com/matlabcentral/fileexchange/163796-large-language-models-llms-with-matlab) 
 
-This repository contains code to connect MATLAB to the [OpenAI™ Chat Completions API](https://platform.openai.com/docs/guides/text-generation/chat-completions-api) (which powers ChatGPT™), OpenAI Images API (which powers DALL·E™), [Azure® OpenAI Service](https://learn.microsoft.com/en-us/azure/ai-services/openai/), and local [Ollama®](https://ollama.com/) models. This allows you to leverage the natural language processing capabilities of large language models directly within your MATLAB environment.
+This repository contains code to connect MATLAB to the [OpenAI™ Chat Completions API](https://platform.openai.com/docs/guides/text-generation/chat-completions-api) (which powers ChatGPT™), OpenAI Images API (which powers DALL·E™), [Azure® OpenAI Service](https://learn.microsoft.com/en-us/azure/ai-services/openai/), and both local and nonlocal [Ollama®](https://ollama.com/) models. This allows you to leverage the natural language processing capabilities of large language models directly within your MATLAB environment.
 
 ## Requirements
 
@@ -15,7 +15,7 @@ This repository contains code to connect MATLAB to the [OpenAI™ Chat Completio
 
 - For OpenAI connections: An active OpenAI API subscription and API key.
 - For Azure OpenAI Services: An active Azure subscription with OpenAI access, deployment, and API key.
-- For Ollama: A local Ollama installation. Currently, only connections on `localhost` are supported, i.e., Ollama and MATLAB must run on the same machine.
+- For Ollama: An Ollama installation.
 
 ## Setup
 

doc/Ollama.md

@@ -1,6 +1,6 @@
 # Ollama
 
-This repository contains code to connect MATLAB to a local [Ollama®](https://ollama.com) server, running large language models (LLMs).
+This repository contains code to connect MATLAB to an [Ollama®](https://ollama.com) server, running large language models (LLMs).
 
 To use local models with Ollama, you will need to install and start an Ollama server, and “pull” models into it. Please follow the Ollama documentation for details. You should be familiar with the limitations and risks associated with using this technology, and you agree that you shall be solely responsible for full compliance with any terms that may apply to your use of any specific model.
 
@@ -13,7 +13,7 @@ Some of the [LLMs currently supported out of the box on Ollama](https://ollama.c
 - gemma, codegemma
 - command-r
 
-## Establishing a connection to local LLMs using Ollama
+## Establishing a connection to LLMs using Ollama
 
 To create the chat assistant, call `ollamaChat` and specify the LLM you want to use:
 ```matlab
@@ -95,3 +95,10 @@ chat = ollamaChat("mistral", StreamFun=sf);
 txt = generate(chat,"What is Model-Based Design and how is it related to Digital Twin?");
 % Should stream the response token by token
 ```
+
+## Establishing a connection to remote LLMs using Ollama
+
+To connect to a remote Ollama server, use the `Endpoint` name-value pair. Include the server name and port number. Ollama starts on 11434 by default.
+```matlab
+chat = ollamaChat("mistral",Endpoint="ollamaServer:11434");
+```

ollamaChat.m

@@ -64,8 +64,9 @@
 % Copyright 2024 The MathWorks, Inc.
 
     properties
-        Model     (1,1) string
-        TopK (1,1) {mustBeReal,mustBePositive} = Inf
+        Model             (1,1) string
+        Endpoint          (1,1) string
+        TopK              (1,1) {mustBeReal,mustBePositive} = Inf
         TailFreeSamplingZ (1,1) {mustBeReal} = 1
     end
 
@@ -82,6 +83,7 @@
                 nvp.TimeOut                  (1,1) {mustBeReal,mustBePositive} = 120
                 nvp.TailFreeSamplingZ        (1,1) {mustBeReal} = 1
                 nvp.StreamFun                (1,1) {mustBeA(nvp.StreamFun,'function_handle')}
+                nvp.Endpoint                 (1,1) string = "127.0.0.1:11434"
             end
 
             if isfield(nvp,"StreamFun")
@@ -105,6 +107,7 @@
             this.TailFreeSamplingZ = nvp.TailFreeSamplingZ;
             this.StopSequences = nvp.StopSequences;
             this.TimeOut = nvp.TimeOut;
+            this.Endpoint = nvp.Endpoint;
         end
 
         function [text, message, response] = generate(this, messages, nvp)
@@ -147,7 +150,8 @@
                 TailFreeSamplingZ=this.TailFreeSamplingZ,...
                 StopSequences=this.StopSequences, MaxNumTokens=nvp.MaxNumTokens, ...
                 ResponseFormat=this.ResponseFormat,Seed=nvp.Seed, ...
-                TimeOut=this.TimeOut, StreamFun=this.StreamFun);
+                TimeOut=this.TimeOut, StreamFun=this.StreamFun, ...
+                Endpoint=this.Endpoint);
 
             if isfield(response.Body.Data,"error")
                 err = response.Body.Data.error;

tests/tollamaChat.m

@@ -98,7 +98,6 @@ function seedFixesResult(testCase)
             testCase.verifyEqual(response1,response2);
         end
 
-
         function streamFunc(testCase)
             function seen = sf(str)
                 persistent data;
@@ -118,6 +117,19 @@ function streamFunc(testCase)
             testCase.verifyGreaterThan(numel(sf("")), 1);
         end
 
+        function reactToEndpoint(testCase)
+            testCase.assumeTrue(isenv("SECOND_OLLAMA_ENDPOINT"),...
+                "Test point assumes a second Ollama server is running " + ...
+                "and $SECOND_OLLAMA_ENDPOINT points to it.");
+            chat = ollamaChat("qwen2:0.5b",Endpoint=getenv("SECOND_OLLAMA_ENDPOINT"));
+            testCase.verifyWarningFree(@() generate(chat,"dummy"));
+            % also make sure "http://" can be included
+            chat = ollamaChat("qwen2:0.5b",Endpoint="http://" + getenv("SECOND_OLLAMA_ENDPOINT"));
+            response = generate(chat,"some input");
+            testCase.verifyClass(response,'string');
+            testCase.verifyGreaterThan(strlength(response),0);
+        end
+
         function doReturnErrors(testCase)
             testCase.assumeFalse( ...
                 any(startsWith(ollamaChat.models,"abcdefghijklmnop")), ...
@@ -126,7 +138,6 @@ function doReturnErrors(testCase)
             testCase.verifyError(@() generate(chat,"hi!"), "llms:apiReturnedError");
         end
 
-
         function invalidInputsConstructor(testCase, InvalidConstructorInput)
             testCase.verifyError(@() ollamaChat("mistral", InvalidConstructorInput.Input{:}), InvalidConstructorInput.Error);
         end