Add llmman as a self-hosted host for chat and embedding providers (#947)

Co-authored-by: Thorsten Sommer <SommerEngineering@users.noreply.github.com>

app/MindWork AI Studio/Provider/LLMProvidersExtensions.cs

@@ -557,21 +557,7 @@ public static class LLMProvidersExtensions
     public static bool CanLoadModels(this LLMProviders provider, Host host, string? apiKey)
     {
         if (provider is LLMProviders.SELF_HOSTED)
-        {
-            switch (host)
-            {
-                case Host.NONE:
-                case Host.WHISPER_CPP:
-                default:
-                    return false;
-
-                case Host.LLAMA_CPP:
-                case Host.OLLAMA:
-                case Host.LM_STUDIO:
-                case Host.VLLM:
-                    return true;
-            }
-        }
+            return host.IsModelListSupported();
 
         if(provider is LLMProviders.NONE)
             return false;

app/MindWork AI Studio/Provider/Reasoning/ReasoningDispatcher.cs

@@ -125,6 +125,13 @@ public static class ReasoningDispatcher
 
             Host.VLLM => [ReasoningDialect.OPEN_AI_COMPATIBLE, ReasoningDialect.REASONING_EFFORT, ReasoningDialect.QWEN_THINKING, ReasoningDialect.GOOGLE_THINKING, ReasoningDialect.VLLM],
 
+            //
+            // llmman starts llama.cpp for GGUF models and vLLM, SGLang, or MLX for the others, and
+            // hands their fields through unchanged. Which engine serves a model is up to llmman, so
+            // it is read in everything llama.cpp and vLLM are read in:
+            //
+            Host.LLMMAN => [ReasoningDialect.OPEN_AI_COMPATIBLE, ReasoningDialect.REASONING_EFFORT, ReasoningDialect.QWEN_THINKING, ReasoningDialect.GOOGLE_THINKING, ReasoningDialect.LLAMA_CPP, ReasoningDialect.VLLM],
+
             _ => EVERYTHING_A_GATEWAY_MIGHT_SERVE,
         },
 

app/MindWork AI Studio/Provider/SelfHosted/Host.cs

@@ -9,4 +9,5 @@ public enum Host
     WHISPER_CPP,
     OLLAMA,
     VLLM,
+    LLMMAN,
 }

app/MindWork AI Studio/Provider/SelfHosted/HostExtensions.cs

@@ -11,6 +11,7 @@ public static class HostExtensions
         Host.WHISPER_CPP => "whisper.cpp",
         Host.OLLAMA => "ollama",
         Host.VLLM => "vLLM",
+        Host.LLMMAN => "llmman",
 
         _ => "Unknown",
     };
@@ -47,6 +48,7 @@ public static class HostExtensions
             case Host.VLLM:
             case Host.LM_STUDIO:
             case Host.LLAMA_CPP:
+            case Host.LLMMAN:
                 return true;
         }
     }
@@ -58,6 +60,7 @@ public static class HostExtensions
             case Host.LM_STUDIO:
             case Host.OLLAMA:
             case Host.VLLM:
+            case Host.LLMMAN:
                 return true;
             
             default:
@@ -73,12 +76,44 @@ public static class HostExtensions
             case Host.OLLAMA:
             case Host.VLLM:
             case Host.WHISPER_CPP:
+            case Host.LLMMAN:
                 return true;
-            
+
             default:
             case Host.LM_STUDIO:
             case Host.LLAMA_CPP:
                 return false;
         }
     }
+
+    /// <summary>
+    /// Whether the engine answers the models route with the models it serves.
+    /// </summary>
+    /// <remarks>
+    /// This is the only place which says so. The provider asks it before it loads the models of a
+    /// chat, embedding, or transcription list, and the dialogs ask it before they offer to reload
+    /// them; a new engine therefore needs no more than its line here. whisper.cpp is not among
+    /// them: it serves the one model it was started with and has no route to list it, so the
+    /// provider offers that model under a fixed name instead. llama.cpp is, even though builds
+    /// older than the route answer it with 404; the provider handles those builds itself.
+    /// </remarks>
+    /// <param name="host">The engine.</param>
+    /// <returns>True when the models can be loaded from the engine; otherwise, false.</returns>
+    public static bool IsModelListSupported(this Host host)
+    {
+        switch (host)
+        {
+            case Host.LLAMA_CPP:
+            case Host.OLLAMA:
+            case Host.LM_STUDIO:
+            case Host.VLLM:
+            case Host.LLMMAN:
+                return true;
+
+            default:
+            case Host.NONE:
+            case Host.WHISPER_CPP:
+                return false;
+        }
+    }
 }

app/MindWork AI Studio/Provider/SelfHosted/Model.cs

@@ -8,8 +8,8 @@ namespace AIStudio.Provider.SelfHosted;
 /// <remarks>
 /// The context window is vLLM's addition to that route: it reports the window the operator started
 /// the engine with, which is the one number no rule about the weights could ever know. Ollama,
-/// LM Studio, and llama.cpp answer the same route without it, so it stays unknown there instead of
-/// being guessed.
+/// LM Studio, llama.cpp, and llmman answer the same route without it, so it stays unknown there
+/// instead of being guessed.
 ///
 /// vLLM calls that field max_model_len, which reads like a limit on the model rather than on a
 /// conversation. The wire keeps their spelling, and this record says what the number means, so that

app/MindWork AI Studio/Provider/SelfHosted/ProviderSelfHosted.cs

@@ -25,7 +25,7 @@ public sealed class ProviderSelfHosted(Host host, string hostname) : BaseProvide
     public override string InstanceName { get; set; } = "Self-hosted";
 
     /// <inheritdoc />
-    public override bool HasModelLoadingCapability => host is Host.OLLAMA or Host.LM_STUDIO or Host.VLLM or Host.LLAMA_CPP;
+    public override bool HasModelLoadingCapability => host.IsModelListSupported();
     
     /// <inheritdoc />
     public override async IAsyncEnumerable<ContentStreamChunk> StreamChatCompletion(Provider.Model chatModel, ChatThread chatThread, SettingsManager settingsManager, [EnumeratorCancellation] CancellationToken token = default)
@@ -40,7 +40,7 @@ public sealed class ProviderSelfHosted(Host host, string hostname) : BaseProvide
                            {
                                // Build the list of messages. The image format depends on the host:
                                // - Ollama uses the direct image URL format: { "type": "image_url", "image_url": "data:..." }
-                               // - LM Studio, vLLM, and llama.cpp use the nested image URL format: { "type": "image_url", "image_url": { "url": "data:..." } }
+                               // - LM Studio, vLLM, llama.cpp, and llmman use the nested image URL format: { "type": "image_url", "image_url": { "url": "data:..." } }
                                var messages = host switch
                                {
                                    Host.OLLAMA => await chatThread.Blocks.BuildMessagesUsingDirectImageUrlAsync(this.CreateSettingsProvider(effectiveChatModel)),
@@ -94,22 +94,17 @@ public sealed class ProviderSelfHosted(Host host, string hostname) : BaseProvide
     {
         try
         {
-            switch (host)
+            if (host is Host.LLAMA_CPP)
+                return await this.LoadLlamaCppTextModels(apiKeyProvisional, token);
+
+            if (!host.IsChatSupported() || !host.IsModelListSupported())
+                return ModelLoadResult.FromModels([]);
+
+            var result = await this.LoadModels(SecretStoreType.LLM_PROVIDER, apiKeyProvisional, token);
+            return result with
             {
-                case Host.LLAMA_CPP:
-                    return await this.LoadLlamaCppTextModels(apiKeyProvisional, token);
-
-                case Host.LM_STUDIO:
-                case Host.OLLAMA:
-                case Host.VLLM:
-                    var result = await this.LoadModels(SecretStoreType.LLM_PROVIDER, apiKeyProvisional, token);
-                    return result with
-                    {
-                        Models = [..result.Models.Where(model => model.IsChatModel(this.Provider))]
-                    };
-            }
-
-            return ModelLoadResult.FromModels([]);
+                Models = [..result.Models.Where(model => model.IsChatModel(this.Provider))]
+            };
         }
         catch(Exception e)
         {
@@ -128,19 +123,14 @@ public sealed class ProviderSelfHosted(Host host, string hostname) : BaseProvide
     {
         try
         {
-            switch (host)
-            {
-                case Host.LM_STUDIO:
-                case Host.OLLAMA:
-                case Host.VLLM:
-                    var result = await this.LoadModels(SecretStoreType.EMBEDDING_PROVIDER, apiKeyProvisional, token);
-                    return result with
-                    {
-                        Models = [..result.Models.Where(model => model.IsEmbeddingModel(this.Provider))]
-                    };
-            }
+            if (!host.IsEmbeddingSupported() || !host.IsModelListSupported())
+                return ModelLoadResult.FromModels([]);
 
-            return ModelLoadResult.FromModels([]);
+            var result = await this.LoadModels(SecretStoreType.EMBEDDING_PROVIDER, apiKeyProvisional, token);
+            return result with
+            {
+                Models = [..result.Models.Where(model => model.IsEmbeddingModel(this.Provider))]
+            };
         }
         catch(Exception e)
         {
@@ -154,33 +144,28 @@ public sealed class ProviderSelfHosted(Host host, string hostname) : BaseProvide
     {
         try
         {
-            switch (host)
-            {
-                case Host.WHISPER_CPP:
-                    return ModelLoadResult.FromModels(
-                    [
-                        new Provider.Model("loaded-model", TB("Model as configured by whisper.cpp")),
-                    ]);
-                
-                //
-                // These two answer the models endpoint with everything they serve, and nothing in
-                // that answer says which of them listens. Asking what each model is made for is the
-                // only thing standing between this list and every chat and embedding model of the
-                // installation, which is what it used to hold. An engine running no speech model at
-                // all therefore offers nothing here, and says so, rather than offering models which
-                // would fail the moment audio reaches them.
-                //
-                case Host.OLLAMA:
-                case Host.VLLM:
-                    var result = await this.LoadModels(SecretStoreType.TRANSCRIPTION_PROVIDER, apiKeyProvisional, token);
-                    return result with
-                    {
-                        Models = [..result.Models.Where(model => model.IsTranscriptionModel(this.Provider))]
-                    };
+            if (host is Host.WHISPER_CPP)
+                return ModelLoadResult.FromModels(
+                [
+                    new Provider.Model("loaded-model", TB("Model as configured by whisper.cpp")),
+                ]);
 
-                default:
-                    return ModelLoadResult.FromModels([]);
-            }
+            if (!host.IsTranscriptionSupported() || !host.IsModelListSupported())
+                return ModelLoadResult.FromModels([]);
+
+            //
+            // The engines which get here answer the models endpoint with everything they serve, and
+            // nothing in that answer says which of them listens. Asking what each model is made for
+            // is the only thing standing between this list and every chat and embedding model of
+            // the installation, which is what it used to hold. An engine running no speech model at
+            // all therefore offers nothing here, and says so, rather than offering models which
+            // would fail the moment audio reaches them.
+            //
+            var result = await this.LoadModels(SecretStoreType.TRANSCRIPTION_PROVIDER, apiKeyProvisional, token);
+            return result with
+            {
+                Models = [..result.Models.Where(model => model.IsTranscriptionModel(this.Provider))]
+            };
         }
         catch (Exception e)
         {