Merge branch 'ollama:main' into main

llm/ext_server/server.cpp

@@ -1223,9 +1223,7 @@ struct llama_server_context
 
                 res.result_json = json
                 {
-                    {"id", res.id},
                     {"embedding", std::vector<float>(embd, embd + n_embd)},
-                    {"timings",             slot.get_formated_timings()},
                 };
             }
         }
@@ -3194,41 +3192,17 @@ int main(int argc, char **argv) {
                     prompt = "";
                 }
 
-                if (prompt.size() == 1) {
-                    prompt = prompt[0];
-                }
-
                 // create and queue the task
-                json responses;
-                {
-                    const int id_task = llama.queue_tasks.get_new_id();
-                    llama.queue_results.add_waiting_task_id(id_task);
-                    llama.request_completion(id_task, {{"prompt", prompt}}, true, -1);
+                const int task_id = llama.queue_tasks.get_new_id();
+                llama.queue_results.add_waiting_task_id(task_id);
+                llama.request_completion(task_id, {{"prompt", prompt}}, true, -1);
 
-                    // get the result
-                    task_result result = llama.queue_results.recv(id_task);
-                    llama.queue_results.remove_waiting_task_id(id_task);
-                    if (result.error) {
-                        return res.set_content(result.result_json.dump(), "application/json; charset=utf-8");
-                    }
+                // get the result
+                task_result result = llama.queue_results.recv(task_id);
+                llama.queue_results.remove_waiting_task_id(task_id);
 
-                    responses = result.result_json.value("results", std::vector<json>{result.result_json});
-                    std::sort(responses.begin(), responses.end(), [](const json& a, const json& b) {
-                        return a["id"] < b["id"];
-                    });
-
-                    json embeddings = json::array();
-
-                    int prompt_n = 0;
-                    for (auto & elem : responses) {
-                        embeddings.push_back(elem.at("embedding"));
-                        prompt_n += elem.at("timings").at("prompt_n").get<int>();
-                    }
-
-                    // send the result
-                    json embedding_res = json{{"embedding", embeddings}, {"prompt_n", prompt_n}};
-                    return res.set_content(embedding_res.dump(), "application/json; charset=utf-8");
-                }
+                // send the result
+                return res.set_content(result.result_json.dump(), "application/json; charset=utf-8");
             });
 
     // GG: if I put the main loop inside a thread, it crashes on the first request when build in Debug!?

llm/ggml.go

@@ -157,6 +157,14 @@ type Tensor struct {
 	io.WriterTo `json:"-"`
 }
 
+func (t Tensor) block() (n int) {
+	if _, err := fmt.Sscanf(t.Name, "blk.%d.", &n); err != nil {
+		return -1
+	}
+
+	return
+}
+
 func (t Tensor) blockSize() uint64 {
 	switch t.Kind {
 	case 0, 1, 24, 25, 26, 27, 28, 30: // F32, F16, I8, I16, I32, I64, F64, BF16

llm/gguf.go

@@ -532,15 +532,14 @@ func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
 		}
 	}
 
-	slices.SortFunc(ts, func(a, b Tensor) int {
-		var i, j int
-		if n, err := fmt.Sscanf(a.Name, "blk.%d", &i); err != nil || n != 1 {
-			return cmp.Compare(a.Name, b.Name)
-		} else if n, err := fmt.Sscanf(b.Name, "blk.%d", &j); err != nil || n != 1 {
-			return cmp.Compare(a.Name, b.Name)
+	slices.SortStableFunc(ts, func(a, b Tensor) int {
+		if i, j := a.block(), b.block(); i < 0 && j > 0 {
+			return 1
+		} else if i > 0 && j < 0 {
+			return -1
+		} else {
+			return cmp.Compare(i, j)
 		}
-
-		return cmp.Compare(i, j)
 	})
 
 	var s uint64

llm/server.go

@@ -33,7 +33,7 @@ type LlamaServer interface {
 	Ping(ctx context.Context) error
 	WaitUntilRunning(ctx context.Context) error
 	Completion(ctx context.Context, req CompletionRequest, fn func(CompletionResponse)) error
-	Embed(ctx context.Context, input []string) (*EmbedResponse, error)
+	Embedding(ctx context.Context, input string) ([]float32, error)
 	Tokenize(ctx context.Context, content string) ([]int, error)
 	Detokenize(ctx context.Context, tokens []int) (string, error)
 	Close() error
@@ -125,8 +125,9 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		}
 	}
 
-	// On linux, over-allocating CPU memory will almost always result in an error
-	if runtime.GOOS == "linux" {
+	// On linux and windows, over-allocating CPU memory will almost always result in an error
+	// Darwin has fully dynamic swap so has no direct concept of free swap space
+	if runtime.GOOS != "darwin" {
 		systemMemoryRequired := estimate.TotalSize - estimate.VRAMSize
 		available := systemFreeMemory + systemSwapFreeMemory
 		if systemMemoryRequired > available {
@@ -882,24 +883,20 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 	return nil
 }
 
-type EmbedRequest struct {
-	Content []string `json:"content"`
+type EmbeddingRequest struct {
+	Content string `json:"content"`
 }
 
-type EmbedResponse struct {
-	Embedding       [][]float32 `json:"embedding"`
-	PromptEvalCount int         `json:"prompt_n"`
+type EmbeddingResponse struct {
+	Embedding []float32 `json:"embedding"`
 }
 
-func (s *llmServer) Embed(ctx context.Context, input []string) (*EmbedResponse, error) {
-	// each input will use a slot, so we need to acquire the semaphore for
-	// the number of inputs up to numParallel
-	slots := int64(min(len(input), s.numParallel))
-	if err := s.sem.Acquire(ctx, slots); err != nil {
+func (s *llmServer) Embedding(ctx context.Context, input string) ([]float32, error) {
+	if err := s.sem.Acquire(ctx, 1); err != nil {
 		slog.Error("Failed to acquire semaphore", "error", err)
 		return nil, err
 	}
-	defer s.sem.Release(slots)
+	defer s.sem.Release(1)
 
 	// Make sure the server is ready
 	status, err := s.getServerStatusRetry(ctx)
@@ -909,18 +906,18 @@ func (s *llmServer) Embed(ctx context.Context, input []string) (*EmbedResponse,
 		return nil, fmt.Errorf("unexpected server status: %s", status.ToString())
 	}
 
-	data, err := json.Marshal(EmbedRequest{Content: input})
+	data, err := json.Marshal(EmbeddingRequest{Content: input})
 	if err != nil {
 		return nil, fmt.Errorf("error marshaling embed data: %w", err)
 	}
 
-	req, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/embedding", s.port), bytes.NewBuffer(data))
+	r, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/embedding", s.port), bytes.NewBuffer(data))
 	if err != nil {
 		return nil, fmt.Errorf("error creating embed request: %w", err)
 	}
-	req.Header.Set("Content-Type", "application/json")
+	r.Header.Set("Content-Type", "application/json")
 
-	resp, err := http.DefaultClient.Do(req)
+	resp, err := http.DefaultClient.Do(r)
 	if err != nil {
 		return nil, fmt.Errorf("do embedding request: %w", err)
 	}
@@ -936,12 +933,12 @@ func (s *llmServer) Embed(ctx context.Context, input []string) (*EmbedResponse,
 		return nil, fmt.Errorf("%s", body)
 	}
 
-	var e EmbedResponse
+	var e EmbeddingResponse
 	if err := json.Unmarshal(body, &e); err != nil {
 		return nil, fmt.Errorf("unmarshal tokenize response: %w", err)
 	}
 
-	return &e, nil
+	return e.Embedding, nil
 }
 
 type TokenizeRequest struct {

llm/status.go

@@ -26,6 +26,7 @@ var errorPrefixes = []string{
 	"cudaMalloc failed",
 	"\"ERR\"",
 	"error loading model",
+	"GGML_ASSERT",
 }
 
 func (w *StatusWriter) Write(b []byte) (int, error) {