Merge branch 'ollama:main' into main

llm/ext_server/server.cpp

@@ -1186,8 +1186,6 @@ struct llama_server_context
             {"model",               params.model_alias},
             {"tokens_predicted",    slot.n_decoded},
             {"tokens_evaluated",    slot.n_prompt_tokens},
-            {"generation_settings", get_formated_generation(slot)},
-            {"prompt",              slot.prompt},
             {"truncated",           slot.truncated},
             {"stopped_eos",         slot.stopped_eos},
             {"stopped_word",        slot.stopped_word},

llm/memory.go

@@ -3,12 +3,11 @@ package llm
 import (
 	"fmt"
 	"log/slog"
-	"os"
-	"strconv"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/gpu"
+	"github.com/ollama/ollama/server/envconfig"
 )
 
 // This algorithm looks for a complete fit to determine if we need to unload other models
@@ -50,15 +49,8 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 	for _, info := range gpus {
 		memoryAvailable += info.FreeMemory
 	}
-	userLimit := os.Getenv("OLLAMA_MAX_VRAM")
-	if userLimit != "" {
-		avail, err := strconv.ParseUint(userLimit, 10, 64)
-		if err != nil {
-			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_VRAM", userLimit, "error", err)
-		} else {
-			slog.Info("user override memory limit", "OLLAMA_MAX_VRAM", avail, "actual", memoryAvailable)
-			memoryAvailable = avail
-		}
+	if envconfig.MaxVRAM > 0 {
+		memoryAvailable = envconfig.MaxVRAM
 	}
 
 	slog.Debug("evaluating", "library", gpus[0].Library, "gpu_count", len(gpus), "available", format.HumanBytes2(memoryAvailable))

llm/patches/05-clip-fix.diff

@@ -0,0 +1,24 @@
+diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
+index e3c9bcd4..b43f892d 100644
+--- a/examples/llava/clip.cpp
++++ b/examples/llava/clip.cpp
+@@ -573,14 +573,16 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
+     struct ggml_tensor * embeddings = inp;
+     if (ctx->has_class_embedding) {
+         embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
++    }
++    ggml_set_name(embeddings, "embeddings");
++    ggml_set_input(embeddings);
++
++    if (ctx->has_class_embedding) {
+         embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
+                 embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
+         embeddings = ggml_acc(ctx0, embeddings, inp,
+                 embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
+     }
+-    ggml_set_name(embeddings, "embeddings");
+-    ggml_set_input(embeddings);
+-
+ 
+     struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
+     ggml_set_name(positions, "positions");

llm/server.go

@@ -26,6 +26,7 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/gpu"
+	"github.com/ollama/ollama/server/envconfig"
 )
 
 type LlamaServer interface {
@@ -124,7 +125,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	} else {
 		servers = serversForGpu(gpus[0]) // All GPUs in the list are matching Library and Variant
 	}
-	demandLib := strings.Trim(os.Getenv("OLLAMA_LLM_LIBRARY"), "\"' ")
+	demandLib := envconfig.LLMLibrary
 	if demandLib != "" {
 		serverPath := availableServers[demandLib]
 		if serverPath == "" {
@@ -145,7 +146,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		"--batch-size", fmt.Sprintf("%d", opts.NumBatch),
 		"--embedding",
 	}
-	if debug := os.Getenv("OLLAMA_DEBUG"); debug != "" {
+	if envconfig.Debug {
 		params = append(params, "--log-format", "json")
 	} else {
 		params = append(params, "--log-disable")
@@ -155,7 +156,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		params = append(params, "--n-gpu-layers", fmt.Sprintf("%d", opts.NumGPU))
 	}
 
-	if debug := os.Getenv("OLLAMA_DEBUG"); debug != "" {
+	if envconfig.Debug {
 		params = append(params, "--verbose")
 	}
 
@@ -193,16 +194,15 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		params = append(params, "--numa")
 	}
 
-	// "--cont-batching", // TODO - doesn't seem to have any noticeable perf change for multiple requests
-	numParallel := 1
-	if onp := os.Getenv("OLLAMA_NUM_PARALLEL"); onp != "" {
-		numParallel, err = strconv.Atoi(onp)
-		if err != nil || numParallel <= 0 {
-			err = fmt.Errorf("invalid OLLAMA_NUM_PARALLEL=%s must be greater than zero - %w", onp, err)
-			slog.Error("misconfiguration", "error", err)
-			return nil, err
-		}
+	numParallel := envconfig.NumParallel
+
+	// TODO (jmorganca): multimodal models don't support parallel yet
+	// see https://github.com/ollama/ollama/issues/4165
+	if len(projectors) > 0 {
+		numParallel = 1
+		slog.Warn("multimodal models don't support parallel requests yet")
 	}
+
 	params = append(params, "--parallel", fmt.Sprintf("%d", numParallel))
 
 	for i := 0; i < len(servers); i++ {