Tighten up memory prediction logging

Prior to this change, we logged the memory prediction multiple times
as the scheduler iterates to find a suitable configuration, which can be
confusing since only the last log before the server starts is actually valid.
This now logs once just before starting the server on the final configuration.
It also reports what library instead of always saying "offloading to gpu" when
using CPU.

llm/memory.go

@@ -49,6 +49,18 @@ type MemoryEstimate struct {
 
 	// For multi-GPU scenarios, this is the size in bytes per GPU
 	GPUSizes []uint64
+
+	// internal fields for logging purposes
+	inferenceLibrary    string
+	layersRequested     int
+	layersModel         int
+	availableList       []string
+	kv                  uint64
+	allocationsList     []string
+	memoryWeights       uint64
+	memoryLayerOutput   uint64
+	graphFullOffload    uint64
+	graphPartialOffload uint64
 }
 
 // Given a model and one or more GPU targets, predict how many layers and bytes we can load, and the total size
@@ -252,78 +264,86 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		allocationsList = append(allocationsList, format.HumanBytes2(a))
 	}
 
+	estimate := MemoryEstimate{
+		TotalSize: memoryRequiredTotal,
+		Layers:    0,
+		Graph:     0,
+		VRAMSize:  0,
+		GPUSizes:  []uint64{},
+
+		inferenceLibrary:    gpus[0].Library,
+		layersRequested:     opts.NumGPU,
+		layersModel:         int(ggml.KV().BlockCount()) + 1,
+		availableList:       availableList,
+		kv:                  kv,
+		allocationsList:     allocationsList,
+		memoryWeights:       memoryWeights,
+		memoryLayerOutput:   memoryLayerOutput,
+		graphFullOffload:    graphFullOffload,
+		graphPartialOffload: graphPartialOffload,
+	}
+
+	if gpus[0].Library == "cpu" {
+		return estimate
+	}
+	if layerCount == 0 {
+		slog.Debug("insufficient VRAM to load any model layers")
+		return estimate
+	}
+	estimate.Layers = layerCount
+	estimate.Graph = graphOffload
+	estimate.VRAMSize = memoryRequiredPartial
+	estimate.TotalSize = memoryRequiredTotal
+	estimate.TensorSplit = tensorSplit
+	estimate.GPUSizes = gpuAllocations
+	return estimate
+}
+
+func (m MemoryEstimate) log() {
 	slog.Info(
-		"offload to gpu",
+		"offload to "+m.inferenceLibrary,
 		slog.Group(
 			"layers",
 			// requested number of layers to offload
-			"requested", opts.NumGPU,
+			"requested", m.layersRequested,
 			// The number of layers the model has (including output)
-			"model", int(ggml.KV().BlockCount())+1,
+			"model", m.layersModel,
 			// estimated number of layers that can be offloaded
-			"offload", layerCount,
-			// multi-gpu split for tesnors
-			"split", tensorSplit,
+			"offload", m.Layers,
+			// multi-gpu split for tensors
+			"split", m.TensorSplit,
 		),
 		slog.Group(
 			"memory",
 			// memory available by GPU for offloading
-			"available", availableList,
+			"available", m.availableList,
 			slog.Group(
 				"required",
 				// memory required for full offloading
-				"full", format.HumanBytes2(memoryRequiredTotal),
+				"full", format.HumanBytes2(m.TotalSize),
 				// memory required to offload layers.estimate layers
-				"partial", format.HumanBytes2(memoryRequiredPartial),
+				"partial", format.HumanBytes2(m.VRAMSize),
 				// memory of KV cache
-				"kv", format.HumanBytes2(kv),
+				"kv", format.HumanBytes2(m.kv),
 				// Allocations across the GPUs
-				"allocations", allocationsList,
+				"allocations", m.allocationsList,
 			),
 			slog.Group(
 				"weights",
 				// memory of the weights
-				"total", format.HumanBytes2(memoryWeights),
+				"total", format.HumanBytes2(m.memoryWeights),
 				// memory of repeating layers
-				"repeating", format.HumanBytes2(memoryWeights-memoryLayerOutput),
+				"repeating", format.HumanBytes2(m.memoryWeights-m.memoryLayerOutput),
 				// memory of non-repeating layers
-				"nonrepeating", format.HumanBytes2(memoryLayerOutput),
+				"nonrepeating", format.HumanBytes2(m.memoryLayerOutput),
 			),
 			slog.Group(
 				"graph",
 				// memory of graph when fully offloaded
-				"full", format.HumanBytes2(graphFullOffload),
+				"full", format.HumanBytes2(m.graphFullOffload),
 				// memory of graph when not fully offloaded
-				"partial", format.HumanBytes2(graphPartialOffload),
+				"partial", format.HumanBytes2(m.graphPartialOffload),
 			),
 		),
 	)
-	if gpus[0].Library == "cpu" {
-		return MemoryEstimate{
-			Layers:    0,
-			Graph:     0,
-			VRAMSize:  0,
-			TotalSize: memoryRequiredTotal,
-			GPUSizes:  []uint64{},
-		}
-	}
-	if layerCount == 0 {
-		slog.Debug("insufficient VRAM to load any model layers")
-		return MemoryEstimate{
-			Layers:    0,
-			Graph:     0,
-			VRAMSize:  0,
-			TotalSize: memoryRequiredTotal,
-			GPUSizes:  []uint64{},
-		}
-	}
-
-	return MemoryEstimate{
-		Layers:      layerCount,
-		Graph:       graphOffload,
-		VRAMSize:    memoryRequiredPartial,
-		TotalSize:   memoryRequiredTotal,
-		TensorSplit: tensorSplit,
-		GPUSizes:    gpuAllocations,
-	}
 }

llm/server.go

@@ -116,6 +116,8 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		}
 	}
 
+	estimate.log()
+
 	// Loop through potential servers
 	finalErr := errors.New("no suitable llama servers found")