ggml: Support heterogeneous KV cache layer sizes in memory estimation

Gemma3 uses sliding windows for its context on 5/6 layers, significantly
reducing memory usage but leading to uneven usage across layers,
which makes allocation to the correct GPU difficult. We currently
estimate very conservatively by assuming all layers are consistent
at the max size.

Llama3.2-vision is also inconsistent between self attention and cross
attention layers - at moment, we calculate the correct total size
and then average this across layers. In some cases, this may lead
to crashes if a large layer is placed on a GPU sized by the average.

This allows memory estimation to calculate per-layer KV cache size
and take this account when placing layers onto GPUs. We already do
this for weights that vary per-tensor, so this is a logical extension.

Fixes #9730
Fixes #9890

fs/ggml/ggml.go

@@ -413,7 +413,7 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
 	}, offset, nil
 }
 
-func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialOffload, fullOffload uint64) {
+func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string) (kv []uint64, partialOffload, fullOffload uint64) {
 	embedding := f.KV().EmbeddingLength()
 	heads := f.KV().HeadCount()
 	headsKV := f.KV().HeadCountKV()
@@ -426,7 +426,10 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 	layers := f.Tensors().GroupLayers()
 
 	bytesPerElement := kvCacheBytesPerElement(kvCacheType)
-	kv = uint64(float64(context*f.KV().BlockCount()*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+	kv = make([]uint64, f.KV().BlockCount())
+	for i := range kv {
+		kv[i] = uint64(float64(context*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+	}
 
 	switch f.KV().Architecture() {
 	case "llama":
@@ -460,16 +463,14 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 	case "mllama":
 		var visionTokens, tiles uint64 = 1601, 4
 
-		if crossAttentionLayers, ok := f.KV()["mllama.attention.cross_attention_layers"].(*array); ok {
-			kv = headsKV *
-				(embeddingHeadsK + embeddingHeadsV) * // one for K, one for V
-				(2* // sizeof(float16)
-					(f.KV().BlockCount()-uint64(crossAttentionLayers.size))* // num non-cross attention layers
-					context +
-					4* // sizeof(float32)
-						uint64(crossAttentionLayers.size)* // num cross attention layers
-						visionTokens*
-						tiles)
+		crossAttentionLayers := f.KV().Uints("attention.cross_attention_layers")
+		for i := range kv {
+			if slices.Contains(crossAttentionLayers, uint32(i)) {
+				kv[i] = headsKV * (embeddingHeadsK + embeddingHeadsV) *
+					4 * // sizeof(float32)
+					visionTokens *
+					tiles
+			}
 		}
 
 		fullOffload = max(
@@ -505,6 +506,20 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 				4*embeddingHeadsK*context*8+
 				embedding*embeddingHeadsK*heads*9/16,
 		)
+
+		// Gemma2 also has sliding window attention but we only have an optimized implementation in the Ollama
+		// engine. Gemma3 always uses the Ollama engine.
+		if f.KV().Architecture() == "gemma3" {
+			const gemma3GlobalCacheCount = 6
+			slidingWindow := (uint64(numParallel) * uint64(f.KV().Uint("attention.sliding_window"))) + batch
+			for i := range kv {
+				// Every 6th layer is a global layer, which is the full context size that has already been set. The other
+				// layers are the smaller local (sliding) layers.
+				if (i+1)%gemma3GlobalCacheCount != 0 {
+					kv[i] = uint64(float64(slidingWindow*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+				}
+			}
+		}
 	case "command-r":
 		fullOffload = max(
 			4*batch*(embedding+vocab),

llm/memory.go

@@ -15,12 +15,12 @@ import (
 )
 
 // This algorithm looks for a complete fit to determine if we need to unload other models
-func PredictServerFit(allGpus discover.GpuInfoList, f *ggml.GGML, adapters, projectors []string, opts api.Options) (bool, uint64) {
+func PredictServerFit(allGpus discover.GpuInfoList, f *ggml.GGML, adapters, projectors []string, opts api.Options, numParallel int) (bool, uint64) {
 	// Split up the GPUs by type and try them
 	var estimatedVRAM uint64
 	for _, gpus := range allGpus.ByLibrary() {
 		var layerCount int
-		estimate := EstimateGPULayers(gpus, f, projectors, opts)
+		estimate := EstimateGPULayers(gpus, f, projectors, opts, numParallel)
 		layerCount, estimatedVRAM = estimate.Layers, estimate.VRAMSize
 		if opts.NumGPU < 0 {
 			if layerCount > 0 && layerCount >= int(f.KV().BlockCount()+1) {
@@ -71,7 +71,7 @@ type MemoryEstimate struct {
 
 // Given a model and one or more GPU targets, predict how many layers and bytes we can load, and the total size
 // The GPUs provided must all be the same Library
-func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []string, opts api.Options) MemoryEstimate {
+func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []string, opts api.Options, numParallel int) MemoryEstimate {
 	// Graph size for a partial offload, applies to all GPUs
 	var graphPartialOffload uint64
 
@@ -137,13 +137,19 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		}
 	}
 
-	kv, graphPartialOffload, graphFullOffload := f.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)), kvct)
+	kv, graphPartialOffload, graphFullOffload := f.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)), numParallel, kvct)
 
-	// KV is proportional to the number of layers
-	layerSize += kv / f.KV().BlockCount()
+	if len(kv) > 0 {
+		layerSize += kv[0]
+	}
+
+	var kvTotal uint64
+	for _, kvLayer := range kv {
+		kvTotal += kvLayer
+	}
 
 	if graphPartialOffload == 0 {
-		graphPartialOffload = f.KV().GQA() * kv / 6
+		graphPartialOffload = f.KV().GQA() * kvTotal / 6
 	}
 	if graphFullOffload == 0 {
 		graphFullOffload = graphPartialOffload
@@ -217,7 +223,7 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		// Some models have inconsistent layer sizes
 		if blk, ok := layers[fmt.Sprintf("blk.%d", i)]; ok {
 			layerSize = blk.Size()
-			layerSize += kv / f.KV().BlockCount()
+			layerSize += kv[i]
 			memoryWeights += blk.Size()
 		}
 
@@ -315,7 +321,7 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		layersRequested:     opts.NumGPU,
 		layersModel:         int(f.KV().BlockCount()) + 1,
 		availableList:       availableList,
-		kv:                  kv,
+		kv:                  kvTotal,
 		allocationsList:     allocationsList,
 		memoryWeights:       memoryWeights,
 		memoryLayerOutput:   memoryLayerOutput,

llm/memory_test.go

@@ -61,7 +61,7 @@ func TestEstimateGPULayers(t *testing.T) {
 	projectors := []string{}
 	opts := api.DefaultOptions()
 	t.Run("cpu", func(t *testing.T) {
-		estimate := EstimateGPULayers(gpus, ggml, projectors, opts)
+		estimate := EstimateGPULayers(gpus, ggml, projectors, opts, 1)
 		assert.Equal(t, 0, estimate.Layers)
 		assert.Equal(t, uint64(0), estimate.Graph)
 	})
@@ -112,7 +112,7 @@ func TestEstimateGPULayers(t *testing.T) {
 			gpus[1].FreeMemory += gpuMinimumMemory + layerSize + s.layer1*layerSize + 1
 			gpus[0].FreeMemory += max(graphFullOffload, graphPartialOffload)
 			gpus[1].FreeMemory += max(graphFullOffload, graphPartialOffload)
-			estimate := EstimateGPULayers(gpus, ggml, projectors, opts)
+			estimate := EstimateGPULayers(gpus, ggml, projectors, opts, 1)
 			assert.Equal(t, int(s.expect0+s.expect1), estimate.Layers, "scenario %d: %v", i, s)
 			assert.Equal(t, fmt.Sprintf("%d,%d", s.expect0, s.expect1), estimate.TensorSplit, "scenario %d: %v", i, s)
 			var layerSums uint64

llm/server.go

@@ -109,7 +109,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		gpus = discover.GetCPUInfo()
 	}
 
-	estimate := EstimateGPULayers(gpus, f, projectors, opts)
+	estimate := EstimateGPULayers(gpus, f, projectors, opts, numParallel)
 	if len(gpus) > 1 || gpus[0].Library != "cpu" {
 		switch {
 		case gpus[0].Library == "metal" && estimate.VRAMSize > systemTotalMemory:

server/sched.go

@@ -711,7 +711,7 @@ func pickBestFullFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.GpuIn
 			req.opts.NumCtx = req.origNumCtx * p
 			if !envconfig.SchedSpread() {
 				for _, g := range sgl {
-					if ok, estimatedVRAM = llm.PredictServerFit([]discover.GpuInfo{g}, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+					if ok, estimatedVRAM = llm.PredictServerFit([]discover.GpuInfo{g}, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, p); ok {
 						slog.Info("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "parallel", p, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM))
 						*numParallel = p
 						return []discover.GpuInfo{g}
@@ -727,7 +727,7 @@ func pickBestFullFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.GpuIn
 		// Now try all the GPUs
 		for _, p := range numParallelToTry {
 			req.opts.NumCtx = req.origNumCtx * p
-			if ok, estimatedVRAM = llm.PredictServerFit(sgl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+			if ok, estimatedVRAM = llm.PredictServerFit(sgl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, p); ok {
 				slog.Info("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "parallel", p, "required", format.HumanBytes2(estimatedVRAM))
 				*numParallel = p
 				return sgl
@@ -750,7 +750,7 @@ func pickBestPartialFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.Gp
 	var bestEstimate uint64
 	var bestFit int
 	for i, gl := range byLibrary {
-		_, estimatedVRAM := llm.PredictServerFit(gl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
+		_, estimatedVRAM := llm.PredictServerFit(gl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, *numParallel)
 		if estimatedVRAM > bestEstimate {
 			bestEstimate = estimatedVRAM
 			bestFit = i
@@ -825,7 +825,7 @@ func (s *Scheduler) expireRunner(model *Model) {
 // If not, pick a runner to unload, else return nil and the request can be loaded
 func (s *Scheduler) maybeFindCPURunnerToUnload(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList) *runnerRef {
 	slog.Debug("evaluating if CPU model load will fit in available system memory")
-	estimate := llm.EstimateGPULayers(gpus, f, req.model.ProjectorPaths, req.opts)
+	estimate := llm.EstimateGPULayers(gpus, f, req.model.ProjectorPaths, req.opts, req.opts.NumCtx/req.origNumCtx)
 	if estimate.TotalSize <= gpus[0].FreeMemory {
 		slog.Debug("cpu inference mode, model fits in available system memory", "model", format.HumanBytes2(estimate.TotalSize), "available", format.HumanBytes2(gpus[0].FreeMemory))
 		return nil