memory.go 12.6 KB
Newer Older
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1
2
3
package llm

import (
4
	"fmt"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
5
	"log/slog"
6
	"os"
7
8
	"strconv"
	"strings"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
9
10

	"github.com/ollama/ollama/api"
11
	"github.com/ollama/ollama/discover"
12
	"github.com/ollama/ollama/envconfig"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
13
14
15
16
	"github.com/ollama/ollama/format"
)

// This algorithm looks for a complete fit to determine if we need to unload other models
17
func PredictServerFit(allGpus discover.GpuInfoList, ggml *GGML, adapters, projectors []string, opts api.Options) (bool, uint64) {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
18
	// Split up the GPUs by type and try them
19
	var estimatedVRAM uint64
Daniel Hiltgen's avatar
Daniel Hiltgen committed
20
21
	for _, gpus := range allGpus.ByLibrary() {
		var layerCount int
22
23
		estimate := EstimateGPULayers(gpus, ggml, projectors, opts)
		layerCount, estimatedVRAM = estimate.Layers, estimate.VRAMSize
Daniel Hiltgen's avatar
Daniel Hiltgen committed
24
25
26
27
28
29
30
31
32
33
34
35
36
		if opts.NumGPU < 0 {
			if layerCount > 0 && layerCount >= int(ggml.KV().BlockCount()+1) {
				return true, estimatedVRAM
			}
		} else {
			if layerCount > 0 && layerCount >= opts.NumGPU {
				return true, estimatedVRAM
			}
		}
	}
	return false, estimatedVRAM
}

37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
type MemoryEstimate struct {
	// How many layers we predict we can load
	Layers int

	// The size of the graph which occupies the main GPU
	Graph uint64

	// How much VRAM will be allocated given the number of layers we predict
	VRAMSize uint64

	// The total size of the model if loaded into VRAM.  If all layers are loaded, VRAMSize == TotalSize
	TotalSize uint64

	// For multi-GPU scenarios, this provides the tensor split parameter
	TensorSplit string

	// For multi-GPU scenarios, this is the size in bytes per GPU
	GPUSizes []uint64
55
56
57
58
59
60
61
62
63
64
65
66

	// 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
67
68

	projectorWeights, projectorGraph uint64
69
70
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
71
// Given a model and one or more GPU targets, predict how many layers and bytes we can load, and the total size
Daniel Hiltgen's avatar
Daniel Hiltgen committed
72
// The GPUs provided must all be the same Library
73
func EstimateGPULayers(gpus []discover.GpuInfo, ggml *GGML, projectors []string, opts api.Options) MemoryEstimate {
74
75
76
77
78
79
80
81
82
83
	// Graph size for a partial offload, applies to all GPUs
	var graphPartialOffload uint64

	// Graph size when all layers are offloaded, applies to all GPUs
	var graphFullOffload uint64

	// Final graph offload once we know full or partial
	var graphOffload uint64

	// Projectors loaded into GPU0 only
84
85
	var projectorWeights uint64
	var projectorGraph uint64
86
87
88
89

	// Conditional output size on GPU 0
	var memoryLayerOutput uint64

Daniel Hiltgen's avatar
Daniel Hiltgen committed
90
91
	// The sizes of a layer
	var layerSize uint64
Daniel Hiltgen's avatar
Daniel Hiltgen committed
92

93
94
95
96
97
98
99
100
101
	// The sum of all the layer sizes (just for logging)
	var memoryWeights uint64

	// True if all the layers are loaded
	var fullyLoaded bool

	// Overflow that didn't fit into the GPU
	var overflow uint64

102
	overhead := envconfig.GpuOverhead()
103
104
105
106
107
	availableList := make([]string, len(gpus))
	for i, gpu := range gpus {
		availableList[i] = format.HumanBytes2(gpu.FreeMemory)
	}
	slog.Debug("evaluating", "library", gpus[0].Library, "gpu_count", len(gpus), "available", availableList)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
108
109

	for _, projector := range projectors {
110
111
112
		weight, graph := projectorMemoryRequirements(projector)
		projectorWeights += weight
		projectorGraph += graph
Daniel Hiltgen's avatar
Daniel Hiltgen committed
113
114
115
116
117

		// multimodal models require at least 2048 context
		opts.NumCtx = max(opts.NumCtx, 2048)
	}

Michael Yang's avatar
Michael Yang committed
118
	layers := ggml.Tensors().Layers()
Michael Yang's avatar
typo  
Michael Yang committed
119
120
	// add one layer worth of memory as a buffer
	if blk0, ok := layers["blk.0"]; ok {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
121
122
123
		layerSize = blk0.size()
	} else {
		slog.Warn("model missing blk.0 layer size")
Michael Yang's avatar
typo  
Michael Yang committed
124
	}
Michael Yang's avatar
Michael Yang committed
125

126
127
128
129
130
131
	fa := envconfig.FlashAttention() &&
		discover.GetGPUInfo().FlashAttentionSupported() &&
		ggml.SupportsFlashAttention()

	var kvct string
	if fa {
132
		requested := strings.ToLower(envconfig.KvCacheType())
133
134
135
136
137
138
139
140
141
142
		if requested != "" && ggml.SupportsKVCacheType(requested) {
			kvct = requested
		}
	}

	kv, graphPartialOffload, graphFullOffload := ggml.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)), kvct)

	// KV is proportional to the number of layers
	layerSize += kv / ggml.KV().BlockCount()

Daniel Hiltgen's avatar
Daniel Hiltgen committed
143
144
145
146
147
148
149
	if graphPartialOffload == 0 {
		graphPartialOffload = ggml.KV().GQA() * kv / 6
	}
	if graphFullOffload == 0 {
		graphFullOffload = graphPartialOffload
	}

150
151
152
	// on metal there's no partial offload overhead
	if gpus[0].Library == "metal" {
		graphPartialOffload = graphFullOffload
Daniel Hiltgen's avatar
Daniel Hiltgen committed
153
154
155
	} else if len(gpus) > 1 {
		// multigpu should always use the partial graph size
		graphFullOffload = graphPartialOffload
156
157
	}

158
159
160
161
162
163
164
	if layer, ok := layers["output_norm"]; ok {
		memoryLayerOutput += layer.size()
	}
	if layer, ok := layers["output"]; ok {
		memoryLayerOutput += layer.size()
	} else if layer, ok := layers["token_embd"]; ok {
		memoryLayerOutput += layer.size()
Michael Yang's avatar
Michael Yang committed
165
166
	}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
167
	// Output layer handled at the end if we have space
168
	gpuZeroOverhead := projectorWeights + projectorGraph
169
170

	// Reduce set of GPUs to only those that have sufficient space to fit overhead and at least one layer
Michael Yang's avatar
Michael Yang committed
171
	var layerCount int
172
173
174
175
	layerCounts := make([]int, len(gpus))
	gpuAllocations := make([]uint64, len(gpus))
	type gs struct {
		i int
176
		g *discover.GpuInfo
177
178
179
180
181
182
183
184
	}
	gpusWithSpace := []gs{}
	for i := range gpus {
		var gzo uint64
		if len(gpusWithSpace) == 0 {
			gzo = gpuZeroOverhead
		}
		// Only include GPUs that can fit the graph, gpu minimum, the layer buffer and at least more layer
185
		if (gpus[i].FreeMemory - overhead) < gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory+2*layerSize {
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
			slog.Debug("gpu has too little memory to allocate any layers",
				"id", gpus[i].ID,
				"library", gpus[i].Library,
				"variant", gpus[i].Variant,
				"compute", gpus[i].Compute,
				"driver", fmt.Sprintf("%d.%d", gpus[i].DriverMajor, gpus[i].DriverMinor),
				"name", gpus[i].Name,
				"total", format.HumanBytes2(gpus[i].TotalMemory),
				"available", format.HumanBytes2(gpus[i].FreeMemory),
				"minimum_memory", gpus[i].MinimumMemory,
				"layer_size", format.HumanBytes2(layerSize),
				"gpu_zer_overhead", format.HumanBytes2(gzo),
				"partial_offload", format.HumanBytes2(graphPartialOffload),
				"full_offload", format.HumanBytes2(graphFullOffload),
			)
201
202
203
			continue
		}
		gpusWithSpace = append(gpusWithSpace, gs{i, &gpus[i]})
Daniel Hiltgen's avatar
Daniel Hiltgen committed
204
		gpuAllocations[i] += gpus[i].MinimumMemory + layerSize // We hold off on graph until we know partial vs. full
205
206
207
208
209
210
211
212
	}

	var gpuZeroID int
	if len(gpusWithSpace) > 0 {
		gpuZeroID = gpusWithSpace[0].i
		gpuAllocations[gpuZeroID] += gpuZeroOverhead
	}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
213
	// For all the layers, find where they can fit on the GPU(s)
Michael Yang's avatar
lint  
Michael Yang committed
214
	for i := range int(ggml.KV().BlockCount()) {
215
216
217
218
219
		// Some models have inconsistent layer sizes
		if blk, ok := layers[fmt.Sprintf("blk.%d", i)]; ok {
			layerSize = blk.size()
			layerSize += kv / ggml.KV().BlockCount()
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
220
		memoryWeights += layerSize
Daniel Hiltgen's avatar
Daniel Hiltgen committed
221

222
223
224
225
226
227
228
229
230
		if opts.NumGPU >= 0 && layerCount >= opts.NumGPU {
			// Stop allocating on GPU(s) once we hit the users target NumGPU
			continue
		}

		// distribute the layers across the GPU(s) that have space
		for j := len(gpusWithSpace); j > 0; j-- {
			g := gpusWithSpace[i%j]
			used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
231
			if (g.g.FreeMemory - overhead) > used+layerSize {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
232
				gpuAllocations[g.i] += layerSize
233
				layerCounts[g.i]++
Michael Yang's avatar
typo  
Michael Yang committed
234
				layerCount++
235
236
237
				break
			} else {
				gpusWithSpace = append(gpusWithSpace[:i%j], gpusWithSpace[i%j+1:]...)
Michael Yang's avatar
typo  
Michael Yang committed
238
			}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
239
		}
240
241
242
243
244
	}
	if layerCount >= int(ggml.KV().BlockCount()) {
		fullyLoaded = true
	} else {
		for i := layerCount; i < int(ggml.KV().BlockCount()); i++ {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
245
			overflow += layerSize
246
247
		}
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
248
249

	// Determine if we need to consider output then find where it fits
250
	if memoryLayerOutput > 0 && (opts.NumGPU < 0 || layerCount < opts.NumGPU) {
251
252
253
		for j := len(gpusWithSpace); j > 0; j-- {
			g := gpusWithSpace[layerCount%j]
			used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
254
			if (g.g.FreeMemory - overhead) > used+memoryLayerOutput {
255
256
257
258
259
260
				gpuAllocations[g.i] += memoryLayerOutput
				layerCounts[g.i]++
				layerCount++
				break
			}
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
261

262
263
264
265
		if layerCount < int(ggml.KV().BlockCount())+1 {
			fullyLoaded = false
			overflow += memoryLayerOutput
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
266
267
	}

268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
	// Add the applicable (full or partial) graph allocations
	for i := range gpus {
		if layerCounts[i] <= 0 {
			continue
		}
		if fullyLoaded {
			gpuAllocations[i] += graphFullOffload
		} else {
			gpuAllocations[i] += graphPartialOffload
		}
	}
	if fullyLoaded {
		graphOffload = graphFullOffload
	} else {
		graphOffload = graphPartialOffload
Daniel Hiltgen's avatar
Daniel Hiltgen committed
283
284
	}

285
286
287
288
	// Summaries for the log
	var memoryRequiredPartial, memoryRequiredTotal uint64
	for i := range gpuAllocations {
		memoryRequiredPartial += gpuAllocations[i]
Daniel Hiltgen's avatar
Daniel Hiltgen committed
289
	}
290
	memoryRequiredTotal = memoryRequiredPartial + overflow
Daniel Hiltgen's avatar
Daniel Hiltgen committed
291

292
293
294
295
296
297
298
299
300
301
302
303
	tensorSplit := ""
	if len(gpus) > 1 {
		splits := make([]string, len(gpus))
		for i, count := range layerCounts {
			splits[i] = strconv.Itoa(count)
		}
		tensorSplit = strings.Join(splits, ",")
	}
	allocationsList := []string{}
	for _, a := range gpuAllocations {
		allocationsList = append(allocationsList, format.HumanBytes2(a))
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
304

305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
	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,
322
323
		projectorWeights:    projectorWeights,
		projectorGraph:      projectorGraph,
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
	}

	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() {
343
	overhead := envconfig.GpuOverhead()
344
345
346
347
348
349
350
351
352
353
354
355
356

	log := slog.With()
	if m.projectorWeights > 0 {
		log = log.With(
			slog.Group(
				"projector",
				"weights", format.HumanBytes2(m.projectorWeights),
				"graph", format.HumanBytes2(m.projectorGraph),
			),
		)
	}

	log.Info(
357
		"offload to "+m.inferenceLibrary,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
358
359
		slog.Group(
			"layers",
Michael Yang's avatar
Michael Yang committed
360
			// requested number of layers to offload
361
			"requested", m.layersRequested,
362
			// The number of layers the model has (including output)
363
			"model", m.layersModel,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
364
			// estimated number of layers that can be offloaded
365
366
367
			"offload", m.Layers,
			// multi-gpu split for tensors
			"split", m.TensorSplit,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
368
369
370
		),
		slog.Group(
			"memory",
371
			// memory available by GPU for offloading
372
			"available", m.availableList,
373
			"gpu_overhead", format.HumanBytes2(overhead),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
374
375
376
			slog.Group(
				"required",
				// memory required for full offloading
377
				"full", format.HumanBytes2(m.TotalSize),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
378
				// memory required to offload layers.estimate layers
379
				"partial", format.HumanBytes2(m.VRAMSize),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
380
				// memory of KV cache
381
				"kv", format.HumanBytes2(m.kv),
382
				// Allocations across the GPUs
383
				"allocations", m.allocationsList,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
384
385
386
387
			),
			slog.Group(
				"weights",
				// memory of the weights
388
				"total", format.HumanBytes2(m.memoryWeights),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
389
				// memory of repeating layers
390
				"repeating", format.HumanBytes2(m.memoryWeights-m.memoryLayerOutput),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
391
				// memory of non-repeating layers
392
				"nonrepeating", format.HumanBytes2(m.memoryLayerOutput),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
393
394
395
396
			),
			slog.Group(
				"graph",
				// memory of graph when fully offloaded
397
				"full", format.HumanBytes2(m.graphFullOffload),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
398
				// memory of graph when not fully offloaded
399
				"partial", format.HumanBytes2(m.graphPartialOffload),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
400
401
402
403
			),
		),
	)
}
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452

func projectorMemoryRequirements(filename string) (weights, graphSize uint64) {
	file, err := os.Open(filename)
	if err != nil {
		return 0, 0
	}
	defer file.Close()

	ggml, _, err := DecodeGGML(file, 0)
	if err != nil {
		return 0, 0
	}

	for _, layer := range ggml.Tensors().Layers() {
		weights += layer.size()
	}

	switch arch := ggml.KV().Architecture(); arch {
	case "mllama":
		kv := func(n string) uint64 {
			if v, ok := ggml.KV()[arch+".vision."+n].(uint32); ok {
				return uint64(v)
			}

			return 0
		}

		imageSize := kv("image_size")

		maxNumTiles := kv("max_num_tiles")
		embeddingLength := kv("embedding_length")
		headCount := kv("attention.head_count")

		numPatches := (imageSize / kv("patch_size")) * (imageSize / kv("patch_size"))
		if _, ok := ggml.Tensors().Layers()["v"]["class_embd"]; ok {
			numPatches++
		}

		numPaddedPatches := numPatches + 8 - (numPatches%8)%8

		graphSize = 4 * (8 +
			imageSize*imageSize*kv("num_channels")*maxNumTiles +
			embeddingLength*numPatches*maxNumTiles +
			9*embeddingLength*numPaddedPatches*maxNumTiles +
			numPaddedPatches*maxNumTiles*numPaddedPatches*maxNumTiles*headCount)
	}

	return weights, graphSize
}