memory.go 12.3 KB
Newer Older
mashun1's avatar
v1  
mashun1 committed
1
2
3
4
5
package llm

import (
	"fmt"
	"log/slog"
xuxzh1's avatar
update  
xuxzh1 committed
6
	"os"
xuxzh1's avatar
init  
xuxzh1 committed
7
8
	"strconv"
	"strings"
mashun1's avatar
v1  
mashun1 committed
9
10

	"github.com/ollama/ollama/api"
xuxzh1's avatar
update  
xuxzh1 committed
11
12
	"github.com/ollama/ollama/discover"
	"github.com/ollama/ollama/envconfig"
mashun1's avatar
v1  
mashun1 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
xuxzh1's avatar
update  
xuxzh1 committed
17
func PredictServerFit(allGpus discover.GpuInfoList, ggml *GGML, adapters, projectors []string, opts api.Options) (bool, uint64) {
mashun1's avatar
v1  
mashun1 committed
18
19
20
21
	// Split up the GPUs by type and try them
	var estimatedVRAM uint64
	for _, gpus := range allGpus.ByLibrary() {
		var layerCount int
xuxzh1's avatar
init  
xuxzh1 committed
22
23
		estimate := EstimateGPULayers(gpus, ggml, projectors, opts)
		layerCount, estimatedVRAM = estimate.Layers, estimate.VRAMSize
mashun1's avatar
v1  
mashun1 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
}

xuxzh1's avatar
init  
xuxzh1 committed
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
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

	// 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
xuxzh1's avatar
update  
xuxzh1 committed
67
68

	projectorWeights, projectorGraph uint64
xuxzh1's avatar
init  
xuxzh1 committed
69
70
}

mashun1's avatar
v1  
mashun1 committed
71
72
// 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
xuxzh1's avatar
update  
xuxzh1 committed
73
func EstimateGPULayers(gpus []discover.GpuInfo, ggml *GGML, projectors []string, opts api.Options) MemoryEstimate {
xuxzh1's avatar
init  
xuxzh1 committed
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
xuxzh1's avatar
update  
xuxzh1 committed
84
85
	var projectorWeights uint64
	var projectorGraph uint64
xuxzh1's avatar
init  
xuxzh1 committed
86
87
88
89
90
91
92
93
94

	// Conditional output size on GPU 0
	var memoryLayerOutput uint64

	// The sizes of a layer
	var layerSize uint64

	// The sum of all the layer sizes (just for logging)
	var memoryWeights uint64
mashun1's avatar
v1  
mashun1 committed
95

xuxzh1's avatar
init  
xuxzh1 committed
96
97
	// True if all the layers are loaded
	var fullyLoaded bool
mashun1's avatar
v1  
mashun1 committed
98

xuxzh1's avatar
init  
xuxzh1 committed
99
100
101
	// Overflow that didn't fit into the GPU
	var overflow uint64

xuxzh1's avatar
update  
xuxzh1 committed
102
	overhead := envconfig.GpuOverhead()
xuxzh1's avatar
init  
xuxzh1 committed
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)
mashun1's avatar
v1  
mashun1 committed
108
109

	for _, projector := range projectors {
xuxzh1's avatar
update  
xuxzh1 committed
110
111
112
		weight, graph := projectorMemoryRequirements(projector)
		projectorWeights += weight
		projectorGraph += graph
mashun1's avatar
v1  
mashun1 committed
113
114
115
116
117
118
119
120

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

	layers := ggml.Tensors().Layers()
	// add one layer worth of memory as a buffer
	if blk0, ok := layers["blk.0"]; ok {
xuxzh1's avatar
init  
xuxzh1 committed
121
122
123
		layerSize = blk0.size()
	} else {
		slog.Warn("model missing blk.0 layer size")
mashun1's avatar
v1  
mashun1 committed
124
125
	}

xuxzh1's avatar
update  
xuxzh1 committed
126
	kv, graphPartialOffload, graphFullOffload := ggml.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)))
mashun1's avatar
v1  
mashun1 committed
127
128
129
130
131
132
133
	if graphPartialOffload == 0 {
		graphPartialOffload = ggml.KV().GQA() * kv / 6
	}
	if graphFullOffload == 0 {
		graphFullOffload = graphPartialOffload
	}

xuxzh1's avatar
update  
xuxzh1 committed
134
135
136
	// KV is proportional to the number of layers
	layerSize += kv / ggml.KV().BlockCount()

mashun1's avatar
v1  
mashun1 committed
137
138
139
	// on metal there's no partial offload overhead
	if gpus[0].Library == "metal" {
		graphPartialOffload = graphFullOffload
xuxzh1's avatar
init  
xuxzh1 committed
140
141
142
	} else if len(gpus) > 1 {
		// multigpu should always use the partial graph size
		graphFullOffload = graphPartialOffload
mashun1's avatar
v1  
mashun1 committed
143
144
145
146
147
148
149
150
151
152
153
	}

	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()
	}

xuxzh1's avatar
init  
xuxzh1 committed
154
	// Output layer handled at the end if we have space
xuxzh1's avatar
update  
xuxzh1 committed
155
	gpuZeroOverhead := projectorWeights + projectorGraph
mashun1's avatar
v1  
mashun1 committed
156

xuxzh1's avatar
init  
xuxzh1 committed
157
	// Reduce set of GPUs to only those that have sufficient space to fit overhead and at least one layer
mashun1's avatar
v1  
mashun1 committed
158
	var layerCount int
xuxzh1's avatar
init  
xuxzh1 committed
159
160
161
162
	layerCounts := make([]int, len(gpus))
	gpuAllocations := make([]uint64, len(gpus))
	type gs struct {
		i int
xuxzh1's avatar
update  
xuxzh1 committed
163
		g *discover.GpuInfo
xuxzh1's avatar
init  
xuxzh1 committed
164
165
166
167
168
169
170
171
	}
	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
xuxzh1's avatar
update  
xuxzh1 committed
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
		if (gpus[i].FreeMemory - overhead) < gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory+2*layerSize {
			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),
			)
xuxzh1's avatar
init  
xuxzh1 committed
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
			continue
		}
		gpusWithSpace = append(gpusWithSpace, gs{i, &gpus[i]})
		gpuAllocations[i] += gpus[i].MinimumMemory + layerSize // We hold off on graph until we know partial vs. full
	}

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

	// For all the layers, find where they can fit on the GPU(s)
	for i := range int(ggml.KV().BlockCount()) {
		// Some models have inconsistent layer sizes
mashun1's avatar
v1  
mashun1 committed
203
		if blk, ok := layers[fmt.Sprintf("blk.%d", i)]; ok {
xuxzh1's avatar
init  
xuxzh1 committed
204
205
206
207
			layerSize = blk.size()
			layerSize += kv / ggml.KV().BlockCount()
		}
		memoryWeights += layerSize
mashun1's avatar
v1  
mashun1 committed
208

xuxzh1's avatar
init  
xuxzh1 committed
209
210
211
212
		if opts.NumGPU >= 0 && layerCount >= opts.NumGPU {
			// Stop allocating on GPU(s) once we hit the users target NumGPU
			continue
		}
mashun1's avatar
v1  
mashun1 committed
213

xuxzh1's avatar
init  
xuxzh1 committed
214
215
216
217
		// 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)
xuxzh1's avatar
update  
xuxzh1 committed
218
			if (g.g.FreeMemory - overhead) > used+layerSize {
xuxzh1's avatar
init  
xuxzh1 committed
219
220
				gpuAllocations[g.i] += layerSize
				layerCounts[g.i]++
mashun1's avatar
v1  
mashun1 committed
221
				layerCount++
xuxzh1's avatar
init  
xuxzh1 committed
222
223
224
				break
			} else {
				gpusWithSpace = append(gpusWithSpace[:i%j], gpusWithSpace[i%j+1:]...)
mashun1's avatar
v1  
mashun1 committed
225
226
227
			}
		}
	}
xuxzh1's avatar
init  
xuxzh1 committed
228
229
230
231
232
233
234
235
236
237
238
239
240
	if layerCount >= int(ggml.KV().BlockCount()) {
		fullyLoaded = true
	} else {
		for i := layerCount; i < int(ggml.KV().BlockCount()); i++ {
			overflow += layerSize
		}
	}

	// Determine if we need to consider output then find where it fits
	if memoryLayerOutput > 0 && (opts.NumGPU < 0 || layerCount < opts.NumGPU) {
		for j := len(gpusWithSpace); j > 0; j-- {
			g := gpusWithSpace[layerCount%j]
			used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
xuxzh1's avatar
update  
xuxzh1 committed
241
			if (g.g.FreeMemory - overhead) > used+memoryLayerOutput {
xuxzh1's avatar
init  
xuxzh1 committed
242
243
244
245
246
247
				gpuAllocations[g.i] += memoryLayerOutput
				layerCounts[g.i]++
				layerCount++
				break
			}
		}
mashun1's avatar
v1  
mashun1 committed
248

xuxzh1's avatar
init  
xuxzh1 committed
249
250
251
252
		if layerCount < int(ggml.KV().BlockCount())+1 {
			fullyLoaded = false
			overflow += memoryLayerOutput
		}
mashun1's avatar
v1  
mashun1 committed
253
254
	}

xuxzh1's avatar
init  
xuxzh1 committed
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
	// 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
mashun1's avatar
v1  
mashun1 committed
270
271
	}

xuxzh1's avatar
init  
xuxzh1 committed
272
273
274
275
276
277
	// Summaries for the log
	var memoryRequiredPartial, memoryRequiredTotal uint64
	for i := range gpuAllocations {
		memoryRequiredPartial += gpuAllocations[i]
	}
	memoryRequiredTotal = memoryRequiredPartial + overflow
mashun1's avatar
v1  
mashun1 committed
278

xuxzh1's avatar
init  
xuxzh1 committed
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
	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))
	}

	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,
xuxzh1's avatar
update  
xuxzh1 committed
309
310
		projectorWeights:    projectorWeights,
		projectorGraph:      projectorGraph,
xuxzh1's avatar
init  
xuxzh1 committed
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
	}

	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() {
xuxzh1's avatar
update  
xuxzh1 committed
330
331
332
333
334
335
336
337
338
339
340
341
342
343
	overhead := envconfig.GpuOverhead()

	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(
xuxzh1's avatar
init  
xuxzh1 committed
344
		"offload to "+m.inferenceLibrary,
mashun1's avatar
v1  
mashun1 committed
345
346
347
		slog.Group(
			"layers",
			// requested number of layers to offload
xuxzh1's avatar
init  
xuxzh1 committed
348
349
350
			"requested", m.layersRequested,
			// The number of layers the model has (including output)
			"model", m.layersModel,
mashun1's avatar
v1  
mashun1 committed
351
			// estimated number of layers that can be offloaded
xuxzh1's avatar
init  
xuxzh1 committed
352
353
354
			"offload", m.Layers,
			// multi-gpu split for tensors
			"split", m.TensorSplit,
mashun1's avatar
v1  
mashun1 committed
355
356
357
		),
		slog.Group(
			"memory",
xuxzh1's avatar
init  
xuxzh1 committed
358
359
			// memory available by GPU for offloading
			"available", m.availableList,
xuxzh1's avatar
update  
xuxzh1 committed
360
			"gpu_overhead", format.HumanBytes2(overhead),
mashun1's avatar
v1  
mashun1 committed
361
362
363
			slog.Group(
				"required",
				// memory required for full offloading
xuxzh1's avatar
init  
xuxzh1 committed
364
				"full", format.HumanBytes2(m.TotalSize),
mashun1's avatar
v1  
mashun1 committed
365
				// memory required to offload layers.estimate layers
xuxzh1's avatar
init  
xuxzh1 committed
366
				"partial", format.HumanBytes2(m.VRAMSize),
mashun1's avatar
v1  
mashun1 committed
367
				// memory of KV cache
xuxzh1's avatar
init  
xuxzh1 committed
368
369
370
				"kv", format.HumanBytes2(m.kv),
				// Allocations across the GPUs
				"allocations", m.allocationsList,
mashun1's avatar
v1  
mashun1 committed
371
372
373
374
			),
			slog.Group(
				"weights",
				// memory of the weights
xuxzh1's avatar
init  
xuxzh1 committed
375
				"total", format.HumanBytes2(m.memoryWeights),
mashun1's avatar
v1  
mashun1 committed
376
				// memory of repeating layers
xuxzh1's avatar
init  
xuxzh1 committed
377
				"repeating", format.HumanBytes2(m.memoryWeights-m.memoryLayerOutput),
mashun1's avatar
v1  
mashun1 committed
378
				// memory of non-repeating layers
xuxzh1's avatar
init  
xuxzh1 committed
379
				"nonrepeating", format.HumanBytes2(m.memoryLayerOutput),
mashun1's avatar
v1  
mashun1 committed
380
381
382
383
			),
			slog.Group(
				"graph",
				// memory of graph when fully offloaded
xuxzh1's avatar
init  
xuxzh1 committed
384
				"full", format.HumanBytes2(m.graphFullOffload),
mashun1's avatar
v1  
mashun1 committed
385
				// memory of graph when not fully offloaded
xuxzh1's avatar
init  
xuxzh1 committed
386
				"partial", format.HumanBytes2(m.graphPartialOffload),
mashun1's avatar
v1  
mashun1 committed
387
388
389
390
			),
		),
	)
}
xuxzh1's avatar
update  
xuxzh1 committed
391
392
393
394
395
396
397
398
399
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

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
}