memory.go 12.3 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

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

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

137
138
139
	// on metal there's no partial offload overhead
	if gpus[0].Library == "metal" {
		graphPartialOffload = graphFullOffload
Daniel Hiltgen's avatar
Daniel Hiltgen committed
140
141
142
	} else if len(gpus) > 1 {
		// multigpu should always use the partial graph size
		graphFullOffload = graphPartialOffload
143
144
	}

145
146
147
148
149
150
151
	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
152
153
	}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
154
	// Output layer handled at the end if we have space
155
	gpuZeroOverhead := projectorWeights + projectorGraph
156
157

	// 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
158
	var layerCount int
159
160
161
162
	layerCounts := make([]int, len(gpus))
	gpuAllocations := make([]uint64, len(gpus))
	type gs struct {
		i int
163
		g *discover.GpuInfo
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
172
		if (gpus[i].FreeMemory - overhead) < gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory+2*layerSize {
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
			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),
			)
188
189
190
			continue
		}
		gpusWithSpace = append(gpusWithSpace, gs{i, &gpus[i]})
Daniel Hiltgen's avatar
Daniel Hiltgen committed
191
		gpuAllocations[i] += gpus[i].MinimumMemory + layerSize // We hold off on graph until we know partial vs. full
192
193
194
195
196
197
198
199
	}

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

Daniel Hiltgen's avatar
Daniel Hiltgen committed
200
	// For all the layers, find where they can fit on the GPU(s)
Michael Yang's avatar
lint  
Michael Yang committed
201
	for i := range int(ggml.KV().BlockCount()) {
202
203
204
205
206
		// 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
207
		memoryWeights += layerSize
Daniel Hiltgen's avatar
Daniel Hiltgen committed
208

209
210
211
212
213
214
215
216
217
		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)
218
			if (g.g.FreeMemory - overhead) > used+layerSize {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
219
				gpuAllocations[g.i] += layerSize
220
				layerCounts[g.i]++
Michael Yang's avatar
typo  
Michael Yang committed
221
				layerCount++
222
223
224
				break
			} else {
				gpusWithSpace = append(gpusWithSpace[:i%j], gpusWithSpace[i%j+1:]...)
Michael Yang's avatar
typo  
Michael Yang committed
225
			}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
226
		}
227
228
229
230
231
	}
	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
232
			overflow += layerSize
233
234
		}
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
235
236

	// Determine if we need to consider output then find where it fits
237
	if memoryLayerOutput > 0 && (opts.NumGPU < 0 || layerCount < opts.NumGPU) {
238
239
240
		for j := len(gpusWithSpace); j > 0; j-- {
			g := gpusWithSpace[layerCount%j]
			used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
241
			if (g.g.FreeMemory - overhead) > used+memoryLayerOutput {
242
243
244
245
246
247
				gpuAllocations[g.i] += memoryLayerOutput
				layerCounts[g.i]++
				layerCount++
				break
			}
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
248

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

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
Daniel Hiltgen's avatar
Daniel Hiltgen committed
270
271
	}

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

279
280
281
282
283
284
285
286
287
288
289
290
	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
291

292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
	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,
309
310
		projectorWeights:    projectorWeights,
		projectorGraph:      projectorGraph,
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() {
330
	overhead := envconfig.GpuOverhead()
331
332
333
334
335
336
337
338
339
340
341
342
343

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