memory.go 9.71 KB
Newer Older
Daniel Hiltgen's avatar
Daniel Hiltgen committed
1
2
3
4
5
package llm

import (
	"fmt"
	"log/slog"
6
7
	"strconv"
	"strings"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
8
9
10
11
12
13
14
15
16

	"github.com/ollama/ollama/api"
	"github.com/ollama/ollama/format"
	"github.com/ollama/ollama/gpu"
)

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

35
36
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
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
55
// 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
56
// The GPUs provided must all be the same Library
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts api.Options) MemoryEstimate {
	// 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
	var projectorSize uint64

	// Conditional output size on GPU 0
	var memoryLayerOutput uint64
	var includeOutput bool

	// One extra layer as a pad for each GPU
	var layerBuffer uint64
Daniel Hiltgen's avatar
Daniel Hiltgen committed
76

77
78
	// The sizes of the main layers
	var layerSizes []uint64
Daniel Hiltgen's avatar
Daniel Hiltgen committed
79

80
81
82
83
84
85
86
87
88
89
90
91
92
93
	// 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

	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
94
95

	for _, projector := range projectors {
96
		projectorSize += projectorMemoryRequirements(projector)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
97
98
99
100
101

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

Michael Yang's avatar
Michael Yang committed
102
	layers := ggml.Tensors().Layers()
Michael Yang's avatar
typo  
Michael Yang committed
103
104
	// add one layer worth of memory as a buffer
	if blk0, ok := layers["blk.0"]; ok {
105
		layerBuffer = blk0.size()
Michael Yang's avatar
typo  
Michael Yang committed
106
	}
Michael Yang's avatar
Michael Yang committed
107

Daniel Hiltgen's avatar
Daniel Hiltgen committed
108
109
110
	// fp16 k,v = (1 (k) + 1 (v)) * sizeof(float16) * n_ctx * n_layer * n_embd / n_head * n_head_kv
	var kv uint64 = 2 * 2 * uint64(opts.NumCtx) * ggml.KV().BlockCount() * ggml.KV().EmbeddingLength() / ggml.KV().HeadCount() * ggml.KV().HeadCountKV()

111
	graphPartialOffload, graphFullOffload = ggml.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)))
Daniel Hiltgen's avatar
Daniel Hiltgen committed
112
113
114
115
116
117
118
	if graphPartialOffload == 0 {
		graphPartialOffload = ggml.KV().GQA() * kv / 6
	}
	if graphFullOffload == 0 {
		graphFullOffload = graphPartialOffload
	}

119
120
121
122
123
	// on metal there's no partial offload overhead
	if gpus[0].Library == "metal" {
		graphPartialOffload = graphFullOffload
	}

124
125
126
127
128
129
130
	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
131
132
133
	}

	if gpus[0].Library == "metal" && opts.UseMMap {
134
135
136
		includeOutput = true
	} else if gpus[0].Library != "metal" || !opts.UseMMap {
		includeOutput = true
Michael Yang's avatar
Michael Yang committed
137
138
	}

139
140
141
142
143
144
	gpuZeroOverhead := projectorSize
	if includeOutput {
		gpuZeroOverhead += memoryLayerOutput
	}

	// 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
145
	var layerCount int
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
	layerCounts := make([]int, len(gpus))
	gpuAllocations := make([]uint64, len(gpus))
	type gs struct {
		i int
		g *gpu.GpuInfo
	}
	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
		if gpus[i].FreeMemory < gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory+2*layerBuffer {
			slog.Debug("gpu has too little memory to allocate any layers", "gpu", gpus[i])
			continue
		}
		gpusWithSpace = append(gpusWithSpace, gs{i, &gpus[i]})
		gpuAllocations[i] += gpus[i].MinimumMemory + layerBuffer // 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
	}

	layerSizes = make([]uint64, int(ggml.KV().BlockCount()))
Michael Yang's avatar
lint  
Michael Yang committed
174
	for i := range int(ggml.KV().BlockCount()) {
Michael Yang's avatar
typo  
Michael Yang committed
175
176
		if blk, ok := layers[fmt.Sprintf("blk.%d", i)]; ok {
			memoryLayer := blk.size()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
177

Michael Yang's avatar
typo  
Michael Yang committed
178
179
			// KV is proportional to the number of layers
			memoryLayer += kv / ggml.KV().BlockCount()
180
181
182
183
			layerSizes[i] = memoryLayer
			memoryWeights += memoryLayer
		}
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
184

185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
	// For all the layers, find where they can fit on the GPU(s)
	for i := range layerSizes {
		if layerSizes[i] == 0 {
			continue
		}
		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)
			if g.g.FreeMemory > used+layerSizes[i] {
				gpuAllocations[g.i] += layerSizes[i]
				layerCounts[g.i]++
Michael Yang's avatar
typo  
Michael Yang committed
202
				layerCount++
203
204
205
				break
			} else {
				gpusWithSpace = append(gpusWithSpace[:i%j], gpusWithSpace[i%j+1:]...)
Michael Yang's avatar
typo  
Michael Yang committed
206
			}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
207
		}
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232

	}
	if layerCount >= int(ggml.KV().BlockCount()) {
		fullyLoaded = true
	} else {
		for i := layerCount; i < int(ggml.KV().BlockCount()); i++ {
			overflow += layerSizes[i]
		}
	}
	// Find where the output fits
	if includeOutput && 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)
			if g.g.FreeMemory > used+memoryLayerOutput {
				gpuAllocations[g.i] += memoryLayerOutput
				layerCounts[g.i]++
				layerCount++
				break
			}
		}
		if layerCount < int(ggml.KV().BlockCount())+1 {
			fullyLoaded = false
			overflow += memoryLayerOutput
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
233
234
	}

235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
	// 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
250
251
	}

252
253
254
255
256
	// Summaries for the log
	var memoryRequiredPartial, memoryRequiredTotal uint64
	for i := range gpuAllocations {
		memoryRequiredPartial += gpuAllocations[i]

Daniel Hiltgen's avatar
Daniel Hiltgen committed
257
	}
258
	memoryRequiredTotal = memoryRequiredPartial + overflow
Daniel Hiltgen's avatar
Daniel Hiltgen committed
259

260
261
262
263
264
265
266
267
268
269
270
271
	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
272
273
274
275
276

	slog.Info(
		"offload to gpu",
		slog.Group(
			"layers",
Michael Yang's avatar
Michael Yang committed
277
278
			// requested number of layers to offload
			"requested", opts.NumGPU,
279
280
			// The number of layers the model has (including output)
			"model", int(ggml.KV().BlockCount())+1,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
281
			// estimated number of layers that can be offloaded
282
283
284
			"offload", layerCount,
			// multi-gpu split for tesnors
			"split", tensorSplit,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
285
286
287
		),
		slog.Group(
			"memory",
288
289
			// memory available by GPU for offloading
			"available", availableList,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
290
291
292
293
294
295
296
297
			slog.Group(
				"required",
				// memory required for full offloading
				"full", format.HumanBytes2(memoryRequiredTotal),
				// memory required to offload layers.estimate layers
				"partial", format.HumanBytes2(memoryRequiredPartial),
				// memory of KV cache
				"kv", format.HumanBytes2(kv),
298
299
				// Allocations across the GPUs
				"allocations", allocationsList,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
			),
			slog.Group(
				"weights",
				// memory of the weights
				"total", format.HumanBytes2(memoryWeights),
				// memory of repeating layers
				"repeating", format.HumanBytes2(memoryWeights-memoryLayerOutput),
				// memory of non-repeating layers
				"nonrepeating", format.HumanBytes2(memoryLayerOutput),
			),
			slog.Group(
				"graph",
				// memory of graph when fully offloaded
				"full", format.HumanBytes2(graphFullOffload),
				// memory of graph when not fully offloaded
				"partial", format.HumanBytes2(graphPartialOffload),
			),
		),
	)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
319
	if gpus[0].Library == "cpu" {
320
321
322
323
324
325
326
		return MemoryEstimate{
			Layers:    0,
			Graph:     0,
			VRAMSize:  0,
			TotalSize: memoryRequiredTotal,
			GPUSizes:  []uint64{},
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
327
	}
328
	if layerCount == 0 {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
329
		slog.Debug("insufficient VRAM to load any model layers")
330
331
332
333
334
335
336
		return MemoryEstimate{
			Layers:    0,
			Graph:     0,
			VRAMSize:  0,
			TotalSize: memoryRequiredTotal,
			GPUSizes:  []uint64{},
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
337
338
	}

339
340
341
342
343
344
345
346
	return MemoryEstimate{
		Layers:      layerCount,
		Graph:       graphOffload,
		VRAMSize:    memoryRequiredPartial,
		TotalSize:   memoryRequiredTotal,
		TensorSplit: tensorSplit,
		GPUSizes:    gpuAllocations,
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
347
}