memory.go 11 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
7
	"strconv"
	"strings"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
8
9

	"github.com/ollama/ollama/api"
10
	"github.com/ollama/ollama/envconfig"
Daniel Hiltgen's avatar
Daniel Hiltgen committed
11
12
13
14
15
16
17
	"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
18
	var estimatedVRAM uint64
Daniel Hiltgen's avatar
Daniel Hiltgen committed
19
20
	for _, gpus := range allGpus.ByLibrary() {
		var layerCount int
21
22
		estimate := EstimateGPULayers(gpus, ggml, projectors, opts)
		layerCount, estimatedVRAM = estimate.Layers, estimate.VRAMSize
Daniel Hiltgen's avatar
Daniel Hiltgen committed
23
24
25
26
27
28
29
30
31
32
33
34
35
		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
}

36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
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
54
55
56
57
58
59
60
61
62
63
64
65

	// 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
66
67
}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
68
// 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
69
// The GPUs provided must all be the same Library
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
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

Daniel Hiltgen's avatar
Daniel Hiltgen committed
86
87
	// The sizes of a layer
	var layerSize uint64
Daniel Hiltgen's avatar
Daniel Hiltgen committed
88

89
90
91
92
93
94
95
96
97
	// 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

98
	overhead := envconfig.GpuOverhead()
99
100
101
102
103
	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
104
105

	for _, projector := range projectors {
106
		projectorSize += projectorMemoryRequirements(projector)
Daniel Hiltgen's avatar
Daniel Hiltgen committed
107
108
109
110
111

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

Michael Yang's avatar
Michael Yang committed
112
	layers := ggml.Tensors().Layers()
Michael Yang's avatar
typo  
Michael Yang committed
113
114
	// add one layer worth of memory as a buffer
	if blk0, ok := layers["blk.0"]; ok {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
115
116
117
		layerSize = blk0.size()
	} else {
		slog.Warn("model missing blk.0 layer size")
Michael Yang's avatar
typo  
Michael Yang committed
118
	}
Michael Yang's avatar
Michael Yang committed
119

Michael Yang's avatar
Michael Yang committed
120
121
	// fp16 k,v = sizeof(float16) * n_ctx * n_layer * (n_embd_head_k + n_embd_head_v) * n_head_kv
	var kv uint64 = 2 * uint64(opts.NumCtx) * ggml.KV().BlockCount() * (ggml.KV().EmbeddingHeadCountK() + ggml.KV().EmbeddingHeadCountV()) * ggml.KV().HeadCountKV()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
122

Daniel Hiltgen's avatar
Daniel Hiltgen committed
123
124
125
	// KV is proportional to the number of layers
	layerSize += kv / ggml.KV().BlockCount()

126
	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
	}

134
135
136
	// on metal there's no partial offload overhead
	if gpus[0].Library == "metal" {
		graphPartialOffload = graphFullOffload
Daniel Hiltgen's avatar
Daniel Hiltgen committed
137
138
139
	} else if len(gpus) > 1 {
		// multigpu should always use the partial graph size
		graphFullOffload = graphPartialOffload
140
141
	}

142
143
144
145
146
147
148
	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
149
150
	}

Daniel Hiltgen's avatar
Daniel Hiltgen committed
151
	// Output layer handled at the end if we have space
152
153
154
	gpuZeroOverhead := projectorSize

	// 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
155
	var layerCount int
156
157
158
159
160
161
162
163
164
165
166
167
168
	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
169
		if (gpus[i].FreeMemory - overhead) < gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory+2*layerSize {
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
			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),
			)
185
186
187
			continue
		}
		gpusWithSpace = append(gpusWithSpace, gs{i, &gpus[i]})
Daniel Hiltgen's avatar
Daniel Hiltgen committed
188
		gpuAllocations[i] += gpus[i].MinimumMemory + layerSize // We hold off on graph until we know partial vs. full
189
190
191
192
193
194
195
196
	}

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

Daniel Hiltgen's avatar
Daniel Hiltgen committed
197
	// For all the layers, find where they can fit on the GPU(s)
Michael Yang's avatar
lint  
Michael Yang committed
198
	for i := range int(ggml.KV().BlockCount()) {
199
200
201
202
203
		// 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
204
		memoryWeights += layerSize
Daniel Hiltgen's avatar
Daniel Hiltgen committed
205

206
207
208
209
210
211
212
213
214
		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)
215
			if (g.g.FreeMemory - overhead) > used+layerSize {
Daniel Hiltgen's avatar
Daniel Hiltgen committed
216
				gpuAllocations[g.i] += layerSize
217
				layerCounts[g.i]++
Michael Yang's avatar
typo  
Michael Yang committed
218
				layerCount++
219
220
221
				break
			} else {
				gpusWithSpace = append(gpusWithSpace[:i%j], gpusWithSpace[i%j+1:]...)
Michael Yang's avatar
typo  
Michael Yang committed
222
			}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
223
		}
224
225
226
227
228
	}
	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
229
			overflow += layerSize
230
231
		}
	}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
232
233

	// Determine if we need to consider output then find where it fits
234
	if memoryLayerOutput > 0 && (opts.NumGPU < 0 || layerCount < opts.NumGPU) {
235
236
237
		for j := len(gpusWithSpace); j > 0; j-- {
			g := gpusWithSpace[layerCount%j]
			used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
238
			if (g.g.FreeMemory - overhead) > used+memoryLayerOutput {
239
240
241
242
243
244
				gpuAllocations[g.i] += memoryLayerOutput
				layerCounts[g.i]++
				layerCount++
				break
			}
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
245

246
247
248
249
		if layerCount < int(ggml.KV().BlockCount())+1 {
			fullyLoaded = false
			overflow += memoryLayerOutput
		}
Daniel Hiltgen's avatar
Daniel Hiltgen committed
250
251
	}

252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
	// 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
267
268
	}

269
270
271
272
	// Summaries for the log
	var memoryRequiredPartial, memoryRequiredTotal uint64
	for i := range gpuAllocations {
		memoryRequiredPartial += gpuAllocations[i]
Daniel Hiltgen's avatar
Daniel Hiltgen committed
273
	}
274
	memoryRequiredTotal = memoryRequiredPartial + overflow
Daniel Hiltgen's avatar
Daniel Hiltgen committed
275

276
277
278
279
280
281
282
283
284
285
286
287
	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
288

289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
	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,
	}

	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() {
325
	overhead := envconfig.GpuOverhead()
Daniel Hiltgen's avatar
Daniel Hiltgen committed
326
	slog.Info(
327
		"offload to "+m.inferenceLibrary,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
328
329
		slog.Group(
			"layers",
Michael Yang's avatar
Michael Yang committed
330
			// requested number of layers to offload
331
			"requested", m.layersRequested,
332
			// The number of layers the model has (including output)
333
			"model", m.layersModel,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
334
			// estimated number of layers that can be offloaded
335
336
337
			"offload", m.Layers,
			// multi-gpu split for tensors
			"split", m.TensorSplit,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
338
339
340
		),
		slog.Group(
			"memory",
341
			// memory available by GPU for offloading
342
			"available", m.availableList,
343
			"gpu_overhead", format.HumanBytes2(overhead),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
344
345
346
			slog.Group(
				"required",
				// memory required for full offloading
347
				"full", format.HumanBytes2(m.TotalSize),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
348
				// memory required to offload layers.estimate layers
349
				"partial", format.HumanBytes2(m.VRAMSize),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
350
				// memory of KV cache
351
				"kv", format.HumanBytes2(m.kv),
352
				// Allocations across the GPUs
353
				"allocations", m.allocationsList,
Daniel Hiltgen's avatar
Daniel Hiltgen committed
354
355
356
357
			),
			slog.Group(
				"weights",
				// memory of the weights
358
				"total", format.HumanBytes2(m.memoryWeights),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
359
				// memory of repeating layers
360
				"repeating", format.HumanBytes2(m.memoryWeights-m.memoryLayerOutput),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
361
				// memory of non-repeating layers
362
				"nonrepeating", format.HumanBytes2(m.memoryLayerOutput),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
363
364
365
366
			),
			slog.Group(
				"graph",
				// memory of graph when fully offloaded
367
				"full", format.HumanBytes2(m.graphFullOffload),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
368
				// memory of graph when not fully offloaded
369
				"partial", format.HumanBytes2(m.graphPartialOffload),
Daniel Hiltgen's avatar
Daniel Hiltgen committed
370
371
372
373
			),
		),
	)
}