image.go 2.94 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
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
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
package main

import (
	"errors"
	"fmt"
	"hash/maphash"
	"log/slog"
	"sync"
	"time"

	"github.com/ollama/ollama/llama"
)

const imageCacheSize = 4

type ImageContext struct {
	// mu is required to be held when generating embeddings or accessing the cache
	mu sync.Mutex

	clip   *llama.ClipContext
	mllama *llama.MllamaContext

	// cache of images to embeddings
	images    []imageCache
	imageHash maphash.Hash
}

func NewImageContext(llamaContext *llama.Context, modelPath string) (*ImageContext, error) {
	arch, err := llama.GetModelArch(modelPath)
	if err != nil {
		return nil, fmt.Errorf("unable to determine vision architecture: %w (%s)", err, modelPath)
	}

	var c ImageContext
	if arch == "clip" {
		c.clip, err = llama.NewClipContext(llamaContext, modelPath)
	} else if arch == "mllama" {
		c.mllama, err = llama.NewMllamaContext(llamaContext, modelPath)
	} else {
		return nil, fmt.Errorf("unknown vision model architecture: %s", arch)
	}

	if err != nil {
		return nil, err
	}

	c.images = make([]imageCache, imageCacheSize)

	return &c, nil
}

func (c *ImageContext) Free(modelPath string) {
	if c == nil {
		return
	}

	if c.clip != nil {
		c.clip.Free()
	}
	if c.mllama != nil {
		c.mllama.Free()
	}
}

func (c *ImageContext) NewEmbed(llamaContext *llama.Context, data []byte, aspectRatioId int) [][]float32 {
	if c == nil {
		return nil
	}

	hash := c.hashImage(data)

	c.mu.Lock()
	defer c.mu.Unlock()

	embed, err := c.findImage(hash)
	if err != nil {
		if c.mllama != nil {
			embed = c.mllama.NewEmbed(llamaContext, data, aspectRatioId)
		} else if c.clip != nil {
			embed = c.clip.NewEmbed(llamaContext, data)
		} else {
			return nil
		}

		c.addImage(hash, embed)
	}

	return embed
}

func (c *ImageContext) EmbedSize(llamaContext *llama.Context) int {
	if c != nil && c.mllama != nil {
		return c.mllama.EmbedSize(llamaContext)
	} else {
		return llamaContext.Model().NEmbd()
	}
}

type imageCache struct {
	key      uint64
	val      [][]float32
	lastUsed time.Time
}

func (c *ImageContext) hashImage(image []byte) uint64 {
	c.imageHash.Reset()
	_, _ = c.imageHash.Write(image)
	return c.imageHash.Sum64()
}

var errImageNotFound = errors.New("image not found in cache")

func (c *ImageContext) findImage(hash uint64) ([][]float32, error) {
	for i := range c.images {
		if c.images[i].key == hash {
			slog.Debug("loading image embeddings from cache", "entry", i)
			c.images[i].lastUsed = time.Now()
			return c.images[i].val, nil
		}
	}

	return nil, errImageNotFound
}

func (c *ImageContext) addImage(hash uint64, embed [][]float32) {
	best := time.Now()
	var bestImage int

	for i := range c.images {
		if c.images[i].key == hash {
			bestImage = i
			break
		}

		if c.images[i].lastUsed.Compare(best) < 0 {
			best = c.images[i].lastUsed
			bestImage = i
		}
	}

	slog.Debug("storing image embeddings in cache", "entry", bestImage, "used", c.images[bestImage].lastUsed)
	c.images[bestImage].key = hash
	c.images[bestImage].val = embed
	c.images[bestImage].lastUsed = time.Now()
}