forked from ollama/ollama
-
Notifications
You must be signed in to change notification settings - Fork 0
/
convert.go
331 lines (280 loc) · 9.15 KB
/
convert.go
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
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
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
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
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
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
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
package convert
import (
"bytes"
"cmp"
"encoding/binary"
"encoding/json"
"fmt"
"io"
"log/slog"
"os"
"path/filepath"
"regexp"
"slices"
"github.com/mitchellh/mapstructure"
"google.golang.org/protobuf/proto"
"github.com/jmorganca/ollama/convert/sentencepiece"
"github.com/jmorganca/ollama/llm"
)
type Params struct {
Architectures []string `json:"architectures"`
VocabSize int `json:"vocab_size"`
HiddenSize int `json:"hidden_size"` // n_embd
HiddenLayers int `json:"num_hidden_layers"` // n_layer
ContextSize int `json:"max_position_embeddings"`
IntermediateSize int `json:"intermediate_size"`
AttentionHeads int `json:"num_attention_heads"` // n_head
KeyValHeads int `json:"num_key_value_heads"`
NormEPS float64 `json:"rms_norm_eps"`
RopeFreqBase float64 `json:"rope_theta"`
BoSTokenID int `json:"bos_token_id"`
EoSTokenID int `json:"eos_token_id"`
}
type MetaData struct {
Type string `mapstructure:"dtype"`
Shape []int `mapstructure:"shape"`
Offsets []int `mapstructure:"data_offsets"`
}
func ReadSafeTensors(fn string, offset uint64) ([]llm.Tensor, uint64, error) {
f, err := os.Open(fn)
if err != nil {
return []llm.Tensor{}, 0, err
}
defer f.Close()
var jsonSize uint64
binary.Read(f, binary.LittleEndian, &jsonSize)
buf := make([]byte, jsonSize)
_, err = io.ReadFull(f, buf)
if err != nil {
return []llm.Tensor{}, 0, err
}
d := json.NewDecoder(bytes.NewBuffer(buf))
d.UseNumber()
var parsed map[string]interface{}
if err = d.Decode(&parsed); err != nil {
return []llm.Tensor{}, 0, err
}
var keys []string
for k := range parsed {
keys = append(keys, k)
}
slices.Sort(keys)
slog.Info("converting layers")
var tensors []llm.Tensor
for _, k := range keys {
vals := parsed[k].(map[string]interface{})
var data MetaData
if err = mapstructure.Decode(vals, &data); err != nil {
return []llm.Tensor{}, 0, err
}
var size uint64
var kind uint32
switch len(data.Shape) {
case 0:
// metadata
continue
case 1:
// convert to float32
kind = 0
size = uint64(data.Shape[0] * 4)
case 2:
// convert to float16
kind = 1
size = uint64(data.Shape[0] * data.Shape[1] * 2)
}
ggufName, err := GetTensorName(k)
if err != nil {
slog.Error("%v", err)
return []llm.Tensor{}, 0, err
}
shape := []uint64{0, 0, 0, 0}
for i := range data.Shape {
shape[i] = uint64(data.Shape[i])
}
t := llm.Tensor{
Name: ggufName,
Kind: kind,
Offset: offset,
Shape: shape[:],
FileName: fn,
OffsetPadding: 8 + jsonSize,
FileOffsets: []uint64{uint64(data.Offsets[0]), uint64(data.Offsets[1])},
}
slog.Debug(fmt.Sprintf("%v", t))
tensors = append(tensors, t)
offset += size
}
return tensors, offset, nil
}
func GetSafeTensors(dirpath string) ([]llm.Tensor, error) {
var tensors []llm.Tensor
files, err := filepath.Glob(filepath.Join(dirpath, "/model-*.safetensors"))
if err != nil {
return []llm.Tensor{}, err
}
var offset uint64
for _, f := range files {
var t []llm.Tensor
var err error
t, offset, err = ReadSafeTensors(f, offset)
if err != nil {
slog.Error("%v", err)
return []llm.Tensor{}, err
}
tensors = append(tensors, t...)
}
return tensors, nil
}
func GetParams(dirpath string) (*Params, error) {
f, err := os.Open(filepath.Join(dirpath, "config.json"))
if err != nil {
return nil, err
}
defer f.Close()
var params Params
d := json.NewDecoder(f)
err = d.Decode(¶ms)
if err != nil {
return nil, err
}
return ¶ms, nil
}
// Details on gguf's tokenizer can be found at:
// https://github.com/ggerganov/ggml/blob/master/docs/gguf.md#tokenizer
type Vocab struct {
Tokens []string
Scores []float32
Types []int32
}
func LoadTokens(dirpath string) (*Vocab, error) {
slog.Info(fmt.Sprintf("reading vocab from %s", filepath.Join(dirpath, "tokenizer.model")))
in, err := os.ReadFile(filepath.Join(dirpath, "tokenizer.model"))
if err != nil {
return nil, err
}
// To regenerate sentencepiece from the protobufs use:
// protoc -I=./ --go_out=./ sentencepiece_model.proto
modelProto := &sentencepiece.ModelProto{}
if err := proto.Unmarshal(in, modelProto); err != nil {
return nil, err
}
v := &Vocab{
Tokens: make([]string, 0),
Scores: make([]float32, 0),
Types: make([]int32, 0),
}
pieces := modelProto.GetPieces()
for _, p := range pieces {
v.Tokens = append(v.Tokens, p.GetPiece())
v.Scores = append(v.Scores, p.GetScore())
t := p.GetType()
v.Types = append(v.Types, int32(t))
}
slog.Info(fmt.Sprintf("vocab size: %d", len(v.Tokens)))
// add any additional tokens
addIn, err := os.ReadFile(filepath.Join(dirpath, "added_tokens.json"))
if os.IsNotExist(err) {
return v, nil
} else if err != nil {
return nil, err
}
slog.Info("reading user defined tokens")
var extraTokenData map[string]int
if err := json.Unmarshal(addIn, &extraTokenData); err != nil {
return nil, err
}
type token struct {
key string
pos int
}
extraTokens := make([]token, 0)
for k, id := range extraTokenData {
extraTokens = append(extraTokens, token{k, id})
}
slices.SortFunc(extraTokens, func(a, b token) int {
return cmp.Compare(a.pos, b.pos)
})
numToks := len(v.Tokens)
for cnt, t := range extraTokens {
// the token id should match the specific index for the total number of tokens
if t.pos != cnt+numToks {
return nil, fmt.Errorf("token ID '%d' for '%s' doesn't match total token size", t.pos, t.key)
}
v.Tokens = append(v.Tokens, t.key)
v.Scores = append(v.Scores, -1000.0)
v.Types = append(v.Types, int32(llm.GGUFTokenUserDefined))
}
slog.Info(fmt.Sprintf("vocab size w/ extra tokens: %d", len(v.Tokens)))
return v, nil
}
func GetTensorName(n string) (string, error) {
tMap := map[string]string{
"model.embed_tokens.weight": "token_embd.weight",
"model.layers.(\\d+).input_layernorm.weight": "blk.$1.attn_norm.weight",
"model.layers.(\\d+).mlp.down_proj.weight": "blk.$1.ffn_down.weight",
"model.layers.(\\d+).mlp.gate_proj.weight": "blk.$1.ffn_gate.weight",
"model.layers.(\\d+).mlp.up_proj.weight": "blk.$1.ffn_up.weight",
"model.layers.(\\d+).post_attention_layernorm.weight": "blk.$1.ffn_norm.weight",
"model.layers.(\\d+).self_attn.k_proj.weight": "blk.$1.attn_k.weight",
"model.layers.(\\d+).self_attn.o_proj.weight": "blk.$1.attn_output.weight",
"model.layers.(\\d+).self_attn.q_proj.weight": "blk.$1.attn_q.weight",
"model.layers.(\\d+).self_attn.v_proj.weight": "blk.$1.attn_v.weight",
"lm_head.weight": "output.weight",
"model.norm.weight": "output_norm.weight",
}
v, ok := tMap[n]
if ok {
return v, nil
}
// quick hack to rename the layers to gguf format
for k, v := range tMap {
re := regexp.MustCompile(k)
newName := re.ReplaceAllString(n, v)
if newName != n {
return newName, nil
}
}
return "", fmt.Errorf("couldn't find a layer name for '%s'", n)
}
func WriteGGUF(name string, tensors []llm.Tensor, params *Params, vocab *Vocab) (string, error) {
c := llm.ContainerGGUF{
ByteOrder: binary.LittleEndian,
}
m := llm.NewGGUFModel(&c)
m.Tensors = tensors
m.KV["general.architecture"] = "llama"
m.KV["general.name"] = name
m.KV["llama.context_length"] = uint32(params.ContextSize)
m.KV["llama.embedding_length"] = uint32(params.HiddenSize)
m.KV["llama.block_count"] = uint32(params.HiddenLayers)
m.KV["llama.feed_forward_length"] = uint32(params.IntermediateSize)
m.KV["llama.rope.dimension_count"] = uint32(128)
m.KV["llama.attention.head_count"] = uint32(params.AttentionHeads)
m.KV["llama.attention.head_count_kv"] = uint32(params.KeyValHeads)
m.KV["llama.attention.layer_norm_rms_epsilon"] = float32(params.NormEPS)
m.KV["llama.rope.freq_base"] = float32(params.RopeFreqBase)
m.KV["general.file_type"] = uint32(1)
m.KV["tokenizer.ggml.model"] = "llama"
m.KV["tokenizer.ggml.tokens"] = vocab.Tokens
m.KV["tokenizer.ggml.scores"] = vocab.Scores
m.KV["tokenizer.ggml.token_type"] = vocab.Types
m.KV["tokenizer.ggml.bos_token_id"] = uint32(params.BoSTokenID)
m.KV["tokenizer.ggml.eos_token_id"] = uint32(params.EoSTokenID)
m.KV["tokenizer.ggml.unknown_token_id"] = uint32(0)
m.KV["tokenizer.ggml.add_bos_token"] = true
m.KV["tokenizer.ggml.add_eos_token"] = false
// llamacpp sets the chat template, however we don't need to set it since we pass it in through a layer
// m.KV["tokenizer.chat_template"] = "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}" // XXX removeme
c.V3.NumTensor = uint64(len(tensors))
c.V3.NumKV = uint64(len(m.KV))
f, err := os.CreateTemp("", "ollama-gguf")
if err != nil {
return "", err
}
defer f.Close()
err = m.Encode(f)
if err != nil {
return "", err
}
return f.Name(), nil
}