-
Notifications
You must be signed in to change notification settings - Fork 20
/
bce.cpp
254 lines (209 loc) · 13 KB
/
bce.cpp
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
namespace embedding
{
struct Config : public BaseConfig
{
};
class Tokenizer : public BaseTokenizer
{
public:
Tokenizer(const Config &config)
: BaseTokenizer::BaseTokenizer(config, nullptr)
{
sys_prompt = "";
}
size_t load(tokenizer::DataReader *buffer, int n_vocab) override;
void encode(const std::string &text, std::vector<int> &ids) const override;
protected:
void encode(const std::string &text, std::vector<int> &ids, bool add_bos, bool add_eos, int max_length) const;
};
size_t Tokenizer::load(tokenizer::DataReader *buffer, int n_vocab)
{
tp = new tokenizer::UnigramProcessor(eos_token_id + 1);
tp->RegisterPreprocessor(new tokenizer::TextPrepNewlineToSpaces());
tp->RegisterPreprocessor(new tokenizer::TextPrepDeleteMultiSpaces());
tp->RegisterPreprocessor(new tokenizer::TextPrepAddLeadingSpace());
size_t size = tp->Load(buffer, n_vocab);
return size;
}
void Tokenizer::encode(const std::string &text, std::vector<int> &ids, bool add_bos, bool add_eos, int max_length) const
{
if (add_bos) max_length--;
if (add_eos) max_length--;
if (add_bos)
ids.push_back(bos_token_id);
size_t start = ids.size();
BaseTokenizer::encode(text, ids);
size_t length = ids.size() - start;
if ((max_length > 0) && ((int)length > max_length))
ids.resize(start + max_length);
if (add_eos)
ids.push_back(eos_token_id);
}
void Tokenizer::encode(const std::string &text, std::vector<int> &ids) const
{
// position embedding offset = 2
encode(text, ids, true, true, max_length - 2);
}
class ConditionalGeneration : public BaseModelForConditionalGeneration<
EmbeddingModel<Config, RobertaEmbedding, RobertaBlock, BCEFinalNorm, int, int, int, int, int>>
{
public:
ConditionalGeneration() = default;
ConditionalGeneration(const Config &config, ModelType type, size_t mem_size, size_t scratch_size)
: BaseModelForConditionalGeneration<
EmbeddingModel<Config, RobertaEmbedding, RobertaBlock, BCEFinalNorm, int, int, int, int, int>>(type, config, mem_size, scratch_size),
config(config)
{
constexpr size_t tensor_ovhd = GGML_TENSOR_SIZE + GGML_OBJECT_SIZE;
const size_t num_tensors = 5 + config.num_hidden_layers * 19;
const size_t ctx_size = num_tensors * tensor_ovhd;
w_ctx_.gctx = GGMLContext({.mem_size = ctx_size, .mem_buffer = nullptr, .no_alloc = true});
w_ctx_.dtype = config.dtype;
transformer = new EmbeddingModel<Config, RobertaEmbedding, RobertaBlock, BCEFinalNorm, int, int, int, int, int>(&w_ctx_, config,
config.hidden_size, config.num_attention_heads,
config.intermediate_size, config.num_attention_heads, config.max_length);
}
ConditionalGeneration(const Config &config)
: ConditionalGeneration(config, MODEL_TYPE_BCE_Embedding, MEM_SIZE, SCRATCH_SIZE)
{}
void load(ModelLoader &loader) override
{
loader.read_tensor("embeddings.word_embeddings.weight", transformer->word_embeddings.word_weight);
loader.read_tensor("embeddings.position_embeddings.weight", transformer->word_embeddings.position_weight);
loader.read_tensor("embeddings.LayerNorm.weight", transformer->word_embeddings.ln.weight);
loader.read_tensor("embeddings.LayerNorm.bias", transformer->word_embeddings.ln.bias);
for (int i = 0; i < config.num_hidden_layers; i++)
{
std::string layer_prefix = "encoder.layer." + std::to_string(layer_ids[i]) + '.';
loader.read_tensor(layer_prefix + "attention.self.query.weight", transformer->layers[i].attention.q_proj.weight);
loader.read_tensor(layer_prefix + "attention.self.query.bias", transformer->layers[i].attention.q_proj.bias);
loader.read_tensor(layer_prefix + "attention.self.key.weight", transformer->layers[i].attention.k_proj.weight);
loader.read_tensor(layer_prefix + "attention.self.key.bias", transformer->layers[i].attention.k_proj.bias);
loader.read_tensor(layer_prefix + "attention.self.value.weight", transformer->layers[i].attention.v_proj.weight);
loader.read_tensor(layer_prefix + "attention.self.value.bias", transformer->layers[i].attention.v_proj.bias);
loader.read_tensor(layer_prefix + "attention.output.dense.weight", transformer->layers[i].attention.o_proj.weight);
loader.read_tensor(layer_prefix + "attention.output.dense.bias", transformer->layers[i].attention.o_proj.bias);
loader.read_tensor(layer_prefix + "attention.output.LayerNorm.weight", transformer->layers[i].post_attention_layernorm.weight);
loader.read_tensor(layer_prefix + "attention.output.LayerNorm.bias", transformer->layers[i].post_attention_layernorm.bias);
loader.read_tensor(layer_prefix + "intermediate.dense.weight", transformer->layers[i].mlp.intermediate.weight);
loader.read_tensor(layer_prefix + "intermediate.dense.bias", transformer->layers[i].mlp.intermediate.bias);
loader.read_tensor(layer_prefix + "output.dense.weight", transformer->layers[i].mlp.output.dense.weight);
loader.read_tensor(layer_prefix + "output.dense.bias", transformer->layers[i].mlp.output.dense.bias);
loader.read_tensor(layer_prefix + "output.LayerNorm.weight", transformer->layers[i].mlp.output.norm.weight);
loader.read_tensor(layer_prefix + "output.LayerNorm.bias", transformer->layers[i].mlp.output.norm.bias);
}
CHATLLM_CHECK(ggml_used_mem(w_ctx_.gctx.get()) == ggml_get_mem_size(w_ctx_.gctx.get()))
<< "corrupted model weights";
}
int get_text_embedding_dim(void) const override
{
return config.hidden_size;
}
public:
static constexpr size_t MEM_SIZE = 812ull * 1024 * 1024;
static constexpr size_t SCRATCH_SIZE = 44ull * 1024 * 1024;
Config config;
private:
// hold ggml_context & kv_cache
InitContext w_ctx_; // weight context
};
}
namespace ranker
{
struct Config : public embedding::Config
{
};
class Tokenizer : public embedding::Tokenizer
{
public:
Tokenizer(const Config &config)
: embedding::Tokenizer(config)
{
}
void encode_qa(const std::string &q, const std::string &a, std::vector<int> &ids) const override
{
const int max_length = this->max_length - 2;
std::vector<int> ids_q;
std::vector<int> ids_a;
BaseTokenizer::encode(q, ids_q);
BaseTokenizer::encode(a, ids_a);
int total = (int)ids_q.size() + (int)ids_a.size();
// this is bad
if (total > max_length - 4)
{
int remain = max_length - 4 - (int)ids_q.size();
CHATLLM_CHECK(remain > 0) << "query is TOOOO long.";
ids_a.resize(remain);
}
ids.push_back(bos_token_id);
ids.insert(std::end(ids), std::begin(ids_q), std::end(ids_q));
ids.push_back(eos_token_id);
ids.push_back(eos_token_id);
ids.insert(std::end(ids), std::begin(ids_a), std::end(ids_a));
ids.push_back(eos_token_id);
}
};
class ConditionalGeneration : public BaseModelForConditionalGeneration<
EmbeddingModel<Config, RobertaEmbedding, RobertaBlock, RobertaClassificationHead, int, int, int, int, int>>
{
public:
ConditionalGeneration() = default;
ConditionalGeneration(const Config &config)
: ConditionalGeneration(config, MODEL_TYPE_BCE_ReRanker, MEM_SIZE, SCRATCH_SIZE)
{}
ConditionalGeneration(const Config &config, ModelType type, size_t mem_size, size_t scratch_size)
: BaseModelForConditionalGeneration<
EmbeddingModel<Config, RobertaEmbedding, RobertaBlock, RobertaClassificationHead, int, int, int, int, int>>(type, config, mem_size, scratch_size),
config(config)
{
constexpr size_t tensor_ovhd = GGML_TENSOR_SIZE + GGML_OBJECT_SIZE;
const size_t num_tensors = 9 + config.num_hidden_layers * 19;
const size_t ctx_size = num_tensors * tensor_ovhd;
w_ctx_.gctx = GGMLContext({.mem_size = ctx_size, .mem_buffer = nullptr, .no_alloc = true});
w_ctx_.dtype = config.dtype;
transformer = new EmbeddingModel<Config, RobertaEmbedding, RobertaBlock, RobertaClassificationHead, int, int, int, int, int>(&w_ctx_, config,
config.hidden_size, config.num_attention_heads,
config.intermediate_size, config.num_attention_heads, config.max_length);
}
void load(ModelLoader &loader) override
{
loader.read_tensor("embeddings.word_embeddings.weight", transformer->word_embeddings.word_weight);
loader.read_tensor("embeddings.position_embeddings.weight", transformer->word_embeddings.position_weight);
loader.read_tensor("embeddings.LayerNorm.weight", transformer->word_embeddings.ln.weight);
loader.read_tensor("embeddings.LayerNorm.bias", transformer->word_embeddings.ln.bias);
for (int i = 0; i < config.num_hidden_layers; i++)
{
std::string layer_prefix = "encoder.layer." + std::to_string(layer_ids[i]) + '.';
loader.read_tensor(layer_prefix + "attention.self.query.weight", transformer->layers[i].attention.q_proj.weight);
loader.read_tensor(layer_prefix + "attention.self.query.bias", transformer->layers[i].attention.q_proj.bias);
loader.read_tensor(layer_prefix + "attention.self.key.weight", transformer->layers[i].attention.k_proj.weight);
loader.read_tensor(layer_prefix + "attention.self.key.bias", transformer->layers[i].attention.k_proj.bias);
loader.read_tensor(layer_prefix + "attention.self.value.weight", transformer->layers[i].attention.v_proj.weight);
loader.read_tensor(layer_prefix + "attention.self.value.bias", transformer->layers[i].attention.v_proj.bias);
loader.read_tensor(layer_prefix + "attention.output.dense.weight", transformer->layers[i].attention.o_proj.weight);
loader.read_tensor(layer_prefix + "attention.output.dense.bias", transformer->layers[i].attention.o_proj.bias);
loader.read_tensor(layer_prefix + "attention.output.LayerNorm.weight", transformer->layers[i].post_attention_layernorm.weight);
loader.read_tensor(layer_prefix + "attention.output.LayerNorm.bias", transformer->layers[i].post_attention_layernorm.bias);
loader.read_tensor(layer_prefix + "intermediate.dense.weight", transformer->layers[i].mlp.intermediate.weight);
loader.read_tensor(layer_prefix + "intermediate.dense.bias", transformer->layers[i].mlp.intermediate.bias);
loader.read_tensor(layer_prefix + "output.dense.weight", transformer->layers[i].mlp.output.dense.weight);
loader.read_tensor(layer_prefix + "output.dense.bias", transformer->layers[i].mlp.output.dense.bias);
loader.read_tensor(layer_prefix + "output.LayerNorm.weight", transformer->layers[i].mlp.output.norm.weight);
loader.read_tensor(layer_prefix + "output.LayerNorm.bias", transformer->layers[i].mlp.output.norm.bias);
}
loader.read_tensor("classifier.dense.weight", transformer->final_layernorm.dense.weight);
loader.read_tensor("classifier.dense.bias", transformer->final_layernorm.dense.bias);
loader.read_tensor("classifier.out_proj.weight", transformer->final_layernorm.out_proj.weight);
loader.read_tensor("classifier.out_proj.bias", transformer->final_layernorm.out_proj.bias);
CHATLLM_CHECK(ggml_used_mem(w_ctx_.gctx.get()) == ggml_get_mem_size(w_ctx_.gctx.get()))
<< "corrupted model weights";
}
public:
static constexpr size_t MEM_SIZE = 812ull * 1024 * 1024;
static constexpr size_t SCRATCH_SIZE = 44ull * 1024 * 1024;
Config config;
private:
// hold ggml_context & kv_cache
InitContext w_ctx_; // weight context
};
}