Skip to content

Commit dbc7ac5

Browse files
ggerganovggerganov
committed
speculative : simplify
ggml-ci
1 parent 0f878a6 commit dbc7ac5

File tree

3 files changed

+124
-110
lines changed

3 files changed

+124
-110
lines changed

‎common/speculative.cpp

Lines changed: 115 additions & 49 deletions
Original file line numberDiff line numberDiff line change
@@ -4,21 +4,29 @@
44
#include "common.h"
55
#include "sampling.h"
66

7+
#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE 128
8+
#define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
9+
710
struct common_speculative {
811
struct common_speculative_params params;
912

10-
llama_batch batch_dft;
13+
llama_batch batch;
1114

15+
struct llama_context * ctx;
1216
struct common_sampler * smpl;
1317

14-
llama_tokens prompt_last;
18+
llama_tokens prompt;
1519
};
1620

17-
struct common_speculative * common_speculative_init(struct common_speculative_params params) {
21+
struct common_speculative * common_speculative_init(
22+
struct common_speculative_params params,
23+
struct llama_context * ctx_dft) {
1824
auto * result = new common_speculative {
19-
/* .params = */ params,
20-
/* .batch_dft = */ llama_batch_init(llama_n_batch(params.ctx_dft), 0, 1),
21-
/* .smpl = */ nullptr,
25+
/* .params = */ params,
26+
/* .batch = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1),
27+
/* .ctx = */ ctx_dft,
28+
/* .smpl = */ nullptr,
29+
/* .prompt = */ {},
2230
};
2331

2432
// TODO: optimize or pass from outside?
@@ -36,7 +44,7 @@ struct common_speculative * common_speculative_init(struct common_speculative_pa
3644
COMMON_SAMPLER_TYPE_INFILL,
3745
};
3846

39-
result->smpl = common_sampler_init(params.model_dft, sparams);
47+
result->smpl = common_sampler_init(llama_get_model(ctx_dft), sparams);
4048
}
4149
#else
4250
{
@@ -49,46 +57,104 @@ struct common_speculative * common_speculative_init(struct common_speculative_pa
4957
COMMON_SAMPLER_TYPE_TOP_K,
5058
};
5159
52-
result->smpl = common_sampler_init(params.model_dft, sparams);
60+
result->smpl = common_sampler_init(llama_get_model(ctx_dft), sparams);
5361
}
5462
#endif
5563

56-
result->batch_dft = llama_batch_init(llama_n_batch(params.ctx_dft), 0, 1);
57-
5864
return result;
5965
}
6066

6167
void common_speculative_free(struct common_speculative * spec) {
6268
common_sampler_free(spec->smpl);
6369

64-
llama_batch_free(spec->batch_dft);
70+
llama_batch_free(spec->batch);
6571

6672
delete spec;
6773
}
6874

75+
bool common_speculative_are_compatible(
76+
const struct llama_context * ctx_tgt,
77+
const struct llama_context * ctx_dft) {
78+
const struct llama_model * model_tgt = llama_get_model(ctx_tgt);
79+
const struct llama_model * model_dft = llama_get_model(ctx_dft);
80+
81+
const bool vocab_type_tgt = llama_vocab_type(model_tgt);
82+
LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
83+
84+
const bool vocab_type_dft = llama_vocab_type(model_dft);
85+
LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
86+
87+
if (vocab_type_tgt != vocab_type_dft) {
88+
LOG_ERR("%s: draft model vocab type must match target model to use speculation but "
89+
"vocab_type_dft = %d while vocab_type_tgt = %d\n", __func__, vocab_type_dft, vocab_type_tgt);
90+
return false;
91+
}
92+
93+
if (llama_add_bos_token(model_tgt) != llama_add_bos_token(model_dft) ||
94+
llama_add_eos_token(model_tgt) != llama_add_eos_token(model_dft) ||
95+
llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
96+
llama_token_eos(model_tgt) != llama_token_eos(model_dft)
97+
) {
98+
LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
99+
return false;
100+
}
101+
102+
{
103+
const int n_vocab_tgt = llama_n_vocab(model_tgt);
104+
const int n_vocab_dft = llama_n_vocab(model_dft);
105+
106+
const int vocab_diff = std::abs(n_vocab_tgt - n_vocab_dft);
107+
108+
if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
109+
LOG_ERR("%s: draft model vocab must closely match target model to use speculation but "
110+
"target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
111+
__func__, n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
112+
return false;
113+
}
114+
115+
for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
116+
const char * token_text_tgt = llama_token_get_text(model_tgt, i);
117+
const char * token_text_dft = llama_token_get_text(model_dft, i);
118+
if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
119+
LOG_ERR("%s: draft model vocab must match target model to use speculation but "
120+
"token %d content differs - target '%s', draft '%s'\n", __func__, i,
121+
common_token_to_piece(ctx_tgt, i).c_str(),
122+
common_token_to_piece(ctx_dft, i).c_str());
123+
return false;
124+
}
125+
}
126+
}
127+
128+
return true;
129+
}
130+
69131
void common_speculative_add_draft(
70132
struct common_speculative * spec,
71133
struct llama_batch & batch_tgt,
72-
const llama_tokens & prompt,
134+
const llama_tokens & prompt_tgt,
73135
llama_token id_last,
74136
llama_token n_past_tgt) {
137+
auto & batch = spec->batch;
138+
auto & ctx = spec->ctx;
139+
auto & smpl = spec->smpl;
140+
auto & prompt = spec->prompt;
75141

76142
int reuse_i = 0;
77143
int reuse_n = 0;
78144

79-
const int n_ctx = llama_n_ctx(spec->params.ctx_dft) - spec->params.n_draft;
145+
const int n_ctx = llama_n_ctx(ctx) - spec->params.n_draft;
80146

81-
const int i_start = std::max<int>(0, (int) prompt.size() - n_ctx);
147+
const int i_start = std::max<int>(0, (int) prompt_tgt.size() - n_ctx);
82148

83-
for (int i = 0; i < (int) spec->prompt_last.size(); ++i) {
149+
for (int i = 0; i < (int) prompt.size(); ++i) {
84150
int cur = 0;
85-
while (i_start + cur < (int) prompt.size() &&
86-
i + cur < (int) spec->prompt_last.size() &&
87-
prompt[i_start + cur] == spec->prompt_last[i + cur]) {
151+
while (i_start + cur < (int) prompt_tgt.size() &&
152+
i + cur < (int) prompt.size() &&
153+
prompt_tgt[i_start + cur] == prompt[i + cur]) {
88154
cur++;
89155
}
90156

91-
if ((cur >= spec->params.n_reuse || prompt.size() <= n_ctx) && cur > reuse_n) {
157+
if ((cur >= spec->params.n_reuse || prompt_tgt.size() <= n_ctx) && cur > reuse_n) {
92158
reuse_i = i;
93159
reuse_n = cur;
94160
}
@@ -97,59 +163,59 @@ void common_speculative_add_draft(
97163
LOG_DBG("%s: reuse_i = %d, reuse_n = %d\n", __func__, reuse_i, reuse_n);
98164

99165
if (reuse_n == 0) {
100-
llama_kv_cache_clear(spec->params.ctx_dft);
166+
llama_kv_cache_clear(ctx);
101167

102-
spec->prompt_last.clear();
168+
prompt.clear();
103169
} else {
104-
llama_kv_cache_seq_rm (spec->params.ctx_dft, 0, 0, reuse_i);
105-
llama_kv_cache_seq_rm (spec->params.ctx_dft, 0, reuse_i + reuse_n, -1);
106-
llama_kv_cache_seq_add(spec->params.ctx_dft, 0, reuse_i, -1, -reuse_i);
170+
llama_kv_cache_seq_rm (ctx, 0, 0, reuse_i);
171+
llama_kv_cache_seq_rm (ctx, 0, reuse_i + reuse_n, -1);
172+
llama_kv_cache_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
107173

108-
spec->prompt_last.erase(spec->prompt_last.begin(), spec->prompt_last.begin() + reuse_i);
109-
spec->prompt_last.erase(spec->prompt_last.begin() + reuse_n, spec->prompt_last.end());
174+
prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
175+
prompt.erase(prompt.begin() + reuse_n, prompt.end());
110176
}
111177

112-
common_batch_clear(spec->batch_dft);
178+
common_batch_clear(batch);
113179

114-
for (int i = i_start + reuse_n; i < (int) prompt.size(); ++i) {
115-
//LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt[i]);
116-
common_batch_add(spec->batch_dft, prompt[i], i - i_start, { 0 }, false);
180+
for (int i = i_start + reuse_n; i < (int) prompt_tgt.size(); ++i) {
181+
//LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt_tgt[i]);
182+
common_batch_add(batch, prompt_tgt[i], i - i_start, { 0 }, false);
117183

118-
spec->prompt_last.push_back(prompt[i]);
184+
prompt.push_back(prompt_tgt[i]);
119185
}
120186

121-
const llama_pos n_past = prompt.size() - i_start;
187+
const llama_pos n_past = prompt_tgt.size() - i_start;
122188

123189
LOG_DBG("%s: n_past = %d\n", __func__, n_past);
124190

125-
if (spec->batch_dft.n_tokens > 0) {
126-
LOG_DBG("%s: draft batch: %s\n", __func__, string_from(spec->params.ctx_dft, spec->batch_dft).c_str());
191+
if (batch.n_tokens > 0) {
192+
LOG_DBG("%s: draft batch: %s\n", __func__, string_from(ctx, batch).c_str());
127193

128-
llama_decode(spec->params.ctx_dft, spec->batch_dft);
194+
llama_decode(ctx, batch);
129195
}
130196

131-
common_batch_clear(spec->batch_dft);
132-
common_batch_add (spec->batch_dft, id_last, n_past, { 0 }, true);
197+
common_batch_clear(batch);
198+
common_batch_add (batch, id_last, n_past, { 0 }, true);
133199

134-
spec->prompt_last.push_back(id_last);
200+
prompt.push_back(id_last);
135201

136-
LOG_DBG("%s: prompt_last: %s\n", __func__, string_from(spec->params.ctx_dft, spec->prompt_last).c_str());
202+
LOG_DBG("%s: prompt_last: %s\n", __func__, string_from(ctx, prompt).c_str());
137203

138-
llama_decode(spec->params.ctx_dft, spec->batch_dft);
204+
llama_decode(ctx, batch);
139205

140-
common_sampler_reset(spec->smpl);
206+
common_sampler_reset(smpl);
141207

142208
// sample n_draft tokens from the draft model
143209
for (int i = 0; i < spec->params.n_draft; ++i) {
144-
common_batch_clear(spec->batch_dft);
210+
common_batch_clear(batch);
145211

146-
common_sampler_sample(spec->smpl, spec->params.ctx_dft, 0, true);
212+
common_sampler_sample(smpl, ctx, 0, true);
147213

148-
const auto * cur_p = common_sampler_get_candidates(spec->smpl);
214+
const auto * cur_p = common_sampler_get_candidates(smpl);
149215

150216
for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
151217
LOG_DBG(" - draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
152-
k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(spec->params.ctx_dft, cur_p->data[k].id).c_str());
218+
k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx, cur_p->data[k].id).c_str());
153219
}
154220

155221
// add drafted token for each sequence
@@ -160,20 +226,20 @@ void common_speculative_add_draft(
160226
break;
161227
}
162228

163-
common_sampler_accept(spec->smpl, id, true);
229+
common_sampler_accept(smpl, id, true);
164230

165231
common_batch_add(batch_tgt, id, n_past_tgt + i, { 0 }, true);
166232

167233
if (batch_tgt.n_tokens > spec->params.n_draft) {
168234
break;
169235
}
170236

171-
common_batch_add(spec->batch_dft, id, n_past + i + 1, { 0 }, true);
237+
common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
172238

173239
// evaluate the drafted tokens on the draft model
174-
llama_decode(spec->params.ctx_dft, spec->batch_dft);
240+
llama_decode(ctx, batch);
175241

176-
spec->prompt_last.push_back(id);
242+
prompt.push_back(id);
177243
}
178244

179245
// don't waste time on small batches

‎common/speculative.h

Lines changed: 7 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -11,16 +11,18 @@ struct common_speculative_params {
1111
int n_reuse = 256;
1212

1313
float p_min = 0.9f;
14-
15-
struct llama_model * model_dft = nullptr;
16-
17-
struct llama_context * ctx_dft = nullptr;
1814
};
1915

20-
struct common_speculative * common_speculative_init(struct common_speculative_params params);
16+
struct common_speculative * common_speculative_init(
17+
struct common_speculative_params params,
18+
struct llama_context * ctx_dft);
2119

2220
void common_speculative_free(struct common_speculative * spec);
2321

22+
bool common_speculative_are_compatible(
23+
const struct llama_context * ctx_tgt,
24+
const struct llama_context * ctx_dft);
25+
2426
// sample up to n_draft tokens and add them to the batch using the draft model
2527
//
2628
void common_speculative_add_draft(

‎examples/speculative-simple/speculative-simple.cpp

Lines changed: 2 additions & 56 deletions
Original file line numberDiff line numberDiff line change
@@ -5,21 +5,14 @@
55
#include "log.h"
66
#include "llama.h"
77

8-
#include <algorithm>
98
#include <cstdio>
109
#include <cstring>
1110
#include <string>
1211
#include <vector>
1312

14-
#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE 128
15-
#define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
16-
1713
int main(int argc, char ** argv) {
1814
common_params params;
1915

20-
// needed to get candidate probs even for temp <= 0.0
21-
params.sparams.n_probs = 128;
22-
2316
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
2417
return 1;
2518
}
@@ -63,55 +56,10 @@ int main(int argc, char ** argv) {
6356
model_dft = llama_init_dft.model;
6457
ctx_dft = llama_init_dft.context;
6558

66-
const bool vocab_type_tgt = llama_vocab_type(model_tgt);
67-
LOG_DBG("vocab_type tgt: %d\n", vocab_type_tgt);
68-
69-
const bool vocab_type_dft = llama_vocab_type(model_dft);
70-
LOG_DBG("vocab_type dft: %d\n", vocab_type_dft);
71-
72-
if (vocab_type_tgt != vocab_type_dft) {
73-
LOG_ERR("%s: draft model vocab type must match target model to use speculation but ", __func__);
74-
LOG_ERR("vocab_type_dft = %d while vocab_type_tgt = %d\n", vocab_type_dft, vocab_type_tgt);
75-
return 1;
76-
}
77-
78-
if (
79-
llama_add_bos_token(model_tgt) != llama_add_bos_token(model_dft) ||
80-
llama_add_eos_token(model_tgt) != llama_add_eos_token(model_dft) ||
81-
llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
82-
llama_token_eos(model_tgt) != llama_token_eos(model_dft)
83-
) {
84-
LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
59+
if (!common_speculative_are_compatible(ctx_tgt, ctx_dft)) {
8560
return 1;
8661
}
8762

88-
{
89-
const int n_vocab_tgt = llama_n_vocab(model_tgt);
90-
const int n_vocab_dft = llama_n_vocab(model_dft);
91-
const int vocab_diff = n_vocab_tgt > n_vocab_dft
92-
? n_vocab_tgt - n_vocab_dft
93-
: n_vocab_dft - n_vocab_tgt;
94-
95-
if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
96-
LOG_ERR("%s: draft model vocab must closely match target model to use speculation but ", __func__);
97-
LOG_ERR("target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
98-
n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
99-
return 1;
100-
}
101-
102-
for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
103-
const char * token_text_tgt = llama_token_get_text(model_tgt, i);
104-
const char * token_text_dft = llama_token_get_text(model_dft, i);
105-
if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
106-
LOG_ERR("%s: draft model vocab must match target model to use speculation but ", __func__);
107-
LOG_ERR("token %d content differs - target '%s', draft '%s'\n", i,
108-
common_token_to_piece(ctx_tgt, i).c_str(),
109-
common_token_to_piece(ctx_dft, i).c_str());
110-
return 1;
111-
}
112-
}
113-
}
114-
11563
// Tokenize the prompt
11664
std::vector<llama_token> inp;
11765
inp = common_tokenize(ctx_tgt, params.prompt, true, true);
@@ -167,10 +115,8 @@ int main(int argc, char ** argv) {
167115
params_spec.n_min = 5;
168116
params_spec.n_reuse = 256;
169117
params_spec.p_min = 0.9f;
170-
params_spec.model_dft = model_dft;
171-
params_spec.ctx_dft = ctx_dft;
172118

173-
struct common_speculative * spec = common_speculative_init(params_spec);
119+
struct common_speculative * spec = common_speculative_init(params_spec, ctx_dft);
174120

175121
llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1);
176122

0 commit comments

Comments
 (0)