Skip to content

Commit e4c122b

Browse files
ggerganovggerganov
committed
speculative : simplify
ggml-ci
1 parent 0f878a6 commit e4c122b

File tree

3 files changed

+126
-110
lines changed

3 files changed

+126
-110
lines changed

‎common/speculative.cpp‎

Lines changed: 117 additions & 49 deletions
Original file line numberDiff line numberDiff line change
@@ -4,21 +4,31 @@
44
#include "common.h"
55
#include "sampling.h"
66

7+
#include <cstring>
8+
9+
#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE 128
10+
#define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
11+
712
struct common_speculative {
813
struct common_speculative_params params;
914

10-
llama_batch batch_dft;
15+
llama_batch batch;
1116

17+
struct llama_context * ctx;
1218
struct common_sampler * smpl;
1319

14-
llama_tokens prompt_last;
20+
llama_tokens prompt;
1521
};
1622

17-
struct common_speculative * common_speculative_init(struct common_speculative_params params) {
23+
struct common_speculative * common_speculative_init(
24+
struct common_speculative_params params,
25+
struct llama_context * ctx_dft) {
1826
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,
27+
/* .params = */ params,
28+
/* .batch = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1),
29+
/* .ctx = */ ctx_dft,
30+
/* .smpl = */ nullptr,
31+
/* .prompt = */ {},
2232
};
2333

2434
// TODO: optimize or pass from outside?
@@ -36,7 +46,7 @@ struct common_speculative * common_speculative_init(struct common_speculative_pa
3646
COMMON_SAMPLER_TYPE_INFILL,
3747
};
3848

39-
result->smpl = common_sampler_init(params.model_dft, sparams);
49+
result->smpl = common_sampler_init(llama_get_model(ctx_dft), sparams);
4050
}
4151
#else
4252
{
@@ -49,46 +59,104 @@ struct common_speculative * common_speculative_init(struct common_speculative_pa
4959
COMMON_SAMPLER_TYPE_TOP_K,
5060
};
5161
52-
result->smpl = common_sampler_init(params.model_dft, sparams);
62+
result->smpl = common_sampler_init(llama_get_model(ctx_dft), sparams);
5363
}
5464
#endif
5565

56-
result->batch_dft = llama_batch_init(llama_n_batch(params.ctx_dft), 0, 1);
57-
5866
return result;
5967
}
6068

6169
void common_speculative_free(struct common_speculative * spec) {
6270
common_sampler_free(spec->smpl);
6371

64-
llama_batch_free(spec->batch_dft);
72+
llama_batch_free(spec->batch);
6573

6674
delete spec;
6775
}
6876

77+
bool common_speculative_are_compatible(
78+
const struct llama_context * ctx_tgt,
79+
const struct llama_context * ctx_dft) {
80+
const struct llama_model * model_tgt = llama_get_model(ctx_tgt);
81+
const struct llama_model * model_dft = llama_get_model(ctx_dft);
82+
83+
const bool vocab_type_tgt = llama_vocab_type(model_tgt);
84+
LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
85+
86+
const bool vocab_type_dft = llama_vocab_type(model_dft);
87+
LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
88+
89+
if (vocab_type_tgt != vocab_type_dft) {
90+
LOG_ERR("%s: draft model vocab type must match target model to use speculation but "
91+
"vocab_type_dft = %d while vocab_type_tgt = %d\n", __func__, vocab_type_dft, vocab_type_tgt);
92+
return false;
93+
}
94+
95+
if (llama_add_bos_token(model_tgt) != llama_add_bos_token(model_dft) ||
96+
llama_add_eos_token(model_tgt) != llama_add_eos_token(model_dft) ||
97+
llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
98+
llama_token_eos(model_tgt) != llama_token_eos(model_dft)
99+
) {
100+
LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
101+
return false;
102+
}
103+
104+
{
105+
const int n_vocab_tgt = llama_n_vocab(model_tgt);
106+
const int n_vocab_dft = llama_n_vocab(model_dft);
107+
108+
const int vocab_diff = std::abs(n_vocab_tgt - n_vocab_dft);
109+
110+
if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
111+
LOG_ERR("%s: draft model vocab must closely match target model to use speculation but "
112+
"target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
113+
__func__, n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
114+
return false;
115+
}
116+
117+
for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
118+
const char * token_text_tgt = llama_token_get_text(model_tgt, i);
119+
const char * token_text_dft = llama_token_get_text(model_dft, i);
120+
if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
121+
LOG_ERR("%s: draft model vocab must match target model to use speculation but "
122+
"token %d content differs - target '%s', draft '%s'\n", __func__, i,
123+
common_token_to_piece(ctx_tgt, i).c_str(),
124+
common_token_to_piece(ctx_dft, i).c_str());
125+
return false;
126+
}
127+
}
128+
}
129+
130+
return true;
131+
}
132+
69133
void common_speculative_add_draft(
70134
struct common_speculative * spec,
71135
struct llama_batch & batch_tgt,
72-
const llama_tokens & prompt,
136+
const llama_tokens & prompt_tgt,
73137
llama_token id_last,
74138
llama_token n_past_tgt) {
139+
auto & batch = spec->batch;
140+
auto & ctx = spec->ctx;
141+
auto & smpl = spec->smpl;
142+
auto & prompt = spec->prompt;
75143

76144
int reuse_i = 0;
77145
int reuse_n = 0;
78146

79-
const int n_ctx = llama_n_ctx(spec->params.ctx_dft) - spec->params.n_draft;
147+
const int n_ctx = llama_n_ctx(ctx) - spec->params.n_draft;
80148

81-
const int i_start = std::max<int>(0, (int) prompt.size() - n_ctx);
149+
const int i_start = std::max<int>(0, (int) prompt_tgt.size() - n_ctx);
82150

83-
for (int i = 0; i < (int) spec->prompt_last.size(); ++i) {
151+
for (int i = 0; i < (int) prompt.size(); ++i) {
84152
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]) {
153+
while (i_start + cur < (int) prompt_tgt.size() &&
154+
i + cur < (int) prompt.size() &&
155+
prompt_tgt[i_start + cur] == prompt[i + cur]) {
88156
cur++;
89157
}
90158

91-
if ((cur >= spec->params.n_reuse || prompt.size() <= n_ctx) && cur > reuse_n) {
159+
if ((cur >= spec->params.n_reuse || prompt_tgt.size() <= n_ctx) && cur > reuse_n) {
92160
reuse_i = i;
93161
reuse_n = cur;
94162
}
@@ -97,59 +165,59 @@ void common_speculative_add_draft(
97165
LOG_DBG("%s: reuse_i = %d, reuse_n = %d\n", __func__, reuse_i, reuse_n);
98166

99167
if (reuse_n == 0) {
100-
llama_kv_cache_clear(spec->params.ctx_dft);
168+
llama_kv_cache_clear(ctx);
101169

102-
spec->prompt_last.clear();
170+
prompt.clear();
103171
} 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);
172+
llama_kv_cache_seq_rm (ctx, 0, 0, reuse_i);
173+
llama_kv_cache_seq_rm (ctx, 0, reuse_i + reuse_n, -1);
174+
llama_kv_cache_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
107175

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());
176+
prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
177+
prompt.erase(prompt.begin() + reuse_n, prompt.end());
110178
}
111179

112-
common_batch_clear(spec->batch_dft);
180+
common_batch_clear(batch);
113181

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);
182+
for (int i = i_start + reuse_n; i < (int) prompt_tgt.size(); ++i) {
183+
//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]);
184+
common_batch_add(batch, prompt_tgt[i], i - i_start, { 0 }, false);
117185

118-
spec->prompt_last.push_back(prompt[i]);
186+
prompt.push_back(prompt_tgt[i]);
119187
}
120188

121-
const llama_pos n_past = prompt.size() - i_start;
189+
const llama_pos n_past = prompt_tgt.size() - i_start;
122190

123191
LOG_DBG("%s: n_past = %d\n", __func__, n_past);
124192

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());
193+
if (batch.n_tokens > 0) {
194+
LOG_DBG("%s: draft batch: %s\n", __func__, string_from(ctx, batch).c_str());
127195

128-
llama_decode(spec->params.ctx_dft, spec->batch_dft);
196+
llama_decode(ctx, batch);
129197
}
130198

131-
common_batch_clear(spec->batch_dft);
132-
common_batch_add (spec->batch_dft, id_last, n_past, { 0 }, true);
199+
common_batch_clear(batch);
200+
common_batch_add (batch, id_last, n_past, { 0 }, true);
133201

134-
spec->prompt_last.push_back(id_last);
202+
prompt.push_back(id_last);
135203

136-
LOG_DBG("%s: prompt_last: %s\n", __func__, string_from(spec->params.ctx_dft, spec->prompt_last).c_str());
204+
LOG_DBG("%s: prompt_last: %s\n", __func__, string_from(ctx, prompt).c_str());
137205

138-
llama_decode(spec->params.ctx_dft, spec->batch_dft);
206+
llama_decode(ctx, batch);
139207

140-
common_sampler_reset(spec->smpl);
208+
common_sampler_reset(smpl);
141209

142210
// sample n_draft tokens from the draft model
143211
for (int i = 0; i < spec->params.n_draft; ++i) {
144-
common_batch_clear(spec->batch_dft);
212+
common_batch_clear(batch);
145213

146-
common_sampler_sample(spec->smpl, spec->params.ctx_dft, 0, true);
214+
common_sampler_sample(smpl, ctx, 0, true);
147215

148-
const auto * cur_p = common_sampler_get_candidates(spec->smpl);
216+
const auto * cur_p = common_sampler_get_candidates(smpl);
149217

150218
for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
151219
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());
220+
k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx, cur_p->data[k].id).c_str());
153221
}
154222

155223
// add drafted token for each sequence
@@ -160,20 +228,20 @@ void common_speculative_add_draft(
160228
break;
161229
}
162230

163-
common_sampler_accept(spec->smpl, id, true);
231+
common_sampler_accept(smpl, id, true);
164232

165233
common_batch_add(batch_tgt, id, n_past_tgt + i, { 0 }, true);
166234

167235
if (batch_tgt.n_tokens > spec->params.n_draft) {
168236
break;
169237
}
170238

171-
common_batch_add(spec->batch_dft, id, n_past + i + 1, { 0 }, true);
239+
common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
172240

173241
// evaluate the drafted tokens on the draft model
174-
llama_decode(spec->params.ctx_dft, spec->batch_dft);
242+
llama_decode(ctx, batch);
175243

176-
spec->prompt_last.push_back(id);
244+
prompt.push_back(id);
177245
}
178246

179247
// 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)