llama-rb 0.1.0 → 0.2.0
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/Gemfile +1 -1
- data/Gemfile.lock +1 -3
- data/bin/console +7 -0
- data/ext/Makefile +4 -0
- data/ext/extconf.rb +10 -0
- data/lib/llama/model.rb +36 -64
- data/lib/llama/version.rb +1 -1
- data/lib/llama.rb +0 -1
- data/llama-rb.gemspec +25 -25
- data/llama.cpp/LICENSE +21 -0
- data/llama.cpp/Makefile +175 -0
- data/llama.cpp/README.md +389 -0
- data/{ext/llama → llama.cpp/examples}/common.cpp +10 -3
- data/llama.cpp/examples/main/main.cpp +460 -0
- data/{ext/llama → llama.cpp}/ggml.c +587 -485
- data/{ext/llama → llama.cpp}/ggml.h +36 -26
- data/{ext/llama → llama.cpp}/llama.cpp +85 -35
- data/{ext/llama → llama.cpp}/llama.h +17 -0
- metadata +18 -27
- data/ext/llama/extconf.rb +0 -12
- data/ext/llama/model.cpp +0 -192
- /data/{ext/llama → llama.cpp/examples}/common.h +0 -0
@@ -258,11 +258,11 @@ struct ggml_tensor {
|
|
258
258
|
enum ggml_type type;
|
259
259
|
|
260
260
|
int n_dims;
|
261
|
-
|
262
|
-
size_t
|
263
|
-
|
264
|
-
|
265
|
-
|
261
|
+
int64_t ne[GGML_MAX_DIMS]; // number of elements
|
262
|
+
size_t nb[GGML_MAX_DIMS]; // stride in bytes:
|
263
|
+
// nb[0] = sizeof(type)
|
264
|
+
// nb[1] = nb[0] * ne[0] + padding
|
265
|
+
// nb[i] = nb[i-1] * ne[i-1]
|
266
266
|
|
267
267
|
// compute data
|
268
268
|
enum ggml_op op;
|
@@ -328,8 +328,8 @@ int64_t ggml_cycles_per_ms(void);
|
|
328
328
|
void ggml_print_object (const struct ggml_object * obj);
|
329
329
|
void ggml_print_objects(const struct ggml_context * ctx);
|
330
330
|
|
331
|
-
|
332
|
-
size_t
|
331
|
+
int64_t ggml_nelements(const struct ggml_tensor * tensor);
|
332
|
+
size_t ggml_nbytes (const struct ggml_tensor * tensor);
|
333
333
|
|
334
334
|
int ggml_blck_size (enum ggml_type type);
|
335
335
|
size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
|
@@ -355,33 +355,33 @@ struct ggml_tensor * ggml_new_tensor(
|
|
355
355
|
struct ggml_context * ctx,
|
356
356
|
enum ggml_type type,
|
357
357
|
int n_dims,
|
358
|
-
const
|
358
|
+
const int64_t *ne);
|
359
359
|
|
360
360
|
struct ggml_tensor * ggml_new_tensor_1d(
|
361
361
|
struct ggml_context * ctx,
|
362
362
|
enum ggml_type type,
|
363
|
-
|
363
|
+
int64_t ne0);
|
364
364
|
|
365
365
|
struct ggml_tensor * ggml_new_tensor_2d(
|
366
366
|
struct ggml_context * ctx,
|
367
367
|
enum ggml_type type,
|
368
|
-
|
369
|
-
|
368
|
+
int64_t ne0,
|
369
|
+
int64_t ne1);
|
370
370
|
|
371
371
|
struct ggml_tensor * ggml_new_tensor_3d(
|
372
372
|
struct ggml_context * ctx,
|
373
373
|
enum ggml_type type,
|
374
|
-
|
375
|
-
|
376
|
-
|
374
|
+
int64_t ne0,
|
375
|
+
int64_t ne1,
|
376
|
+
int64_t ne2);
|
377
377
|
|
378
378
|
struct ggml_tensor * ggml_new_tensor_4d(
|
379
379
|
struct ggml_context * ctx,
|
380
380
|
enum ggml_type type,
|
381
|
-
|
382
|
-
|
383
|
-
|
384
|
-
|
381
|
+
int64_t ne0,
|
382
|
+
int64_t ne1,
|
383
|
+
int64_t ne2,
|
384
|
+
int64_t ne3);
|
385
385
|
|
386
386
|
struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
|
387
387
|
struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
|
@@ -531,33 +531,43 @@ struct ggml_tensor * ggml_reshape(
|
|
531
531
|
struct ggml_tensor * ggml_reshape_2d(
|
532
532
|
struct ggml_context * ctx,
|
533
533
|
struct ggml_tensor * a,
|
534
|
-
|
535
|
-
|
534
|
+
int64_t ne0,
|
535
|
+
int64_t ne1);
|
536
536
|
|
537
537
|
// return view(a)
|
538
538
|
// TODO: when we start computing gradient, make a copy instead of view
|
539
539
|
struct ggml_tensor * ggml_reshape_3d(
|
540
540
|
struct ggml_context * ctx,
|
541
541
|
struct ggml_tensor * a,
|
542
|
-
|
543
|
-
|
544
|
-
|
542
|
+
int64_t ne0,
|
543
|
+
int64_t ne1,
|
544
|
+
int64_t ne2);
|
545
545
|
|
546
546
|
// offset in bytes
|
547
547
|
struct ggml_tensor * ggml_view_1d(
|
548
548
|
struct ggml_context * ctx,
|
549
549
|
struct ggml_tensor * a,
|
550
|
-
|
550
|
+
int64_t ne0,
|
551
551
|
size_t offset);
|
552
552
|
|
553
553
|
struct ggml_tensor * ggml_view_2d(
|
554
554
|
struct ggml_context * ctx,
|
555
555
|
struct ggml_tensor * a,
|
556
|
-
|
557
|
-
|
556
|
+
int64_t ne0,
|
557
|
+
int64_t ne1,
|
558
558
|
size_t nb1, // row stride in bytes
|
559
559
|
size_t offset);
|
560
560
|
|
561
|
+
struct ggml_tensor * ggml_view_3d(
|
562
|
+
struct ggml_context * ctx,
|
563
|
+
struct ggml_tensor * a,
|
564
|
+
int64_t ne0,
|
565
|
+
int64_t ne1,
|
566
|
+
int64_t ne2,
|
567
|
+
size_t nb1, // row stride in bytes
|
568
|
+
size_t nb2, // slice stride in bytes
|
569
|
+
size_t offset);
|
570
|
+
|
561
571
|
struct ggml_tensor * ggml_permute(
|
562
572
|
struct ggml_context * ctx,
|
563
573
|
struct ggml_tensor * a,
|
@@ -256,8 +256,8 @@ static bool kv_cache_init(
|
|
256
256
|
const int n_embd = hparams.n_embd;
|
257
257
|
const int n_layer = hparams.n_layer;
|
258
258
|
|
259
|
-
const
|
260
|
-
const
|
259
|
+
const int64_t n_mem = (int64_t)n_layer*n_ctx;
|
260
|
+
const int64_t n_elements = n_embd*n_mem;
|
261
261
|
|
262
262
|
cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
|
263
263
|
|
@@ -679,7 +679,7 @@ static bool llama_model_load(
|
|
679
679
|
return false;
|
680
680
|
}
|
681
681
|
if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
|
682
|
-
fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%
|
682
|
+
fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%" PRId64 ", %" PRId64 "], expected [%d, %d]\n",
|
683
683
|
__func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
|
684
684
|
return false;
|
685
685
|
}
|
@@ -810,37 +810,35 @@ static bool llama_eval_internal(
|
|
810
810
|
|
811
811
|
// self-attention
|
812
812
|
{
|
813
|
-
|
814
|
-
struct ggml_tensor *
|
815
|
-
struct ggml_tensor *
|
813
|
+
// compute Q and K and RoPE them
|
814
|
+
struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0);
|
815
|
+
struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0);
|
816
816
|
|
817
817
|
// store key and value to memory
|
818
|
-
|
818
|
+
{
|
819
|
+
// compute the transposed [N, n_embd] V matrix
|
820
|
+
struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, cur), n_embd, N));
|
821
|
+
|
819
822
|
struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
|
820
|
-
struct ggml_tensor * v =
|
823
|
+
struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
|
824
|
+
( n_ctx)*ggml_element_size(kv_self.v),
|
825
|
+
(il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
|
821
826
|
|
827
|
+
// important: storing RoPE-ed version of K in the KV cache!
|
822
828
|
ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
|
823
829
|
ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
|
824
830
|
}
|
825
831
|
|
826
|
-
// Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
|
827
832
|
struct ggml_tensor * Q =
|
828
833
|
ggml_permute(ctx0,
|
829
|
-
|
830
|
-
ggml_cpy(ctx0,
|
831
|
-
Qcur,
|
832
|
-
ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
|
833
|
-
n_past, n_rot, 0),
|
834
|
+
Qcur,
|
834
835
|
0, 2, 1, 3);
|
835
836
|
|
836
|
-
// K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
|
837
837
|
struct ggml_tensor * K =
|
838
838
|
ggml_permute(ctx0,
|
839
|
-
|
840
|
-
|
841
|
-
|
842
|
-
n_embd/n_head, n_head, n_past + N),
|
843
|
-
n_past, n_rot, 1),
|
839
|
+
ggml_reshape_3d(ctx0,
|
840
|
+
ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd),
|
841
|
+
n_embd/n_head, n_head, n_past + N),
|
844
842
|
0, 2, 1, 3);
|
845
843
|
|
846
844
|
// K * Q
|
@@ -858,18 +856,23 @@ static bool llama_eval_internal(
|
|
858
856
|
// KQ = soft_max(KQ_masked)
|
859
857
|
struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
|
860
858
|
|
861
|
-
//
|
862
|
-
struct ggml_tensor *
|
863
|
-
|
864
|
-
|
865
|
-
|
866
|
-
|
867
|
-
|
868
|
-
1, 2, 0, 3),
|
869
|
-
ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd/n_head, n_head));
|
859
|
+
// split cached V into n_head heads
|
860
|
+
struct ggml_tensor * V =
|
861
|
+
ggml_view_3d(ctx0, kv_self.v,
|
862
|
+
n_past + N, n_embd/n_head, n_head,
|
863
|
+
n_ctx*ggml_element_size(kv_self.v),
|
864
|
+
n_ctx*ggml_element_size(kv_self.v)*n_embd/n_head,
|
865
|
+
il*n_ctx*ggml_element_size(kv_self.v)*n_embd);
|
870
866
|
|
871
|
-
|
872
|
-
struct ggml_tensor * KQV = ggml_mul_mat(ctx0,
|
867
|
+
#if 1
|
868
|
+
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
|
869
|
+
#else
|
870
|
+
// make V contiguous in memory to speed up the matmul, however we waste time on the copy
|
871
|
+
// on M1 this is faster for the perplexity computation, but ~5% slower for the single-token generation
|
872
|
+
// is there a better way?
|
873
|
+
struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd/n_head, n_head));
|
874
|
+
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max);
|
875
|
+
#endif
|
873
876
|
|
874
877
|
// KQV_merged = KQV.permute(0, 2, 1, 3)
|
875
878
|
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
|
@@ -955,9 +958,13 @@ static bool llama_eval_internal(
|
|
955
958
|
ggml_build_forward_expand(&gf, inpL);
|
956
959
|
ggml_graph_compute (ctx0, &gf);
|
957
960
|
|
961
|
+
// print timing information per ggml operation (for debugging purposes)
|
962
|
+
// requires GGML_PERF to be defined
|
963
|
+
//ggml_graph_print(&gf);
|
964
|
+
|
965
|
+
// plot the computation graph in dot format (for debugging purposes)
|
958
966
|
//if (n_past%100 == 0) {
|
959
|
-
//
|
960
|
-
// ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
|
967
|
+
// ggml_graph_dump_dot(&gf, NULL, "llama.dot");
|
961
968
|
//}
|
962
969
|
|
963
970
|
//embd_w.resize(n_vocab*N);
|
@@ -1194,6 +1201,20 @@ static llama_vocab::id llama_sample_top_p_top_k(
|
|
1194
1201
|
const auto & logits = lctx.logits;
|
1195
1202
|
const auto * plogits = logits.data() + logits.size() - n_logits;
|
1196
1203
|
|
1204
|
+
if (temp <= 0) {
|
1205
|
+
// select the token with the highest logit directly
|
1206
|
+
float max_logit = plogits[0];
|
1207
|
+
llama_vocab::id max_id = 0;
|
1208
|
+
|
1209
|
+
for (int i = 1; i < n_logits; ++i) {
|
1210
|
+
if (plogits[i] > max_logit) {
|
1211
|
+
max_logit = plogits[i];
|
1212
|
+
max_id = i;
|
1213
|
+
}
|
1214
|
+
}
|
1215
|
+
return max_id;
|
1216
|
+
}
|
1217
|
+
|
1197
1218
|
std::vector<std::pair<float, llama_vocab::id>> logits_id;
|
1198
1219
|
logits_id.reserve(n_logits);
|
1199
1220
|
|
@@ -1215,7 +1236,9 @@ static llama_vocab::id llama_sample_top_p_top_k(
|
|
1215
1236
|
}
|
1216
1237
|
}
|
1217
1238
|
|
1218
|
-
|
1239
|
+
if (top_k > 0 && top_k < n_logits) {
|
1240
|
+
sample_top_k(logits_id, top_k);
|
1241
|
+
}
|
1219
1242
|
|
1220
1243
|
float maxl = -std::numeric_limits<float>::infinity();
|
1221
1244
|
for (const auto & kv : logits_id) {
|
@@ -1608,7 +1631,7 @@ struct llama_context * llama_init_from_file(
|
|
1608
1631
|
}
|
1609
1632
|
|
1610
1633
|
// reserve memory for context buffers
|
1611
|
-
{
|
1634
|
+
if (!params.vocab_only) {
|
1612
1635
|
if (!kv_cache_init(ctx->model.hparams, ctx->model.kv_self, memory_type, ctx->model.hparams.n_ctx)) {
|
1613
1636
|
fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
|
1614
1637
|
llama_free(ctx);
|
@@ -1668,6 +1691,33 @@ int llama_model_quantize(
|
|
1668
1691
|
return 0;
|
1669
1692
|
}
|
1670
1693
|
|
1694
|
+
// Returns the KV cache that will contain the context for the
|
1695
|
+
// ongoing prediction with the model.
|
1696
|
+
const uint8_t * llama_get_kv_cache(struct llama_context * ctx) {
|
1697
|
+
return ctx->model.kv_self.buf.data();
|
1698
|
+
}
|
1699
|
+
|
1700
|
+
// Returns the size of the KV cache
|
1701
|
+
size_t llama_get_kv_cache_size(struct llama_context * ctx) {
|
1702
|
+
return ctx->model.kv_self.buf.size();
|
1703
|
+
}
|
1704
|
+
|
1705
|
+
int llama_get_kv_cache_token_count(struct llama_context * ctx) {
|
1706
|
+
return ctx->model.kv_self.n;
|
1707
|
+
}
|
1708
|
+
|
1709
|
+
// Sets the KV cache containing the current context for the model
|
1710
|
+
void llama_set_kv_cache(
|
1711
|
+
struct llama_context * ctx,
|
1712
|
+
const uint8_t * kv_cache,
|
1713
|
+
size_t n_size,
|
1714
|
+
int n_token_count) {
|
1715
|
+
// Make sure we have the same kv cache setup
|
1716
|
+
LLAMA_ASSERT(ctx->model.kv_self.buf.size() == n_size);
|
1717
|
+
memcpy(ctx->model.kv_self.buf.data(), kv_cache, n_size);
|
1718
|
+
ctx->model.kv_self.n = n_token_count;
|
1719
|
+
}
|
1720
|
+
|
1671
1721
|
int llama_eval(
|
1672
1722
|
struct llama_context * ctx,
|
1673
1723
|
const llama_token * tokens,
|
@@ -83,6 +83,23 @@ extern "C" {
|
|
83
83
|
const char * fname_out,
|
84
84
|
int itype);
|
85
85
|
|
86
|
+
// Returns the KV cache that will contain the context for the
|
87
|
+
// ongoing prediction with the model.
|
88
|
+
LLAMA_API const uint8_t * llama_get_kv_cache(struct llama_context * ctx);
|
89
|
+
|
90
|
+
// Returns the size of the KV cache
|
91
|
+
LLAMA_API size_t llama_get_kv_cache_size(struct llama_context * ctx);
|
92
|
+
|
93
|
+
// Returns the number of tokens in the KV cache
|
94
|
+
LLAMA_API int llama_get_kv_cache_token_count(struct llama_context * ctx);
|
95
|
+
|
96
|
+
// Sets the KV cache containing the current context for the model
|
97
|
+
LLAMA_API void llama_set_kv_cache(
|
98
|
+
struct llama_context * ctx,
|
99
|
+
const uint8_t * kv_cache,
|
100
|
+
size_t n_size,
|
101
|
+
int n_token_count);
|
102
|
+
|
86
103
|
// Run the llama inference to obtain the logits and probabilities for the next token.
|
87
104
|
// tokens + n_tokens is the provided batch of new tokens to process
|
88
105
|
// n_past is the number of tokens to use from previous eval calls
|
metadata
CHANGED
@@ -1,35 +1,21 @@
|
|
1
1
|
--- !ruby/object:Gem::Specification
|
2
2
|
name: llama-rb
|
3
3
|
version: !ruby/object:Gem::Version
|
4
|
-
version: 0.
|
4
|
+
version: 0.2.0
|
5
5
|
platform: ruby
|
6
6
|
authors:
|
7
7
|
- zfletch
|
8
8
|
autorequire:
|
9
9
|
bindir: exe
|
10
10
|
cert_chain: []
|
11
|
-
date: 2023-04-
|
12
|
-
dependencies:
|
13
|
-
- !ruby/object:Gem::Dependency
|
14
|
-
name: rice
|
15
|
-
requirement: !ruby/object:Gem::Requirement
|
16
|
-
requirements:
|
17
|
-
- - "~>"
|
18
|
-
- !ruby/object:Gem::Version
|
19
|
-
version: 4.0.4
|
20
|
-
type: :runtime
|
21
|
-
prerelease: false
|
22
|
-
version_requirements: !ruby/object:Gem::Requirement
|
23
|
-
requirements:
|
24
|
-
- - "~>"
|
25
|
-
- !ruby/object:Gem::Version
|
26
|
-
version: 4.0.4
|
11
|
+
date: 2023-04-06 00:00:00.000000000 Z
|
12
|
+
dependencies: []
|
27
13
|
description: ggerganov/llama.cpp with Ruby hooks
|
28
14
|
email:
|
29
|
-
-
|
15
|
+
- zf.rubygems@gmail.com
|
30
16
|
executables: []
|
31
17
|
extensions:
|
32
|
-
- ext/
|
18
|
+
- ext/extconf.rb
|
33
19
|
extra_rdoc_files: []
|
34
20
|
files:
|
35
21
|
- Gemfile
|
@@ -37,18 +23,23 @@ files:
|
|
37
23
|
- LICENSE
|
38
24
|
- README.md
|
39
25
|
- Rakefile
|
40
|
-
-
|
41
|
-
- ext/
|
42
|
-
- ext/
|
43
|
-
- ext/llama/ggml.c
|
44
|
-
- ext/llama/ggml.h
|
45
|
-
- ext/llama/llama.cpp
|
46
|
-
- ext/llama/llama.h
|
47
|
-
- ext/llama/model.cpp
|
26
|
+
- bin/console
|
27
|
+
- ext/Makefile
|
28
|
+
- ext/extconf.rb
|
48
29
|
- lib/llama.rb
|
49
30
|
- lib/llama/model.rb
|
50
31
|
- lib/llama/version.rb
|
51
32
|
- llama-rb.gemspec
|
33
|
+
- llama.cpp/LICENSE
|
34
|
+
- llama.cpp/Makefile
|
35
|
+
- llama.cpp/README.md
|
36
|
+
- llama.cpp/examples/common.cpp
|
37
|
+
- llama.cpp/examples/common.h
|
38
|
+
- llama.cpp/examples/main/main.cpp
|
39
|
+
- llama.cpp/ggml.c
|
40
|
+
- llama.cpp/ggml.h
|
41
|
+
- llama.cpp/llama.cpp
|
42
|
+
- llama.cpp/llama.h
|
52
43
|
- models/.gitkeep
|
53
44
|
homepage: https://github.com/zfletch/llama-rb
|
54
45
|
licenses:
|
data/ext/llama/extconf.rb
DELETED
@@ -1,12 +0,0 @@
|
|
1
|
-
require 'mkmf-rice'
|
2
|
-
|
3
|
-
# Compile llama.cpp
|
4
|
-
# root = File.expand_path(File.join(File.dirname(__FILE__), "..", ".."))
|
5
|
-
# llama_cpp = File.join(root, 'llama.cpp')
|
6
|
-
#
|
7
|
-
# Dir.chdir(llama_cpp) do
|
8
|
-
# system("make", exception: true)
|
9
|
-
# end
|
10
|
-
|
11
|
-
# Create Makefile for Ruby bindings
|
12
|
-
create_makefile 'llama/model'
|
data/ext/llama/model.cpp
DELETED
@@ -1,192 +0,0 @@
|
|
1
|
-
#include <rice/rice.hpp>
|
2
|
-
|
3
|
-
#include "common.h"
|
4
|
-
#include "llama.h"
|
5
|
-
|
6
|
-
#include <cassert>
|
7
|
-
#include <cinttypes>
|
8
|
-
#include <cmath>
|
9
|
-
#include <cstdio>
|
10
|
-
#include <cstring>
|
11
|
-
#include <fstream>
|
12
|
-
#include <iostream>
|
13
|
-
#include <string>
|
14
|
-
#include <vector>
|
15
|
-
|
16
|
-
class ModelCpp
|
17
|
-
{
|
18
|
-
public:
|
19
|
-
llama_context *ctx;
|
20
|
-
ModelCpp()
|
21
|
-
{
|
22
|
-
ctx = NULL;
|
23
|
-
}
|
24
|
-
void model_initialize(
|
25
|
-
const char *model,
|
26
|
-
const int32_t n_ctx,
|
27
|
-
const int32_t n_parts,
|
28
|
-
const int32_t seed,
|
29
|
-
const bool memory_f16,
|
30
|
-
const bool use_mlock
|
31
|
-
);
|
32
|
-
Rice::Object model_predict(
|
33
|
-
const char *prompt,
|
34
|
-
const int32_t n_predict
|
35
|
-
);
|
36
|
-
~ModelCpp()
|
37
|
-
{
|
38
|
-
|
39
|
-
if (ctx != NULL) {
|
40
|
-
llama_free(ctx);
|
41
|
-
}
|
42
|
-
}
|
43
|
-
};
|
44
|
-
|
45
|
-
void ModelCpp::model_initialize(
|
46
|
-
const char *model, // path to model file, e.g. "models/7B/ggml-model-q4_0.bin"
|
47
|
-
const int32_t n_ctx, // context size
|
48
|
-
const int32_t n_parts, // amount of model parts (-1 = determine from model dimensions)
|
49
|
-
const int32_t seed, // RNG seed
|
50
|
-
const bool memory_f16, // use f16 instead of f32 for memory kv
|
51
|
-
const bool use_mlock // use mlock to keep model in memory
|
52
|
-
)
|
53
|
-
{
|
54
|
-
auto lparams = llama_context_default_params();
|
55
|
-
|
56
|
-
lparams.n_ctx = n_ctx;
|
57
|
-
lparams.n_parts = n_parts;
|
58
|
-
lparams.seed = seed;
|
59
|
-
lparams.f16_kv = memory_f16;
|
60
|
-
lparams.use_mlock = use_mlock;
|
61
|
-
|
62
|
-
ctx = llama_init_from_file(model, lparams);
|
63
|
-
}
|
64
|
-
|
65
|
-
Rice::Object ModelCpp::model_predict(
|
66
|
-
const char *prompt, // string used as prompt
|
67
|
-
const int32_t n_predict // number of tokens to predict
|
68
|
-
)
|
69
|
-
{
|
70
|
-
std::string return_val = "";
|
71
|
-
|
72
|
-
gpt_params params;
|
73
|
-
params.prompt = prompt;
|
74
|
-
params.n_predict = n_predict;
|
75
|
-
|
76
|
-
// add a space in front of the first character to match OG llama tokenizer behavior
|
77
|
-
params.prompt.insert(0, 1, ' ');
|
78
|
-
|
79
|
-
// tokenize the prompt
|
80
|
-
auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
|
81
|
-
const int n_ctx = llama_n_ctx(ctx);
|
82
|
-
|
83
|
-
// determine newline token
|
84
|
-
auto llama_token_newline = ::llama_tokenize(ctx, "\n", false);
|
85
|
-
|
86
|
-
// generate output
|
87
|
-
{
|
88
|
-
std::vector<llama_token> last_n_tokens(n_ctx);
|
89
|
-
std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
|
90
|
-
|
91
|
-
int n_past = 0;
|
92
|
-
int n_remain = params.n_predict;
|
93
|
-
int n_consumed = 0;
|
94
|
-
|
95
|
-
std::vector<llama_token> embd;
|
96
|
-
|
97
|
-
while (n_remain != 0) {
|
98
|
-
if (embd.size() > 0) {
|
99
|
-
// infinite text generation via context swapping
|
100
|
-
// if we run out of context:
|
101
|
-
// - take the n_keep first tokens from the original prompt (via n_past)
|
102
|
-
// - take half of the last (n_ctx - n_keep) tokens and recompute the logits in a batch
|
103
|
-
if (n_past + (int) embd.size() > n_ctx) {
|
104
|
-
const int n_left = n_past - params.n_keep;
|
105
|
-
|
106
|
-
n_past = params.n_keep;
|
107
|
-
|
108
|
-
// insert n_left/2 tokens at the start of embd from last_n_tokens
|
109
|
-
embd.insert(embd.begin(), last_n_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_n_tokens.end() - embd.size());
|
110
|
-
}
|
111
|
-
|
112
|
-
if (llama_eval(ctx, embd.data(), embd.size(), n_past, params.n_threads)) {
|
113
|
-
throw Rice::Exception(rb_eRuntimeError, "Failed to eval");
|
114
|
-
}
|
115
|
-
}
|
116
|
-
|
117
|
-
|
118
|
-
n_past += embd.size();
|
119
|
-
embd.clear();
|
120
|
-
|
121
|
-
if ((int) embd_inp.size() <= n_consumed) {
|
122
|
-
// out of user input, sample next token
|
123
|
-
const int32_t top_k = params.top_k;
|
124
|
-
const float top_p = params.top_p;
|
125
|
-
const float temp = params.temp;
|
126
|
-
const float repeat_penalty = params.repeat_penalty;
|
127
|
-
|
128
|
-
llama_token id = 0;
|
129
|
-
|
130
|
-
{
|
131
|
-
auto logits = llama_get_logits(ctx);
|
132
|
-
|
133
|
-
if (params.ignore_eos) {
|
134
|
-
logits[llama_token_eos()] = 0;
|
135
|
-
}
|
136
|
-
|
137
|
-
id = llama_sample_top_p_top_k(ctx,
|
138
|
-
last_n_tokens.data() + n_ctx - params.repeat_last_n,
|
139
|
-
params.repeat_last_n, top_k, top_p, temp, repeat_penalty);
|
140
|
-
|
141
|
-
last_n_tokens.erase(last_n_tokens.begin());
|
142
|
-
last_n_tokens.push_back(id);
|
143
|
-
}
|
144
|
-
|
145
|
-
// replace end of text token with newline token when in interactive mode
|
146
|
-
if (id == llama_token_eos() && params.interactive && !params.instruct) {
|
147
|
-
id = llama_token_newline.front();
|
148
|
-
if (params.antiprompt.size() != 0) {
|
149
|
-
// tokenize and inject first reverse prompt
|
150
|
-
const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false);
|
151
|
-
embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
|
152
|
-
}
|
153
|
-
}
|
154
|
-
|
155
|
-
// add it to the context
|
156
|
-
embd.push_back(id);
|
157
|
-
|
158
|
-
// decrement remaining sampling budget
|
159
|
-
--n_remain;
|
160
|
-
} else {
|
161
|
-
// some user input remains from prompt or interaction, forward it to processing
|
162
|
-
while ((int) embd_inp.size() > n_consumed) {
|
163
|
-
embd.push_back(embd_inp[n_consumed]);
|
164
|
-
last_n_tokens.erase(last_n_tokens.begin());
|
165
|
-
last_n_tokens.push_back(embd_inp[n_consumed]);
|
166
|
-
++n_consumed;
|
167
|
-
if ((int) embd.size() >= params.n_batch) {
|
168
|
-
break;
|
169
|
-
}
|
170
|
-
}
|
171
|
-
}
|
172
|
-
|
173
|
-
for (auto id : embd) {
|
174
|
-
return_val += llama_token_to_str(ctx, id);
|
175
|
-
}
|
176
|
-
}
|
177
|
-
}
|
178
|
-
|
179
|
-
Rice::String ruby_return_val(return_val);
|
180
|
-
return ruby_return_val;
|
181
|
-
}
|
182
|
-
|
183
|
-
extern "C"
|
184
|
-
void Init_model()
|
185
|
-
{
|
186
|
-
Rice::Module rb_mLlama = Rice::define_module("Llama");
|
187
|
-
Rice::Data_Type<ModelCpp> rb_cModel =Rice::define_class_under<ModelCpp>(rb_mLlama, "Model");
|
188
|
-
|
189
|
-
rb_cModel.define_constructor(Rice::Constructor<ModelCpp>());
|
190
|
-
rb_cModel.define_method("initialize_cpp", &ModelCpp::model_initialize);
|
191
|
-
rb_cModel.define_method("predict_cpp", &ModelCpp::model_predict);
|
192
|
-
}
|
File without changes
|