cui-llama.rn 1.4.0 → 1.4.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (73) hide show
  1. package/android/src/main/jni.cpp +9 -9
  2. package/cpp/common.cpp +163 -60
  3. package/cpp/common.h +43 -12
  4. package/cpp/ggml-alloc.c +1042 -1037
  5. package/cpp/ggml-backend-impl.h +255 -256
  6. package/cpp/ggml-backend-reg.cpp +582 -582
  7. package/cpp/ggml-backend.cpp +2002 -2002
  8. package/cpp/ggml-backend.h +354 -352
  9. package/cpp/ggml-common.h +1853 -1853
  10. package/cpp/ggml-cpp.h +39 -39
  11. package/cpp/ggml-cpu-aarch64.cpp +4247 -4247
  12. package/cpp/ggml-cpu-aarch64.h +8 -8
  13. package/cpp/ggml-cpu-impl.h +386 -386
  14. package/cpp/ggml-cpu-quants.c +10920 -10839
  15. package/cpp/ggml-cpu-traits.cpp +36 -36
  16. package/cpp/ggml-cpu-traits.h +38 -38
  17. package/cpp/ggml-cpu.c +329 -60
  18. package/cpp/ggml-cpu.cpp +10 -2
  19. package/cpp/ggml-cpu.h +135 -135
  20. package/cpp/ggml-impl.h +567 -567
  21. package/cpp/ggml-metal-impl.h +17 -17
  22. package/cpp/ggml-metal.m +4884 -4884
  23. package/cpp/ggml-quants.c +5238 -5238
  24. package/cpp/ggml-threading.h +14 -14
  25. package/cpp/ggml.c +6514 -6448
  26. package/cpp/ggml.h +2194 -2163
  27. package/cpp/gguf.cpp +1329 -1325
  28. package/cpp/gguf.h +202 -202
  29. package/cpp/json-schema-to-grammar.cpp +1045 -1045
  30. package/cpp/json-schema-to-grammar.h +8 -8
  31. package/cpp/json.hpp +24766 -24766
  32. package/cpp/llama-adapter.cpp +347 -346
  33. package/cpp/llama-adapter.h +74 -73
  34. package/cpp/llama-arch.cpp +1487 -1434
  35. package/cpp/llama-arch.h +400 -395
  36. package/cpp/llama-batch.cpp +368 -368
  37. package/cpp/llama-batch.h +88 -88
  38. package/cpp/llama-chat.cpp +578 -567
  39. package/cpp/llama-chat.h +52 -51
  40. package/cpp/llama-context.cpp +1775 -1771
  41. package/cpp/llama-context.h +128 -128
  42. package/cpp/llama-cparams.cpp +1 -1
  43. package/cpp/llama-cparams.h +37 -37
  44. package/cpp/llama-cpp.h +30 -30
  45. package/cpp/llama-grammar.cpp +1139 -1139
  46. package/cpp/llama-grammar.h +143 -143
  47. package/cpp/llama-hparams.cpp +71 -71
  48. package/cpp/llama-hparams.h +139 -140
  49. package/cpp/llama-impl.cpp +167 -167
  50. package/cpp/llama-impl.h +61 -61
  51. package/cpp/llama-kv-cache.cpp +718 -718
  52. package/cpp/llama-kv-cache.h +218 -218
  53. package/cpp/llama-mmap.cpp +2 -1
  54. package/cpp/llama-mmap.h +67 -67
  55. package/cpp/llama-model-loader.cpp +1124 -1011
  56. package/cpp/llama-model-loader.h +167 -158
  57. package/cpp/llama-model.cpp +3997 -2202
  58. package/cpp/llama-model.h +370 -391
  59. package/cpp/llama-sampling.cpp +2408 -2406
  60. package/cpp/llama-sampling.h +32 -48
  61. package/cpp/llama-vocab.cpp +3247 -1982
  62. package/cpp/llama-vocab.h +125 -182
  63. package/cpp/llama.cpp +416 -2886
  64. package/cpp/llama.h +1323 -1285
  65. package/cpp/log.cpp +401 -401
  66. package/cpp/log.h +121 -121
  67. package/cpp/rn-llama.hpp +18 -12
  68. package/cpp/sampling.cpp +505 -500
  69. package/cpp/sgemm.cpp +2597 -2597
  70. package/cpp/speculative.cpp +277 -274
  71. package/cpp/speculative.h +28 -28
  72. package/cpp/unicode.cpp +2 -3
  73. package/package.json +1 -1
package/cpp/llama-adapter.h CHANGED
@@ -1,73 +1,74 @@
1
- #pragma once
2
-
3
- #include "llama-impl.h"
4
- #include "llama-hparams.h"
5
-
6
- #include "ggml-cpp.h"
7
-
8
- #include <unordered_map>
9
- #include <vector>
10
-
11
- //
12
- // llama_adapter_cvec
13
- //
14
-
15
- // TODO: rename to llama_adapter_cvec
16
- struct llama_control_vector {
17
- std::vector<lm_ggml_context_ptr> ctxs;
18
- std::vector<lm_ggml_backend_buffer_ptr> bufs;
19
-
20
- std::vector<struct lm_ggml_tensor *> tensors; // per layer
21
-
22
- int32_t layer_start = -1;
23
- int32_t layer_end = -1;
24
-
25
- struct lm_ggml_tensor * tensor_for(int il) const;
26
-
27
- struct lm_ggml_tensor * apply_to(struct lm_ggml_context * ctx, struct lm_ggml_tensor * cur, int il) const;
28
- };
29
-
30
- int32_t llama_control_vector_apply(
31
- struct llama_control_vector & cvec,
32
- const llama_model & model,
33
- const float * data,
34
- size_t len,
35
- int32_t n_embd,
36
- int32_t il_start,
37
- int32_t il_end);
38
-
39
- //
40
- // llama_adapter_lora
41
- //
42
-
43
- // TODO: rename to llama_adapter_lora_weight
44
- struct llama_lora_weight {
45
- struct lm_ggml_tensor * a = nullptr;
46
- struct lm_ggml_tensor * b = nullptr;
47
-
48
- // get actual scale based on rank and alpha
49
- float get_scale(float alpha, float adapter_scale) {
50
- const float rank = (float) b->ne[0];
51
- const float scale = alpha ? adapter_scale * alpha / rank : adapter_scale;
52
- return scale;
53
- }
54
-
55
- llama_lora_weight() = default;
56
- llama_lora_weight(struct lm_ggml_tensor * a, struct lm_ggml_tensor * b) : a(a), b(b) {}
57
- };
58
-
59
- // TODO: rename to llama_adapter_lora
60
- struct llama_lora_adapter {
61
- // map tensor name to lora_a_b
62
- std::unordered_map<std::string, struct llama_lora_weight> ab_map;
63
-
64
- std::vector<lm_ggml_context_ptr> ctxs;
65
- std::vector<lm_ggml_backend_buffer_ptr> bufs;
66
-
67
- float alpha;
68
-
69
- llama_lora_adapter() = default;
70
- ~llama_lora_adapter() = default;
71
-
72
- llama_lora_weight * get_weight(struct lm_ggml_tensor * w);
73
- };
1
+ #pragma once
2
+
3
+ #include "llama.h"
4
+
5
+ #include "ggml-cpp.h"
6
+
7
+ #include <string>
8
+ #include <unordered_map>
9
+ #include <vector>
10
+
11
+ // TODO: pimpl
12
+
13
+ //
14
+ // llama_adapter_cvec
15
+ //
16
+
17
+ struct llama_adapter_cvec {
18
+ struct lm_ggml_tensor * tensor_for(int il) const;
19
+
20
+ struct lm_ggml_tensor * apply_to(struct lm_ggml_context * ctx, struct lm_ggml_tensor * cur, int il) const;
21
+
22
+ int32_t apply(
23
+ const llama_model & model,
24
+ const float * data,
25
+ size_t len,
26
+ int32_t n_embd,
27
+ int32_t il_start,
28
+ int32_t il_end);
29
+
30
+ private:
31
+ bool init(const llama_model & model);
32
+
33
+ int32_t layer_start = -1;
34
+ int32_t layer_end = -1;
35
+
36
+ std::vector<lm_ggml_context_ptr> ctxs;
37
+ std::vector<lm_ggml_backend_buffer_ptr> bufs;
38
+
39
+ std::vector<struct lm_ggml_tensor *> tensors; // per layer
40
+ };
41
+
42
+ //
43
+ // llama_adapter_lora
44
+ //
45
+
46
+ struct llama_adapter_lora_weight {
47
+ struct lm_ggml_tensor * a = nullptr;
48
+ struct lm_ggml_tensor * b = nullptr;
49
+
50
+ // get actual scale based on rank and alpha
51
+ float get_scale(float alpha, float adapter_scale) const {
52
+ const float rank = (float) b->ne[0];
53
+ const float scale = alpha ? adapter_scale * alpha / rank : adapter_scale;
54
+ return scale;
55
+ }
56
+
57
+ llama_adapter_lora_weight() = default;
58
+ llama_adapter_lora_weight(struct lm_ggml_tensor * a, struct lm_ggml_tensor * b) : a(a), b(b) {}
59
+ };
60
+
61
+ struct llama_adapter_lora {
62
+ // map tensor name to lora_a_b
63
+ std::unordered_map<std::string, struct llama_adapter_lora_weight> ab_map;
64
+
65
+ std::vector<lm_ggml_context_ptr> ctxs;
66
+ std::vector<lm_ggml_backend_buffer_ptr> bufs;
67
+
68
+ float alpha;
69
+
70
+ llama_adapter_lora() = default;
71
+ ~llama_adapter_lora() = default;
72
+
73
+ llama_adapter_lora_weight * get_weight(struct lm_ggml_tensor * w);
74
+ };