llama_cpp 0.16.2 → 0.17.0

Sign up to get free protection for your applications and to get access to all the features.
Files changed (177) hide show
  1. checksums.yaml +4 -4
  2. data/CHANGELOG.md +18 -0
  3. data/README.md +7 -12
  4. data/ext/llama_cpp/extconf.rb +2 -43
  5. data/ext/llama_cpp/llama_cpp.cpp +8 -0
  6. data/lib/llama_cpp/version.rb +3 -3
  7. data/sig/llama_cpp.rbs +3 -0
  8. metadata +2 -171
  9. data/vendor/include/.gitkeep +0 -0
  10. data/vendor/lib/.gitkeep +0 -0
  11. data/vendor/tmp/llama.cpp/LICENSE +0 -21
  12. data/vendor/tmp/llama.cpp/Makefile +0 -1124
  13. data/vendor/tmp/llama.cpp/ggml-alloc.c +0 -1041
  14. data/vendor/tmp/llama.cpp/ggml-alloc.h +0 -76
  15. data/vendor/tmp/llama.cpp/ggml-backend-impl.h +0 -153
  16. data/vendor/tmp/llama.cpp/ggml-backend.c +0 -2225
  17. data/vendor/tmp/llama.cpp/ggml-backend.h +0 -236
  18. data/vendor/tmp/llama.cpp/ggml-blas.cpp +0 -363
  19. data/vendor/tmp/llama.cpp/ggml-blas.h +0 -23
  20. data/vendor/tmp/llama.cpp/ggml-common.h +0 -1805
  21. data/vendor/tmp/llama.cpp/ggml-cuda/acc.cu +0 -47
  22. data/vendor/tmp/llama.cpp/ggml-cuda/arange.cu +0 -34
  23. data/vendor/tmp/llama.cpp/ggml-cuda/argsort.cu +0 -104
  24. data/vendor/tmp/llama.cpp/ggml-cuda/binbcast.cu +0 -280
  25. data/vendor/tmp/llama.cpp/ggml-cuda/clamp.cu +0 -34
  26. data/vendor/tmp/llama.cpp/ggml-cuda/concat.cu +0 -196
  27. data/vendor/tmp/llama.cpp/ggml-cuda/convert.cu +0 -686
  28. data/vendor/tmp/llama.cpp/ggml-cuda/cpy.cu +0 -490
  29. data/vendor/tmp/llama.cpp/ggml-cuda/diagmask.cu +0 -40
  30. data/vendor/tmp/llama.cpp/ggml-cuda/dmmv.cu +0 -674
  31. data/vendor/tmp/llama.cpp/ggml-cuda/fattn-tile-f16.cu +0 -319
  32. data/vendor/tmp/llama.cpp/ggml-cuda/fattn-tile-f32.cu +0 -312
  33. data/vendor/tmp/llama.cpp/ggml-cuda/fattn.cu +0 -345
  34. data/vendor/tmp/llama.cpp/ggml-cuda/getrows.cu +0 -178
  35. data/vendor/tmp/llama.cpp/ggml-cuda/im2col.cu +0 -104
  36. data/vendor/tmp/llama.cpp/ggml-cuda/mmq.cu +0 -88
  37. data/vendor/tmp/llama.cpp/ggml-cuda/mmvq.cu +0 -419
  38. data/vendor/tmp/llama.cpp/ggml-cuda/norm.cu +0 -221
  39. data/vendor/tmp/llama.cpp/ggml-cuda/pad.cu +0 -49
  40. data/vendor/tmp/llama.cpp/ggml-cuda/pool2d.cu +0 -94
  41. data/vendor/tmp/llama.cpp/ggml-cuda/quantize.cu +0 -112
  42. data/vendor/tmp/llama.cpp/ggml-cuda/rope.cu +0 -271
  43. data/vendor/tmp/llama.cpp/ggml-cuda/scale.cu +0 -31
  44. data/vendor/tmp/llama.cpp/ggml-cuda/softmax.cu +0 -206
  45. data/vendor/tmp/llama.cpp/ggml-cuda/sumrows.cu +0 -40
  46. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu +0 -5
  47. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu +0 -5
  48. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu +0 -5
  49. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu +0 -5
  50. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu +0 -5
  51. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu +0 -5
  52. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu +0 -5
  53. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu +0 -5
  54. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu +0 -5
  55. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu +0 -5
  56. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu +0 -5
  57. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu +0 -5
  58. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu +0 -5
  59. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu +0 -5
  60. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu +0 -5
  61. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu +0 -5
  62. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu +0 -5
  63. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu +0 -5
  64. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu +0 -5
  65. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu +0 -5
  66. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu +0 -5
  67. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu +0 -5
  68. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu +0 -5
  69. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu +0 -5
  70. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu +0 -5
  71. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu +0 -5
  72. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu +0 -5
  73. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu +0 -5
  74. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu +0 -5
  75. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu +0 -5
  76. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu +0 -5
  77. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu +0 -5
  78. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu +0 -5
  79. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu +0 -5
  80. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu +0 -5
  81. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu +0 -5
  82. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu +0 -5
  83. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu +0 -5
  84. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu +0 -5
  85. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu +0 -5
  86. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu +0 -5
  87. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu +0 -5
  88. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu +0 -5
  89. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu +0 -5
  90. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu +0 -5
  91. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu +0 -5
  92. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu +0 -5
  93. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu +0 -5
  94. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu +0 -5
  95. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu +0 -5
  96. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu +0 -5
  97. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu +0 -5
  98. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu +0 -5
  99. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu +0 -5
  100. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu +0 -5
  101. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu +0 -5
  102. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu +0 -5
  103. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu +0 -5
  104. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu +0 -5
  105. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu +0 -5
  106. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu +0 -5
  107. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu +0 -5
  108. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu +0 -5
  109. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu +0 -5
  110. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu +0 -5
  111. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu +0 -5
  112. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu +0 -5
  113. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu +0 -5
  114. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu +0 -5
  115. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu +0 -5
  116. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu +0 -5
  117. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu +0 -5
  118. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu +0 -5
  119. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu +0 -5
  120. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu +0 -5
  121. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu +0 -5
  122. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu +0 -5
  123. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu +0 -5
  124. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu +0 -5
  125. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu +0 -5
  126. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu +0 -5
  127. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu +0 -5
  128. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu +0 -5
  129. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu +0 -5
  130. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu +0 -5
  131. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu +0 -5
  132. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu +0 -10
  133. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu +0 -9
  134. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu +0 -10
  135. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu +0 -10
  136. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu +0 -8
  137. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q2_k.cu +0 -5
  138. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q3_k.cu +0 -5
  139. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q4_0.cu +0 -5
  140. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q4_1.cu +0 -5
  141. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q4_k.cu +0 -5
  142. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q5_0.cu +0 -5
  143. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q5_1.cu +0 -5
  144. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q5_k.cu +0 -5
  145. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q6_k.cu +0 -5
  146. data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q8_0.cu +0 -5
  147. data/vendor/tmp/llama.cpp/ggml-cuda/tsembd.cu +0 -47
  148. data/vendor/tmp/llama.cpp/ggml-cuda/unary.cu +0 -314
  149. data/vendor/tmp/llama.cpp/ggml-cuda/upscale.cu +0 -51
  150. data/vendor/tmp/llama.cpp/ggml-cuda.cu +0 -3069
  151. data/vendor/tmp/llama.cpp/ggml-cuda.h +0 -44
  152. data/vendor/tmp/llama.cpp/ggml-impl.h +0 -651
  153. data/vendor/tmp/llama.cpp/ggml-kompute.cpp +0 -2038
  154. data/vendor/tmp/llama.cpp/ggml-kompute.h +0 -46
  155. data/vendor/tmp/llama.cpp/ggml-metal.h +0 -66
  156. data/vendor/tmp/llama.cpp/ggml-metal.m +0 -3273
  157. data/vendor/tmp/llama.cpp/ggml-metal.metal +0 -6540
  158. data/vendor/tmp/llama.cpp/ggml-quants.c +0 -14994
  159. data/vendor/tmp/llama.cpp/ggml-quants.h +0 -133
  160. data/vendor/tmp/llama.cpp/ggml-rpc.cpp +0 -1178
  161. data/vendor/tmp/llama.cpp/ggml-rpc.h +0 -24
  162. data/vendor/tmp/llama.cpp/ggml-sycl.cpp +0 -6351
  163. data/vendor/tmp/llama.cpp/ggml-sycl.h +0 -40
  164. data/vendor/tmp/llama.cpp/ggml-vulkan-shaders.hpp +0 -144508
  165. data/vendor/tmp/llama.cpp/ggml-vulkan.cpp +0 -7183
  166. data/vendor/tmp/llama.cpp/ggml-vulkan.h +0 -29
  167. data/vendor/tmp/llama.cpp/ggml.c +0 -22506
  168. data/vendor/tmp/llama.cpp/ggml.h +0 -2458
  169. data/vendor/tmp/llama.cpp/llama.cpp +0 -18985
  170. data/vendor/tmp/llama.cpp/llama.h +0 -1147
  171. data/vendor/tmp/llama.cpp/scripts/get-flags.mk +0 -38
  172. data/vendor/tmp/llama.cpp/sgemm.cpp +0 -1032
  173. data/vendor/tmp/llama.cpp/sgemm.h +0 -14
  174. data/vendor/tmp/llama.cpp/unicode-data.cpp +0 -7033
  175. data/vendor/tmp/llama.cpp/unicode-data.h +0 -20
  176. data/vendor/tmp/llama.cpp/unicode.cpp +0 -810
  177. data/vendor/tmp/llama.cpp/unicode.h +0 -63
@@ -1,47 +0,0 @@
1
- #include "tsembd.cuh"
2
-
3
- static __global__ void timestep_embedding_f32(const float * timesteps, float * dst, const int nb1, const int dim, const int max_period) {
4
- // blockIDx.y: idx of timesteps->ne[0]
5
- // blockIDx.x: idx of ((dim + 1) / 2) / BLOCK_SIZE
6
- int i = blockIdx.y;
7
- int j = threadIdx.x + blockIdx.x * blockDim.x;
8
- float * embed_data = (float *)((char *)dst + i*nb1);
9
-
10
- if (dim % 2 != 0 && j == ((dim + 1) / 2)) {
11
- embed_data[dim] = 0.f;
12
- }
13
-
14
- int half = dim / 2;
15
- if (j >= half) {
16
- return;
17
- }
18
-
19
- float timestep = timesteps[i];
20
- float freq = (float)expf(-logf(max_period) * j / half);
21
- float arg = timestep * freq;
22
- embed_data[j] = cosf(arg);
23
- embed_data[j + half] = sinf(arg);
24
- }
25
-
26
- static void timestep_embedding_f32_cuda(const float * x, float * dst, const int ne00, const int nb1,
27
- const int dim, const int max_period, cudaStream_t stream) {
28
- int half_ceil = (dim + 1) / 2;
29
- int num_blocks = (half_ceil + CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE;
30
- dim3 gridDim(num_blocks, ne00, 1);
31
- timestep_embedding_f32<<<gridDim, CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE, 0, stream>>>(x, dst, nb1, dim, max_period);
32
- }
33
-
34
- void ggml_cuda_op_timestep_embedding(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
35
- const ggml_tensor * src0 = dst->src[0];
36
- const float * src0_d = (const float *)src0->data;
37
- float * dst_d = (float *)dst->data;
38
- cudaStream_t stream = ctx.stream();
39
-
40
- GGML_ASSERT(src0->type == GGML_TYPE_F32);
41
- GGML_ASSERT(dst->type == GGML_TYPE_F32);
42
-
43
- const int dim = dst->op_params[0];
44
- const int max_period = dst->op_params[1];
45
-
46
- timestep_embedding_f32_cuda(src0_d, dst_d, src0->ne[0], dst->nb[1], dim, max_period, stream);
47
- }
@@ -1,314 +0,0 @@
1
- #include "unary.cuh"
2
-
3
- static __global__ void gelu_f32(const float * x, float * dst, const int k) {
4
- const float GELU_COEF_A = 0.044715f;
5
- const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
6
- const int i = blockDim.x*blockIdx.x + threadIdx.x;
7
-
8
- if (i >= k) {
9
- return;
10
- }
11
-
12
- float xi = x[i];
13
- dst[i] = 0.5f*xi*(1.0f + tanhf(SQRT_2_OVER_PI*xi*(1.0f + GELU_COEF_A*xi*xi)));
14
- }
15
-
16
- static __global__ void gelu_quick_f32(const float * x, float * dst, int k) {
17
- const float GELU_QUICK_COEF = -1.702f;
18
- const int i = blockDim.x*blockIdx.x + threadIdx.x;
19
- if (i >= k) {
20
- return;
21
- }
22
- dst[i] = x[i] * (1.0f / (1.0f + expf(GELU_QUICK_COEF * x[i])));
23
- }
24
-
25
- static __global__ void silu_f32(const float * x, float * dst, const int k) {
26
- const int i = blockDim.x*blockIdx.x + threadIdx.x;
27
-
28
- if (i >= k) {
29
- return;
30
- }
31
- dst[i] = x[i] / (1.0f + expf(-x[i]));
32
- }
33
-
34
- static __global__ void tanh_f32(const float * x, float * dst, int k) {
35
- const int i = blockDim.x*blockIdx.x + threadIdx.x;
36
- if (i >= k) {
37
- return;
38
- }
39
- dst[i] = tanhf(x[i]);
40
- }
41
-
42
- static __global__ void relu_f32(const float * x, float * dst, const int k) {
43
- const int i = blockDim.x*blockIdx.x + threadIdx.x;
44
-
45
- if (i >= k) {
46
- return;
47
- }
48
- dst[i] = fmaxf(x[i], 0);
49
- }
50
-
51
- static __global__ void sigmoid_f32(const float * x, float * dst, const int k) {
52
- const int i = blockDim.x*blockIdx.x + threadIdx.x;
53
-
54
- if (i >= k) {
55
- return;
56
- }
57
- dst[i] = 1.0f / (1.0f + expf(-x[i]));
58
- }
59
-
60
- static __global__ void hardsigmoid_f32(const float * x, float * dst, const int k) {
61
- const int i = blockDim.x*blockIdx.x + threadIdx.x;
62
-
63
- if (i >= k) {
64
- return;
65
- }
66
- dst[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f));
67
- }
68
-
69
- static __global__ void hardswish_f32(const float * x, float * dst, const int k) {
70
- const int i = blockDim.x*blockIdx.x + threadIdx.x;
71
-
72
- if (i >= k) {
73
- return;
74
- }
75
- dst[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f));
76
- }
77
-
78
- static __global__ void leaky_relu_f32(const float * x, float * dst, const int k, const float negative_slope) {
79
- const int i = blockDim.x*blockIdx.x + threadIdx.x;
80
- if (i >= k) {
81
- return;
82
- }
83
- dst[i] = fmaxf(x[i], 0) + fminf(x[i], 0.0f) * negative_slope;
84
- }
85
-
86
- static __global__ void sqr_f32(const float * x, float * dst, const int k) {
87
- const int i = blockDim.x*blockIdx.x + threadIdx.x;
88
-
89
- if (i >= k) {
90
- return;
91
- }
92
- dst[i] = x[i] * x[i];
93
- }
94
-
95
- static __global__ void sqrt_f32(const float * x, float * dst, const int k) {
96
- const int i = blockDim.x*blockIdx.x + threadIdx.x;
97
-
98
- if (i >= k) {
99
- return;
100
- }
101
- dst[i] = sqrtf(x[i]);
102
- }
103
-
104
- static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
105
- const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
106
- gelu_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
107
- }
108
-
109
- static void gelu_quick_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
110
- const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
111
- gelu_quick_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
112
- }
113
-
114
- static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
115
- const int num_blocks = (k + CUDA_SILU_BLOCK_SIZE - 1) / CUDA_SILU_BLOCK_SIZE;
116
- silu_f32<<<num_blocks, CUDA_SILU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
117
- }
118
-
119
- static void tanh_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
120
- const int num_blocks = (k + CUDA_TANH_BLOCK_SIZE - 1) / CUDA_TANH_BLOCK_SIZE;
121
- tanh_f32<<<num_blocks, CUDA_TANH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
122
- }
123
-
124
- static void relu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
125
- const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
126
- relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
127
- }
128
-
129
- static void sigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
130
- const int num_blocks = (k + CUDA_SIGMOID_BLOCK_SIZE - 1) / CUDA_SIGMOID_BLOCK_SIZE;
131
- sigmoid_f32<<<num_blocks, CUDA_SIGMOID_BLOCK_SIZE, 0, stream>>>(x, dst, k);
132
- }
133
-
134
- static void hardsigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
135
- const int num_blocks = (k + CUDA_HARDSIGMOID_BLOCK_SIZE - 1) / CUDA_HARDSIGMOID_BLOCK_SIZE;
136
- hardsigmoid_f32<<<num_blocks, CUDA_HARDSIGMOID_BLOCK_SIZE, 0, stream>>>(x, dst, k);
137
- }
138
-
139
- static void hardswish_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
140
- const int num_blocks = (k + CUDA_HARDSWISH_BLOCK_SIZE - 1) / CUDA_HARDSWISH_BLOCK_SIZE;
141
- hardswish_f32<<<num_blocks, CUDA_HARDSWISH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
142
- }
143
-
144
- static void leaky_relu_f32_cuda(const float * x, float * dst, const int k, const float negative_slope, cudaStream_t stream) {
145
- const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
146
- leaky_relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k, negative_slope);
147
- }
148
-
149
- static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
150
- const int num_blocks = (k + CUDA_SQR_BLOCK_SIZE - 1) / CUDA_SQR_BLOCK_SIZE;
151
- sqr_f32<<<num_blocks, CUDA_SQR_BLOCK_SIZE, 0, stream>>>(x, dst, k);
152
- }
153
-
154
- static void sqrt_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
155
- const int num_blocks = (k + CUDA_SQRT_BLOCK_SIZE - 1) / CUDA_SQRT_BLOCK_SIZE;
156
- sqrt_f32<<<num_blocks, CUDA_SQRT_BLOCK_SIZE, 0, stream>>>(x, dst, k);
157
- }
158
-
159
- void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
160
- const ggml_tensor * src0 = dst->src[0];
161
- const float * src0_d = (const float *)src0->data;
162
- float * dst_d = (float *)dst->data;
163
- cudaStream_t stream = ctx.stream();
164
-
165
- GGML_ASSERT(ggml_is_contiguous(src0));
166
-
167
- GGML_ASSERT(src0->type == GGML_TYPE_F32);
168
- GGML_ASSERT( dst->type == GGML_TYPE_F32);
169
-
170
- gelu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
171
- }
172
-
173
- void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
174
- const ggml_tensor * src0 = dst->src[0];
175
- const float * src0_d = (const float *)src0->data;
176
- float * dst_d = (float *)dst->data;
177
- cudaStream_t stream = ctx.stream();
178
-
179
- GGML_ASSERT(ggml_is_contiguous(src0));
180
-
181
- GGML_ASSERT(src0->type == GGML_TYPE_F32);
182
- GGML_ASSERT( dst->type == GGML_TYPE_F32);
183
-
184
- silu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
185
- }
186
-
187
- void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
188
- const ggml_tensor * src0 = dst->src[0];
189
- const float * src0_d = (const float *)src0->data;
190
- float * dst_d = (float *)dst->data;
191
- cudaStream_t stream = ctx.stream();
192
-
193
- GGML_ASSERT(ggml_is_contiguous(src0));
194
-
195
- GGML_ASSERT(src0->type == GGML_TYPE_F32);
196
- GGML_ASSERT( dst->type == GGML_TYPE_F32);
197
-
198
- gelu_quick_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
199
- }
200
-
201
- void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
202
- const ggml_tensor * src0 = dst->src[0];
203
- const float * src0_d = (const float *)src0->data;
204
- float * dst_d = (float *)dst->data;
205
- cudaStream_t stream = ctx.stream();
206
-
207
- GGML_ASSERT(ggml_is_contiguous(src0));
208
-
209
- GGML_ASSERT(src0->type == GGML_TYPE_F32);
210
- GGML_ASSERT( dst->type == GGML_TYPE_F32);
211
-
212
- tanh_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
213
- }
214
-
215
- void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
216
- const ggml_tensor * src0 = dst->src[0];
217
- const float * src0_d = (const float *)src0->data;
218
- float * dst_d = (float *)dst->data;
219
- cudaStream_t stream = ctx.stream();
220
-
221
- GGML_ASSERT(ggml_is_contiguous(src0));
222
-
223
- GGML_ASSERT(src0->type == GGML_TYPE_F32);
224
- GGML_ASSERT( dst->type == GGML_TYPE_F32);
225
-
226
- relu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
227
- }
228
-
229
- void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
230
- const ggml_tensor * src0 = dst->src[0];
231
- const float * src0_d = (const float *)src0->data;
232
- float * dst_d = (float *)dst->data;
233
- cudaStream_t stream = ctx.stream();
234
-
235
- GGML_ASSERT(ggml_is_contiguous(src0));
236
-
237
- GGML_ASSERT(src0->type == GGML_TYPE_F32);
238
- GGML_ASSERT( dst->type == GGML_TYPE_F32);
239
-
240
- sigmoid_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
241
- }
242
-
243
- void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
244
- const ggml_tensor * src0 = dst->src[0];
245
- const float * src0_d = (const float *)src0->data;
246
- float * dst_d = (float *)dst->data;
247
- cudaStream_t stream = ctx.stream();
248
-
249
- GGML_ASSERT(ggml_is_contiguous(src0));
250
-
251
- GGML_ASSERT(src0->type == GGML_TYPE_F32);
252
- GGML_ASSERT( dst->type == GGML_TYPE_F32);
253
-
254
- hardsigmoid_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
255
- }
256
-
257
- void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
258
- const ggml_tensor * src0 = dst->src[0];
259
- const float * src0_d = (const float *)src0->data;
260
- float * dst_d = (float *)dst->data;
261
- cudaStream_t stream = ctx.stream();
262
-
263
- GGML_ASSERT(ggml_is_contiguous(src0));
264
-
265
- GGML_ASSERT(src0->type == GGML_TYPE_F32);
266
- GGML_ASSERT( dst->type == GGML_TYPE_F32);
267
-
268
- hardswish_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
269
- }
270
-
271
- void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
272
- const ggml_tensor * src0 = dst->src[0];
273
- const float * src0_d = (const float *)src0->data;
274
- float * dst_d = (float *)dst->data;
275
- cudaStream_t stream = ctx.stream();
276
-
277
- GGML_ASSERT(ggml_is_contiguous(src0));
278
-
279
- GGML_ASSERT(src0->type == GGML_TYPE_F32);
280
- GGML_ASSERT( dst->type == GGML_TYPE_F32);
281
-
282
- float negative_slope;
283
- memcpy(&negative_slope, dst->op_params, sizeof(float));
284
-
285
- leaky_relu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), negative_slope, stream);
286
- }
287
-
288
- void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
289
- const ggml_tensor * src0 = dst->src[0];
290
- const float * src0_d = (const float *)src0->data;
291
- float * dst_d = (float *)dst->data;
292
- cudaStream_t stream = ctx.stream();
293
-
294
- GGML_ASSERT(ggml_is_contiguous(src0));
295
-
296
- GGML_ASSERT(src0->type == GGML_TYPE_F32);
297
- GGML_ASSERT( dst->type == GGML_TYPE_F32);
298
-
299
- sqr_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
300
- }
301
-
302
- void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
303
- const ggml_tensor * src0 = dst->src[0];
304
- const float * src0_d = (const float *)src0->data;
305
- float * dst_d = (float *)dst->data;
306
- cudaStream_t stream = ctx.stream();
307
-
308
- GGML_ASSERT(ggml_is_contiguous(src0));
309
-
310
- GGML_ASSERT(src0->type == GGML_TYPE_F32);
311
- GGML_ASSERT( dst->type == GGML_TYPE_F32);
312
-
313
- sqrt_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
314
- }
@@ -1,51 +0,0 @@
1
- #include "upscale.cuh"
2
-
3
- static __global__ void upscale_f32(const float * x, float * dst,
4
- const int nb00, const int nb01, const int nb02, const int nb03,
5
- const int ne10, const int ne11, const int ne12, const int ne13,
6
- const float sf0, const float sf1, const float sf2, const float sf3) {
7
- int index = threadIdx.x + blockIdx.x * blockDim.x;
8
- if (index >= ne10 * ne11 * ne12 * ne13) {
9
- return;
10
- }
11
-
12
- int i10 = index % ne10;
13
- int i11 = (index / ne10) % ne11;
14
- int i12 = (index / (ne10 * ne11)) % ne12;
15
- int i13 = (index / (ne10 * ne11 * ne12)) % ne13;
16
-
17
- int i00 = i10 / sf0;
18
- int i01 = i11 / sf1;
19
- int i02 = i12 / sf2;
20
- int i03 = i13 / sf3;
21
-
22
- dst[index] = *(float *)((char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
23
- }
24
-
25
- static void upscale_f32_cuda(const float * x, float * dst,
26
- const int nb00, const int nb01, const int nb02, const int nb03,
27
- const int ne10, const int ne11, const int ne12, const int ne13,
28
- const float sf0, const float sf1, const float sf2, const float sf3,
29
- cudaStream_t stream) {
30
- int dst_size = ne10 * ne11 * ne12 * ne13;
31
- int num_blocks = (dst_size + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
32
-
33
- upscale_f32<<<num_blocks, CUDA_UPSCALE_BLOCK_SIZE,0,stream>>>(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3);
34
- }
35
-
36
- void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
37
- const ggml_tensor * src0 = dst->src[0];
38
- const float * src0_d = (const float *)src0->data;
39
- float * dst_d = (float *)dst->data;
40
- cudaStream_t stream = ctx.stream();
41
-
42
- GGML_ASSERT(src0->type == GGML_TYPE_F32);
43
- GGML_ASSERT( dst->type == GGML_TYPE_F32);
44
-
45
- const float sf0 = (float)dst->ne[0]/src0->ne[0];
46
- const float sf1 = (float)dst->ne[1]/src0->ne[1];
47
- const float sf2 = (float)dst->ne[2]/src0->ne[2];
48
- const float sf3 = (float)dst->ne[3]/src0->ne[3];
49
-
50
- upscale_f32_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, stream);
51
- }