whispercpp 1.3.1 → 1.3.3

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 (857) hide show
  1. checksums.yaml +4 -4
  2. data/.gitignore +7 -3
  3. data/README.md +161 -43
  4. data/Rakefile +45 -13
  5. data/ext/.gitignore +4 -8
  6. data/ext/dependencies.rb +73 -0
  7. data/ext/extconf.rb +21 -198
  8. data/ext/options.rb +85 -0
  9. data/ext/ruby_whisper.c +177 -0
  10. data/ext/ruby_whisper.h +17 -2
  11. data/ext/ruby_whisper_context.c +672 -0
  12. data/ext/ruby_whisper_error.c +52 -0
  13. data/ext/ruby_whisper_model.c +232 -0
  14. data/ext/ruby_whisper_params.c +1303 -0
  15. data/ext/ruby_whisper_segment.c +220 -0
  16. data/ext/ruby_whisper_transcribe.cpp +93 -0
  17. data/ext/ruby_whisper_vad_params.c +288 -0
  18. data/ext/sources/CMakeGraphVizOptions.cmake +8 -0
  19. data/ext/sources/CMakeLists.txt +255 -0
  20. data/ext/sources/bindings/javascript/CMakeLists.txt +41 -0
  21. data/ext/sources/bindings/javascript/emscripten.cpp +93 -0
  22. data/ext/sources/bindings/javascript/libwhisper.worker.js +1 -0
  23. data/ext/sources/bindings/javascript/package-tmpl.json +26 -0
  24. data/ext/sources/bindings/javascript/package.json +26 -0
  25. data/ext/sources/bindings/javascript/whisper.js +19 -0
  26. data/ext/sources/build-xcframework.sh +547 -0
  27. data/ext/sources/cmake/DefaultTargetOptions.cmake +16 -0
  28. data/ext/sources/cmake/FindFFmpeg.cmake +163 -0
  29. data/ext/sources/cmake/build-info.cmake +60 -0
  30. data/ext/sources/cmake/git-vars.cmake +22 -0
  31. data/ext/sources/cmake/whisper-config.cmake.in +65 -0
  32. data/ext/sources/cmake/whisper.pc.in +10 -0
  33. data/ext/sources/examples/CMakeLists.txt +124 -0
  34. data/ext/sources/examples/addon.node/CMakeLists.txt +31 -0
  35. data/ext/sources/examples/addon.node/__test__/whisper.spec.js +133 -0
  36. data/ext/sources/examples/addon.node/addon.cpp +557 -0
  37. data/ext/sources/examples/addon.node/index.js +57 -0
  38. data/ext/sources/examples/addon.node/package.json +16 -0
  39. data/ext/sources/examples/addon.node/vad-example.js +132 -0
  40. data/ext/sources/examples/bench/CMakeLists.txt +8 -0
  41. data/ext/sources/examples/bench/bench.cpp +176 -0
  42. data/ext/sources/examples/bench.wasm/CMakeLists.txt +49 -0
  43. data/ext/sources/examples/bench.wasm/emscripten.cpp +87 -0
  44. data/ext/sources/examples/bench.wasm/index-tmpl.html +284 -0
  45. data/ext/sources/examples/cli/CMakeLists.txt +8 -0
  46. data/ext/sources/examples/cli/cli.cpp +1295 -0
  47. data/ext/sources/examples/coi-serviceworker.js +146 -0
  48. data/ext/sources/examples/command/CMakeLists.txt +10 -0
  49. data/ext/sources/examples/command/command.cpp +800 -0
  50. data/ext/sources/examples/command/commands.txt +9 -0
  51. data/ext/sources/examples/command.wasm/CMakeLists.txt +50 -0
  52. data/ext/sources/examples/command.wasm/emscripten.cpp +327 -0
  53. data/ext/sources/examples/command.wasm/index-tmpl.html +414 -0
  54. data/ext/sources/examples/common-ggml.cpp +238 -0
  55. data/ext/sources/examples/common-ggml.h +18 -0
  56. data/ext/sources/examples/common-sdl.cpp +227 -0
  57. data/ext/sources/examples/common-sdl.h +49 -0
  58. data/ext/sources/examples/common-whisper.cpp +175 -0
  59. data/ext/sources/examples/common-whisper.h +24 -0
  60. data/ext/sources/examples/common.cpp +675 -0
  61. data/ext/sources/examples/common.h +322 -0
  62. data/ext/sources/examples/deprecation-warning/CMakeLists.txt +6 -0
  63. data/ext/sources/examples/deprecation-warning/deprecation-warning.cpp +38 -0
  64. data/ext/sources/examples/ffmpeg-transcode.cpp +368 -0
  65. data/ext/sources/examples/generate-karaoke.sh +57 -0
  66. data/ext/sources/examples/grammar-parser.cpp +423 -0
  67. data/ext/sources/examples/grammar-parser.h +29 -0
  68. data/ext/sources/examples/helpers.js +191 -0
  69. data/ext/sources/examples/json.hpp +24596 -0
  70. data/ext/sources/examples/livestream.sh +112 -0
  71. data/ext/sources/examples/lsp/CMakeLists.txt +9 -0
  72. data/ext/sources/examples/lsp/lsp.cpp +469 -0
  73. data/ext/sources/examples/lsp/whisper.vim +362 -0
  74. data/ext/sources/examples/miniaudio.h +93468 -0
  75. data/ext/sources/examples/python/test_whisper_processor.py +7 -0
  76. data/ext/sources/examples/python/whisper_processor.py +54 -0
  77. data/ext/sources/examples/quantize/CMakeLists.txt +6 -0
  78. data/ext/sources/examples/quantize/quantize.cpp +226 -0
  79. data/ext/sources/examples/server/CMakeLists.txt +15 -0
  80. data/ext/sources/examples/server/bench.js +29 -0
  81. data/ext/sources/examples/server/httplib.h +10497 -0
  82. data/ext/sources/examples/server/server.cpp +1238 -0
  83. data/ext/sources/examples/server.py +115 -0
  84. data/ext/sources/examples/stb_vorbis.c +5584 -0
  85. data/ext/sources/examples/stream/CMakeLists.txt +10 -0
  86. data/ext/sources/examples/stream/stream.cpp +435 -0
  87. data/ext/sources/examples/stream.wasm/CMakeLists.txt +49 -0
  88. data/ext/sources/examples/stream.wasm/emscripten.cpp +216 -0
  89. data/ext/sources/examples/stream.wasm/index-tmpl.html +414 -0
  90. data/ext/sources/examples/sycl/CMakeLists.txt +9 -0
  91. data/ext/sources/examples/sycl/build.sh +22 -0
  92. data/ext/sources/examples/sycl/ls-sycl-device.cpp +11 -0
  93. data/ext/sources/examples/sycl/run-whisper.sh +17 -0
  94. data/ext/sources/examples/talk-llama/CMakeLists.txt +43 -0
  95. data/ext/sources/examples/talk-llama/eleven-labs.py +80 -0
  96. data/ext/sources/examples/talk-llama/llama-adapter.cpp +388 -0
  97. data/ext/sources/examples/talk-llama/llama-adapter.h +76 -0
  98. data/ext/sources/examples/talk-llama/llama-arch.cpp +1914 -0
  99. data/ext/sources/examples/talk-llama/llama-arch.h +464 -0
  100. data/ext/sources/examples/talk-llama/llama-batch.cpp +843 -0
  101. data/ext/sources/examples/talk-llama/llama-batch.h +147 -0
  102. data/ext/sources/examples/talk-llama/llama-chat.cpp +685 -0
  103. data/ext/sources/examples/talk-llama/llama-chat.h +59 -0
  104. data/ext/sources/examples/talk-llama/llama-context.cpp +2845 -0
  105. data/ext/sources/examples/talk-llama/llama-context.h +297 -0
  106. data/ext/sources/examples/talk-llama/llama-cparams.cpp +5 -0
  107. data/ext/sources/examples/talk-llama/llama-cparams.h +41 -0
  108. data/ext/sources/examples/talk-llama/llama-grammar.cpp +1229 -0
  109. data/ext/sources/examples/talk-llama/llama-grammar.h +173 -0
  110. data/ext/sources/examples/talk-llama/llama-graph.cpp +1693 -0
  111. data/ext/sources/examples/talk-llama/llama-graph.h +710 -0
  112. data/ext/sources/examples/talk-llama/llama-hparams.cpp +103 -0
  113. data/ext/sources/examples/talk-llama/llama-hparams.h +207 -0
  114. data/ext/sources/examples/talk-llama/llama-impl.cpp +167 -0
  115. data/ext/sources/examples/talk-llama/llama-impl.h +61 -0
  116. data/ext/sources/examples/talk-llama/llama-io.cpp +15 -0
  117. data/ext/sources/examples/talk-llama/llama-io.h +35 -0
  118. data/ext/sources/examples/talk-llama/llama-kv-cache-unified-iswa.cpp +279 -0
  119. data/ext/sources/examples/talk-llama/llama-kv-cache-unified-iswa.h +128 -0
  120. data/ext/sources/examples/talk-llama/llama-kv-cache-unified.cpp +1841 -0
  121. data/ext/sources/examples/talk-llama/llama-kv-cache-unified.h +303 -0
  122. data/ext/sources/examples/talk-llama/llama-kv-cache.h +44 -0
  123. data/ext/sources/examples/talk-llama/llama-kv-cells.h +439 -0
  124. data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +246 -0
  125. data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +138 -0
  126. data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +1125 -0
  127. data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +183 -0
  128. data/ext/sources/examples/talk-llama/llama-memory.cpp +59 -0
  129. data/ext/sources/examples/talk-llama/llama-memory.h +116 -0
  130. data/ext/sources/examples/talk-llama/llama-mmap.cpp +600 -0
  131. data/ext/sources/examples/talk-llama/llama-mmap.h +68 -0
  132. data/ext/sources/examples/talk-llama/llama-model-loader.cpp +1163 -0
  133. data/ext/sources/examples/talk-llama/llama-model-loader.h +169 -0
  134. data/ext/sources/examples/talk-llama/llama-model-saver.cpp +282 -0
  135. data/ext/sources/examples/talk-llama/llama-model-saver.h +37 -0
  136. data/ext/sources/examples/talk-llama/llama-model.cpp +15114 -0
  137. data/ext/sources/examples/talk-llama/llama-model.h +452 -0
  138. data/ext/sources/examples/talk-llama/llama-quant.cpp +1049 -0
  139. data/ext/sources/examples/talk-llama/llama-quant.h +1 -0
  140. data/ext/sources/examples/talk-llama/llama-sampling.cpp +2575 -0
  141. data/ext/sources/examples/talk-llama/llama-sampling.h +32 -0
  142. data/ext/sources/examples/talk-llama/llama-vocab.cpp +3377 -0
  143. data/ext/sources/examples/talk-llama/llama-vocab.h +132 -0
  144. data/ext/sources/examples/talk-llama/llama.cpp +358 -0
  145. data/ext/sources/examples/talk-llama/llama.h +1484 -0
  146. data/ext/sources/examples/talk-llama/prompts/talk-alpaca.txt +23 -0
  147. data/ext/sources/examples/talk-llama/speak +40 -0
  148. data/ext/sources/examples/talk-llama/speak.bat +1 -0
  149. data/ext/sources/examples/talk-llama/speak.ps1 +14 -0
  150. data/ext/sources/examples/talk-llama/talk-llama.cpp +810 -0
  151. data/ext/sources/examples/talk-llama/unicode-data.cpp +7034 -0
  152. data/ext/sources/examples/talk-llama/unicode-data.h +20 -0
  153. data/ext/sources/examples/talk-llama/unicode.cpp +854 -0
  154. data/ext/sources/examples/talk-llama/unicode.h +66 -0
  155. data/ext/sources/examples/vad-speech-segments/CMakeLists.txt +8 -0
  156. data/ext/sources/examples/vad-speech-segments/speech.cpp +149 -0
  157. data/ext/sources/examples/wchess/CMakeLists.txt +10 -0
  158. data/ext/sources/examples/wchess/libwchess/CMakeLists.txt +19 -0
  159. data/ext/sources/examples/wchess/libwchess/Chessboard.cpp +803 -0
  160. data/ext/sources/examples/wchess/libwchess/Chessboard.h +33 -0
  161. data/ext/sources/examples/wchess/libwchess/WChess.cpp +193 -0
  162. data/ext/sources/examples/wchess/libwchess/WChess.h +63 -0
  163. data/ext/sources/examples/wchess/libwchess/test-chessboard.cpp +117 -0
  164. data/ext/sources/examples/wchess/wchess.cmd/CMakeLists.txt +8 -0
  165. data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +251 -0
  166. data/ext/sources/examples/whisper.wasm/CMakeLists.txt +50 -0
  167. data/ext/sources/examples/whisper.wasm/emscripten.cpp +118 -0
  168. data/ext/sources/examples/whisper.wasm/index-tmpl.html +658 -0
  169. data/ext/sources/ggml/CMakeLists.txt +435 -0
  170. data/ext/sources/ggml/cmake/BuildTypes.cmake +54 -0
  171. data/ext/sources/ggml/cmake/GitVars.cmake +22 -0
  172. data/ext/sources/ggml/cmake/common.cmake +50 -0
  173. data/ext/sources/ggml/cmake/ggml-config.cmake.in +152 -0
  174. data/ext/{ggml → sources/ggml}/include/ggml-alloc.h +1 -1
  175. data/ext/{ggml → sources/ggml}/include/ggml-backend.h +10 -8
  176. data/ext/{ggml → sources/ggml}/include/ggml-cpp.h +2 -1
  177. data/ext/{ggml → sources/ggml}/include/ggml-cpu.h +11 -1
  178. data/ext/{ggml → sources/ggml}/include/ggml-metal.h +1 -1
  179. data/ext/{ggml → sources/ggml}/include/ggml-opt.h +49 -28
  180. data/ext/{ggml → sources/ggml}/include/ggml-rpc.h +6 -1
  181. data/ext/{ggml → sources/ggml}/include/ggml-vulkan.h +0 -2
  182. data/ext/{ggml → sources/ggml}/include/ggml.h +325 -269
  183. data/ext/sources/ggml/include/gguf.h +202 -0
  184. data/ext/sources/ggml/src/CMakeLists.txt +404 -0
  185. data/ext/{ggml → sources/ggml}/src/ggml-alloc.c +34 -29
  186. data/ext/sources/ggml/src/ggml-amx/CMakeLists.txt +107 -0
  187. data/ext/{ggml → sources/ggml}/src/ggml-backend-impl.h +1 -2
  188. data/ext/{ggml → sources/ggml}/src/ggml-backend-reg.cpp +92 -53
  189. data/ext/{ggml → sources/ggml}/src/ggml-backend.cpp +69 -34
  190. data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +87 -0
  191. data/ext/sources/ggml/src/ggml-cann/CMakeLists.txt +75 -0
  192. data/ext/sources/ggml/src/ggml-cann/Doxyfile +2579 -0
  193. data/ext/{ggml → sources/ggml}/src/ggml-cann/acl_tensor.cpp +10 -4
  194. data/ext/{ggml → sources/ggml}/src/ggml-cann/acl_tensor.h +5 -5
  195. data/ext/{ggml → sources/ggml}/src/ggml-cann/aclnn_ops.cpp +1272 -1506
  196. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +1125 -0
  197. data/ext/{ggml → sources/ggml}/src/ggml-cann/common.h +140 -1
  198. data/ext/{ggml → sources/ggml}/src/ggml-cann/ggml-cann.cpp +588 -146
  199. data/ext/sources/ggml/src/ggml-cann/kernels/CMakeLists.txt +30 -0
  200. data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/dup.cpp +3 -5
  201. data/ext/{ggml → sources/ggml}/src/ggml-common.h +16 -8
  202. data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +597 -0
  203. data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/amx.cpp +3 -2
  204. data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/mmq.cpp +11 -10
  205. data/ext/sources/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp +94 -0
  206. data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +4114 -0
  207. data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +2163 -0
  208. data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +2639 -0
  209. data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp +82 -0
  210. data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +2732 -0
  211. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +2069 -0
  212. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +397 -0
  213. data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +1300 -0
  214. data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +1481 -0
  215. data/ext/{ggml/src/ggml-cpu/cpu-feats-x86.cpp → sources/ggml/src/ggml-cpu/arch/x86/cpu-feats.cpp} +5 -1
  216. data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +4311 -0
  217. data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +3285 -0
  218. data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +184 -0
  219. data/ext/sources/ggml/src/ggml-cpu/binary-ops.cpp +158 -0
  220. data/ext/sources/ggml/src/ggml-cpu/binary-ops.h +16 -0
  221. data/ext/sources/ggml/src/ggml-cpu/cmake/FindSIMD.cmake +100 -0
  222. data/ext/sources/ggml/src/ggml-cpu/common.h +73 -0
  223. data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu-impl.h +172 -41
  224. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +3551 -0
  225. data/ext/{ggml → sources/ggml}/src/ggml-cpu/ggml-cpu.cpp +78 -25
  226. data/ext/{ggml/src/ggml-cpu/ggml-cpu-hbm.cpp → sources/ggml/src/ggml-cpu/hbm.cpp} +1 -1
  227. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +337 -0
  228. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.h +95 -0
  229. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +482 -0
  230. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.h +17 -0
  231. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +3594 -0
  232. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.h +19 -0
  233. data/ext/sources/ggml/src/ggml-cpu/ops.cpp +9786 -0
  234. data/ext/sources/ggml/src/ggml-cpu/ops.h +118 -0
  235. data/ext/sources/ggml/src/ggml-cpu/quants.c +1158 -0
  236. data/ext/{ggml/src/ggml-cpu/ggml-cpu-quants.h → sources/ggml/src/ggml-cpu/quants.h} +26 -0
  237. data/ext/sources/ggml/src/ggml-cpu/repack.cpp +1571 -0
  238. data/ext/sources/ggml/src/ggml-cpu/repack.h +98 -0
  239. data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +1184 -0
  240. data/ext/{ggml/src/ggml-cpu/ggml-cpu-traits.cpp → sources/ggml/src/ggml-cpu/traits.cpp} +1 -1
  241. data/ext/sources/ggml/src/ggml-cpu/unary-ops.cpp +186 -0
  242. data/ext/sources/ggml/src/ggml-cpu/unary-ops.h +28 -0
  243. data/ext/sources/ggml/src/ggml-cpu/vec.cpp +345 -0
  244. data/ext/sources/ggml/src/ggml-cpu/vec.h +1027 -0
  245. data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +184 -0
  246. data/ext/sources/ggml/src/ggml-cuda/acc.cu +61 -0
  247. data/ext/sources/ggml/src/ggml-cuda/acc.cuh +5 -0
  248. data/ext/sources/ggml/src/ggml-cuda/arange.cu +34 -0
  249. data/ext/sources/ggml/src/ggml-cuda/arange.cuh +5 -0
  250. data/ext/sources/ggml/src/ggml-cuda/argmax.cu +91 -0
  251. data/ext/sources/ggml/src/ggml-cuda/argmax.cuh +3 -0
  252. data/ext/sources/ggml/src/ggml-cuda/argsort.cu +104 -0
  253. data/ext/sources/ggml/src/ggml-cuda/argsort.cuh +3 -0
  254. data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +363 -0
  255. data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +9 -0
  256. data/ext/sources/ggml/src/ggml-cuda/clamp.cu +45 -0
  257. data/ext/sources/ggml/src/ggml-cuda/clamp.cuh +5 -0
  258. data/ext/sources/ggml/src/ggml-cuda/common.cuh +851 -0
  259. data/ext/sources/ggml/src/ggml-cuda/concat.cu +221 -0
  260. data/ext/sources/ggml/src/ggml-cuda/concat.cuh +5 -0
  261. data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cu +89 -0
  262. data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cuh +5 -0
  263. data/ext/sources/ggml/src/ggml-cuda/conv2d-dw.cu +161 -0
  264. data/ext/sources/ggml/src/ggml-cuda/conv2d-dw.cuh +5 -0
  265. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cu +91 -0
  266. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cuh +4 -0
  267. data/ext/sources/ggml/src/ggml-cuda/convert.cu +752 -0
  268. data/ext/sources/ggml/src/ggml-cuda/convert.cuh +31 -0
  269. data/ext/sources/ggml/src/ggml-cuda/count-equal.cu +64 -0
  270. data/ext/sources/ggml/src/ggml-cuda/count-equal.cuh +5 -0
  271. data/ext/sources/ggml/src/ggml-cuda/cp-async.cuh +57 -0
  272. data/ext/sources/ggml/src/ggml-cuda/cpy.cu +705 -0
  273. data/ext/sources/ggml/src/ggml-cuda/cpy.cuh +11 -0
  274. data/ext/sources/ggml/src/ggml-cuda/cross-entropy-loss.cu +189 -0
  275. data/ext/sources/ggml/src/ggml-cuda/cross-entropy-loss.cuh +7 -0
  276. data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +103 -0
  277. data/ext/sources/ggml/src/ggml-cuda/diagmask.cu +40 -0
  278. data/ext/sources/ggml/src/ggml-cuda/diagmask.cuh +5 -0
  279. data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +881 -0
  280. data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +1474 -0
  281. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cu +357 -0
  282. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cuh +3 -0
  283. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cu +365 -0
  284. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cuh +3 -0
  285. data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f16.cuh +482 -0
  286. data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f32.cuh +472 -0
  287. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +638 -0
  288. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cuh +3 -0
  289. data/ext/sources/ggml/src/ggml-cuda/fattn.cu +346 -0
  290. data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +3 -0
  291. data/ext/sources/ggml/src/ggml-cuda/getrows.cu +275 -0
  292. data/ext/sources/ggml/src/ggml-cuda/getrows.cuh +15 -0
  293. data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +3647 -0
  294. data/ext/sources/ggml/src/ggml-cuda/gla.cu +93 -0
  295. data/ext/sources/ggml/src/ggml-cuda/gla.cuh +3 -0
  296. data/ext/sources/ggml/src/ggml-cuda/im2col.cu +103 -0
  297. data/ext/sources/ggml/src/ggml-cuda/im2col.cuh +5 -0
  298. data/ext/sources/ggml/src/ggml-cuda/mean.cu +19 -0
  299. data/ext/sources/ggml/src/ggml-cuda/mean.cuh +3 -0
  300. data/ext/sources/ggml/src/ggml-cuda/mma.cuh +396 -0
  301. data/ext/sources/ggml/src/ggml-cuda/mmq.cu +324 -0
  302. data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +3217 -0
  303. data/ext/sources/ggml/src/ggml-cuda/mmv.cu +506 -0
  304. data/ext/sources/ggml/src/ggml-cuda/mmv.cuh +11 -0
  305. data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +595 -0
  306. data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +12 -0
  307. data/ext/sources/ggml/src/ggml-cuda/norm.cu +458 -0
  308. data/ext/sources/ggml/src/ggml-cuda/norm.cuh +11 -0
  309. data/ext/sources/ggml/src/ggml-cuda/opt-step-adamw.cu +78 -0
  310. data/ext/sources/ggml/src/ggml-cuda/opt-step-adamw.cuh +5 -0
  311. data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +68 -0
  312. data/ext/sources/ggml/src/ggml-cuda/out-prod.cuh +3 -0
  313. data/ext/sources/ggml/src/ggml-cuda/pad.cu +49 -0
  314. data/ext/sources/ggml/src/ggml-cuda/pad.cuh +5 -0
  315. data/ext/sources/ggml/src/ggml-cuda/pool2d.cu +94 -0
  316. data/ext/sources/ggml/src/ggml-cuda/pool2d.cuh +5 -0
  317. data/ext/sources/ggml/src/ggml-cuda/quantize.cu +190 -0
  318. data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +27 -0
  319. data/ext/sources/ggml/src/ggml-cuda/rope.cu +456 -0
  320. data/ext/sources/ggml/src/ggml-cuda/rope.cuh +7 -0
  321. data/ext/sources/ggml/src/ggml-cuda/scale.cu +31 -0
  322. data/ext/sources/ggml/src/ggml-cuda/scale.cuh +5 -0
  323. data/ext/sources/ggml/src/ggml-cuda/softmax.cu +283 -0
  324. data/ext/sources/ggml/src/ggml-cuda/softmax.cuh +7 -0
  325. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +148 -0
  326. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +3 -0
  327. data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +155 -0
  328. data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cuh +3 -0
  329. data/ext/sources/ggml/src/ggml-cuda/sum.cu +45 -0
  330. data/ext/sources/ggml/src/ggml-cuda/sum.cuh +5 -0
  331. data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +26 -0
  332. data/ext/sources/ggml/src/ggml-cuda/sumrows.cuh +4 -0
  333. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +5 -0
  334. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +10 -0
  335. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu +10 -0
  336. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu +10 -0
  337. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +10 -0
  338. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +5 -0
  339. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +10 -0
  340. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +10 -0
  341. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu +10 -0
  342. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu +10 -0
  343. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +5 -0
  344. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu +10 -0
  345. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +10 -0
  346. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +10 -0
  347. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu +10 -0
  348. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu +10 -0
  349. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu +10 -0
  350. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +10 -0
  351. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +10 -0
  352. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu +5 -0
  353. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu +5 -0
  354. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu +5 -0
  355. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu +5 -0
  356. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu +5 -0
  357. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu +5 -0
  358. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu +5 -0
  359. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu +5 -0
  360. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu +5 -0
  361. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu +5 -0
  362. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu +5 -0
  363. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu +5 -0
  364. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu +5 -0
  365. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu +5 -0
  366. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu +5 -0
  367. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu +5 -0
  368. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu +5 -0
  369. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu +5 -0
  370. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu +5 -0
  371. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu +5 -0
  372. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu +5 -0
  373. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu +5 -0
  374. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu +5 -0
  375. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu +5 -0
  376. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu +5 -0
  377. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu +5 -0
  378. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu +5 -0
  379. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu +5 -0
  380. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu +5 -0
  381. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu +5 -0
  382. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu +5 -0
  383. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu +5 -0
  384. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu +5 -0
  385. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu +5 -0
  386. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu +5 -0
  387. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu +5 -0
  388. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu +5 -0
  389. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu +5 -0
  390. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu +5 -0
  391. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu +5 -0
  392. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu +5 -0
  393. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu +5 -0
  394. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu +5 -0
  395. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu +5 -0
  396. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu +5 -0
  397. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu +5 -0
  398. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu +5 -0
  399. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu +5 -0
  400. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu +5 -0
  401. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu +5 -0
  402. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu +5 -0
  403. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu +5 -0
  404. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu +5 -0
  405. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu +5 -0
  406. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu +5 -0
  407. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu +5 -0
  408. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu +5 -0
  409. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu +5 -0
  410. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu +5 -0
  411. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu +5 -0
  412. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu +5 -0
  413. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu +5 -0
  414. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu +5 -0
  415. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu +5 -0
  416. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu +5 -0
  417. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu +5 -0
  418. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu +5 -0
  419. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu +5 -0
  420. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu +5 -0
  421. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu +5 -0
  422. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu +5 -0
  423. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu +5 -0
  424. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu +5 -0
  425. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu +5 -0
  426. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu +5 -0
  427. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu +5 -0
  428. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu +5 -0
  429. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu +5 -0
  430. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu +5 -0
  431. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu +5 -0
  432. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu +5 -0
  433. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu +5 -0
  434. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu +5 -0
  435. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu +5 -0
  436. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu +5 -0
  437. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu +5 -0
  438. data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +78 -0
  439. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s.cu +5 -0
  440. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_s.cu +5 -0
  441. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu +5 -0
  442. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu +5 -0
  443. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_s.cu +5 -0
  444. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu +5 -0
  445. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu +5 -0
  446. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu +5 -0
  447. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_k.cu +5 -0
  448. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q3_k.cu +5 -0
  449. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_0.cu +5 -0
  450. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_1.cu +5 -0
  451. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_k.cu +5 -0
  452. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_0.cu +5 -0
  453. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_1.cu +5 -0
  454. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_k.cu +5 -0
  455. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q6_k.cu +5 -0
  456. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q8_0.cu +5 -0
  457. data/ext/sources/ggml/src/ggml-cuda/tsembd.cu +47 -0
  458. data/ext/sources/ggml/src/ggml-cuda/tsembd.cuh +5 -0
  459. data/ext/sources/ggml/src/ggml-cuda/unary.cu +378 -0
  460. data/ext/sources/ggml/src/ggml-cuda/unary.cuh +66 -0
  461. data/ext/sources/ggml/src/ggml-cuda/upscale.cu +51 -0
  462. data/ext/sources/ggml/src/ggml-cuda/upscale.cuh +5 -0
  463. data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +1135 -0
  464. data/ext/{ggml → sources/ggml}/src/ggml-cuda/vendors/cuda.h +1 -0
  465. data/ext/{ggml → sources/ggml}/src/ggml-cuda/vendors/hip.h +57 -0
  466. data/ext/{ggml → sources/ggml}/src/ggml-cuda/vendors/musa.h +7 -1
  467. data/ext/sources/ggml/src/ggml-cuda/wkv.cu +199 -0
  468. data/ext/sources/ggml/src/ggml-cuda/wkv.cuh +7 -0
  469. data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +135 -0
  470. data/ext/{ggml → sources/ggml}/src/ggml-impl.h +147 -158
  471. data/ext/sources/ggml/src/ggml-kompute/CMakeLists.txt +166 -0
  472. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/common.comp +112 -0
  473. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_add.comp +58 -0
  474. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp +25 -0
  475. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp +52 -0
  476. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp +52 -0
  477. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp +52 -0
  478. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp +52 -0
  479. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp +30 -0
  480. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp +22 -0
  481. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp +17 -0
  482. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp +31 -0
  483. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp +31 -0
  484. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp +38 -0
  485. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp +39 -0
  486. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp +44 -0
  487. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp +52 -0
  488. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp +69 -0
  489. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp +51 -0
  490. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp +33 -0
  491. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp +35 -0
  492. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp +140 -0
  493. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp +106 -0
  494. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp +73 -0
  495. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp +52 -0
  496. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp +28 -0
  497. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp +84 -0
  498. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp +21 -0
  499. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp +53 -0
  500. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp +52 -0
  501. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp +52 -0
  502. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp +52 -0
  503. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp +52 -0
  504. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp +19 -0
  505. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp +23 -0
  506. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp +22 -0
  507. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp +72 -0
  508. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp +71 -0
  509. data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +121 -0
  510. data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +649 -0
  511. data/ext/{ggml → sources/ggml}/src/ggml-metal/ggml-metal.m +2504 -1108
  512. data/ext/{ggml → sources/ggml}/src/ggml-metal/ggml-metal.metal +2102 -1463
  513. data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +113 -0
  514. data/ext/sources/ggml/src/ggml-musa/mudnn.cu +112 -0
  515. data/ext/sources/ggml/src/ggml-musa/mudnn.cuh +12 -0
  516. data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +110 -0
  517. data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +6494 -0
  518. data/ext/sources/ggml/src/ggml-opencl/kernels/add.cl +83 -0
  519. data/ext/sources/ggml/src/ggml-opencl/kernels/argsort.cl +86 -0
  520. data/ext/sources/ggml/src/ggml-opencl/kernels/clamp.cl +20 -0
  521. data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +109 -0
  522. data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +184 -0
  523. data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +118 -0
  524. data/ext/sources/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl +58 -0
  525. data/ext/sources/ggml/src/ggml-opencl/kernels/div.cl +72 -0
  526. data/ext/sources/ggml/src/ggml-opencl/kernels/embed_kernel.py +26 -0
  527. data/ext/sources/ggml/src/ggml-opencl/kernels/gelu.cl +62 -0
  528. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle.cl +268 -0
  529. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general.cl +274 -0
  530. data/ext/sources/ggml/src/ggml-opencl/kernels/get_rows.cl +163 -0
  531. data/ext/sources/ggml/src/ggml-opencl/kernels/glu.cl +201 -0
  532. data/ext/sources/ggml/src/ggml-opencl/kernels/group_norm.cl +72 -0
  533. data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f16.cl +57 -0
  534. data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f32.cl +57 -0
  535. data/ext/sources/ggml/src/ggml-opencl/kernels/mul.cl +79 -0
  536. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mat_Ab_Bi_8x4.cl +139 -0
  537. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl +118 -0
  538. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl +118 -0
  539. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl +94 -0
  540. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl +84 -0
  541. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl +118 -0
  542. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl +283 -0
  543. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl +192 -0
  544. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl +307 -0
  545. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl +265 -0
  546. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl +272 -0
  547. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl +254 -0
  548. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k.cl +190 -0
  549. data/ext/sources/ggml/src/ggml-opencl/kernels/norm.cl +81 -0
  550. data/ext/sources/ggml/src/ggml-opencl/kernels/pad.cl +30 -0
  551. data/ext/sources/ggml/src/ggml-opencl/kernels/relu.cl +16 -0
  552. data/ext/sources/ggml/src/ggml-opencl/kernels/repeat.cl +39 -0
  553. data/ext/sources/ggml/src/ggml-opencl/kernels/rms_norm.cl +96 -0
  554. data/ext/sources/ggml/src/ggml-opencl/kernels/rope.cl +721 -0
  555. data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +16 -0
  556. data/ext/sources/ggml/src/ggml-opencl/kernels/sigmoid.cl +29 -0
  557. data/ext/sources/ggml/src/ggml-opencl/kernels/silu.cl +30 -0
  558. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +87 -0
  559. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +87 -0
  560. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f16.cl +86 -0
  561. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f32.cl +86 -0
  562. data/ext/sources/ggml/src/ggml-opencl/kernels/sub.cl +72 -0
  563. data/ext/sources/ggml/src/ggml-opencl/kernels/sum_rows.cl +39 -0
  564. data/ext/sources/ggml/src/ggml-opencl/kernels/tanh.cl +63 -0
  565. data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +84 -0
  566. data/ext/sources/ggml/src/ggml-opencl/kernels/tsembd.cl +48 -0
  567. data/ext/sources/ggml/src/ggml-opencl/kernels/upscale.cl +121 -0
  568. data/ext/{ggml → sources/ggml}/src/ggml-opt.cpp +373 -190
  569. data/ext/{ggml → sources/ggml}/src/ggml-quants.c +120 -128
  570. data/ext/sources/ggml/src/ggml-rpc/CMakeLists.txt +9 -0
  571. data/ext/{ggml → sources/ggml}/src/ggml-rpc/ggml-rpc.cpp +494 -84
  572. data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +189 -0
  573. data/ext/sources/ggml/src/ggml-sycl/backend.hpp +37 -0
  574. data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +344 -0
  575. data/ext/sources/ggml/src/ggml-sycl/binbcast.hpp +39 -0
  576. data/ext/{ggml → sources/ggml}/src/ggml-sycl/common.cpp +20 -32
  577. data/ext/sources/ggml/src/ggml-sycl/common.hpp +561 -0
  578. data/ext/{ggml → sources/ggml}/src/ggml-sycl/concat.cpp +56 -70
  579. data/ext/sources/ggml/src/ggml-sycl/concat.hpp +20 -0
  580. data/ext/{ggml → sources/ggml}/src/ggml-sycl/conv.cpp +8 -12
  581. data/ext/sources/ggml/src/ggml-sycl/conv.hpp +20 -0
  582. data/ext/sources/ggml/src/ggml-sycl/convert.cpp +575 -0
  583. data/ext/sources/ggml/src/ggml-sycl/convert.hpp +34 -0
  584. data/ext/sources/ggml/src/ggml-sycl/cpy.cpp +839 -0
  585. data/ext/sources/ggml/src/ggml-sycl/cpy.hpp +11 -0
  586. data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +823 -0
  587. data/ext/{ggml → sources/ggml}/src/ggml-sycl/dmmv.cpp +188 -67
  588. data/ext/sources/ggml/src/ggml-sycl/dmmv.hpp +27 -0
  589. data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +2987 -0
  590. data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +1120 -0
  591. data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +84 -0
  592. data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +102 -0
  593. data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +212 -0
  594. data/ext/sources/ggml/src/ggml-sycl/getrows.hpp +20 -0
  595. data/ext/{ggml → sources/ggml}/src/ggml-sycl/ggml-sycl.cpp +1197 -1295
  596. data/ext/sources/ggml/src/ggml-sycl/gla.cpp +106 -0
  597. data/ext/sources/ggml/src/ggml-sycl/gla.hpp +8 -0
  598. data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +136 -0
  599. data/ext/sources/ggml/src/ggml-sycl/im2col.hpp +21 -0
  600. data/ext/{ggml → sources/ggml}/src/ggml-sycl/mmq.cpp +60 -81
  601. data/ext/sources/ggml/src/ggml-sycl/mmq.hpp +33 -0
  602. data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +1065 -0
  603. data/ext/sources/ggml/src/ggml-sycl/mmvq.hpp +27 -0
  604. data/ext/sources/ggml/src/ggml-sycl/norm.cpp +482 -0
  605. data/ext/sources/ggml/src/ggml-sycl/norm.hpp +26 -0
  606. data/ext/{ggml → sources/ggml}/src/ggml-sycl/outprod.cpp +8 -17
  607. data/ext/sources/ggml/src/ggml-sycl/outprod.hpp +10 -0
  608. data/ext/sources/ggml/src/ggml-sycl/presets.hpp +74 -0
  609. data/ext/sources/ggml/src/ggml-sycl/quants.hpp +111 -0
  610. data/ext/sources/ggml/src/ggml-sycl/rope.cpp +472 -0
  611. data/ext/sources/ggml/src/ggml-sycl/rope.hpp +20 -0
  612. data/ext/{ggml → sources/ggml}/src/ggml-sycl/softmax.cpp +38 -28
  613. data/ext/sources/ggml/src/ggml-sycl/softmax.hpp +20 -0
  614. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +15 -0
  615. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +26 -0
  616. data/ext/{ggml → sources/ggml}/src/ggml-sycl/tsembd.cpp +6 -11
  617. data/ext/sources/ggml/src/ggml-sycl/tsembd.hpp +20 -0
  618. data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +1307 -0
  619. data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +289 -0
  620. data/ext/sources/ggml/src/ggml-sycl/wkv.hpp +10 -0
  621. data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +200 -0
  622. data/ext/sources/ggml/src/ggml-vulkan/cmake/host-toolchain.cmake.in +15 -0
  623. data/ext/{ggml → sources/ggml}/src/ggml-vulkan/ggml-vulkan.cpp +3822 -1335
  624. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +31 -0
  625. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +29 -0
  626. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add.comp +29 -0
  627. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp +51 -0
  628. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp +69 -0
  629. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/clamp.comp +17 -0
  630. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/concat.comp +41 -0
  631. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +49 -0
  632. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp +105 -0
  633. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp +98 -0
  634. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +23 -0
  635. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +51 -0
  636. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +242 -0
  637. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/cos.comp +17 -0
  638. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/count_equal.comp +31 -0
  639. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_f32.comp +20 -0
  640. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +462 -0
  641. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp +699 -0
  642. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_head.comp +13 -0
  643. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp +42 -0
  644. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp +35 -0
  645. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp +44 -0
  646. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xs.comp +43 -0
  647. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp +48 -0
  648. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp +39 -0
  649. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp +49 -0
  650. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp +32 -0
  651. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_xs.comp +34 -0
  652. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp +34 -0
  653. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp +42 -0
  654. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_0.comp +30 -0
  655. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_1.comp +32 -0
  656. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp +68 -0
  657. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_0.comp +34 -0
  658. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_1.comp +35 -0
  659. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp +70 -0
  660. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp +33 -0
  661. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q8_0.comp +31 -0
  662. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp +34 -0
  663. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/div.comp +27 -0
  664. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +337 -0
  665. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp +162 -0
  666. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +360 -0
  667. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +267 -0
  668. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +59 -0
  669. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/geglu.comp +13 -0
  670. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu.comp +25 -0
  671. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_quick.comp +23 -0
  672. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp +64 -0
  673. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_head.comp +9 -0
  674. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.comp +76 -0
  675. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +33 -0
  676. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +41 -0
  677. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp +15 -0
  678. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.comp +29 -0
  679. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp +66 -0
  680. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +100 -0
  681. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +41 -0
  682. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/leaky_relu.comp +22 -0
  683. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul.comp +27 -0
  684. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp +48 -0
  685. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +169 -0
  686. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp +118 -0
  687. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp +82 -0
  688. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp +79 -0
  689. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp +90 -0
  690. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp +87 -0
  691. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp +87 -0
  692. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp +90 -0
  693. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp +88 -0
  694. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +118 -0
  695. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp +154 -0
  696. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +130 -0
  697. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp +132 -0
  698. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +136 -0
  699. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +167 -0
  700. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp +130 -0
  701. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +868 -0
  702. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +441 -0
  703. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +442 -0
  704. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp +99 -0
  705. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/norm.comp +44 -0
  706. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_adamw.comp +42 -0
  707. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/pad.comp +28 -0
  708. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/pool2d.comp +74 -0
  709. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp +77 -0
  710. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/reglu.comp +9 -0
  711. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/relu.comp +21 -0
  712. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/repeat.comp +26 -0
  713. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp +37 -0
  714. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +61 -0
  715. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_back.comp +55 -0
  716. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp +58 -0
  717. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +60 -0
  718. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +43 -0
  719. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +43 -0
  720. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +47 -0
  721. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/scale.comp +24 -0
  722. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp +20 -0
  723. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/silu.comp +22 -0
  724. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/silu_back.comp +26 -0
  725. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sin.comp +17 -0
  726. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +173 -0
  727. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp +50 -0
  728. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/square.comp +17 -0
  729. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sub.comp +29 -0
  730. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp +37 -0
  731. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/swiglu.comp +9 -0
  732. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp +20 -0
  733. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_bfloat16_support.comp +7 -0
  734. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat2_support.comp +7 -0
  735. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat_support.comp +7 -0
  736. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_integer_dot_support.comp +7 -0
  737. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp +41 -0
  738. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +1373 -0
  739. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +36 -0
  740. data/ext/{ggml → sources/ggml}/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +203 -36
  741. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/wkv6.comp +87 -0
  742. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/wkv7.comp +91 -0
  743. data/ext/{ggml → sources/ggml}/src/ggml.c +918 -1782
  744. data/ext/sources/ggml/src/ggml.cpp +26 -0
  745. data/ext/sources/ggml/src/gguf.cpp +1351 -0
  746. data/ext/{include → sources/include}/whisper.h +70 -2
  747. data/ext/sources/src/CMakeLists.txt +145 -0
  748. data/ext/sources/src/coreml/whisper-compat.h +10 -0
  749. data/ext/sources/src/coreml/whisper-compat.m +35 -0
  750. data/ext/{src → sources/src}/coreml/whisper-decoder-impl.h +27 -15
  751. data/ext/{src → sources/src}/coreml/whisper-decoder-impl.m +36 -10
  752. data/ext/{src → sources/src}/coreml/whisper-encoder-impl.h +21 -9
  753. data/ext/{src → sources/src}/coreml/whisper-encoder-impl.m +29 -3
  754. data/ext/sources/src/coreml/whisper-encoder.mm +73 -0
  755. data/ext/sources/src/whisper-arch.h +197 -0
  756. data/ext/{src → sources/src}/whisper.cpp +1966 -386
  757. data/ext/sources/tests/CMakeLists.txt +105 -0
  758. data/ext/sources/tests/earnings21/eval.mk +58 -0
  759. data/ext/sources/tests/earnings21/eval.py +68 -0
  760. data/ext/sources/tests/earnings21/normalizers/__init__.py +2 -0
  761. data/ext/sources/tests/earnings21/normalizers/basic.py +80 -0
  762. data/ext/sources/tests/earnings21/normalizers/english.json +1741 -0
  763. data/ext/sources/tests/earnings21/normalizers/english.py +550 -0
  764. data/ext/sources/tests/earnings21/requirements.txt +6 -0
  765. data/ext/sources/tests/en-0-ref.txt +1 -0
  766. data/ext/sources/tests/en-1-ref.txt +1 -0
  767. data/ext/sources/tests/en-2-ref.txt +1 -0
  768. data/ext/sources/tests/es-0-ref.txt +1 -0
  769. data/ext/sources/tests/librispeech/eval.mk +39 -0
  770. data/ext/sources/tests/librispeech/eval.py +47 -0
  771. data/ext/sources/tests/librispeech/normalizers/__init__.py +2 -0
  772. data/ext/sources/tests/librispeech/normalizers/basic.py +80 -0
  773. data/ext/sources/tests/librispeech/normalizers/english.json +1741 -0
  774. data/ext/sources/tests/librispeech/normalizers/english.py +550 -0
  775. data/ext/sources/tests/librispeech/requirements.txt +6 -0
  776. data/ext/sources/tests/run-tests.sh +130 -0
  777. data/ext/sources/tests/test-c.c +3 -0
  778. data/ext/sources/tests/test-vad-full.cpp +54 -0
  779. data/ext/sources/tests/test-vad.cpp +83 -0
  780. data/ext/sources/tests/test-whisper.js +58 -0
  781. data/extsources.rb +39 -5
  782. data/lib/whisper/context.rb +15 -0
  783. data/lib/whisper/model/uri.rb +202 -126
  784. data/lib/whisper/segment.rb +58 -0
  785. data/sig/whisper.rbs +510 -0
  786. data/test/helper.rb +24 -0
  787. data/{tests → test}/test_callback.rb +45 -3
  788. data/{tests → test}/test_error.rb +2 -2
  789. data/{tests → test}/test_model.rb +47 -0
  790. data/test/test_package.rb +51 -0
  791. data/test/test_params.rb +297 -0
  792. data/test/test_segment.rb +146 -0
  793. data/test/test_vad.rb +19 -0
  794. data/test/test_vad_params.rb +103 -0
  795. data/{tests → test}/test_whisper.rb +106 -36
  796. data/whispercpp.gemspec +5 -5
  797. metadata +837 -134
  798. data/ext/cpu.mk +0 -9
  799. data/ext/examples/dr_wav.h +0 -8815
  800. data/ext/ggml/src/ggml-cann/aclnn_ops.h +0 -592
  801. data/ext/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +0 -4262
  802. data/ext/ggml/src/ggml-cpu/ggml-cpu-aarch64.h +0 -8
  803. data/ext/ggml/src/ggml-cpu/ggml-cpu-quants.c +0 -10835
  804. data/ext/ggml/src/ggml-cpu/ggml-cpu.c +0 -14123
  805. data/ext/ggml/src/ggml-cpu/llamafile/sgemm.cpp +0 -1884
  806. data/ext/ggml/src/ggml-cpu/llamafile/sgemm.h +0 -14
  807. data/ext/ggml/src/ggml-metal/ggml-metal-impl.h +0 -288
  808. data/ext/ggml/src/ggml-sycl/convert.cpp +0 -547
  809. data/ext/ggml/src/ggml-sycl/element_wise.cpp +0 -1030
  810. data/ext/ggml/src/ggml-sycl/im2col.cpp +0 -126
  811. data/ext/ggml/src/ggml-sycl/mmvq.cpp +0 -1015
  812. data/ext/ggml/src/ggml-sycl/norm.cpp +0 -378
  813. data/ext/ggml/src/ggml-sycl/rope.cpp +0 -276
  814. data/ext/ggml/src/ggml-sycl/wkv6.cpp +0 -141
  815. data/ext/metal-embed.mk +0 -17
  816. data/ext/metal.mk +0 -6
  817. data/ext/ruby_whisper.cpp +0 -1909
  818. data/ext/scripts/get-flags.mk +0 -38
  819. data/lib/whisper.rb +0 -2
  820. data/tests/helper.rb +0 -7
  821. data/tests/test_package.rb +0 -31
  822. data/tests/test_params.rb +0 -160
  823. data/tests/test_segment.rb +0 -83
  824. /data/ext/{ggml → sources/ggml}/include/ggml-blas.h +0 -0
  825. /data/ext/{ggml → sources/ggml}/include/ggml-cann.h +0 -0
  826. /data/ext/{ggml → sources/ggml}/include/ggml-cuda.h +0 -0
  827. /data/ext/{ggml → sources/ggml}/include/ggml-kompute.h +0 -0
  828. /data/ext/{ggml → sources/ggml}/include/ggml-opencl.h +0 -0
  829. /data/ext/{ggml → sources/ggml}/include/ggml-sycl.h +0 -0
  830. /data/ext/{ggml → sources/ggml}/src/ggml-amx/common.h +0 -0
  831. /data/ext/{ggml → sources/ggml}/src/ggml-amx/ggml-amx.cpp +0 -0
  832. /data/ext/{ggml → sources/ggml}/src/ggml-amx/mmq.cpp +0 -0
  833. /data/ext/{ggml → sources/ggml}/src/ggml-amx/mmq.h +0 -0
  834. /data/ext/{ggml → sources/ggml}/src/ggml-blas/ggml-blas.cpp +0 -0
  835. /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/ascendc_kernels.h +0 -0
  836. /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/get_row_f16.cpp +0 -0
  837. /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/get_row_f32.cpp +0 -0
  838. /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/get_row_q4_0.cpp +0 -0
  839. /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/get_row_q8_0.cpp +0 -0
  840. /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/quantize_f16_q8_0.cpp +0 -0
  841. /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/quantize_f32_q8_0.cpp +0 -0
  842. /data/ext/{ggml → sources/ggml}/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp +0 -0
  843. /data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/amx.h +0 -0
  844. /data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/common.h +0 -0
  845. /data/ext/{ggml → sources/ggml}/src/ggml-cpu/amx/mmq.h +0 -0
  846. /data/ext/{ggml/src/ggml-cpu/ggml-cpu-hbm.h → sources/ggml/src/ggml-cpu/hbm.h} +0 -0
  847. /data/ext/{ggml/src/ggml-cpu/ggml-cpu-traits.h → sources/ggml/src/ggml-cpu/traits.h} +0 -0
  848. /data/ext/{ggml → sources/ggml}/src/ggml-kompute/ggml-kompute.cpp +0 -0
  849. /data/ext/{ggml → sources/ggml}/src/ggml-quants.h +0 -0
  850. /data/ext/{ggml → sources/ggml}/src/ggml-threading.cpp +0 -0
  851. /data/ext/{ggml → sources/ggml}/src/ggml-threading.h +0 -0
  852. /data/ext/{src → sources/src}/coreml/whisper-encoder.h +0 -0
  853. /data/ext/{src → sources/src}/openvino/whisper-openvino-encoder.cpp +0 -0
  854. /data/ext/{src → sources/src}/openvino/whisper-openvino-encoder.h +0 -0
  855. /data/{tests → test}/jfk_reader/.gitignore +0 -0
  856. /data/{tests → test}/jfk_reader/extconf.rb +0 -0
  857. /data/{tests → test}/jfk_reader/jfk_reader.c +0 -0
@@ -0,0 +1,118 @@
1
+ #pragma once
2
+
3
+ #include "ggml.h"
4
+
5
+ //
6
+ // cache line
7
+ //
8
+
9
+ #if defined(__cpp_lib_hardware_interference_size)
10
+ #define CACHE_LINE_SIZE std::hardware_destructive_interference_size
11
+ #else
12
+ #if defined(__POWER9_VECTOR__)
13
+ #define CACHE_LINE_SIZE 128
14
+ #elif defined(__VXE__) || defined(__VXE2__)
15
+ #define CACHE_LINE_SIZE 256
16
+ #else
17
+ #define CACHE_LINE_SIZE 64
18
+ #endif
19
+ #endif
20
+
21
+ static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
22
+
23
+ // Work buffer size for im2col operations in CONV2D
24
+ #define GGML_IM2COL_WORK_SIZE (16 * 1024 * 1024)
25
+
26
+ #ifdef __cplusplus
27
+ extern "C" {
28
+ #endif
29
+
30
+ void ggml_compute_forward_dup(const struct ggml_compute_params * params, struct ggml_tensor * dst);
31
+ void ggml_compute_forward_add(const struct ggml_compute_params * params, struct ggml_tensor * dst);
32
+ void ggml_compute_forward_add1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
33
+ void ggml_compute_forward_acc(const struct ggml_compute_params * params, struct ggml_tensor * dst);
34
+ void ggml_compute_forward_sum(const struct ggml_compute_params * params, struct ggml_tensor * dst);
35
+ void ggml_compute_forward_sum_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
36
+ void ggml_compute_forward_mean(const struct ggml_compute_params * params, struct ggml_tensor * dst);
37
+ void ggml_compute_forward_argmax(const struct ggml_compute_params * params, struct ggml_tensor * dst);
38
+ void ggml_compute_forward_count_equal(const struct ggml_compute_params * params, struct ggml_tensor * dst);
39
+ void ggml_compute_forward_repeat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
40
+ void ggml_compute_forward_repeat_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
41
+ void ggml_compute_forward_concat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
42
+ void ggml_compute_forward_silu_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
43
+ void ggml_compute_forward_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
44
+ void ggml_compute_forward_rms_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
45
+ void ggml_compute_forward_rms_norm_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
46
+ void ggml_compute_forward_group_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
47
+ void ggml_compute_forward_l2_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst);
48
+ void ggml_compute_forward_out_prod(const struct ggml_compute_params * params, struct ggml_tensor * dst);
49
+ void ggml_compute_forward_scale(const struct ggml_compute_params * params, struct ggml_tensor * dst);
50
+ void ggml_compute_forward_set(const struct ggml_compute_params * params, struct ggml_tensor * dst);
51
+ void ggml_compute_forward_cpy(const struct ggml_compute_params * params, struct ggml_tensor * dst);
52
+ void ggml_compute_forward_cont(const struct ggml_compute_params * params, struct ggml_tensor * dst);
53
+ void ggml_compute_forward_reshape(const struct ggml_compute_params * params, struct ggml_tensor * dst);
54
+ void ggml_compute_forward_view(const struct ggml_compute_params * params, struct ggml_tensor * dst);
55
+ void ggml_compute_forward_permute(const struct ggml_compute_params * params, struct ggml_tensor * dst);
56
+ void ggml_compute_forward_transpose(const struct ggml_compute_params * params, struct ggml_tensor * dst);
57
+ void ggml_compute_forward_get_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
58
+ void ggml_compute_forward_get_rows_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
59
+ void ggml_compute_forward_set_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
60
+ void ggml_compute_forward_diag(const struct ggml_compute_params * params, struct ggml_tensor * dst);
61
+ void ggml_compute_forward_diag_mask_inf(const struct ggml_compute_params * params, struct ggml_tensor * dst);
62
+ void ggml_compute_forward_diag_mask_zero(const struct ggml_compute_params * params, struct ggml_tensor * dst);
63
+ void ggml_compute_forward_soft_max(const struct ggml_compute_params * params, struct ggml_tensor * dst);
64
+ void ggml_compute_forward_soft_max_ext_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
65
+ void ggml_compute_forward_rope(const struct ggml_compute_params * params, struct ggml_tensor * dst);
66
+ void ggml_compute_forward_rope_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
67
+ void ggml_compute_forward_clamp(const struct ggml_compute_params * params, struct ggml_tensor * dst);
68
+ void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
69
+ void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
70
+ void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
71
+ void ggml_compute_forward_conv_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
72
+ void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
73
+ void ggml_compute_forward_conv_2d_dw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
74
+ void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
75
+ void ggml_compute_forward_pool_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
76
+ void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
77
+ void ggml_compute_forward_upscale(const struct ggml_compute_params * params, struct ggml_tensor * dst);
78
+ void ggml_compute_forward_pad(const struct ggml_compute_params * params, struct ggml_tensor * dst);
79
+ void ggml_compute_forward_pad_reflect_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
80
+ void ggml_compute_forward_roll(const struct ggml_compute_params * params, struct ggml_tensor * dst);
81
+ void ggml_compute_forward_arange(const struct ggml_compute_params * params, struct ggml_tensor * dst);
82
+ void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
83
+ void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);
84
+ void ggml_compute_forward_leaky_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
85
+ void ggml_compute_forward_flash_attn_ext(
86
+ const struct ggml_compute_params * params,
87
+ const struct ggml_tensor * q,
88
+ const struct ggml_tensor * k,
89
+ const struct ggml_tensor * v,
90
+ const struct ggml_tensor * mask,
91
+ struct ggml_tensor * dst);
92
+ void ggml_compute_forward_flash_attn_back(
93
+ const struct ggml_compute_params * params,
94
+ const bool masked,
95
+ struct ggml_tensor * dst);
96
+ void ggml_compute_forward_ssm_conv(const struct ggml_compute_params * params, struct ggml_tensor * dst);
97
+ void ggml_compute_forward_ssm_scan(const struct ggml_compute_params * params, struct ggml_tensor * dst);
98
+ void ggml_compute_forward_win_part(const struct ggml_compute_params * params, struct ggml_tensor * dst);
99
+ void ggml_compute_forward_win_unpart(const struct ggml_compute_params * params, struct ggml_tensor * dst);
100
+ void ggml_compute_forward_unary(const struct ggml_compute_params * params, struct ggml_tensor * dst);
101
+ void ggml_compute_forward_glu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
102
+ void ggml_compute_forward_get_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
103
+ void ggml_compute_forward_add_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
104
+ void ggml_compute_forward_rwkv_wkv6(const struct ggml_compute_params * params, struct ggml_tensor * dst);
105
+ void ggml_compute_forward_rwkv_wkv7(const struct ggml_compute_params * params, struct ggml_tensor * dst);
106
+ void ggml_compute_forward_gla(const struct ggml_compute_params * params, struct ggml_tensor * dst);
107
+ void ggml_compute_forward_map_custom1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
108
+ void ggml_compute_forward_map_custom2(const struct ggml_compute_params * params, struct ggml_tensor * dst);
109
+ void ggml_compute_forward_map_custom3(const struct ggml_compute_params * params, struct ggml_tensor * dst);
110
+ void ggml_compute_forward_custom(const struct ggml_compute_params * params, struct ggml_tensor * dst);
111
+ void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params * params, struct ggml_tensor * dst);
112
+ void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
113
+ void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
114
+ void ggml_compute_forward_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
115
+
116
+ #ifdef __cplusplus
117
+ }
118
+ #endif