whispercpp 1.3.0 → 1.3.2

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 (787) hide show
  1. checksums.yaml +4 -4
  2. data/.gitignore +6 -0
  3. data/LICENSE +1 -1
  4. data/README.md +216 -424
  5. data/Rakefile +79 -11
  6. data/ext/.gitignore +11 -0
  7. data/ext/dependencies.rb +61 -0
  8. data/ext/extconf.rb +18 -26
  9. data/ext/options.rb +221 -0
  10. data/ext/ruby_whisper.c +159 -0
  11. data/ext/ruby_whisper.h +27 -2
  12. data/ext/ruby_whisper_context.c +641 -0
  13. data/ext/ruby_whisper_error.c +52 -0
  14. data/ext/ruby_whisper_model.c +232 -0
  15. data/ext/ruby_whisper_params.c +1301 -0
  16. data/ext/ruby_whisper_segment.c +143 -0
  17. data/ext/ruby_whisper_transcribe.cpp +87 -0
  18. data/ext/ruby_whisper_vad_params.c +288 -0
  19. data/ext/sources/.dockerignore +3 -0
  20. data/ext/sources/.github/workflows/bindings-ruby.yml +21 -0
  21. data/ext/sources/CMakeGraphVizOptions.cmake +8 -0
  22. data/ext/sources/CMakeLists.txt +251 -0
  23. data/ext/sources/bindings/javascript/CMakeLists.txt +41 -0
  24. data/ext/sources/bindings/javascript/emscripten.cpp +93 -0
  25. data/ext/sources/bindings/javascript/libwhisper.worker.js +1 -0
  26. data/ext/sources/bindings/javascript/package-tmpl.json +26 -0
  27. data/ext/sources/bindings/javascript/package.json +26 -0
  28. data/ext/sources/bindings/javascript/whisper.js +19 -0
  29. data/ext/sources/build-xcframework.sh +547 -0
  30. data/ext/sources/ci/run.sh +336 -0
  31. data/ext/sources/close-issue.yml +28 -0
  32. data/ext/sources/cmake/DefaultTargetOptions.cmake +16 -0
  33. data/ext/sources/cmake/FindFFmpeg.cmake +163 -0
  34. data/ext/sources/cmake/build-info.cmake +60 -0
  35. data/ext/sources/cmake/git-vars.cmake +22 -0
  36. data/ext/sources/cmake/whisper-config.cmake.in +65 -0
  37. data/ext/sources/cmake/whisper.pc.in +10 -0
  38. data/ext/sources/examples/CMakeLists.txt +124 -0
  39. data/ext/sources/examples/addon.node/CMakeLists.txt +31 -0
  40. data/ext/sources/examples/addon.node/__test__/whisper.spec.js +37 -0
  41. data/ext/sources/examples/addon.node/addon.cpp +438 -0
  42. data/ext/sources/examples/addon.node/index.js +54 -0
  43. data/ext/sources/examples/addon.node/package.json +16 -0
  44. data/ext/sources/examples/bench/CMakeLists.txt +8 -0
  45. data/ext/sources/examples/bench/bench.cpp +175 -0
  46. data/ext/sources/examples/bench.wasm/CMakeLists.txt +49 -0
  47. data/ext/sources/examples/bench.wasm/emscripten.cpp +87 -0
  48. data/ext/sources/examples/bench.wasm/index-tmpl.html +284 -0
  49. data/ext/sources/examples/cli/CMakeLists.txt +8 -0
  50. data/ext/sources/examples/cli/cli.cpp +1294 -0
  51. data/ext/sources/examples/coi-serviceworker.js +146 -0
  52. data/ext/sources/examples/command/CMakeLists.txt +10 -0
  53. data/ext/sources/examples/command/command.cpp +776 -0
  54. data/ext/sources/examples/command/commands.txt +9 -0
  55. data/ext/sources/examples/command.wasm/CMakeLists.txt +50 -0
  56. data/ext/sources/examples/command.wasm/emscripten.cpp +327 -0
  57. data/ext/sources/examples/command.wasm/index-tmpl.html +414 -0
  58. data/ext/sources/examples/common-ggml.cpp +238 -0
  59. data/ext/sources/examples/common-ggml.h +18 -0
  60. data/ext/sources/examples/common-sdl.cpp +227 -0
  61. data/ext/sources/examples/common-sdl.h +49 -0
  62. data/ext/sources/examples/common-whisper.cpp +168 -0
  63. data/ext/sources/examples/common-whisper.h +24 -0
  64. data/ext/sources/examples/common.cpp +675 -0
  65. data/ext/sources/examples/common.h +322 -0
  66. data/ext/sources/examples/deprecation-warning/CMakeLists.txt +6 -0
  67. data/ext/sources/examples/deprecation-warning/deprecation-warning.cpp +38 -0
  68. data/ext/sources/examples/ffmpeg-transcode.cpp +368 -0
  69. data/ext/sources/examples/generate-karaoke.sh +57 -0
  70. data/ext/sources/examples/grammar-parser.cpp +423 -0
  71. data/ext/sources/examples/grammar-parser.h +29 -0
  72. data/ext/sources/examples/helpers.js +191 -0
  73. data/ext/sources/examples/json.hpp +24596 -0
  74. data/ext/sources/examples/livestream.sh +112 -0
  75. data/ext/sources/examples/lsp/CMakeLists.txt +9 -0
  76. data/ext/sources/examples/lsp/lsp.cpp +467 -0
  77. data/ext/sources/examples/lsp/whisper.vim +362 -0
  78. data/ext/sources/examples/miniaudio.h +93468 -0
  79. data/ext/sources/examples/python/test_whisper_processor.py +7 -0
  80. data/ext/sources/examples/python/whisper_processor.py +54 -0
  81. data/ext/sources/examples/quantize/CMakeLists.txt +6 -0
  82. data/ext/sources/examples/quantize/quantize.cpp +223 -0
  83. data/ext/sources/examples/server/CMakeLists.txt +12 -0
  84. data/ext/sources/examples/server/bench.js +29 -0
  85. data/ext/sources/examples/server/httplib.h +10497 -0
  86. data/ext/sources/examples/server/server.cpp +1091 -0
  87. data/ext/sources/examples/server.py +115 -0
  88. data/ext/sources/examples/stb_vorbis.c +5584 -0
  89. data/ext/sources/examples/stream/CMakeLists.txt +10 -0
  90. data/ext/sources/examples/stream/stream.cpp +429 -0
  91. data/ext/sources/examples/stream.wasm/CMakeLists.txt +49 -0
  92. data/ext/sources/examples/stream.wasm/emscripten.cpp +216 -0
  93. data/ext/sources/examples/stream.wasm/index-tmpl.html +414 -0
  94. data/ext/sources/examples/sycl/CMakeLists.txt +9 -0
  95. data/ext/sources/examples/sycl/build.sh +22 -0
  96. data/ext/sources/examples/sycl/ls-sycl-device.cpp +11 -0
  97. data/ext/sources/examples/sycl/run-whisper.sh +17 -0
  98. data/ext/sources/examples/talk-llama/CMakeLists.txt +40 -0
  99. data/ext/sources/examples/talk-llama/eleven-labs.py +80 -0
  100. data/ext/sources/examples/talk-llama/llama-adapter.cpp +388 -0
  101. data/ext/sources/examples/talk-llama/llama-adapter.h +76 -0
  102. data/ext/sources/examples/talk-llama/llama-arch.cpp +1746 -0
  103. data/ext/sources/examples/talk-llama/llama-arch.h +437 -0
  104. data/ext/sources/examples/talk-llama/llama-batch.cpp +374 -0
  105. data/ext/sources/examples/talk-llama/llama-batch.h +89 -0
  106. data/ext/sources/examples/talk-llama/llama-chat.cpp +663 -0
  107. data/ext/sources/examples/talk-llama/llama-chat.h +58 -0
  108. data/ext/sources/examples/talk-llama/llama-context.cpp +2676 -0
  109. data/ext/sources/examples/talk-llama/llama-context.h +276 -0
  110. data/ext/sources/examples/talk-llama/llama-cparams.cpp +5 -0
  111. data/ext/sources/examples/talk-llama/llama-cparams.h +41 -0
  112. data/ext/sources/examples/talk-llama/llama-grammar.cpp +1229 -0
  113. data/ext/sources/examples/talk-llama/llama-grammar.h +173 -0
  114. data/ext/sources/examples/talk-llama/llama-graph.cpp +1618 -0
  115. data/ext/sources/examples/talk-llama/llama-graph.h +640 -0
  116. data/ext/sources/examples/talk-llama/llama-hparams.cpp +95 -0
  117. data/ext/sources/examples/talk-llama/llama-hparams.h +190 -0
  118. data/ext/sources/examples/talk-llama/llama-impl.cpp +167 -0
  119. data/ext/sources/examples/talk-llama/llama-impl.h +61 -0
  120. data/ext/sources/examples/talk-llama/llama-io.cpp +15 -0
  121. data/ext/sources/examples/talk-llama/llama-io.h +35 -0
  122. data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +2739 -0
  123. data/ext/sources/examples/talk-llama/llama-kv-cache.h +502 -0
  124. data/ext/sources/examples/talk-llama/llama-kv-cells.h +379 -0
  125. data/ext/sources/examples/talk-llama/llama-memory.cpp +1 -0
  126. data/ext/sources/examples/talk-llama/llama-memory.h +32 -0
  127. data/ext/sources/examples/talk-llama/llama-mmap.cpp +600 -0
  128. data/ext/sources/examples/talk-llama/llama-mmap.h +68 -0
  129. data/ext/sources/examples/talk-llama/llama-model-loader.cpp +1138 -0
  130. data/ext/sources/examples/talk-llama/llama-model-loader.h +169 -0
  131. data/ext/sources/examples/talk-llama/llama-model-saver.cpp +281 -0
  132. data/ext/sources/examples/talk-llama/llama-model-saver.h +37 -0
  133. data/ext/sources/examples/talk-llama/llama-model.cpp +13814 -0
  134. data/ext/sources/examples/talk-llama/llama-model.h +425 -0
  135. data/ext/sources/examples/talk-llama/llama-quant.cpp +966 -0
  136. data/ext/sources/examples/talk-llama/llama-quant.h +1 -0
  137. data/ext/sources/examples/talk-llama/llama-sampling.cpp +2575 -0
  138. data/ext/sources/examples/talk-llama/llama-sampling.h +32 -0
  139. data/ext/sources/examples/talk-llama/llama-vocab.cpp +3340 -0
  140. data/ext/sources/examples/talk-llama/llama-vocab.h +131 -0
  141. data/ext/sources/examples/talk-llama/llama.cpp +354 -0
  142. data/ext/sources/examples/talk-llama/llama.h +1377 -0
  143. data/ext/sources/examples/talk-llama/prompts/talk-alpaca.txt +23 -0
  144. data/ext/sources/examples/talk-llama/speak +40 -0
  145. data/ext/sources/examples/talk-llama/speak.bat +1 -0
  146. data/ext/sources/examples/talk-llama/speak.ps1 +14 -0
  147. data/ext/sources/examples/talk-llama/talk-llama.cpp +808 -0
  148. data/ext/sources/examples/talk-llama/unicode-data.cpp +7034 -0
  149. data/ext/sources/examples/talk-llama/unicode-data.h +20 -0
  150. data/ext/sources/examples/talk-llama/unicode.cpp +849 -0
  151. data/ext/sources/examples/talk-llama/unicode.h +66 -0
  152. data/ext/sources/examples/vad-speech-segments/CMakeLists.txt +8 -0
  153. data/ext/sources/examples/vad-speech-segments/speech.cpp +143 -0
  154. data/ext/sources/examples/wchess/CMakeLists.txt +10 -0
  155. data/ext/sources/examples/wchess/libwchess/CMakeLists.txt +19 -0
  156. data/ext/sources/examples/wchess/libwchess/Chessboard.cpp +803 -0
  157. data/ext/sources/examples/wchess/libwchess/Chessboard.h +33 -0
  158. data/ext/sources/examples/wchess/libwchess/WChess.cpp +193 -0
  159. data/ext/sources/examples/wchess/libwchess/WChess.h +63 -0
  160. data/ext/sources/examples/wchess/libwchess/test-chessboard.cpp +117 -0
  161. data/ext/sources/examples/wchess/wchess.cmd/CMakeLists.txt +8 -0
  162. data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +249 -0
  163. data/ext/sources/examples/whisper.wasm/CMakeLists.txt +50 -0
  164. data/ext/sources/examples/whisper.wasm/emscripten.cpp +118 -0
  165. data/ext/sources/examples/whisper.wasm/index-tmpl.html +658 -0
  166. data/ext/sources/ggml/CMakeLists.txt +390 -0
  167. data/ext/sources/ggml/cmake/BuildTypes.cmake +54 -0
  168. data/ext/sources/ggml/cmake/GitVars.cmake +22 -0
  169. data/ext/sources/ggml/cmake/common.cmake +26 -0
  170. data/ext/sources/ggml/cmake/ggml-config.cmake.in +152 -0
  171. data/ext/sources/ggml/include/ggml-alloc.h +76 -0
  172. data/ext/sources/ggml/include/ggml-backend.h +354 -0
  173. data/ext/sources/ggml/include/ggml-blas.h +25 -0
  174. data/ext/sources/ggml/include/ggml-cann.h +123 -0
  175. data/ext/sources/ggml/include/ggml-cpp.h +39 -0
  176. data/ext/sources/ggml/include/ggml-cpu.h +143 -0
  177. data/ext/sources/ggml/include/ggml-cuda.h +47 -0
  178. data/ext/sources/ggml/include/ggml-kompute.h +50 -0
  179. data/ext/sources/ggml/include/ggml-metal.h +66 -0
  180. data/ext/sources/ggml/include/ggml-opencl.h +26 -0
  181. data/ext/sources/ggml/include/ggml-opt.h +237 -0
  182. data/ext/sources/ggml/include/ggml-rpc.h +33 -0
  183. data/ext/sources/ggml/include/ggml-sycl.h +49 -0
  184. data/ext/sources/ggml/include/ggml-vulkan.h +29 -0
  185. data/ext/{ggml.h → sources/ggml/include/ggml.h} +621 -821
  186. data/ext/sources/ggml/include/gguf.h +202 -0
  187. data/ext/sources/ggml/src/CMakeLists.txt +346 -0
  188. data/ext/sources/ggml/src/ggml-alloc.c +1042 -0
  189. data/ext/sources/ggml/src/ggml-amx/CMakeLists.txt +107 -0
  190. data/ext/sources/ggml/src/ggml-amx/common.h +94 -0
  191. data/ext/sources/ggml/src/ggml-amx/ggml-amx.cpp +446 -0
  192. data/ext/sources/ggml/src/ggml-amx/mmq.cpp +2510 -0
  193. data/ext/sources/ggml/src/ggml-amx/mmq.h +17 -0
  194. data/ext/sources/ggml/src/ggml-backend-impl.h +255 -0
  195. data/ext/sources/ggml/src/ggml-backend-reg.cpp +586 -0
  196. data/ext/sources/ggml/src/ggml-backend.cpp +2011 -0
  197. data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +87 -0
  198. data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +517 -0
  199. data/ext/sources/ggml/src/ggml-cann/CMakeLists.txt +74 -0
  200. data/ext/sources/ggml/src/ggml-cann/Doxyfile +2579 -0
  201. data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +181 -0
  202. data/ext/sources/ggml/src/ggml-cann/acl_tensor.h +258 -0
  203. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +3193 -0
  204. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +1125 -0
  205. data/ext/sources/ggml/src/ggml-cann/common.h +420 -0
  206. data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +2606 -0
  207. data/ext/sources/ggml/src/ggml-cann/kernels/CMakeLists.txt +30 -0
  208. data/ext/sources/ggml/src/ggml-cann/kernels/ascendc_kernels.h +19 -0
  209. data/ext/sources/ggml/src/ggml-cann/kernels/dup.cpp +234 -0
  210. data/ext/sources/ggml/src/ggml-cann/kernels/get_row_f16.cpp +197 -0
  211. data/ext/sources/ggml/src/ggml-cann/kernels/get_row_f32.cpp +190 -0
  212. data/ext/sources/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp +204 -0
  213. data/ext/sources/ggml/src/ggml-cann/kernels/get_row_q8_0.cpp +191 -0
  214. data/ext/sources/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp +218 -0
  215. data/ext/sources/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp +216 -0
  216. data/ext/sources/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp +295 -0
  217. data/ext/sources/ggml/src/ggml-common.h +1857 -0
  218. data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +504 -0
  219. data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +221 -0
  220. data/ext/sources/ggml/src/ggml-cpu/amx/amx.h +8 -0
  221. data/ext/sources/ggml/src/ggml-cpu/amx/common.h +91 -0
  222. data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +2511 -0
  223. data/ext/sources/ggml/src/ggml-cpu/amx/mmq.h +10 -0
  224. data/ext/sources/ggml/src/ggml-cpu/binary-ops.cpp +158 -0
  225. data/ext/sources/ggml/src/ggml-cpu/binary-ops.h +16 -0
  226. data/ext/sources/ggml/src/ggml-cpu/cmake/FindSIMD.cmake +100 -0
  227. data/ext/sources/ggml/src/ggml-cpu/common.h +72 -0
  228. data/ext/sources/ggml/src/ggml-cpu/cpu-feats-x86.cpp +327 -0
  229. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +6431 -0
  230. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-aarch64.h +8 -0
  231. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-hbm.cpp +55 -0
  232. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-hbm.h +8 -0
  233. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +508 -0
  234. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-quants.c +13747 -0
  235. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-quants.h +63 -0
  236. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-traits.cpp +36 -0
  237. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-traits.h +38 -0
  238. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +3510 -0
  239. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +671 -0
  240. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +337 -0
  241. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.h +95 -0
  242. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +482 -0
  243. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.h +17 -0
  244. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +3544 -0
  245. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.h +14 -0
  246. data/ext/sources/ggml/src/ggml-cpu/ops.cpp +8903 -0
  247. data/ext/sources/ggml/src/ggml-cpu/ops.h +110 -0
  248. data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +892 -0
  249. data/ext/sources/ggml/src/ggml-cpu/unary-ops.cpp +186 -0
  250. data/ext/sources/ggml/src/ggml-cpu/unary-ops.h +28 -0
  251. data/ext/sources/ggml/src/ggml-cpu/vec.cpp +252 -0
  252. data/ext/sources/ggml/src/ggml-cpu/vec.h +818 -0
  253. data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +184 -0
  254. data/ext/sources/ggml/src/ggml-cuda/acc.cu +61 -0
  255. data/ext/sources/ggml/src/ggml-cuda/acc.cuh +5 -0
  256. data/ext/sources/ggml/src/ggml-cuda/arange.cu +34 -0
  257. data/ext/sources/ggml/src/ggml-cuda/arange.cuh +5 -0
  258. data/ext/sources/ggml/src/ggml-cuda/argmax.cu +91 -0
  259. data/ext/sources/ggml/src/ggml-cuda/argmax.cuh +3 -0
  260. data/ext/sources/ggml/src/ggml-cuda/argsort.cu +104 -0
  261. data/ext/sources/ggml/src/ggml-cuda/argsort.cuh +3 -0
  262. data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +363 -0
  263. data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +9 -0
  264. data/ext/sources/ggml/src/ggml-cuda/clamp.cu +45 -0
  265. data/ext/sources/ggml/src/ggml-cuda/clamp.cuh +5 -0
  266. data/ext/sources/ggml/src/ggml-cuda/common.cuh +828 -0
  267. data/ext/sources/ggml/src/ggml-cuda/concat.cu +221 -0
  268. data/ext/sources/ggml/src/ggml-cuda/concat.cuh +5 -0
  269. data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cu +89 -0
  270. data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cuh +5 -0
  271. data/ext/sources/ggml/src/ggml-cuda/convert.cu +730 -0
  272. data/ext/sources/ggml/src/ggml-cuda/convert.cuh +26 -0
  273. data/ext/sources/ggml/src/ggml-cuda/count-equal.cu +64 -0
  274. data/ext/sources/ggml/src/ggml-cuda/count-equal.cuh +5 -0
  275. data/ext/sources/ggml/src/ggml-cuda/cp-async.cuh +57 -0
  276. data/ext/sources/ggml/src/ggml-cuda/cpy.cu +705 -0
  277. data/ext/sources/ggml/src/ggml-cuda/cpy.cuh +11 -0
  278. data/ext/sources/ggml/src/ggml-cuda/cross-entropy-loss.cu +189 -0
  279. data/ext/sources/ggml/src/ggml-cuda/cross-entropy-loss.cuh +7 -0
  280. data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +103 -0
  281. data/ext/sources/ggml/src/ggml-cuda/diagmask.cu +40 -0
  282. data/ext/sources/ggml/src/ggml-cuda/diagmask.cuh +5 -0
  283. data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +881 -0
  284. data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +1471 -0
  285. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cu +357 -0
  286. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cuh +3 -0
  287. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cu +365 -0
  288. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cuh +3 -0
  289. data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f16.cuh +482 -0
  290. data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f32.cuh +472 -0
  291. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +634 -0
  292. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cuh +3 -0
  293. data/ext/sources/ggml/src/ggml-cuda/fattn.cu +346 -0
  294. data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +3 -0
  295. data/ext/sources/ggml/src/ggml-cuda/getrows.cu +275 -0
  296. data/ext/sources/ggml/src/ggml-cuda/getrows.cuh +15 -0
  297. data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +3505 -0
  298. data/ext/sources/ggml/src/ggml-cuda/gla.cu +93 -0
  299. data/ext/sources/ggml/src/ggml-cuda/gla.cuh +3 -0
  300. data/ext/sources/ggml/src/ggml-cuda/im2col.cu +103 -0
  301. data/ext/sources/ggml/src/ggml-cuda/im2col.cuh +5 -0
  302. data/ext/sources/ggml/src/ggml-cuda/mma.cuh +396 -0
  303. data/ext/sources/ggml/src/ggml-cuda/mmq.cu +324 -0
  304. data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +3217 -0
  305. data/ext/sources/ggml/src/ggml-cuda/mmv.cu +336 -0
  306. data/ext/sources/ggml/src/ggml-cuda/mmv.cuh +12 -0
  307. data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +595 -0
  308. data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +12 -0
  309. data/ext/sources/ggml/src/ggml-cuda/norm.cu +458 -0
  310. data/ext/sources/ggml/src/ggml-cuda/norm.cuh +11 -0
  311. data/ext/sources/ggml/src/ggml-cuda/opt-step-adamw.cu +78 -0
  312. data/ext/sources/ggml/src/ggml-cuda/opt-step-adamw.cuh +5 -0
  313. data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +68 -0
  314. data/ext/sources/ggml/src/ggml-cuda/out-prod.cuh +3 -0
  315. data/ext/sources/ggml/src/ggml-cuda/pad.cu +49 -0
  316. data/ext/sources/ggml/src/ggml-cuda/pad.cuh +5 -0
  317. data/ext/sources/ggml/src/ggml-cuda/pool2d.cu +94 -0
  318. data/ext/sources/ggml/src/ggml-cuda/pool2d.cuh +5 -0
  319. data/ext/sources/ggml/src/ggml-cuda/quantize.cu +190 -0
  320. data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +27 -0
  321. data/ext/sources/ggml/src/ggml-cuda/rope.cu +456 -0
  322. data/ext/sources/ggml/src/ggml-cuda/rope.cuh +7 -0
  323. data/ext/sources/ggml/src/ggml-cuda/scale.cu +31 -0
  324. data/ext/sources/ggml/src/ggml-cuda/scale.cuh +5 -0
  325. data/ext/sources/ggml/src/ggml-cuda/softmax.cu +283 -0
  326. data/ext/sources/ggml/src/ggml-cuda/softmax.cuh +7 -0
  327. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +148 -0
  328. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +3 -0
  329. data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +153 -0
  330. data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cuh +3 -0
  331. data/ext/sources/ggml/src/ggml-cuda/sum.cu +45 -0
  332. data/ext/sources/ggml/src/ggml-cuda/sum.cuh +5 -0
  333. data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +39 -0
  334. data/ext/sources/ggml/src/ggml-cuda/sumrows.cuh +5 -0
  335. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +5 -0
  336. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +10 -0
  337. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu +10 -0
  338. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu +10 -0
  339. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +10 -0
  340. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +5 -0
  341. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +10 -0
  342. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +10 -0
  343. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu +10 -0
  344. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu +10 -0
  345. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +5 -0
  346. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu +10 -0
  347. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +10 -0
  348. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +10 -0
  349. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu +10 -0
  350. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu +10 -0
  351. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu +10 -0
  352. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +10 -0
  353. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +10 -0
  354. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu +5 -0
  355. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu +5 -0
  356. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu +5 -0
  357. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu +5 -0
  358. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu +5 -0
  359. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu +5 -0
  360. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu +5 -0
  361. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu +5 -0
  362. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu +5 -0
  363. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu +5 -0
  364. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu +5 -0
  365. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu +5 -0
  366. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu +5 -0
  367. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu +5 -0
  368. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu +5 -0
  369. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu +5 -0
  370. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu +5 -0
  371. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu +5 -0
  372. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu +5 -0
  373. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu +5 -0
  374. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu +5 -0
  375. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu +5 -0
  376. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu +5 -0
  377. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu +5 -0
  378. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu +5 -0
  379. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu +5 -0
  380. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu +5 -0
  381. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu +5 -0
  382. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu +5 -0
  383. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu +5 -0
  384. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu +5 -0
  385. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu +5 -0
  386. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu +5 -0
  387. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu +5 -0
  388. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu +5 -0
  389. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu +5 -0
  390. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu +5 -0
  391. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu +5 -0
  392. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu +5 -0
  393. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu +5 -0
  394. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu +5 -0
  395. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu +5 -0
  396. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu +5 -0
  397. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu +5 -0
  398. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu +5 -0
  399. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu +5 -0
  400. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu +5 -0
  401. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu +5 -0
  402. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu +5 -0
  403. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu +5 -0
  404. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu +5 -0
  405. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu +5 -0
  406. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu +5 -0
  407. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu +5 -0
  408. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu +5 -0
  409. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu +5 -0
  410. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu +5 -0
  411. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu +5 -0
  412. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu +5 -0
  413. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu +5 -0
  414. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu +5 -0
  415. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu +5 -0
  416. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu +5 -0
  417. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu +5 -0
  418. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu +5 -0
  419. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu +5 -0
  420. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu +5 -0
  421. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu +5 -0
  422. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu +5 -0
  423. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu +5 -0
  424. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu +5 -0
  425. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu +5 -0
  426. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu +5 -0
  427. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu +5 -0
  428. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu +5 -0
  429. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu +5 -0
  430. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu +5 -0
  431. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu +5 -0
  432. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu +5 -0
  433. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu +5 -0
  434. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu +5 -0
  435. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu +5 -0
  436. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu +5 -0
  437. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu +5 -0
  438. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu +5 -0
  439. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu +5 -0
  440. data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +78 -0
  441. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s.cu +5 -0
  442. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_s.cu +5 -0
  443. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu +5 -0
  444. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu +5 -0
  445. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_s.cu +5 -0
  446. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu +5 -0
  447. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu +5 -0
  448. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu +5 -0
  449. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_k.cu +5 -0
  450. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q3_k.cu +5 -0
  451. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_0.cu +5 -0
  452. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_1.cu +5 -0
  453. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_k.cu +5 -0
  454. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_0.cu +5 -0
  455. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_1.cu +5 -0
  456. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_k.cu +5 -0
  457. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q6_k.cu +5 -0
  458. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q8_0.cu +5 -0
  459. data/ext/sources/ggml/src/ggml-cuda/tsembd.cu +47 -0
  460. data/ext/sources/ggml/src/ggml-cuda/tsembd.cuh +5 -0
  461. data/ext/sources/ggml/src/ggml-cuda/unary.cu +289 -0
  462. data/ext/sources/ggml/src/ggml-cuda/unary.cuh +59 -0
  463. data/ext/sources/ggml/src/ggml-cuda/upscale.cu +51 -0
  464. data/ext/sources/ggml/src/ggml-cuda/upscale.cuh +5 -0
  465. data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +1135 -0
  466. data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +15 -0
  467. data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +243 -0
  468. data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +140 -0
  469. data/ext/sources/ggml/src/ggml-cuda/wkv.cu +199 -0
  470. data/ext/sources/ggml/src/ggml-cuda/wkv.cuh +7 -0
  471. data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +131 -0
  472. data/ext/sources/ggml/src/ggml-impl.h +601 -0
  473. data/ext/sources/ggml/src/ggml-kompute/CMakeLists.txt +166 -0
  474. data/ext/sources/ggml/src/ggml-kompute/ggml-kompute.cpp +2251 -0
  475. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/common.comp +112 -0
  476. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_add.comp +58 -0
  477. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp +25 -0
  478. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp +52 -0
  479. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp +52 -0
  480. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp +52 -0
  481. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp +52 -0
  482. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp +30 -0
  483. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp +22 -0
  484. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp +17 -0
  485. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp +31 -0
  486. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp +31 -0
  487. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp +38 -0
  488. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp +39 -0
  489. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp +44 -0
  490. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp +52 -0
  491. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp +69 -0
  492. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp +51 -0
  493. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp +33 -0
  494. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp +35 -0
  495. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp +140 -0
  496. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp +106 -0
  497. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp +73 -0
  498. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp +52 -0
  499. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp +28 -0
  500. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp +84 -0
  501. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp +21 -0
  502. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp +53 -0
  503. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp +52 -0
  504. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp +52 -0
  505. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp +52 -0
  506. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp +52 -0
  507. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp +19 -0
  508. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp +23 -0
  509. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp +22 -0
  510. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp +72 -0
  511. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp +71 -0
  512. data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +120 -0
  513. data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +622 -0
  514. data/ext/sources/ggml/src/ggml-metal/ggml-metal.m +5998 -0
  515. data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +7089 -0
  516. data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +113 -0
  517. data/ext/sources/ggml/src/ggml-musa/mudnn.cu +112 -0
  518. data/ext/sources/ggml/src/ggml-musa/mudnn.cuh +12 -0
  519. data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +96 -0
  520. data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +5124 -0
  521. data/ext/sources/ggml/src/ggml-opencl/kernels/add.cl +83 -0
  522. data/ext/sources/ggml/src/ggml-opencl/kernels/clamp.cl +20 -0
  523. data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +184 -0
  524. data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +118 -0
  525. data/ext/sources/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl +58 -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/im2col_f16.cl +57 -0
  532. data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f32.cl +57 -0
  533. data/ext/sources/ggml/src/ggml-opencl/kernels/mul.cl +79 -0
  534. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mat_Ab_Bi_8x4.cl +139 -0
  535. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl +118 -0
  536. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl +118 -0
  537. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl +94 -0
  538. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl +84 -0
  539. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl +118 -0
  540. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl +192 -0
  541. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl +307 -0
  542. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl +265 -0
  543. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl +272 -0
  544. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl +254 -0
  545. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k.cl +190 -0
  546. data/ext/sources/ggml/src/ggml-opencl/kernels/norm.cl +81 -0
  547. data/ext/sources/ggml/src/ggml-opencl/kernels/relu.cl +16 -0
  548. data/ext/sources/ggml/src/ggml-opencl/kernels/rms_norm.cl +96 -0
  549. data/ext/sources/ggml/src/ggml-opencl/kernels/rope.cl +721 -0
  550. data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +16 -0
  551. data/ext/sources/ggml/src/ggml-opencl/kernels/silu.cl +30 -0
  552. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +87 -0
  553. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +87 -0
  554. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f16.cl +86 -0
  555. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f32.cl +86 -0
  556. data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +84 -0
  557. data/ext/sources/ggml/src/ggml-opt.cpp +1037 -0
  558. data/ext/sources/ggml/src/ggml-quants.c +5232 -0
  559. data/ext/sources/ggml/src/ggml-quants.h +100 -0
  560. data/ext/sources/ggml/src/ggml-rpc/CMakeLists.txt +9 -0
  561. data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +1813 -0
  562. data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +189 -0
  563. data/ext/sources/ggml/src/ggml-sycl/backend.hpp +37 -0
  564. data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +345 -0
  565. data/ext/sources/ggml/src/ggml-sycl/binbcast.hpp +39 -0
  566. data/ext/sources/ggml/src/ggml-sycl/common.cpp +83 -0
  567. data/ext/sources/ggml/src/ggml-sycl/common.hpp +589 -0
  568. data/ext/sources/ggml/src/ggml-sycl/concat.cpp +195 -0
  569. data/ext/sources/ggml/src/ggml-sycl/concat.hpp +20 -0
  570. data/ext/sources/ggml/src/ggml-sycl/conv.cpp +101 -0
  571. data/ext/sources/ggml/src/ggml-sycl/conv.hpp +20 -0
  572. data/ext/sources/ggml/src/ggml-sycl/convert.cpp +623 -0
  573. data/ext/sources/ggml/src/ggml-sycl/convert.hpp +34 -0
  574. data/ext/sources/ggml/src/ggml-sycl/cpy.cpp +700 -0
  575. data/ext/sources/ggml/src/ggml-sycl/cpy.hpp +11 -0
  576. data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +791 -0
  577. data/ext/sources/ggml/src/ggml-sycl/dmmv.cpp +1162 -0
  578. data/ext/sources/ggml/src/ggml-sycl/dmmv.hpp +27 -0
  579. data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +2957 -0
  580. data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +1511 -0
  581. data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +75 -0
  582. data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +99 -0
  583. data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +309 -0
  584. data/ext/sources/ggml/src/ggml-sycl/getrows.hpp +20 -0
  585. data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +4493 -0
  586. data/ext/sources/ggml/src/ggml-sycl/gla.cpp +106 -0
  587. data/ext/sources/ggml/src/ggml-sycl/gla.hpp +8 -0
  588. data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +136 -0
  589. data/ext/sources/ggml/src/ggml-sycl/im2col.hpp +21 -0
  590. data/ext/sources/ggml/src/ggml-sycl/mmq.cpp +3030 -0
  591. data/ext/sources/ggml/src/ggml-sycl/mmq.hpp +33 -0
  592. data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +1110 -0
  593. data/ext/sources/ggml/src/ggml-sycl/mmvq.hpp +27 -0
  594. data/ext/sources/ggml/src/ggml-sycl/norm.cpp +501 -0
  595. data/ext/sources/ggml/src/ggml-sycl/norm.hpp +26 -0
  596. data/ext/sources/ggml/src/ggml-sycl/outprod.cpp +47 -0
  597. data/ext/sources/ggml/src/ggml-sycl/outprod.hpp +10 -0
  598. data/ext/sources/ggml/src/ggml-sycl/presets.hpp +74 -0
  599. data/ext/sources/ggml/src/ggml-sycl/quants.hpp +83 -0
  600. data/ext/sources/ggml/src/ggml-sycl/rope.cpp +361 -0
  601. data/ext/sources/ggml/src/ggml-sycl/rope.hpp +20 -0
  602. data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +261 -0
  603. data/ext/sources/ggml/src/ggml-sycl/softmax.hpp +20 -0
  604. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +13 -0
  605. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +23 -0
  606. data/ext/sources/ggml/src/ggml-sycl/tsembd.cpp +72 -0
  607. data/ext/sources/ggml/src/ggml-sycl/tsembd.hpp +20 -0
  608. data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +1215 -0
  609. data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +293 -0
  610. data/ext/sources/ggml/src/ggml-sycl/wkv.hpp +10 -0
  611. data/ext/sources/ggml/src/ggml-threading.cpp +12 -0
  612. data/ext/sources/ggml/src/ggml-threading.h +14 -0
  613. data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +196 -0
  614. data/ext/sources/ggml/src/ggml-vulkan/cmake/host-toolchain.cmake.in +15 -0
  615. data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +10700 -0
  616. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +39 -0
  617. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +29 -0
  618. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add.comp +29 -0
  619. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp +51 -0
  620. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp +69 -0
  621. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/clamp.comp +17 -0
  622. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/concat.comp +41 -0
  623. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +49 -0
  624. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp +105 -0
  625. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +23 -0
  626. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +51 -0
  627. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +242 -0
  628. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/cos.comp +17 -0
  629. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/count_equal.comp +31 -0
  630. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_f32.comp +20 -0
  631. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +462 -0
  632. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp +699 -0
  633. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_head.comp +13 -0
  634. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp +42 -0
  635. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp +35 -0
  636. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp +44 -0
  637. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xs.comp +43 -0
  638. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp +48 -0
  639. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp +39 -0
  640. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp +49 -0
  641. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp +32 -0
  642. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_xs.comp +34 -0
  643. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp +34 -0
  644. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp +42 -0
  645. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_0.comp +30 -0
  646. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_1.comp +32 -0
  647. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp +68 -0
  648. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_0.comp +34 -0
  649. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_1.comp +35 -0
  650. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp +70 -0
  651. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp +33 -0
  652. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q8_0.comp +31 -0
  653. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp +34 -0
  654. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/div.comp +27 -0
  655. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +337 -0
  656. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp +162 -0
  657. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +360 -0
  658. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +267 -0
  659. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +59 -0
  660. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu.comp +25 -0
  661. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_quick.comp +23 -0
  662. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp +64 -0
  663. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_head.comp +9 -0
  664. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.comp +76 -0
  665. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +33 -0
  666. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +41 -0
  667. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp +66 -0
  668. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +100 -0
  669. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +41 -0
  670. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/leaky_relu.comp +22 -0
  671. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul.comp +27 -0
  672. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp +48 -0
  673. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +169 -0
  674. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp +118 -0
  675. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp +82 -0
  676. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp +79 -0
  677. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp +90 -0
  678. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp +87 -0
  679. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp +87 -0
  680. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp +90 -0
  681. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp +88 -0
  682. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +118 -0
  683. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp +154 -0
  684. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +130 -0
  685. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp +132 -0
  686. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +136 -0
  687. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +167 -0
  688. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp +130 -0
  689. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +868 -0
  690. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +441 -0
  691. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +442 -0
  692. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp +99 -0
  693. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/norm.comp +44 -0
  694. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_adamw.comp +42 -0
  695. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/pad.comp +28 -0
  696. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/pool2d.comp +74 -0
  697. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp +77 -0
  698. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/relu.comp +21 -0
  699. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/repeat.comp +26 -0
  700. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp +37 -0
  701. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +52 -0
  702. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_back.comp +55 -0
  703. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp +58 -0
  704. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +60 -0
  705. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +43 -0
  706. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +43 -0
  707. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +47 -0
  708. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/scale.comp +24 -0
  709. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp +20 -0
  710. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/silu.comp +22 -0
  711. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/silu_back.comp +26 -0
  712. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sin.comp +17 -0
  713. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +173 -0
  714. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp +50 -0
  715. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/square.comp +17 -0
  716. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sub.comp +29 -0
  717. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp +37 -0
  718. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp +20 -0
  719. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_bfloat16_support.comp +7 -0
  720. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat2_support.comp +7 -0
  721. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat_support.comp +7 -0
  722. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/test_integer_dot_support.comp +7 -0
  723. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp +41 -0
  724. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +1373 -0
  725. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +36 -0
  726. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +751 -0
  727. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/wkv6.comp +87 -0
  728. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/wkv7.comp +91 -0
  729. data/ext/sources/ggml/src/ggml.c +6550 -0
  730. data/ext/sources/ggml/src/gguf.cpp +1330 -0
  731. data/ext/{whisper.h → sources/include/whisper.h} +91 -24
  732. data/ext/sources/src/CMakeLists.txt +143 -0
  733. data/ext/sources/src/coreml/whisper-decoder-impl.h +158 -0
  734. data/ext/sources/src/coreml/whisper-decoder-impl.m +226 -0
  735. data/ext/sources/src/coreml/whisper-encoder-impl.h +154 -0
  736. data/ext/sources/src/coreml/whisper-encoder-impl.m +222 -0
  737. data/ext/sources/src/coreml/whisper-encoder.h +26 -0
  738. data/ext/sources/src/coreml/whisper-encoder.mm +73 -0
  739. data/ext/sources/src/openvino/whisper-openvino-encoder.cpp +108 -0
  740. data/ext/sources/src/openvino/whisper-openvino-encoder.h +31 -0
  741. data/ext/sources/src/whisper-arch.h +197 -0
  742. data/ext/{whisper.cpp → sources/src/whisper.cpp} +2535 -835
  743. data/ext/sources/tests/CMakeLists.txt +105 -0
  744. data/ext/sources/tests/earnings21/eval.mk +58 -0
  745. data/ext/sources/tests/earnings21/eval.py +68 -0
  746. data/ext/sources/tests/earnings21/normalizers/__init__.py +2 -0
  747. data/ext/sources/tests/earnings21/normalizers/basic.py +80 -0
  748. data/ext/sources/tests/earnings21/normalizers/english.json +1741 -0
  749. data/ext/sources/tests/earnings21/normalizers/english.py +550 -0
  750. data/ext/sources/tests/earnings21/requirements.txt +6 -0
  751. data/ext/sources/tests/en-0-ref.txt +1 -0
  752. data/ext/sources/tests/en-1-ref.txt +1 -0
  753. data/ext/sources/tests/en-2-ref.txt +1 -0
  754. data/ext/sources/tests/es-0-ref.txt +1 -0
  755. data/ext/sources/tests/librispeech/eval.mk +39 -0
  756. data/ext/sources/tests/librispeech/eval.py +47 -0
  757. data/ext/sources/tests/librispeech/normalizers/__init__.py +2 -0
  758. data/ext/sources/tests/librispeech/normalizers/basic.py +80 -0
  759. data/ext/sources/tests/librispeech/normalizers/english.json +1741 -0
  760. data/ext/sources/tests/librispeech/normalizers/english.py +550 -0
  761. data/ext/sources/tests/librispeech/requirements.txt +6 -0
  762. data/ext/sources/tests/run-tests.sh +130 -0
  763. data/ext/sources/tests/test-c.c +3 -0
  764. data/ext/sources/tests/test-vad-full.cpp +54 -0
  765. data/ext/sources/tests/test-vad.cpp +83 -0
  766. data/ext/sources/tests/test-whisper.js +58 -0
  767. data/extsources.rb +34 -0
  768. data/lib/whisper/model/uri.rb +178 -0
  769. data/sig/whisper.rbs +480 -0
  770. data/tests/helper.rb +35 -0
  771. data/tests/jfk_reader/.gitignore +5 -0
  772. data/tests/jfk_reader/extconf.rb +3 -0
  773. data/tests/jfk_reader/jfk_reader.c +68 -0
  774. data/tests/test_callback.rb +202 -0
  775. data/tests/test_error.rb +20 -0
  776. data/tests/test_model.rb +109 -0
  777. data/tests/test_package.rb +46 -0
  778. data/tests/test_params.rb +297 -0
  779. data/tests/test_segment.rb +74 -0
  780. data/tests/test_vad.rb +19 -0
  781. data/tests/test_vad_params.rb +103 -0
  782. data/tests/test_whisper.rb +212 -124
  783. data/whispercpp.gemspec +37 -0
  784. metadata +794 -13
  785. data/ext/dr_wav.h +0 -6434
  786. data/ext/ggml.c +0 -21755
  787. data/ext/ruby_whisper.cpp +0 -426
@@ -0,0 +1,586 @@
1
+ #include "ggml-backend-impl.h"
2
+ #include "ggml-backend.h"
3
+ #include "ggml-impl.h"
4
+ #include <algorithm>
5
+ #include <cstring>
6
+ #include <filesystem>
7
+ #include <memory>
8
+ #include <string>
9
+ #include <type_traits>
10
+ #include <vector>
11
+ #include <cctype>
12
+
13
+ #ifdef _WIN32
14
+ # define WIN32_LEAN_AND_MEAN
15
+ # ifndef NOMINMAX
16
+ # define NOMINMAX
17
+ # endif
18
+ # include <windows.h>
19
+ #elif defined(__APPLE__)
20
+ # include <mach-o/dyld.h>
21
+ # include <dlfcn.h>
22
+ #else
23
+ # include <dlfcn.h>
24
+ # include <unistd.h>
25
+ #endif
26
+
27
+ // Backend registry
28
+ #ifdef GGML_USE_CPU
29
+ #include "ggml-cpu.h"
30
+ #endif
31
+
32
+ #ifdef GGML_USE_CUDA
33
+ #include "ggml-cuda.h"
34
+ #endif
35
+
36
+ #ifdef GGML_USE_METAL
37
+ #include "ggml-metal.h"
38
+ #endif
39
+
40
+ #ifdef GGML_USE_SYCL
41
+ #include "ggml-sycl.h"
42
+ #endif
43
+
44
+ #ifdef GGML_USE_VULKAN
45
+ #include "ggml-vulkan.h"
46
+ #endif
47
+
48
+ #ifdef GGML_USE_OPENCL
49
+ #include "ggml-opencl.h"
50
+ #endif
51
+
52
+ #ifdef GGML_USE_BLAS
53
+ #include "ggml-blas.h"
54
+ #endif
55
+
56
+ #ifdef GGML_USE_RPC
57
+ #include "ggml-rpc.h"
58
+ #endif
59
+
60
+ #ifdef GGML_USE_CANN
61
+ #include "ggml-cann.h"
62
+ #endif
63
+
64
+ #ifdef GGML_USE_KOMPUTE
65
+ #include "ggml-kompute.h"
66
+ #endif
67
+
68
+ // disable C++17 deprecation warning for std::codecvt_utf8
69
+ #if defined(__clang__)
70
+ # pragma clang diagnostic push
71
+ # pragma clang diagnostic ignored "-Wdeprecated-declarations"
72
+ #endif
73
+
74
+ namespace fs = std::filesystem;
75
+
76
+ static std::string path_str(const fs::path & path) {
77
+ std::string u8path;
78
+ try {
79
+ #if defined(__cpp_lib_char8_t)
80
+ // C++20 and later: u8string() returns std::u8string
81
+ std::u8string u8str = path.u8string();
82
+ u8path = std::string(reinterpret_cast<const char*>(u8str.c_str()));
83
+ #else
84
+ // C++17: u8string() returns std::string
85
+ u8path = path.u8string();
86
+ #endif
87
+ } catch (...) {
88
+ }
89
+ return u8path;
90
+ }
91
+
92
+ #if defined(__clang__)
93
+ # pragma clang diagnostic pop
94
+ #endif
95
+
96
+ #ifdef _WIN32
97
+
98
+ using dl_handle = std::remove_pointer_t<HMODULE>;
99
+
100
+ struct dl_handle_deleter {
101
+ void operator()(HMODULE handle) {
102
+ FreeLibrary(handle);
103
+ }
104
+ };
105
+
106
+ static dl_handle * dl_load_library(const fs::path & path) {
107
+ // suppress error dialogs for missing DLLs
108
+ DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
109
+ SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
110
+
111
+ HMODULE handle = LoadLibraryW(path.wstring().c_str());
112
+
113
+ SetErrorMode(old_mode);
114
+
115
+ return handle;
116
+ }
117
+
118
+ static void * dl_get_sym(dl_handle * handle, const char * name) {
119
+ DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
120
+ SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
121
+
122
+ void * p = (void *) GetProcAddress(handle, name);
123
+
124
+ SetErrorMode(old_mode);
125
+
126
+ return p;
127
+ }
128
+
129
+ #else
130
+
131
+ using dl_handle = void;
132
+
133
+ struct dl_handle_deleter {
134
+ void operator()(void * handle) {
135
+ dlclose(handle);
136
+ }
137
+ };
138
+
139
+ static void * dl_load_library(const fs::path & path) {
140
+ dl_handle * handle = dlopen(path.string().c_str(), RTLD_NOW | RTLD_LOCAL);
141
+
142
+ return handle;
143
+ }
144
+
145
+ static void * dl_get_sym(dl_handle * handle, const char * name) {
146
+ return dlsym(handle, name);
147
+ }
148
+
149
+ #endif
150
+
151
+ using dl_handle_ptr = std::unique_ptr<dl_handle, dl_handle_deleter>;
152
+
153
+ struct ggml_backend_reg_entry {
154
+ ggml_backend_reg_t reg;
155
+ dl_handle_ptr handle;
156
+ };
157
+
158
+ struct ggml_backend_registry {
159
+ std::vector<ggml_backend_reg_entry> backends;
160
+ std::vector<ggml_backend_dev_t> devices;
161
+
162
+ ggml_backend_registry() {
163
+ #ifdef GGML_USE_CUDA
164
+ register_backend(ggml_backend_cuda_reg());
165
+ #endif
166
+ #ifdef GGML_USE_METAL
167
+ register_backend(ggml_backend_metal_reg());
168
+ #endif
169
+ #ifdef GGML_USE_SYCL
170
+ register_backend(ggml_backend_sycl_reg());
171
+ #endif
172
+ #ifdef GGML_USE_VULKAN
173
+ register_backend(ggml_backend_vk_reg());
174
+ #endif
175
+ #ifdef GGML_USE_OPENCL
176
+ register_backend(ggml_backend_opencl_reg());
177
+ #endif
178
+ #ifdef GGML_USE_CANN
179
+ register_backend(ggml_backend_cann_reg());
180
+ #endif
181
+ #ifdef GGML_USE_BLAS
182
+ register_backend(ggml_backend_blas_reg());
183
+ #endif
184
+ #ifdef GGML_USE_RPC
185
+ register_backend(ggml_backend_rpc_reg());
186
+ #endif
187
+ #ifdef GGML_USE_KOMPUTE
188
+ register_backend(ggml_backend_kompute_reg());
189
+ #endif
190
+ #ifdef GGML_USE_CPU
191
+ register_backend(ggml_backend_cpu_reg());
192
+ #endif
193
+ }
194
+
195
+ ~ggml_backend_registry() {
196
+ // FIXME: backends cannot be safely unloaded without a function to destroy all the backend resources,
197
+ // since backend threads may still be running and accessing resources from the dynamic library
198
+ for (auto & entry : backends) {
199
+ if (entry.handle) {
200
+ entry.handle.release(); // NOLINT
201
+ }
202
+ }
203
+ }
204
+
205
+ void register_backend(ggml_backend_reg_t reg, dl_handle_ptr handle = nullptr) {
206
+ if (!reg) {
207
+ return;
208
+ }
209
+
210
+ #ifndef NDEBUG
211
+ GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
212
+ __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
213
+ #endif
214
+ backends.push_back({ reg, std::move(handle) });
215
+ for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
216
+ register_device(ggml_backend_reg_dev_get(reg, i));
217
+ }
218
+ }
219
+
220
+ void register_device(ggml_backend_dev_t device) {
221
+ #ifndef NDEBUG
222
+ GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
223
+ #endif
224
+ devices.push_back(device);
225
+ }
226
+
227
+ ggml_backend_reg_t load_backend(const fs::path & path, bool silent) {
228
+ dl_handle_ptr handle { dl_load_library(path) };
229
+ if (!handle) {
230
+ if (!silent) {
231
+ GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path_str(path).c_str());
232
+ }
233
+ return nullptr;
234
+ }
235
+
236
+ auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
237
+ if (score_fn && score_fn() == 0) {
238
+ if (!silent) {
239
+ GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, path_str(path).c_str());
240
+ }
241
+ return nullptr;
242
+ }
243
+
244
+ auto backend_init_fn = (ggml_backend_init_t) dl_get_sym(handle.get(), "ggml_backend_init");
245
+ if (!backend_init_fn) {
246
+ if (!silent) {
247
+ GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, path_str(path).c_str());
248
+ }
249
+ return nullptr;
250
+ }
251
+
252
+ ggml_backend_reg_t reg = backend_init_fn();
253
+ if (!reg || reg->api_version != GGML_BACKEND_API_VERSION) {
254
+ if (!silent) {
255
+ if (!reg) {
256
+ GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n",
257
+ __func__, path_str(path).c_str());
258
+ } else {
259
+ GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n",
260
+ __func__, path_str(path).c_str(), reg->api_version, GGML_BACKEND_API_VERSION);
261
+ }
262
+ }
263
+ return nullptr;
264
+ }
265
+
266
+ GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path_str(path).c_str());
267
+
268
+ register_backend(reg, std::move(handle));
269
+
270
+ return reg;
271
+ }
272
+
273
+ void unload_backend(ggml_backend_reg_t reg, bool silent) {
274
+ auto it = std::find_if(backends.begin(), backends.end(),
275
+ [reg](const ggml_backend_reg_entry & entry) { return entry.reg == reg; });
276
+
277
+ if (it == backends.end()) {
278
+ if (!silent) {
279
+ GGML_LOG_ERROR("%s: backend not found\n", __func__);
280
+ }
281
+ return;
282
+ }
283
+
284
+ if (!silent) {
285
+ GGML_LOG_DEBUG("%s: unloading %s backend\n", __func__, ggml_backend_reg_name(reg));
286
+ }
287
+
288
+ // remove devices
289
+ devices.erase(
290
+ std::remove_if(devices.begin(), devices.end(),
291
+ [reg](ggml_backend_dev_t dev) { return ggml_backend_dev_backend_reg(dev) == reg; }),
292
+ devices.end());
293
+
294
+ // remove backend
295
+ backends.erase(it);
296
+ }
297
+ };
298
+
299
+ static ggml_backend_registry & get_reg() {
300
+ static ggml_backend_registry reg;
301
+ return reg;
302
+ }
303
+
304
+ // Internal API
305
+ void ggml_backend_register(ggml_backend_reg_t reg) {
306
+ get_reg().register_backend(reg);
307
+ }
308
+
309
+ void ggml_backend_device_register(ggml_backend_dev_t device) {
310
+ get_reg().register_device(device);
311
+ }
312
+
313
+ // Backend (reg) enumeration
314
+ static bool striequals(const char * a, const char * b) {
315
+ for (; *a && *b; a++, b++) {
316
+ if (std::tolower(*a) != std::tolower(*b)) {
317
+ return false;
318
+ }
319
+ }
320
+ return *a == *b;
321
+ }
322
+
323
+ size_t ggml_backend_reg_count() {
324
+ return get_reg().backends.size();
325
+ }
326
+
327
+ ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
328
+ GGML_ASSERT(index < ggml_backend_reg_count());
329
+ return get_reg().backends[index].reg;
330
+ }
331
+
332
+ ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
333
+ for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
334
+ ggml_backend_reg_t reg = ggml_backend_reg_get(i);
335
+ if (striequals(ggml_backend_reg_name(reg), name)) {
336
+ return reg;
337
+ }
338
+ }
339
+ return nullptr;
340
+ }
341
+
342
+ // Device enumeration
343
+ size_t ggml_backend_dev_count() {
344
+ return get_reg().devices.size();
345
+ }
346
+
347
+ ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
348
+ GGML_ASSERT(index < ggml_backend_dev_count());
349
+ return get_reg().devices[index];
350
+ }
351
+
352
+ ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
353
+ for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
354
+ ggml_backend_dev_t dev = ggml_backend_dev_get(i);
355
+ if (striequals(ggml_backend_dev_name(dev), name)) {
356
+ return dev;
357
+ }
358
+ }
359
+ return nullptr;
360
+ }
361
+
362
+ ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
363
+ for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
364
+ ggml_backend_dev_t dev = ggml_backend_dev_get(i);
365
+ if (ggml_backend_dev_type(dev) == type) {
366
+ return dev;
367
+ }
368
+ }
369
+ return nullptr;
370
+ }
371
+
372
+ // Convenience functions
373
+ ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
374
+ ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
375
+ if (!dev) {
376
+ return nullptr;
377
+ }
378
+ return ggml_backend_dev_init(dev, params);
379
+ }
380
+
381
+ ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
382
+ ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
383
+ if (!dev) {
384
+ return nullptr;
385
+ }
386
+ return ggml_backend_dev_init(dev, params);
387
+ }
388
+
389
+ ggml_backend_t ggml_backend_init_best(void) {
390
+ ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
391
+ if (!dev) {
392
+ dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
393
+ }
394
+ if (!dev) {
395
+ return nullptr;
396
+ }
397
+ return ggml_backend_dev_init(dev, nullptr);
398
+ }
399
+
400
+ // Dynamic loading
401
+ ggml_backend_reg_t ggml_backend_load(const char * path) {
402
+ return get_reg().load_backend(path, false);
403
+ }
404
+
405
+ void ggml_backend_unload(ggml_backend_reg_t reg) {
406
+ get_reg().unload_backend(reg, true);
407
+ }
408
+
409
+ static fs::path get_executable_path() {
410
+ #if defined(__APPLE__)
411
+ // get executable path
412
+ std::vector<char> path;
413
+ uint32_t size;
414
+ while (true) {
415
+ size = path.size();
416
+ if (_NSGetExecutablePath(path.data(), &size) == 0) {
417
+ break;
418
+ }
419
+ path.resize(size);
420
+ }
421
+ std::string base_path(path.data(), size);
422
+ // remove executable name
423
+ auto last_slash = base_path.find_last_of('/');
424
+ if (last_slash != std::string::npos) {
425
+ base_path = base_path.substr(0, last_slash);
426
+ }
427
+ return base_path + "/";
428
+ #elif defined(__linux__) || defined(__FreeBSD__)
429
+ std::string base_path = ".";
430
+ std::vector<char> path(1024);
431
+ while (true) {
432
+ // get executable path
433
+ # if defined(__linux__)
434
+ ssize_t len = readlink("/proc/self/exe", path.data(), path.size());
435
+ # elif defined(__FreeBSD__)
436
+ ssize_t len = readlink("/proc/curproc/file", path.data(), path.size());
437
+ # endif
438
+ if (len == -1) {
439
+ break;
440
+ }
441
+ if (len < (ssize_t) path.size()) {
442
+ base_path = std::string(path.data(), len);
443
+ // remove executable name
444
+ auto last_slash = base_path.find_last_of('/');
445
+ if (last_slash != std::string::npos) {
446
+ base_path = base_path.substr(0, last_slash);
447
+ }
448
+ break;
449
+ }
450
+ path.resize(path.size() * 2);
451
+ }
452
+
453
+ return base_path + "/";
454
+ #elif defined(_WIN32)
455
+ std::vector<wchar_t> path(MAX_PATH);
456
+ DWORD len = GetModuleFileNameW(NULL, path.data(), path.size());
457
+ if (len == 0) {
458
+ return {};
459
+ }
460
+ std::wstring base_path(path.data(), len);
461
+ // remove executable name
462
+ auto last_slash = base_path.find_last_of('\\');
463
+ if (last_slash != std::string::npos) {
464
+ base_path = base_path.substr(0, last_slash);
465
+ }
466
+ return base_path + L"\\";
467
+ #else
468
+ return {};
469
+ #endif
470
+ }
471
+
472
+ static fs::path backend_filename_prefix() {
473
+ #ifdef _WIN32
474
+ return fs::u8path("ggml-");
475
+ #else
476
+ return fs::u8path("libggml-");
477
+ #endif
478
+ }
479
+
480
+ static fs::path backend_filename_extension() {
481
+ #ifdef _WIN32
482
+ return fs::u8path(".dll");
483
+ #else
484
+ return fs::u8path(".so");
485
+ #endif
486
+ }
487
+
488
+ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, const char * user_search_path) {
489
+ // enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths
490
+ const fs::path name_path = fs::u8path(name);
491
+ const fs::path file_prefix = backend_filename_prefix().native() + name_path.native() + fs::u8path("-").native();
492
+ const fs::path file_extension = backend_filename_extension();
493
+
494
+ std::vector<fs::path> search_paths;
495
+ if (user_search_path == nullptr) {
496
+ // default search paths: executable directory, current directory
497
+ search_paths.push_back(get_executable_path());
498
+ search_paths.push_back(fs::current_path());
499
+ } else {
500
+ search_paths.push_back(fs::u8path(user_search_path));
501
+ }
502
+
503
+ int best_score = 0;
504
+ fs::path best_path;
505
+
506
+ for (const auto & search_path : search_paths) {
507
+ if (!fs::exists(search_path)) {
508
+ GGML_LOG_DEBUG("%s: search path %s does not exist\n", __func__, path_str(search_path).c_str());
509
+ continue;
510
+ }
511
+ fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
512
+ for (const auto & entry : dir_it) {
513
+ if (entry.is_regular_file()) {
514
+ auto filename = entry.path().filename();
515
+ auto ext = entry.path().extension();
516
+ if (filename.native().find(file_prefix) == 0 && ext == file_extension) {
517
+ dl_handle_ptr handle { dl_load_library(entry) };
518
+ if (!handle && !silent) {
519
+ GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path_str(entry.path()).c_str());
520
+ }
521
+ if (handle) {
522
+ auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
523
+ if (score_fn) {
524
+ int s = score_fn();
525
+ #ifndef NDEBUG
526
+ GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, path_str(entry.path()).c_str(), s);
527
+ #endif
528
+ if (s > best_score) {
529
+ best_score = s;
530
+ best_path = entry.path();
531
+ }
532
+ } else {
533
+ if (!silent) {
534
+ GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, path_str(entry.path()).c_str());
535
+ }
536
+ }
537
+ }
538
+ }
539
+ }
540
+ }
541
+ }
542
+
543
+ if (best_score == 0) {
544
+ // try to load the base backend
545
+ for (const auto & search_path : search_paths) {
546
+ fs::path filename = backend_filename_prefix().native() + name_path.native() + backend_filename_extension().native();
547
+ fs::path path = search_path / filename;
548
+ if (fs::exists(path)) {
549
+ return get_reg().load_backend(path, silent);
550
+ }
551
+ }
552
+ return nullptr;
553
+ }
554
+
555
+ return get_reg().load_backend(best_path, silent);
556
+ }
557
+
558
+ void ggml_backend_load_all() {
559
+ ggml_backend_load_all_from_path(nullptr);
560
+ }
561
+
562
+ void ggml_backend_load_all_from_path(const char * dir_path) {
563
+ #ifdef NDEBUG
564
+ bool silent = true;
565
+ #else
566
+ bool silent = false;
567
+ #endif
568
+
569
+ ggml_backend_load_best("blas", silent, dir_path);
570
+ ggml_backend_load_best("cann", silent, dir_path);
571
+ ggml_backend_load_best("cuda", silent, dir_path);
572
+ ggml_backend_load_best("hip", silent, dir_path);
573
+ ggml_backend_load_best("kompute", silent, dir_path);
574
+ ggml_backend_load_best("metal", silent, dir_path);
575
+ ggml_backend_load_best("rpc", silent, dir_path);
576
+ ggml_backend_load_best("sycl", silent, dir_path);
577
+ ggml_backend_load_best("vulkan", silent, dir_path);
578
+ ggml_backend_load_best("opencl", silent, dir_path);
579
+ ggml_backend_load_best("musa", silent, dir_path);
580
+ ggml_backend_load_best("cpu", silent, dir_path);
581
+ // check the environment variable GGML_BACKEND_PATH to load an out-of-tree backend
582
+ const char * backend_path = std::getenv("GGML_BACKEND_PATH");
583
+ if (backend_path) {
584
+ ggml_backend_load(backend_path);
585
+ }
586
+ }