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 +1,504 @@
1
+ function(ggml_add_cpu_backend_variant_impl tag_name)
2
+ if (tag_name)
3
+ set(GGML_CPU_NAME ggml-cpu-${tag_name})
4
+ else()
5
+ set(GGML_CPU_NAME ggml-cpu)
6
+ endif()
7
+
8
+ ggml_add_backend_library(${GGML_CPU_NAME})
9
+
10
+ list (APPEND GGML_CPU_SOURCES
11
+ ggml-cpu/ggml-cpu.c
12
+ ggml-cpu/ggml-cpu.cpp
13
+ ggml-cpu/ggml-cpu-aarch64.cpp
14
+ ggml-cpu/ggml-cpu-aarch64.h
15
+ ggml-cpu/ggml-cpu-hbm.cpp
16
+ ggml-cpu/ggml-cpu-hbm.h
17
+ ggml-cpu/ggml-cpu-quants.c
18
+ ggml-cpu/ggml-cpu-quants.h
19
+ ggml-cpu/ggml-cpu-traits.cpp
20
+ ggml-cpu/ggml-cpu-traits.h
21
+ ggml-cpu/amx/amx.cpp
22
+ ggml-cpu/amx/amx.h
23
+ ggml-cpu/amx/mmq.cpp
24
+ ggml-cpu/amx/mmq.h
25
+ ggml-cpu/ggml-cpu-impl.h
26
+ ggml-cpu/common.h
27
+ ggml-cpu/binary-ops.h
28
+ ggml-cpu/binary-ops.cpp
29
+ ggml-cpu/unary-ops.h
30
+ ggml-cpu/unary-ops.cpp
31
+ ggml-cpu/simd-mappings.h
32
+ ggml-cpu/vec.h
33
+ ggml-cpu/vec.cpp
34
+ ggml-cpu/ops.h
35
+ ggml-cpu/ops.cpp
36
+ )
37
+
38
+ target_compile_features(${GGML_CPU_NAME} PRIVATE c_std_11 cxx_std_17)
39
+ target_include_directories(${GGML_CPU_NAME} PRIVATE . ggml-cpu)
40
+
41
+ if (APPLE AND GGML_ACCELERATE)
42
+ find_library(ACCELERATE_FRAMEWORK Accelerate)
43
+ if (ACCELERATE_FRAMEWORK)
44
+ message(STATUS "Accelerate framework found")
45
+
46
+ target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_ACCELERATE)
47
+ target_compile_definitions(${GGML_CPU_NAME} PRIVATE ACCELERATE_NEW_LAPACK)
48
+ target_compile_definitions(${GGML_CPU_NAME} PRIVATE ACCELERATE_LAPACK_ILP64)
49
+
50
+ target_link_libraries(${GGML_CPU_NAME} PRIVATE ${ACCELERATE_FRAMEWORK})
51
+ else()
52
+ message(WARNING "Accelerate framework not found")
53
+ endif()
54
+ endif()
55
+
56
+ if (GGML_OPENMP)
57
+ find_package(OpenMP)
58
+ if (OpenMP_FOUND)
59
+ target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_OPENMP)
60
+
61
+ target_link_libraries(${GGML_CPU_NAME} PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
62
+ else()
63
+ message(WARNING "OpenMP not found")
64
+ endif()
65
+ endif()
66
+
67
+ if (GGML_LLAMAFILE)
68
+ target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_LLAMAFILE)
69
+
70
+ list(APPEND GGML_CPU_SOURCES
71
+ ggml-cpu/llamafile/sgemm.cpp
72
+ ggml-cpu/llamafile/sgemm.h)
73
+ endif()
74
+
75
+ if (GGML_CPU_HBM)
76
+ find_library(memkind memkind REQUIRED)
77
+
78
+ message(STATUS "Using memkind for CPU HBM")
79
+
80
+ target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_HBM)
81
+
82
+ target_link_libraries(${GGML_CPU_NAME} PUBLIC memkind)
83
+ endif()
84
+
85
+ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR
86
+ CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
87
+ (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
88
+ CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
89
+
90
+ message(STATUS "ARM detected")
91
+
92
+ if (MSVC AND NOT CMAKE_C_COMPILER_ID STREQUAL "Clang")
93
+ message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
94
+ else()
95
+ check_cxx_compiler_flag(-mfp16-format=ieee GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E)
96
+ if (NOT "${GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
97
+ list(APPEND ARCH_FLAGS -mfp16-format=ieee)
98
+ endif()
99
+
100
+ if (GGML_NATIVE)
101
+ # -mcpu=native does not always enable all the features in some compilers,
102
+ # so we check for them manually and enable them if available
103
+
104
+ execute_process(
105
+ COMMAND ${CMAKE_C_COMPILER} -mcpu=native -E -v -
106
+ INPUT_FILE "/dev/null"
107
+ OUTPUT_QUIET
108
+ ERROR_VARIABLE ARM_MCPU
109
+ RESULT_VARIABLE ARM_MCPU_RESULT
110
+ )
111
+ if (NOT ARM_MCPU_RESULT)
112
+ string(REGEX MATCH "-mcpu=[^ ']+" ARM_MCPU_FLAG "${ARM_MCPU}")
113
+ endif()
114
+ if ("${ARM_MCPU_FLAG}" STREQUAL "")
115
+ set(ARM_MCPU_FLAG -mcpu=native)
116
+ message(STATUS "ARM -mcpu not found, -mcpu=native will be used")
117
+ endif()
118
+
119
+ include(CheckCXXSourceRuns)
120
+
121
+ function(check_arm_feature tag code)
122
+ set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
123
+ set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+${tag}")
124
+ check_cxx_source_runs("${code}" GGML_MACHINE_SUPPORTS_${tag})
125
+ if (GGML_MACHINE_SUPPORTS_${tag})
126
+ set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+${tag}" PARENT_SCOPE)
127
+ else()
128
+ set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+no${tag}")
129
+ check_cxx_source_compiles("int main() { return 0; }" GGML_MACHINE_SUPPORTS_no${tag})
130
+ if (GGML_MACHINE_SUPPORTS_no${tag})
131
+ set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE)
132
+ endif()
133
+ endif()
134
+ set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
135
+ endfunction()
136
+
137
+ check_arm_feature(dotprod "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }")
138
+ check_arm_feature(i8mm "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }")
139
+ check_arm_feature(sve "#include <arm_sve.h>\nint main() { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }")
140
+ check_arm_feature(sme "#include <arm_sme.h>\n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }")
141
+
142
+ list(APPEND ARCH_FLAGS "${ARM_MCPU_FLAG}${ARM_MCPU_FLAG_FIX}")
143
+ else()
144
+ if (GGML_CPU_ARM_ARCH)
145
+ list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
146
+ endif()
147
+ endif()
148
+
149
+ # show enabled features
150
+ if (CMAKE_HOST_SYSTEM_NAME STREQUAL "Windows")
151
+ set(FEAT_INPUT_FILE "NUL")
152
+ else()
153
+ set(FEAT_INPUT_FILE "/dev/null")
154
+ endif()
155
+
156
+ execute_process(
157
+ COMMAND ${CMAKE_C_COMPILER} ${ARCH_FLAGS} -dM -E -
158
+ INPUT_FILE ${FEAT_INPUT_FILE}
159
+ OUTPUT_VARIABLE ARM_FEATURE
160
+ RESULT_VARIABLE ARM_FEATURE_RESULT
161
+ )
162
+ if (ARM_FEATURE_RESULT)
163
+ message(WARNING "Failed to get ARM features")
164
+ else()
165
+ foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
166
+ string(FIND "${ARM_FEATURE}" "__ARM_FEATURE_${feature} 1" feature_pos)
167
+ if (NOT ${feature_pos} EQUAL -1)
168
+ message(STATUS "ARM feature ${feature} enabled")
169
+ endif()
170
+ endforeach()
171
+ endif()
172
+ endif()
173
+ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
174
+ (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
175
+ CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
176
+
177
+ message(STATUS "x86 detected")
178
+
179
+ if (MSVC)
180
+ # instruction set detection for MSVC only
181
+ if (GGML_NATIVE)
182
+ include(ggml-cpu/cmake/FindSIMD.cmake)
183
+ endif ()
184
+ if (GGML_AVX512)
185
+ list(APPEND ARCH_FLAGS /arch:AVX512)
186
+ # /arch:AVX512 includes: __AVX512F__, __AVX512CD__, __AVX512BW__, __AVX512DQ__, and __AVX512VL__
187
+ # MSVC has no compile-time flags enabling specific
188
+ # AVX512 extensions, neither it defines the
189
+ # macros corresponding to the extensions.
190
+ # Do it manually.
191
+ list(APPEND ARCH_DEFINITIONS GGML_AVX512)
192
+ if (GGML_AVX512_VBMI)
193
+ list(APPEND ARCH_DEFINITIONS __AVX512VBMI__)
194
+ if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
195
+ list(APPEND ARCH_FLAGS -mavx512vbmi)
196
+ endif()
197
+ endif()
198
+ if (GGML_AVX512_VNNI)
199
+ list(APPEND ARCH_DEFINITIONS __AVX512VNNI__ GGML_AVX512_VNNI)
200
+ if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
201
+ list(APPEND ARCH_FLAGS -mavx512vnni)
202
+ endif()
203
+ endif()
204
+ if (GGML_AVX512_BF16)
205
+ list(APPEND ARCH_DEFINITIONS __AVX512BF16__ GGML_AVX512_BF16)
206
+ if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
207
+ list(APPEND ARCH_FLAGS -mavx512bf16)
208
+ endif()
209
+ endif()
210
+ if (GGML_AMX_TILE)
211
+ list(APPEND ARCH_DEFINITIONS __AMX_TILE__ GGML_AMX_TILE)
212
+ endif()
213
+ if (GGML_AMX_INT8)
214
+ list(APPEND ARCH_DEFINITIONS __AMX_INT8__ GGML_AMX_INT8)
215
+ endif()
216
+ if (GGML_AMX_BF16)
217
+ list(APPEND ARCH_DEFINITIONS __AMX_BF16__ GGML_AMX_BF16)
218
+ endif()
219
+ elseif (GGML_AVX2)
220
+ list(APPEND ARCH_FLAGS /arch:AVX2)
221
+ list(APPEND ARCH_DEFINITIONS GGML_AVX2 GGML_FMA GGML_F16C)
222
+ elseif (GGML_AVX)
223
+ list(APPEND ARCH_FLAGS /arch:AVX)
224
+ list(APPEND ARCH_DEFINITIONS GGML_AVX)
225
+ elseif (GGML_SSE42)
226
+ list(APPEND ARCH_FLAGS /arch:SSE4.2)
227
+ list(APPEND ARCH_DEFINITIONS GGML_SSE42)
228
+ endif()
229
+ if (GGML_AVX_VNNI)
230
+ list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI)
231
+ endif()
232
+ if (GGML_BMI2)
233
+ # MSVC does not define macro __BMI2__
234
+ list(APPEND ARCH_DEFINITIONS __BMI2__ GGML_BMI2)
235
+ endif()
236
+ else ()
237
+ if (GGML_NATIVE)
238
+ list(APPEND ARCH_FLAGS -march=native)
239
+ else ()
240
+ if (GGML_SSE42)
241
+ list(APPEND ARCH_FLAGS -msse4.2)
242
+ list(APPEND ARCH_DEFINITIONS GGML_SSE42)
243
+ endif()
244
+ if (GGML_F16C)
245
+ list(APPEND ARCH_FLAGS -mf16c)
246
+ list(APPEND ARCH_DEFINITIONS GGML_F16C)
247
+ endif()
248
+ if (GGML_FMA)
249
+ list(APPEND ARCH_FLAGS -mfma)
250
+ list(APPEND ARCH_DEFINITIONS GGML_FMA)
251
+ endif()
252
+ if (GGML_BMI2)
253
+ list(APPEND ARCH_FLAGS -mbmi2)
254
+ list(APPEND ARCH_DEFINITIONS GGML_BMI2)
255
+ endif()
256
+ if (GGML_AVX)
257
+ list(APPEND ARCH_FLAGS -mavx)
258
+ list(APPEND ARCH_DEFINITIONS GGML_AVX)
259
+ endif()
260
+ if (GGML_AVX2)
261
+ list(APPEND ARCH_FLAGS -mavx2)
262
+ list(APPEND ARCH_DEFINITIONS GGML_AVX2)
263
+ endif()
264
+ if (GGML_AVX_VNNI)
265
+ list(APPEND ARCH_FLAGS -mavxvnni)
266
+ list(APPEND ARCH_DEFINITIONS GGML_AVX_VNNI)
267
+ endif()
268
+ if (GGML_AVX512)
269
+ list(APPEND ARCH_FLAGS -mavx512f)
270
+ list(APPEND ARCH_FLAGS -mavx512cd)
271
+ list(APPEND ARCH_FLAGS -mavx512vl)
272
+ list(APPEND ARCH_FLAGS -mavx512dq)
273
+ list(APPEND ARCH_FLAGS -mavx512bw)
274
+ list(APPEND ARCH_DEFINITIONS GGML_AVX512)
275
+ endif()
276
+ if (GGML_AVX512_VBMI)
277
+ list(APPEND ARCH_FLAGS -mavx512vbmi)
278
+ list(APPEND ARCH_DEFINITIONS GGML_AVX512_VBMI)
279
+ endif()
280
+ if (GGML_AVX512_VNNI)
281
+ list(APPEND ARCH_FLAGS -mavx512vnni)
282
+ list(APPEND ARCH_DEFINITIONS GGML_AVX512_VNNI)
283
+ endif()
284
+ if (GGML_AVX512_BF16)
285
+ list(APPEND ARCH_FLAGS -mavx512bf16)
286
+ list(APPEND ARCH_DEFINITIONS GGML_AVX512_BF16)
287
+ endif()
288
+ if (GGML_AMX_TILE)
289
+ list(APPEND ARCH_FLAGS -mamx-tile)
290
+ list(APPEND ARCH_DEFINITIONS GGML_AMX_TILE)
291
+ endif()
292
+ if (GGML_AMX_INT8)
293
+ list(APPEND ARCH_FLAGS -mamx-int8)
294
+ list(APPEND ARCH_DEFINITIONS GGML_AMX_INT8)
295
+ endif()
296
+ if (GGML_AMX_BF16)
297
+ list(APPEND ARCH_FLAGS -mamx-bf16)
298
+ list(APPEND ARCH_DEFINITIONS GGML_AMX_BF16)
299
+ endif()
300
+ endif()
301
+ endif()
302
+
303
+ if (GGML_BACKEND_DL)
304
+ if (GGML_NATIVE)
305
+ # the feature check relies on ARCH_DEFINITIONS, but it is not set with GGML_NATIVE
306
+ message(FATAL_ERROR "GGML_NATIVE is not compatible with GGML_BACKEND_DL, consider using GGML_CPU_ALL_VARIANTS")
307
+ endif()
308
+
309
+ # The feature detection code is compiled as a separate target so that
310
+ # it can be built without the architecture flags
311
+ # Since multiple variants of the CPU backend may be included in the same
312
+ # build, using set_source_files_properties() to set the arch flags is not possible
313
+ set(GGML_CPU_FEATS_NAME ${GGML_CPU_NAME}-feats)
314
+ add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/cpu-feats-x86.cpp)
315
+ target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
316
+ target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARCH_DEFINITIONS})
317
+ target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
318
+ set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
319
+ target_link_libraries(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_FEATS_NAME})
320
+ endif()
321
+ elseif ("${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "ppc64le " OR "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
322
+ message(STATUS "PowerPC detected")
323
+ if (GGML_NATIVE)
324
+ if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
325
+ file(READ "/proc/cpuinfo" POWER10_M)
326
+ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "powerpc")
327
+ execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M)
328
+ endif()
329
+
330
+ string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M}")
331
+ string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
332
+
333
+ if (EXTRACTED_NUMBER GREATER_EQUAL 10)
334
+ list(APPEND ARCH_FLAGS -mcpu=power10 -mpowerpc64)
335
+ elseif (EXTRACTED_NUMBER EQUAL 9)
336
+ list(APPEND ARCH_FLAGS -mcpu=power9 -mpowerpc64)
337
+ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
338
+ list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
339
+ else()
340
+ list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
341
+ endif()
342
+ else()
343
+ if (GGML_CPU_POWERPC_CPUTYPE)
344
+ list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE})
345
+ endif()
346
+ endif()
347
+ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
348
+ message(STATUS "loongarch64 detected")
349
+
350
+ list(APPEND ARCH_FLAGS -march=loongarch64)
351
+ if (GGML_LASX)
352
+ list(APPEND ARCH_FLAGS -mlasx)
353
+ endif()
354
+ if (GGML_LSX)
355
+ list(APPEND ARCH_FLAGS -mlsx)
356
+ endif()
357
+ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64")
358
+ message(STATUS "RISC-V detected")
359
+ if (GGML_RVV)
360
+ if (GGML_XTHEADVECTOR)
361
+ list(APPEND ARCH_FLAGS -march=rv64gc_xtheadvector -mabi=lp64d)
362
+ elseif (GGML_RV_ZFH)
363
+ list(APPEND ARCH_FLAGS -march=rv64gcv_zfhmin -mabi=lp64d)
364
+ else()
365
+ list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
366
+ endif()
367
+ endif()
368
+ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
369
+ message(STATUS "s390x detected")
370
+ file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
371
+ string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
372
+
373
+ # TODO: Separation to determine activation of VX/VXE/VXE2
374
+ if (${S390X_M} MATCHES "8561|8562")
375
+ message(STATUS "z15 target")
376
+ list(APPEND ARCH_FLAGS -march=z15)
377
+ elseif (${S390X_M} MATCHES "3931")
378
+ message(STATUS "z16 target")
379
+ list(APPEND ARCH_FLAGS -march=z16)
380
+ elseif (${S390X_M} MATCHES "9175|9176")
381
+ # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
382
+ message(STATUS "z17 target")
383
+ list(APPEND ARCH_FLAGS -march=z17)
384
+ else()
385
+ message(STATUS "Unknown target")
386
+ message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
387
+ list(APPEND ARCH_FLAGS -march=native -mtune=native)
388
+ endif()
389
+
390
+ if (GGML_VXE)
391
+ list(APPEND ARCH_FLAGS -mvx -mzvector)
392
+ endif()
393
+ else()
394
+ message(STATUS "Unknown architecture")
395
+ endif()
396
+
397
+ if (GGML_CPU_AARCH64)
398
+ target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_AARCH64)
399
+ endif()
400
+
401
+ if (GGML_CPU_KLEIDIAI)
402
+ message(STATUS "Using KleidiAI optimized kernels if applicable")
403
+
404
+ # Disable the KleidiAI tests
405
+ set(KLEIDIAI_BUILD_TESTS OFF)
406
+
407
+ # Fetch KleidiAI sources:
408
+ include(FetchContent)
409
+ set(KLEIDIAI_COMMIT_TAG "v1.6.0")
410
+ set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
411
+ set(KLEIDIAI_ARCHIVE_MD5 "75b4ad68f25ab673dcc01065e5a0b05f")
412
+
413
+ if (POLICY CMP0135)
414
+ cmake_policy(SET CMP0135 NEW)
415
+ endif()
416
+
417
+ FetchContent_Declare(KleidiAI_Download
418
+ URL ${KLEIDIAI_DOWNLOAD_URL}
419
+ DOWNLOAD_EXTRACT_TIMESTAMP NEW
420
+ URL_HASH MD5=${KLEIDIAI_ARCHIVE_MD5})
421
+
422
+ FetchContent_MakeAvailable(KleidiAI_Download)
423
+ FetchContent_GetProperties(KleidiAI_Download
424
+ SOURCE_DIR KLEIDIAI_SRC
425
+ POPULATED KLEIDIAI_POPULATED)
426
+
427
+ if (NOT KLEIDIAI_POPULATED)
428
+ message(FATAL_ERROR "KleidiAI source downloaded failed.")
429
+ endif()
430
+
431
+ add_compile_definitions(GGML_USE_CPU_KLEIDIAI)
432
+
433
+ # Remove kleidiai target after fetching it
434
+ if (TARGET kleidiai)
435
+ set_target_properties(kleidiai PROPERTIES EXCLUDE_FROM_ALL TRUE)
436
+ endif()
437
+
438
+ list(APPEND GGML_CPU_SOURCES
439
+ ggml-cpu/kleidiai/kleidiai.cpp
440
+ ggml-cpu/kleidiai/kernels.cpp
441
+ ggml-cpu/kleidiai/kleidiai.h
442
+ ggml-cpu/kleidiai/kernels.h
443
+ )
444
+
445
+ # KleidiAI
446
+ include_directories(
447
+ ${KLEIDIAI_SRC}/
448
+ ${KLEIDIAI_SRC}/kai/
449
+ ${KLEIDIAI_SRC}/kai/ukernels/
450
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/
451
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
452
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
453
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)
454
+
455
+ set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}")
456
+ if (NOT ARCH_FLAGS_TEMP)
457
+ string(REGEX MATCH "-march=[^ ]+" ARCH_FLAGS_TEMP "${CMAKE_C_FLAGS}")
458
+ endif()
459
+ string(FIND "${ARCH_FLAGS_TEMP}" "+dotprod" DOTPROD_ENABLED)
460
+ string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED)
461
+ string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED)
462
+
463
+ set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS_TEMP})
464
+
465
+ list(APPEND GGML_KLEIDIAI_SOURCES
466
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c
467
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c
468
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c
469
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)
470
+
471
+ if (NOT DOTPROD_ENABLED MATCHES -1)
472
+ list(APPEND GGML_KLEIDIAI_SOURCES
473
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c
474
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c
475
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c)
476
+ endif()
477
+
478
+ if (NOT I8MM_ENABLED MATCHES -1)
479
+ list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c)
480
+ endif()
481
+
482
+ if (NOT SME_ENABLED MATCHES -1)
483
+ list(APPEND GGML_KLEIDIAI_SOURCES
484
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c
485
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
486
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
487
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
488
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c)
489
+ set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2")
490
+ endif()
491
+
492
+ set_source_files_properties(${GGML_KLEIDIAI_SOURCES} PROPERTIES COMPILE_OPTIONS "${PRIVATE_ARCH_FLAGS}")
493
+ list(APPEND GGML_CPU_SOURCES ${GGML_KLEIDIAI_SOURCES})
494
+ endif()
495
+
496
+ message(STATUS "Adding CPU backend variant ${GGML_CPU_NAME}: ${ARCH_FLAGS} ${ARCH_DEFINITIONS}")
497
+ target_sources(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_SOURCES})
498
+ target_compile_options(${GGML_CPU_NAME} PRIVATE ${ARCH_FLAGS})
499
+ target_compile_definitions(${GGML_CPU_NAME} PRIVATE ${ARCH_DEFINITIONS})
500
+
501
+ if (EMSCRIPTEN)
502
+ set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
503
+ endif()
504
+ endfunction()