whispercpp 1.3.6 → 1.3.7

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 (828) hide show
  1. checksums.yaml +4 -4
  2. data/.document +3 -0
  3. data/.rdoc_options +2 -0
  4. data/README.md +38 -5
  5. data/Rakefile +18 -3
  6. data/ext/dependencies.rb +10 -4
  7. data/ext/dependencies_for_windows.rb +17 -0
  8. data/ext/extconf.rb +20 -8
  9. data/ext/options.rb +54 -14
  10. data/ext/options_for_windows.rb +51 -0
  11. data/ext/ruby_whisper.c +36 -42
  12. data/ext/ruby_whisper.h +135 -0
  13. data/ext/ruby_whisper_context.c +107 -28
  14. data/ext/ruby_whisper_log_queue.c +180 -0
  15. data/ext/ruby_whisper_log_settable.h +47 -0
  16. data/ext/ruby_whisper_parakeet.c +49 -0
  17. data/ext/ruby_whisper_parakeet_context.c +304 -0
  18. data/ext/ruby_whisper_parakeet_context_params.c +117 -0
  19. data/ext/ruby_whisper_parakeet_model.c +84 -0
  20. data/ext/ruby_whisper_parakeet_params.c +548 -0
  21. data/ext/ruby_whisper_parakeet_segment.c +157 -0
  22. data/ext/ruby_whisper_parakeet_token.c +188 -0
  23. data/ext/ruby_whisper_parakeet_transcribe.cpp +58 -0
  24. data/ext/ruby_whisper_params.c +256 -65
  25. data/ext/ruby_whisper_segment.c +6 -6
  26. data/ext/ruby_whisper_transcribe.cpp +42 -15
  27. data/ext/sources/CMakeLists.txt +41 -3
  28. data/ext/sources/CMakePresets.json +95 -0
  29. data/ext/sources/cmake/parakeet-config.cmake.in +30 -0
  30. data/ext/sources/cmake/parakeet.pc.in +10 -0
  31. data/ext/sources/cmake/whisper.pc.in +1 -1
  32. data/ext/sources/examples/CMakeLists.txt +4 -2
  33. data/ext/sources/examples/bench/bench.cpp +1 -1
  34. data/ext/sources/examples/cli/cli.cpp +43 -9
  35. data/ext/sources/examples/common-ggml.cpp +2 -0
  36. data/ext/sources/examples/common-whisper.cpp +139 -67
  37. data/ext/sources/examples/common-whisper.h +11 -0
  38. data/ext/sources/examples/ffmpeg-transcode.cpp +211 -341
  39. data/ext/sources/examples/parakeet-cli/CMakeLists.txt +8 -0
  40. data/ext/sources/examples/parakeet-cli/parakeet-cli.cpp +243 -0
  41. data/ext/sources/examples/parakeet-quantize/CMakeLists.txt +7 -0
  42. data/ext/sources/examples/parakeet-quantize/parakeet-quantize.cpp +230 -0
  43. data/ext/sources/examples/server/server.cpp +199 -163
  44. data/ext/sources/ggml/CMakeLists.txt +21 -13
  45. data/ext/sources/ggml/cmake/FindNCCL.cmake +36 -0
  46. data/ext/sources/ggml/cmake/ggml-config.cmake.in +12 -2
  47. data/ext/sources/ggml/include/ggml-alloc.h +1 -0
  48. data/ext/sources/ggml/include/ggml-backend.h +72 -10
  49. data/ext/sources/ggml/include/ggml-cuda.h +3 -0
  50. data/ext/sources/ggml/include/ggml-rpc.h +3 -3
  51. data/ext/sources/ggml/include/ggml.h +101 -9
  52. data/ext/sources/ggml/include/gguf.h +10 -2
  53. data/ext/sources/ggml/src/CMakeLists.txt +22 -5
  54. data/ext/sources/ggml/src/ggml-alloc.c +5 -1
  55. data/ext/sources/ggml/src/ggml-backend-impl.h +22 -2
  56. data/ext/sources/ggml/src/ggml-backend-meta.cpp +2263 -0
  57. data/ext/sources/ggml/src/ggml-backend-reg.cpp +12 -0
  58. data/ext/sources/ggml/src/ggml-backend.cpp +110 -9
  59. data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +4 -0
  60. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +672 -257
  61. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +71 -0
  62. data/ext/sources/ggml/src/ggml-cann/common.h +20 -10
  63. data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +211 -30
  64. data/ext/sources/ggml/src/ggml-common.h +11 -0
  65. data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +58 -29
  66. data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +2 -0
  67. data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +16 -16
  68. data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +116 -7
  69. data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +65 -0
  70. data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +151 -1
  71. data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +0 -1
  72. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +4279 -1292
  73. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +5 -35
  74. data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +0 -1
  75. data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +72 -1
  76. data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +177 -27
  77. data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +1 -1
  78. data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +5 -0
  79. data/ext/sources/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
  80. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +10 -0
  81. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +95 -5
  82. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +2 -0
  83. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +146 -134
  84. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +88 -70
  85. data/ext/sources/ggml/src/ggml-cpu/ops.cpp +372 -73
  86. data/ext/sources/ggml/src/ggml-cpu/ops.h +3 -0
  87. data/ext/sources/ggml/src/ggml-cpu/quants.c +55 -0
  88. data/ext/sources/ggml/src/ggml-cpu/quants.h +3 -0
  89. data/ext/sources/ggml/src/ggml-cpu/repack.cpp +3 -0
  90. data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +90 -0
  91. data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +3 -16
  92. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +1402 -687
  93. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.h +8 -0
  94. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +597 -2766
  95. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp +5768 -0
  96. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.cpp +320 -0
  97. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.h +55 -0
  98. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_kernels.h +182 -19
  99. data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.cpp +1795 -0
  100. data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.h +14 -0
  101. data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp +3178 -0
  102. data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.h +95 -0
  103. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_barrier.h +34 -0
  104. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp +760 -0
  105. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h +32 -0
  106. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_tcm.h +409 -0
  107. data/ext/sources/ggml/src/ggml-cpu/vec.cpp +37 -53
  108. data/ext/sources/ggml/src/ggml-cpu/vec.h +225 -240
  109. data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +17 -7
  110. data/ext/sources/ggml/src/ggml-cuda/allreduce.cu +971 -0
  111. data/ext/sources/ggml/src/ggml-cuda/allreduce.cuh +29 -0
  112. data/ext/sources/ggml/src/ggml-cuda/argsort.cu +62 -26
  113. data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +44 -18
  114. data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +1 -0
  115. data/ext/sources/ggml/src/ggml-cuda/common.cuh +242 -28
  116. data/ext/sources/ggml/src/ggml-cuda/concat.cu +120 -114
  117. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cu +45 -21
  118. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cuh +1 -0
  119. data/ext/sources/ggml/src/ggml-cuda/convert.cu +53 -0
  120. data/ext/sources/ggml/src/ggml-cuda/convert.cuh +10 -0
  121. data/ext/sources/ggml/src/ggml-cuda/cpy.cu +14 -6
  122. data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +22 -0
  123. data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +278 -44
  124. data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +331 -130
  125. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +12 -0
  126. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +126 -27
  127. data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +40 -15
  128. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +18 -9
  129. data/ext/sources/ggml/src/ggml-cuda/fattn.cu +152 -49
  130. data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +2 -0
  131. data/ext/sources/ggml/src/ggml-cuda/fwht.cu +101 -0
  132. data/ext/sources/ggml/src/ggml-cuda/fwht.cuh +4 -0
  133. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cu +84 -35
  134. data/ext/sources/ggml/src/ggml-cuda/getrows.cu +34 -12
  135. data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +1069 -609
  136. data/ext/sources/ggml/src/ggml-cuda/im2col.cu +32 -29
  137. data/ext/sources/ggml/src/ggml-cuda/mean.cu +4 -2
  138. data/ext/sources/ggml/src/ggml-cuda/mma.cuh +242 -195
  139. data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +3 -3
  140. data/ext/sources/ggml/src/ggml-cuda/mmq.cu +18 -12
  141. data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +502 -423
  142. data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +19 -12
  143. data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +485 -57
  144. data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +6 -1
  145. data/ext/sources/ggml/src/ggml-cuda/norm.cu +36 -10
  146. data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +23 -7
  147. data/ext/sources/ggml/src/ggml-cuda/quantize.cu +133 -26
  148. data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +1 -1
  149. data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +5 -1
  150. data/ext/sources/ggml/src/ggml-cuda/rope.cu +11 -4
  151. data/ext/sources/ggml/src/ggml-cuda/scale.cu +4 -1
  152. data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +14 -6
  153. data/ext/sources/ggml/src/ggml-cuda/snake.cu +72 -0
  154. data/ext/sources/ggml/src/ggml-cuda/snake.cuh +8 -0
  155. data/ext/sources/ggml/src/ggml-cuda/softcap.cu +4 -1
  156. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +45 -13
  157. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +1 -1
  158. data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +40 -18
  159. data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +8 -4
  160. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +1 -0
  161. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +1 -0
  162. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +2 -0
  163. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +1 -0
  164. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +1 -0
  165. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +1 -0
  166. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +1 -0
  167. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +2 -0
  168. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +1 -0
  169. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +1 -0
  170. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +2 -0
  171. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +1 -0
  172. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +2 -0
  173. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu +5 -0
  174. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu +5 -0
  175. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +5 -0
  176. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu +7 -0
  177. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu +7 -0
  178. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu +7 -0
  179. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu +7 -0
  180. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu +7 -0
  181. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu +7 -0
  182. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu +7 -0
  183. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu +7 -0
  184. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu +7 -0
  185. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu +7 -0
  186. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu +7 -0
  187. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu +7 -0
  188. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu +7 -0
  189. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu +5 -0
  190. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu +5 -0
  191. data/ext/sources/ggml/src/ggml-cuda/top-k.cu +5 -4
  192. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +26 -23
  193. data/ext/sources/ggml/src/ggml-cuda/unary.cu +31 -2
  194. data/ext/sources/ggml/src/ggml-cuda/unary.cuh +2 -0
  195. data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +80 -0
  196. data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +7 -2
  197. data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +22 -4
  198. data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +3 -0
  199. data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +2 -1
  200. data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +1428 -743
  201. data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +45 -7
  202. data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +53 -84
  203. data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +25 -12
  204. data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +165 -184
  205. data/ext/sources/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +5 -5
  206. data/ext/sources/ggml/src/ggml-hexagon/htp/concat-ops.c +277 -0
  207. data/ext/sources/ggml/src/ggml-hexagon/htp/cpy-ops.c +170 -127
  208. data/ext/sources/ggml/src/ggml-hexagon/htp/cumsum-ops.c +270 -0
  209. data/ext/sources/ggml/src/ggml-hexagon/htp/diag-ops.c +216 -0
  210. data/ext/sources/ggml/src/ggml-hexagon/htp/fill-ops.c +123 -0
  211. data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +125 -97
  212. data/ext/sources/ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c +1148 -0
  213. data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +148 -42
  214. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.c +2 -2
  215. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.h +252 -62
  216. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +9 -0
  217. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +87 -1
  218. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-flash-attn-ops.c +1878 -0
  219. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-matmul-ops.c +2066 -0
  220. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-ops.c +6 -0
  221. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-ops.h +88 -0
  222. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-profile.h +34 -0
  223. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.c +158 -0
  224. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.h +134 -0
  225. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-utils.h +200 -0
  226. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +96 -13
  227. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +182 -57
  228. data/ext/sources/ggml/src/ggml-hexagon/htp/htp_iface.idl +9 -3
  229. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +71 -3
  230. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +27 -10
  231. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +63 -23
  232. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +9 -8
  233. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -0
  234. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-log.h +65 -0
  235. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-pow.h +42 -0
  236. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-repl.h +74 -0
  237. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +1 -0
  238. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
  239. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +5 -8
  240. data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +529 -815
  241. data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +2522 -234
  242. data/ext/sources/ggml/src/ggml-hexagon/htp/pad-ops.c +547 -0
  243. data/ext/sources/ggml/src/ggml-hexagon/htp/repeat-ops.c +148 -0
  244. data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +291 -95
  245. data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +59 -37
  246. data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +121 -133
  247. data/ext/sources/ggml/src/ggml-hexagon/htp/solve-tri-ops.c +267 -0
  248. data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c +244 -151
  249. data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +6 -6
  250. data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +719 -45
  251. data/ext/sources/ggml/src/ggml-hexagon/htp/vtcm-utils.h +16 -0
  252. data/ext/sources/ggml/src/ggml-hexagon/htp-opnode.h +272 -0
  253. data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +3 -1
  254. data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +22 -9
  255. data/ext/sources/ggml/src/ggml-impl.h +6 -1
  256. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +138 -13
  257. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +32 -1
  258. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +164 -28
  259. data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +80 -0
  260. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +190 -19
  261. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +2 -0
  262. data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +39 -26
  263. data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +823 -322
  264. data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +5 -6
  265. data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +54 -5
  266. data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +12248 -5907
  267. data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +67 -0
  268. data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +59 -0
  269. data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +1819 -112
  270. data/ext/sources/ggml/src/ggml-opencl/kernels/gated_delta_net.cl +249 -0
  271. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl +306 -0
  272. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl +256 -0
  273. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl +258 -0
  274. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl +283 -0
  275. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl +260 -0
  276. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl +262 -0
  277. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl +288 -0
  278. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl +267 -0
  279. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_f32.cl +150 -0
  280. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mat_Ab_Bi_8x4.cl → gemm_noshuffle_q4_0_f32.cl} +1 -1
  281. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl +172 -0
  282. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl +131 -0
  283. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl +134 -0
  284. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl +176 -0
  285. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl +140 -0
  286. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mm_q8_0_f32_8x4.cl → gemm_noshuffle_q8_0_f32.cl} +1 -1
  287. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_xmem_f16_f32_os8.cl +233 -0
  288. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl +165 -0
  289. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl +120 -0
  290. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl +123 -0
  291. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl +155 -0
  292. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl +123 -0
  293. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl +125 -0
  294. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl +160 -0
  295. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl +141 -0
  296. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +302 -0
  297. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general.cl → gemv_noshuffle_q4_0_f32.cl} +5 -5
  298. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle.cl → gemv_noshuffle_q4_0_f32_spec.cl} +5 -5
  299. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +318 -0
  300. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +291 -0
  301. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +294 -0
  302. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +326 -0
  303. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +293 -0
  304. data/ext/sources/ggml/src/ggml-opencl/kernels/get_rows.cl +15 -9
  305. data/ext/sources/ggml/src/ggml-opencl/kernels/moe_reorder_b.cl +30 -0
  306. data/ext/sources/ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl +82 -0
  307. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_iq4_nl_f32_l4_lm.cl +171 -0
  308. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl +179 -0
  309. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_0_f32_l4_lm.cl +173 -0
  310. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_1_f32_l4_lm.cl +175 -0
  311. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_k_f32_l4_lm.cl +192 -0
  312. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32.cl +164 -0
  313. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32_flat.cl +202 -0
  314. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl +196 -0
  315. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32.cl +241 -0
  316. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32_flat.cl +243 -0
  317. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32.cl +243 -0
  318. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32_flat.cl +247 -0
  319. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32.cl +187 -0
  320. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32_flat.cl +203 -0
  321. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +48 -64
  322. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +15 -5
  323. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +18 -11
  324. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +35 -13
  325. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +264 -192
  326. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +33 -7
  327. data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_gelu.cpp +25 -0
  328. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +1 -0
  329. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +1 -0
  330. data/ext/sources/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp +41 -0
  331. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.cpp +27 -3
  332. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +67 -36
  333. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +1 -0
  334. data/ext/sources/ggml/src/ggml-openvino/utils.cpp +101 -44
  335. data/ext/sources/ggml/src/ggml-openvino/utils.h +23 -3
  336. data/ext/sources/ggml/src/ggml-opt.cpp +1 -0
  337. data/ext/sources/ggml/src/ggml-quants.c +289 -114
  338. data/ext/sources/ggml/src/ggml-quants.h +3 -0
  339. data/ext/sources/ggml/src/ggml-rpc/CMakeLists.txt +24 -0
  340. data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +167 -311
  341. data/ext/sources/ggml/src/ggml-rpc/transport.cpp +683 -0
  342. data/ext/sources/ggml/src/ggml-rpc/transport.h +34 -0
  343. data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +50 -4
  344. data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +1 -1
  345. data/ext/sources/ggml/src/ggml-sycl/backend.hpp +3 -1
  346. data/ext/sources/ggml/src/ggml-sycl/common.cpp +74 -2
  347. data/ext/sources/ggml/src/ggml-sycl/common.hpp +41 -1
  348. data/ext/sources/ggml/src/ggml-sycl/convert.cpp +115 -13
  349. data/ext/sources/ggml/src/ggml-sycl/convert.hpp +9 -0
  350. data/ext/sources/ggml/src/ggml-sycl/cumsum.cpp +148 -0
  351. data/ext/sources/ggml/src/ggml-sycl/cumsum.hpp +5 -0
  352. data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +663 -0
  353. data/ext/sources/ggml/src/ggml-sycl/diag.cpp +67 -0
  354. data/ext/sources/ggml/src/ggml-sycl/diag.hpp +5 -0
  355. data/ext/sources/ggml/src/ggml-sycl/dmmv.cpp +586 -6
  356. data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +1 -90
  357. data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +0 -2
  358. data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.cpp +56 -0
  359. data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.hpp +63 -0
  360. data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp +7 -5
  361. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +4 -0
  362. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +76 -168
  363. data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +7 -0
  364. data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +3 -1
  365. data/ext/sources/ggml/src/ggml-sycl/fill.cpp +55 -0
  366. data/ext/sources/ggml/src/ggml-sycl/fill.hpp +5 -0
  367. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.cpp +69 -31
  368. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +1 -0
  369. data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +3 -0
  370. data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +79 -3
  371. data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +823 -190
  372. data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +353 -89
  373. data/ext/sources/ggml/src/ggml-sycl/im2col.hpp +5 -3
  374. data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +1344 -26
  375. data/ext/sources/ggml/src/ggml-sycl/mmvq.hpp +16 -0
  376. data/ext/sources/ggml/src/ggml-sycl/pad.cpp +27 -27
  377. data/ext/sources/ggml/src/ggml-sycl/quants.hpp +71 -0
  378. data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +7 -1
  379. data/ext/sources/ggml/src/ggml-sycl/solve_tri.cpp +172 -0
  380. data/ext/sources/ggml/src/ggml-sycl/solve_tri.hpp +8 -0
  381. data/ext/sources/ggml/src/ggml-sycl/ssm_conv.cpp +6 -1
  382. data/ext/sources/ggml/src/ggml-sycl/ssm_scan.cpp +156 -0
  383. data/ext/sources/ggml/src/ggml-sycl/ssm_scan.hpp +5 -0
  384. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +62 -10
  385. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +18 -6
  386. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp +6 -0
  387. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +1 -0
  388. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +1 -0
  389. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +1 -0
  390. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +1 -0
  391. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +1 -0
  392. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +1 -0
  393. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +1 -0
  394. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +1 -0
  395. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +1 -0
  396. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +1 -0
  397. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +1 -0
  398. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +1 -0
  399. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +1 -0
  400. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +1 -0
  401. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +1 -0
  402. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +1 -0
  403. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +1 -0
  404. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +1 -0
  405. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +1 -0
  406. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +1 -0
  407. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +1 -0
  408. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +1 -0
  409. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +1 -0
  410. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +1 -0
  411. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +1 -0
  412. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +1 -0
  413. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +1 -0
  414. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +1 -0
  415. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +1 -0
  416. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +1 -0
  417. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +1 -0
  418. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +1 -0
  419. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +1 -0
  420. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +1 -0
  421. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +1 -0
  422. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +1 -0
  423. data/ext/sources/ggml/src/ggml-sycl/type.hpp +112 -0
  424. data/ext/sources/ggml/src/ggml-sycl/upscale.cpp +410 -0
  425. data/ext/sources/ggml/src/ggml-sycl/upscale.hpp +9 -0
  426. data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +215 -53
  427. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +4 -0
  428. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +2 -0
  429. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +2 -0
  430. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +1 -0
  431. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +1 -0
  432. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +0 -2
  433. data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +11 -0
  434. data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +2060 -535
  435. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +4 -0
  436. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +6 -2
  437. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +146 -13
  438. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +3 -1
  439. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +1 -1
  440. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +25 -1
  441. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +88 -0
  442. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl +643 -1
  443. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp +32 -0
  444. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q1_0.comp +29 -0
  445. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +0 -1
  446. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl +27 -0
  447. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +0 -1
  448. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2_decode_vector.comp +7 -0
  449. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +197 -48
  450. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +60 -59
  451. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +115 -113
  452. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +122 -31
  453. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_dequant.glsl +131 -0
  454. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mmq_funcs.glsl +203 -0
  455. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp +115 -0
  456. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +125 -64
  457. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl +0 -1
  458. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl +10 -1
  459. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl +16 -6
  460. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +76 -54
  461. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +0 -1
  462. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +0 -1
  463. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +122 -27
  464. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +6 -6
  465. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +1 -1
  466. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +1 -1
  467. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +1 -1
  468. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +1 -0
  469. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +88 -55
  470. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +11 -17
  471. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +43 -10
  472. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +159 -125
  473. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +8 -8
  474. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +24 -9
  475. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +0 -1
  476. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +5 -2
  477. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +0 -1
  478. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +3 -2
  479. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/snake.comp +49 -0
  480. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +11 -1
  481. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +0 -1
  482. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +79 -2
  483. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +171 -147
  484. data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +5 -2
  485. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +2202 -283
  486. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +2610 -1403
  487. data/ext/sources/ggml/src/ggml-webgpu/pre_wgsl.hpp +37 -7
  488. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_id.wgsl +64 -0
  489. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +8 -7
  490. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +76 -95
  491. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +19 -1
  492. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/conv2d.wgsl +165 -0
  493. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{cpy.tmpl.wgsl → cpy.wgsl} +25 -50
  494. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +107 -184
  495. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_quant_staging.tmpl +124 -0
  496. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl +397 -0
  497. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_blk.wgsl +101 -0
  498. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_reduce.wgsl +84 -0
  499. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl +619 -0
  500. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl +149 -0
  501. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl +183 -78
  502. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl +155 -0
  503. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/im2col.wgsl +101 -0
  504. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +655 -495
  505. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id.wgsl +195 -0
  506. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_gather.wgsl +52 -0
  507. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl +154 -0
  508. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +8 -6
  509. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.wgsl +5 -1
  510. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +80 -409
  511. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl +1432 -0
  512. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl +303 -0
  513. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl +21 -0
  514. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl +173 -0
  515. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl +152 -0
  516. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{rope.tmpl.wgsl → rope.wgsl} +71 -142
  517. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/row_norm.wgsl +153 -0
  518. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/scale.wgsl +6 -4
  519. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set.wgsl +109 -0
  520. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +2 -3
  521. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows_quant.wgsl +224 -0
  522. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{soft_max.tmpl.wgsl → soft_max.wgsl} +106 -206
  523. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/solve_tri.wgsl +121 -0
  524. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_conv.wgsl +65 -0
  525. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl +193 -0
  526. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +68 -48
  527. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
  528. data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +18 -14
  529. data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +1 -1
  530. data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +244 -10
  531. data/ext/sources/ggml/src/ggml.c +110 -28
  532. data/ext/sources/ggml/src/gguf.cpp +173 -28
  533. data/ext/sources/include/parakeet.h +342 -0
  534. data/ext/sources/include/whisper.h +10 -0
  535. data/ext/sources/media/matmul.png +0 -0
  536. data/ext/sources/src/CMakeLists.txt +23 -0
  537. data/ext/sources/src/parakeet-arch.h +188 -0
  538. data/ext/sources/src/parakeet.cpp +3838 -0
  539. data/ext/sources/src/whisper.cpp +56 -12
  540. data/extsources.rb +26 -10
  541. data/lib/whisper/log_settable.rb +36 -0
  542. data/lib/whisper/model/uri.rb +13 -1
  543. data/lib/whisper/output.rb +74 -0
  544. data/sig/whisper.rbs +411 -62
  545. data/test/helper.rb +2 -0
  546. data/test/jfk_reader/jfk_reader.c +50 -7
  547. data/test/test_callback.rb +1 -0
  548. data/test/test_package.rb +6 -5
  549. data/test/test_parakeet.rb +28 -0
  550. data/test/test_parakeet_callback.rb +107 -0
  551. data/test/test_parakeet_context.rb +116 -0
  552. data/test/test_parakeet_context_params.rb +24 -0
  553. data/test/test_parakeet_model.rb +21 -0
  554. data/test/test_parakeet_params.rb +78 -0
  555. data/test/test_parakeet_segment.rb +42 -0
  556. data/test/test_parakeet_token.rb +73 -0
  557. data/test/test_params.rb +2 -0
  558. data/test/test_vad_segment.rb +1 -1
  559. data/test/test_whisper.rb +24 -6
  560. data/whispercpp.gemspec +2 -2
  561. metadata +215 -281
  562. data/ext/sources/bindings/javascript/CMakeLists.txt +0 -41
  563. data/ext/sources/bindings/javascript/emscripten.cpp +0 -93
  564. data/ext/sources/bindings/javascript/libwhisper.worker.js +0 -1
  565. data/ext/sources/bindings/javascript/package.json +0 -26
  566. data/ext/sources/bindings/javascript/whisper.js +0 -19
  567. data/ext/sources/examples/addon.node/CMakeLists.txt +0 -31
  568. data/ext/sources/examples/addon.node/__test__/whisper.spec.js +0 -133
  569. data/ext/sources/examples/addon.node/addon.cpp +0 -557
  570. data/ext/sources/examples/addon.node/index.js +0 -59
  571. data/ext/sources/examples/addon.node/package.json +0 -16
  572. data/ext/sources/examples/addon.node/vad-example.js +0 -132
  573. data/ext/sources/examples/bench.wasm/CMakeLists.txt +0 -49
  574. data/ext/sources/examples/bench.wasm/emscripten.cpp +0 -87
  575. data/ext/sources/examples/bench.wasm/index-tmpl.html +0 -285
  576. data/ext/sources/examples/coi-serviceworker.js +0 -146
  577. data/ext/sources/examples/command/CMakeLists.txt +0 -10
  578. data/ext/sources/examples/command/command.cpp +0 -802
  579. data/ext/sources/examples/command/commands.txt +0 -9
  580. data/ext/sources/examples/command.wasm/CMakeLists.txt +0 -50
  581. data/ext/sources/examples/command.wasm/emscripten.cpp +0 -327
  582. data/ext/sources/examples/command.wasm/index-tmpl.html +0 -415
  583. data/ext/sources/examples/generate-karaoke.sh +0 -57
  584. data/ext/sources/examples/helpers.js +0 -191
  585. data/ext/sources/examples/livestream.sh +0 -112
  586. data/ext/sources/examples/lsp/CMakeLists.txt +0 -10
  587. data/ext/sources/examples/lsp/lsp.cpp +0 -471
  588. data/ext/sources/examples/lsp/whisper.vim +0 -362
  589. data/ext/sources/examples/python/test_whisper_processor.py +0 -7
  590. data/ext/sources/examples/python/whisper_processor.py +0 -54
  591. data/ext/sources/examples/server/bench.js +0 -29
  592. data/ext/sources/examples/server.py +0 -120
  593. data/ext/sources/examples/stream/CMakeLists.txt +0 -10
  594. data/ext/sources/examples/stream/stream.cpp +0 -437
  595. data/ext/sources/examples/stream.wasm/CMakeLists.txt +0 -49
  596. data/ext/sources/examples/stream.wasm/emscripten.cpp +0 -216
  597. data/ext/sources/examples/stream.wasm/index-tmpl.html +0 -491
  598. data/ext/sources/examples/sycl/CMakeLists.txt +0 -9
  599. data/ext/sources/examples/sycl/build.sh +0 -22
  600. data/ext/sources/examples/sycl/ls-sycl-device.cpp +0 -11
  601. data/ext/sources/examples/sycl/run-whisper.sh +0 -17
  602. data/ext/sources/examples/talk-llama/CMakeLists.txt +0 -48
  603. data/ext/sources/examples/talk-llama/eleven-labs.py +0 -80
  604. data/ext/sources/examples/talk-llama/llama-adapter.cpp +0 -488
  605. data/ext/sources/examples/talk-llama/llama-adapter.h +0 -89
  606. data/ext/sources/examples/talk-llama/llama-arch.cpp +0 -2877
  607. data/ext/sources/examples/talk-llama/llama-arch.h +0 -628
  608. data/ext/sources/examples/talk-llama/llama-batch.cpp +0 -919
  609. data/ext/sources/examples/talk-llama/llama-batch.h +0 -173
  610. data/ext/sources/examples/talk-llama/llama-chat.cpp +0 -896
  611. data/ext/sources/examples/talk-llama/llama-chat.h +0 -71
  612. data/ext/sources/examples/talk-llama/llama-context.cpp +0 -3633
  613. data/ext/sources/examples/talk-llama/llama-context.h +0 -359
  614. data/ext/sources/examples/talk-llama/llama-cparams.cpp +0 -5
  615. data/ext/sources/examples/talk-llama/llama-cparams.h +0 -47
  616. data/ext/sources/examples/talk-llama/llama-ext.h +0 -12
  617. data/ext/sources/examples/talk-llama/llama-grammar.cpp +0 -1464
  618. data/ext/sources/examples/talk-llama/llama-grammar.h +0 -194
  619. data/ext/sources/examples/talk-llama/llama-graph.cpp +0 -2735
  620. data/ext/sources/examples/talk-llama/llama-graph.h +0 -1031
  621. data/ext/sources/examples/talk-llama/llama-hparams.cpp +0 -258
  622. data/ext/sources/examples/talk-llama/llama-hparams.h +0 -353
  623. data/ext/sources/examples/talk-llama/llama-impl.cpp +0 -171
  624. data/ext/sources/examples/talk-llama/llama-impl.h +0 -75
  625. data/ext/sources/examples/talk-llama/llama-io.cpp +0 -15
  626. data/ext/sources/examples/talk-llama/llama-io.h +0 -35
  627. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +0 -330
  628. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.h +0 -137
  629. data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +0 -2285
  630. data/ext/sources/examples/talk-llama/llama-kv-cache.h +0 -389
  631. data/ext/sources/examples/talk-llama/llama-kv-cells.h +0 -533
  632. data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.cpp +0 -275
  633. data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.h +0 -140
  634. data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +0 -268
  635. data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +0 -139
  636. data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +0 -1165
  637. data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +0 -182
  638. data/ext/sources/examples/talk-llama/llama-memory.cpp +0 -59
  639. data/ext/sources/examples/talk-llama/llama-memory.h +0 -122
  640. data/ext/sources/examples/talk-llama/llama-mmap.cpp +0 -752
  641. data/ext/sources/examples/talk-llama/llama-mmap.h +0 -73
  642. data/ext/sources/examples/talk-llama/llama-model-loader.cpp +0 -1655
  643. data/ext/sources/examples/talk-llama/llama-model-loader.h +0 -206
  644. data/ext/sources/examples/talk-llama/llama-model-saver.cpp +0 -299
  645. data/ext/sources/examples/talk-llama/llama-model-saver.h +0 -40
  646. data/ext/sources/examples/talk-llama/llama-model.cpp +0 -9056
  647. data/ext/sources/examples/talk-llama/llama-model.h +0 -597
  648. data/ext/sources/examples/talk-llama/llama-quant.cpp +0 -1304
  649. data/ext/sources/examples/talk-llama/llama-quant.h +0 -1
  650. data/ext/sources/examples/talk-llama/llama-sampler.cpp +0 -3885
  651. data/ext/sources/examples/talk-llama/llama-sampler.h +0 -42
  652. data/ext/sources/examples/talk-llama/llama-vocab.cpp +0 -3970
  653. data/ext/sources/examples/talk-llama/llama-vocab.h +0 -187
  654. data/ext/sources/examples/talk-llama/llama.cpp +0 -1194
  655. data/ext/sources/examples/talk-llama/llama.h +0 -1573
  656. data/ext/sources/examples/talk-llama/models/afmoe.cpp +0 -190
  657. data/ext/sources/examples/talk-llama/models/apertus.cpp +0 -125
  658. data/ext/sources/examples/talk-llama/models/arcee.cpp +0 -135
  659. data/ext/sources/examples/talk-llama/models/arctic.cpp +0 -137
  660. data/ext/sources/examples/talk-llama/models/arwkv7.cpp +0 -86
  661. data/ext/sources/examples/talk-llama/models/baichuan.cpp +0 -123
  662. data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +0 -143
  663. data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +0 -133
  664. data/ext/sources/examples/talk-llama/models/bert.cpp +0 -184
  665. data/ext/sources/examples/talk-llama/models/bitnet.cpp +0 -145
  666. data/ext/sources/examples/talk-llama/models/bloom.cpp +0 -101
  667. data/ext/sources/examples/talk-llama/models/chameleon.cpp +0 -178
  668. data/ext/sources/examples/talk-llama/models/chatglm.cpp +0 -132
  669. data/ext/sources/examples/talk-llama/models/codeshell.cpp +0 -111
  670. data/ext/sources/examples/talk-llama/models/cogvlm.cpp +0 -102
  671. data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +0 -134
  672. data/ext/sources/examples/talk-llama/models/command-r.cpp +0 -122
  673. data/ext/sources/examples/talk-llama/models/dbrx.cpp +0 -122
  674. data/ext/sources/examples/talk-llama/models/deci.cpp +0 -135
  675. data/ext/sources/examples/talk-llama/models/deepseek.cpp +0 -142
  676. data/ext/sources/examples/talk-llama/models/deepseek2.cpp +0 -262
  677. data/ext/sources/examples/talk-llama/models/delta-net-base.cpp +0 -445
  678. data/ext/sources/examples/talk-llama/models/dots1.cpp +0 -132
  679. data/ext/sources/examples/talk-llama/models/dream.cpp +0 -105
  680. data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +0 -148
  681. data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +0 -110
  682. data/ext/sources/examples/talk-llama/models/eurobert.cpp +0 -97
  683. data/ext/sources/examples/talk-llama/models/exaone-moe.cpp +0 -145
  684. data/ext/sources/examples/talk-llama/models/exaone.cpp +0 -114
  685. data/ext/sources/examples/talk-llama/models/exaone4.cpp +0 -123
  686. data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +0 -111
  687. data/ext/sources/examples/talk-llama/models/falcon.cpp +0 -120
  688. data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +0 -116
  689. data/ext/sources/examples/talk-llama/models/gemma.cpp +0 -112
  690. data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +0 -128
  691. data/ext/sources/examples/talk-llama/models/gemma3.cpp +0 -155
  692. data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +0 -384
  693. data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +0 -170
  694. data/ext/sources/examples/talk-llama/models/glm4.cpp +0 -157
  695. data/ext/sources/examples/talk-llama/models/gpt2.cpp +0 -105
  696. data/ext/sources/examples/talk-llama/models/gptneox.cpp +0 -144
  697. data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +0 -195
  698. data/ext/sources/examples/talk-llama/models/granite.cpp +0 -210
  699. data/ext/sources/examples/talk-llama/models/grok.cpp +0 -159
  700. data/ext/sources/examples/talk-llama/models/grovemoe.cpp +0 -139
  701. data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +0 -132
  702. data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +0 -153
  703. data/ext/sources/examples/talk-llama/models/internlm2.cpp +0 -120
  704. data/ext/sources/examples/talk-llama/models/jais.cpp +0 -86
  705. data/ext/sources/examples/talk-llama/models/jais2.cpp +0 -123
  706. data/ext/sources/examples/talk-llama/models/jamba.cpp +0 -106
  707. data/ext/sources/examples/talk-llama/models/kimi-linear.cpp +0 -381
  708. data/ext/sources/examples/talk-llama/models/lfm2.cpp +0 -196
  709. data/ext/sources/examples/talk-llama/models/llada-moe.cpp +0 -122
  710. data/ext/sources/examples/talk-llama/models/llada.cpp +0 -99
  711. data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +0 -178
  712. data/ext/sources/examples/talk-llama/models/llama.cpp +0 -175
  713. data/ext/sources/examples/talk-llama/models/maincoder.cpp +0 -117
  714. data/ext/sources/examples/talk-llama/models/mamba-base.cpp +0 -289
  715. data/ext/sources/examples/talk-llama/models/mamba.cpp +0 -54
  716. data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +0 -129
  717. data/ext/sources/examples/talk-llama/models/minicpm3.cpp +0 -200
  718. data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +0 -123
  719. data/ext/sources/examples/talk-llama/models/mistral3.cpp +0 -160
  720. data/ext/sources/examples/talk-llama/models/models.h +0 -704
  721. data/ext/sources/examples/talk-llama/models/modern-bert.cpp +0 -109
  722. data/ext/sources/examples/talk-llama/models/mpt.cpp +0 -126
  723. data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +0 -162
  724. data/ext/sources/examples/talk-llama/models/nemotron.cpp +0 -122
  725. data/ext/sources/examples/talk-llama/models/neo-bert.cpp +0 -104
  726. data/ext/sources/examples/talk-llama/models/olmo.cpp +0 -121
  727. data/ext/sources/examples/talk-llama/models/olmo2.cpp +0 -150
  728. data/ext/sources/examples/talk-llama/models/olmoe.cpp +0 -124
  729. data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +0 -127
  730. data/ext/sources/examples/talk-llama/models/openelm.cpp +0 -124
  731. data/ext/sources/examples/talk-llama/models/orion.cpp +0 -123
  732. data/ext/sources/examples/talk-llama/models/paddleocr.cpp +0 -122
  733. data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +0 -121
  734. data/ext/sources/examples/talk-llama/models/phi2.cpp +0 -121
  735. data/ext/sources/examples/talk-llama/models/phi3.cpp +0 -152
  736. data/ext/sources/examples/talk-llama/models/plamo.cpp +0 -110
  737. data/ext/sources/examples/talk-llama/models/plamo2.cpp +0 -320
  738. data/ext/sources/examples/talk-llama/models/plamo3.cpp +0 -128
  739. data/ext/sources/examples/talk-llama/models/plm.cpp +0 -169
  740. data/ext/sources/examples/talk-llama/models/qwen.cpp +0 -108
  741. data/ext/sources/examples/talk-llama/models/qwen2.cpp +0 -126
  742. data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +0 -151
  743. data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +0 -117
  744. data/ext/sources/examples/talk-llama/models/qwen3.cpp +0 -120
  745. data/ext/sources/examples/talk-llama/models/qwen35.cpp +0 -381
  746. data/ext/sources/examples/talk-llama/models/qwen35moe.cpp +0 -422
  747. data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +0 -131
  748. data/ext/sources/examples/talk-llama/models/qwen3next.cpp +0 -525
  749. data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +0 -140
  750. data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +0 -132
  751. data/ext/sources/examples/talk-llama/models/refact.cpp +0 -94
  752. data/ext/sources/examples/talk-llama/models/rnd1.cpp +0 -126
  753. data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +0 -164
  754. data/ext/sources/examples/talk-llama/models/rwkv6.cpp +0 -94
  755. data/ext/sources/examples/talk-llama/models/rwkv6qwen2.cpp +0 -86
  756. data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +0 -137
  757. data/ext/sources/examples/talk-llama/models/rwkv7.cpp +0 -90
  758. data/ext/sources/examples/talk-llama/models/seed-oss.cpp +0 -124
  759. data/ext/sources/examples/talk-llama/models/smallthinker.cpp +0 -126
  760. data/ext/sources/examples/talk-llama/models/smollm3.cpp +0 -128
  761. data/ext/sources/examples/talk-llama/models/stablelm.cpp +0 -146
  762. data/ext/sources/examples/talk-llama/models/starcoder.cpp +0 -100
  763. data/ext/sources/examples/talk-llama/models/starcoder2.cpp +0 -121
  764. data/ext/sources/examples/talk-llama/models/step35-iswa.cpp +0 -165
  765. data/ext/sources/examples/talk-llama/models/t5-dec.cpp +0 -166
  766. data/ext/sources/examples/talk-llama/models/t5-enc.cpp +0 -96
  767. data/ext/sources/examples/talk-llama/models/wavtokenizer-dec.cpp +0 -149
  768. data/ext/sources/examples/talk-llama/models/xverse.cpp +0 -108
  769. data/ext/sources/examples/talk-llama/prompts/talk-alpaca.txt +0 -23
  770. data/ext/sources/examples/talk-llama/speak +0 -40
  771. data/ext/sources/examples/talk-llama/speak.bat +0 -1
  772. data/ext/sources/examples/talk-llama/speak.ps1 +0 -14
  773. data/ext/sources/examples/talk-llama/talk-llama.cpp +0 -813
  774. data/ext/sources/examples/talk-llama/unicode-data.cpp +0 -7034
  775. data/ext/sources/examples/talk-llama/unicode-data.h +0 -20
  776. data/ext/sources/examples/talk-llama/unicode.cpp +0 -1103
  777. data/ext/sources/examples/talk-llama/unicode.h +0 -111
  778. data/ext/sources/examples/wchess/CMakeLists.txt +0 -10
  779. data/ext/sources/examples/wchess/libwchess/CMakeLists.txt +0 -19
  780. data/ext/sources/examples/wchess/libwchess/Chessboard.cpp +0 -803
  781. data/ext/sources/examples/wchess/libwchess/Chessboard.h +0 -33
  782. data/ext/sources/examples/wchess/libwchess/WChess.cpp +0 -193
  783. data/ext/sources/examples/wchess/libwchess/WChess.h +0 -63
  784. data/ext/sources/examples/wchess/libwchess/test-chessboard.cpp +0 -117
  785. data/ext/sources/examples/wchess/wchess.cmd/CMakeLists.txt +0 -8
  786. data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +0 -253
  787. data/ext/sources/examples/whisper.wasm/CMakeLists.txt +0 -50
  788. data/ext/sources/examples/whisper.wasm/emscripten.cpp +0 -118
  789. data/ext/sources/examples/whisper.wasm/index-tmpl.html +0 -659
  790. data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +0 -99
  791. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +0 -155
  792. data/ext/sources/ggml/src/ggml-hexagon/op-desc.h +0 -153
  793. data/ext/sources/ggml/src/ggml-opencl/kernels/embed_kernel.py +0 -26
  794. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.cpp +0 -123
  795. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.h +0 -17
  796. data/ext/sources/ggml/src/ggml-virtgpu/regenerate_remoting.py +0 -333
  797. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.glsl +0 -5
  798. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +0 -182
  799. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.tmpl.wgsl +0 -323
  800. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl +0 -718
  801. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +0 -123
  802. data/ext/sources/tests/CMakeLists.txt +0 -112
  803. data/ext/sources/tests/earnings21/eval.mk +0 -58
  804. data/ext/sources/tests/earnings21/eval.py +0 -68
  805. data/ext/sources/tests/earnings21/normalizers/__init__.py +0 -2
  806. data/ext/sources/tests/earnings21/normalizers/basic.py +0 -80
  807. data/ext/sources/tests/earnings21/normalizers/english.json +0 -1741
  808. data/ext/sources/tests/earnings21/normalizers/english.py +0 -550
  809. data/ext/sources/tests/earnings21/requirements.txt +0 -6
  810. data/ext/sources/tests/en-0-ref.txt +0 -1
  811. data/ext/sources/tests/en-1-ref.txt +0 -1
  812. data/ext/sources/tests/en-2-ref.txt +0 -1
  813. data/ext/sources/tests/es-0-ref.txt +0 -1
  814. data/ext/sources/tests/librispeech/eval.mk +0 -39
  815. data/ext/sources/tests/librispeech/eval.py +0 -47
  816. data/ext/sources/tests/librispeech/normalizers/__init__.py +0 -2
  817. data/ext/sources/tests/librispeech/normalizers/basic.py +0 -80
  818. data/ext/sources/tests/librispeech/normalizers/english.json +0 -1741
  819. data/ext/sources/tests/librispeech/normalizers/english.py +0 -550
  820. data/ext/sources/tests/librispeech/requirements.txt +0 -6
  821. data/ext/sources/tests/run-tests.sh +0 -130
  822. data/ext/sources/tests/test-c.c +0 -3
  823. data/ext/sources/tests/test-vad-full.cpp +0 -56
  824. data/ext/sources/tests/test-vad.cpp +0 -83
  825. data/ext/sources/tests/test-whisper.js +0 -58
  826. data/lib/whisper/context.rb +0 -15
  827. data/lib/whisper/segment.rb +0 -58
  828. /data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general_q8_0_f32.cl → gemv_noshuffle_q8_0_f32.cl} +0 -0
@@ -1,752 +0,0 @@
1
- #include "llama-mmap.h"
2
-
3
- #include "llama-impl.h"
4
-
5
- #include "ggml.h"
6
-
7
- #include <cstring>
8
- #include <climits>
9
- #include <stdexcept>
10
- #include <cerrno>
11
- #include <algorithm>
12
-
13
- #ifdef __has_include
14
- #if __has_include(<unistd.h>)
15
- #include <unistd.h>
16
- #include <fcntl.h>
17
- #include <sys/stat.h>
18
- #if defined(_POSIX_MAPPED_FILES)
19
- #include <sys/mman.h>
20
- #endif
21
- #if defined(_POSIX_MEMLOCK_RANGE)
22
- #include <sys/resource.h>
23
- #endif
24
- #endif
25
- #endif
26
-
27
- #if defined(_WIN32)
28
- #define WIN32_LEAN_AND_MEAN
29
- #ifndef NOMINMAX
30
- #define NOMINMAX
31
- #endif
32
- #include <windows.h>
33
- #ifndef PATH_MAX
34
- #define PATH_MAX MAX_PATH
35
- #endif
36
- #include <io.h>
37
- #endif
38
-
39
- #if defined(__APPLE__)
40
- #include <TargetConditionals.h>
41
- #endif
42
-
43
- // TODO: consider moving to llama-impl.h if needed in more places
44
- #if defined(_WIN32)
45
- static std::string llama_format_win_err(DWORD err) {
46
- LPSTR buf;
47
- size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
48
- NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
49
- if (!size) {
50
- return "FormatMessageA failed";
51
- }
52
- std::string ret(buf, size);
53
- LocalFree(buf);
54
- return ret;
55
- }
56
- #endif
57
-
58
- // llama_file
59
-
60
- struct llama_file::impl {
61
- #if defined(_WIN32)
62
- HANDLE fp_win32;
63
- std::string GetErrorMessageWin32(DWORD error_code) const {
64
- std::string ret;
65
- LPSTR lpMsgBuf = NULL;
66
- DWORD bufLen = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
67
- NULL, error_code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&lpMsgBuf, 0, NULL);
68
- if (!bufLen) {
69
- ret = format("Win32 error code: %lx", error_code);
70
- } else {
71
- ret = lpMsgBuf;
72
- LocalFree(lpMsgBuf);
73
- }
74
-
75
- return ret;
76
- }
77
-
78
- impl(const char * fname, const char * mode, [[maybe_unused]] const bool use_direct_io = false) {
79
- fp = ggml_fopen(fname, mode);
80
- if (fp == NULL) {
81
- throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
82
- }
83
- fp_win32 = (HANDLE) _get_osfhandle(_fileno(fp));
84
- seek(0, SEEK_END);
85
- size = tell();
86
- seek(0, SEEK_SET);
87
- }
88
-
89
- size_t tell() const {
90
- LARGE_INTEGER li;
91
- li.QuadPart = 0;
92
- BOOL ret = SetFilePointerEx(fp_win32, li, &li, FILE_CURRENT);
93
- if (!ret) {
94
- throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
95
- }
96
-
97
- return li.QuadPart;
98
- }
99
-
100
- void seek(size_t offset, int whence) const {
101
- static_assert(SEEK_SET == FILE_BEGIN, "SEEK_SET != FILE_BEGIN");
102
- static_assert(SEEK_CUR == FILE_CURRENT, "SEEK_CUR != FILE_CURRENT");
103
- static_assert(SEEK_END == FILE_END, "SEEK_END != FILE_END");
104
-
105
- LARGE_INTEGER li;
106
- li.QuadPart = offset;
107
- BOOL ret = SetFilePointerEx(fp_win32, li, NULL, whence);
108
- if (!ret) {
109
- throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
110
- }
111
- }
112
-
113
- void read_raw(void * ptr, size_t len) {
114
- size_t bytes_read = 0;
115
- while (bytes_read < len) {
116
- size_t chunk_size = std::min<size_t>(len - bytes_read, 64*1024*1024);
117
- DWORD chunk_read = 0;
118
- BOOL result = ReadFile(fp_win32, reinterpret_cast<char*>(ptr) + bytes_read, chunk_size, &chunk_read, NULL);
119
- if (!result) {
120
- throw std::runtime_error(format("read error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
121
- }
122
- if (chunk_read < chunk_size || chunk_read == 0) {
123
- throw std::runtime_error("unexpectedly reached end of file");
124
- }
125
-
126
- bytes_read += chunk_read;
127
- }
128
- }
129
-
130
- uint32_t read_u32() {
131
- uint32_t val;
132
- read_raw(&val, sizeof(val));
133
- return val;
134
- }
135
-
136
- void write_raw(const void * ptr, size_t len) const {
137
- size_t bytes_written = 0;
138
- while (bytes_written < len) {
139
- size_t chunk_size = std::min<size_t>(len - bytes_written, 64*1024*1024);
140
- DWORD chunk_written = 0;
141
- BOOL result = WriteFile(fp_win32, reinterpret_cast<char const*>(ptr) + bytes_written, chunk_size, &chunk_written, NULL);
142
- if (!result) {
143
- throw std::runtime_error(format("write error: %s", GetErrorMessageWin32(GetLastError()).c_str()));
144
- }
145
- if (chunk_written < chunk_size || chunk_written == 0) {
146
- throw std::runtime_error("unexpectedly failed to write bytes");
147
- }
148
-
149
- bytes_written += chunk_written;
150
- }
151
- }
152
-
153
- void write_u32(uint32_t val) const {
154
- write_raw(&val, sizeof(val));
155
- }
156
-
157
- bool has_direct_io() const {
158
- return true;
159
- }
160
-
161
- ~impl() {
162
- if (fp) {
163
- std::fclose(fp);
164
- }
165
- }
166
- #else
167
- impl(const char * fname, const char * mode, [[maybe_unused]] const bool use_direct_io = false) : fname(fname) {
168
- #ifdef __linux__
169
- // Try unbuffered I/O for read only
170
- if (use_direct_io && std::strcmp(mode, "rb") == 0) {
171
- if (init_fd()) {
172
- return;
173
- }
174
- LLAMA_LOG_WARN("Failed to open file '%s' with error: %s. Falling back to buffered I/O",
175
- fname, strerror(errno));
176
- }
177
- #endif
178
- init_fp(mode);
179
- }
180
-
181
- #ifdef __linux__
182
- bool init_fd() {
183
- fd = open(fname.c_str(), O_RDONLY | O_DIRECT);
184
-
185
- if (fd != -1) {
186
- struct stat file_stats{};
187
- fstat(fd, &file_stats);
188
-
189
- size = file_stats.st_size;
190
- alignment = file_stats.st_blksize;
191
-
192
- off_t ret = lseek(fd, 0, SEEK_SET);
193
- if (ret == -1) {
194
- throw std::runtime_error(format("seek error: %s", strerror(errno)));
195
- }
196
- return true;
197
- }
198
- return false;
199
- }
200
- #endif
201
-
202
- void init_fp(const char * mode) {
203
- fp = ggml_fopen(fname.c_str(), mode);
204
- if (fp == NULL) {
205
- throw std::runtime_error(format("failed to open %s: %s", fname.c_str(), strerror(errno)));
206
- }
207
- seek(0, SEEK_END);
208
- size = tell();
209
- seek(0, SEEK_SET);
210
- }
211
-
212
- size_t tell() const {
213
- if (fd == -1) {
214
- long ret = std::ftell(fp);
215
- if (ret == -1) {
216
- throw std::runtime_error(format("ftell error: %s", strerror(errno)));
217
- }
218
-
219
- return (size_t) ret;
220
- }
221
-
222
- off_t pos = lseek(fd, 0, SEEK_CUR);
223
- if (pos == -1) {
224
- throw std::runtime_error(format("lseek error: %s", strerror(errno)));
225
- }
226
- return (size_t) pos;
227
- }
228
-
229
- void seek(size_t offset, int whence) const {
230
- off_t ret = 0;
231
- if (fd == -1) {
232
- ret = std::fseek(fp, (long) offset, whence);
233
- } else {
234
- ret = lseek(fd, offset, whence);
235
- }
236
- if (ret == -1) {
237
- throw std::runtime_error(format("seek error: %s", strerror(errno)));
238
- }
239
- }
240
-
241
- void read_raw_unsafe(void * ptr, size_t len) {
242
- if (len == 0) {
243
- return;
244
- }
245
- errno = 0;
246
- if (fd == -1) {
247
- const size_t curr_off = tell();
248
- const size_t to_read = std::min(len, size - curr_off);
249
-
250
- std::size_t ret = std::fread(ptr, to_read, 1, fp);
251
- if (ferror(fp)) {
252
- throw std::runtime_error(format("read error: %s", strerror(errno)));
253
- }
254
- if (to_read > 0 && ret != 1) {
255
- throw std::runtime_error("unexpectedly reached end of file");
256
- }
257
- } else {
258
- size_t bytes_read = 0;
259
- while (bytes_read < len) {
260
- const size_t to_read = len - bytes_read;
261
- ssize_t ret = ::read(fd, reinterpret_cast<char *>(ptr) + bytes_read, to_read);
262
-
263
- if (ret == -1) {
264
- if (errno == EINTR) {
265
- continue; // Interrupted by signal, retry
266
- }
267
- // Fallback to std::fread in case the DMA controller cannot access the buffer
268
- if (errno == EFAULT || errno == EINVAL) {
269
- LLAMA_LOG_WARN("%s: Falling back to buffered IO due to %s\n", __func__, strerror(errno));
270
- auto curr_off = tell();
271
- close(fd);
272
- fd = -1;
273
- alignment = 1;
274
- init_fp("rb");
275
- seek(curr_off, SEEK_SET);
276
- read_raw_unsafe(ptr, len);
277
- return;
278
- }
279
- throw std::runtime_error(format("read error: %s", strerror(errno)));
280
- }
281
- if (ret == 0) {
282
- // EOF: allow if this read was only pulling alignment padding past file end
283
- off_t pos = lseek(fd, 0, SEEK_CUR);
284
- if (pos != -1 && (size_t) pos == size) {
285
- std::memset(reinterpret_cast<char *>(ptr) + bytes_read, 0, len - bytes_read);
286
- return;
287
- }
288
- throw std::runtime_error("unexpectedly reached end of file");
289
- }
290
-
291
- bytes_read += (size_t) ret;
292
- }
293
- }
294
- }
295
-
296
- void read_aligned_chunk(void * dest, size_t size) {
297
- size_t offset = tell();
298
- off_t aligned_offset = offset & ~(alignment - 1);
299
- off_t offset_from_alignment = offset - aligned_offset;
300
- size_t bytes_to_read = (offset_from_alignment + size + alignment - 1) & ~(alignment - 1);
301
-
302
- void * raw_buffer = nullptr;
303
- int ret = posix_memalign(&raw_buffer, alignment, bytes_to_read);
304
- if (ret != 0) {
305
- throw std::runtime_error(format("posix_memalign failed with error %d", ret));
306
- }
307
-
308
- struct aligned_buffer_deleter {
309
- void operator()(void * p) const { free(p); }
310
- };
311
- std::unique_ptr<void, aligned_buffer_deleter> buffer(raw_buffer);
312
-
313
- seek(aligned_offset, SEEK_SET);
314
- read_raw_unsafe(buffer.get(), bytes_to_read);
315
-
316
- uintptr_t actual_data = reinterpret_cast<uintptr_t>(buffer.get()) + offset_from_alignment;
317
- memcpy(dest, reinterpret_cast<void *>(actual_data), size);
318
- }
319
-
320
- void read_raw(void * ptr, size_t len) {
321
- if (has_direct_io()) {
322
- read_aligned_chunk(ptr, len);
323
- } else {
324
- read_raw_unsafe(ptr, len);
325
- }
326
- }
327
-
328
- uint32_t read_u32() {
329
- uint32_t ret;
330
- read_raw(&ret, sizeof(ret));
331
- return ret;
332
- }
333
-
334
- void write_raw(const void * ptr, size_t len) const {
335
- if (len == 0) {
336
- return;
337
- }
338
- errno = 0;
339
- size_t ret = std::fwrite(ptr, len, 1, fp);
340
- if (ret != 1) {
341
- throw std::runtime_error(format("write error: %s", strerror(errno)));
342
- }
343
- }
344
-
345
- void write_u32(uint32_t val) const {
346
- write_raw(&val, sizeof(val));
347
- }
348
-
349
- bool has_direct_io() const {
350
- return fd != -1 && alignment > 1;
351
- }
352
-
353
- ~impl() {
354
- if (fd != -1) {
355
- close(fd);
356
- } else {
357
- std::fclose(fp);
358
- }
359
- }
360
- int fd = -1;
361
- std::string fname;
362
- #endif
363
-
364
- size_t read_alignment() const {
365
- return alignment;
366
- }
367
-
368
- size_t alignment = 1;
369
-
370
- FILE * fp{};
371
- size_t size{};
372
- };
373
-
374
- llama_file::llama_file(const char * fname, const char * mode, const bool use_direct_io) :
375
- pimpl(std::make_unique<impl>(fname, mode, use_direct_io)) {}
376
- llama_file::~llama_file() = default;
377
-
378
- size_t llama_file::tell() const { return pimpl->tell(); }
379
- size_t llama_file::size() const { return pimpl->size; }
380
-
381
- size_t llama_file::read_alignment() const { return pimpl->read_alignment(); }
382
- bool llama_file::has_direct_io() const { return pimpl->has_direct_io(); }
383
-
384
- int llama_file::file_id() const {
385
- #ifdef _WIN32
386
- return _fileno(pimpl->fp);
387
- #else
388
- if (pimpl->fd != -1) {
389
- return pimpl->fd;
390
- }
391
- #if defined(fileno)
392
- return fileno(pimpl->fp);
393
- #else
394
- return ::fileno(pimpl->fp);
395
- #endif
396
- #endif
397
- }
398
-
399
- void llama_file::seek(size_t offset, int whence) const { pimpl->seek(offset, whence); }
400
- void llama_file::read_raw(void * ptr, size_t len) { pimpl->read_raw(ptr, len); }
401
- #ifdef _WIN32
402
- void llama_file::read_raw_unsafe(void * ptr, size_t len) { pimpl->read_raw(ptr, len); }
403
- #else
404
- void llama_file::read_raw_unsafe(void * ptr, size_t len) { pimpl->read_raw_unsafe(ptr, len); }
405
- #endif
406
-
407
- uint32_t llama_file::read_u32() { return pimpl->read_u32(); }
408
-
409
- void llama_file::write_raw(const void * ptr, size_t len) const { pimpl->write_raw(ptr, len); }
410
- void llama_file::write_u32(uint32_t val) const { pimpl->write_u32(val); }
411
-
412
- // llama_mmap
413
-
414
- struct llama_mmap::impl {
415
- #ifdef _POSIX_MAPPED_FILES
416
- std::vector<std::pair<size_t, size_t>> mapped_fragments;
417
-
418
- impl(struct llama_file * file, size_t prefetch, bool numa) {
419
- size = file->size();
420
- int fd = file->file_id();
421
- int flags = MAP_SHARED;
422
- if (numa) { prefetch = 0; }
423
- #ifdef __linux__
424
- if (posix_fadvise(fd, 0, 0, POSIX_FADV_SEQUENTIAL)) {
425
- LLAMA_LOG_WARN("warning: posix_fadvise(.., POSIX_FADV_SEQUENTIAL) failed: %s\n",
426
- strerror(errno));
427
- }
428
- if (prefetch) { flags |= MAP_POPULATE; }
429
- #endif
430
- addr = mmap(NULL, file->size(), PROT_READ, flags, fd, 0);
431
- if (addr == MAP_FAILED) {
432
- throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
433
- }
434
-
435
- if (prefetch > 0) {
436
- if (posix_madvise(addr, std::min(file->size(), prefetch), POSIX_MADV_WILLNEED)) {
437
- LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n",
438
- strerror(errno));
439
- }
440
- }
441
- if (numa) {
442
- if (posix_madvise(addr, file->size(), POSIX_MADV_RANDOM)) {
443
- LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_RANDOM) failed: %s\n",
444
- strerror(errno));
445
- }
446
- }
447
-
448
- mapped_fragments.emplace_back(0, file->size());
449
- }
450
-
451
- static void align_range(size_t * first, size_t * last, size_t page_size) {
452
- size_t offset_in_page = *first & (page_size - 1);
453
- size_t offset_to_page = offset_in_page == 0 ? 0 : page_size - offset_in_page;
454
- *first += offset_to_page;
455
-
456
- *last = *last & ~(page_size - 1);
457
-
458
- if (*last <= *first) {
459
- *last = *first;
460
- }
461
- }
462
-
463
- void unmap_fragment(size_t first, size_t last) {
464
- int page_size = sysconf(_SC_PAGESIZE);
465
- align_range(&first, &last, page_size);
466
- size_t len = last - first;
467
-
468
- if (len == 0) {
469
- return;
470
- }
471
-
472
- GGML_ASSERT(first % page_size == 0);
473
- GGML_ASSERT(last % page_size == 0);
474
- GGML_ASSERT(last > first);
475
-
476
- void * next_page_start = (uint8_t *) addr + first;
477
-
478
- if (munmap(next_page_start, len)) {
479
- LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno));
480
- }
481
-
482
- std::vector<std::pair<size_t, size_t>> new_mapped_fragments;
483
- for (const auto & frag : mapped_fragments) {
484
- if (frag.first < first && frag.second > last) {
485
- new_mapped_fragments.emplace_back(frag.first, first);
486
- new_mapped_fragments.emplace_back(last, frag.second);
487
- } else if (frag.first < first && frag.second > first) {
488
- new_mapped_fragments.emplace_back(frag.first, first);
489
- } else if (frag.first < last && frag.second > last) {
490
- new_mapped_fragments.emplace_back(last, frag.second);
491
- } else if (frag.first >= first && frag.second <= last) {
492
- } else {
493
- new_mapped_fragments.push_back(frag);
494
- }
495
- }
496
- mapped_fragments = std::move(new_mapped_fragments);
497
- }
498
-
499
- ~impl() {
500
- for (const auto & frag : mapped_fragments) {
501
- if (munmap((char *) addr + frag.first, frag.second - frag.first)) {
502
- LLAMA_LOG_WARN("warning: munmap failed: %s\n", strerror(errno));
503
- }
504
- }
505
- }
506
- #elif defined(_WIN32)
507
- HANDLE hMapping = nullptr;
508
-
509
- impl(struct llama_file * file, size_t prefetch, bool numa) {
510
- GGML_UNUSED(numa);
511
-
512
- size = file->size();
513
-
514
- HANDLE hFile = (HANDLE) _get_osfhandle(file->file_id());
515
-
516
- hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
517
-
518
- if (hMapping == NULL) {
519
- DWORD error = GetLastError();
520
- throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
521
- }
522
-
523
- addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
524
- DWORD error = GetLastError();
525
-
526
- if (addr == NULL) {
527
- CloseHandle(hMapping);
528
- throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
529
- }
530
-
531
- if (prefetch > 0) {
532
- #if _WIN32_WINNT >= 0x602
533
- BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG);
534
- HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
535
-
536
- pPrefetchVirtualMemory = (decltype(pPrefetchVirtualMemory))(void *) GetProcAddress(hKernel32, "PrefetchVirtualMemory");
537
-
538
- if (pPrefetchVirtualMemory) {
539
- WIN32_MEMORY_RANGE_ENTRY range;
540
- range.VirtualAddress = addr;
541
- range.NumberOfBytes = (SIZE_T) std::min(size, prefetch);
542
- if (!pPrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
543
- LLAMA_LOG_WARN("warning: PrefetchVirtualMemory failed: %s\n",
544
- llama_format_win_err(GetLastError()).c_str());
545
- }
546
- }
547
- #else
548
- LLAMA_LOG_DEBUG("skipping PrefetchVirtualMemory because _WIN32_WINNT < 0x602\n");
549
- #endif
550
- }
551
- }
552
-
553
- void unmap_fragment(size_t first, size_t last) {
554
- GGML_UNUSED(first);
555
- GGML_UNUSED(last);
556
- }
557
-
558
- ~impl() {
559
- if (hMapping) {
560
- if (addr) {
561
- if (!UnmapViewOfFile(addr)) {
562
- LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n",
563
- llama_format_win_err(GetLastError()).c_str());
564
- }
565
- }
566
- if (!CloseHandle(hMapping)) {
567
- LLAMA_LOG_WARN("warning: CloseHandle failed: %s\n",
568
- llama_format_win_err(GetLastError()).c_str());
569
- }
570
- }
571
- }
572
- #else
573
- impl(struct llama_file * file, size_t prefetch, bool numa) {
574
- GGML_UNUSED(file);
575
- GGML_UNUSED(prefetch);
576
- GGML_UNUSED(numa);
577
-
578
- throw std::runtime_error("mmap not supported");
579
- }
580
-
581
- void unmap_fragment(size_t first, size_t last) {
582
- GGML_UNUSED(first);
583
- GGML_UNUSED(last);
584
-
585
- throw std::runtime_error("mmap not supported");
586
- }
587
- #endif
588
-
589
- void * addr;
590
- size_t size;
591
- };
592
-
593
- llama_mmap::llama_mmap(struct llama_file * file, size_t prefetch, bool numa) : pimpl(std::make_unique<impl>(file, prefetch, numa)) {}
594
- llama_mmap::~llama_mmap() = default;
595
-
596
- size_t llama_mmap::size() const { return pimpl->size; }
597
- void * llama_mmap::addr() const { return pimpl->addr; }
598
-
599
- void llama_mmap::unmap_fragment(size_t first, size_t last) { pimpl->unmap_fragment(first, last); }
600
-
601
- #if defined(_POSIX_MEMLOCK_RANGE) || defined(_WIN32)
602
- const bool llama_mmap::SUPPORTED = true;
603
- #else
604
- const bool llama_mmap::SUPPORTED = false;
605
- #endif
606
-
607
- // llama_mlock
608
-
609
- struct llama_mlock::impl {
610
- #ifdef _POSIX_MEMLOCK_RANGE
611
- static size_t lock_granularity() {
612
- return (size_t) sysconf(_SC_PAGESIZE);
613
- }
614
-
615
- bool raw_lock(const void * addr, size_t size) const {
616
- if (!mlock(addr, size)) {
617
- return true;
618
- }
619
-
620
- #ifdef __APPLE__
621
- #define MLOCK_SUGGESTION \
622
- "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
623
- "decreasing 'vm.global_no_user_wire_amount'. Also try increasing RLIMIT_MEMLOCK (ulimit -l).\n"
624
- #else
625
- #define MLOCK_SUGGESTION \
626
- "Try increasing RLIMIT_MEMLOCK ('ulimit -l' as root).\n"
627
- #endif
628
-
629
- char* errmsg = std::strerror(errno);
630
- bool suggest = (errno == ENOMEM);
631
- #if defined(TARGET_OS_VISION) || defined(TARGET_OS_TV) || defined(_AIX) || defined(__HAIKU__)
632
- // visionOS/tvOS/Haiku don't support RLIMIT_MEMLOCK
633
- // Skip resource limit checks on these platforms
634
- suggest = false;
635
- #else
636
- struct rlimit lock_limit;
637
- if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit)) {
638
- suggest = false;
639
- }
640
- if (suggest && ((uint64_t)lock_limit.rlim_max > (uint64_t)lock_limit.rlim_cur + size)) {
641
- suggest = false;
642
- }
643
- #endif
644
-
645
- LLAMA_LOG_WARN("warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
646
- size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
647
- return false;
648
- }
649
-
650
- static void raw_unlock(void * addr, size_t size) {
651
- if (munlock(addr, size)) {
652
- LLAMA_LOG_WARN("warning: failed to munlock buffer: %s\n", std::strerror(errno));
653
- }
654
- }
655
- #elif defined(_WIN32)
656
- static size_t lock_granularity() {
657
- SYSTEM_INFO si;
658
- GetSystemInfo(&si);
659
- return (size_t) si.dwPageSize;
660
- }
661
-
662
- bool raw_lock(void * ptr, size_t len) const {
663
- for (int tries = 1; ; tries++) {
664
- if (VirtualLock(ptr, len)) {
665
- return true;
666
- }
667
- if (tries == 2) {
668
- LLAMA_LOG_WARN("warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
669
- len, size, llama_format_win_err(GetLastError()).c_str());
670
- return false;
671
- }
672
-
673
- SIZE_T min_ws_size, max_ws_size;
674
- if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
675
- LLAMA_LOG_WARN("warning: GetProcessWorkingSetSize failed: %s\n",
676
- llama_format_win_err(GetLastError()).c_str());
677
- return false;
678
- }
679
- size_t increment = len + 1048576;
680
- min_ws_size += increment;
681
- max_ws_size += increment;
682
- if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
683
- LLAMA_LOG_WARN("warning: SetProcessWorkingSetSize failed: %s\n",
684
- llama_format_win_err(GetLastError()).c_str());
685
- return false;
686
- }
687
- }
688
- }
689
-
690
- static void raw_unlock(void * ptr, size_t len) {
691
- if (!VirtualUnlock(ptr, len)) {
692
- LLAMA_LOG_WARN("warning: failed to VirtualUnlock buffer: %s\n",
693
- llama_format_win_err(GetLastError()).c_str());
694
- }
695
- }
696
- #else
697
- static size_t lock_granularity() {
698
- return (size_t) 65536;
699
- }
700
-
701
- bool raw_lock(const void * addr, size_t len) const {
702
- LLAMA_LOG_WARN("warning: mlock not supported on this system\n");
703
- return false;
704
- }
705
-
706
- static void raw_unlock(const void * addr, size_t len) {}
707
- #endif
708
-
709
- impl() : addr(NULL), size(0), failed_already(false) {}
710
-
711
- void init(void * ptr) {
712
- GGML_ASSERT(addr == NULL && size == 0);
713
- addr = ptr;
714
- }
715
-
716
- void grow_to(size_t target_size) {
717
- GGML_ASSERT(addr);
718
- if (failed_already) {
719
- return;
720
- }
721
- size_t granularity = lock_granularity();
722
- target_size = (target_size + granularity - 1) & ~(granularity - 1);
723
- if (target_size > size) {
724
- if (raw_lock((uint8_t *) addr + size, target_size - size)) {
725
- size = target_size;
726
- } else {
727
- failed_already = true;
728
- }
729
- }
730
- }
731
-
732
- void * addr;
733
- size_t size;
734
-
735
- bool failed_already;
736
- };
737
-
738
- llama_mlock::llama_mlock() : pimpl(std::make_unique<impl>()) {}
739
- llama_mlock::~llama_mlock() = default;
740
-
741
- void llama_mlock::init(void * ptr) { pimpl->init(ptr); }
742
- void llama_mlock::grow_to(size_t target_size) { pimpl->grow_to(target_size); }
743
-
744
- #if defined(_POSIX_MEMLOCK_RANGE) || defined(_WIN32)
745
- const bool llama_mlock::SUPPORTED = true;
746
- #else
747
- const bool llama_mlock::SUPPORTED = false;
748
- #endif
749
-
750
- size_t llama_path_max() {
751
- return PATH_MAX;
752
- }