whispercpp 1.3.5 → 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 (1017) hide show
  1. checksums.yaml +4 -4
  2. data/.document +3 -0
  3. data/.rdoc_options +2 -0
  4. data/LICENSE +1 -1
  5. data/README.md +133 -3
  6. data/Rakefile +18 -3
  7. data/ext/dependencies.rb +10 -4
  8. data/ext/dependencies_for_windows.rb +17 -0
  9. data/ext/extconf.rb +20 -7
  10. data/ext/options.rb +54 -14
  11. data/ext/options_for_windows.rb +51 -0
  12. data/ext/ruby_whisper.c +56 -46
  13. data/ext/ruby_whisper.h +165 -2
  14. data/ext/ruby_whisper_context.c +297 -126
  15. data/ext/ruby_whisper_context_params.c +163 -0
  16. data/ext/ruby_whisper_log_queue.c +180 -0
  17. data/ext/ruby_whisper_log_settable.h +47 -0
  18. data/ext/ruby_whisper_model.c +0 -1
  19. data/ext/ruby_whisper_parakeet.c +49 -0
  20. data/ext/ruby_whisper_parakeet_context.c +304 -0
  21. data/ext/ruby_whisper_parakeet_context_params.c +117 -0
  22. data/ext/ruby_whisper_parakeet_model.c +84 -0
  23. data/ext/ruby_whisper_parakeet_params.c +548 -0
  24. data/ext/ruby_whisper_parakeet_segment.c +157 -0
  25. data/ext/ruby_whisper_parakeet_token.c +188 -0
  26. data/ext/ruby_whisper_parakeet_transcribe.cpp +58 -0
  27. data/ext/ruby_whisper_params.c +256 -66
  28. data/ext/ruby_whisper_segment.c +6 -7
  29. data/ext/ruby_whisper_token.c +29 -9
  30. data/ext/ruby_whisper_transcribe.cpp +46 -16
  31. data/ext/ruby_whisper_vad_context.c +48 -1
  32. data/ext/ruby_whisper_vad_context_detect.cpp +6 -5
  33. data/ext/ruby_whisper_vad_params.c +0 -1
  34. data/ext/ruby_whisper_vad_segment.c +0 -1
  35. data/ext/ruby_whisper_vad_segments.c +0 -1
  36. data/ext/sources/CMakeLists.txt +41 -3
  37. data/ext/sources/CMakePresets.json +95 -0
  38. data/ext/sources/cmake/parakeet-config.cmake.in +30 -0
  39. data/ext/sources/cmake/parakeet.pc.in +10 -0
  40. data/ext/sources/cmake/whisper-config.cmake.in +5 -40
  41. data/ext/sources/cmake/whisper.pc.in +1 -1
  42. data/ext/sources/examples/CMakeLists.txt +4 -2
  43. data/ext/sources/examples/bench/bench.cpp +24 -19
  44. data/ext/sources/examples/cli/cli.cpp +51 -9
  45. data/ext/sources/examples/common-ggml.cpp +4 -0
  46. data/ext/sources/examples/common-whisper.cpp +139 -67
  47. data/ext/sources/examples/common-whisper.h +11 -0
  48. data/ext/sources/examples/ffmpeg-transcode.cpp +211 -341
  49. data/ext/sources/examples/miniaudio.h +4507 -2131
  50. data/ext/sources/examples/parakeet-cli/CMakeLists.txt +8 -0
  51. data/ext/sources/examples/parakeet-cli/parakeet-cli.cpp +243 -0
  52. data/ext/sources/examples/parakeet-quantize/CMakeLists.txt +7 -0
  53. data/ext/sources/examples/parakeet-quantize/parakeet-quantize.cpp +230 -0
  54. data/ext/sources/examples/server/server.cpp +213 -163
  55. data/ext/sources/ggml/CMakeLists.txt +29 -15
  56. data/ext/sources/ggml/cmake/FindNCCL.cmake +36 -0
  57. data/ext/sources/ggml/cmake/ggml-config.cmake.in +12 -2
  58. data/ext/sources/ggml/include/ggml-alloc.h +1 -0
  59. data/ext/sources/ggml/include/ggml-backend.h +73 -11
  60. data/ext/sources/ggml/include/ggml-cann.h +1 -1
  61. data/ext/sources/ggml/include/ggml-cpu.h +5 -0
  62. data/ext/sources/ggml/include/ggml-cuda.h +3 -0
  63. data/ext/sources/ggml/include/ggml-openvino.h +37 -0
  64. data/ext/sources/ggml/include/ggml-opt.h +1 -1
  65. data/ext/sources/ggml/include/ggml-rpc.h +8 -3
  66. data/ext/sources/ggml/include/ggml-virtgpu.h +14 -0
  67. data/ext/sources/ggml/include/ggml.h +155 -16
  68. data/ext/sources/ggml/include/gguf.h +10 -2
  69. data/ext/sources/ggml/src/CMakeLists.txt +25 -5
  70. data/ext/sources/ggml/src/ggml-alloc.c +9 -10
  71. data/ext/sources/ggml/src/ggml-backend-dl.cpp +48 -0
  72. data/ext/sources/ggml/src/ggml-backend-dl.h +45 -0
  73. data/ext/sources/ggml/src/ggml-backend-impl.h +22 -2
  74. data/ext/sources/ggml/src/ggml-backend-meta.cpp +2263 -0
  75. data/ext/sources/ggml/src/ggml-backend-reg.cpp +40 -86
  76. data/ext/sources/ggml/src/ggml-backend.cpp +114 -10
  77. data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +1 -1
  78. data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +10 -2
  79. data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +1 -1
  80. data/ext/sources/ggml/src/ggml-cann/acl_tensor.h +1 -1
  81. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +1016 -442
  82. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +111 -85
  83. data/ext/sources/ggml/src/ggml-cann/common.h +23 -14
  84. data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +255 -92
  85. data/ext/sources/ggml/src/ggml-common.h +22 -0
  86. data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +68 -34
  87. data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +44 -19
  88. data/ext/sources/ggml/src/ggml-cpu/amx/common.h +34 -10
  89. data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +101 -101
  90. data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +194 -1
  91. data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +2874 -613
  92. data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +151 -1
  93. data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +0 -1
  94. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +5480 -840
  95. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +1361 -0
  96. data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +8 -11
  97. data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +72 -1
  98. data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +186 -36
  99. data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +119 -19
  100. data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +112 -26
  101. data/ext/sources/ggml/src/ggml-cpu/binary-ops.cpp +2 -6
  102. data/ext/sources/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
  103. data/ext/sources/ggml/src/ggml-cpu/common.h +8 -0
  104. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +13 -0
  105. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +153 -16
  106. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +17 -0
  107. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +21 -20
  108. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +976 -251
  109. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +671 -266
  110. data/ext/sources/ggml/src/ggml-cpu/ops.cpp +1277 -263
  111. data/ext/sources/ggml/src/ggml-cpu/ops.h +4 -0
  112. data/ext/sources/ggml/src/ggml-cpu/quants.c +95 -0
  113. data/ext/sources/ggml/src/ggml-cpu/quants.h +6 -0
  114. data/ext/sources/ggml/src/ggml-cpu/repack.cpp +2893 -679
  115. data/ext/sources/ggml/src/ggml-cpu/repack.h +119 -8
  116. data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +226 -0
  117. data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +114 -19
  118. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +1402 -687
  119. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.h +8 -0
  120. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +597 -2766
  121. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp +5768 -0
  122. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.cpp +320 -0
  123. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.h +55 -0
  124. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_kernels.h +182 -19
  125. data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.cpp +1795 -0
  126. data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.h +14 -0
  127. data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp +3178 -0
  128. data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.h +95 -0
  129. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_barrier.h +34 -0
  130. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp +760 -0
  131. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h +32 -0
  132. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_tcm.h +409 -0
  133. data/ext/sources/ggml/src/ggml-cpu/unary-ops.cpp +1 -1
  134. data/ext/sources/ggml/src/ggml-cpu/vec.cpp +54 -53
  135. data/ext/sources/ggml/src/ggml-cpu/vec.h +225 -240
  136. data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +18 -8
  137. data/ext/sources/ggml/src/ggml-cuda/allreduce.cu +971 -0
  138. data/ext/sources/ggml/src/ggml-cuda/allreduce.cuh +29 -0
  139. data/ext/sources/ggml/src/ggml-cuda/argsort.cu +73 -28
  140. data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +69 -41
  141. data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +1 -0
  142. data/ext/sources/ggml/src/ggml-cuda/common.cuh +359 -29
  143. data/ext/sources/ggml/src/ggml-cuda/concat.cu +120 -114
  144. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cu +45 -21
  145. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cuh +1 -0
  146. data/ext/sources/ggml/src/ggml-cuda/convert.cu +94 -27
  147. data/ext/sources/ggml/src/ggml-cuda/convert.cuh +10 -0
  148. data/ext/sources/ggml/src/ggml-cuda/cpy.cu +20 -9
  149. data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +22 -0
  150. data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +333 -85
  151. data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +632 -190
  152. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +12 -0
  153. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +162 -49
  154. data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +43 -18
  155. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +44 -14
  156. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cuh +1 -1
  157. data/ext/sources/ggml/src/ggml-cuda/fattn.cu +241 -23
  158. data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +2 -0
  159. data/ext/sources/ggml/src/ggml-cuda/fwht.cu +101 -0
  160. data/ext/sources/ggml/src/ggml-cuda/fwht.cuh +4 -0
  161. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cu +312 -0
  162. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cuh +4 -0
  163. data/ext/sources/ggml/src/ggml-cuda/getrows.cu +34 -12
  164. data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +1454 -599
  165. data/ext/sources/ggml/src/ggml-cuda/im2col.cu +32 -29
  166. data/ext/sources/ggml/src/ggml-cuda/mean.cu +13 -10
  167. data/ext/sources/ggml/src/ggml-cuda/mma.cuh +397 -183
  168. data/ext/sources/ggml/src/ggml-cuda/mmf.cu +30 -10
  169. data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +161 -88
  170. data/ext/sources/ggml/src/ggml-cuda/mmq.cu +18 -12
  171. data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +522 -431
  172. data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +139 -72
  173. data/ext/sources/ggml/src/ggml-cuda/mmvf.cuh +2 -0
  174. data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +608 -88
  175. data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +6 -0
  176. data/ext/sources/ggml/src/ggml-cuda/norm.cu +47 -79
  177. data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +23 -7
  178. data/ext/sources/ggml/src/ggml-cuda/pad.cu +13 -10
  179. data/ext/sources/ggml/src/ggml-cuda/quantize.cu +134 -27
  180. data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +1 -1
  181. data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +7 -17
  182. data/ext/sources/ggml/src/ggml-cuda/rope.cu +244 -137
  183. data/ext/sources/ggml/src/ggml-cuda/scale.cu +4 -1
  184. data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +14 -6
  185. data/ext/sources/ggml/src/ggml-cuda/snake.cu +72 -0
  186. data/ext/sources/ggml/src/ggml-cuda/snake.cuh +8 -0
  187. data/ext/sources/ggml/src/ggml-cuda/softcap.cu +4 -1
  188. data/ext/sources/ggml/src/ggml-cuda/softmax.cu +8 -83
  189. data/ext/sources/ggml/src/ggml-cuda/solve_tri.cu +1 -1
  190. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +96 -40
  191. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +1 -1
  192. data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +40 -18
  193. data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +8 -4
  194. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +1 -0
  195. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +6 -0
  196. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +2 -0
  197. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +2 -0
  198. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +1 -0
  199. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +6 -0
  200. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +2 -0
  201. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +2 -0
  202. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +1 -0
  203. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +2 -0
  204. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +2 -0
  205. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +2 -0
  206. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +2 -0
  207. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu +5 -0
  208. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu +5 -0
  209. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +5 -0
  210. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu +7 -0
  211. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu +7 -0
  212. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu +7 -0
  213. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu +7 -0
  214. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu +7 -0
  215. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu +7 -0
  216. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu +7 -0
  217. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu +7 -0
  218. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu +7 -0
  219. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu +7 -0
  220. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu +7 -0
  221. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu +7 -0
  222. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu +7 -0
  223. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu +5 -0
  224. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu +5 -0
  225. data/ext/sources/ggml/src/ggml-cuda/top-k.cu +5 -5
  226. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +202 -135
  227. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cuh +20 -14
  228. data/ext/sources/ggml/src/ggml-cuda/unary.cu +86 -2
  229. data/ext/sources/ggml/src/ggml-cuda/unary.cuh +4 -0
  230. data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +111 -17
  231. data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +7 -2
  232. data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +30 -2
  233. data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +3 -0
  234. data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +84 -46
  235. data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +1612 -753
  236. data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +51 -11
  237. data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +361 -261
  238. data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +294 -0
  239. data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +753 -241
  240. data/ext/sources/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +5 -5
  241. data/ext/sources/ggml/src/ggml-hexagon/htp/concat-ops.c +277 -0
  242. data/ext/sources/ggml/src/ggml-hexagon/htp/cpy-ops.c +295 -0
  243. data/ext/sources/ggml/src/ggml-hexagon/htp/cumsum-ops.c +270 -0
  244. data/ext/sources/ggml/src/ggml-hexagon/htp/diag-ops.c +216 -0
  245. data/ext/sources/ggml/src/ggml-hexagon/htp/fill-ops.c +123 -0
  246. data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +471 -296
  247. data/ext/sources/ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c +1148 -0
  248. data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +159 -53
  249. data/ext/sources/ggml/src/ggml-hexagon/htp/{htp-dma.c → hex-dma.c} +3 -3
  250. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.h +372 -0
  251. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +86 -0
  252. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
  253. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +137 -0
  254. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-flash-attn-ops.c +1878 -0
  255. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-matmul-ops.c +2066 -0
  256. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-ops.c +6 -0
  257. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-ops.h +88 -0
  258. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-profile.h +34 -0
  259. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.c +158 -0
  260. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.h +134 -0
  261. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-utils.h +200 -0
  262. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +97 -14
  263. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +163 -67
  264. data/ext/sources/ggml/src/ggml-hexagon/htp/htp_iface.idl +9 -3
  265. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-arith.h +443 -0
  266. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +308 -0
  267. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +262 -0
  268. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +291 -0
  269. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
  270. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +216 -0
  271. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -0
  272. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
  273. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.h +210 -0
  274. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-log.h +65 -0
  275. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-pow.h +42 -0
  276. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-reduce.h +296 -0
  277. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-repl.h +74 -0
  278. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-scale.h +133 -0
  279. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +142 -0
  280. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
  281. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +126 -0
  282. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
  283. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +18 -1348
  284. data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +547 -635
  285. data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +3556 -1101
  286. data/ext/sources/ggml/src/ggml-hexagon/htp/pad-ops.c +547 -0
  287. data/ext/sources/ggml/src/ggml-hexagon/htp/repeat-ops.c +148 -0
  288. data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +475 -269
  289. data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +94 -72
  290. data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +222 -217
  291. data/ext/sources/ggml/src/ggml-hexagon/htp/solve-tri-ops.c +267 -0
  292. data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c +432 -0
  293. data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +128 -0
  294. data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +886 -117
  295. data/ext/sources/ggml/src/ggml-hexagon/htp/vtcm-utils.h +16 -0
  296. data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.c +1 -5
  297. data/ext/sources/ggml/src/ggml-hexagon/htp-drv.cpp +418 -0
  298. data/ext/sources/ggml/src/ggml-hexagon/htp-drv.h +121 -0
  299. data/ext/sources/ggml/src/ggml-hexagon/htp-opnode.h +272 -0
  300. data/ext/sources/ggml/src/ggml-hexagon/libdl.h +79 -0
  301. data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +40 -0
  302. data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +28 -9
  303. data/ext/sources/ggml/src/ggml-impl.h +68 -1
  304. data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +10 -10
  305. data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp +13 -2
  306. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.h +8 -0
  307. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.m +147 -17
  308. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +409 -83
  309. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +54 -5
  310. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +254 -52
  311. data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +254 -23
  312. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +756 -285
  313. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +7 -4
  314. data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +359 -133
  315. data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +1867 -1123
  316. data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +5 -6
  317. data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +71 -4
  318. data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +14127 -5314
  319. data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +97 -88
  320. data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +104 -0
  321. data/ext/sources/ggml/src/ggml-opencl/kernels/cumsum.cl +139 -0
  322. data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +1978 -67
  323. data/ext/sources/ggml/src/ggml-opencl/kernels/diag.cl +27 -0
  324. data/ext/sources/ggml/src/ggml-opencl/kernels/exp.cl +125 -0
  325. data/ext/sources/ggml/src/ggml-opencl/kernels/expm1.cl +87 -56
  326. data/ext/sources/ggml/src/ggml-opencl/kernels/gated_delta_net.cl +249 -0
  327. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl +306 -0
  328. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl +256 -0
  329. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl +258 -0
  330. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl +283 -0
  331. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl +260 -0
  332. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl +262 -0
  333. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl +288 -0
  334. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl +267 -0
  335. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_f32.cl +150 -0
  336. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mat_Ab_Bi_8x4.cl → gemm_noshuffle_q4_0_f32.cl} +1 -1
  337. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_1_f32.cl +132 -0
  338. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl +172 -0
  339. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl +131 -0
  340. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl +134 -0
  341. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl +176 -0
  342. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl +140 -0
  343. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q8_0_f32.cl +129 -0
  344. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_xmem_f16_f32_os8.cl +233 -0
  345. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl +165 -0
  346. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl +120 -0
  347. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl +123 -0
  348. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl +155 -0
  349. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl +123 -0
  350. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl +125 -0
  351. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl +160 -0
  352. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl +141 -0
  353. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +302 -0
  354. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general.cl → gemv_noshuffle_q4_0_f32.cl} +5 -5
  355. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle.cl → gemv_noshuffle_q4_0_f32_spec.cl} +5 -5
  356. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl +283 -0
  357. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +318 -0
  358. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +291 -0
  359. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +294 -0
  360. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +326 -0
  361. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +293 -0
  362. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl +195 -0
  363. data/ext/sources/ggml/src/ggml-opencl/kernels/get_rows.cl +15 -9
  364. data/ext/sources/ggml/src/ggml-opencl/kernels/l2_norm.cl +71 -0
  365. data/ext/sources/ggml/src/ggml-opencl/kernels/mean.cl +114 -13
  366. data/ext/sources/ggml/src/ggml-opencl/kernels/moe_reorder_b.cl +30 -0
  367. data/ext/sources/ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl +82 -0
  368. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_iq4_nl_f32_l4_lm.cl +171 -0
  369. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
  370. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
  371. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl +179 -0
  372. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_0_f32_l4_lm.cl +173 -0
  373. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_1_f32_l4_lm.cl +175 -0
  374. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_k_f32_l4_lm.cl +192 -0
  375. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
  376. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32.cl +164 -0
  377. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32_flat.cl +202 -0
  378. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
  379. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
  380. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +180 -0
  381. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl +196 -0
  382. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32.cl +241 -0
  383. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32_flat.cl +243 -0
  384. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32.cl +243 -0
  385. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32_flat.cl +247 -0
  386. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32.cl +187 -0
  387. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32_flat.cl +203 -0
  388. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mv_q6_k.cl → mul_mv_q6_k_f32.cl} +4 -0
  389. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +178 -0
  390. data/ext/sources/ggml/src/ggml-opencl/kernels/neg.cl +125 -0
  391. data/ext/sources/ggml/src/ggml-opencl/kernels/repeat.cl +31 -32
  392. data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +14 -4
  393. data/ext/sources/ggml/src/ggml-opencl/kernels/softplus.cl +88 -60
  394. data/ext/sources/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
  395. data/ext/sources/ggml/src/ggml-opencl/kernels/sum_rows.cl +114 -13
  396. data/ext/sources/ggml/src/ggml-opencl/kernels/tanh.cl +94 -48
  397. data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +26 -0
  398. data/ext/sources/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
  399. data/ext/sources/ggml/src/ggml-openvino/.clang-format +154 -0
  400. data/ext/sources/ggml/src/ggml-openvino/CMakeLists.txt +22 -0
  401. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +985 -0
  402. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.h +294 -0
  403. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +380 -0
  404. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.h +182 -0
  405. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +1132 -0
  406. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +956 -0
  407. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.h +153 -0
  408. data/ext/sources/ggml/src/ggml-openvino/openvino/decoder.h +74 -0
  409. data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.cpp +27 -0
  410. data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.h +23 -0
  411. data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.cpp +17 -0
  412. data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.h +29 -0
  413. data/ext/sources/ggml/src/ggml-openvino/openvino/node_context.h +112 -0
  414. data/ext/sources/ggml/src/ggml-openvino/openvino/op/cont.cpp +48 -0
  415. data/ext/sources/ggml/src/ggml-openvino/openvino/op/cpy.cpp +21 -0
  416. data/ext/sources/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +90 -0
  417. data/ext/sources/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +69 -0
  418. data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +61 -0
  419. data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +62 -0
  420. data/ext/sources/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +90 -0
  421. data/ext/sources/ggml/src/ggml-openvino/openvino/op/permute.cpp +102 -0
  422. data/ext/sources/ggml/src/ggml-openvino/openvino/op/reshape.cpp +83 -0
  423. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +46 -0
  424. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +149 -0
  425. data/ext/sources/ggml/src/ggml-openvino/openvino/op/scale.cpp +41 -0
  426. data/ext/sources/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +76 -0
  427. data/ext/sources/ggml/src/ggml-openvino/openvino/op/softmax.cpp +89 -0
  428. data/ext/sources/ggml/src/ggml-openvino/openvino/op/transpose.cpp +23 -0
  429. data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_gelu.cpp +25 -0
  430. data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +27 -0
  431. data/ext/sources/ggml/src/ggml-openvino/openvino/op/view.cpp +53 -0
  432. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +47 -0
  433. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +40 -0
  434. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.cpp +60 -0
  435. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.h +17 -0
  436. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +29 -0
  437. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp +58 -0
  438. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.h +17 -0
  439. data/ext/sources/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp +41 -0
  440. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.cpp +317 -0
  441. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.h +28 -0
  442. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +257 -0
  443. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +86 -0
  444. data/ext/sources/ggml/src/ggml-openvino/utils.cpp +880 -0
  445. data/ext/sources/ggml/src/ggml-openvino/utils.h +143 -0
  446. data/ext/sources/ggml/src/ggml-opt.cpp +1 -0
  447. data/ext/sources/ggml/src/ggml-quants.c +385 -119
  448. data/ext/sources/ggml/src/ggml-quants.h +6 -0
  449. data/ext/sources/ggml/src/ggml-rpc/CMakeLists.txt +24 -0
  450. data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +167 -311
  451. data/ext/sources/ggml/src/ggml-rpc/transport.cpp +683 -0
  452. data/ext/sources/ggml/src/ggml-rpc/transport.h +34 -0
  453. data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +64 -91
  454. data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +5 -1
  455. data/ext/sources/ggml/src/ggml-sycl/backend.hpp +4 -1
  456. data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +21 -20
  457. data/ext/sources/ggml/src/ggml-sycl/common.cpp +74 -2
  458. data/ext/sources/ggml/src/ggml-sycl/common.hpp +356 -11
  459. data/ext/sources/ggml/src/ggml-sycl/convert.cpp +184 -14
  460. data/ext/sources/ggml/src/ggml-sycl/convert.hpp +31 -1
  461. data/ext/sources/ggml/src/ggml-sycl/count-equal.cpp +1 -1
  462. data/ext/sources/ggml/src/ggml-sycl/cumsum.cpp +148 -0
  463. data/ext/sources/ggml/src/ggml-sycl/cumsum.hpp +5 -0
  464. data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +663 -0
  465. data/ext/sources/ggml/src/ggml-sycl/diag.cpp +67 -0
  466. data/ext/sources/ggml/src/ggml-sycl/diag.hpp +5 -0
  467. data/ext/sources/ggml/src/ggml-sycl/dmmv.cpp +586 -6
  468. data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +791 -47
  469. data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +77 -156
  470. data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -2
  471. data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.cpp +56 -0
  472. data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.hpp +63 -0
  473. data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp +1181 -0
  474. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +59 -0
  475. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +1246 -0
  476. data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +674 -0
  477. data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +227 -0
  478. data/ext/sources/ggml/src/ggml-sycl/fattn.hpp +22 -0
  479. data/ext/sources/ggml/src/ggml-sycl/fill.cpp +55 -0
  480. data/ext/sources/ggml/src/ggml-sycl/fill.hpp +5 -0
  481. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.cpp +347 -0
  482. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +9 -0
  483. data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +3 -0
  484. data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +79 -3
  485. data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +1134 -236
  486. data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +353 -89
  487. data/ext/sources/ggml/src/ggml-sycl/im2col.hpp +5 -3
  488. data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +1344 -26
  489. data/ext/sources/ggml/src/ggml-sycl/mmvq.hpp +16 -0
  490. data/ext/sources/ggml/src/ggml-sycl/norm.cpp +65 -66
  491. data/ext/sources/ggml/src/ggml-sycl/outprod.cpp +3 -3
  492. data/ext/sources/ggml/src/ggml-sycl/pad.cpp +27 -27
  493. data/ext/sources/ggml/src/ggml-sycl/presets.hpp +3 -0
  494. data/ext/sources/ggml/src/ggml-sycl/quants.hpp +72 -1
  495. data/ext/sources/ggml/src/ggml-sycl/rope.cpp +450 -287
  496. data/ext/sources/ggml/src/ggml-sycl/rope.hpp +6 -0
  497. data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +7 -1
  498. data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +6 -6
  499. data/ext/sources/ggml/src/ggml-sycl/solve_tri.cpp +172 -0
  500. data/ext/sources/ggml/src/ggml-sycl/solve_tri.hpp +8 -0
  501. data/ext/sources/ggml/src/ggml-sycl/ssm_conv.cpp +6 -1
  502. data/ext/sources/ggml/src/ggml-sycl/ssm_scan.cpp +156 -0
  503. data/ext/sources/ggml/src/ggml-sycl/ssm_scan.hpp +5 -0
  504. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +62 -10
  505. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +18 -6
  506. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq112-dv112.cpp +5 -0
  507. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq128-dv128.cpp +5 -0
  508. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq256-dv256.cpp +5 -0
  509. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq40-dv40.cpp +5 -0
  510. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp +6 -0
  511. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq576-dv512.cpp +5 -0
  512. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq64-dv64.cpp +5 -0
  513. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq72-dv72.cpp +5 -0
  514. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq80-dv80.cpp +5 -0
  515. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq96-dv96.cpp +5 -0
  516. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +8 -0
  517. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +8 -0
  518. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +8 -0
  519. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +8 -0
  520. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +8 -0
  521. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +8 -0
  522. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +8 -0
  523. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +8 -0
  524. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +8 -0
  525. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +8 -0
  526. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +8 -0
  527. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +8 -0
  528. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +8 -0
  529. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +8 -0
  530. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +8 -0
  531. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +8 -0
  532. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +8 -0
  533. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +8 -0
  534. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +8 -0
  535. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +8 -0
  536. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +8 -0
  537. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +8 -0
  538. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +8 -0
  539. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +8 -0
  540. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +8 -0
  541. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +8 -0
  542. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +8 -0
  543. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +8 -0
  544. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +8 -0
  545. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +8 -0
  546. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +8 -0
  547. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +8 -0
  548. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +8 -0
  549. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +8 -0
  550. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +8 -0
  551. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +8 -0
  552. data/ext/sources/ggml/src/ggml-sycl/type.hpp +112 -0
  553. data/ext/sources/ggml/src/ggml-sycl/upscale.cpp +410 -0
  554. data/ext/sources/ggml/src/ggml-sycl/upscale.hpp +9 -0
  555. data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +228 -53
  556. data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +1 -1
  557. data/ext/sources/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
  558. data/ext/sources/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
  559. data/ext/sources/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
  560. data/ext/sources/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
  561. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
  562. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
  563. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
  564. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
  565. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +148 -0
  566. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
  567. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
  568. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
  569. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
  570. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
  571. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
  572. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
  573. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
  574. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
  575. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
  576. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
  577. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
  578. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +123 -0
  579. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +160 -0
  580. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
  581. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +71 -0
  582. data/ext/sources/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
  583. data/ext/sources/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
  584. data/ext/sources/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
  585. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
  586. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
  587. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
  588. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
  589. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +192 -0
  590. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
  591. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +53 -0
  592. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +99 -0
  593. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
  594. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
  595. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
  596. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +545 -0
  597. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +115 -0
  598. data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +12 -1
  599. data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +3250 -940
  600. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +4 -0
  601. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +16 -8
  602. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +6 -2
  603. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +146 -13
  604. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +3 -1
  605. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +1 -1
  606. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +25 -1
  607. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +88 -0
  608. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl +643 -1
  609. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp +32 -0
  610. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q1_0.comp +29 -0
  611. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +0 -1
  612. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl +27 -0
  613. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/elu.comp +27 -0
  614. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +0 -1
  615. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2_decode_vector.comp +7 -0
  616. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +533 -180
  617. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +113 -68
  618. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +412 -222
  619. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +222 -83
  620. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_dequant.glsl +131 -0
  621. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mask_opt.comp +162 -0
  622. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mmq_funcs.glsl +203 -0
  623. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +9 -8
  624. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp +115 -0
  625. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +189 -0
  626. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl +0 -1
  627. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl +10 -1
  628. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl +16 -6
  629. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +76 -54
  630. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +0 -1
  631. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +12 -9
  632. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +0 -1
  633. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +122 -27
  634. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl +20 -17
  635. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +6 -6
  636. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +1 -1
  637. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +1 -1
  638. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +1 -1
  639. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +1 -0
  640. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +88 -55
  641. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +22 -20
  642. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +51 -14
  643. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +159 -125
  644. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +3 -1
  645. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +5 -3
  646. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +8 -8
  647. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +24 -9
  648. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +0 -1
  649. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +2 -3
  650. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +39 -63
  651. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +0 -1
  652. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +7 -4
  653. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +7 -4
  654. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +7 -4
  655. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +13 -7
  656. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +7 -4
  657. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sgn.comp +21 -0
  658. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/snake.comp +49 -0
  659. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +27 -11
  660. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +0 -1
  661. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +79 -2
  662. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +193 -149
  663. data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +5 -2
  664. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +3221 -97
  665. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +3493 -1997
  666. data/ext/sources/ggml/src/ggml-webgpu/pre_wgsl.hpp +37 -7
  667. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_id.wgsl +64 -0
  668. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
  669. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +106 -0
  670. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
  671. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +142 -0
  672. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +115 -141
  673. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +93 -0
  674. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/conv2d.wgsl +165 -0
  675. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{cpy.tmpl.wgsl → cpy.wgsl} +25 -44
  676. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
  677. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +198 -230
  678. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_quant_staging.tmpl +124 -0
  679. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl +397 -0
  680. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_blk.wgsl +101 -0
  681. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_reduce.wgsl +84 -0
  682. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl +619 -0
  683. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl +149 -0
  684. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{get_rows.tmpl.wgsl → get_rows.wgsl} +234 -335
  685. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl +155 -0
  686. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/im2col.wgsl +101 -0
  687. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +871 -42
  688. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id.wgsl +195 -0
  689. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_gather.wgsl +52 -0
  690. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl +154 -0
  691. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +149 -0
  692. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat_subgroup_matrix.tmpl.wgsl → mul_mat_subgroup_matrix.wgsl} +36 -138
  693. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +151 -0
  694. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl +1432 -0
  695. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl +303 -0
  696. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
  697. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl +21 -0
  698. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl +173 -0
  699. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl +67 -0
  700. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl +152 -0
  701. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{rope.tmpl.wgsl → rope.wgsl} +71 -142
  702. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/row_norm.wgsl +153 -0
  703. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{scale.tmpl.wgsl → scale.wgsl} +15 -40
  704. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set.wgsl +109 -0
  705. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +39 -12
  706. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows_quant.wgsl +224 -0
  707. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{soft_max.tmpl.wgsl → soft_max.wgsl} +106 -206
  708. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/solve_tri.wgsl +121 -0
  709. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_conv.wgsl +65 -0
  710. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl +193 -0
  711. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
  712. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +213 -0
  713. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
  714. data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +24 -15
  715. data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +31 -32
  716. data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +253 -16
  717. data/ext/sources/ggml/src/ggml.c +268 -52
  718. data/ext/sources/ggml/src/gguf.cpp +377 -47
  719. data/ext/sources/include/parakeet.h +342 -0
  720. data/ext/sources/include/whisper.h +10 -0
  721. data/ext/sources/media/matmul.png +0 -0
  722. data/ext/sources/src/CMakeLists.txt +23 -0
  723. data/ext/sources/src/parakeet-arch.h +188 -0
  724. data/ext/sources/src/parakeet.cpp +3838 -0
  725. data/ext/sources/src/whisper.cpp +62 -40
  726. data/extsources.rb +26 -10
  727. data/lib/whisper/log_settable.rb +36 -0
  728. data/lib/whisper/model/uri.rb +13 -1
  729. data/lib/whisper/output.rb +74 -0
  730. data/sig/whisper.rbs +445 -55
  731. data/test/helper.rb +2 -0
  732. data/test/jfk_reader/jfk_reader.c +50 -7
  733. data/test/test_callback.rb +1 -0
  734. data/test/test_context_params.rb +82 -0
  735. data/test/test_package.rb +6 -5
  736. data/test/test_parakeet.rb +28 -0
  737. data/test/test_parakeet_callback.rb +107 -0
  738. data/test/test_parakeet_context.rb +116 -0
  739. data/test/test_parakeet_context_params.rb +24 -0
  740. data/test/test_parakeet_model.rb +21 -0
  741. data/test/test_parakeet_params.rb +78 -0
  742. data/test/test_parakeet_segment.rb +42 -0
  743. data/test/test_parakeet_token.rb +73 -0
  744. data/test/test_params.rb +2 -0
  745. data/test/test_token.rb +11 -0
  746. data/test/test_vad_context.rb +58 -8
  747. data/test/test_vad_segment.rb +1 -1
  748. data/test/test_whisper.rb +44 -6
  749. data/whispercpp.gemspec +2 -2
  750. metadata +426 -280
  751. data/ext/sources/bindings/javascript/CMakeLists.txt +0 -41
  752. data/ext/sources/bindings/javascript/emscripten.cpp +0 -93
  753. data/ext/sources/bindings/javascript/libwhisper.worker.js +0 -1
  754. data/ext/sources/bindings/javascript/package.json +0 -26
  755. data/ext/sources/bindings/javascript/whisper.js +0 -19
  756. data/ext/sources/examples/addon.node/CMakeLists.txt +0 -31
  757. data/ext/sources/examples/addon.node/__test__/whisper.spec.js +0 -133
  758. data/ext/sources/examples/addon.node/addon.cpp +0 -557
  759. data/ext/sources/examples/addon.node/index.js +0 -59
  760. data/ext/sources/examples/addon.node/package.json +0 -16
  761. data/ext/sources/examples/addon.node/vad-example.js +0 -132
  762. data/ext/sources/examples/bench.wasm/CMakeLists.txt +0 -49
  763. data/ext/sources/examples/bench.wasm/emscripten.cpp +0 -87
  764. data/ext/sources/examples/bench.wasm/index-tmpl.html +0 -285
  765. data/ext/sources/examples/coi-serviceworker.js +0 -146
  766. data/ext/sources/examples/command/CMakeLists.txt +0 -10
  767. data/ext/sources/examples/command/command.cpp +0 -802
  768. data/ext/sources/examples/command/commands.txt +0 -9
  769. data/ext/sources/examples/command.wasm/CMakeLists.txt +0 -50
  770. data/ext/sources/examples/command.wasm/emscripten.cpp +0 -327
  771. data/ext/sources/examples/command.wasm/index-tmpl.html +0 -415
  772. data/ext/sources/examples/generate-karaoke.sh +0 -57
  773. data/ext/sources/examples/helpers.js +0 -191
  774. data/ext/sources/examples/livestream.sh +0 -112
  775. data/ext/sources/examples/lsp/CMakeLists.txt +0 -10
  776. data/ext/sources/examples/lsp/lsp.cpp +0 -471
  777. data/ext/sources/examples/lsp/whisper.vim +0 -362
  778. data/ext/sources/examples/python/test_whisper_processor.py +0 -7
  779. data/ext/sources/examples/python/whisper_processor.py +0 -54
  780. data/ext/sources/examples/server/bench.js +0 -29
  781. data/ext/sources/examples/server.py +0 -120
  782. data/ext/sources/examples/stream/CMakeLists.txt +0 -10
  783. data/ext/sources/examples/stream/stream.cpp +0 -437
  784. data/ext/sources/examples/stream.wasm/CMakeLists.txt +0 -49
  785. data/ext/sources/examples/stream.wasm/emscripten.cpp +0 -216
  786. data/ext/sources/examples/stream.wasm/index-tmpl.html +0 -491
  787. data/ext/sources/examples/sycl/CMakeLists.txt +0 -9
  788. data/ext/sources/examples/sycl/build.sh +0 -22
  789. data/ext/sources/examples/sycl/ls-sycl-device.cpp +0 -11
  790. data/ext/sources/examples/sycl/run-whisper.sh +0 -17
  791. data/ext/sources/examples/talk-llama/CMakeLists.txt +0 -47
  792. data/ext/sources/examples/talk-llama/eleven-labs.py +0 -80
  793. data/ext/sources/examples/talk-llama/llama-adapter.cpp +0 -494
  794. data/ext/sources/examples/talk-llama/llama-adapter.h +0 -88
  795. data/ext/sources/examples/talk-llama/llama-arch.cpp +0 -2559
  796. data/ext/sources/examples/talk-llama/llama-arch.h +0 -586
  797. data/ext/sources/examples/talk-llama/llama-batch.cpp +0 -917
  798. data/ext/sources/examples/talk-llama/llama-batch.h +0 -173
  799. data/ext/sources/examples/talk-llama/llama-chat.cpp +0 -876
  800. data/ext/sources/examples/talk-llama/llama-chat.h +0 -70
  801. data/ext/sources/examples/talk-llama/llama-context.cpp +0 -3645
  802. data/ext/sources/examples/talk-llama/llama-context.h +0 -360
  803. data/ext/sources/examples/talk-llama/llama-cparams.cpp +0 -5
  804. data/ext/sources/examples/talk-llama/llama-cparams.h +0 -42
  805. data/ext/sources/examples/talk-llama/llama-grammar.cpp +0 -1464
  806. data/ext/sources/examples/talk-llama/llama-grammar.h +0 -194
  807. data/ext/sources/examples/talk-llama/llama-graph.cpp +0 -2282
  808. data/ext/sources/examples/talk-llama/llama-graph.h +0 -910
  809. data/ext/sources/examples/talk-llama/llama-hparams.cpp +0 -241
  810. data/ext/sources/examples/talk-llama/llama-hparams.h +0 -284
  811. data/ext/sources/examples/talk-llama/llama-impl.cpp +0 -171
  812. data/ext/sources/examples/talk-llama/llama-impl.h +0 -63
  813. data/ext/sources/examples/talk-llama/llama-io.cpp +0 -15
  814. data/ext/sources/examples/talk-llama/llama-io.h +0 -35
  815. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +0 -328
  816. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.h +0 -137
  817. data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +0 -2100
  818. data/ext/sources/examples/talk-llama/llama-kv-cache.h +0 -390
  819. data/ext/sources/examples/talk-llama/llama-kv-cells.h +0 -533
  820. data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +0 -268
  821. data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +0 -139
  822. data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +0 -1167
  823. data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +0 -182
  824. data/ext/sources/examples/talk-llama/llama-memory.cpp +0 -59
  825. data/ext/sources/examples/talk-llama/llama-memory.h +0 -122
  826. data/ext/sources/examples/talk-llama/llama-mmap.cpp +0 -735
  827. data/ext/sources/examples/talk-llama/llama-mmap.h +0 -73
  828. data/ext/sources/examples/talk-llama/llama-model-loader.cpp +0 -1247
  829. data/ext/sources/examples/talk-llama/llama-model-loader.h +0 -176
  830. data/ext/sources/examples/talk-llama/llama-model-saver.cpp +0 -285
  831. data/ext/sources/examples/talk-llama/llama-model-saver.h +0 -37
  832. data/ext/sources/examples/talk-llama/llama-model.cpp +0 -8338
  833. data/ext/sources/examples/talk-llama/llama-model.h +0 -544
  834. data/ext/sources/examples/talk-llama/llama-quant.cpp +0 -1072
  835. data/ext/sources/examples/talk-llama/llama-quant.h +0 -1
  836. data/ext/sources/examples/talk-llama/llama-sampling.cpp +0 -3771
  837. data/ext/sources/examples/talk-llama/llama-sampling.h +0 -44
  838. data/ext/sources/examples/talk-llama/llama-vocab.cpp +0 -3900
  839. data/ext/sources/examples/talk-llama/llama-vocab.h +0 -182
  840. data/ext/sources/examples/talk-llama/llama.cpp +0 -1140
  841. data/ext/sources/examples/talk-llama/llama.h +0 -1540
  842. data/ext/sources/examples/talk-llama/models/afmoe.cpp +0 -191
  843. data/ext/sources/examples/talk-llama/models/apertus.cpp +0 -125
  844. data/ext/sources/examples/talk-llama/models/arcee.cpp +0 -135
  845. data/ext/sources/examples/talk-llama/models/arctic.cpp +0 -138
  846. data/ext/sources/examples/talk-llama/models/arwkv7.cpp +0 -86
  847. data/ext/sources/examples/talk-llama/models/baichuan.cpp +0 -122
  848. data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +0 -144
  849. data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +0 -135
  850. data/ext/sources/examples/talk-llama/models/bert.cpp +0 -178
  851. data/ext/sources/examples/talk-llama/models/bitnet.cpp +0 -160
  852. data/ext/sources/examples/talk-llama/models/bloom.cpp +0 -101
  853. data/ext/sources/examples/talk-llama/models/chameleon.cpp +0 -178
  854. data/ext/sources/examples/talk-llama/models/chatglm.cpp +0 -132
  855. data/ext/sources/examples/talk-llama/models/codeshell.cpp +0 -111
  856. data/ext/sources/examples/talk-llama/models/cogvlm.cpp +0 -102
  857. data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +0 -134
  858. data/ext/sources/examples/talk-llama/models/command-r.cpp +0 -122
  859. data/ext/sources/examples/talk-llama/models/dbrx.cpp +0 -123
  860. data/ext/sources/examples/talk-llama/models/deci.cpp +0 -135
  861. data/ext/sources/examples/talk-llama/models/deepseek.cpp +0 -144
  862. data/ext/sources/examples/talk-llama/models/deepseek2.cpp +0 -259
  863. data/ext/sources/examples/talk-llama/models/dots1.cpp +0 -134
  864. data/ext/sources/examples/talk-llama/models/dream.cpp +0 -105
  865. data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +0 -150
  866. data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +0 -110
  867. data/ext/sources/examples/talk-llama/models/exaone.cpp +0 -114
  868. data/ext/sources/examples/talk-llama/models/exaone4.cpp +0 -123
  869. data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +0 -113
  870. data/ext/sources/examples/talk-llama/models/falcon.cpp +0 -120
  871. data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +0 -116
  872. data/ext/sources/examples/talk-llama/models/gemma.cpp +0 -112
  873. data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +0 -128
  874. data/ext/sources/examples/talk-llama/models/gemma3.cpp +0 -155
  875. data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +0 -384
  876. data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +0 -170
  877. data/ext/sources/examples/talk-llama/models/glm4.cpp +0 -150
  878. data/ext/sources/examples/talk-llama/models/gpt2.cpp +0 -105
  879. data/ext/sources/examples/talk-llama/models/gptneox.cpp +0 -144
  880. data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +0 -196
  881. data/ext/sources/examples/talk-llama/models/granite.cpp +0 -211
  882. data/ext/sources/examples/talk-llama/models/graph-context-mamba.cpp +0 -283
  883. data/ext/sources/examples/talk-llama/models/grok.cpp +0 -159
  884. data/ext/sources/examples/talk-llama/models/grovemoe.cpp +0 -141
  885. data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +0 -132
  886. data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +0 -154
  887. data/ext/sources/examples/talk-llama/models/internlm2.cpp +0 -120
  888. data/ext/sources/examples/talk-llama/models/jais.cpp +0 -86
  889. data/ext/sources/examples/talk-llama/models/jamba.cpp +0 -106
  890. data/ext/sources/examples/talk-llama/models/lfm2.cpp +0 -175
  891. data/ext/sources/examples/talk-llama/models/llada-moe.cpp +0 -122
  892. data/ext/sources/examples/talk-llama/models/llada.cpp +0 -99
  893. data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +0 -178
  894. data/ext/sources/examples/talk-llama/models/llama.cpp +0 -168
  895. data/ext/sources/examples/talk-llama/models/maincoder.cpp +0 -117
  896. data/ext/sources/examples/talk-llama/models/mamba.cpp +0 -55
  897. data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +0 -123
  898. data/ext/sources/examples/talk-llama/models/minicpm3.cpp +0 -199
  899. data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +0 -124
  900. data/ext/sources/examples/talk-llama/models/mistral3.cpp +0 -160
  901. data/ext/sources/examples/talk-llama/models/models.h +0 -569
  902. data/ext/sources/examples/talk-llama/models/modern-bert.cpp +0 -116
  903. data/ext/sources/examples/talk-llama/models/mpt.cpp +0 -126
  904. data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +0 -150
  905. data/ext/sources/examples/talk-llama/models/nemotron.cpp +0 -122
  906. data/ext/sources/examples/talk-llama/models/neo-bert.cpp +0 -104
  907. data/ext/sources/examples/talk-llama/models/olmo.cpp +0 -121
  908. data/ext/sources/examples/talk-llama/models/olmo2.cpp +0 -150
  909. data/ext/sources/examples/talk-llama/models/olmoe.cpp +0 -124
  910. data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +0 -127
  911. data/ext/sources/examples/talk-llama/models/openelm.cpp +0 -124
  912. data/ext/sources/examples/talk-llama/models/orion.cpp +0 -123
  913. data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +0 -121
  914. data/ext/sources/examples/talk-llama/models/phi2.cpp +0 -121
  915. data/ext/sources/examples/talk-llama/models/phi3.cpp +0 -152
  916. data/ext/sources/examples/talk-llama/models/plamo.cpp +0 -110
  917. data/ext/sources/examples/talk-llama/models/plamo2.cpp +0 -316
  918. data/ext/sources/examples/talk-llama/models/plamo3.cpp +0 -128
  919. data/ext/sources/examples/talk-llama/models/plm.cpp +0 -168
  920. data/ext/sources/examples/talk-llama/models/qwen.cpp +0 -108
  921. data/ext/sources/examples/talk-llama/models/qwen2.cpp +0 -126
  922. data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +0 -151
  923. data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +0 -117
  924. data/ext/sources/examples/talk-llama/models/qwen3.cpp +0 -117
  925. data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +0 -124
  926. data/ext/sources/examples/talk-llama/models/qwen3next.cpp +0 -873
  927. data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +0 -149
  928. data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +0 -141
  929. data/ext/sources/examples/talk-llama/models/refact.cpp +0 -94
  930. data/ext/sources/examples/talk-llama/models/rnd1.cpp +0 -126
  931. data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +0 -162
  932. data/ext/sources/examples/talk-llama/models/rwkv6.cpp +0 -94
  933. data/ext/sources/examples/talk-llama/models/rwkv6qwen2.cpp +0 -86
  934. data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +0 -135
  935. data/ext/sources/examples/talk-llama/models/rwkv7.cpp +0 -90
  936. data/ext/sources/examples/talk-llama/models/seed-oss.cpp +0 -124
  937. data/ext/sources/examples/talk-llama/models/smallthinker.cpp +0 -126
  938. data/ext/sources/examples/talk-llama/models/smollm3.cpp +0 -128
  939. data/ext/sources/examples/talk-llama/models/stablelm.cpp +0 -146
  940. data/ext/sources/examples/talk-llama/models/starcoder.cpp +0 -100
  941. data/ext/sources/examples/talk-llama/models/starcoder2.cpp +0 -121
  942. data/ext/sources/examples/talk-llama/models/t5-dec.cpp +0 -166
  943. data/ext/sources/examples/talk-llama/models/t5-enc.cpp +0 -96
  944. data/ext/sources/examples/talk-llama/models/wavtokenizer-dec.cpp +0 -149
  945. data/ext/sources/examples/talk-llama/models/xverse.cpp +0 -108
  946. data/ext/sources/examples/talk-llama/prompts/talk-alpaca.txt +0 -23
  947. data/ext/sources/examples/talk-llama/speak +0 -40
  948. data/ext/sources/examples/talk-llama/speak.bat +0 -1
  949. data/ext/sources/examples/talk-llama/speak.ps1 +0 -14
  950. data/ext/sources/examples/talk-llama/talk-llama.cpp +0 -813
  951. data/ext/sources/examples/talk-llama/unicode-data.cpp +0 -7034
  952. data/ext/sources/examples/talk-llama/unicode-data.h +0 -20
  953. data/ext/sources/examples/talk-llama/unicode.cpp +0 -1147
  954. data/ext/sources/examples/talk-llama/unicode.h +0 -111
  955. data/ext/sources/examples/wchess/CMakeLists.txt +0 -10
  956. data/ext/sources/examples/wchess/libwchess/CMakeLists.txt +0 -19
  957. data/ext/sources/examples/wchess/libwchess/Chessboard.cpp +0 -803
  958. data/ext/sources/examples/wchess/libwchess/Chessboard.h +0 -33
  959. data/ext/sources/examples/wchess/libwchess/WChess.cpp +0 -193
  960. data/ext/sources/examples/wchess/libwchess/WChess.h +0 -63
  961. data/ext/sources/examples/wchess/libwchess/test-chessboard.cpp +0 -117
  962. data/ext/sources/examples/wchess/wchess.cmd/CMakeLists.txt +0 -8
  963. data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +0 -253
  964. data/ext/sources/examples/whisper.wasm/CMakeLists.txt +0 -50
  965. data/ext/sources/examples/whisper.wasm/emscripten.cpp +0 -118
  966. data/ext/sources/examples/whisper.wasm/index-tmpl.html +0 -659
  967. data/ext/sources/ggml/cmake/BuildTypes.cmake +0 -54
  968. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm-ppc.h +0 -333
  969. data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +0 -99
  970. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-dma.h +0 -157
  971. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +0 -165
  972. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.c +0 -94
  973. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.c +0 -72
  974. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.c +0 -49
  975. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.c +0 -1020
  976. data/ext/sources/ggml/src/ggml-hexagon/htp/ops-utils.h +0 -149
  977. data/ext/sources/ggml/src/ggml-hexagon/htp-utils.c +0 -454
  978. data/ext/sources/ggml/src/ggml-hexagon/htp-utils.h +0 -221
  979. data/ext/sources/ggml/src/ggml-hexagon/op-desc.h +0 -153
  980. data/ext/sources/ggml/src/ggml-opencl/kernels/embed_kernel.py +0 -26
  981. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.glsl +0 -5
  982. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/bin_op.tmpl.wgsl +0 -188
  983. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary_head.tmpl +0 -45
  984. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +0 -147
  985. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.tmpl.wgsl +0 -323
  986. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl +0 -907
  987. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.tmpl.wgsl +0 -247
  988. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.tmpl.wgsl +0 -267
  989. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +0 -123
  990. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.tmpl.wgsl +0 -112
  991. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary_op.wgsl +0 -483
  992. data/ext/sources/tests/CMakeLists.txt +0 -112
  993. data/ext/sources/tests/earnings21/eval.mk +0 -58
  994. data/ext/sources/tests/earnings21/eval.py +0 -68
  995. data/ext/sources/tests/earnings21/normalizers/__init__.py +0 -2
  996. data/ext/sources/tests/earnings21/normalizers/basic.py +0 -80
  997. data/ext/sources/tests/earnings21/normalizers/english.json +0 -1741
  998. data/ext/sources/tests/earnings21/normalizers/english.py +0 -550
  999. data/ext/sources/tests/earnings21/requirements.txt +0 -6
  1000. data/ext/sources/tests/en-0-ref.txt +0 -1
  1001. data/ext/sources/tests/en-1-ref.txt +0 -1
  1002. data/ext/sources/tests/en-2-ref.txt +0 -1
  1003. data/ext/sources/tests/es-0-ref.txt +0 -1
  1004. data/ext/sources/tests/librispeech/eval.mk +0 -39
  1005. data/ext/sources/tests/librispeech/eval.py +0 -47
  1006. data/ext/sources/tests/librispeech/normalizers/__init__.py +0 -2
  1007. data/ext/sources/tests/librispeech/normalizers/basic.py +0 -80
  1008. data/ext/sources/tests/librispeech/normalizers/english.json +0 -1741
  1009. data/ext/sources/tests/librispeech/normalizers/english.py +0 -550
  1010. data/ext/sources/tests/librispeech/requirements.txt +0 -6
  1011. data/ext/sources/tests/run-tests.sh +0 -130
  1012. data/ext/sources/tests/test-c.c +0 -3
  1013. data/ext/sources/tests/test-vad-full.cpp +0 -56
  1014. data/ext/sources/tests/test-vad.cpp +0 -83
  1015. data/ext/sources/tests/test-whisper.js +0 -58
  1016. data/lib/whisper/context.rb +0 -15
  1017. data/lib/whisper/segment.rb +0 -58
@@ -1,240 +1,70 @@
1
- #define(VARIANTS)
2
-
3
- [
4
- {
5
- "SHADER_SUFFIX": "f32_vec",
6
- "REPLS": {
7
- "TYPE" : "vec4<f32>",
8
- "DST_TYPE": "vec4<f32>",
9
- "BLOCK_SIZE": 4
10
- },
11
- "DECLS": ["F32_VEC"]
12
- },
13
- {
14
- "REPLS": {
15
- "TYPE" : "f32",
16
- "DST_TYPE": "f32",
17
- "BLOCK_SIZE": 1
18
- },
19
- "DECLS": ["F32"]
20
- },
21
- {
22
- "REPLS": {
23
- "TYPE" : "f16",
24
- "DST_TYPE": "f32",
25
- "BLOCK_SIZE": 1
26
- },
27
- "DECLS": ["F16"]
28
- },
29
- {
30
- "REPLS": {
31
- "TYPE" : "i32",
32
- "DST_TYPE": "i32",
33
- "BLOCK_SIZE": 1
34
- },
35
- "DECLS": ["I32"]
36
- },
37
- {
38
- "REPLS": {
39
- "TYPE" : "q4_0",
40
- "DST_TYPE": "f32",
41
- "BLOCK_SIZE": 32
42
- },
43
- "DECLS": ["BYTE_HELPERS", "Q4_0_T", "Q4_0"]
44
- },
45
- {
46
- "REPLS": {
47
- "TYPE" : "q4_1",
48
- "DST_TYPE": "f32",
49
- "BLOCK_SIZE": 32
50
- },
51
- "DECLS": ["BYTE_HELPERS", "Q4_1_T", "Q4_1"]
52
- },
53
- {
54
- "REPLS": {
55
- "TYPE" : "q5_0",
56
- "DST_TYPE": "f32",
57
- "BLOCK_SIZE": 32
58
- },
59
- "DECLS": ["BYTE_HELPERS", "Q5_0_T", "Q5_0"]
60
- },
61
- {
62
- "REPLS": {
63
- "TYPE" : "q5_1",
64
- "DST_TYPE": "f32",
65
- "BLOCK_SIZE": 32
66
- },
67
- "DECLS": ["BYTE_HELPERS", "Q5_1_T", "Q5_1"]
68
- },
69
- {
70
- "REPLS": {
71
- "TYPE" : "q8_0",
72
- "DST_TYPE": "f32",
73
- "BLOCK_SIZE": 32
74
- },
75
- "DECLS": ["BYTE_HELPERS", "Q8_0_T", "Q8_0"]
76
- },
77
- {
78
- "REPLS": {
79
- "TYPE" : "q2_k",
80
- "DST_TYPE": "f32",
81
- "BLOCK_SIZE": 256
82
- },
83
- "DECLS": ["BYTE_HELPERS", "Q2_K_T", "Q2_K"]
84
- },
85
- {
86
- "REPLS": {
87
- "TYPE" : "q3_k",
88
- "DST_TYPE": "f32",
89
- "BLOCK_SIZE": 256
90
- },
91
- "DECLS": ["BYTE_HELPERS", "Q3_K_T", "Q3_K"]
92
- },
93
- {
94
- "REPLS": {
95
- "TYPE" : "q4_k",
96
- "DST_TYPE": "f32",
97
- "BLOCK_SIZE": 256
98
- },
99
- "DECLS": ["Q45_K_SCALE_MIN", "BYTE_HELPERS", "Q4_K_T", "Q4_K"]
100
- },
101
- {
102
- "REPLS": {
103
- "TYPE" : "q5_k",
104
- "DST_TYPE": "f32",
105
- "BLOCK_SIZE": 256
106
- },
107
- "DECLS": ["Q45_K_SCALE_MIN", "BYTE_HELPERS", "Q5_K_T", "Q5_K"]
108
- },
109
- {
110
- "REPLS": {
111
- "TYPE" : "q6_k",
112
- "DST_TYPE": "f32",
113
- "BLOCK_SIZE": 256
114
- },
115
- "DECLS": ["BYTE_HELPERS", "Q6_K_T", "Q6_K"]
116
- },
117
- {
118
- "REPLS": {
119
- "TYPE" : "iq2_xxs",
120
- "DST_TYPE": "f32",
121
- "BLOCK_SIZE": 256
122
- },
123
- "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_XXS_GRID", "IQ2_XXS_T", "IQ2_XXS"]
124
- },
125
- {
126
- "REPLS": {
127
- "TYPE" : "iq2_xs",
128
- "DST_TYPE": "f32",
129
- "BLOCK_SIZE": 256
130
- },
131
- "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_XS_GRID", "IQ2_XS_T", "IQ2_XS"]
132
- },
133
- {
134
- "REPLS": {
135
- "TYPE": "iq2_s",
136
- "DST_TYPE": "f32",
137
- "BLOCK_SIZE": 256
138
- },
139
- "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_S_GRID", "IQ2_S_T", "IQ2_S"]
140
- },
141
- {
142
- "REPLS": {
143
- "TYPE": "iq3_xxs",
144
- "DST_TYPE": "f32",
145
- "BLOCK_SIZE": 256
146
- },
147
- "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ3_XSS_GRID", "IQ3_XSS_T", "IQ3_XSS"]
148
- },
149
- {
150
- "REPLS": {
151
- "TYPE": "iq3_s",
152
- "DST_TYPE": "f32",
153
- "BLOCK_SIZE": 256
154
- },
155
- "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ3_S_GRID", "IQ3_S_T", "IQ3_S"]
156
- },
157
- {
158
- "REPLS": {
159
- "TYPE": "iq1_s",
160
- "DST_TYPE": "f32",
161
- "BLOCK_SIZE": 256
162
- },
163
- "DECLS": ["BYTE_HELPERS", "IQ1_GRID", "IQ1_S_T", "IQ1_S"]
164
- },
165
- {
166
- "REPLS": {
167
- "TYPE": "iq1_m",
168
- "DST_TYPE": "f32",
169
- "BLOCK_SIZE": 256
170
- },
171
- "DECLS": ["BYTE_HELPERS", "IQ1_GRID", "IQ1_M_T", "IQ1_M"]
172
- },
173
- {
174
- "REPLS": {
175
- "TYPE": "iq4_nl",
176
- "DST_TYPE": "f32",
177
- "BLOCK_SIZE": 32,
178
- },
179
- "DECLS": ["BYTE_HELPERS", "IQ4_GRID", "IQ4_NL_T", "IQ4_NL"]
180
- },
181
- {
182
- "REPLS": {
183
- "TYPE": "iq4_xs",
184
- "DST_TYPE": "f32",
185
- "BLOCK_SIZE": 256,
186
- },
187
- "DECLS": ["BYTE_HELPERS", "IQ4_GRID", "IQ4_XS_T", "IQ4_XS"]
188
- }
189
- ]
190
-
191
- #end(VARIANTS)
192
-
193
- #define(DECLS)
194
-
195
- #decl(F32_VEC)
1
+ enable f16;
2
+ #define DECLARE_BYTE_LOADERS_SRC
3
+ #include "common_decls.tmpl"
4
+
5
+
6
+ #ifdef F32_VEC
196
7
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
197
8
  dst[(dst_base / 4) + offset] = src[(src_base / 4) + offset];
198
9
  }
199
- #enddecl(F32_VEC)
10
+ #endif
200
11
 
201
- #decl(F32)
12
+ #ifdef F32
202
13
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
203
14
  dst[dst_base + offset] = src[src_base + offset];
204
15
  }
205
- #enddecl(F32)
16
+ #endif
206
17
 
207
- #decl(F16)
18
+ #ifdef F16
208
19
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
209
20
  dst[dst_base + offset] = f32(src[src_base + offset]);
210
21
  }
211
- #enddecl(F16)
22
+ #endif
212
23
 
213
- #decl(I32)
24
+ #ifdef I32
214
25
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
215
26
  dst[dst_base + offset] = src[src_base + offset];
216
27
  }
217
- #enddecl(I32)
28
+ #endif
218
29
 
219
- #decl(Q4_0)
30
+ #ifdef Q1_0
220
31
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
221
- let block_q4_0 = src[src_base + offset];
222
- let d = f32(block_q4_0.d);
223
- for (var j: u32 = 0; j < 4; j++) {
224
- let q_packed = bitcast<u32>(vec2(block_q4_0.qs[2 * j], block_q4_0.qs[2 * j + 1]));
32
+ let block_byte_base = (src_base + offset) * 18;
33
+ let d = load_f16_as_f32_at_src(block_byte_base);
34
+ for (var j: u32 = 0u; j < 4u; j++) {
35
+ let q_packed = load_u32_at_src(block_byte_base + 2u + j * 4u);
36
+ let dst_base128 = dst_base + offset * 128u + j * 32u;
37
+ for (var k: u32 = 0; k < 4u; k++) {
38
+ let q_byte = get_byte(q_packed, k);
39
+ for (var bit: u32 = 0; bit < 8u; bit++) {
40
+ let w = select(-d, d, ((q_byte >> bit) & 1u) != 0u);
41
+ dst[dst_base128 + k * 8u + bit] = w;
42
+ }
43
+ }
44
+ }
45
+ }
46
+ #endif
47
+
48
+ #ifdef Q4_0
49
+ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
50
+ let block_byte_base = (src_base + offset) * 18; // Block stride: 18 bytes
51
+ let d = load_f16_as_f32_at_src(block_byte_base);
52
+ for (var j: u32 = 0u; j < 4; j++) {
53
+ let q_byte_offset = block_byte_base + 2 + j * 4;
54
+ let q_packed = load_u32_at_src(q_byte_offset);
225
55
  for (var k: u32 = 0; k < 4; k++) {
226
56
  let q_byte = get_byte(q_packed, k);
227
- let q_hi = (f32((q_byte >> 4) & 0xF) - 8.0f) * d;
228
- let q_lo = (f32(q_byte & 0xF) - 8.0f) * d;
57
+ let q_hi = (f32((q_byte >> 4) & 0xF) - 8.0) * d;
58
+ let q_lo = (f32(q_byte & 0xFu) - 8.0) * d;
229
59
  let dst_offset = dst_base + offset * 32 + j * 4 + k;
230
60
  dst[dst_offset] = q_lo;
231
- dst[dst_offset + 16] = q_hi;
61
+ dst[dst_offset + 16u] = q_hi;
232
62
  }
233
63
  }
234
64
  }
235
- #enddecl(Q4_0)
65
+ #endif
236
66
 
237
- #decl(Q4_1)
67
+ #ifdef Q4_1
238
68
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
239
69
  let block_q4_1 = src[src_base + offset];
240
70
  let d = f32(block_q4_1.d);
@@ -251,31 +81,35 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
251
81
  }
252
82
  }
253
83
  }
254
- #enddecl(Q4_1)
84
+ #endif
255
85
 
256
- #decl(Q5_0)
86
+ #ifdef Q5_0
257
87
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
258
- let block_q5_0 = src[src_base + offset];
259
- let d = f32(block_q5_0.d);
260
- let qh_packed = bitcast<u32>(vec2(block_q5_0.qh[0], block_q5_0.qh[1]));
88
+ let block_byte_base = (src_base + offset) * 22; // Block stride: 22 bytes
89
+ let d = load_f16_as_f32_at_src(block_byte_base);
90
+ let qh_packed = load_u32_at_src(block_byte_base + 2);
261
91
  for (var j: u32 = 0; j < 4; j++) {
262
- let q_packed = bitcast<u32>(vec2(block_q5_0.qs[2 * j], block_q5_0.qs[2 * j + 1]));
92
+ let q_byte_offset = block_byte_base + 6 + j * 4;
93
+ let q_packed = load_u32_at_src(q_byte_offset);
94
+
263
95
  for (var k: u32 = 0; k < 4; k++) {
264
96
  let q_byte = get_byte(q_packed, k);
97
+
265
98
  let qh_hi = (qh_packed >> (j * 4 + k + 12)) & 0x10;
266
99
  let q_hi = (f32(((q_byte >> 4) & 0xF) | qh_hi) - 16.0) * d;
100
+
267
101
  let qh_lo = ((qh_packed >> (j * 4 + k)) << 4) & 0x10;
268
102
  let q_lo = (f32((q_byte & 0xF) | qh_lo) - 16.0) * d;
103
+
269
104
  let dst_offset = dst_base + offset * 32 + j * 4 + k;
270
105
  dst[dst_offset] = q_lo;
271
106
  dst[dst_offset + 16] = q_hi;
272
107
  }
273
108
  }
274
109
  }
110
+ #endif
275
111
 
276
- #enddecl(Q5_0)
277
-
278
- #decl(Q5_1)
112
+ #ifdef Q5_1
279
113
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
280
114
  let block_q5_1 = src[src_base + offset];
281
115
  let d = f32(block_q5_1.d);
@@ -294,25 +128,26 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
294
128
  }
295
129
  }
296
130
  }
297
- #enddecl(Q5_1)
131
+ #endif
298
132
 
299
- #decl(Q8_0)
133
+ #ifdef Q8_0
300
134
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
301
- let block_q8_0 = src[src_base + offset];
302
- let d = f32(block_q8_0.d);
303
- for (var j: u32 = 0; j < 8; j++) {
304
- let q_packed = bitcast<u32>(vec2(block_q8_0.qs[2 * j], block_q8_0.qs[2 * j + 1]));
305
- for (var k: u32 = 0; k < 4; k++) {
135
+ let block_byte_base = (src_base + offset) * 34; // Block stride: 34 bytes
136
+ let d = load_f16_as_f32_at_src(block_byte_base);
137
+ for (var j: u32 = 0u; j < 8u; j++) {
138
+ let q_byte_offset = block_byte_base + 2u + j * 4u;
139
+ let q_packed = load_u32_at_src(q_byte_offset);
140
+ for (var k: u32 = 0u; k < 4u; k++) {
306
141
  let q_byte = get_byte_i32(q_packed, k);
307
142
  let q_val = f32(q_byte) * d;
308
- let dst_offset = dst_base + offset * 32 + j * 4 + k;
143
+ let dst_offset = dst_base + offset * 32u + j * 4u + k;
309
144
  dst[dst_offset] = q_val;
310
145
  }
311
146
  }
312
147
  }
313
- #enddecl(Q8_0)
148
+ #endif
314
149
 
315
- #decl(Q2_K)
150
+ #ifdef Q2_K
316
151
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
317
152
  let block = src[src_base + offset];
318
153
  let d = f32(block.d);
@@ -340,40 +175,46 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
340
175
  }
341
176
  }
342
177
  }
343
- #enddecl(Q2_K)
178
+ #endif
344
179
 
345
- #decl(Q3_K)
180
+ #ifdef Q3_K
346
181
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
347
- let block = src[src_base + offset];
348
- let d = f32(block.d);
182
+ let block_byte_base = (src_base + offset) * 110; // Block stride: 110 bytes
183
+
184
+ // Bytes 108-109: f16 scale 'd'
185
+ let d = load_f16_as_f32_at_src(block_byte_base + 108);
349
186
 
350
- // extract 6-bit scales, which consist of 4-bits from first 8 bytes of scale,
351
- // and 2-bits from the last 4 bytes
187
+ // Bytes 96-107: 12 bytes of scales (3 u32s)
352
188
  let kmask1: u32 = 0x03030303;
353
189
  let kmask2: u32 = 0x0f0f0f0f;
190
+
354
191
  var scale_vals: array<u32, 4>;
355
- for (var i: u32 = 0; i < 4; i++) {
356
- scale_vals[i] = bitcast<u32>(vec2(block.scales[2 * i], block.scales[2 * i + 1]));
357
- }
192
+ scale_vals[0] = load_u32_at_src(block_byte_base + 96);
193
+ scale_vals[1] = load_u32_at_src(block_byte_base + 100);
194
+ scale_vals[2] = load_u32_at_src(block_byte_base + 104);
195
+
358
196
  var tmp: u32 = scale_vals[2];
359
197
  scale_vals[2] = ((scale_vals[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
360
198
  scale_vals[3] = ((scale_vals[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
361
199
  scale_vals[0] = (scale_vals[0] & kmask2) | ((tmp & kmask1) << 4);
362
200
  scale_vals[1] = (scale_vals[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
363
201
 
364
- // convert arrays of f16 -> u32
202
+ // Bytes 0-31: 32 bytes of hmask (8 u32s)
365
203
  var hmask_vals: array<u32, 8>;
366
204
  for (var i: u32 = 0; i < 8; i++) {
367
- hmask_vals[i] = bitcast<u32>(vec2(block.hmask[2 * i], block.hmask[2 * i + 1]));
205
+ hmask_vals[i] = load_u32_at_src(block_byte_base + i * 4);
368
206
  }
207
+
208
+ // Bytes 32-95: 64 bytes of qs (16 u32s)
369
209
  var qs_vals: array<u32, 16>;
370
- for (var i: u32 = 0; i < 16; i++) {
371
- qs_vals[i] = bitcast<u32>(vec2(block.qs[2 * i], block.qs[2 * i + 1]));
210
+ for (var i: u32 = 0u; i < 16; i++) {
211
+ qs_vals[i] = load_u32_at_src(block_byte_base + 32 + i * 4);
372
212
  }
373
213
 
374
214
  var dst_i = dst_base + offset * 256;
375
215
  var is: u32 = 0;
376
216
  var m: u32 = 1;
217
+
377
218
  // 2 halves of the block (128 elements each)
378
219
  for (var q_b_idx: u32 = 0; q_b_idx < 64; q_b_idx += 32) {
379
220
  // 4 groups (each group has 2 blocks of 16 elements)
@@ -383,11 +224,13 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
383
224
  let sc = get_byte(scale_vals[is / 4], is % 4);
384
225
  is++;
385
226
  let dl = d * (f32(sc) - 32.0);
386
- for (var l: u32 = 0u; l < 16u; l++) {
227
+
228
+ for (var l: u32 = 0; l < 16; l++) {
387
229
  let q_idx = q_b_idx + k + l;
388
230
  let hm_idx = k + l;
389
231
  let q_byte = get_byte(qs_vals[q_idx / 4], q_idx % 4);
390
232
  let hmask_byte = get_byte(hmask_vals[hm_idx / 4], hm_idx % 4);
233
+
391
234
  let hm = select(4.0, 0.0, (hmask_byte & m) != 0);
392
235
  let qs_val = (q_byte >> shift) & 3;
393
236
  dst[dst_i] = (f32(qs_val) - hm) * dl;
@@ -398,9 +241,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
398
241
  }
399
242
  }
400
243
  }
401
- #enddecl(Q3_K)
244
+ #endif
402
245
 
403
- #decl(Q4_K)
246
+ #ifdef Q4_K
404
247
  // 8 blocks of 32 elements each
405
248
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
406
249
  let block = src[src_base + offset];
@@ -425,9 +268,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
425
268
  }
426
269
  }
427
270
  }
428
- #enddecl(Q4_K)
271
+ #endif
429
272
 
430
- #decl(Q5_K)
273
+ #ifdef Q5_K
431
274
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
432
275
  let block = src[src_base + offset];
433
276
  let d = f32(block.d);
@@ -455,26 +298,32 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
455
298
  }
456
299
  }
457
300
  }
458
- #enddecl(Q5_K)
301
+ #endif
459
302
 
460
- #decl(Q6_K)
303
+ #ifdef Q6_K
461
304
  // 16 blocks of 16 elements each
462
305
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
463
- let block = src[src_base + offset];
464
- let d = f32(block.d);
306
+ let block_byte_base = (src_base + offset) * 210; // Block stride: 210 bytes
465
307
 
466
- // convert arrays of f16 -> u32
308
+ // Bytes 208-209: f16 scale 'd'
309
+ let d = load_f16_as_f32_at_src(block_byte_base + 208);
310
+
311
+ // Bytes 0-127: 128 bytes of ql (32 u32s)
467
312
  var ql_vals: array<u32, 32>;
468
313
  for (var i: u32 = 0; i < 32; i++) {
469
- ql_vals[i] = bitcast<u32>(vec2(block.ql[2 * i], block.ql[2 * i + 1]));
314
+ ql_vals[i] = load_u32_at_src(block_byte_base + i * 4);
470
315
  }
316
+
317
+ // Bytes 128-191: 64 bytes of qh (16 u32s)
471
318
  var qh_vals: array<u32, 16>;
472
- for (var i: u32 = 0; i < 16; i++) {
473
- qh_vals[i] = bitcast<u32>(vec2(block.qh[2 * i], block.qh[2 * i + 1]));
319
+ for (var i: u32 = 0; i < 16u; i++) {
320
+ qh_vals[i] = load_u32_at_src(block_byte_base + 128 + i * 4u);
474
321
  }
322
+
323
+ // Bytes 192-207: 16 bytes of scales (4 u32s)
475
324
  var scale_vals: array<u32, 4>;
476
325
  for (var i: u32 = 0; i < 4; i++) {
477
- scale_vals[i] = bitcast<u32>(vec2(block.scales[2 * i], block.scales[2 * i + 1]));
326
+ scale_vals[i] = load_u32_at_src(block_byte_base + 192 + i * 4);
478
327
  }
479
328
 
480
329
  var dst_i = dst_base + offset * 256;
@@ -511,17 +360,18 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
511
360
  sc_b_idx += 8;
512
361
  }
513
362
  }
363
+ #endif
514
364
 
515
- #enddecl(Q6_K)
516
-
517
- #decl(IQ2_XXS)
365
+ #ifdef IQ2_XXS
518
366
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
519
- let block = src[src_base + offset];
520
- let d = f32(block.d);
367
+ let block_byte_base = (src_base + offset) * 66; // Block stride: 66 bytes
368
+ let d = load_f16_as_f32_at_src(block_byte_base);
521
369
  var dst_i = dst_base + offset * 256;
522
370
  for (var ib: u32 = 0; ib < 32; ib += 4) {
523
- let aux0 = bitcast<u32>(vec2(block.qs[ib], block.qs[ib + 1]));
524
- let aux1 = bitcast<u32>(vec2(block.qs[ib + 2], block.qs[ib + 3]));
371
+ let aux0_offset = block_byte_base + 2 + ib * 2;
372
+ let aux1_offset = block_byte_base + 2 + (ib + 2) * 2;
373
+ let aux0 = load_u32_at_src(aux0_offset);
374
+ let aux1 = load_u32_at_src(aux1_offset);
525
375
  let db = d * (0.5 + f32(aux1 >> 28)) * 0.25;
526
376
  for (var l: u32 = 0; l < 4; l++) {
527
377
  let ig = get_byte(aux0, l) * 8;
@@ -536,17 +386,21 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
536
386
  }
537
387
  }
538
388
  }
539
- #enddecl(IQ2_XXS)
389
+ #endif
390
+
391
+
540
392
 
541
- #decl(IQ2_XS)
393
+ #ifdef IQ2_XS
542
394
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
543
- let block = src[src_base + offset];
544
- let d = f32(block.d);
395
+ let block_byte_base = (src_base + offset) * 74; // Block stride: 74 bytes
396
+ let d = load_f16_as_f32_at_src(block_byte_base);
545
397
  var dst_i = dst_base + offset * 256;
398
+
546
399
  var scale_vals = array<u32, 2>(
547
- bitcast<u32>(vec2(block.scales[0], block.scales[1])),
548
- bitcast<u32>(vec2(block.scales[2], block.scales[3]))
400
+ load_u32_at_src(block_byte_base + 66),
401
+ load_u32_at_src(block_byte_base + 70)
549
402
  );
403
+
550
404
  for (var ib: u32 = 0; ib < 32; ib += 4) {
551
405
  let s = get_byte(scale_vals[ib / 16], (ib % 16) / 4);
552
406
  let db = array<f32, 2>(
@@ -554,7 +408,8 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
554
408
  d * (0.5 + f32(s >> 4)) * 0.25
555
409
  );
556
410
  for (var l: u32 = 0; l < 4; l++) {
557
- let qs_val = bitcast<u32>(vec2(block.qs[ib + l], 0.0));
411
+ let qs_offset = block_byte_base + 2 + (ib + l) * 2;
412
+ let qs_val = load_u32_at_src(qs_offset) & 0xFFFF;
558
413
  let ig = (qs_val & 511) * 8;
559
414
  let is = qs_val >> 9;
560
415
  let signs = get_byte(ksigns_iq2xs[is / 4], is % 4);
@@ -568,25 +423,27 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
568
423
  }
569
424
  }
570
425
  }
571
- #enddecl(IQ2_XS)
426
+ #endif
572
427
 
573
- #decl(IQ2_S)
428
+ #ifdef IQ2_S
574
429
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
575
- let block = src[src_base + offset];
576
- let d = f32(block.d);
430
+ let block_byte_base = (src_base + offset) * 82; // Block stride: 82 bytes
431
+ let d = load_f16_as_f32_at_src(block_byte_base);
577
432
  var dst_i = dst_base + offset * 256;
433
+
578
434
  var qs_vals : array<u32, 16>;
579
435
  for (var i: u32 = 0; i < 16; i++) {
580
- qs_vals[i] = bitcast<u32>(vec2(block.qs[i * 2], block.qs[i * 2 + 1]));
436
+ qs_vals[i] = load_u32_at_src(block_byte_base + 2 + i * 4);
581
437
  }
582
- var qh_vals = array<u32, 2>(
583
- bitcast<u32>(vec2(block.qh[0], block.qh[1])),
584
- bitcast<u32>(vec2(block.qh[2], block.qh[3]))
585
- );
586
- var scale_vals = array<u32, 2>(
587
- bitcast<u32>(vec2(block.scales[0], block.scales[1])),
588
- bitcast<u32>(vec2(block.scales[2], block.scales[3]))
589
- );
438
+
439
+ var qh_vals: array<u32, 2>;
440
+ qh_vals[0] = load_u32_at_src(block_byte_base + 66);
441
+ qh_vals[1] = load_u32_at_src(block_byte_base + 70);
442
+
443
+ var scale_vals: array<u32, 2>;
444
+ scale_vals[0] = load_u32_at_src(block_byte_base + 74);
445
+ scale_vals[1] = load_u32_at_src(block_byte_base + 78);
446
+
590
447
  for (var ib: u32 = 0; ib < 8; ib ++) {
591
448
  let s = get_byte(scale_vals[ib / 4], ib % 4);
592
449
  let db = array<f32, 2>(
@@ -608,21 +465,21 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
608
465
  }
609
466
  }
610
467
  }
468
+ #endif
611
469
 
612
- #enddecl(IQ2_S)
613
-
614
- #decl(IQ3_XSS)
470
+ #ifdef IQ3_XXS
615
471
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
616
- let block = src[src_base + offset];
617
- let d = f32(block.d);
472
+ let block_byte_base = (src_base + offset) * 98; // Block stride: 98 bytes
473
+ let d = load_f16_as_f32_at_src(block_byte_base);
618
474
  var dst_i = dst_base + offset * 256;
619
475
  for (var ib: u32 = 0; ib < 16; ib += 2) {
620
- let sc_sign = bitcast<u32>(vec2(block.qs[ib + 32], block.qs[ib + 33]));
476
+ let sc_sign_offset = block_byte_base + 2 + (ib + 32) * 2;
477
+ let sc_sign = load_u32_at_src(sc_sign_offset);
621
478
  let db = d * (0.5 + f32(sc_sign >> 28)) * 0.5;
622
479
  for (var l: u32 = 0; l < 4; l++) {
623
480
  let is = (sc_sign >> (7 * l)) & 127;
624
481
  let signs = get_byte(ksigns_iq2xs[is / 4], is % 4);
625
- let ig_val = bitcast<u32>(vec2(block.qs[ib * 2 + l], 0.0));
482
+ let ig_val = load_u32_at_src(block_byte_base + 2 + (ib * 2 + l) * 2) & 0xFFFF;
626
483
  let ig1 = get_byte(ig_val, 0);
627
484
  let ig2 = get_byte(ig_val, 1);
628
485
  for (var j: u32 = 0; j < 4; j++) {
@@ -638,22 +495,26 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
638
495
  }
639
496
  }
640
497
  }
641
- #enddecl(IQ3_XSS)
498
+ #endif
642
499
 
643
- #decl(IQ3_S)
500
+ #ifdef IQ3_S
644
501
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
645
- let block = src[src_base + offset];
646
- let d = f32(block.d);
502
+ let block_byte_base = (src_base + offset) * 110; // Block stride: 110 bytes
503
+ let d = load_f16_as_f32_at_src(block_byte_base);
647
504
  var dst_i = dst_base + offset * 256;
505
+
648
506
  var qh_vals = array<u32, 2>(
649
- bitcast<u32>(vec2(block.qh[0], block.qh[1])),
650
- bitcast<u32>(vec2(block.qh[2], block.qh[3]))
507
+ load_u32_at_src(block_byte_base + 66),
508
+ load_u32_at_src(block_byte_base + 70)
651
509
  );
510
+
652
511
  var sign_vals: array<u32, 8>;
653
512
  for (var i: u32 = 0; i < 8; i++) {
654
- sign_vals[i] = bitcast<u32>(vec2(block.signs[i * 2], block.signs[i * 2 + 1]));
513
+ sign_vals[i] = load_u32_at_src(block_byte_base + 74 + i * 4);
655
514
  }
656
- var scale_vals = bitcast<u32>(vec2(block.scales[0], block.scales[1]));
515
+
516
+ var scale_vals = load_u32_at_src(block_byte_base + 106);
517
+
657
518
  for (var ib: u32 = 0; ib < 4; ib++) {
658
519
  let s = get_byte(scale_vals, ib);
659
520
  let db = array<f32, 2>(
@@ -666,7 +527,7 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
666
527
  let sign_w = sign_vals[ib * 2 + k];
667
528
  for (var l: u32 = 0; l < 4; l++) {
668
529
  let signs = get_byte(sign_w, l);
669
- let ig_val = bitcast<u32>(vec2(block.qs[ib * 8 + k * 4 + l], 0.0));
530
+ let ig_val = load_u32_at_src(block_byte_base + 2 + (ib * 8 + k * 4 + l) * 2) & 0xFFFF;
670
531
  let ig1 = get_byte(ig_val, 0) | ((qh_byte << ((8 - (2 * l)))) & 256);
671
532
  let ig2 = get_byte(ig_val, 1) | ((qh_byte << ((7 - (2 * l)))) & 256);
672
533
  for (var j: u32 = 0; j < 4; j++) {
@@ -683,18 +544,18 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
683
544
  }
684
545
  }
685
546
  }
686
- #enddecl(IQ3_S)
547
+ #endif
687
548
 
688
- #decl(IQ1_S)
549
+ #ifdef IQ1_S
689
550
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
690
- let block = src[src_base + offset];
691
- let d = f32(block.d);
551
+ let block_byte_base = (src_base + offset) * 50; // Block stride: 50 bytes
552
+ let d = load_f16_as_f32_at_src(block_byte_base);
692
553
  var dst_i = dst_base + offset * 256;
693
554
  for (var ib: u32 = 0; ib < 8; ib++) {
694
- let qh = bitcast<u32>(vec2(block.qh[ib], 0.0));
695
- let dl = d * (2 * f32((qh >> 12) & 7) + 1);
555
+ let qh = load_u32_at_src(block_byte_base + 34 + ib * 2) & 0xFFFF;
556
+ let dl = d * (2.0 * f32((qh >> 12) & 7) + 1.0);
696
557
  let delta = select(IQ1_DELTA, -IQ1_DELTA, (qh & 0x8000) != 0);
697
- let qs_w = bitcast<u32>(vec2(block.qs[ib * 2], block.qs[ib * 2 + 1]));
558
+ let qs_w = load_u32_at_src(block_byte_base + 2 + ib * 4);
698
559
  for (var l: u32 = 0; l < 4; l++) {
699
560
  let ig = (get_byte(qs_w, l) | (((qh >> (3 * l)) & 7) << 8)) * 8;
700
561
  for (var j: u32 = 0; j < 8; j++) {
@@ -707,10 +568,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
707
568
  }
708
569
  }
709
570
  }
571
+ #endif
710
572
 
711
- #enddecl(IQ1_S)
712
-
713
- #decl(IQ1_M)
573
+ #ifdef IQ1_M
714
574
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
715
575
  let block = src[src_base + offset];
716
576
 
@@ -751,17 +611,16 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
751
611
  }
752
612
  }
753
613
  }
614
+ #endif
754
615
 
755
- #enddecl(IQ1_M)
756
-
757
- #decl(IQ4_NL)
616
+ #ifdef IQ4_NL
758
617
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
759
- let block = src[src_base + offset];
760
- let d = f32(block.d);
618
+ let block_byte_base = (src_base + offset) * 18; // Block stride: 18 bytes
619
+ let d = load_f16_as_f32_at_src(block_byte_base);
761
620
  var dst_i = dst_base + offset * 32;
762
621
  var qs: array<u32, 4>;
763
622
  for (var i: u32 = 0; i < 4; i++) {
764
- qs[i] = bitcast<u32>(vec2(block.qs[i * 2], block.qs[i * 2 + 1]));
623
+ qs[i] = load_u32_at_src(block_byte_base + 2 + i * 4);
765
624
  }
766
625
  for (var j: u32 = 0; j < 16; j++) {
767
626
  let qsb = get_byte(qs[j / 4], j % 4);
@@ -770,13 +629,13 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
770
629
  dst_i++;
771
630
  }
772
631
  }
773
- #enddecl(IQ4_NL)
632
+ #endif
774
633
 
775
- #decl(IQ4_XS)
634
+ #ifdef IQ4_XS
776
635
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
777
636
  let block = src[src_base + offset];
778
- let d = f32(block.d);
779
- let scales_h = bitcast<u32>(vec2(block.scales_h, 0.0));
637
+ let d = unpack2x16float(block.d_scales_h)[0];
638
+ let scales_h = block.d_scales_h >> 16;
780
639
  var dst_i = dst_base + offset * 256;
781
640
  for (var ib: u32 = 0; ib < 8; ib++) {
782
641
  let ls = ((get_byte(block.scales_l, ib / 2) >> (4 * (ib % 2))) & 0xF) | (((scales_h >> (2 * ib)) & 3) << 4);
@@ -791,24 +650,37 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
791
650
  dst_i += 16;
792
651
  }
793
652
  }
794
- #enddecl(IQ4_XS)
795
-
796
- #end(DECLS)
653
+ #endif
797
654
 
798
- #define(SHADER)
799
-
800
- enable f16;
655
+ #ifdef MXFP4
656
+ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
657
+ let block_byte_base = (src_base + offset) * 17;
658
+ let eu8 = get_byte(load_u32_at_src(block_byte_base), 0);
659
+ let d = ldexp(1.0, i32(eu8) - 128);
660
+ for (var j: u32 = 0u; j < 4; j++) {
661
+ let q_byte_offset = block_byte_base + 1 + j * 4;
662
+ let q_packed = load_u32_at_src(q_byte_offset);
663
+ for (var k: u32 = 0; k < 4; k++) {
664
+ let q_byte = get_byte(q_packed, k);
665
+ let q_hi = f32(kvalues_mxfp4[(q_byte >> 4) & 0xF]) * d;
666
+ let q_lo = f32(kvalues_mxfp4[q_byte & 0xFu]) * d;
667
+ let dst_offset = dst_base + offset * 32 + j * 4 + k;
668
+ dst[dst_offset] = q_lo;
669
+ dst[dst_offset + 16u] = q_hi;
670
+ }
671
+ }
672
+ }
673
+ #endif
801
674
 
802
- DECLS
803
675
 
804
676
  @group(0) @binding(0)
805
- var<storage, read_write> src: array<{{TYPE}}>;
677
+ var<storage, read_write> src: array<SRC_TYPE>;
806
678
 
807
679
  @group(0) @binding(1)
808
680
  var<storage, read_write> idx: array<i32>;
809
681
 
810
682
  @group(0) @binding(2)
811
- var<storage, read_write> dst: array<{{DST_TYPE}}>;
683
+ var<storage, read_write> dst: array<DST_TYPE>;
812
684
 
813
685
  struct Params {
814
686
  offset_src: u32, // in elements
@@ -842,9 +714,37 @@ struct Params {
842
714
  @group(0) @binding(3)
843
715
  var<uniform> params: Params;
844
716
 
845
- override wg_size: u32;
846
- @compute @workgroup_size(wg_size)
717
+ @compute @workgroup_size(WG_SIZE)
847
718
  fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
719
+ #ifdef FLOAT_PARALLEL
720
+ let blocks_per_row = params.ne0 / BLOCK_SIZE;
721
+ let row_count = params.n_rows * params.ne2 * params.ne3;
722
+
723
+ if (gid.x >= blocks_per_row * row_count) {
724
+ return;
725
+ }
726
+
727
+ let block_idx = gid.x % blocks_per_row;
728
+ var row_idx = gid.x / blocks_per_row;
729
+ let i_dst3 = row_idx / (params.ne2 * params.n_rows);
730
+
731
+ row_idx = row_idx % (params.ne2 * params.n_rows);
732
+ let i_dst2 = row_idx / params.n_rows;
733
+ let i_dst1 = row_idx % params.n_rows;
734
+
735
+ let i_idx2 = i_dst3 % params.idx2;
736
+ let i_idx1 = i_dst2 % params.idx1;
737
+ let i_idx0 = i_dst1;
738
+
739
+ let i_idx = params.offset_idx + i_idx0 * params.stride_idx0 + i_idx1 * params.stride_idx1 + i_idx2 * params.stride_idx2;
740
+
741
+ let idx_val = u32(idx[i_idx]);
742
+
743
+ let i_src_row = params.offset_src + idx_val * params.stride_src1 + i_dst2 * params.stride_src2 + i_dst3 * params.stride_src3;
744
+ let i_dst_row = params.offset_dst + i_dst1 * params.stride_dst1 + i_dst2 * params.stride_dst2 + i_dst3 * params.stride_dst3;
745
+
746
+ copy_elements(i_src_row, i_dst_row, block_idx);
747
+ #else
848
748
  if (gid.x >= params.n_rows * params.ne2 * params.ne3) {
849
749
  return;
850
750
  }
@@ -866,9 +766,8 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
866
766
  let i_src_row = params.offset_src + idx_val * params.stride_src1 + i_dst2 * params.stride_src2 + i_dst3 * params.stride_src3;
867
767
  let i_dst_row = params.offset_dst + i_dst1 * params.stride_dst1 + i_dst2 * params.stride_dst2 + i_dst3 * params.stride_dst3;
868
768
 
869
- for (var i: u32 = 0; i < params.ne0/{{BLOCK_SIZE}}; i++) {
769
+ for (var i: u32 = 0; i < params.ne0/BLOCK_SIZE; i++) {
870
770
  copy_elements(i_src_row, i_dst_row, i);
871
771
  }
772
+ #endif
872
773
  }
873
-
874
- #end(SHADER)