whispercpp 1.3.6 → 1.3.7

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (828) hide show
  1. checksums.yaml +4 -4
  2. data/.document +3 -0
  3. data/.rdoc_options +2 -0
  4. data/README.md +38 -5
  5. data/Rakefile +18 -3
  6. data/ext/dependencies.rb +10 -4
  7. data/ext/dependencies_for_windows.rb +17 -0
  8. data/ext/extconf.rb +20 -8
  9. data/ext/options.rb +54 -14
  10. data/ext/options_for_windows.rb +51 -0
  11. data/ext/ruby_whisper.c +36 -42
  12. data/ext/ruby_whisper.h +135 -0
  13. data/ext/ruby_whisper_context.c +107 -28
  14. data/ext/ruby_whisper_log_queue.c +180 -0
  15. data/ext/ruby_whisper_log_settable.h +47 -0
  16. data/ext/ruby_whisper_parakeet.c +49 -0
  17. data/ext/ruby_whisper_parakeet_context.c +304 -0
  18. data/ext/ruby_whisper_parakeet_context_params.c +117 -0
  19. data/ext/ruby_whisper_parakeet_model.c +84 -0
  20. data/ext/ruby_whisper_parakeet_params.c +548 -0
  21. data/ext/ruby_whisper_parakeet_segment.c +157 -0
  22. data/ext/ruby_whisper_parakeet_token.c +188 -0
  23. data/ext/ruby_whisper_parakeet_transcribe.cpp +58 -0
  24. data/ext/ruby_whisper_params.c +256 -65
  25. data/ext/ruby_whisper_segment.c +6 -6
  26. data/ext/ruby_whisper_transcribe.cpp +42 -15
  27. data/ext/sources/CMakeLists.txt +41 -3
  28. data/ext/sources/CMakePresets.json +95 -0
  29. data/ext/sources/cmake/parakeet-config.cmake.in +30 -0
  30. data/ext/sources/cmake/parakeet.pc.in +10 -0
  31. data/ext/sources/cmake/whisper.pc.in +1 -1
  32. data/ext/sources/examples/CMakeLists.txt +4 -2
  33. data/ext/sources/examples/bench/bench.cpp +1 -1
  34. data/ext/sources/examples/cli/cli.cpp +43 -9
  35. data/ext/sources/examples/common-ggml.cpp +2 -0
  36. data/ext/sources/examples/common-whisper.cpp +139 -67
  37. data/ext/sources/examples/common-whisper.h +11 -0
  38. data/ext/sources/examples/ffmpeg-transcode.cpp +211 -341
  39. data/ext/sources/examples/parakeet-cli/CMakeLists.txt +8 -0
  40. data/ext/sources/examples/parakeet-cli/parakeet-cli.cpp +243 -0
  41. data/ext/sources/examples/parakeet-quantize/CMakeLists.txt +7 -0
  42. data/ext/sources/examples/parakeet-quantize/parakeet-quantize.cpp +230 -0
  43. data/ext/sources/examples/server/server.cpp +199 -163
  44. data/ext/sources/ggml/CMakeLists.txt +21 -13
  45. data/ext/sources/ggml/cmake/FindNCCL.cmake +36 -0
  46. data/ext/sources/ggml/cmake/ggml-config.cmake.in +12 -2
  47. data/ext/sources/ggml/include/ggml-alloc.h +1 -0
  48. data/ext/sources/ggml/include/ggml-backend.h +72 -10
  49. data/ext/sources/ggml/include/ggml-cuda.h +3 -0
  50. data/ext/sources/ggml/include/ggml-rpc.h +3 -3
  51. data/ext/sources/ggml/include/ggml.h +101 -9
  52. data/ext/sources/ggml/include/gguf.h +10 -2
  53. data/ext/sources/ggml/src/CMakeLists.txt +22 -5
  54. data/ext/sources/ggml/src/ggml-alloc.c +5 -1
  55. data/ext/sources/ggml/src/ggml-backend-impl.h +22 -2
  56. data/ext/sources/ggml/src/ggml-backend-meta.cpp +2263 -0
  57. data/ext/sources/ggml/src/ggml-backend-reg.cpp +12 -0
  58. data/ext/sources/ggml/src/ggml-backend.cpp +110 -9
  59. data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +4 -0
  60. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +672 -257
  61. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +71 -0
  62. data/ext/sources/ggml/src/ggml-cann/common.h +20 -10
  63. data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +211 -30
  64. data/ext/sources/ggml/src/ggml-common.h +11 -0
  65. data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +58 -29
  66. data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +2 -0
  67. data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +16 -16
  68. data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +116 -7
  69. data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +65 -0
  70. data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +151 -1
  71. data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +0 -1
  72. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +4279 -1292
  73. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +5 -35
  74. data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +0 -1
  75. data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +72 -1
  76. data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +177 -27
  77. data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +1 -1
  78. data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +5 -0
  79. data/ext/sources/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
  80. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +10 -0
  81. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +95 -5
  82. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +2 -0
  83. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +146 -134
  84. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +88 -70
  85. data/ext/sources/ggml/src/ggml-cpu/ops.cpp +372 -73
  86. data/ext/sources/ggml/src/ggml-cpu/ops.h +3 -0
  87. data/ext/sources/ggml/src/ggml-cpu/quants.c +55 -0
  88. data/ext/sources/ggml/src/ggml-cpu/quants.h +3 -0
  89. data/ext/sources/ggml/src/ggml-cpu/repack.cpp +3 -0
  90. data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +90 -0
  91. data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +3 -16
  92. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +1402 -687
  93. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.h +8 -0
  94. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +597 -2766
  95. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp +5768 -0
  96. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.cpp +320 -0
  97. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.h +55 -0
  98. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_kernels.h +182 -19
  99. data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.cpp +1795 -0
  100. data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.h +14 -0
  101. data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp +3178 -0
  102. data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.h +95 -0
  103. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_barrier.h +34 -0
  104. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp +760 -0
  105. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h +32 -0
  106. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_tcm.h +409 -0
  107. data/ext/sources/ggml/src/ggml-cpu/vec.cpp +37 -53
  108. data/ext/sources/ggml/src/ggml-cpu/vec.h +225 -240
  109. data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +17 -7
  110. data/ext/sources/ggml/src/ggml-cuda/allreduce.cu +971 -0
  111. data/ext/sources/ggml/src/ggml-cuda/allreduce.cuh +29 -0
  112. data/ext/sources/ggml/src/ggml-cuda/argsort.cu +62 -26
  113. data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +44 -18
  114. data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +1 -0
  115. data/ext/sources/ggml/src/ggml-cuda/common.cuh +242 -28
  116. data/ext/sources/ggml/src/ggml-cuda/concat.cu +120 -114
  117. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cu +45 -21
  118. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cuh +1 -0
  119. data/ext/sources/ggml/src/ggml-cuda/convert.cu +53 -0
  120. data/ext/sources/ggml/src/ggml-cuda/convert.cuh +10 -0
  121. data/ext/sources/ggml/src/ggml-cuda/cpy.cu +14 -6
  122. data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +22 -0
  123. data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +278 -44
  124. data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +331 -130
  125. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +12 -0
  126. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +126 -27
  127. data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +40 -15
  128. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +18 -9
  129. data/ext/sources/ggml/src/ggml-cuda/fattn.cu +152 -49
  130. data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +2 -0
  131. data/ext/sources/ggml/src/ggml-cuda/fwht.cu +101 -0
  132. data/ext/sources/ggml/src/ggml-cuda/fwht.cuh +4 -0
  133. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cu +84 -35
  134. data/ext/sources/ggml/src/ggml-cuda/getrows.cu +34 -12
  135. data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +1069 -609
  136. data/ext/sources/ggml/src/ggml-cuda/im2col.cu +32 -29
  137. data/ext/sources/ggml/src/ggml-cuda/mean.cu +4 -2
  138. data/ext/sources/ggml/src/ggml-cuda/mma.cuh +242 -195
  139. data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +3 -3
  140. data/ext/sources/ggml/src/ggml-cuda/mmq.cu +18 -12
  141. data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +502 -423
  142. data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +19 -12
  143. data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +485 -57
  144. data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +6 -1
  145. data/ext/sources/ggml/src/ggml-cuda/norm.cu +36 -10
  146. data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +23 -7
  147. data/ext/sources/ggml/src/ggml-cuda/quantize.cu +133 -26
  148. data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +1 -1
  149. data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +5 -1
  150. data/ext/sources/ggml/src/ggml-cuda/rope.cu +11 -4
  151. data/ext/sources/ggml/src/ggml-cuda/scale.cu +4 -1
  152. data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +14 -6
  153. data/ext/sources/ggml/src/ggml-cuda/snake.cu +72 -0
  154. data/ext/sources/ggml/src/ggml-cuda/snake.cuh +8 -0
  155. data/ext/sources/ggml/src/ggml-cuda/softcap.cu +4 -1
  156. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +45 -13
  157. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +1 -1
  158. data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +40 -18
  159. data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +8 -4
  160. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +1 -0
  161. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +1 -0
  162. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +2 -0
  163. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +1 -0
  164. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +1 -0
  165. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +1 -0
  166. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +1 -0
  167. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +2 -0
  168. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +1 -0
  169. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +1 -0
  170. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +2 -0
  171. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +1 -0
  172. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +2 -0
  173. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu +5 -0
  174. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu +5 -0
  175. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +5 -0
  176. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu +7 -0
  177. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu +7 -0
  178. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu +7 -0
  179. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu +7 -0
  180. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu +7 -0
  181. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu +7 -0
  182. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu +7 -0
  183. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu +7 -0
  184. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu +7 -0
  185. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu +7 -0
  186. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu +7 -0
  187. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu +7 -0
  188. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu +7 -0
  189. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu +5 -0
  190. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu +5 -0
  191. data/ext/sources/ggml/src/ggml-cuda/top-k.cu +5 -4
  192. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +26 -23
  193. data/ext/sources/ggml/src/ggml-cuda/unary.cu +31 -2
  194. data/ext/sources/ggml/src/ggml-cuda/unary.cuh +2 -0
  195. data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +80 -0
  196. data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +7 -2
  197. data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +22 -4
  198. data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +3 -0
  199. data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +2 -1
  200. data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +1428 -743
  201. data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +45 -7
  202. data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +53 -84
  203. data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +25 -12
  204. data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +165 -184
  205. data/ext/sources/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +5 -5
  206. data/ext/sources/ggml/src/ggml-hexagon/htp/concat-ops.c +277 -0
  207. data/ext/sources/ggml/src/ggml-hexagon/htp/cpy-ops.c +170 -127
  208. data/ext/sources/ggml/src/ggml-hexagon/htp/cumsum-ops.c +270 -0
  209. data/ext/sources/ggml/src/ggml-hexagon/htp/diag-ops.c +216 -0
  210. data/ext/sources/ggml/src/ggml-hexagon/htp/fill-ops.c +123 -0
  211. data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +125 -97
  212. data/ext/sources/ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c +1148 -0
  213. data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +148 -42
  214. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.c +2 -2
  215. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.h +252 -62
  216. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +9 -0
  217. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +87 -1
  218. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-flash-attn-ops.c +1878 -0
  219. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-matmul-ops.c +2066 -0
  220. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-ops.c +6 -0
  221. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-ops.h +88 -0
  222. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-profile.h +34 -0
  223. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.c +158 -0
  224. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.h +134 -0
  225. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-utils.h +200 -0
  226. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +96 -13
  227. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +182 -57
  228. data/ext/sources/ggml/src/ggml-hexagon/htp/htp_iface.idl +9 -3
  229. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +71 -3
  230. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +27 -10
  231. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +63 -23
  232. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +9 -8
  233. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -0
  234. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-log.h +65 -0
  235. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-pow.h +42 -0
  236. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-repl.h +74 -0
  237. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +1 -0
  238. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
  239. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +5 -8
  240. data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +529 -815
  241. data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +2522 -234
  242. data/ext/sources/ggml/src/ggml-hexagon/htp/pad-ops.c +547 -0
  243. data/ext/sources/ggml/src/ggml-hexagon/htp/repeat-ops.c +148 -0
  244. data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +291 -95
  245. data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +59 -37
  246. data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +121 -133
  247. data/ext/sources/ggml/src/ggml-hexagon/htp/solve-tri-ops.c +267 -0
  248. data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c +244 -151
  249. data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +6 -6
  250. data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +719 -45
  251. data/ext/sources/ggml/src/ggml-hexagon/htp/vtcm-utils.h +16 -0
  252. data/ext/sources/ggml/src/ggml-hexagon/htp-opnode.h +272 -0
  253. data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +3 -1
  254. data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +22 -9
  255. data/ext/sources/ggml/src/ggml-impl.h +6 -1
  256. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +138 -13
  257. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +32 -1
  258. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +164 -28
  259. data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +80 -0
  260. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +190 -19
  261. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +2 -0
  262. data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +39 -26
  263. data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +823 -322
  264. data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +5 -6
  265. data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +54 -5
  266. data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +12248 -5907
  267. data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +67 -0
  268. data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +59 -0
  269. data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +1819 -112
  270. data/ext/sources/ggml/src/ggml-opencl/kernels/gated_delta_net.cl +249 -0
  271. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl +306 -0
  272. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl +256 -0
  273. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl +258 -0
  274. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl +283 -0
  275. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl +260 -0
  276. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl +262 -0
  277. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl +288 -0
  278. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl +267 -0
  279. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_f32.cl +150 -0
  280. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mat_Ab_Bi_8x4.cl → gemm_noshuffle_q4_0_f32.cl} +1 -1
  281. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl +172 -0
  282. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl +131 -0
  283. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl +134 -0
  284. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl +176 -0
  285. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl +140 -0
  286. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mm_q8_0_f32_8x4.cl → gemm_noshuffle_q8_0_f32.cl} +1 -1
  287. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_xmem_f16_f32_os8.cl +233 -0
  288. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl +165 -0
  289. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl +120 -0
  290. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl +123 -0
  291. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl +155 -0
  292. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl +123 -0
  293. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl +125 -0
  294. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl +160 -0
  295. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl +141 -0
  296. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +302 -0
  297. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general.cl → gemv_noshuffle_q4_0_f32.cl} +5 -5
  298. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle.cl → gemv_noshuffle_q4_0_f32_spec.cl} +5 -5
  299. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +318 -0
  300. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +291 -0
  301. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +294 -0
  302. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +326 -0
  303. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +293 -0
  304. data/ext/sources/ggml/src/ggml-opencl/kernels/get_rows.cl +15 -9
  305. data/ext/sources/ggml/src/ggml-opencl/kernels/moe_reorder_b.cl +30 -0
  306. data/ext/sources/ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl +82 -0
  307. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_iq4_nl_f32_l4_lm.cl +171 -0
  308. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl +179 -0
  309. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_0_f32_l4_lm.cl +173 -0
  310. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_1_f32_l4_lm.cl +175 -0
  311. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_k_f32_l4_lm.cl +192 -0
  312. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32.cl +164 -0
  313. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32_flat.cl +202 -0
  314. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl +196 -0
  315. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32.cl +241 -0
  316. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32_flat.cl +243 -0
  317. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32.cl +243 -0
  318. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32_flat.cl +247 -0
  319. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32.cl +187 -0
  320. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32_flat.cl +203 -0
  321. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +48 -64
  322. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +15 -5
  323. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +18 -11
  324. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +35 -13
  325. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +264 -192
  326. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +33 -7
  327. data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_gelu.cpp +25 -0
  328. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +1 -0
  329. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +1 -0
  330. data/ext/sources/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp +41 -0
  331. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.cpp +27 -3
  332. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +67 -36
  333. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +1 -0
  334. data/ext/sources/ggml/src/ggml-openvino/utils.cpp +101 -44
  335. data/ext/sources/ggml/src/ggml-openvino/utils.h +23 -3
  336. data/ext/sources/ggml/src/ggml-opt.cpp +1 -0
  337. data/ext/sources/ggml/src/ggml-quants.c +289 -114
  338. data/ext/sources/ggml/src/ggml-quants.h +3 -0
  339. data/ext/sources/ggml/src/ggml-rpc/CMakeLists.txt +24 -0
  340. data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +167 -311
  341. data/ext/sources/ggml/src/ggml-rpc/transport.cpp +683 -0
  342. data/ext/sources/ggml/src/ggml-rpc/transport.h +34 -0
  343. data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +50 -4
  344. data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +1 -1
  345. data/ext/sources/ggml/src/ggml-sycl/backend.hpp +3 -1
  346. data/ext/sources/ggml/src/ggml-sycl/common.cpp +74 -2
  347. data/ext/sources/ggml/src/ggml-sycl/common.hpp +41 -1
  348. data/ext/sources/ggml/src/ggml-sycl/convert.cpp +115 -13
  349. data/ext/sources/ggml/src/ggml-sycl/convert.hpp +9 -0
  350. data/ext/sources/ggml/src/ggml-sycl/cumsum.cpp +148 -0
  351. data/ext/sources/ggml/src/ggml-sycl/cumsum.hpp +5 -0
  352. data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +663 -0
  353. data/ext/sources/ggml/src/ggml-sycl/diag.cpp +67 -0
  354. data/ext/sources/ggml/src/ggml-sycl/diag.hpp +5 -0
  355. data/ext/sources/ggml/src/ggml-sycl/dmmv.cpp +586 -6
  356. data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +1 -90
  357. data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +0 -2
  358. data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.cpp +56 -0
  359. data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.hpp +63 -0
  360. data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp +7 -5
  361. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +4 -0
  362. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +76 -168
  363. data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +7 -0
  364. data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +3 -1
  365. data/ext/sources/ggml/src/ggml-sycl/fill.cpp +55 -0
  366. data/ext/sources/ggml/src/ggml-sycl/fill.hpp +5 -0
  367. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.cpp +69 -31
  368. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +1 -0
  369. data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +3 -0
  370. data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +79 -3
  371. data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +823 -190
  372. data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +353 -89
  373. data/ext/sources/ggml/src/ggml-sycl/im2col.hpp +5 -3
  374. data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +1344 -26
  375. data/ext/sources/ggml/src/ggml-sycl/mmvq.hpp +16 -0
  376. data/ext/sources/ggml/src/ggml-sycl/pad.cpp +27 -27
  377. data/ext/sources/ggml/src/ggml-sycl/quants.hpp +71 -0
  378. data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +7 -1
  379. data/ext/sources/ggml/src/ggml-sycl/solve_tri.cpp +172 -0
  380. data/ext/sources/ggml/src/ggml-sycl/solve_tri.hpp +8 -0
  381. data/ext/sources/ggml/src/ggml-sycl/ssm_conv.cpp +6 -1
  382. data/ext/sources/ggml/src/ggml-sycl/ssm_scan.cpp +156 -0
  383. data/ext/sources/ggml/src/ggml-sycl/ssm_scan.hpp +5 -0
  384. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +62 -10
  385. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +18 -6
  386. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp +6 -0
  387. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +1 -0
  388. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +1 -0
  389. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +1 -0
  390. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +1 -0
  391. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +1 -0
  392. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +1 -0
  393. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +1 -0
  394. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +1 -0
  395. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +1 -0
  396. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +1 -0
  397. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +1 -0
  398. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +1 -0
  399. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +1 -0
  400. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +1 -0
  401. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +1 -0
  402. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +1 -0
  403. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +1 -0
  404. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +1 -0
  405. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +1 -0
  406. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +1 -0
  407. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +1 -0
  408. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +1 -0
  409. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +1 -0
  410. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +1 -0
  411. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +1 -0
  412. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +1 -0
  413. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +1 -0
  414. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +1 -0
  415. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +1 -0
  416. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +1 -0
  417. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +1 -0
  418. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +1 -0
  419. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +1 -0
  420. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +1 -0
  421. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +1 -0
  422. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +1 -0
  423. data/ext/sources/ggml/src/ggml-sycl/type.hpp +112 -0
  424. data/ext/sources/ggml/src/ggml-sycl/upscale.cpp +410 -0
  425. data/ext/sources/ggml/src/ggml-sycl/upscale.hpp +9 -0
  426. data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +215 -53
  427. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +4 -0
  428. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +2 -0
  429. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +2 -0
  430. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +1 -0
  431. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +1 -0
  432. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +0 -2
  433. data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +11 -0
  434. data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +2060 -535
  435. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +4 -0
  436. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +6 -2
  437. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +146 -13
  438. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +3 -1
  439. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +1 -1
  440. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +25 -1
  441. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +88 -0
  442. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl +643 -1
  443. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp +32 -0
  444. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q1_0.comp +29 -0
  445. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +0 -1
  446. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl +27 -0
  447. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +0 -1
  448. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2_decode_vector.comp +7 -0
  449. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +197 -48
  450. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +60 -59
  451. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +115 -113
  452. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +122 -31
  453. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_dequant.glsl +131 -0
  454. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mmq_funcs.glsl +203 -0
  455. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp +115 -0
  456. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +125 -64
  457. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl +0 -1
  458. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl +10 -1
  459. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl +16 -6
  460. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +76 -54
  461. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +0 -1
  462. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +0 -1
  463. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +122 -27
  464. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +6 -6
  465. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +1 -1
  466. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +1 -1
  467. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +1 -1
  468. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +1 -0
  469. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +88 -55
  470. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +11 -17
  471. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +43 -10
  472. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +159 -125
  473. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +8 -8
  474. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +24 -9
  475. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +0 -1
  476. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +5 -2
  477. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +0 -1
  478. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +3 -2
  479. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/snake.comp +49 -0
  480. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +11 -1
  481. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +0 -1
  482. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +79 -2
  483. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +171 -147
  484. data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +5 -2
  485. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +2202 -283
  486. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +2610 -1403
  487. data/ext/sources/ggml/src/ggml-webgpu/pre_wgsl.hpp +37 -7
  488. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_id.wgsl +64 -0
  489. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +8 -7
  490. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +76 -95
  491. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +19 -1
  492. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/conv2d.wgsl +165 -0
  493. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{cpy.tmpl.wgsl → cpy.wgsl} +25 -50
  494. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +107 -184
  495. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_quant_staging.tmpl +124 -0
  496. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl +397 -0
  497. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_blk.wgsl +101 -0
  498. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_reduce.wgsl +84 -0
  499. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl +619 -0
  500. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl +149 -0
  501. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl +183 -78
  502. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl +155 -0
  503. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/im2col.wgsl +101 -0
  504. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +655 -495
  505. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id.wgsl +195 -0
  506. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_gather.wgsl +52 -0
  507. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl +154 -0
  508. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +8 -6
  509. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.wgsl +5 -1
  510. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +80 -409
  511. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl +1432 -0
  512. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl +303 -0
  513. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl +21 -0
  514. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl +173 -0
  515. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl +152 -0
  516. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{rope.tmpl.wgsl → rope.wgsl} +71 -142
  517. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/row_norm.wgsl +153 -0
  518. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/scale.wgsl +6 -4
  519. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set.wgsl +109 -0
  520. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +2 -3
  521. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows_quant.wgsl +224 -0
  522. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{soft_max.tmpl.wgsl → soft_max.wgsl} +106 -206
  523. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/solve_tri.wgsl +121 -0
  524. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_conv.wgsl +65 -0
  525. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl +193 -0
  526. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +68 -48
  527. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
  528. data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +18 -14
  529. data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +1 -1
  530. data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +244 -10
  531. data/ext/sources/ggml/src/ggml.c +110 -28
  532. data/ext/sources/ggml/src/gguf.cpp +173 -28
  533. data/ext/sources/include/parakeet.h +342 -0
  534. data/ext/sources/include/whisper.h +10 -0
  535. data/ext/sources/media/matmul.png +0 -0
  536. data/ext/sources/src/CMakeLists.txt +23 -0
  537. data/ext/sources/src/parakeet-arch.h +188 -0
  538. data/ext/sources/src/parakeet.cpp +3838 -0
  539. data/ext/sources/src/whisper.cpp +56 -12
  540. data/extsources.rb +26 -10
  541. data/lib/whisper/log_settable.rb +36 -0
  542. data/lib/whisper/model/uri.rb +13 -1
  543. data/lib/whisper/output.rb +74 -0
  544. data/sig/whisper.rbs +411 -62
  545. data/test/helper.rb +2 -0
  546. data/test/jfk_reader/jfk_reader.c +50 -7
  547. data/test/test_callback.rb +1 -0
  548. data/test/test_package.rb +6 -5
  549. data/test/test_parakeet.rb +28 -0
  550. data/test/test_parakeet_callback.rb +107 -0
  551. data/test/test_parakeet_context.rb +116 -0
  552. data/test/test_parakeet_context_params.rb +24 -0
  553. data/test/test_parakeet_model.rb +21 -0
  554. data/test/test_parakeet_params.rb +78 -0
  555. data/test/test_parakeet_segment.rb +42 -0
  556. data/test/test_parakeet_token.rb +73 -0
  557. data/test/test_params.rb +2 -0
  558. data/test/test_vad_segment.rb +1 -1
  559. data/test/test_whisper.rb +24 -6
  560. data/whispercpp.gemspec +2 -2
  561. metadata +215 -281
  562. data/ext/sources/bindings/javascript/CMakeLists.txt +0 -41
  563. data/ext/sources/bindings/javascript/emscripten.cpp +0 -93
  564. data/ext/sources/bindings/javascript/libwhisper.worker.js +0 -1
  565. data/ext/sources/bindings/javascript/package.json +0 -26
  566. data/ext/sources/bindings/javascript/whisper.js +0 -19
  567. data/ext/sources/examples/addon.node/CMakeLists.txt +0 -31
  568. data/ext/sources/examples/addon.node/__test__/whisper.spec.js +0 -133
  569. data/ext/sources/examples/addon.node/addon.cpp +0 -557
  570. data/ext/sources/examples/addon.node/index.js +0 -59
  571. data/ext/sources/examples/addon.node/package.json +0 -16
  572. data/ext/sources/examples/addon.node/vad-example.js +0 -132
  573. data/ext/sources/examples/bench.wasm/CMakeLists.txt +0 -49
  574. data/ext/sources/examples/bench.wasm/emscripten.cpp +0 -87
  575. data/ext/sources/examples/bench.wasm/index-tmpl.html +0 -285
  576. data/ext/sources/examples/coi-serviceworker.js +0 -146
  577. data/ext/sources/examples/command/CMakeLists.txt +0 -10
  578. data/ext/sources/examples/command/command.cpp +0 -802
  579. data/ext/sources/examples/command/commands.txt +0 -9
  580. data/ext/sources/examples/command.wasm/CMakeLists.txt +0 -50
  581. data/ext/sources/examples/command.wasm/emscripten.cpp +0 -327
  582. data/ext/sources/examples/command.wasm/index-tmpl.html +0 -415
  583. data/ext/sources/examples/generate-karaoke.sh +0 -57
  584. data/ext/sources/examples/helpers.js +0 -191
  585. data/ext/sources/examples/livestream.sh +0 -112
  586. data/ext/sources/examples/lsp/CMakeLists.txt +0 -10
  587. data/ext/sources/examples/lsp/lsp.cpp +0 -471
  588. data/ext/sources/examples/lsp/whisper.vim +0 -362
  589. data/ext/sources/examples/python/test_whisper_processor.py +0 -7
  590. data/ext/sources/examples/python/whisper_processor.py +0 -54
  591. data/ext/sources/examples/server/bench.js +0 -29
  592. data/ext/sources/examples/server.py +0 -120
  593. data/ext/sources/examples/stream/CMakeLists.txt +0 -10
  594. data/ext/sources/examples/stream/stream.cpp +0 -437
  595. data/ext/sources/examples/stream.wasm/CMakeLists.txt +0 -49
  596. data/ext/sources/examples/stream.wasm/emscripten.cpp +0 -216
  597. data/ext/sources/examples/stream.wasm/index-tmpl.html +0 -491
  598. data/ext/sources/examples/sycl/CMakeLists.txt +0 -9
  599. data/ext/sources/examples/sycl/build.sh +0 -22
  600. data/ext/sources/examples/sycl/ls-sycl-device.cpp +0 -11
  601. data/ext/sources/examples/sycl/run-whisper.sh +0 -17
  602. data/ext/sources/examples/talk-llama/CMakeLists.txt +0 -48
  603. data/ext/sources/examples/talk-llama/eleven-labs.py +0 -80
  604. data/ext/sources/examples/talk-llama/llama-adapter.cpp +0 -488
  605. data/ext/sources/examples/talk-llama/llama-adapter.h +0 -89
  606. data/ext/sources/examples/talk-llama/llama-arch.cpp +0 -2877
  607. data/ext/sources/examples/talk-llama/llama-arch.h +0 -628
  608. data/ext/sources/examples/talk-llama/llama-batch.cpp +0 -919
  609. data/ext/sources/examples/talk-llama/llama-batch.h +0 -173
  610. data/ext/sources/examples/talk-llama/llama-chat.cpp +0 -896
  611. data/ext/sources/examples/talk-llama/llama-chat.h +0 -71
  612. data/ext/sources/examples/talk-llama/llama-context.cpp +0 -3633
  613. data/ext/sources/examples/talk-llama/llama-context.h +0 -359
  614. data/ext/sources/examples/talk-llama/llama-cparams.cpp +0 -5
  615. data/ext/sources/examples/talk-llama/llama-cparams.h +0 -47
  616. data/ext/sources/examples/talk-llama/llama-ext.h +0 -12
  617. data/ext/sources/examples/talk-llama/llama-grammar.cpp +0 -1464
  618. data/ext/sources/examples/talk-llama/llama-grammar.h +0 -194
  619. data/ext/sources/examples/talk-llama/llama-graph.cpp +0 -2735
  620. data/ext/sources/examples/talk-llama/llama-graph.h +0 -1031
  621. data/ext/sources/examples/talk-llama/llama-hparams.cpp +0 -258
  622. data/ext/sources/examples/talk-llama/llama-hparams.h +0 -353
  623. data/ext/sources/examples/talk-llama/llama-impl.cpp +0 -171
  624. data/ext/sources/examples/talk-llama/llama-impl.h +0 -75
  625. data/ext/sources/examples/talk-llama/llama-io.cpp +0 -15
  626. data/ext/sources/examples/talk-llama/llama-io.h +0 -35
  627. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +0 -330
  628. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.h +0 -137
  629. data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +0 -2285
  630. data/ext/sources/examples/talk-llama/llama-kv-cache.h +0 -389
  631. data/ext/sources/examples/talk-llama/llama-kv-cells.h +0 -533
  632. data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.cpp +0 -275
  633. data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.h +0 -140
  634. data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +0 -268
  635. data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +0 -139
  636. data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +0 -1165
  637. data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +0 -182
  638. data/ext/sources/examples/talk-llama/llama-memory.cpp +0 -59
  639. data/ext/sources/examples/talk-llama/llama-memory.h +0 -122
  640. data/ext/sources/examples/talk-llama/llama-mmap.cpp +0 -752
  641. data/ext/sources/examples/talk-llama/llama-mmap.h +0 -73
  642. data/ext/sources/examples/talk-llama/llama-model-loader.cpp +0 -1655
  643. data/ext/sources/examples/talk-llama/llama-model-loader.h +0 -206
  644. data/ext/sources/examples/talk-llama/llama-model-saver.cpp +0 -299
  645. data/ext/sources/examples/talk-llama/llama-model-saver.h +0 -40
  646. data/ext/sources/examples/talk-llama/llama-model.cpp +0 -9056
  647. data/ext/sources/examples/talk-llama/llama-model.h +0 -597
  648. data/ext/sources/examples/talk-llama/llama-quant.cpp +0 -1304
  649. data/ext/sources/examples/talk-llama/llama-quant.h +0 -1
  650. data/ext/sources/examples/talk-llama/llama-sampler.cpp +0 -3885
  651. data/ext/sources/examples/talk-llama/llama-sampler.h +0 -42
  652. data/ext/sources/examples/talk-llama/llama-vocab.cpp +0 -3970
  653. data/ext/sources/examples/talk-llama/llama-vocab.h +0 -187
  654. data/ext/sources/examples/talk-llama/llama.cpp +0 -1194
  655. data/ext/sources/examples/talk-llama/llama.h +0 -1573
  656. data/ext/sources/examples/talk-llama/models/afmoe.cpp +0 -190
  657. data/ext/sources/examples/talk-llama/models/apertus.cpp +0 -125
  658. data/ext/sources/examples/talk-llama/models/arcee.cpp +0 -135
  659. data/ext/sources/examples/talk-llama/models/arctic.cpp +0 -137
  660. data/ext/sources/examples/talk-llama/models/arwkv7.cpp +0 -86
  661. data/ext/sources/examples/talk-llama/models/baichuan.cpp +0 -123
  662. data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +0 -143
  663. data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +0 -133
  664. data/ext/sources/examples/talk-llama/models/bert.cpp +0 -184
  665. data/ext/sources/examples/talk-llama/models/bitnet.cpp +0 -145
  666. data/ext/sources/examples/talk-llama/models/bloom.cpp +0 -101
  667. data/ext/sources/examples/talk-llama/models/chameleon.cpp +0 -178
  668. data/ext/sources/examples/talk-llama/models/chatglm.cpp +0 -132
  669. data/ext/sources/examples/talk-llama/models/codeshell.cpp +0 -111
  670. data/ext/sources/examples/talk-llama/models/cogvlm.cpp +0 -102
  671. data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +0 -134
  672. data/ext/sources/examples/talk-llama/models/command-r.cpp +0 -122
  673. data/ext/sources/examples/talk-llama/models/dbrx.cpp +0 -122
  674. data/ext/sources/examples/talk-llama/models/deci.cpp +0 -135
  675. data/ext/sources/examples/talk-llama/models/deepseek.cpp +0 -142
  676. data/ext/sources/examples/talk-llama/models/deepseek2.cpp +0 -262
  677. data/ext/sources/examples/talk-llama/models/delta-net-base.cpp +0 -445
  678. data/ext/sources/examples/talk-llama/models/dots1.cpp +0 -132
  679. data/ext/sources/examples/talk-llama/models/dream.cpp +0 -105
  680. data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +0 -148
  681. data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +0 -110
  682. data/ext/sources/examples/talk-llama/models/eurobert.cpp +0 -97
  683. data/ext/sources/examples/talk-llama/models/exaone-moe.cpp +0 -145
  684. data/ext/sources/examples/talk-llama/models/exaone.cpp +0 -114
  685. data/ext/sources/examples/talk-llama/models/exaone4.cpp +0 -123
  686. data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +0 -111
  687. data/ext/sources/examples/talk-llama/models/falcon.cpp +0 -120
  688. data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +0 -116
  689. data/ext/sources/examples/talk-llama/models/gemma.cpp +0 -112
  690. data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +0 -128
  691. data/ext/sources/examples/talk-llama/models/gemma3.cpp +0 -155
  692. data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +0 -384
  693. data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +0 -170
  694. data/ext/sources/examples/talk-llama/models/glm4.cpp +0 -157
  695. data/ext/sources/examples/talk-llama/models/gpt2.cpp +0 -105
  696. data/ext/sources/examples/talk-llama/models/gptneox.cpp +0 -144
  697. data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +0 -195
  698. data/ext/sources/examples/talk-llama/models/granite.cpp +0 -210
  699. data/ext/sources/examples/talk-llama/models/grok.cpp +0 -159
  700. data/ext/sources/examples/talk-llama/models/grovemoe.cpp +0 -139
  701. data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +0 -132
  702. data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +0 -153
  703. data/ext/sources/examples/talk-llama/models/internlm2.cpp +0 -120
  704. data/ext/sources/examples/talk-llama/models/jais.cpp +0 -86
  705. data/ext/sources/examples/talk-llama/models/jais2.cpp +0 -123
  706. data/ext/sources/examples/talk-llama/models/jamba.cpp +0 -106
  707. data/ext/sources/examples/talk-llama/models/kimi-linear.cpp +0 -381
  708. data/ext/sources/examples/talk-llama/models/lfm2.cpp +0 -196
  709. data/ext/sources/examples/talk-llama/models/llada-moe.cpp +0 -122
  710. data/ext/sources/examples/talk-llama/models/llada.cpp +0 -99
  711. data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +0 -178
  712. data/ext/sources/examples/talk-llama/models/llama.cpp +0 -175
  713. data/ext/sources/examples/talk-llama/models/maincoder.cpp +0 -117
  714. data/ext/sources/examples/talk-llama/models/mamba-base.cpp +0 -289
  715. data/ext/sources/examples/talk-llama/models/mamba.cpp +0 -54
  716. data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +0 -129
  717. data/ext/sources/examples/talk-llama/models/minicpm3.cpp +0 -200
  718. data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +0 -123
  719. data/ext/sources/examples/talk-llama/models/mistral3.cpp +0 -160
  720. data/ext/sources/examples/talk-llama/models/models.h +0 -704
  721. data/ext/sources/examples/talk-llama/models/modern-bert.cpp +0 -109
  722. data/ext/sources/examples/talk-llama/models/mpt.cpp +0 -126
  723. data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +0 -162
  724. data/ext/sources/examples/talk-llama/models/nemotron.cpp +0 -122
  725. data/ext/sources/examples/talk-llama/models/neo-bert.cpp +0 -104
  726. data/ext/sources/examples/talk-llama/models/olmo.cpp +0 -121
  727. data/ext/sources/examples/talk-llama/models/olmo2.cpp +0 -150
  728. data/ext/sources/examples/talk-llama/models/olmoe.cpp +0 -124
  729. data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +0 -127
  730. data/ext/sources/examples/talk-llama/models/openelm.cpp +0 -124
  731. data/ext/sources/examples/talk-llama/models/orion.cpp +0 -123
  732. data/ext/sources/examples/talk-llama/models/paddleocr.cpp +0 -122
  733. data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +0 -121
  734. data/ext/sources/examples/talk-llama/models/phi2.cpp +0 -121
  735. data/ext/sources/examples/talk-llama/models/phi3.cpp +0 -152
  736. data/ext/sources/examples/talk-llama/models/plamo.cpp +0 -110
  737. data/ext/sources/examples/talk-llama/models/plamo2.cpp +0 -320
  738. data/ext/sources/examples/talk-llama/models/plamo3.cpp +0 -128
  739. data/ext/sources/examples/talk-llama/models/plm.cpp +0 -169
  740. data/ext/sources/examples/talk-llama/models/qwen.cpp +0 -108
  741. data/ext/sources/examples/talk-llama/models/qwen2.cpp +0 -126
  742. data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +0 -151
  743. data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +0 -117
  744. data/ext/sources/examples/talk-llama/models/qwen3.cpp +0 -120
  745. data/ext/sources/examples/talk-llama/models/qwen35.cpp +0 -381
  746. data/ext/sources/examples/talk-llama/models/qwen35moe.cpp +0 -422
  747. data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +0 -131
  748. data/ext/sources/examples/talk-llama/models/qwen3next.cpp +0 -525
  749. data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +0 -140
  750. data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +0 -132
  751. data/ext/sources/examples/talk-llama/models/refact.cpp +0 -94
  752. data/ext/sources/examples/talk-llama/models/rnd1.cpp +0 -126
  753. data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +0 -164
  754. data/ext/sources/examples/talk-llama/models/rwkv6.cpp +0 -94
  755. data/ext/sources/examples/talk-llama/models/rwkv6qwen2.cpp +0 -86
  756. data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +0 -137
  757. data/ext/sources/examples/talk-llama/models/rwkv7.cpp +0 -90
  758. data/ext/sources/examples/talk-llama/models/seed-oss.cpp +0 -124
  759. data/ext/sources/examples/talk-llama/models/smallthinker.cpp +0 -126
  760. data/ext/sources/examples/talk-llama/models/smollm3.cpp +0 -128
  761. data/ext/sources/examples/talk-llama/models/stablelm.cpp +0 -146
  762. data/ext/sources/examples/talk-llama/models/starcoder.cpp +0 -100
  763. data/ext/sources/examples/talk-llama/models/starcoder2.cpp +0 -121
  764. data/ext/sources/examples/talk-llama/models/step35-iswa.cpp +0 -165
  765. data/ext/sources/examples/talk-llama/models/t5-dec.cpp +0 -166
  766. data/ext/sources/examples/talk-llama/models/t5-enc.cpp +0 -96
  767. data/ext/sources/examples/talk-llama/models/wavtokenizer-dec.cpp +0 -149
  768. data/ext/sources/examples/talk-llama/models/xverse.cpp +0 -108
  769. data/ext/sources/examples/talk-llama/prompts/talk-alpaca.txt +0 -23
  770. data/ext/sources/examples/talk-llama/speak +0 -40
  771. data/ext/sources/examples/talk-llama/speak.bat +0 -1
  772. data/ext/sources/examples/talk-llama/speak.ps1 +0 -14
  773. data/ext/sources/examples/talk-llama/talk-llama.cpp +0 -813
  774. data/ext/sources/examples/talk-llama/unicode-data.cpp +0 -7034
  775. data/ext/sources/examples/talk-llama/unicode-data.h +0 -20
  776. data/ext/sources/examples/talk-llama/unicode.cpp +0 -1103
  777. data/ext/sources/examples/talk-llama/unicode.h +0 -111
  778. data/ext/sources/examples/wchess/CMakeLists.txt +0 -10
  779. data/ext/sources/examples/wchess/libwchess/CMakeLists.txt +0 -19
  780. data/ext/sources/examples/wchess/libwchess/Chessboard.cpp +0 -803
  781. data/ext/sources/examples/wchess/libwchess/Chessboard.h +0 -33
  782. data/ext/sources/examples/wchess/libwchess/WChess.cpp +0 -193
  783. data/ext/sources/examples/wchess/libwchess/WChess.h +0 -63
  784. data/ext/sources/examples/wchess/libwchess/test-chessboard.cpp +0 -117
  785. data/ext/sources/examples/wchess/wchess.cmd/CMakeLists.txt +0 -8
  786. data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +0 -253
  787. data/ext/sources/examples/whisper.wasm/CMakeLists.txt +0 -50
  788. data/ext/sources/examples/whisper.wasm/emscripten.cpp +0 -118
  789. data/ext/sources/examples/whisper.wasm/index-tmpl.html +0 -659
  790. data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +0 -99
  791. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +0 -155
  792. data/ext/sources/ggml/src/ggml-hexagon/op-desc.h +0 -153
  793. data/ext/sources/ggml/src/ggml-opencl/kernels/embed_kernel.py +0 -26
  794. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.cpp +0 -123
  795. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.h +0 -17
  796. data/ext/sources/ggml/src/ggml-virtgpu/regenerate_remoting.py +0 -333
  797. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.glsl +0 -5
  798. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +0 -182
  799. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.tmpl.wgsl +0 -323
  800. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl +0 -718
  801. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +0 -123
  802. data/ext/sources/tests/CMakeLists.txt +0 -112
  803. data/ext/sources/tests/earnings21/eval.mk +0 -58
  804. data/ext/sources/tests/earnings21/eval.py +0 -68
  805. data/ext/sources/tests/earnings21/normalizers/__init__.py +0 -2
  806. data/ext/sources/tests/earnings21/normalizers/basic.py +0 -80
  807. data/ext/sources/tests/earnings21/normalizers/english.json +0 -1741
  808. data/ext/sources/tests/earnings21/normalizers/english.py +0 -550
  809. data/ext/sources/tests/earnings21/requirements.txt +0 -6
  810. data/ext/sources/tests/en-0-ref.txt +0 -1
  811. data/ext/sources/tests/en-1-ref.txt +0 -1
  812. data/ext/sources/tests/en-2-ref.txt +0 -1
  813. data/ext/sources/tests/es-0-ref.txt +0 -1
  814. data/ext/sources/tests/librispeech/eval.mk +0 -39
  815. data/ext/sources/tests/librispeech/eval.py +0 -47
  816. data/ext/sources/tests/librispeech/normalizers/__init__.py +0 -2
  817. data/ext/sources/tests/librispeech/normalizers/basic.py +0 -80
  818. data/ext/sources/tests/librispeech/normalizers/english.json +0 -1741
  819. data/ext/sources/tests/librispeech/normalizers/english.py +0 -550
  820. data/ext/sources/tests/librispeech/requirements.txt +0 -6
  821. data/ext/sources/tests/run-tests.sh +0 -130
  822. data/ext/sources/tests/test-c.c +0 -3
  823. data/ext/sources/tests/test-vad-full.cpp +0 -56
  824. data/ext/sources/tests/test-vad.cpp +0 -83
  825. data/ext/sources/tests/test-whisper.js +0 -58
  826. data/lib/whisper/context.rb +0 -15
  827. data/lib/whisper/segment.rb +0 -58
  828. /data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general_q8_0_f32.cl → gemv_noshuffle_q8_0_f32.cl} +0 -0
data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp CHANGED
@@ -1,15 +1,67 @@
1
1
  #include "sycl_hw.hpp"
2
2
 
3
- // TODO: currently not used
4
- /*
5
- sycl_hw_info get_device_hw_info(sycl::device *device_ptr) {
6
- sycl_hw_info res;
7
- int32_t id = device_ptr->get_info<sycl::ext::intel::info::device::device_id>();
8
- res.device_id = id;
3
+ using namespace std;
9
4
 
10
- syclex::architecture arch = device_ptr->get_info<syclex::info::device::architecture>();
11
- res.arch = arch;
5
+ /*defined in
6
+ * /opt/intel/oneapi/compiler/latest/include/sycl/ext/oneapi/experimental/device_architecture.def
7
+ */
8
+ static map<gpu_arch, std::pair<const char*, sycl_intel_gpu_family>> arch2name = {
9
+ {gpu_arch::intel_gpu_bdw, {"intel_gpu_bdw", GPU_FAMILY_IGPU_NON_XE}},
10
+ {gpu_arch::intel_gpu_skl, {"intel_gpu_skl", GPU_FAMILY_IGPU_NON_XE}},
11
+ {gpu_arch::intel_gpu_kbl, {"intel_gpu_kbl", GPU_FAMILY_IGPU_NON_XE}},
12
+ {gpu_arch::intel_gpu_cfl, {"intel_gpu_cfl", GPU_FAMILY_IGPU_NON_XE}},
13
+ {gpu_arch::intel_gpu_apl, {"intel_gpu_apl", GPU_FAMILY_IGPU_NON_XE}},
14
+ {gpu_arch::intel_gpu_glk, {"intel_gpu_glk", GPU_FAMILY_IGPU_NON_XE}},
15
+ {gpu_arch::intel_gpu_whl, {"intel_gpu_whl", GPU_FAMILY_IGPU_NON_XE}},
16
+ {gpu_arch::intel_gpu_aml, {"intel_gpu_aml", GPU_FAMILY_IGPU_NON_XE}},
17
+ {gpu_arch::intel_gpu_cml, {"intel_gpu_cml", GPU_FAMILY_IGPU_NON_XE}},
18
+ {gpu_arch::intel_gpu_icllp, {"intel_gpu_icllp", GPU_FAMILY_IGPU_NON_XE}},
19
+ {gpu_arch::intel_gpu_ehl, {"intel_gpu_ehl", GPU_FAMILY_IGPU_NON_XE}},
20
+ {gpu_arch::intel_gpu_tgllp, {"intel_gpu_tgllp", GPU_FAMILY_IGPU_NON_XE}},
21
+ {gpu_arch::intel_gpu_rkl, {"intel_gpu_rkl", GPU_FAMILY_IGPU_NON_XE}},
22
+ {gpu_arch::intel_gpu_adl_s, {"intel_gpu_adl_s", GPU_FAMILY_IGPU_NON_XE}},
23
+ {gpu_arch::intel_gpu_adl_p, {"intel_gpu_adl_p", GPU_FAMILY_IGPU_NON_XE}},
24
+ {gpu_arch::intel_gpu_adl_n, {"intel_gpu_adl_n", GPU_FAMILY_IGPU_NON_XE}},
25
+ {gpu_arch::intel_gpu_dg1, {"intel_gpu_dg1", GPU_FAMILY_DGPU_CLIENT_GAME}},
26
+ {gpu_arch::intel_gpu_acm_g10, {"intel_gpu_acm_g10", GPU_FAMILY_DGPU_CLIENT_GAME}},
27
+ {gpu_arch::intel_gpu_acm_g11, {"intel_gpu_acm_g11", GPU_FAMILY_DGPU_CLIENT_GAME}},
28
+ {gpu_arch::intel_gpu_acm_g12, {"intel_gpu_acm_g12", GPU_FAMILY_DGPU_CLIENT_GAME}},
29
+ {gpu_arch::intel_gpu_pvc, {"intel_gpu_pvc", GPU_FAMILY_DGPU_CLOUD}},
30
+ {gpu_arch::intel_gpu_pvc_vg, {"intel_gpu_pvc_vg", GPU_FAMILY_DGPU_CLOUD}},
31
+ {gpu_arch::intel_gpu_mtl_u, {"intel_gpu_mtl_u", GPU_FAMILY_IGPU_XE}},
32
+ {gpu_arch::intel_gpu_mtl_h, {"intel_gpu_mtl_h", GPU_FAMILY_IGPU_XE}},
33
+ {gpu_arch::intel_gpu_arl_h, {"intel_gpu_arl_h", GPU_FAMILY_IGPU_XE}},
34
+ {gpu_arch::intel_gpu_bmg_g21, {"intel_gpu_bmg_g21", GPU_FAMILY_DGPU_CLIENT_GAME}},
35
+ {gpu_arch::intel_gpu_bmg_g31, {"intel_gpu_bmg_g31", GPU_FAMILY_DGPU_CLIENT_GAME}},
36
+ {gpu_arch::intel_gpu_lnl_m, {"intel_gpu_lnl_m", GPU_FAMILY_IGPU_XE}},
37
+ {gpu_arch::intel_gpu_ptl_h, {"intel_gpu_ptl_h", GPU_FAMILY_IGPU_XE}},
38
+ {gpu_arch::intel_gpu_ptl_u, {"intel_gpu_ptl_u", GPU_FAMILY_IGPU_XE}},
39
+ {gpu_arch::intel_gpu_wcl, {"intel_gpu_wcl", GPU_FAMILY_IGPU_XE}}
40
+ };
41
+
42
+
43
+ sycl_hw_info get_device_hw_info(sycl::device* device_ptr) {
44
+ sycl_hw_info res;
45
+ int32_t id =
46
+ device_ptr->get_info<sycl::ext::intel::info::device::device_id>();
47
+ res.device_id = id;
48
+
49
+ res.name = device_ptr->get_info<sycl::info::device::name>();
12
50
 
13
- return res;
51
+ syclex::architecture arch =
52
+ device_ptr->get_info<syclex::info::device::architecture>();
53
+ res.arch = arch;
54
+
55
+ map<syclex::architecture,
56
+ std::pair<const char*, sycl_intel_gpu_family>>::iterator it =
57
+ arch2name.find(res.arch);
58
+ if (it != arch2name.end()) {
59
+ res.arch_name = it->second.first;
60
+ res.gpu_family = it->second.second;
61
+ } else {
62
+ res.arch_name = "unknown";
63
+ res.gpu_family = GPU_FAMILY_UKNOWN;
64
+ }
65
+
66
+ return res;
14
67
  }
15
- */
data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp CHANGED
@@ -9,18 +9,30 @@
9
9
  #include <sycl/sycl.hpp>
10
10
 
11
11
  namespace syclex = sycl::ext::oneapi::experimental;
12
+ using gpu_arch = sycl::ext::oneapi::experimental::architecture;
13
+
14
+ // It's used to mark the GPU computing capacity
15
+ // The value must flow the order of performance.
16
+ enum sycl_intel_gpu_family {
17
+ GPU_FAMILY_UKNOWN = -1,
18
+ // iGPU without Xe core, before Meteor Lake iGPU(Xe)
19
+ GPU_FAMILY_IGPU_NON_XE = 0,
20
+ // iGPU with Xe core, Meteor Lake iGPU or newer.
21
+ GPU_FAMILY_IGPU_XE = 1,
22
+ // dGPU for gaming in client/data center (DG1/FLex 140 or newer).
23
+ GPU_FAMILY_DGPU_CLIENT_GAME = 2,
24
+ // dGPU for AI in cloud, PVC or newer.
25
+ GPU_FAMILY_DGPU_CLOUD = 3
26
+ };
12
27
 
13
- // TODO: currently not used
14
- /*
15
28
  struct sycl_hw_info {
16
29
  syclex::architecture arch;
30
+ const char* arch_name;
17
31
  int32_t device_id;
32
+ std::string name;
33
+ sycl_intel_gpu_family gpu_family;
18
34
  };
19
35
 
20
- bool is_in_vector(std::vector<int> &vec, int item);
21
-
22
36
  sycl_hw_info get_device_hw_info(sycl::device *device_ptr);
23
- */
24
-
25
37
 
26
38
  #endif // SYCL_HW_HPP
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp ADDED
@@ -0,0 +1,6 @@
1
+ // This file has been autogenerated by generate_cu_files.py, do not edit manually.
2
+
3
+ #include "../fattn-tile.hpp"
4
+
5
+ DECL_FATTN_TILE_CASE(512, 512);
6
+
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_F16, GGML_TYPE_F16);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q4_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q4_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_F16, GGML_TYPE_Q4_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q4_1);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q4_1);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_F16, GGML_TYPE_Q4_1);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q5_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q5_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_F16, GGML_TYPE_Q5_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q5_1);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q5_1);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_F16, GGML_TYPE_Q5_1);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q8_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_F16, GGML_TYPE_Q8_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_F16, GGML_TYPE_Q8_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_F16);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_F16);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_0, GGML_TYPE_F16);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_F16);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_F16);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_1, GGML_TYPE_F16);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_F16);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_F16);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_0, GGML_TYPE_F16);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_F16);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_F16);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_1, GGML_TYPE_F16);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_F16);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_F16);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q8_0, GGML_TYPE_F16);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp CHANGED
@@ -5,3 +5,4 @@
5
5
  DECL_FATTN_VEC_CASE( 64, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
6
6
  DECL_FATTN_VEC_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
7
7
  DECL_FATTN_VEC_CASE(256, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
8
+ DECL_FATTN_VEC_CASE(512, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
data/ext/sources/ggml/src/ggml-sycl/type.hpp ADDED
@@ -0,0 +1,112 @@
1
+ #pragma once
2
+
3
+ #include <sycl/sycl.hpp>
4
+ #include <cstdint>
5
+ #include <limits>
6
+
7
+ inline uint8_t float_to_e4m3(float f)
8
+ {
9
+ if (sycl::isnan(f)) {
10
+ return 0x7F; // Canonical NaN (positive)
11
+ }
12
+
13
+ uint32_t bits = sycl::bit_cast<uint32_t>(f);
14
+ uint32_t sign = (bits >> 31) & 0x1u;
15
+ uint32_t exp = (bits >> 23) & 0xFFu;
16
+ uint32_t mant = bits & 0x7FFFFFu;
17
+
18
+ // Zero
19
+ if (exp == 0 && mant == 0) {
20
+ return static_cast<uint8_t>(sign << 7);
21
+ }
22
+
23
+ // Extract biased exponent and mantissa for FP8
24
+ int e = static_cast<int>(exp) - 127; // true exponent (IEEE bias 127)
25
+ uint32_t m = mant;
26
+
27
+ // Handle very large values → NaN (NVIDIA behavior for E4M3)
28
+ if (e > 7) { // max exponent for E4M3 is 7 (biased 14)
29
+ return static_cast<uint8_t>((sign << 7) | 0x7F);
30
+ }
31
+
32
+ // Handle subnormals and normal numbers
33
+ if (e < -6) { // smallest normal exponent is -6
34
+ // Subnormal in FP8: shift mantissa right
35
+ int shift = -6 - e;
36
+ m = (m | 0x800000u) >> (shift + 1); // +1 because we lose the implicit 1 position
37
+ if (shift > 23) m = 0;
38
+ } else {
39
+ // Normal number: adjust exponent bias from 127 to 7
40
+ int new_exp = e + 7;
41
+ m = (m >> 20) & 0x7u; // take top 3 mantissa bits (after implicit 1)
42
+ m |= (static_cast<uint32_t>(new_exp) << 3);
43
+ }
44
+
45
+ // Round-to-nearest-even (simple guard + round bit)
46
+ // For better accuracy you can add sticky bit, but this is sufficient for most use cases
47
+ uint32_t round_bit = (mant >> 19) & 0x1u; // bit after the 3 mantissa bits
48
+ if (round_bit) {
49
+ m += 1;
50
+ // Carry into exponent if mantissa overflows
51
+ if ((m & 0x8u) != 0) {
52
+ m = (m & 0x7u) | ((m & 0x38u) << 1); // simple carry handling
53
+ // If exponent overflows after carry → NaN
54
+ if ((m >> 3) > 14) {
55
+ return static_cast<uint8_t>((sign << 7) | 0x7F);
56
+ }
57
+ }
58
+ }
59
+
60
+ uint8_t result = static_cast<uint8_t>((sign << 7) | (m & 0x7F));
61
+ return result;
62
+ }
63
+
64
+ inline float e4m3_to_float(uint8_t x)
65
+ {
66
+ if (x == 0) return 0.0f;
67
+
68
+ uint8_t sign = (x >> 7) & 0x1u;
69
+ uint8_t exp = (x >> 3) & 0xFu;
70
+ uint8_t mant = x & 0x7u;
71
+
72
+ // NaN (NVIDIA uses 0x7F / 0xFF as NaN)
73
+ if (exp == 0xF && mant != 0) {
74
+ return std::numeric_limits<float>::quiet_NaN();
75
+ }
76
+ if (exp == 0xF) { // 0x7F or 0xFF treated as NaN
77
+ return std::numeric_limits<float>::quiet_NaN();
78
+ }
79
+
80
+ float val;
81
+
82
+ if (exp == 0) {
83
+ // Subnormal
84
+ val = mant * (1.0f / 8.0f) * sycl::pow(2.0f, -6.0f);
85
+ } else {
86
+ // Normal: implicit leading 1 + bias 7
87
+ val = (1.0f + mant / 8.0f) * sycl::pow(2.0f, static_cast<float>(exp) - 7.0f);
88
+ }
89
+
90
+ return sign ? -val : val;
91
+ }
92
+
93
+ // The actual type definition
94
+ struct __nv_fp8_e4m3 {
95
+ uint8_t raw;
96
+
97
+ __nv_fp8_e4m3() = default;
98
+
99
+ explicit __nv_fp8_e4m3(float f) : raw(float_to_e4m3(f)) {}
100
+ explicit __nv_fp8_e4m3(sycl::half h) : raw(float_to_e4m3(static_cast<float>(h))) {}
101
+
102
+ operator float() const { return e4m3_to_float(raw); }
103
+ operator sycl::half() const { return static_cast<sycl::half>(static_cast<float>(*this)); }
104
+
105
+ // Allow direct access for vector loads/stores
106
+ operator uint8_t&() { return raw; }
107
+ operator uint8_t() const { return raw; }
108
+ };
109
+
110
+ using __nv_fp8x2_e4m3 = sycl::vec<__nv_fp8_e4m3, 2>;
111
+ using __nv_fp8x4_e4m3 = sycl::vec<__nv_fp8_e4m3, 4>;
112
+