whispercpp 1.3.6 → 1.3.8

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 (965) hide show
  1. checksums.yaml +4 -4
  2. data/.document +3 -0
  3. data/.rdoc_options +2 -0
  4. data/README.md +43 -9
  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 +35 -42
  12. data/ext/ruby_whisper.h +141 -0
  13. data/ext/ruby_whisper_context.c +157 -29
  14. data/ext/ruby_whisper_log_queue.c +180 -0
  15. data/ext/ruby_whisper_log_settable.h +46 -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 +265 -73
  25. data/ext/ruby_whisper_segment.c +6 -6
  26. data/ext/ruby_whisper_transcribe.cpp +23 -15
  27. data/ext/ruby_whisper_vad_context.c +30 -10
  28. data/ext/ruby_whisper_vad_context_detect.cpp +8 -9
  29. data/ext/ruby_whisper_vad_params.c +4 -4
  30. data/ext/ruby_whisper_vad_segment.c +2 -2
  31. data/ext/sources/CMakeLists.txt +42 -3
  32. data/ext/sources/CMakePresets.json +95 -0
  33. data/ext/sources/cmake/parakeet-config.cmake.in +30 -0
  34. data/ext/sources/cmake/parakeet.pc.in +10 -0
  35. data/ext/sources/cmake/whisper.pc.in +2 -2
  36. data/ext/sources/examples/CMakeLists.txt +4 -2
  37. data/ext/sources/examples/bench/bench.cpp +1 -1
  38. data/ext/sources/examples/cli/cli.cpp +52 -10
  39. data/ext/sources/examples/common-ggml.cpp +4 -0
  40. data/ext/sources/examples/common-whisper.cpp +139 -67
  41. data/ext/sources/examples/common-whisper.h +11 -0
  42. data/ext/sources/examples/ffmpeg-transcode.cpp +211 -341
  43. data/ext/sources/examples/parakeet-cli/CMakeLists.txt +8 -0
  44. data/ext/sources/examples/parakeet-cli/parakeet-cli.cpp +243 -0
  45. data/ext/sources/examples/parakeet-quantize/CMakeLists.txt +7 -0
  46. data/ext/sources/examples/parakeet-quantize/parakeet-quantize.cpp +230 -0
  47. data/ext/sources/examples/server/server.cpp +199 -163
  48. data/ext/sources/examples/vad-speech-segments/speech.cpp +3 -2
  49. data/ext/sources/ggml/CMakeLists.txt +21 -14
  50. data/ext/sources/ggml/cmake/FindNCCL.cmake +36 -0
  51. data/ext/sources/ggml/cmake/ggml-config.cmake.in +12 -2
  52. data/ext/sources/ggml/include/ggml-alloc.h +1 -0
  53. data/ext/sources/ggml/include/ggml-backend.h +72 -10
  54. data/ext/sources/ggml/include/ggml-cuda.h +2 -2
  55. data/ext/sources/ggml/include/ggml-rpc.h +3 -3
  56. data/ext/sources/ggml/include/ggml-sycl.h +8 -0
  57. data/ext/sources/ggml/include/ggml.h +103 -9
  58. data/ext/sources/ggml/include/gguf.h +10 -2
  59. data/ext/sources/ggml/src/CMakeLists.txt +30 -6
  60. data/ext/sources/ggml/src/ggml-alloc.c +5 -1
  61. data/ext/sources/ggml/src/ggml-backend-impl.h +22 -2
  62. data/ext/sources/ggml/src/ggml-backend-meta.cpp +2266 -0
  63. data/ext/sources/ggml/src/ggml-backend-reg.cpp +12 -0
  64. data/ext/sources/ggml/src/ggml-backend.cpp +110 -9
  65. data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +4 -0
  66. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +672 -257
  67. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +71 -0
  68. data/ext/sources/ggml/src/ggml-cann/common.h +20 -10
  69. data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +211 -30
  70. data/ext/sources/ggml/src/ggml-common.h +24 -2
  71. data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +59 -30
  72. data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +2 -0
  73. data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +21 -22
  74. data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +194 -11
  75. data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +65 -0
  76. data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +151 -1
  77. data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +0 -1
  78. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +4279 -1292
  79. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +5 -35
  80. data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +0 -1
  81. data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +72 -1
  82. data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +319 -31
  83. data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +1 -1
  84. data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +12 -2
  85. data/ext/sources/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
  86. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +10 -0
  87. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +109 -5
  88. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +2 -0
  89. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +146 -134
  90. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +107 -82
  91. data/ext/sources/ggml/src/ggml-cpu/ops.cpp +501 -119
  92. data/ext/sources/ggml/src/ggml-cpu/ops.h +3 -0
  93. data/ext/sources/ggml/src/ggml-cpu/quants.c +106 -0
  94. data/ext/sources/ggml/src/ggml-cpu/quants.h +6 -0
  95. data/ext/sources/ggml/src/ggml-cpu/repack.cpp +3 -0
  96. data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +91 -1
  97. data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +14 -16
  98. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +1402 -687
  99. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.h +8 -0
  100. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +597 -2766
  101. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp +5768 -0
  102. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.cpp +320 -0
  103. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.h +55 -0
  104. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_kernels.h +182 -19
  105. data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.cpp +1795 -0
  106. data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.h +14 -0
  107. data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp +3178 -0
  108. data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.h +95 -0
  109. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_barrier.h +34 -0
  110. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp +760 -0
  111. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h +32 -0
  112. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_tcm.h +409 -0
  113. data/ext/sources/ggml/src/ggml-cpu/vec.cpp +39 -55
  114. data/ext/sources/ggml/src/ggml-cpu/vec.h +225 -240
  115. data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +17 -7
  116. data/ext/sources/ggml/src/ggml-cuda/allreduce.cu +971 -0
  117. data/ext/sources/ggml/src/ggml-cuda/allreduce.cuh +29 -0
  118. data/ext/sources/ggml/src/ggml-cuda/argsort.cu +62 -26
  119. data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +134 -64
  120. data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +1 -0
  121. data/ext/sources/ggml/src/ggml-cuda/col2im-1d.cu +81 -0
  122. data/ext/sources/ggml/src/ggml-cuda/col2im-1d.cuh +3 -0
  123. data/ext/sources/ggml/src/ggml-cuda/common.cuh +246 -28
  124. data/ext/sources/ggml/src/ggml-cuda/concat.cu +134 -116
  125. data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cu +14 -12
  126. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cu +45 -21
  127. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cuh +1 -0
  128. data/ext/sources/ggml/src/ggml-cuda/convert.cu +139 -34
  129. data/ext/sources/ggml/src/ggml-cuda/convert.cuh +10 -0
  130. data/ext/sources/ggml/src/ggml-cuda/cpy.cu +88 -29
  131. data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +22 -0
  132. data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +287 -49
  133. data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +335 -130
  134. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +12 -0
  135. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +127 -24
  136. data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +40 -15
  137. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +18 -9
  138. data/ext/sources/ggml/src/ggml-cuda/fattn.cu +169 -60
  139. data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +2 -0
  140. data/ext/sources/ggml/src/ggml-cuda/fwht.cu +101 -0
  141. data/ext/sources/ggml/src/ggml-cuda/fwht.cuh +4 -0
  142. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cu +109 -45
  143. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cuh +10 -0
  144. data/ext/sources/ggml/src/ggml-cuda/getrows.cu +48 -23
  145. data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +2034 -2104
  146. data/ext/sources/ggml/src/ggml-cuda/im2col.cu +32 -29
  147. data/ext/sources/ggml/src/ggml-cuda/mean.cu +4 -2
  148. data/ext/sources/ggml/src/ggml-cuda/mma.cuh +242 -195
  149. data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +3 -3
  150. data/ext/sources/ggml/src/ggml-cuda/mmq.cu +25 -12
  151. data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +502 -423
  152. data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +19 -12
  153. data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +562 -97
  154. data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +6 -1
  155. data/ext/sources/ggml/src/ggml-cuda/norm.cu +36 -10
  156. data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +66 -7
  157. data/ext/sources/ggml/src/ggml-cuda/quantize.cu +133 -26
  158. data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +1 -1
  159. data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +5 -1
  160. data/ext/sources/ggml/src/ggml-cuda/rope.cu +11 -4
  161. data/ext/sources/ggml/src/ggml-cuda/scale.cu +4 -1
  162. data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +78 -10
  163. data/ext/sources/ggml/src/ggml-cuda/snake.cu +72 -0
  164. data/ext/sources/ggml/src/ggml-cuda/snake.cuh +8 -0
  165. data/ext/sources/ggml/src/ggml-cuda/softcap.cu +4 -1
  166. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +45 -13
  167. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +1 -1
  168. data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +40 -18
  169. data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +8 -4
  170. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +1 -0
  171. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +1 -0
  172. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +2 -0
  173. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu +1 -0
  174. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +1 -0
  175. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +1 -0
  176. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +1 -0
  177. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +1 -0
  178. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +2 -0
  179. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu +1 -0
  180. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +1 -0
  181. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu +1 -0
  182. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +1 -0
  183. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +2 -0
  184. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu +1 -0
  185. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +1 -0
  186. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +2 -0
  187. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu +5 -0
  188. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu +5 -0
  189. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +5 -0
  190. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu +7 -0
  191. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu +7 -0
  192. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu +7 -0
  193. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu +7 -0
  194. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu +7 -0
  195. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu +7 -0
  196. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu +7 -0
  197. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu +7 -0
  198. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu +7 -0
  199. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu +7 -0
  200. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu +7 -0
  201. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu +7 -0
  202. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu +7 -0
  203. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu +5 -0
  204. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu +5 -0
  205. data/ext/sources/ggml/src/ggml-cuda/top-k.cu +5 -4
  206. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +33 -24
  207. data/ext/sources/ggml/src/ggml-cuda/unary.cu +31 -2
  208. data/ext/sources/ggml/src/ggml-cuda/unary.cuh +2 -0
  209. data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +80 -0
  210. data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +7 -2
  211. data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +23 -4
  212. data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +4 -0
  213. data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +1 -5
  214. data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +2788 -1762
  215. data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +13 -4
  216. data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +53 -84
  217. data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +25 -12
  218. data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +165 -184
  219. data/ext/sources/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +17 -19
  220. data/ext/sources/ggml/src/ggml-hexagon/htp/concat-ops.c +277 -0
  221. data/ext/sources/ggml/src/ggml-hexagon/htp/cpy-ops.c +170 -127
  222. data/ext/sources/ggml/src/ggml-hexagon/htp/cumsum-ops.c +270 -0
  223. data/ext/sources/ggml/src/ggml-hexagon/htp/diag-ops.c +216 -0
  224. data/ext/sources/ggml/src/ggml-hexagon/htp/fill-ops.c +123 -0
  225. data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +1774 -396
  226. data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.h +303 -0
  227. data/ext/sources/ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c +1148 -0
  228. data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +148 -42
  229. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-common.h +80 -0
  230. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.c +2 -2
  231. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.h +255 -62
  232. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +9 -0
  233. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-profile.h +64 -0
  234. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +25 -21
  235. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-fa-kernels.h +555 -0
  236. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-mm-kernels-tiled.h +1303 -0
  237. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.c +167 -0
  238. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.h +157 -0
  239. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-utils.h +222 -0
  240. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +104 -13
  241. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +222 -57
  242. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-vtcm.h +19 -0
  243. data/ext/sources/ggml/src/ggml-hexagon/htp/htp_iface.idl +10 -3
  244. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +78 -26
  245. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +27 -10
  246. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +63 -23
  247. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +48 -8
  248. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-fa-kernels.h +232 -0
  249. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -0
  250. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-log.h +65 -0
  251. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-mm-kernels-flat.h +1511 -0
  252. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-mm-kernels-tiled.h +1200 -0
  253. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-pow.h +42 -0
  254. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-repl.h +74 -0
  255. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +40 -0
  256. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
  257. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +5 -8
  258. data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +625 -816
  259. data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +3052 -2166
  260. data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.h +650 -0
  261. data/ext/sources/ggml/src/ggml-hexagon/htp/pad-ops.c +547 -0
  262. data/ext/sources/ggml/src/ggml-hexagon/htp/repeat-ops.c +148 -0
  263. data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +337 -106
  264. data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +59 -37
  265. data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +121 -133
  266. data/ext/sources/ggml/src/ggml-hexagon/htp/solve-tri-ops.c +267 -0
  267. data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c +245 -151
  268. data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +6 -6
  269. data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +719 -45
  270. data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.c +15 -3
  271. data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.h +8 -0
  272. data/ext/sources/ggml/src/ggml-hexagon/htp-opnode.h +390 -0
  273. data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +3 -5
  274. data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +27 -9
  275. data/ext/sources/ggml/src/ggml-impl.h +6 -1
  276. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +207 -18
  277. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +36 -2
  278. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +186 -29
  279. data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +118 -0
  280. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +322 -21
  281. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +4 -0
  282. data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +39 -26
  283. data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +1226 -467
  284. data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +5 -6
  285. data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +67 -5
  286. data/ext/sources/ggml/src/ggml-opencl/fa_tune.h +92 -0
  287. data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +16290 -6246
  288. data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +67 -0
  289. data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +59 -0
  290. data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +1997 -92
  291. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl +81 -41
  292. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +88 -39
  293. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl +1995 -96
  294. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32_q4_0.cl +1615 -0
  295. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32_q8_0.cl +1486 -0
  296. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_pre_f16.cl +156 -0
  297. data/ext/sources/ggml/src/ggml-opencl/kernels/gated_delta_net.cl +249 -0
  298. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl +374 -0
  299. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl +324 -0
  300. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl +326 -0
  301. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl +348 -0
  302. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl +328 -0
  303. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl +330 -0
  304. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl +356 -0
  305. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl +335 -0
  306. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_f32.cl +150 -0
  307. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q1_0_f32.cl +94 -0
  308. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mat_Ab_Bi_8x4.cl → gemm_noshuffle_q4_0_f32.cl} +1 -1
  309. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl +172 -0
  310. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl +131 -0
  311. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl +134 -0
  312. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl +176 -0
  313. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl +140 -0
  314. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mm_q8_0_f32_8x4.cl → gemm_noshuffle_q8_0_f32.cl} +1 -1
  315. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_xmem_f16_f32_os8.cl +233 -0
  316. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl +165 -0
  317. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl +120 -0
  318. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl +123 -0
  319. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl +155 -0
  320. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl +123 -0
  321. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl +125 -0
  322. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl +160 -0
  323. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl +141 -0
  324. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +302 -0
  325. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q1_0_f32.cl +121 -0
  326. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general.cl → gemv_noshuffle_q4_0_f32.cl} +5 -5
  327. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle.cl → gemv_noshuffle_q4_0_f32_spec.cl} +5 -5
  328. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +318 -0
  329. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +291 -0
  330. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +294 -0
  331. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +326 -0
  332. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +293 -0
  333. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general_q8_0_f32.cl → gemv_noshuffle_q8_0_f32.cl} +1 -1
  334. data/ext/sources/ggml/src/ggml-opencl/kernels/get_rows.cl +15 -9
  335. data/ext/sources/ggml/src/ggml-opencl/kernels/moe_reorder_b.cl +30 -0
  336. data/ext/sources/ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl +82 -0
  337. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_iq4_nl_f32_l4_lm.cl +171 -0
  338. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q1_0_f32_l4_lm.cl +156 -0
  339. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl +179 -0
  340. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_0_f32_l4_lm.cl +173 -0
  341. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_1_f32_l4_lm.cl +175 -0
  342. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_k_f32_l4_lm.cl +192 -0
  343. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl +1149 -0
  344. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32.cl +164 -0
  345. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32_flat.cl +202 -0
  346. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q1_0_f32.cl +141 -0
  347. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q1_0_f32_flat.cl +190 -0
  348. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl +196 -0
  349. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32.cl +241 -0
  350. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32_flat.cl +243 -0
  351. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32.cl +243 -0
  352. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32_flat.cl +247 -0
  353. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32.cl +187 -0
  354. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32_flat.cl +203 -0
  355. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +48 -64
  356. data/ext/sources/ggml/src/ggml-opencl/kernels/norm.cl +5 -2
  357. data/ext/sources/ggml/src/ggml-opencl/kernels/set_rows.cl +500 -0
  358. data/ext/sources/ggml/src/ggml-opencl/libdl.h +79 -0
  359. data/ext/sources/ggml/src/ggml-openvino/.clang-format +0 -5
  360. data/ext/sources/ggml/src/ggml-openvino/CMakeLists.txt +2 -4
  361. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +740 -127
  362. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.h +76 -23
  363. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +75 -14
  364. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.h +29 -8
  365. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +339 -69
  366. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +330 -192
  367. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.h +10 -4
  368. data/ext/sources/ggml/src/ggml-openvino/openvino/decoder.h +56 -16
  369. data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.h +1 -1
  370. data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.h +4 -4
  371. data/ext/sources/ggml/src/ggml-openvino/openvino/node_context.h +94 -37
  372. data/ext/sources/ggml/src/ggml-openvino/openvino/op/add_id.cpp +76 -0
  373. data/ext/sources/ggml/src/ggml-openvino/openvino/op/argsort.cpp +47 -0
  374. data/ext/sources/ggml/src/ggml-openvino/openvino/op/clamp.cpp +33 -0
  375. data/ext/sources/ggml/src/ggml-openvino/openvino/op/concat.cpp +48 -0
  376. data/ext/sources/ggml/src/ggml-openvino/openvino/op/cont.cpp +8 -16
  377. data/ext/sources/ggml/src/ggml-openvino/openvino/op/cpy.cpp +14 -1
  378. data/ext/sources/ggml/src/ggml-openvino/openvino/op/div.cpp +146 -0
  379. data/ext/sources/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +108 -21
  380. data/ext/sources/ggml/src/ggml-openvino/openvino/op/gated_delta_net.cpp +282 -0
  381. data/ext/sources/ggml/src/ggml-openvino/openvino/op/gated_delta_net.hpp +65 -0
  382. data/ext/sources/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +2 -9
  383. data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +21 -7
  384. data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +41 -8
  385. data/ext/sources/ggml/src/ggml-openvino/openvino/op/im2col.cpp +120 -0
  386. data/ext/sources/ggml/src/ggml-openvino/openvino/op/l2_norm.cpp +44 -0
  387. data/ext/sources/ggml/src/ggml-openvino/openvino/op/mul_mat_id.cpp +226 -0
  388. data/ext/sources/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +19 -9
  389. data/ext/sources/ggml/src/ggml-openvino/openvino/op/norm.cpp +58 -0
  390. data/ext/sources/ggml/src/ggml-openvino/openvino/op/pad.cpp +95 -0
  391. data/ext/sources/ggml/src/ggml-openvino/openvino/op/permute.cpp +58 -13
  392. data/ext/sources/ggml/src/ggml-openvino/openvino/op/repeat.cpp +74 -0
  393. data/ext/sources/ggml/src/ggml-openvino/openvino/op/reshape.cpp +13 -6
  394. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +1 -1
  395. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +161 -39
  396. data/ext/sources/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +3 -3
  397. data/ext/sources/ggml/src/ggml-openvino/openvino/op/softmax.cpp +126 -49
  398. data/ext/sources/ggml/src/ggml-openvino/openvino/op/ssm_conv.cpp +59 -0
  399. data/ext/sources/ggml/src/ggml-openvino/openvino/op/sum_rows.cpp +27 -0
  400. data/ext/sources/ggml/src/ggml-openvino/openvino/op/transpose.cpp +32 -1
  401. data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +1 -1
  402. data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_softplus.cpp +38 -0
  403. data/ext/sources/ggml/src/ggml-openvino/openvino/op/view.cpp +90 -25
  404. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +41 -22
  405. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +18 -4
  406. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +1 -1
  407. data/ext/sources/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp +41 -0
  408. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.cpp +70 -43
  409. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.h +5 -4
  410. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +612 -36
  411. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +29 -26
  412. data/ext/sources/ggml/src/ggml-openvino/utils.cpp +460 -114
  413. data/ext/sources/ggml/src/ggml-openvino/utils.h +32 -9
  414. data/ext/sources/ggml/src/ggml-opt.cpp +1 -0
  415. data/ext/sources/ggml/src/ggml-quants.c +365 -114
  416. data/ext/sources/ggml/src/ggml-quants.h +6 -0
  417. data/ext/sources/ggml/src/ggml-rpc/CMakeLists.txt +24 -0
  418. data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +167 -311
  419. data/ext/sources/ggml/src/ggml-rpc/transport.cpp +683 -0
  420. data/ext/sources/ggml/src/ggml-rpc/transport.h +34 -0
  421. data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +50 -4
  422. data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +1 -1
  423. data/ext/sources/ggml/src/ggml-sycl/backend.hpp +5 -1
  424. data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +12 -0
  425. data/ext/sources/ggml/src/ggml-sycl/col2im-1d.cpp +102 -0
  426. data/ext/sources/ggml/src/ggml-sycl/col2im-1d.hpp +8 -0
  427. data/ext/sources/ggml/src/ggml-sycl/common.cpp +72 -2
  428. data/ext/sources/ggml/src/ggml-sycl/common.hpp +59 -2
  429. data/ext/sources/ggml/src/ggml-sycl/concat.cpp +21 -1
  430. data/ext/sources/ggml/src/ggml-sycl/conv2d-dw.cpp +158 -0
  431. data/ext/sources/ggml/src/ggml-sycl/conv2d-dw.hpp +10 -0
  432. data/ext/sources/ggml/src/ggml-sycl/conv2d-transpose.cpp +125 -0
  433. data/ext/sources/ggml/src/ggml-sycl/conv2d-transpose.hpp +10 -0
  434. data/ext/sources/ggml/src/ggml-sycl/conv2d.cpp +150 -0
  435. data/ext/sources/ggml/src/ggml-sycl/conv2d.hpp +10 -0
  436. data/ext/sources/ggml/src/ggml-sycl/conv3d.cpp +224 -0
  437. data/ext/sources/ggml/src/ggml-sycl/conv3d.hpp +8 -0
  438. data/ext/sources/ggml/src/ggml-sycl/convert.cpp +121 -13
  439. data/ext/sources/ggml/src/ggml-sycl/convert.hpp +9 -0
  440. data/ext/sources/ggml/src/ggml-sycl/cpy.cpp +706 -0
  441. data/ext/sources/ggml/src/ggml-sycl/cpy.hpp +281 -0
  442. data/ext/sources/ggml/src/ggml-sycl/cross_entropy_loss.cpp +255 -0
  443. data/ext/sources/ggml/src/ggml-sycl/cross_entropy_loss.hpp +7 -0
  444. data/ext/sources/ggml/src/ggml-sycl/cumsum.cpp +148 -0
  445. data/ext/sources/ggml/src/ggml-sycl/cumsum.hpp +5 -0
  446. data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +678 -0
  447. data/ext/sources/ggml/src/ggml-sycl/diag.cpp +67 -0
  448. data/ext/sources/ggml/src/ggml-sycl/diag.hpp +5 -0
  449. data/ext/sources/ggml/src/ggml-sycl/dmmv.cpp +997 -244
  450. data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +15 -7
  451. data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +215 -204
  452. data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -2
  453. data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.cpp +56 -0
  454. data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.hpp +63 -0
  455. data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp +7 -5
  456. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +4 -0
  457. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +76 -168
  458. data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +7 -0
  459. data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +3 -1
  460. data/ext/sources/ggml/src/ggml-sycl/fill.cpp +55 -0
  461. data/ext/sources/ggml/src/ggml-sycl/fill.hpp +5 -0
  462. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.cpp +69 -31
  463. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +1 -0
  464. data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +3 -0
  465. data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +79 -3
  466. data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +1758 -455
  467. data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +353 -89
  468. data/ext/sources/ggml/src/ggml-sycl/im2col.hpp +5 -3
  469. data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +1542 -39
  470. data/ext/sources/ggml/src/ggml-sycl/mmvq.hpp +33 -0
  471. data/ext/sources/ggml/src/ggml-sycl/norm.cpp +103 -49
  472. data/ext/sources/ggml/src/ggml-sycl/outprod.cpp +45 -9
  473. data/ext/sources/ggml/src/ggml-sycl/pad.cpp +27 -27
  474. data/ext/sources/ggml/src/ggml-sycl/pool.cpp +185 -0
  475. data/ext/sources/ggml/src/ggml-sycl/pool.hpp +22 -0
  476. data/ext/sources/ggml/src/ggml-sycl/presets.hpp +3 -1
  477. data/ext/sources/ggml/src/ggml-sycl/quants.hpp +71 -0
  478. data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +17 -3
  479. data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +9 -10
  480. data/ext/sources/ggml/src/ggml-sycl/solve_tri.cpp +172 -0
  481. data/ext/sources/ggml/src/ggml-sycl/solve_tri.hpp +8 -0
  482. data/ext/sources/ggml/src/ggml-sycl/ssm_conv.cpp +6 -1
  483. data/ext/sources/ggml/src/ggml-sycl/ssm_scan.cpp +156 -0
  484. data/ext/sources/ggml/src/ggml-sycl/ssm_scan.hpp +5 -0
  485. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +62 -10
  486. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +18 -6
  487. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp +6 -0
  488. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +1 -0
  489. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +1 -0
  490. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +1 -0
  491. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +1 -0
  492. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +1 -0
  493. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +1 -0
  494. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +1 -0
  495. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +1 -0
  496. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +1 -0
  497. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +1 -0
  498. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +1 -0
  499. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +1 -0
  500. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +1 -0
  501. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +1 -0
  502. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +1 -0
  503. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +1 -0
  504. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +1 -0
  505. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +1 -0
  506. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +1 -0
  507. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +1 -0
  508. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +1 -0
  509. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +1 -0
  510. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +1 -0
  511. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +1 -0
  512. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +1 -0
  513. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +1 -0
  514. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +1 -0
  515. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +1 -0
  516. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +1 -0
  517. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +1 -0
  518. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +1 -0
  519. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +1 -0
  520. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +1 -0
  521. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +1 -0
  522. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +1 -0
  523. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +1 -0
  524. data/ext/sources/ggml/src/ggml-sycl/type.hpp +112 -0
  525. data/ext/sources/ggml/src/ggml-sycl/upscale.cpp +410 -0
  526. data/ext/sources/ggml/src/ggml-sycl/upscale.hpp +9 -0
  527. data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +242 -45
  528. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +4 -0
  529. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +2 -0
  530. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +2 -0
  531. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +1 -0
  532. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +1 -0
  533. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +0 -2
  534. data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +16 -0
  535. data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +2843 -700
  536. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +4 -0
  537. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/col2im_1d.comp +61 -0
  538. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +6 -2
  539. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +146 -13
  540. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv3d_mm.comp +431 -0
  541. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +3 -1
  542. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +1 -1
  543. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +25 -1
  544. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +88 -0
  545. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl +643 -1
  546. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp +32 -0
  547. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q1_0.comp +29 -0
  548. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +3 -4
  549. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl +27 -0
  550. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2_decode_vector.comp +7 -0
  551. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +198 -48
  552. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +60 -59
  553. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +116 -113
  554. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +122 -31
  555. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_dequant.glsl +131 -0
  556. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mmq_funcs.glsl +203 -0
  557. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp +115 -0
  558. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +125 -64
  559. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl +0 -1
  560. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.glsl +21 -19
  561. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_back.comp +25 -0
  562. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl +29 -1
  563. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl +17 -11
  564. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +76 -54
  565. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +0 -1
  566. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +4 -7
  567. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +0 -1
  568. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +122 -27
  569. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +6 -6
  570. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +1 -1
  571. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +1 -1
  572. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +1 -1
  573. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +22 -24
  574. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +88 -55
  575. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +42 -40
  576. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +49 -15
  577. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +222 -171
  578. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +8 -8
  579. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +24 -9
  580. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +0 -1
  581. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/norm.comp +10 -10
  582. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp +3 -3
  583. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/roll.comp +3 -3
  584. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +5 -2
  585. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +0 -1
  586. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +3 -2
  587. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/snake.comp +49 -0
  588. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +11 -1
  589. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +3 -4
  590. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +79 -2
  591. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/unary.comp +168 -0
  592. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +282 -211
  593. data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +5 -2
  594. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +2209 -283
  595. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +2618 -1416
  596. data/ext/sources/ggml/src/ggml-webgpu/pre_wgsl.hpp +37 -7
  597. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_id.wgsl +64 -0
  598. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +8 -7
  599. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +90 -95
  600. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +19 -1
  601. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/conv2d.wgsl +165 -0
  602. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{cpy.tmpl.wgsl → cpy.wgsl} +25 -50
  603. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +107 -184
  604. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_quant_staging.tmpl +124 -0
  605. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl +397 -0
  606. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_blk.wgsl +101 -0
  607. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_reduce.wgsl +84 -0
  608. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl +619 -0
  609. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl +149 -0
  610. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl +204 -78
  611. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl +155 -0
  612. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/im2col.wgsl +101 -0
  613. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +805 -526
  614. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id.wgsl +195 -0
  615. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_gather.wgsl +52 -0
  616. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl +154 -0
  617. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +8 -6
  618. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.wgsl +5 -1
  619. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +90 -413
  620. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl +1553 -0
  621. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl +297 -0
  622. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl +21 -0
  623. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl +178 -0
  624. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl +152 -0
  625. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{rope.tmpl.wgsl → rope.wgsl} +71 -142
  626. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/row_norm.wgsl +153 -0
  627. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/scale.wgsl +6 -4
  628. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set.wgsl +109 -0
  629. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +2 -3
  630. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows_quant.wgsl +224 -0
  631. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{soft_max.tmpl.wgsl → soft_max.wgsl} +106 -206
  632. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/solve_tri.wgsl +121 -0
  633. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_conv.wgsl +65 -0
  634. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl +193 -0
  635. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +68 -48
  636. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
  637. data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +18 -14
  638. data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +1 -1
  639. data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +244 -10
  640. data/ext/sources/ggml/src/ggml.c +146 -42
  641. data/ext/sources/ggml/src/gguf.cpp +173 -28
  642. data/ext/sources/include/parakeet.h +342 -0
  643. data/ext/sources/include/whisper.h +31 -0
  644. data/ext/sources/media/matmul.png +0 -0
  645. data/ext/sources/src/CMakeLists.txt +23 -0
  646. data/ext/sources/src/parakeet-arch.h +188 -0
  647. data/ext/sources/src/parakeet.cpp +3838 -0
  648. data/ext/sources/src/whisper.cpp +220 -26
  649. data/extsources.rb +26 -10
  650. data/lib/whisper/log_settable.rb +33 -0
  651. data/lib/whisper/model/uri.rb +13 -8
  652. data/lib/whisper/output.rb +74 -0
  653. data/sig/whisper.rbs +417 -62
  654. data/test/helper.rb +2 -0
  655. data/test/jfk_reader/jfk_reader.c +50 -7
  656. data/test/test_callback.rb +1 -0
  657. data/test/test_package.rb +6 -5
  658. data/test/test_parakeet.rb +28 -0
  659. data/test/test_parakeet_callback.rb +107 -0
  660. data/test/test_parakeet_context.rb +116 -0
  661. data/test/test_parakeet_context_params.rb +24 -0
  662. data/test/test_parakeet_model.rb +21 -0
  663. data/test/test_parakeet_params.rb +78 -0
  664. data/test/test_parakeet_segment.rb +42 -0
  665. data/test/test_parakeet_token.rb +73 -0
  666. data/test/test_params.rb +2 -0
  667. data/test/test_vad.rb +9 -0
  668. data/test/test_vad_context.rb +2 -2
  669. data/test/test_vad_segment.rb +1 -1
  670. data/test/test_whisper.rb +24 -6
  671. data/whispercpp.gemspec +2 -2
  672. metadata +263 -304
  673. data/ext/sources/bindings/javascript/CMakeLists.txt +0 -41
  674. data/ext/sources/bindings/javascript/emscripten.cpp +0 -93
  675. data/ext/sources/bindings/javascript/libwhisper.worker.js +0 -1
  676. data/ext/sources/bindings/javascript/package.json +0 -26
  677. data/ext/sources/bindings/javascript/whisper.js +0 -19
  678. data/ext/sources/examples/addon.node/CMakeLists.txt +0 -31
  679. data/ext/sources/examples/addon.node/__test__/whisper.spec.js +0 -133
  680. data/ext/sources/examples/addon.node/addon.cpp +0 -557
  681. data/ext/sources/examples/addon.node/index.js +0 -59
  682. data/ext/sources/examples/addon.node/package.json +0 -16
  683. data/ext/sources/examples/addon.node/vad-example.js +0 -132
  684. data/ext/sources/examples/bench.wasm/CMakeLists.txt +0 -49
  685. data/ext/sources/examples/bench.wasm/emscripten.cpp +0 -87
  686. data/ext/sources/examples/bench.wasm/index-tmpl.html +0 -285
  687. data/ext/sources/examples/coi-serviceworker.js +0 -146
  688. data/ext/sources/examples/command/CMakeLists.txt +0 -10
  689. data/ext/sources/examples/command/command.cpp +0 -802
  690. data/ext/sources/examples/command/commands.txt +0 -9
  691. data/ext/sources/examples/command.wasm/CMakeLists.txt +0 -50
  692. data/ext/sources/examples/command.wasm/emscripten.cpp +0 -327
  693. data/ext/sources/examples/command.wasm/index-tmpl.html +0 -415
  694. data/ext/sources/examples/generate-karaoke.sh +0 -57
  695. data/ext/sources/examples/helpers.js +0 -191
  696. data/ext/sources/examples/livestream.sh +0 -112
  697. data/ext/sources/examples/lsp/CMakeLists.txt +0 -10
  698. data/ext/sources/examples/lsp/lsp.cpp +0 -471
  699. data/ext/sources/examples/lsp/whisper.vim +0 -362
  700. data/ext/sources/examples/python/test_whisper_processor.py +0 -7
  701. data/ext/sources/examples/python/whisper_processor.py +0 -54
  702. data/ext/sources/examples/server/bench.js +0 -29
  703. data/ext/sources/examples/server.py +0 -120
  704. data/ext/sources/examples/stream/CMakeLists.txt +0 -10
  705. data/ext/sources/examples/stream/stream.cpp +0 -437
  706. data/ext/sources/examples/stream.wasm/CMakeLists.txt +0 -49
  707. data/ext/sources/examples/stream.wasm/emscripten.cpp +0 -216
  708. data/ext/sources/examples/stream.wasm/index-tmpl.html +0 -491
  709. data/ext/sources/examples/sycl/CMakeLists.txt +0 -9
  710. data/ext/sources/examples/sycl/build.sh +0 -22
  711. data/ext/sources/examples/sycl/ls-sycl-device.cpp +0 -11
  712. data/ext/sources/examples/sycl/run-whisper.sh +0 -17
  713. data/ext/sources/examples/talk-llama/CMakeLists.txt +0 -48
  714. data/ext/sources/examples/talk-llama/eleven-labs.py +0 -80
  715. data/ext/sources/examples/talk-llama/llama-adapter.cpp +0 -488
  716. data/ext/sources/examples/talk-llama/llama-adapter.h +0 -89
  717. data/ext/sources/examples/talk-llama/llama-arch.cpp +0 -2877
  718. data/ext/sources/examples/talk-llama/llama-arch.h +0 -628
  719. data/ext/sources/examples/talk-llama/llama-batch.cpp +0 -919
  720. data/ext/sources/examples/talk-llama/llama-batch.h +0 -173
  721. data/ext/sources/examples/talk-llama/llama-chat.cpp +0 -896
  722. data/ext/sources/examples/talk-llama/llama-chat.h +0 -71
  723. data/ext/sources/examples/talk-llama/llama-context.cpp +0 -3633
  724. data/ext/sources/examples/talk-llama/llama-context.h +0 -359
  725. data/ext/sources/examples/talk-llama/llama-cparams.cpp +0 -5
  726. data/ext/sources/examples/talk-llama/llama-cparams.h +0 -47
  727. data/ext/sources/examples/talk-llama/llama-ext.h +0 -12
  728. data/ext/sources/examples/talk-llama/llama-grammar.cpp +0 -1464
  729. data/ext/sources/examples/talk-llama/llama-grammar.h +0 -194
  730. data/ext/sources/examples/talk-llama/llama-graph.cpp +0 -2735
  731. data/ext/sources/examples/talk-llama/llama-graph.h +0 -1031
  732. data/ext/sources/examples/talk-llama/llama-hparams.cpp +0 -258
  733. data/ext/sources/examples/talk-llama/llama-hparams.h +0 -353
  734. data/ext/sources/examples/talk-llama/llama-impl.cpp +0 -171
  735. data/ext/sources/examples/talk-llama/llama-impl.h +0 -75
  736. data/ext/sources/examples/talk-llama/llama-io.cpp +0 -15
  737. data/ext/sources/examples/talk-llama/llama-io.h +0 -35
  738. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +0 -330
  739. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.h +0 -137
  740. data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +0 -2285
  741. data/ext/sources/examples/talk-llama/llama-kv-cache.h +0 -389
  742. data/ext/sources/examples/talk-llama/llama-kv-cells.h +0 -533
  743. data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.cpp +0 -275
  744. data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.h +0 -140
  745. data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +0 -268
  746. data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +0 -139
  747. data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +0 -1165
  748. data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +0 -182
  749. data/ext/sources/examples/talk-llama/llama-memory.cpp +0 -59
  750. data/ext/sources/examples/talk-llama/llama-memory.h +0 -122
  751. data/ext/sources/examples/talk-llama/llama-mmap.cpp +0 -752
  752. data/ext/sources/examples/talk-llama/llama-mmap.h +0 -73
  753. data/ext/sources/examples/talk-llama/llama-model-loader.cpp +0 -1655
  754. data/ext/sources/examples/talk-llama/llama-model-loader.h +0 -206
  755. data/ext/sources/examples/talk-llama/llama-model-saver.cpp +0 -299
  756. data/ext/sources/examples/talk-llama/llama-model-saver.h +0 -40
  757. data/ext/sources/examples/talk-llama/llama-model.cpp +0 -9056
  758. data/ext/sources/examples/talk-llama/llama-model.h +0 -597
  759. data/ext/sources/examples/talk-llama/llama-quant.cpp +0 -1304
  760. data/ext/sources/examples/talk-llama/llama-quant.h +0 -1
  761. data/ext/sources/examples/talk-llama/llama-sampler.cpp +0 -3885
  762. data/ext/sources/examples/talk-llama/llama-sampler.h +0 -42
  763. data/ext/sources/examples/talk-llama/llama-vocab.cpp +0 -3970
  764. data/ext/sources/examples/talk-llama/llama-vocab.h +0 -187
  765. data/ext/sources/examples/talk-llama/llama.cpp +0 -1194
  766. data/ext/sources/examples/talk-llama/llama.h +0 -1573
  767. data/ext/sources/examples/talk-llama/models/afmoe.cpp +0 -190
  768. data/ext/sources/examples/talk-llama/models/apertus.cpp +0 -125
  769. data/ext/sources/examples/talk-llama/models/arcee.cpp +0 -135
  770. data/ext/sources/examples/talk-llama/models/arctic.cpp +0 -137
  771. data/ext/sources/examples/talk-llama/models/arwkv7.cpp +0 -86
  772. data/ext/sources/examples/talk-llama/models/baichuan.cpp +0 -123
  773. data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +0 -143
  774. data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +0 -133
  775. data/ext/sources/examples/talk-llama/models/bert.cpp +0 -184
  776. data/ext/sources/examples/talk-llama/models/bitnet.cpp +0 -145
  777. data/ext/sources/examples/talk-llama/models/bloom.cpp +0 -101
  778. data/ext/sources/examples/talk-llama/models/chameleon.cpp +0 -178
  779. data/ext/sources/examples/talk-llama/models/chatglm.cpp +0 -132
  780. data/ext/sources/examples/talk-llama/models/codeshell.cpp +0 -111
  781. data/ext/sources/examples/talk-llama/models/cogvlm.cpp +0 -102
  782. data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +0 -134
  783. data/ext/sources/examples/talk-llama/models/command-r.cpp +0 -122
  784. data/ext/sources/examples/talk-llama/models/dbrx.cpp +0 -122
  785. data/ext/sources/examples/talk-llama/models/deci.cpp +0 -135
  786. data/ext/sources/examples/talk-llama/models/deepseek.cpp +0 -142
  787. data/ext/sources/examples/talk-llama/models/deepseek2.cpp +0 -262
  788. data/ext/sources/examples/talk-llama/models/delta-net-base.cpp +0 -445
  789. data/ext/sources/examples/talk-llama/models/dots1.cpp +0 -132
  790. data/ext/sources/examples/talk-llama/models/dream.cpp +0 -105
  791. data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +0 -148
  792. data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +0 -110
  793. data/ext/sources/examples/talk-llama/models/eurobert.cpp +0 -97
  794. data/ext/sources/examples/talk-llama/models/exaone-moe.cpp +0 -145
  795. data/ext/sources/examples/talk-llama/models/exaone.cpp +0 -114
  796. data/ext/sources/examples/talk-llama/models/exaone4.cpp +0 -123
  797. data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +0 -111
  798. data/ext/sources/examples/talk-llama/models/falcon.cpp +0 -120
  799. data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +0 -116
  800. data/ext/sources/examples/talk-llama/models/gemma.cpp +0 -112
  801. data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +0 -128
  802. data/ext/sources/examples/talk-llama/models/gemma3.cpp +0 -155
  803. data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +0 -384
  804. data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +0 -170
  805. data/ext/sources/examples/talk-llama/models/glm4.cpp +0 -157
  806. data/ext/sources/examples/talk-llama/models/gpt2.cpp +0 -105
  807. data/ext/sources/examples/talk-llama/models/gptneox.cpp +0 -144
  808. data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +0 -195
  809. data/ext/sources/examples/talk-llama/models/granite.cpp +0 -210
  810. data/ext/sources/examples/talk-llama/models/grok.cpp +0 -159
  811. data/ext/sources/examples/talk-llama/models/grovemoe.cpp +0 -139
  812. data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +0 -132
  813. data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +0 -153
  814. data/ext/sources/examples/talk-llama/models/internlm2.cpp +0 -120
  815. data/ext/sources/examples/talk-llama/models/jais.cpp +0 -86
  816. data/ext/sources/examples/talk-llama/models/jais2.cpp +0 -123
  817. data/ext/sources/examples/talk-llama/models/jamba.cpp +0 -106
  818. data/ext/sources/examples/talk-llama/models/kimi-linear.cpp +0 -381
  819. data/ext/sources/examples/talk-llama/models/lfm2.cpp +0 -196
  820. data/ext/sources/examples/talk-llama/models/llada-moe.cpp +0 -122
  821. data/ext/sources/examples/talk-llama/models/llada.cpp +0 -99
  822. data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +0 -178
  823. data/ext/sources/examples/talk-llama/models/llama.cpp +0 -175
  824. data/ext/sources/examples/talk-llama/models/maincoder.cpp +0 -117
  825. data/ext/sources/examples/talk-llama/models/mamba-base.cpp +0 -289
  826. data/ext/sources/examples/talk-llama/models/mamba.cpp +0 -54
  827. data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +0 -129
  828. data/ext/sources/examples/talk-llama/models/minicpm3.cpp +0 -200
  829. data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +0 -123
  830. data/ext/sources/examples/talk-llama/models/mistral3.cpp +0 -160
  831. data/ext/sources/examples/talk-llama/models/models.h +0 -704
  832. data/ext/sources/examples/talk-llama/models/modern-bert.cpp +0 -109
  833. data/ext/sources/examples/talk-llama/models/mpt.cpp +0 -126
  834. data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +0 -162
  835. data/ext/sources/examples/talk-llama/models/nemotron.cpp +0 -122
  836. data/ext/sources/examples/talk-llama/models/neo-bert.cpp +0 -104
  837. data/ext/sources/examples/talk-llama/models/olmo.cpp +0 -121
  838. data/ext/sources/examples/talk-llama/models/olmo2.cpp +0 -150
  839. data/ext/sources/examples/talk-llama/models/olmoe.cpp +0 -124
  840. data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +0 -127
  841. data/ext/sources/examples/talk-llama/models/openelm.cpp +0 -124
  842. data/ext/sources/examples/talk-llama/models/orion.cpp +0 -123
  843. data/ext/sources/examples/talk-llama/models/paddleocr.cpp +0 -122
  844. data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +0 -121
  845. data/ext/sources/examples/talk-llama/models/phi2.cpp +0 -121
  846. data/ext/sources/examples/talk-llama/models/phi3.cpp +0 -152
  847. data/ext/sources/examples/talk-llama/models/plamo.cpp +0 -110
  848. data/ext/sources/examples/talk-llama/models/plamo2.cpp +0 -320
  849. data/ext/sources/examples/talk-llama/models/plamo3.cpp +0 -128
  850. data/ext/sources/examples/talk-llama/models/plm.cpp +0 -169
  851. data/ext/sources/examples/talk-llama/models/qwen.cpp +0 -108
  852. data/ext/sources/examples/talk-llama/models/qwen2.cpp +0 -126
  853. data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +0 -151
  854. data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +0 -117
  855. data/ext/sources/examples/talk-llama/models/qwen3.cpp +0 -120
  856. data/ext/sources/examples/talk-llama/models/qwen35.cpp +0 -381
  857. data/ext/sources/examples/talk-llama/models/qwen35moe.cpp +0 -422
  858. data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +0 -131
  859. data/ext/sources/examples/talk-llama/models/qwen3next.cpp +0 -525
  860. data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +0 -140
  861. data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +0 -132
  862. data/ext/sources/examples/talk-llama/models/refact.cpp +0 -94
  863. data/ext/sources/examples/talk-llama/models/rnd1.cpp +0 -126
  864. data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +0 -164
  865. data/ext/sources/examples/talk-llama/models/rwkv6.cpp +0 -94
  866. data/ext/sources/examples/talk-llama/models/rwkv6qwen2.cpp +0 -86
  867. data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +0 -137
  868. data/ext/sources/examples/talk-llama/models/rwkv7.cpp +0 -90
  869. data/ext/sources/examples/talk-llama/models/seed-oss.cpp +0 -124
  870. data/ext/sources/examples/talk-llama/models/smallthinker.cpp +0 -126
  871. data/ext/sources/examples/talk-llama/models/smollm3.cpp +0 -128
  872. data/ext/sources/examples/talk-llama/models/stablelm.cpp +0 -146
  873. data/ext/sources/examples/talk-llama/models/starcoder.cpp +0 -100
  874. data/ext/sources/examples/talk-llama/models/starcoder2.cpp +0 -121
  875. data/ext/sources/examples/talk-llama/models/step35-iswa.cpp +0 -165
  876. data/ext/sources/examples/talk-llama/models/t5-dec.cpp +0 -166
  877. data/ext/sources/examples/talk-llama/models/t5-enc.cpp +0 -96
  878. data/ext/sources/examples/talk-llama/models/wavtokenizer-dec.cpp +0 -149
  879. data/ext/sources/examples/talk-llama/models/xverse.cpp +0 -108
  880. data/ext/sources/examples/talk-llama/prompts/talk-alpaca.txt +0 -23
  881. data/ext/sources/examples/talk-llama/speak +0 -40
  882. data/ext/sources/examples/talk-llama/speak.bat +0 -1
  883. data/ext/sources/examples/talk-llama/speak.ps1 +0 -14
  884. data/ext/sources/examples/talk-llama/talk-llama.cpp +0 -813
  885. data/ext/sources/examples/talk-llama/unicode-data.cpp +0 -7034
  886. data/ext/sources/examples/talk-llama/unicode-data.h +0 -20
  887. data/ext/sources/examples/talk-llama/unicode.cpp +0 -1103
  888. data/ext/sources/examples/talk-llama/unicode.h +0 -111
  889. data/ext/sources/examples/wchess/CMakeLists.txt +0 -10
  890. data/ext/sources/examples/wchess/libwchess/CMakeLists.txt +0 -19
  891. data/ext/sources/examples/wchess/libwchess/Chessboard.cpp +0 -803
  892. data/ext/sources/examples/wchess/libwchess/Chessboard.h +0 -33
  893. data/ext/sources/examples/wchess/libwchess/WChess.cpp +0 -193
  894. data/ext/sources/examples/wchess/libwchess/WChess.h +0 -63
  895. data/ext/sources/examples/wchess/libwchess/test-chessboard.cpp +0 -117
  896. data/ext/sources/examples/wchess/wchess.cmd/CMakeLists.txt +0 -8
  897. data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +0 -253
  898. data/ext/sources/examples/whisper.wasm/CMakeLists.txt +0 -50
  899. data/ext/sources/examples/whisper.wasm/emscripten.cpp +0 -118
  900. data/ext/sources/examples/whisper.wasm/index-tmpl.html +0 -659
  901. data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +0 -99
  902. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +0 -155
  903. data/ext/sources/ggml/src/ggml-hexagon/op-desc.h +0 -153
  904. data/ext/sources/ggml/src/ggml-opencl/kernels/embed_kernel.py +0 -26
  905. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.cpp +0 -123
  906. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.h +0 -17
  907. data/ext/sources/ggml/src/ggml-virtgpu/regenerate_remoting.py +0 -333
  908. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/abs.comp +0 -21
  909. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ceil.comp +0 -22
  910. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/clamp.comp +0 -17
  911. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/cos.comp +0 -17
  912. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/elu.comp +0 -27
  913. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +0 -21
  914. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/floor.comp +0 -22
  915. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu.comp +0 -25
  916. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_erf.comp +0 -39
  917. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_quick.comp +0 -23
  918. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardsigmoid.comp +0 -22
  919. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardswish.comp +0 -22
  920. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/leaky_relu.comp +0 -22
  921. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/neg.comp +0 -20
  922. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/relu.comp +0 -21
  923. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/round.comp +0 -29
  924. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.glsl +0 -5
  925. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sgn.comp +0 -21
  926. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp +0 -20
  927. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/silu.comp +0 -22
  928. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sin.comp +0 -17
  929. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/softplus.comp +0 -23
  930. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp +0 -17
  931. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/square.comp +0 -17
  932. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/step.comp +0 -22
  933. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp +0 -20
  934. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/trunc.comp +0 -22
  935. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/xielu.comp +0 -35
  936. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +0 -182
  937. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.tmpl.wgsl +0 -323
  938. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl +0 -718
  939. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +0 -123
  940. data/ext/sources/tests/CMakeLists.txt +0 -112
  941. data/ext/sources/tests/earnings21/eval.mk +0 -58
  942. data/ext/sources/tests/earnings21/eval.py +0 -68
  943. data/ext/sources/tests/earnings21/normalizers/__init__.py +0 -2
  944. data/ext/sources/tests/earnings21/normalizers/basic.py +0 -80
  945. data/ext/sources/tests/earnings21/normalizers/english.json +0 -1741
  946. data/ext/sources/tests/earnings21/normalizers/english.py +0 -550
  947. data/ext/sources/tests/earnings21/requirements.txt +0 -6
  948. data/ext/sources/tests/en-0-ref.txt +0 -1
  949. data/ext/sources/tests/en-1-ref.txt +0 -1
  950. data/ext/sources/tests/en-2-ref.txt +0 -1
  951. data/ext/sources/tests/es-0-ref.txt +0 -1
  952. data/ext/sources/tests/librispeech/eval.mk +0 -39
  953. data/ext/sources/tests/librispeech/eval.py +0 -47
  954. data/ext/sources/tests/librispeech/normalizers/__init__.py +0 -2
  955. data/ext/sources/tests/librispeech/normalizers/basic.py +0 -80
  956. data/ext/sources/tests/librispeech/normalizers/english.json +0 -1741
  957. data/ext/sources/tests/librispeech/normalizers/english.py +0 -550
  958. data/ext/sources/tests/librispeech/requirements.txt +0 -6
  959. data/ext/sources/tests/run-tests.sh +0 -130
  960. data/ext/sources/tests/test-c.c +0 -3
  961. data/ext/sources/tests/test-vad-full.cpp +0 -56
  962. data/ext/sources/tests/test-vad.cpp +0 -83
  963. data/ext/sources/tests/test-whisper.js +0 -58
  964. data/lib/whisper/context.rb +0 -15
  965. data/lib/whisper/segment.rb +0 -58
@@ -6,7 +6,7 @@
6
6
  #include <openvino/op/constant.hpp>
7
7
  #include <openvino/runtime/tensor.hpp>
8
8
 
9
- void unpack_32_4(const uint8_t* data, uint8_t* dst);
9
+ void unpack_32_4(const uint8_t * data, uint8_t * dst);
10
10
 
11
11
  void extract_q4_0_data(const ggml_tensor * tensor,
12
12
  ov::Tensor & weights_arr,
@@ -19,12 +19,18 @@ void extract_q4_1_data(const ggml_tensor * tensor,
19
19
  ov::Tensor & zp_arr,
20
20
  bool use_bias = false);
21
21
 
22
+ void extract_q5_1_data(const ggml_tensor * tensor,
23
+ ov::Tensor & weights_arr,
24
+ ov::Tensor & scales_arr,
25
+ ov::Tensor & zp_arr,
26
+ bool use_bias = false);
27
+
22
28
  void extract_q8_0_data(const ggml_tensor * tensor,
23
29
  ov::Tensor & weights_arr,
24
30
  ov::Tensor & scales_arr,
25
31
  ov::Tensor & zp_arr);
26
32
 
27
- void unpack_256_4(const uint8_t* data, uint8_t* dst);
33
+ void unpack_256_4(const uint8_t * data, uint8_t * dst);
28
34
 
29
35
  void extract_q4_k_data(const ggml_tensor * tensor,
30
36
  ov::Tensor & weights_arr,
@@ -145,8 +151,8 @@ namespace ov {
145
151
  namespace op {
146
152
  namespace util {
147
153
  // From <openvino>/src/common/transformations/include/transformations/utils/utils.hpp
148
- bool get_single_value(const std::shared_ptr<ov::op::v0::Constant>& const_node,
149
- float& value,
154
+ bool get_single_value(const std::shared_ptr<ov::op::v0::Constant> & const_node,
155
+ float & value,
150
156
  bool check_value_range = true);
151
157
  } // namespace util
152
158
  } // namespace op
@@ -3,6 +3,8 @@
3
3
  #include <cstdint>
4
4
  #include <map>
5
5
  #include <openvino/core/node.hpp>
6
+ #include <openvino/core/partial_shape.hpp>
7
+ #include <openvino/core/shape.hpp>
6
8
  #include <openvino/frontend/decoder.hpp>
7
9
  #include <string>
8
10
 
@@ -12,22 +14,50 @@ namespace ggml {
12
14
 
13
15
  class GgmlDecoder : public DecoderBase {
14
16
  public:
15
- virtual ov::Any get_attribute(const std::string& name) const = 0;
17
+ virtual ov::Any get_attribute(const std::string & name) const = 0;
16
18
 
17
- virtual PartialShape get_input_shape(int node_idx, const std::string& name) const = 0;
19
+ virtual PartialShape get_input_shape(int node_idx, const std::string & name) const = 0;
18
20
 
19
- virtual std::vector<size_t> get_input_stride(int node_idx, const std::string& name) const = 0;
21
+ virtual std::vector<size_t> get_input_stride(int node_idx, const std::string & name) const = 0;
20
22
 
21
- virtual element::Type get_input_type(int node_idx, const std::string& name) const = 0;
23
+ virtual size_t get_view_input_size(int node_idx, const std::string & name) const = 0;
24
+
25
+ virtual size_t get_view_input_offset(int node_idx, const std::string & name, size_t view_index) const = 0;
26
+
27
+ virtual size_t get_view_input_src_offset(int node_idx, const std::string & name, size_t view_index) const = 0;
28
+
29
+ virtual std::vector<size_t> get_view_input_stride(int node_idx,
30
+ const std::string & name,
31
+ size_t view_index) const = 0;
32
+
33
+ virtual std::vector<size_t> get_view_input_src_stride(int node_idx,
34
+ const std::string & name,
35
+ size_t view_index) const = 0;
36
+
37
+ virtual Shape get_view_input_ggml_shape(int node_idx, const std::string & name, size_t view_index) const = 0;
38
+
39
+ virtual Shape get_view_input_src_ggml_shape(int node_idx, const std::string & name, size_t view_index) const = 0;
40
+
41
+ virtual PartialShape get_view_input_ov_shape(int node_idx, const std::string & name, size_t view_index) const = 0;
42
+
43
+ virtual PartialShape get_view_input_src_ov_shape(int node_idx,
44
+ const std::string & name,
45
+ size_t view_index) const = 0;
46
+
47
+ virtual std::string get_view_input_name(int node_idx, const std::string & name, size_t view_index) const = 0;
48
+
49
+ virtual std::string get_view_input_src_name(int node_idx, const std::string & name, size_t view_index) const = 0;
50
+
51
+ virtual element::Type get_input_type(int node_idx, const std::string & name) const = 0;
22
52
 
23
53
  virtual size_t get_input_size() const = 0;
24
54
 
25
55
  virtual size_t get_input_size(int node_idx) const = 0;
26
56
 
27
57
  virtual void get_input_node(size_t input_port_idx,
28
- std::string& producer_name,
29
- std::string& producer_output_port_name,
30
- size_t& producer_output_port_index) const = 0;
58
+ std::string & producer_name,
59
+ std::string & producer_output_port_name,
60
+ size_t & producer_output_port_index) const = 0;
31
61
 
32
62
  virtual std::vector<std::string> get_input_names(int node_idx) const = 0;
33
63
 
@@ -35,30 +65,36 @@ public:
35
65
 
36
66
  virtual element::Type get_output_type(const int node_idx) const = 0;
37
67
 
38
- virtual int32_t* get_input_op_params(int node_idx, const std::string& name) const = 0;
68
+ virtual std::vector<size_t> get_output_stride(int node_idx) const = 0;
69
+
70
+ virtual int32_t * get_input_op_params(int node_idx, const std::string & name) const = 0;
39
71
 
40
72
  virtual int32_t * get_output_op_params(int node_idx) const = 0;
41
73
 
74
+ virtual size_t get_output_op_offset(int node_idx) const = 0;
75
+
42
76
  virtual std::vector<std::string> get_output_names(int node_idx) const = 0;
43
77
 
44
- virtual const std::string& get_op_type() const = 0;
78
+ virtual const std::string & get_op_type() const = 0;
45
79
 
46
- virtual const std::string& get_op_type(int node_idx) const = 0;
80
+ virtual const std::string & get_op_type(int node_idx) const = 0;
47
81
 
48
- virtual const std::string& get_op_name() const = 0;
82
+ virtual const std::string & get_op_name() const = 0;
49
83
 
50
- virtual const std::string& get_op_name(int node_idx) const = 0;
84
+ virtual const std::string & get_op_name(int node_idx) const = 0;
51
85
 
52
86
  virtual void visit_subgraph(std::function<void(std::shared_ptr<GgmlDecoder>, int node_idx)> node_visitor) const = 0;
53
87
 
54
88
  virtual int get_op_case(int node_idx) const = 0;
55
89
 
56
- virtual const std::map<std::string, std::shared_ptr<ov::Node>>& get_model_inputs() const = 0;
57
- virtual const std::map<std::string, std::shared_ptr<ov::Node>>& get_model_extra_inputs() const = 0;
58
- virtual const std::map<std::string, std::shared_ptr<ov::Node>>& get_model_weights() const = 0;
90
+ virtual const std::map<std::string, std::shared_ptr<ov::Node>> & get_model_inputs() const = 0;
91
+ virtual const std::map<std::string, std::shared_ptr<ov::Node>> & get_model_extra_inputs() const = 0;
92
+ virtual const std::map<std::string, std::shared_ptr<ov::Node>> & get_model_weights() const = 0;
59
93
  virtual std::vector<std::string> get_model_output_names() const = 0;
60
94
 
61
- virtual int32_t* get_rope_params() const = 0;
95
+ virtual int32_t * get_rope_params() const = 0;
96
+
97
+ virtual bool has_mixed_rope_params() const = 0;
62
98
 
63
99
  virtual std::map<std::string, std::string> get_kv_param_res_names() const = 0;
64
100
 
@@ -66,7 +102,11 @@ public:
66
102
 
67
103
  virtual bool is_stateful() const = 0;
68
104
 
105
+ virtual bool is_splited_model() const = 0;
106
+
69
107
  virtual int is_swa_layer(int layer) const = 0;
108
+
109
+ virtual int32_t get_op_dynamic_dim(int node_idx) const = 0;
70
110
  };
71
111
 
72
112
  } // namespace ggml
@@ -15,7 +15,7 @@ public:
15
15
  using Ptr = std::shared_ptr<FrontEnd>;
16
16
  FrontEnd();
17
17
 
18
- static std::shared_ptr<Model> convert(const InputModel::Ptr& model, bool naive = false);
18
+ static std::shared_ptr<Model> convert(const InputModel::Ptr & model, bool naive = false);
19
19
  };
20
20
 
21
21
  } // namespace ggml
@@ -1,9 +1,9 @@
1
1
  #pragma once
2
2
 
3
- #include <openvino/frontend/input_model.hpp>
4
-
5
3
  #include "decoder.h"
6
4
 
5
+ #include <openvino/frontend/input_model.hpp>
6
+
7
7
  namespace ov {
8
8
  namespace frontend {
9
9
  namespace ggml {
@@ -16,9 +16,9 @@ class InputModel : public ov::frontend::InputModel {
16
16
  friend class ::ov::frontend::ggml::FrontEnd;
17
17
 
18
18
  public:
19
- explicit InputModel(const std::shared_ptr<GgmlDecoder>& gdecoder);
19
+ explicit InputModel(const std::shared_ptr<GgmlDecoder> & gdecoder);
20
20
 
21
- const std::shared_ptr<GgmlDecoder>& get_model_decoder() const;
21
+ const std::shared_ptr<GgmlDecoder> & get_model_decoder() const;
22
22
 
23
23
  private:
24
24
  std::shared_ptr<GgmlDecoder> m_decoder;
@@ -1,11 +1,11 @@
1
1
  #pragma once
2
2
 
3
+ #include "decoder.h"
4
+
3
5
  #include <cstdint>
4
6
  #include <openvino/frontend/node_context.hpp>
5
7
  #include <string>
6
8
 
7
- #include "decoder.h"
8
-
9
9
  namespace ov {
10
10
  namespace frontend {
11
11
  namespace ggml {
@@ -16,28 +16,24 @@ typedef std::map<std::string, Output<Node>> TensorMap;
16
16
 
17
17
  class NodeContext : public frontend::NodeContext {
18
18
  public:
19
- NodeContext(const std::shared_ptr<GgmlDecoder>& decoder,
20
- std::shared_ptr<TensorMap>& tensor_map,
19
+ NodeContext(const std::shared_ptr<GgmlDecoder> & decoder,
20
+ std::shared_ptr<TensorMap> & tensor_map,
21
21
  int node_idx,
22
- TranslateSession* translate_session = nullptr)
23
- : ov::frontend::NodeContext(decoder->get_op_type(node_idx)),
24
- m_decoder(decoder),
25
- m_tensor_map(tensor_map),
26
- m_node_idx(node_idx),
27
- m_translate_session(translate_session) {
22
+ TranslateSession * translate_session = nullptr) :
23
+ ov::frontend::NodeContext(decoder->get_op_type(node_idx)),
24
+ m_decoder(decoder),
25
+ m_tensor_map(tensor_map),
26
+ m_node_idx(node_idx),
27
+ m_translate_session(translate_session) {
28
28
  m_input_names = decoder->get_input_names(m_node_idx);
29
29
  m_output_names = decoder->get_output_names(m_node_idx);
30
30
  }
31
31
 
32
- TranslateSession* get_translate_session() const {
33
- return m_translate_session;
34
- }
32
+ TranslateSession * get_translate_session() const { return m_translate_session; }
35
33
 
36
- const std::vector<std::string>& get_input_names() const { return m_input_names; }
34
+ const std::vector<std::string> & get_input_names() const { return m_input_names; }
37
35
 
38
- size_t get_input_size() const override {
39
- return m_decoder->get_input_size(m_node_idx);
40
- }
36
+ size_t get_input_size() const override { return m_decoder->get_input_size(m_node_idx); }
41
37
 
42
38
  ov::element::Type get_input_type(size_t index) const {
43
39
  return m_decoder->get_input_type(m_node_idx, m_input_names[index]);
@@ -55,42 +51,103 @@ public:
55
51
 
56
52
  PartialShape get_output_shape() const { return m_decoder->get_output_shape(m_node_idx); }
57
53
 
58
- int32_t* get_input_op_params(size_t index) const {
54
+ int32_t * get_input_op_params(size_t index) const {
59
55
  return m_decoder->get_input_op_params(m_node_idx, m_input_names[index]);
60
56
  }
61
57
 
62
- int32_t * get_output_op_params() const { return m_decoder->get_output_op_params(m_node_idx); }
58
+ size_t get_view_input_size(size_t index) const {
59
+ return m_decoder->get_view_input_size(m_node_idx, m_input_names[index]);
60
+ }
61
+
62
+ size_t get_view_input_offset(size_t index, size_t view_index) const {
63
+ return m_decoder->get_view_input_offset(m_node_idx, m_input_names[index], view_index);
64
+ }
63
65
 
64
- ov::element::Type get_output_type() const {
65
- return m_decoder->get_output_type(m_node_idx);
66
+ size_t get_view_input_src_offset(size_t index, size_t view_index) const {
67
+ return m_decoder->get_view_input_src_offset(m_node_idx, m_input_names[index], view_index);
66
68
  }
67
69
 
70
+ std::vector<size_t> get_view_input_stride(size_t index, size_t view_index) const {
71
+ return m_decoder->get_view_input_stride(m_node_idx, m_input_names[index], view_index);
72
+ }
73
+
74
+ std::vector<size_t> get_view_input_src_stride(size_t index, size_t view_index) const {
75
+ return m_decoder->get_view_input_src_stride(m_node_idx, m_input_names[index], view_index);
76
+ }
77
+
78
+ ov::Shape get_view_input_ggml_shape(size_t index, size_t view_index) const {
79
+ return m_decoder->get_view_input_ggml_shape(m_node_idx, m_input_names[index], view_index);
80
+ }
81
+
82
+ ov::Shape get_view_input_src_ggml_shape(size_t index, size_t view_index) const {
83
+ return m_decoder->get_view_input_src_ggml_shape(m_node_idx, m_input_names[index], view_index);
84
+ }
85
+
86
+ ov::PartialShape get_view_input_ov_shape(size_t index, size_t view_index) const {
87
+ return m_decoder->get_view_input_ov_shape(m_node_idx, m_input_names[index], view_index);
88
+ }
89
+
90
+ ov::PartialShape get_view_input_src_ov_shape(size_t index, size_t view_index) const {
91
+ return m_decoder->get_view_input_src_ov_shape(m_node_idx, m_input_names[index], view_index);
92
+ }
93
+
94
+ std::string get_view_input_name(size_t index, size_t view_index) const {
95
+ return m_decoder->get_view_input_name(m_node_idx, m_input_names[index], view_index);
96
+ }
97
+
98
+ std::string get_view_input_src_name(size_t index, size_t view_index) const {
99
+ return m_decoder->get_view_input_src_name(m_node_idx, m_input_names[index], view_index);
100
+ }
101
+
102
+ int32_t get_op_dynamic_dim() const { return m_decoder->get_op_dynamic_dim(m_node_idx); }
103
+
104
+ int32_t * get_output_op_params() const { return m_decoder->get_output_op_params(m_node_idx); }
105
+
106
+ size_t get_output_op_offset() const { return m_decoder->get_output_op_offset(m_node_idx); }
107
+
108
+ ov::element::Type get_output_type() const { return m_decoder->get_output_type(m_node_idx); }
109
+
110
+ std::vector<size_t> get_output_stride() const { return m_decoder->get_output_stride(m_node_idx); }
111
+
68
112
  Output<Node> get_input(int idx) const override {
113
+ // Check if this input is a VIEW
114
+ size_t view_input_size = m_decoder->get_view_input_size(m_node_idx, m_input_names[idx]);
115
+ if (view_input_size > 0) {
116
+ // This is a VIEW input, get the base tensor name (last element in the chain)
117
+ std::string base_name =
118
+ m_decoder->get_view_input_src_name(m_node_idx, m_input_names[idx], view_input_size - 1);
119
+ // Check if the VIEW has been resolved (translate_view produced a Slice)
120
+ auto view_it = m_tensor_map->find(m_input_names[idx]);
121
+ if (!base_name.empty() && view_it != m_tensor_map->end()) {
122
+ auto base_it = m_tensor_map->find(base_name);
123
+ if (base_it != m_tensor_map->end() &&
124
+ view_it->second.get_node_shared_ptr() != base_it->second.get_node_shared_ptr()) {
125
+ return view_it->second;
126
+ }
127
+ return base_it->second;
128
+ }
129
+ if (!base_name.empty()) {
130
+ return m_tensor_map->at(base_name);
131
+ }
132
+ }
133
+ // Not a VIEW or failed to get base name, use the original logic
69
134
  return m_tensor_map->at(m_input_names[idx]);
70
135
  }
71
136
 
72
- Output<Node> get_input(const std::string& name) const override {
137
+ Output<Node> get_input(const std::string & name) const override {
73
138
  if (m_tensor_map->find(name) == m_tensor_map->end()) {
74
139
  throw std::runtime_error("'" + name + "' not found in tensor map.");
75
140
  }
76
141
  return m_tensor_map->at(name);
77
142
  }
78
143
 
79
- bool has_input(const std::string& name) const {
80
- return m_tensor_map->find(name) != m_tensor_map->end();
81
- }
144
+ bool has_input(const std::string & name) const { return m_tensor_map->find(name) != m_tensor_map->end(); }
82
145
 
83
- const std::string& get_name() const override {
84
- return m_decoder->get_op_name(m_node_idx);
85
- }
146
+ const std::string & get_name() const override { return m_decoder->get_op_name(m_node_idx); }
86
147
 
87
- ov::Any get_attribute_as_any(const std::string& name) const override {
88
- return m_decoder->get_attribute(name);
89
- }
148
+ ov::Any get_attribute_as_any(const std::string & name) const override { return m_decoder->get_attribute(name); }
90
149
 
91
- int get_op_case() const {
92
- return m_decoder->get_op_case(m_node_idx);
93
- }
150
+ int get_op_case() const { return m_decoder->get_op_case(m_node_idx); }
94
151
 
95
152
  bool is_static() const { return m_decoder->is_static(); }
96
153
 
@@ -98,14 +155,14 @@ public:
98
155
 
99
156
  private:
100
157
  std::shared_ptr<GgmlDecoder> m_decoder;
101
- std::shared_ptr<TensorMap>& m_tensor_map;
158
+ std::shared_ptr<TensorMap> & m_tensor_map;
102
159
  int m_node_idx;
103
- TranslateSession* m_translate_session;
160
+ TranslateSession * m_translate_session;
104
161
  std::vector<std::string> m_input_names;
105
162
  std::vector<std::string> m_output_names;
106
163
  };
107
164
 
108
- using CreatorFunction = std::function<ov::OutputVector(const ov::frontend::ggml::NodeContext&)>;
165
+ using CreatorFunction = std::function<ov::OutputVector(const ov::frontend::ggml::NodeContext &)>;
109
166
 
110
167
  } // namespace ggml
111
168
  } // namespace frontend
@@ -0,0 +1,76 @@
1
+ #include "../node_context.h"
2
+ #include "../op_table.h"
3
+ #include "../utils.h"
4
+
5
+ #include <memory>
6
+ #include <openvino/core/node.hpp>
7
+ #include <openvino/core/node_output.hpp>
8
+ #include <openvino/op/add.hpp>
9
+ #include <openvino/op/constant.hpp>
10
+ #include <openvino/op/convert.hpp>
11
+ #include <openvino/op/gather.hpp>
12
+ #include <openvino/op/reshape.hpp>
13
+ #include <openvino/op/shape_of.hpp>
14
+
15
+ namespace ov {
16
+ namespace frontend {
17
+ namespace ggml {
18
+ namespace op {
19
+
20
+ static ov::Output<ov::Node> reshape_add_id_input_to_2d(const ov::Output<ov::Node> & input,
21
+ const ov::PartialShape & input_shape,
22
+ const std::vector<int> & dims) {
23
+ const auto actual_shape = input.get_partial_shape();
24
+ if (actual_shape.rank().is_static() && actual_shape.rank().get_length() == 2) {
25
+ return input;
26
+ }
27
+
28
+ if (input_shape.rank().is_static() && input_shape.rank().get_length() == 2) {
29
+ return input;
30
+ }
31
+
32
+ auto shape = std::make_shared<ov::op::v3::ShapeOf>(input, ov::element::i64);
33
+ return std::make_shared<ov::op::v1::Reshape>(input, get_dimensions(shape, dims), false);
34
+ }
35
+
36
+ OutputVector translate_add_id(const NodeContext & context) {
37
+ num_inputs_check(context, 3, 3);
38
+
39
+ auto input = process_view_input_new(context, 0);
40
+ auto bias = process_view_input_new(context, 1);
41
+ auto ids = process_view_input_new(context, 2);
42
+
43
+ // OpenVINO uses reversed GGML dimensions:
44
+ // input: [1, n_token, n_used, n_embd]
45
+ // bias: [1, 1, n_expert, n_embd]
46
+ // ids: [1, 1, n_token, n_used]
47
+ // Model bias constants may already be stored as [n_expert, n_embd].
48
+ bias = reshape_add_id_input_to_2d(bias, context.get_input_shape(1), {2, 3});
49
+ ids = reshape_add_id_input_to_2d(ids, context.get_input_shape(2), {2, 3});
50
+
51
+ if (ids.get_element_type() != ov::element::i32 && ids.get_element_type() != ov::element::i64) {
52
+ ids = std::make_shared<ov::op::v0::Convert>(ids, ov::element::i32);
53
+ }
54
+
55
+ auto gather_axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {0});
56
+ ov::Output<ov::Node> selected_bias = std::make_shared<ov::op::v8::Gather>(bias, ids, gather_axis);
57
+ selected_bias = std::make_shared<ov::op::v1::Reshape>(
58
+ selected_bias, std::make_shared<ov::op::v3::ShapeOf>(input, ov::element::i64), false);
59
+
60
+ if (selected_bias.get_element_type() != input.get_element_type()) {
61
+ selected_bias = std::make_shared<ov::op::v0::Convert>(selected_bias, input.get_element_type());
62
+ }
63
+
64
+ ov::Output<ov::Node> res = std::make_shared<ov::op::v1::Add>(input, selected_bias);
65
+ const auto output_type = context.get_output_type();
66
+ if (res.get_element_type() != output_type) {
67
+ res = std::make_shared<ov::op::v0::Convert>(res, output_type);
68
+ }
69
+
70
+ return rename_outputs_with_suffix({res}, context.get_name());
71
+ }
72
+
73
+ } // namespace op
74
+ } // namespace ggml
75
+ } // namespace frontend
76
+ } // namespace ov
@@ -0,0 +1,47 @@
1
+ #include "../node_context.h"
2
+ #include "../op_table.h"
3
+ #include "../utils.h"
4
+ #include "ggml.h"
5
+
6
+ #include <openvino/frontend/exception.hpp>
7
+ #include <openvino/op/constant.hpp>
8
+ #include <openvino/op/squeeze.hpp>
9
+ #include <openvino/op/topk.hpp>
10
+
11
+ namespace ov {
12
+ namespace frontend {
13
+ namespace ggml {
14
+ namespace op {
15
+
16
+ OutputVector translate_argsort(const NodeContext & context) {
17
+ num_inputs_check(context, 1, 1);
18
+
19
+ auto input = process_view_input_new(context, 0);
20
+
21
+ const int32_t order = context.get_output_op_params()[0];
22
+
23
+ ov::op::v11::TopK::Mode mode;
24
+ switch (order) {
25
+ case GGML_SORT_ORDER_ASC:
26
+ mode = ov::op::v11::TopK::Mode::MIN;
27
+ break;
28
+ case GGML_SORT_ORDER_DESC:
29
+ mode = ov::op::v11::TopK::Mode::MAX;
30
+ break;
31
+ default:
32
+ FRONT_END_OP_CONVERSION_CHECK(false, "Unsupported GGML_OP_ARGSORT order: ", order);
33
+ }
34
+
35
+ auto k = std::make_shared<ov::op::v0::Squeeze>(get_dimensions(input.get_node_shared_ptr(), {3}),
36
+ ov::op::v0::Constant::create(ov::element::i64, {1}, {0}));
37
+
38
+ auto topk = std::make_shared<ov::op::v11::TopK>(input, k, 3, mode, ov::op::v11::TopK::SortType::SORT_VALUES,
39
+ context.get_output_type(), false);
40
+
41
+ return rename_outputs_with_suffix({topk->output(1)}, context.get_name());
42
+ }
43
+
44
+ } // namespace op
45
+ } // namespace ggml
46
+ } // namespace frontend
47
+ } // namespace ov
@@ -0,0 +1,33 @@
1
+ #include "../node_context.h"
2
+ #include "../op_table.h"
3
+ #include "../utils.h"
4
+
5
+ #include <cstring>
6
+ #include <openvino/op/clamp.hpp>
7
+
8
+ namespace ov {
9
+ namespace frontend {
10
+ namespace ggml {
11
+ namespace op {
12
+
13
+ OutputVector translate_clamp(const NodeContext & context) {
14
+ num_inputs_check(context, 1, 1);
15
+
16
+ auto input = process_view_input_new(context, 0);
17
+
18
+ const int32_t * op_params = context.get_output_op_params();
19
+ FRONT_END_CHECK_IMPLEMENTED(op_params != nullptr, "CLAMP requires output op params");
20
+
21
+ float min;
22
+ float max;
23
+ std::memcpy(&min, reinterpret_cast<const float *>(op_params) + 0, sizeof(float));
24
+ std::memcpy(&max, reinterpret_cast<const float *>(op_params) + 1, sizeof(float));
25
+
26
+ auto res = std::make_shared<ov::op::v0::Clamp>(input, min, max);
27
+ return rename_outputs_with_suffix({res}, context.get_name());
28
+ }
29
+
30
+ } // namespace op
31
+ } // namespace ggml
32
+ } // namespace frontend
33
+ } // namespace ov
@@ -0,0 +1,48 @@
1
+ #include "../node_context.h"
2
+ #include "../op_table.h"
3
+ #include "../utils.h"
4
+
5
+ #include <memory>
6
+ #include <openvino/frontend/exception.hpp>
7
+ #include <openvino/op/concat.hpp>
8
+ #include <openvino/op/convert.hpp>
9
+
10
+ namespace ov {
11
+ namespace frontend {
12
+ namespace ggml {
13
+ namespace op {
14
+
15
+ OutputVector translate_concat(const NodeContext & context) {
16
+ num_inputs_check(context, 2, 2);
17
+
18
+ const int32_t * op_params = context.get_output_op_params();
19
+ FRONT_END_CHECK_IMPLEMENTED(op_params != nullptr, "CONCAT requires output op params");
20
+
21
+ const auto output_shape = context.get_output_shape();
22
+ FRONT_END_CHECK_IMPLEMENTED(output_shape.rank().is_static(), "CONCAT requires static output rank");
23
+
24
+ const auto rank = output_shape.rank().get_length();
25
+ const int32_t ggml_dim = op_params[0];
26
+ FRONT_END_CHECK_IMPLEMENTED(ggml_dim >= 0 && ggml_dim < rank, "CONCAT axis is out of range");
27
+
28
+ auto input_0 = process_view_input_new(context, 0);
29
+ auto input_1 = process_view_input_new(context, 1);
30
+ const auto output_type = context.get_output_type();
31
+
32
+ if (input_0.get_element_type() != output_type) {
33
+ input_0 = std::make_shared<ov::op::v0::Convert>(input_0, output_type);
34
+ }
35
+ if (input_1.get_element_type() != output_type) {
36
+ input_1 = std::make_shared<ov::op::v0::Convert>(input_1, output_type);
37
+ }
38
+
39
+ const auto axis = static_cast<int64_t>(rank - 1 - ggml_dim);
40
+ auto res = std::make_shared<ov::op::v0::Concat>(OutputVector{input_0, input_1}, axis);
41
+
42
+ return rename_outputs_with_suffix({res}, context.get_name());
43
+ }
44
+
45
+ } // namespace op
46
+ } // namespace ggml
47
+ } // namespace frontend
48
+ } // namespace ov
@@ -18,27 +18,19 @@ namespace op {
18
18
  OutputVector translate_cont(const NodeContext & context) {
19
19
  num_inputs_check(context, 1, 1);
20
20
 
21
- int op_case = context.get_op_case();
22
- FRONT_END_CHECK_IMPLEMENTED(op_case == 1 || op_case == 2 || op_case == 3, "Unsupported CONT case");
23
-
24
21
  auto src_shape = context.get_input_shape(0).to_shape();
25
22
  auto dst_shape = context.get_output_shape().to_shape();
26
- ov::Output<Node> res;
27
23
 
28
- if (op_case == 1) {
29
- // The input comes from a PERMUTE
30
- throw std::runtime_error("Code of this case might be outdated");
31
- dst_shape[1] = -1;
32
- res = std::make_shared<ov::op::v1::Reshape>(
33
- context.get_input(0), ov::op::v0::Constant::create(ov::element::i64, {dst_shape.size()}, dst_shape), false);
34
- } else if (op_case == 2) {
35
- // The input comes from a TRANSPOSE
36
- return {context.get_input(0)};
37
- } else {
38
- // The input comes from a VIEW
39
- res = process_view_input(context, 0);
24
+ if (context.get_op_dynamic_dim() != -1) {
25
+ dst_shape[3 - context.get_op_dynamic_dim()] = -1;
40
26
  }
41
27
 
28
+ auto input = process_view_input_new(context, 0);
29
+
30
+ ov::Output<Node> res;
31
+ res = std::make_shared<ov::op::v1::Reshape>(
32
+ input, ov::op::v0::Constant::create(ov::element::i64, {dst_shape.size()}, dst_shape), false);
33
+
42
34
  return rename_outputs_with_suffix({res}, context.get_name());
43
35
  }
44
36
 
@@ -3,7 +3,9 @@
3
3
  #include "../utils.h"
4
4
 
5
5
  #include <memory>
6
+ #include <openvino/op/constant.hpp>
6
7
  #include <openvino/op/convert.hpp>
8
+ #include <openvino/op/reshape.hpp>
7
9
 
8
10
  namespace ov {
9
11
  namespace frontend {
@@ -11,7 +13,18 @@ namespace ggml {
11
13
  namespace op {
12
14
 
13
15
  OutputVector translate_cpy(const NodeContext & context) {
14
- auto res = std::make_shared<ov::op::v0::Convert>(context.get_input(0), context.get_output_type());
16
+ auto input = process_view_input_new(context, 0);
17
+ auto input_shape = context.get_input_shape(0);
18
+ auto output_shape = context.get_output_shape();
19
+
20
+ // Non-cast CPY may need a reshape (e.g. [3,192,1,1] -> [576,1,1,1])
21
+ if (input_shape != output_shape) {
22
+ auto new_shape = ov::op::v0::Constant::create(
23
+ ov::element::i64, {static_cast<size_t>(output_shape.rank().get_length())}, output_shape.to_shape());
24
+ input = std::make_shared<ov::op::v1::Reshape>(input, new_shape, false);
25
+ }
26
+
27
+ auto res = std::make_shared<ov::op::v0::Convert>(input, context.get_output_type());
15
28
  return rename_outputs_with_suffix({res}, context.get_name());
16
29
  }
17
30