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
data/ext/sources/examples/talk-llama/models/stablelm.cpp DELETED
@@ -1,146 +0,0 @@
1
- #include "models.h"
2
-
3
- llm_build_stablelm::llm_build_stablelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
4
- const int64_t n_embd_head = hparams.n_embd_head_v();
5
-
6
- GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
7
-
8
- ggml_tensor * cur;
9
- ggml_tensor * inpL;
10
-
11
- inpL = build_inp_embd(model.tok_embd);
12
-
13
- // inp_pos - contains the positions
14
- ggml_tensor * inp_pos = build_inp_pos();
15
-
16
- auto * inp_attn = build_attn_inp_kv();
17
-
18
- ggml_tensor * inp_out_ids = build_inp_out_ids();
19
-
20
- for (int il = 0; il < n_layer; ++il) {
21
- // norm
22
- cur = build_norm(inpL,
23
- model.layers[il].attn_norm,
24
- model.layers[il].attn_norm_b,
25
- LLM_NORM, il);
26
- cb(cur, "attn_norm", il);
27
-
28
- ggml_tensor * inpSA = cur;
29
-
30
- // self-attention
31
- {
32
- // compute Q and K and RoPE them
33
- ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
34
- cb(Qcur, "Qcur", il);
35
- if (model.layers[il].bq) {
36
- Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
37
- cb(Qcur, "Qcur", il);
38
- }
39
-
40
- ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
41
- cb(Kcur, "Kcur", il);
42
- if (model.layers[il].bk) {
43
- Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
44
- cb(Kcur, "Kcur", il);
45
- }
46
-
47
- ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
48
- cb(Vcur, "Vcur", il);
49
- if (model.layers[il].bv) {
50
- Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
51
- cb(Vcur, "Vcur", il);
52
- }
53
-
54
- Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
55
- Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
56
- Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
57
-
58
- if (model.layers[il].attn_q_norm) {
59
- Qcur = build_norm(Qcur,
60
- model.layers[il].attn_q_norm,
61
- NULL,
62
- LLM_NORM, il);
63
- cb(Qcur, "Qcur", il);
64
- }
65
- if (model.layers[il].attn_k_norm) {
66
- Kcur = build_norm(Kcur,
67
- model.layers[il].attn_k_norm,
68
- NULL,
69
- LLM_NORM, il);
70
- cb(Kcur, "Kcur", il);
71
- }
72
-
73
- Qcur = ggml_rope_ext(
74
- ctx0, Qcur, inp_pos, nullptr,
75
- n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
76
- ext_factor, attn_factor, beta_fast, beta_slow
77
- );
78
-
79
- Kcur = ggml_rope_ext(
80
- ctx0, Kcur, inp_pos, nullptr,
81
- n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
82
- ext_factor, attn_factor, beta_fast, beta_slow
83
- );
84
-
85
- cb(Qcur, "Qcur", il);
86
- cb(Kcur, "Kcur", il);
87
- cb(Vcur, "Vcur", il);
88
-
89
- cur = build_attn(inp_attn,
90
- model.layers[il].wo, NULL,
91
- Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
92
- }
93
- if (il == n_layer - 1 && inp_out_ids) {
94
- cur = ggml_get_rows(ctx0, cur, inp_out_ids);
95
- inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
96
- inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
97
- }
98
- ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
99
- cb(ffn_inp, "ffn_inp", il);
100
-
101
- // feed-forward network
102
- {
103
- if (model.layers[il].ffn_norm) {
104
- cur = build_norm(ffn_inp,
105
- model.layers[il].ffn_norm,
106
- model.layers[il].ffn_norm_b,
107
- LLM_NORM, il);
108
- cb(cur, "ffn_norm", il);
109
- } else {
110
- // parallel residual
111
- cur = inpSA;
112
- }
113
- cur = build_ffn(cur,
114
- model.layers[il].ffn_up, NULL, NULL,
115
- model.layers[il].ffn_gate, NULL, NULL,
116
- model.layers[il].ffn_down, NULL, NULL,
117
- NULL,
118
- LLM_FFN_SILU, LLM_FFN_PAR, il);
119
- cb(cur, "ffn_out", il);
120
- }
121
- cur = ggml_add(ctx0, cur, ffn_inp);
122
-
123
- cur = build_cvec(cur, il);
124
- cb(cur, "l_out", il);
125
-
126
- // input for next layer
127
- inpL = cur;
128
- }
129
- cur = inpL;
130
-
131
- cur = build_norm(cur,
132
- model.output_norm,
133
- model.output_norm_b,
134
- LLM_NORM, -1);
135
-
136
- cb(cur, "result_norm", -1);
137
- res->t_embd = cur;
138
-
139
- // lm_head
140
- cur = build_lora_mm(model.output, cur);
141
-
142
- cb(cur, "result_output", -1);
143
- res->t_logits = cur;
144
-
145
- ggml_build_forward_expand(gf, cur);
146
- }
data/ext/sources/examples/talk-llama/models/starcoder.cpp DELETED
@@ -1,100 +0,0 @@
1
- #include "models.h"
2
-
3
- llm_build_starcoder::llm_build_starcoder(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
4
- const int64_t n_embd_head = hparams.n_embd_head_v();
5
- const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
6
-
7
- GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
8
-
9
- ggml_tensor * cur;
10
- ggml_tensor * inpL;
11
-
12
- inpL = build_inp_embd(model.tok_embd);
13
-
14
- // inp_pos - contains the positions
15
- ggml_tensor * inp_pos = build_inp_pos();
16
-
17
- auto * inp_attn = build_attn_inp_kv();
18
-
19
- ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
20
- cb(pos, "pos_embd", -1);
21
-
22
- inpL = ggml_add(ctx0, inpL, pos);
23
- cb(inpL, "inpL", -1);
24
-
25
- ggml_tensor * inp_out_ids = build_inp_out_ids();
26
-
27
- for (int il = 0; il < n_layer; ++il) {
28
- cur = build_norm(inpL,
29
- model.layers[il].attn_norm,
30
- model.layers[il].attn_norm_b,
31
- LLM_NORM, il);
32
- cb(cur, "attn_norm", il);
33
-
34
- // self-attention
35
- {
36
- cur = build_lora_mm(model.layers[il].wqkv, cur);
37
- cb(cur, "wqkv", il);
38
-
39
- cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
40
- cb(cur, "bqkv", il);
41
-
42
- ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
43
- ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
44
- ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
45
-
46
- cb(Qcur, "Qcur", il);
47
- cb(Kcur, "Kcur", il);
48
- cb(Vcur, "Vcur", il);
49
-
50
- cur = build_attn(inp_attn,
51
- model.layers[il].wo, model.layers[il].bo,
52
- Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
53
- }
54
- if (il == n_layer - 1 && inp_out_ids) {
55
- cur = ggml_get_rows(ctx0, cur, inp_out_ids);
56
- inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
57
- }
58
- // add the input
59
- ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
60
- cb(ffn_inp, "ffn_inp", il);
61
-
62
- // FF
63
- {
64
- cur = build_norm(ffn_inp,
65
- model.layers[il].ffn_norm,
66
- model.layers[il].ffn_norm_b,
67
- LLM_NORM, il);
68
- cb(cur, "ffn_norm", il);
69
-
70
- cur = build_ffn(cur,
71
- model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
72
- NULL, NULL, NULL,
73
- model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
74
- NULL,
75
- LLM_FFN_GELU, LLM_FFN_SEQ, il);
76
- cb(cur, "ffn_out", il);
77
- }
78
- cur = ggml_add(ctx0, cur, ffn_inp);
79
-
80
- cur = build_cvec(cur, il);
81
- cb(cur, "l_out", il);
82
-
83
- // input for next layer
84
- inpL = cur;
85
- }
86
- cur = build_norm(inpL,
87
- model.output_norm,
88
- model.output_norm_b,
89
- LLM_NORM, -1);
90
-
91
- cb(cur, "result_norm", -1);
92
- res->t_embd = cur;
93
-
94
- cur = build_lora_mm(model.output, cur);
95
-
96
- cb(cur, "result_output", -1);
97
- res->t_logits = cur;
98
-
99
- ggml_build_forward_expand(gf, cur);
100
- }
data/ext/sources/examples/talk-llama/models/starcoder2.cpp DELETED
@@ -1,121 +0,0 @@
1
- #include "models.h"
2
-
3
- llm_build_starcoder2::llm_build_starcoder2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
4
- const int64_t n_embd_head = hparams.n_embd_head_v();
5
-
6
- GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
7
- GGML_ASSERT(n_embd_head == n_rot);
8
-
9
- ggml_tensor * cur;
10
- ggml_tensor * inpL;
11
-
12
- inpL = build_inp_embd(model.tok_embd);
13
-
14
- // inp_pos - contains the positions
15
- ggml_tensor * inp_pos = build_inp_pos();
16
-
17
- auto * inp_attn = build_attn_inp_kv();
18
-
19
- ggml_tensor * inp_out_ids = build_inp_out_ids();
20
-
21
- for (int il = 0; il < n_layer; ++il) {
22
- ggml_tensor * inpSA = inpL;
23
-
24
- // norm
25
- cur = build_norm(inpL,
26
- model.layers[il].attn_norm, model.layers[il].attn_norm_b,
27
- LLM_NORM, il);
28
- cb(cur, "attn_norm", il);
29
-
30
- // self-attention
31
- {
32
- // compute Q and K and RoPE them
33
- ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
34
- cb(Qcur, "Qcur", il);
35
- if (model.layers[il].bq) {
36
- Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
37
- cb(Qcur, "Qcur", il);
38
- }
39
- ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
40
- cb(Kcur, "Kcur", il);
41
- if (model.layers[il].bk) {
42
- Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
43
- cb(Kcur, "Kcur", il);
44
- }
45
- ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
46
- cb(Vcur, "Vcur", il);
47
- if (model.layers[il].bv) {
48
- Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
49
- cb(Vcur, "Vcur", il);
50
- }
51
- Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
52
- Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
53
- Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
54
-
55
- Qcur = ggml_rope_ext(
56
- ctx0, Qcur, inp_pos, nullptr,
57
- n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
58
- ext_factor, attn_factor, beta_fast, beta_slow
59
- );
60
-
61
- Kcur = ggml_rope_ext(
62
- ctx0, Kcur, inp_pos, nullptr,
63
- n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
64
- ext_factor, attn_factor, beta_fast, beta_slow
65
- );
66
-
67
- cb(Qcur, "Qcur", il);
68
- cb(Kcur, "Kcur", il);
69
- cb(Vcur, "Vcur", il);
70
-
71
- cur = build_attn(inp_attn,
72
- model.layers[il].wo, model.layers[il].bo,
73
- Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
74
- }
75
- if (il == n_layer - 1 && inp_out_ids) {
76
- cur = ggml_get_rows(ctx0, cur, inp_out_ids);
77
- inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
78
- }
79
- ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
80
- cb(ffn_inp, "ffn_inp", il);
81
-
82
- // feed-forward network
83
-
84
- cur = build_norm(ffn_inp,
85
- model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,
86
- LLM_NORM, il);
87
- cb(cur, "ffn_norm", il);
88
-
89
- cur = build_ffn(cur,
90
- model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
91
- NULL, NULL, NULL,
92
- model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
93
- NULL,
94
- LLM_FFN_GELU, LLM_FFN_SEQ, il);
95
- cb(cur, "ffn_out", il);
96
-
97
- cur = ggml_add(ctx0, cur, ffn_inp);
98
-
99
- cur = build_cvec(cur, il);
100
- cb(cur, "l_out", il);
101
-
102
- // input for next layer
103
- inpL = cur;
104
- }
105
- cur = inpL;
106
-
107
- cur = build_norm(cur,
108
- model.output_norm, model.output_norm_b,
109
- LLM_NORM, -1);
110
-
111
- cb(cur, "result_norm", -1);
112
- res->t_embd = cur;
113
-
114
- // lm_head
115
- cur = build_lora_mm(model.output, cur);
116
-
117
- cb(cur, "result_output", -1);
118
- res->t_logits = cur;
119
-
120
- ggml_build_forward_expand(gf, cur);
121
- }
data/ext/sources/examples/talk-llama/models/step35-iswa.cpp DELETED
@@ -1,165 +0,0 @@
1
- #include "models.h"
2
-
3
- llm_build_step35_iswa::llm_build_step35_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
4
- ggml_tensor * cur;
5
- ggml_tensor * inpL;
6
-
7
- inpL = build_inp_embd(model.tok_embd);
8
- ggml_tensor * inp_pos = build_inp_pos();
9
- auto * inp_attn = build_attn_inp_kv_iswa();
10
- ggml_tensor * inp_out_ids = build_inp_out_ids();
11
-
12
- for (int il = 0; il < n_layer; ++il) {
13
- ggml_tensor * inpSA = inpL;
14
-
15
- const uint32_t n_head_l = hparams.n_head(il);
16
- const uint32_t n_head_kv_l = hparams.n_head_kv(il);
17
-
18
- const float freq_base_l = model.get_rope_freq_base(cparams, il);
19
- const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
20
-
21
- cur = inpL;
22
-
23
- // dump pre-attn RMSNorm input to pinpoint layer boundary issues
24
- cb(cur, "attn_norm_in", il);
25
-
26
- // self-attention
27
- {
28
- cur = build_norm(cur, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
29
- cb(cur, "attn_norm", il);
30
- ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
31
- ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
32
- ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
33
-
34
- cb(Qcur, "Qcur", il);
35
- cb(Kcur, "Kcur", il);
36
- cb(Vcur, "Vcur", il);
37
-
38
- Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l, n_tokens);
39
- Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens);
40
- Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens);
41
-
42
- // Q/K per-head RMSNorm (Step35 q_norm / k_norm)
43
- if (model.layers[il].attn_q_norm) {
44
- Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il);
45
- cb(Qcur, "Qcur_normed", il);
46
- }
47
- if (model.layers[il].attn_k_norm) {
48
- Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il);
49
- cb(Kcur, "Kcur_normed", il);
50
- }
51
-
52
- // RoPE (partial rotary factors per layer)
53
- const bool is_swa = hparams.is_swa(il);
54
- ggml_tensor * rope_factors = is_swa ? nullptr : model.get_rope_factors(cparams, il);
55
- const int64_t n_rot_l = hparams.n_rot(il);
56
- Qcur = ggml_rope_ext(
57
- ctx0, Qcur, inp_pos, rope_factors,
58
- n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
59
- ext_factor, attn_factor, beta_fast, beta_slow
60
- );
61
- Kcur = ggml_rope_ext(
62
- ctx0, Kcur, inp_pos, rope_factors,
63
- n_rot_l, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
64
- ext_factor, attn_factor, beta_fast, beta_slow
65
- );
66
- cb(Qcur, "Qcur_pos", il);
67
- cb(Kcur, "Kcur_pos", il);
68
-
69
- const float kq_scale = 1.0f / sqrtf(float(n_embd_head_k));
70
- ggml_tensor * attn_out = build_attn(inp_attn,
71
- nullptr, nullptr,
72
- Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
73
- cb(attn_out, "attn_out", il);
74
- // head-wise attention gate: sigmoid(g_proj(x)) in torch
75
- if (model.layers[il].wqkv_gate) {
76
- ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, cur); // [n_head_l, n_tokens]
77
- cb(gate, "attn_gate", il);
78
-
79
- gate = ggml_sigmoid(ctx0, gate);
80
- cb(gate, "attn_gate_sigmoid", il);
81
-
82
- // reshape + broadcast to [n_embd_head_v, n_head_l, n_tokens]
83
- ggml_tensor * attn_3d = ggml_reshape_3d(ctx0, attn_out, n_embd_head_v, n_head_l, n_tokens);
84
- ggml_tensor * gate_3d = ggml_reshape_3d(ctx0, gate, 1, n_head_l, n_tokens);
85
- cb(gate_3d, "attn_gate_3d", il);
86
-
87
- attn_3d = ggml_mul(ctx0, attn_3d, gate_3d);
88
- cb(attn_3d, "attn_gated_3d", il);
89
-
90
- attn_out = ggml_reshape_2d(ctx0, attn_3d, n_embd_head_v * n_head_l, n_tokens);
91
- cb(attn_out, "attn_gated", il);
92
- }
93
-
94
- // output projection
95
- cur = build_lora_mm(model.layers[il].wo, attn_out);
96
- cb(cur, "attn_proj", il);
97
- }
98
-
99
- if (il == n_layer - 1 && inp_out_ids) {
100
- cur = ggml_get_rows(ctx0, cur, inp_out_ids);
101
- inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
102
- }
103
-
104
- ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
105
- cb(ffn_inp, "ffn_inp", il);
106
-
107
- cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
108
- cb(cur, "ffn_norm", il);
109
-
110
- // feed-forward
111
- if (model.layers[il].ffn_gate_inp == nullptr) {
112
- // dense MLP
113
- cur = build_ffn(cur,
114
- model.layers[il].ffn_up, model.layers[il].ffn_up_b, nullptr,
115
- model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, nullptr,
116
- model.layers[il].ffn_down, model.layers[il].ffn_down_b, nullptr,
117
- nullptr,
118
- LLM_FFN_SILU, LLM_FFN_PAR, il);
119
- cb(cur, "ffn_out", il);
120
- } else {
121
- // MoE routed experts
122
- ggml_tensor * moe_out = build_moe_ffn(cur,
123
- model.layers[il].ffn_gate_inp,
124
- model.layers[il].ffn_up_exps,
125
- model.layers[il].ffn_gate_exps,
126
- model.layers[il].ffn_down_exps,
127
- model.layers[il].ffn_exp_probs_b,
128
- n_expert, n_expert_used,
129
- LLM_FFN_SILU, hparams.expert_weights_norm,
130
- hparams.expert_weights_scale,
131
- (llama_expert_gating_func_type) hparams.expert_gating_func,
132
- il);
133
- cb(moe_out, "ffn_moe_out", il);
134
-
135
- // shared expert MLP (always added on MoE layers in Step35)
136
- ggml_tensor * sh_out = build_ffn(cur,
137
- model.layers[il].ffn_up_shexp, nullptr, nullptr,
138
- model.layers[il].ffn_gate_shexp, nullptr, nullptr,
139
- model.layers[il].ffn_down_shexp, nullptr, nullptr,
140
- nullptr,
141
- LLM_FFN_SILU, LLM_FFN_PAR, il);
142
- cb(sh_out, "ffn_shared_out", il);
143
-
144
- cur = ggml_add(ctx0, moe_out, sh_out);
145
- cb(cur, "ffn_out", il);
146
- }
147
- cur = ggml_add(ctx0, cur, ffn_inp);
148
- cur = build_cvec(cur, il);
149
- cb(cur, "l_out", il);
150
-
151
- inpL = cur;
152
- }
153
-
154
- cur = inpL;
155
-
156
- cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
157
- cb(cur, "result_norm", -1);
158
- res->t_embd = cur;
159
-
160
- cur = build_lora_mm(model.output, cur);
161
- cb(cur, "result_output", -1);
162
- res->t_logits = cur;
163
-
164
- ggml_build_forward_expand(gf, cur);
165
- }
data/ext/sources/examples/talk-llama/models/t5-dec.cpp DELETED
@@ -1,166 +0,0 @@
1
- #include "models.h"
2
-
3
- llm_build_t5_dec::llm_build_t5_dec(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
4
- const int64_t n_embd_head = hparams.n_embd_head_v();
5
- //const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
6
-
7
- GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
8
-
9
- ggml_tensor * cur;
10
- ggml_tensor * inpL;
11
-
12
- inpL = build_inp_embd(model.tok_embd);
13
-
14
- ggml_tensor * embd_enc = build_inp_cross_embd();
15
- ggml_tensor * pos_bucket_dec = build_inp_pos_bucket_dec();
16
-
17
- const int64_t n_outputs_enc = embd_enc->ne[1];
18
-
19
- auto * inp_attn_self = build_attn_inp_kv();
20
- auto * inp_attn_cross = build_attn_inp_cross();
21
-
22
- ggml_tensor * inp_out_ids = build_inp_out_ids();
23
-
24
- const int64_t dec_n_layer = hparams.dec_n_layer;
25
-
26
- for (int il = 0; il < dec_n_layer; ++il) {
27
- ggml_tensor * inpSA = inpL;
28
-
29
- // norm
30
- cur = build_norm(inpL,
31
- model.layers[il].attn_norm, NULL,
32
- LLM_NORM_RMS, il);
33
- cb(cur, "attn_norm", il);
34
-
35
- // self-attention
36
- {
37
- ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
38
- cb(Qcur, "Qcur", il);
39
-
40
- ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
41
- cb(Kcur, "Kcur", il);
42
-
43
- ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
44
- cb(Vcur, "Vcur", il);
45
-
46
- Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
47
- Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
48
- Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
49
-
50
- ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b;
51
- ggml_tensor * kq_b = build_pos_bias(pos_bucket_dec, attn_rel_b);
52
-
53
- cur = build_attn(inp_attn_self,
54
- model.layers[il].wo, model.layers[il].bo,
55
- Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);
56
- cb(cur, "kqv_out", il);
57
- }
58
- cur = ggml_add(ctx0, cur, inpSA);
59
- cb(cur, "cross_inp", il);
60
-
61
- ggml_tensor * inpCA = cur;
62
-
63
- // norm
64
- cur = build_norm(cur,
65
- model.layers[il].attn_norm_cross, NULL,
66
- LLM_NORM_RMS, il);
67
- cb(cur, "attn_norm_cross", il);
68
-
69
- // cross-attention
70
- {
71
- ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_cross, cur);
72
- cb(Qcur, "Qcur", il);
73
-
74
- ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_cross, embd_enc);
75
- cb(Kcur, "Kcur", il);
76
-
77
- ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_cross, embd_enc);
78
- cb(Vcur, "Vcur", il);
79
-
80
- Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
81
- Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc);
82
- Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_outputs_enc);
83
-
84
- cur = build_attn(inp_attn_cross,
85
- model.layers[il].wo_cross, nullptr,
86
- Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
87
- cb(cur, "kqv_out", il);
88
-
89
- //ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
90
- //ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
91
-
92
- //ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
93
- //cb(kq, "kq", il);
94
-
95
- //kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias);
96
- //cb(kq, "kq_soft_max_ext", il);
97
-
98
- //ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc)));
99
- //cb(v, "v", il);
100
-
101
- //ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq);
102
- //cb(kqv, "kqv", il);
103
-
104
- //ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
105
- //cb(kqv_merged, "kqv_merged", il);
106
-
107
- //cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
108
- //cb(cur, "kqv_merged_cont", il);
109
-
110
- //ggml_build_forward_expand(gf, cur);
111
-
112
- //cur = build_lora_mm(model.layers[il].wo_cross, cur);
113
- //cb(cur, "kqv_out", il);
114
- }
115
- if (il == dec_n_layer - 1 && inp_out_ids) {
116
- cur = ggml_get_rows(ctx0, cur, inp_out_ids);
117
- inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids);
118
- }
119
- ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA);
120
- cb(ffn_inp, "ffn_inp", il);
121
-
122
- // feed-forward network
123
- {
124
- cur = build_norm(ffn_inp,
125
- model.layers[il].ffn_norm, NULL,
126
- LLM_NORM_RMS, il);
127
- cb(cur, "ffn_norm", il);
128
-
129
- // T5 uses relu, flan-T5 uses gelu-gated
130
- cur = build_ffn(cur,
131
- model.layers[il].ffn_up, NULL, NULL,
132
- model.layers[il].ffn_gate, NULL, NULL,
133
- model.layers[il].ffn_down, NULL, NULL,
134
- NULL,
135
- model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_RELU,
136
- model.layers[il].ffn_gate ? LLM_FFN_PAR : LLM_FFN_SEQ,
137
- il);
138
- cb(cur, "ffn_out", il);
139
- }
140
- cur = ggml_add(ctx0, cur, ffn_inp);
141
- cb(cur, "ffn_out", il);
142
-
143
- cur = build_cvec(cur, il);
144
- cb(cur, "l_out", il);
145
-
146
- // input for next layer
147
- inpL = cur;
148
- }
149
- cur = inpL;
150
- cb(cur, "result_embd", -1);
151
-
152
- cur = build_norm(cur,
153
- model.output_norm, NULL,
154
- LLM_NORM_RMS, -1);
155
-
156
- cb(cur, "result_norm", -1);
157
- res->t_embd = cur;
158
-
159
- // lm_head
160
- cur = build_lora_mm(model.output, cur);
161
-
162
- cb(cur, "result_output", -1);
163
- res->t_logits = cur;
164
-
165
- ggml_build_forward_expand(gf, cur);
166
- }