elelem-ggml 0.1.0

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 (1367) hide show
  1. checksums.yaml +7 -0
  2. data/LICENSE.txt +21 -0
  3. data/exe/elelem-ggml +8 -0
  4. data/ext/elelem/ggml/elelem.c +466 -0
  5. data/ext/elelem/ggml/extconf.rb +52 -0
  6. data/lib/elelem/ggml/chat_ml.rb +23 -0
  7. data/lib/elelem/ggml/gguf.rb +82 -0
  8. data/lib/elelem/ggml/provider.rb +63 -0
  9. data/lib/elelem/ggml/tokenizer.rb +95 -0
  10. data/lib/elelem/ggml/version.rb +7 -0
  11. data/lib/elelem/ggml.rb +37 -0
  12. data/vendor/ggml/AUTHORS +335 -0
  13. data/vendor/ggml/CMakeLists.txt +499 -0
  14. data/vendor/ggml/CONTRIBUTING.md +3 -0
  15. data/vendor/ggml/LICENSE +21 -0
  16. data/vendor/ggml/README.md +49 -0
  17. data/vendor/ggml/cmake/FindNCCL.cmake +36 -0
  18. data/vendor/ggml/cmake/GitVars.cmake +22 -0
  19. data/vendor/ggml/cmake/common.cmake +50 -0
  20. data/vendor/ggml/cmake/ggml-config.cmake.in +216 -0
  21. data/vendor/ggml/ggml.pc.in +10 -0
  22. data/vendor/ggml/include/ggml-alloc.h +86 -0
  23. data/vendor/ggml/include/ggml-backend.h +437 -0
  24. data/vendor/ggml/include/ggml-blas.h +25 -0
  25. data/vendor/ggml/include/ggml-cann.h +123 -0
  26. data/vendor/ggml/include/ggml-cpp.h +39 -0
  27. data/vendor/ggml/include/ggml-cpu.h +152 -0
  28. data/vendor/ggml/include/ggml-cuda.h +47 -0
  29. data/vendor/ggml/include/ggml-et.h +28 -0
  30. data/vendor/ggml/include/ggml-hexagon.h +19 -0
  31. data/vendor/ggml/include/ggml-metal.h +61 -0
  32. data/vendor/ggml/include/ggml-opencl.h +26 -0
  33. data/vendor/ggml/include/ggml-openvino.h +37 -0
  34. data/vendor/ggml/include/ggml-opt.h +256 -0
  35. data/vendor/ggml/include/ggml-rpc.h +35 -0
  36. data/vendor/ggml/include/ggml-sycl.h +57 -0
  37. data/vendor/ggml/include/ggml-virtgpu.h +14 -0
  38. data/vendor/ggml/include/ggml-vulkan.h +29 -0
  39. data/vendor/ggml/include/ggml-webgpu.h +19 -0
  40. data/vendor/ggml/include/ggml-zdnn.h +17 -0
  41. data/vendor/ggml/include/ggml-zendnn.h +22 -0
  42. data/vendor/ggml/include/ggml.h +2963 -0
  43. data/vendor/ggml/include/gguf.h +211 -0
  44. data/vendor/ggml/requirements.txt +12 -0
  45. data/vendor/ggml/scripts/gen-authors.sh +9 -0
  46. data/vendor/ggml/scripts/make-release-checks.sh +96 -0
  47. data/vendor/ggml/scripts/make-release-desc.sh +69 -0
  48. data/vendor/ggml/scripts/make-release-summary.txt +33 -0
  49. data/vendor/ggml/scripts/release.sh +195 -0
  50. data/vendor/ggml/scripts/sync-llama-am.sh +167 -0
  51. data/vendor/ggml/scripts/sync-llama.last +1 -0
  52. data/vendor/ggml/scripts/sync-llama.sh +21 -0
  53. data/vendor/ggml/scripts/sync-whisper-am.sh +138 -0
  54. data/vendor/ggml/scripts/sync-whisper.last +1 -0
  55. data/vendor/ggml/scripts/sync-whisper.sh +17 -0
  56. data/vendor/ggml/src/CMakeLists.txt +634 -0
  57. data/vendor/ggml/src/ggml-alloc.c +1249 -0
  58. data/vendor/ggml/src/ggml-backend-dl.cpp +48 -0
  59. data/vendor/ggml/src/ggml-backend-dl.h +45 -0
  60. data/vendor/ggml/src/ggml-backend-impl.h +291 -0
  61. data/vendor/ggml/src/ggml-backend-meta.cpp +2511 -0
  62. data/vendor/ggml/src/ggml-backend-reg.cpp +605 -0
  63. data/vendor/ggml/src/ggml-backend.cpp +2517 -0
  64. data/vendor/ggml/src/ggml-blas/CMakeLists.txt +101 -0
  65. data/vendor/ggml/src/ggml-blas/ggml-blas.cpp +530 -0
  66. data/vendor/ggml/src/ggml-cann/CMakeLists.txt +89 -0
  67. data/vendor/ggml/src/ggml-cann/acl_tensor.cpp +195 -0
  68. data/vendor/ggml/src/ggml-cann/acl_tensor.h +349 -0
  69. data/vendor/ggml/src/ggml-cann/aclnn_ops.cpp +4479 -0
  70. data/vendor/ggml/src/ggml-cann/aclnn_ops.h +1191 -0
  71. data/vendor/ggml/src/ggml-cann/common.h +651 -0
  72. data/vendor/ggml/src/ggml-cann/ggml-cann.cpp +3070 -0
  73. data/vendor/ggml/src/ggml-common.h +1911 -0
  74. data/vendor/ggml/src/ggml-cpu/CMakeLists.txt +685 -0
  75. data/vendor/ggml/src/ggml-cpu/amx/amx.cpp +249 -0
  76. data/vendor/ggml/src/ggml-cpu/amx/amx.h +8 -0
  77. data/vendor/ggml/src/ggml-cpu/amx/common.h +115 -0
  78. data/vendor/ggml/src/ggml-cpu/amx/mmq.cpp +2511 -0
  79. data/vendor/ggml/src/ggml-cpu/amx/mmq.h +10 -0
  80. data/vendor/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp +41 -0
  81. data/vendor/ggml/src/ggml-cpu/arch/arm/quants.c +4319 -0
  82. data/vendor/ggml/src/ggml-cpu/arch/arm/repack.cpp +5156 -0
  83. data/vendor/ggml/src/ggml-cpu/arch/loongarch/quants.c +2309 -0
  84. data/vendor/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp +82 -0
  85. data/vendor/ggml/src/ggml-cpu/arch/powerpc/quants.c +2304 -0
  86. data/vendor/ggml/src/ggml-cpu/arch/riscv/cpu-feats.cpp +38 -0
  87. data/vendor/ggml/src/ggml-cpu/arch/riscv/quants.c +6596 -0
  88. data/vendor/ggml/src/ggml-cpu/arch/riscv/repack.cpp +1703 -0
  89. data/vendor/ggml/src/ggml-cpu/arch/s390/cpu-feats.cpp +50 -0
  90. data/vendor/ggml/src/ggml-cpu/arch/s390/quants.c +1465 -0
  91. data/vendor/ggml/src/ggml-cpu/arch/wasm/quants.c +1292 -0
  92. data/vendor/ggml/src/ggml-cpu/arch/x86/cpu-feats.cpp +327 -0
  93. data/vendor/ggml/src/ggml-cpu/arch/x86/quants.c +4108 -0
  94. data/vendor/ggml/src/ggml-cpu/arch/x86/repack.cpp +6407 -0
  95. data/vendor/ggml/src/ggml-cpu/arch-fallback.h +353 -0
  96. data/vendor/ggml/src/ggml-cpu/binary-ops.cpp +154 -0
  97. data/vendor/ggml/src/ggml-cpu/binary-ops.h +16 -0
  98. data/vendor/ggml/src/ggml-cpu/cmake/FindSIMD.cmake +100 -0
  99. data/vendor/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
  100. data/vendor/ggml/src/ggml-cpu/common.h +95 -0
  101. data/vendor/ggml/src/ggml-cpu/ggml-cpu-impl.h +539 -0
  102. data/vendor/ggml/src/ggml-cpu/ggml-cpu.c +3936 -0
  103. data/vendor/ggml/src/ggml-cpu/ggml-cpu.cpp +712 -0
  104. data/vendor/ggml/src/ggml-cpu/hbm.cpp +55 -0
  105. data/vendor/ggml/src/ggml-cpu/hbm.h +8 -0
  106. data/vendor/ggml/src/ggml-cpu/iqp.cpp +1253 -0
  107. data/vendor/ggml/src/ggml-cpu/iqp.h +39 -0
  108. data/vendor/ggml/src/ggml-cpu/kleidiai/CMakeLists.txt +14 -0
  109. data/vendor/ggml/src/ggml-cpu/kleidiai/kernels.cpp +1101 -0
  110. data/vendor/ggml/src/ggml-cpu/kleidiai/kernels.h +101 -0
  111. data/vendor/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +1919 -0
  112. data/vendor/ggml/src/ggml-cpu/kleidiai/kleidiai.h +17 -0
  113. data/vendor/ggml/src/ggml-cpu/llamafile/sgemm.cpp +4164 -0
  114. data/vendor/ggml/src/ggml-cpu/llamafile/sgemm.h +25 -0
  115. data/vendor/ggml/src/ggml-cpu/ops.cpp +12164 -0
  116. data/vendor/ggml/src/ggml-cpu/ops.h +124 -0
  117. data/vendor/ggml/src/ggml-cpu/quants.c +1339 -0
  118. data/vendor/ggml/src/ggml-cpu/quants.h +106 -0
  119. data/vendor/ggml/src/ggml-cpu/repack.cpp +4836 -0
  120. data/vendor/ggml/src/ggml-cpu/repack.h +245 -0
  121. data/vendor/ggml/src/ggml-cpu/simd-gemm.h +226 -0
  122. data/vendor/ggml/src/ggml-cpu/simd-mappings.h +1319 -0
  123. data/vendor/ggml/src/ggml-cpu/spacemit/ime.cpp +1742 -0
  124. data/vendor/ggml/src/ggml-cpu/spacemit/ime.h +21 -0
  125. data/vendor/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +1027 -0
  126. data/vendor/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp +5768 -0
  127. data/vendor/ggml/src/ggml-cpu/spacemit/ime_env.cpp +320 -0
  128. data/vendor/ggml/src/ggml-cpu/spacemit/ime_env.h +55 -0
  129. data/vendor/ggml/src/ggml-cpu/spacemit/ime_kernels.h +189 -0
  130. data/vendor/ggml/src/ggml-cpu/spacemit/repack.cpp +1795 -0
  131. data/vendor/ggml/src/ggml-cpu/spacemit/repack.h +14 -0
  132. data/vendor/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp +3178 -0
  133. data/vendor/ggml/src/ggml-cpu/spacemit/rvv_kernels.h +95 -0
  134. data/vendor/ggml/src/ggml-cpu/spacemit/spine_barrier.h +34 -0
  135. data/vendor/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp +760 -0
  136. data/vendor/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h +32 -0
  137. data/vendor/ggml/src/ggml-cpu/spacemit/spine_tcm.h +409 -0
  138. data/vendor/ggml/src/ggml-cpu/traits.cpp +36 -0
  139. data/vendor/ggml/src/ggml-cpu/traits.h +38 -0
  140. data/vendor/ggml/src/ggml-cpu/unary-ops.cpp +337 -0
  141. data/vendor/ggml/src/ggml-cpu/unary-ops.h +35 -0
  142. data/vendor/ggml/src/ggml-cpu/vec.cpp +613 -0
  143. data/vendor/ggml/src/ggml-cpu/vec.h +1570 -0
  144. data/vendor/ggml/src/ggml-cuda/CMakeLists.txt +267 -0
  145. data/vendor/ggml/src/ggml-cuda/acc.cu +61 -0
  146. data/vendor/ggml/src/ggml-cuda/acc.cuh +5 -0
  147. data/vendor/ggml/src/ggml-cuda/add-id.cu +58 -0
  148. data/vendor/ggml/src/ggml-cuda/add-id.cuh +3 -0
  149. data/vendor/ggml/src/ggml-cuda/allreduce.cu +971 -0
  150. data/vendor/ggml/src/ggml-cuda/allreduce.cuh +29 -0
  151. data/vendor/ggml/src/ggml-cuda/arange.cu +34 -0
  152. data/vendor/ggml/src/ggml-cuda/arange.cuh +5 -0
  153. data/vendor/ggml/src/ggml-cuda/argmax.cu +91 -0
  154. data/vendor/ggml/src/ggml-cuda/argmax.cuh +3 -0
  155. data/vendor/ggml/src/ggml-cuda/argsort.cu +292 -0
  156. data/vendor/ggml/src/ggml-cuda/argsort.cuh +20 -0
  157. data/vendor/ggml/src/ggml-cuda/binbcast.cu +574 -0
  158. data/vendor/ggml/src/ggml-cuda/binbcast.cuh +12 -0
  159. data/vendor/ggml/src/ggml-cuda/clamp.cu +45 -0
  160. data/vendor/ggml/src/ggml-cuda/clamp.cuh +5 -0
  161. data/vendor/ggml/src/ggml-cuda/col2im-1d.cu +81 -0
  162. data/vendor/ggml/src/ggml-cuda/col2im-1d.cuh +3 -0
  163. data/vendor/ggml/src/ggml-cuda/common.cuh +1678 -0
  164. data/vendor/ggml/src/ggml-cuda/concat.cu +242 -0
  165. data/vendor/ggml/src/ggml-cuda/concat.cuh +5 -0
  166. data/vendor/ggml/src/ggml-cuda/conv-transpose-1d.cu +88 -0
  167. data/vendor/ggml/src/ggml-cuda/conv-transpose-1d.cuh +5 -0
  168. data/vendor/ggml/src/ggml-cuda/conv2d-dw.cu +161 -0
  169. data/vendor/ggml/src/ggml-cuda/conv2d-dw.cuh +5 -0
  170. data/vendor/ggml/src/ggml-cuda/conv2d-transpose.cu +115 -0
  171. data/vendor/ggml/src/ggml-cuda/conv2d-transpose.cuh +5 -0
  172. data/vendor/ggml/src/ggml-cuda/conv2d.cu +167 -0
  173. data/vendor/ggml/src/ggml-cuda/conv2d.cuh +5 -0
  174. data/vendor/ggml/src/ggml-cuda/convert.cu +705 -0
  175. data/vendor/ggml/src/ggml-cuda/convert.cuh +66 -0
  176. data/vendor/ggml/src/ggml-cuda/count-equal.cu +64 -0
  177. data/vendor/ggml/src/ggml-cuda/count-equal.cuh +5 -0
  178. data/vendor/ggml/src/ggml-cuda/cp-async.cuh +57 -0
  179. data/vendor/ggml/src/ggml-cuda/cpy-utils.cuh +217 -0
  180. data/vendor/ggml/src/ggml-cuda/cpy.cu +617 -0
  181. data/vendor/ggml/src/ggml-cuda/cpy.cuh +7 -0
  182. data/vendor/ggml/src/ggml-cuda/cross-entropy-loss.cu +177 -0
  183. data/vendor/ggml/src/ggml-cuda/cross-entropy-loss.cuh +7 -0
  184. data/vendor/ggml/src/ggml-cuda/cumsum.cu +307 -0
  185. data/vendor/ggml/src/ggml-cuda/cumsum.cuh +5 -0
  186. data/vendor/ggml/src/ggml-cuda/dequantize.cuh +452 -0
  187. data/vendor/ggml/src/ggml-cuda/diag.cu +77 -0
  188. data/vendor/ggml/src/ggml-cuda/diag.cuh +5 -0
  189. data/vendor/ggml/src/ggml-cuda/diagmask.cu +40 -0
  190. data/vendor/ggml/src/ggml-cuda/diagmask.cuh +5 -0
  191. data/vendor/ggml/src/ggml-cuda/dsv4-hc.cu +294 -0
  192. data/vendor/ggml/src/ggml-cuda/dsv4-hc.cuh +6 -0
  193. data/vendor/ggml/src/ggml-cuda/fattn-common.cuh +1289 -0
  194. data/vendor/ggml/src/ggml-cuda/fattn-mma-f16.cuh +2136 -0
  195. data/vendor/ggml/src/ggml-cuda/fattn-swizzle.cuh +126 -0
  196. data/vendor/ggml/src/ggml-cuda/fattn-tile.cu +60 -0
  197. data/vendor/ggml/src/ggml-cuda/fattn-tile.cuh +1355 -0
  198. data/vendor/ggml/src/ggml-cuda/fattn-vec.cuh +611 -0
  199. data/vendor/ggml/src/ggml-cuda/fattn.cu +722 -0
  200. data/vendor/ggml/src/ggml-cuda/fattn.cuh +7 -0
  201. data/vendor/ggml/src/ggml-cuda/fill.cu +37 -0
  202. data/vendor/ggml/src/ggml-cuda/fill.cuh +3 -0
  203. data/vendor/ggml/src/ggml-cuda/fwht.cu +101 -0
  204. data/vendor/ggml/src/ggml-cuda/fwht.cuh +4 -0
  205. data/vendor/ggml/src/ggml-cuda/gated_delta_net.cu +327 -0
  206. data/vendor/ggml/src/ggml-cuda/gated_delta_net.cuh +14 -0
  207. data/vendor/ggml/src/ggml-cuda/getrows.cu +490 -0
  208. data/vendor/ggml/src/ggml-cuda/getrows.cuh +15 -0
  209. data/vendor/ggml/src/ggml-cuda/ggml-cuda.cu +5781 -0
  210. data/vendor/ggml/src/ggml-cuda/gla.cu +93 -0
  211. data/vendor/ggml/src/ggml-cuda/gla.cuh +3 -0
  212. data/vendor/ggml/src/ggml-cuda/im2col.cu +267 -0
  213. data/vendor/ggml/src/ggml-cuda/im2col.cuh +6 -0
  214. data/vendor/ggml/src/ggml-cuda/lightning-indexer.cu +588 -0
  215. data/vendor/ggml/src/ggml-cuda/lightning-indexer.cuh +4 -0
  216. data/vendor/ggml/src/ggml-cuda/mean.cu +77 -0
  217. data/vendor/ggml/src/ggml-cuda/mean.cuh +3 -0
  218. data/vendor/ggml/src/ggml-cuda/mma.cuh +1456 -0
  219. data/vendor/ggml/src/ggml-cuda/mmf.cu +191 -0
  220. data/vendor/ggml/src/ggml-cuda/mmf.cuh +908 -0
  221. data/vendor/ggml/src/ggml-cuda/mmid.cu +180 -0
  222. data/vendor/ggml/src/ggml-cuda/mmid.cuh +5 -0
  223. data/vendor/ggml/src/ggml-cuda/mmq-config-ampere.cuh +383 -0
  224. data/vendor/ggml/src/ggml-cuda/mmq-config-blackwell.cuh +37 -0
  225. data/vendor/ggml/src/ggml-cuda/mmq-config-cdna.cuh +185 -0
  226. data/vendor/ggml/src/ggml-cuda/mmq-config-pascal-dp4a.cuh +273 -0
  227. data/vendor/ggml/src/ggml-cuda/mmq-config-pascal-older.cuh +273 -0
  228. data/vendor/ggml/src/ggml-cuda/mmq-config-rdna2.cuh +273 -0
  229. data/vendor/ggml/src/ggml-cuda/mmq-config-rdna3-5.cuh +290 -0
  230. data/vendor/ggml/src/ggml-cuda/mmq-config-rdna3.cuh +274 -0
  231. data/vendor/ggml/src/ggml-cuda/mmq-config-rdna4.cuh +290 -0
  232. data/vendor/ggml/src/ggml-cuda/mmq-load-tiles.cuh +1768 -0
  233. data/vendor/ggml/src/ggml-cuda/mmq-vec-dot.cuh +1240 -0
  234. data/vendor/ggml/src/ggml-cuda/mmq.cu +386 -0
  235. data/vendor/ggml/src/ggml-cuda/mmq.cuh +1598 -0
  236. data/vendor/ggml/src/ggml-cuda/mmvf.cu +875 -0
  237. data/vendor/ggml/src/ggml-cuda/mmvf.cuh +14 -0
  238. data/vendor/ggml/src/ggml-cuda/mmvq.cu +1504 -0
  239. data/vendor/ggml/src/ggml-cuda/mmvq.cuh +18 -0
  240. data/vendor/ggml/src/ggml-cuda/moe-weighted-reduction.cu +65 -0
  241. data/vendor/ggml/src/ggml-cuda/moe-weighted-reduction.cuh +7 -0
  242. data/vendor/ggml/src/ggml-cuda/norm.cu +698 -0
  243. data/vendor/ggml/src/ggml-cuda/norm.cuh +18 -0
  244. data/vendor/ggml/src/ggml-cuda/opt-step-adamw.cu +78 -0
  245. data/vendor/ggml/src/ggml-cuda/opt-step-adamw.cuh +5 -0
  246. data/vendor/ggml/src/ggml-cuda/opt-step-sgd.cu +49 -0
  247. data/vendor/ggml/src/ggml-cuda/opt-step-sgd.cuh +5 -0
  248. data/vendor/ggml/src/ggml-cuda/out-prod.cu +125 -0
  249. data/vendor/ggml/src/ggml-cuda/out-prod.cuh +3 -0
  250. data/vendor/ggml/src/ggml-cuda/pad.cu +106 -0
  251. data/vendor/ggml/src/ggml-cuda/pad.cuh +5 -0
  252. data/vendor/ggml/src/ggml-cuda/pad_reflect_1d.cu +91 -0
  253. data/vendor/ggml/src/ggml-cuda/pad_reflect_1d.cuh +5 -0
  254. data/vendor/ggml/src/ggml-cuda/pool1d.cu +85 -0
  255. data/vendor/ggml/src/ggml-cuda/pool1d.cuh +5 -0
  256. data/vendor/ggml/src/ggml-cuda/pool2d.cu +94 -0
  257. data/vendor/ggml/src/ggml-cuda/pool2d.cuh +5 -0
  258. data/vendor/ggml/src/ggml-cuda/quantize.cu +697 -0
  259. data/vendor/ggml/src/ggml-cuda/quantize.cuh +70 -0
  260. data/vendor/ggml/src/ggml-cuda/reduce_rows.cuh +43 -0
  261. data/vendor/ggml/src/ggml-cuda/roll.cu +67 -0
  262. data/vendor/ggml/src/ggml-cuda/roll.cuh +5 -0
  263. data/vendor/ggml/src/ggml-cuda/rope.cu +941 -0
  264. data/vendor/ggml/src/ggml-cuda/rope.cuh +11 -0
  265. data/vendor/ggml/src/ggml-cuda/scale.cu +37 -0
  266. data/vendor/ggml/src/ggml-cuda/scale.cuh +5 -0
  267. data/vendor/ggml/src/ggml-cuda/set-rows.cu +398 -0
  268. data/vendor/ggml/src/ggml-cuda/set-rows.cuh +7 -0
  269. data/vendor/ggml/src/ggml-cuda/set.cu +39 -0
  270. data/vendor/ggml/src/ggml-cuda/set.cuh +7 -0
  271. data/vendor/ggml/src/ggml-cuda/snake.cu +72 -0
  272. data/vendor/ggml/src/ggml-cuda/snake.cuh +8 -0
  273. data/vendor/ggml/src/ggml-cuda/softcap.cu +37 -0
  274. data/vendor/ggml/src/ggml-cuda/softcap.cuh +5 -0
  275. data/vendor/ggml/src/ggml-cuda/softmax.cu +480 -0
  276. data/vendor/ggml/src/ggml-cuda/softmax.cuh +7 -0
  277. data/vendor/ggml/src/ggml-cuda/solve_tri.cu +273 -0
  278. data/vendor/ggml/src/ggml-cuda/solve_tri.cuh +3 -0
  279. data/vendor/ggml/src/ggml-cuda/ssm-conv.cu +206 -0
  280. data/vendor/ggml/src/ggml-cuda/ssm-conv.cuh +3 -0
  281. data/vendor/ggml/src/ggml-cuda/ssm-scan.cu +859 -0
  282. data/vendor/ggml/src/ggml-cuda/ssm-scan.cuh +3 -0
  283. data/vendor/ggml/src/ggml-cuda/sum.cu +41 -0
  284. data/vendor/ggml/src/ggml-cuda/sum.cuh +5 -0
  285. data/vendor/ggml/src/ggml-cuda/sumrows.cu +47 -0
  286. data/vendor/ggml/src/ggml-cuda/sumrows.cuh +4 -0
  287. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +6 -0
  288. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +6 -0
  289. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +12 -0
  290. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu +10 -0
  291. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu +11 -0
  292. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +12 -0
  293. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +6 -0
  294. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +6 -0
  295. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +12 -0
  296. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +12 -0
  297. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu +10 -0
  298. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu +11 -0
  299. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +6 -0
  300. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu +11 -0
  301. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +12 -0
  302. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +12 -0
  303. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu +10 -0
  304. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu +10 -0
  305. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu +11 -0
  306. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +12 -0
  307. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +12 -0
  308. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq112-dv112.cu +5 -0
  309. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq128-dv128.cu +5 -0
  310. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu +5 -0
  311. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq256-dv256.cu +5 -0
  312. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu +5 -0
  313. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq40-dv40.cu +5 -0
  314. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +5 -0
  315. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq576-dv512.cu +5 -0
  316. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq64-dv64.cu +5 -0
  317. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq72-dv72.cu +5 -0
  318. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq80-dv80.cu +5 -0
  319. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq96-dv96.cu +5 -0
  320. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu +7 -0
  321. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu +7 -0
  322. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu +7 -0
  323. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu +7 -0
  324. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu +7 -0
  325. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu +7 -0
  326. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu +7 -0
  327. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu +7 -0
  328. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-f16.cu +7 -0
  329. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_0.cu +7 -0
  330. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_1.cu +7 -0
  331. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_0.cu +7 -0
  332. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_1.cu +7 -0
  333. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q8_0.cu +7 -0
  334. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu +7 -0
  335. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-f16.cu +7 -0
  336. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_0.cu +7 -0
  337. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_1.cu +7 -0
  338. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_0.cu +7 -0
  339. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_1.cu +7 -0
  340. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q8_0.cu +7 -0
  341. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu +7 -0
  342. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-f16.cu +7 -0
  343. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_0.cu +7 -0
  344. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_1.cu +7 -0
  345. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_0.cu +7 -0
  346. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_1.cu +7 -0
  347. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q8_0.cu +7 -0
  348. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu +7 -0
  349. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-f16.cu +7 -0
  350. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_0.cu +7 -0
  351. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_1.cu +7 -0
  352. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_0.cu +7 -0
  353. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_1.cu +7 -0
  354. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q8_0.cu +7 -0
  355. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu +7 -0
  356. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-f16.cu +7 -0
  357. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_0.cu +7 -0
  358. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_1.cu +7 -0
  359. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_0.cu +7 -0
  360. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_1.cu +7 -0
  361. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q8_0.cu +7 -0
  362. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu +7 -0
  363. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-f16.cu +7 -0
  364. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_0.cu +7 -0
  365. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_1.cu +7 -0
  366. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_0.cu +7 -0
  367. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_1.cu +7 -0
  368. data/vendor/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q8_0.cu +7 -0
  369. data/vendor/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +111 -0
  370. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_1.cu +5 -0
  371. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_10.cu +5 -0
  372. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_11.cu +5 -0
  373. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_12.cu +5 -0
  374. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_13.cu +5 -0
  375. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_14.cu +5 -0
  376. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_15.cu +5 -0
  377. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_16.cu +5 -0
  378. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_2.cu +5 -0
  379. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_3.cu +5 -0
  380. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_4.cu +5 -0
  381. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_5.cu +5 -0
  382. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_6.cu +5 -0
  383. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_7.cu +5 -0
  384. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_8.cu +5 -0
  385. data/vendor/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_9.cu +5 -0
  386. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s.cu +5 -0
  387. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_s.cu +5 -0
  388. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu +5 -0
  389. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu +5 -0
  390. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_s.cu +5 -0
  391. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu +5 -0
  392. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu +5 -0
  393. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu +5 -0
  394. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu +5 -0
  395. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu +5 -0
  396. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu +5 -0
  397. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_0.cu +5 -0
  398. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_k.cu +5 -0
  399. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q3_k.cu +5 -0
  400. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_0.cu +5 -0
  401. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_1.cu +5 -0
  402. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_k.cu +5 -0
  403. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_0.cu +5 -0
  404. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_1.cu +5 -0
  405. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_k.cu +5 -0
  406. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q6_k.cu +5 -0
  407. data/vendor/ggml/src/ggml-cuda/template-instances/mmq-instance-q8_0.cu +5 -0
  408. data/vendor/ggml/src/ggml-cuda/top-k.cu +275 -0
  409. data/vendor/ggml/src/ggml-cuda/top-k.cuh +3 -0
  410. data/vendor/ggml/src/ggml-cuda/topk-moe.cu +442 -0
  411. data/vendor/ggml/src/ggml-cuda/topk-moe.cuh +31 -0
  412. data/vendor/ggml/src/ggml-cuda/tri.cu +136 -0
  413. data/vendor/ggml/src/ggml-cuda/tri.cuh +5 -0
  414. data/vendor/ggml/src/ggml-cuda/tsembd.cu +47 -0
  415. data/vendor/ggml/src/ggml-cuda/tsembd.cuh +5 -0
  416. data/vendor/ggml/src/ggml-cuda/unary.cu +721 -0
  417. data/vendor/ggml/src/ggml-cuda/unary.cuh +123 -0
  418. data/vendor/ggml/src/ggml-cuda/upscale.cu +293 -0
  419. data/vendor/ggml/src/ggml-cuda/upscale.cuh +5 -0
  420. data/vendor/ggml/src/ggml-cuda/vecdotq.cuh +1371 -0
  421. data/vendor/ggml/src/ggml-cuda/vendors/cuda.h +28 -0
  422. data/vendor/ggml/src/ggml-cuda/vendors/hip.h +301 -0
  423. data/vendor/ggml/src/ggml-cuda/vendors/musa.h +151 -0
  424. data/vendor/ggml/src/ggml-cuda/wkv.cu +253 -0
  425. data/vendor/ggml/src/ggml-cuda/wkv.cuh +7 -0
  426. data/vendor/ggml/src/ggml-et/CMakeLists.txt +246 -0
  427. data/vendor/ggml/src/ggml-et/cmake/embed_one_kernel.cmake +15 -0
  428. data/vendor/ggml/src/ggml-et/cmake/ggml-et-kernels-embed.cpp.in +6 -0
  429. data/vendor/ggml/src/ggml-et/cmake/ggml-et-kernels-embed.hpp.in +12 -0
  430. data/vendor/ggml/src/ggml-et/cmake/ggml-et-uberkernel-kernel-map.cpp.in +18 -0
  431. data/vendor/ggml/src/ggml-et/cmake/ggml-et-uberkernel-kernel-map.h.in +13 -0
  432. data/vendor/ggml/src/ggml-et/et-kernels/CMakeLists.txt +137 -0
  433. data/vendor/ggml/src/ggml-et/et-kernels/scripts/check_unimplemented_instructions.sh +36 -0
  434. data/vendor/ggml/src/ggml-et/et-kernels/src/RunBackend.sh +23 -0
  435. data/vendor/ggml/src/ggml-et/et-kernels/src/block_ops.h +997 -0
  436. data/vendor/ggml/src/ggml-et/et-kernels/src/clamp_f32.c +120 -0
  437. data/vendor/ggml/src/ggml-et/et-kernels/src/concat_f32.c +175 -0
  438. data/vendor/ggml/src/ggml-et/et-kernels/src/cont_f16.c +107 -0
  439. data/vendor/ggml/src/ggml-et/et-kernels/src/cont_f32.c +248 -0
  440. data/vendor/ggml/src/ggml-et/et-kernels/src/conv_2d_f32_me.c +807 -0
  441. data/vendor/ggml/src/ggml-et/et-kernels/src/cpy_f32_f16.c +110 -0
  442. data/vendor/ggml/src/ggml-et/et-kernels/src/crt.S +15 -0
  443. data/vendor/ggml/src/ggml-et/et-kernels/src/cumsum_f32.c +96 -0
  444. data/vendor/ggml/src/ggml-et/et-kernels/src/diag_f32.c +90 -0
  445. data/vendor/ggml/src/ggml-et/et-kernels/src/el_map_f32.c +377 -0
  446. data/vendor/ggml/src/ggml-et/et-kernels/src/fill_f32.c +87 -0
  447. data/vendor/ggml/src/ggml-et/et-kernels/src/flash_attn_ext_f16_me.c +1000 -0
  448. data/vendor/ggml/src/ggml-et/et-kernels/src/flash_attn_ext_f32.c +217 -0
  449. data/vendor/ggml/src/ggml-et/et-kernels/src/gated_delta_net_f32.c +346 -0
  450. data/vendor/ggml/src/ggml-et/et-kernels/src/get_rows_f32.c +612 -0
  451. data/vendor/ggml/src/ggml-et/et-kernels/src/ggml_tensor.h +44 -0
  452. data/vendor/ggml/src/ggml-et/et-kernels/src/glu_f32.c +606 -0
  453. data/vendor/ggml/src/ggml-et/et-kernels/src/group_norm_f32.c +171 -0
  454. data/vendor/ggml/src/ggml-et/et-kernels/src/im2col.c +130 -0
  455. data/vendor/ggml/src/ggml-et/et-kernels/src/l2_norm_f32.c +237 -0
  456. data/vendor/ggml/src/ggml-et/et-kernels/src/linker.ld +85 -0
  457. data/vendor/ggml/src/ggml-et/et-kernels/src/math_fp.h +299 -0
  458. data/vendor/ggml/src/ggml-et/et-kernels/src/mean_f32.c +220 -0
  459. data/vendor/ggml/src/ggml-et/et-kernels/src/memops.c +181 -0
  460. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_Q4_0.c +358 -0
  461. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_Q4_0_matrix_engine.c +368 -0
  462. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_Q8_0.c +413 -0
  463. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f16.c +142 -0
  464. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f16_matrix_engine.c +329 -0
  465. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f32.c +137 -0
  466. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_f32_matrix_engine.c +155 -0
  467. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_id_Q4_0.c +169 -0
  468. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_id_Q8_0.c +160 -0
  469. data/vendor/ggml/src/ggml-et/et-kernels/src/mul_mat_id_f32.c +288 -0
  470. data/vendor/ggml/src/ggml-et/et-kernels/src/norm_f32.c +328 -0
  471. data/vendor/ggml/src/ggml-et/et-kernels/src/pad_f32.c +165 -0
  472. data/vendor/ggml/src/ggml-et/et-kernels/src/platform.h +545 -0
  473. data/vendor/ggml/src/ggml-et/et-kernels/src/quants.h +72 -0
  474. data/vendor/ggml/src/ggml-et/et-kernels/src/repeat_f32.c +118 -0
  475. data/vendor/ggml/src/ggml-et/et-kernels/src/rms_norm_f32.c +270 -0
  476. data/vendor/ggml/src/ggml-et/et-kernels/src/rms_norm_mul_f32.c +290 -0
  477. data/vendor/ggml/src/ggml-et/et-kernels/src/rope_f32.c +656 -0
  478. data/vendor/ggml/src/ggml-et/et-kernels/src/rwkv_wkv6_f32.c +184 -0
  479. data/vendor/ggml/src/ggml-et/et-kernels/src/rwkv_wkv7_f32.c +272 -0
  480. data/vendor/ggml/src/ggml-et/et-kernels/src/scale_f32.c +94 -0
  481. data/vendor/ggml/src/ggml-et/et-kernels/src/set_f32.c +101 -0
  482. data/vendor/ggml/src/ggml-et/et-kernels/src/set_rows_f32.c +394 -0
  483. data/vendor/ggml/src/ggml-et/et-kernels/src/softmax_f32.c +698 -0
  484. data/vendor/ggml/src/ggml-et/et-kernels/src/solve_tri_f32.c +109 -0
  485. data/vendor/ggml/src/ggml-et/et-kernels/src/sqr_f32.c +88 -0
  486. data/vendor/ggml/src/ggml-et/et-kernels/src/ssm_conv_f32.c +129 -0
  487. data/vendor/ggml/src/ggml-et/et-kernels/src/ssm_scan_f32.c +282 -0
  488. data/vendor/ggml/src/ggml-et/et-kernels/src/sum_rows_f32.c +103 -0
  489. data/vendor/ggml/src/ggml-et/et-kernels/src/tensor.h +897 -0
  490. data/vendor/ggml/src/ggml-et/et-kernels/src/tri_f32.c +244 -0
  491. data/vendor/ggml/src/ggml-et/et-kernels/src/uberkernel.c +497 -0
  492. data/vendor/ggml/src/ggml-et/et-kernels/src/unary_f32.c +705 -0
  493. data/vendor/ggml/src/ggml-et/ggml-et-common.h +86 -0
  494. data/vendor/ggml/src/ggml-et/ggml-et-cpu-compare.cpp +502 -0
  495. data/vendor/ggml/src/ggml-et/ggml-et-cpu-compare.h +54 -0
  496. data/vendor/ggml/src/ggml-et/ggml-et-kernels.cpp +508 -0
  497. data/vendor/ggml/src/ggml-et/ggml-et-kernels.h +48 -0
  498. data/vendor/ggml/src/ggml-et/ggml-et-memops.cpp +36 -0
  499. data/vendor/ggml/src/ggml-et/ggml-et-memops.h +18 -0
  500. data/vendor/ggml/src/ggml-et/ggml-et-ops.cpp +2584 -0
  501. data/vendor/ggml/src/ggml-et/ggml-et-ops.h +393 -0
  502. data/vendor/ggml/src/ggml-et/ggml-et-uberkernel-common.h +17 -0
  503. data/vendor/ggml/src/ggml-et/ggml-et.cpp +1880 -0
  504. data/vendor/ggml/src/ggml-feats.h +166 -0
  505. data/vendor/ggml/src/ggml-hexagon/CMakeLists.txt +113 -0
  506. data/vendor/ggml/src/ggml-hexagon/ggml-hexagon.cpp +6523 -0
  507. data/vendor/ggml/src/ggml-hexagon/htp/CMakeLists.txt +58 -0
  508. data/vendor/ggml/src/ggml-hexagon/htp/act-ops.c +585 -0
  509. data/vendor/ggml/src/ggml-hexagon/htp/allreduce-ops.c +398 -0
  510. data/vendor/ggml/src/ggml-hexagon/htp/allreduce-ops.h +40 -0
  511. data/vendor/ggml/src/ggml-hexagon/htp/argsort-ops.c +514 -0
  512. data/vendor/ggml/src/ggml-hexagon/htp/binary-ops.c +873 -0
  513. data/vendor/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +155 -0
  514. data/vendor/ggml/src/ggml-hexagon/htp/concat-ops.c +277 -0
  515. data/vendor/ggml/src/ggml-hexagon/htp/cpy-ops.c +354 -0
  516. data/vendor/ggml/src/ggml-hexagon/htp/cumsum-ops.c +265 -0
  517. data/vendor/ggml/src/ggml-hexagon/htp/diag-ops.c +216 -0
  518. data/vendor/ggml/src/ggml-hexagon/htp/dma-queue.c +104 -0
  519. data/vendor/ggml/src/ggml-hexagon/htp/dma-queue.h +408 -0
  520. data/vendor/ggml/src/ggml-hexagon/htp/fill-ops.c +123 -0
  521. data/vendor/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +2479 -0
  522. data/vendor/ggml/src/ggml-hexagon/htp/flash-attn-ops.h +344 -0
  523. data/vendor/ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c +1157 -0
  524. data/vendor/ggml/src/ggml-hexagon/htp/get-rows-ops.c +260 -0
  525. data/vendor/ggml/src/ggml-hexagon/htp/get-rows-ops.h +77 -0
  526. data/vendor/ggml/src/ggml-hexagon/htp/hex-bitmap.h +24 -0
  527. data/vendor/ggml/src/ggml-hexagon/htp/hex-common.h +80 -0
  528. data/vendor/ggml/src/ggml-hexagon/htp/hex-dma.h +2 -0
  529. data/vendor/ggml/src/ggml-hexagon/htp/hex-dump.h +86 -0
  530. data/vendor/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
  531. data/vendor/ggml/src/ggml-hexagon/htp/hex-profile.h +64 -0
  532. data/vendor/ggml/src/ggml-hexagon/htp/hex-utils.h +65 -0
  533. data/vendor/ggml/src/ggml-hexagon/htp/hmx-fa-kernels.h +555 -0
  534. data/vendor/ggml/src/ggml-hexagon/htp/hmx-mm-kernels-tiled.h +1378 -0
  535. data/vendor/ggml/src/ggml-hexagon/htp/hmx-queue.c +170 -0
  536. data/vendor/ggml/src/ggml-hexagon/htp/hmx-queue.h +160 -0
  537. data/vendor/ggml/src/ggml-hexagon/htp/hmx-utils.h +222 -0
  538. data/vendor/ggml/src/ggml-hexagon/htp/htp-ctx.h +146 -0
  539. data/vendor/ggml/src/ggml-hexagon/htp/htp-ops.h +242 -0
  540. data/vendor/ggml/src/ggml-hexagon/htp/htp-tensor.c +300 -0
  541. data/vendor/ggml/src/ggml-hexagon/htp/htp-tensor.h +29 -0
  542. data/vendor/ggml/src/ggml-hexagon/htp/htp-vtcm.h +19 -0
  543. data/vendor/ggml/src/ggml-hexagon/htp/htp_iface.idl +23 -0
  544. data/vendor/ggml/src/ggml-hexagon/htp/hvx-arith.h +636 -0
  545. data/vendor/ggml/src/ggml-hexagon/htp/hvx-base.h +292 -0
  546. data/vendor/ggml/src/ggml-hexagon/htp/hvx-copy.h +262 -0
  547. data/vendor/ggml/src/ggml-hexagon/htp/hvx-div.h +291 -0
  548. data/vendor/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
  549. data/vendor/ggml/src/ggml-hexagon/htp/hvx-exp.h +255 -0
  550. data/vendor/ggml/src/ggml-hexagon/htp/hvx-fa-kernels.h +303 -0
  551. data/vendor/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -0
  552. data/vendor/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
  553. data/vendor/ggml/src/ggml-hexagon/htp/hvx-inverse.h +210 -0
  554. data/vendor/ggml/src/ggml-hexagon/htp/hvx-log.h +118 -0
  555. data/vendor/ggml/src/ggml-hexagon/htp/hvx-mm-kernels-flat.h +1511 -0
  556. data/vendor/ggml/src/ggml-hexagon/htp/hvx-mm-kernels-tiled.h +1200 -0
  557. data/vendor/ggml/src/ggml-hexagon/htp/hvx-norm.h +454 -0
  558. data/vendor/ggml/src/ggml-hexagon/htp/hvx-pow.h +42 -0
  559. data/vendor/ggml/src/ggml-hexagon/htp/hvx-quant.h +165 -0
  560. data/vendor/ggml/src/ggml-hexagon/htp/hvx-reduce.h +336 -0
  561. data/vendor/ggml/src/ggml-hexagon/htp/hvx-repl.h +74 -0
  562. data/vendor/ggml/src/ggml-hexagon/htp/hvx-scale.h +199 -0
  563. data/vendor/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +181 -0
  564. data/vendor/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
  565. data/vendor/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +189 -0
  566. data/vendor/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
  567. data/vendor/ggml/src/ggml-hexagon/htp/hvx-utils.h +24 -0
  568. data/vendor/ggml/src/ggml-hexagon/htp/im2col-ops.c +306 -0
  569. data/vendor/ggml/src/ggml-hexagon/htp/main.c +1245 -0
  570. data/vendor/ggml/src/ggml-hexagon/htp/matmul-ops.c +4078 -0
  571. data/vendor/ggml/src/ggml-hexagon/htp/matmul-ops.h +778 -0
  572. data/vendor/ggml/src/ggml-hexagon/htp/pad-ops.c +547 -0
  573. data/vendor/ggml/src/ggml-hexagon/htp/repeat-ops.c +148 -0
  574. data/vendor/ggml/src/ggml-hexagon/htp/rope-ops.c +733 -0
  575. data/vendor/ggml/src/ggml-hexagon/htp/set-rows-ops.c +230 -0
  576. data/vendor/ggml/src/ggml-hexagon/htp/set-rows-ops.h +74 -0
  577. data/vendor/ggml/src/ggml-hexagon/htp/softmax-ops.c +407 -0
  578. data/vendor/ggml/src/ggml-hexagon/htp/solve-tri-ops.c +267 -0
  579. data/vendor/ggml/src/ggml-hexagon/htp/ssm-conv.c +433 -0
  580. data/vendor/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +128 -0
  581. data/vendor/ggml/src/ggml-hexagon/htp/unary-ops.c +1264 -0
  582. data/vendor/ggml/src/ggml-hexagon/htp/unary-ops.h +176 -0
  583. data/vendor/ggml/src/ggml-hexagon/htp/work-queue.c +244 -0
  584. data/vendor/ggml/src/ggml-hexagon/htp/work-queue.h +38 -0
  585. data/vendor/ggml/src/ggml-hexagon/htp-drv.cpp +425 -0
  586. data/vendor/ggml/src/ggml-hexagon/htp-drv.h +123 -0
  587. data/vendor/ggml/src/ggml-hexagon/htp-opnode.h +372 -0
  588. data/vendor/ggml/src/ggml-hexagon/libdl.h +79 -0
  589. data/vendor/ggml/src/ggml-hexagon/libggml-htp.inf +36 -0
  590. data/vendor/ggml/src/ggml-hip/CMakeLists.txt +153 -0
  591. data/vendor/ggml/src/ggml-impl.h +783 -0
  592. data/vendor/ggml/src/ggml-metal/CMakeLists.txt +240 -0
  593. data/vendor/ggml/src/ggml-metal/ggml-metal-common.cpp +474 -0
  594. data/vendor/ggml/src/ggml-metal/ggml-metal-common.h +56 -0
  595. data/vendor/ggml/src/ggml-metal/ggml-metal-context.h +41 -0
  596. data/vendor/ggml/src/ggml-metal/ggml-metal-context.m +759 -0
  597. data/vendor/ggml/src/ggml-metal/ggml-metal-device.cpp +2330 -0
  598. data/vendor/ggml/src/ggml-metal/ggml-metal-device.h +355 -0
  599. data/vendor/ggml/src/ggml-metal/ggml-metal-device.m +2473 -0
  600. data/vendor/ggml/src/ggml-metal/ggml-metal-impl.h +1345 -0
  601. data/vendor/ggml/src/ggml-metal/ggml-metal-ops.cpp +5610 -0
  602. data/vendor/ggml/src/ggml-metal/ggml-metal-ops.h +106 -0
  603. data/vendor/ggml/src/ggml-metal/ggml-metal-tuning.cpp +3732 -0
  604. data/vendor/ggml/src/ggml-metal/ggml-metal-tuning.h +77 -0
  605. data/vendor/ggml/src/ggml-metal/ggml-metal.cpp +1006 -0
  606. data/vendor/ggml/src/ggml-metal/kernels/argsort.metal +337 -0
  607. data/vendor/ggml/src/ggml-metal/kernels/binbcast.metal +228 -0
  608. data/vendor/ggml/src/ggml-metal/kernels/common.h +126 -0
  609. data/vendor/ggml/src/ggml-metal/kernels/conv.metal +724 -0
  610. data/vendor/ggml/src/ggml-metal/kernels/dequantize.h +735 -0
  611. data/vendor/ggml/src/ggml-metal/kernels/fa.metal +2445 -0
  612. data/vendor/ggml/src/ggml-metal/kernels/gated_delta_net.metal +250 -0
  613. data/vendor/ggml/src/ggml-metal/kernels/misc.metal +595 -0
  614. data/vendor/ggml/src/ggml-metal/kernels/mul_mm.metal +853 -0
  615. data/vendor/ggml/src/ggml-metal/kernels/mul_mv.metal +3225 -0
  616. data/vendor/ggml/src/ggml-metal/kernels/norm.metal +308 -0
  617. data/vendor/ggml/src/ggml-metal/kernels/pool.metal +148 -0
  618. data/vendor/ggml/src/ggml-metal/kernels/quantize.h +262 -0
  619. data/vendor/ggml/src/ggml-metal/kernels/quantize.metal +480 -0
  620. data/vendor/ggml/src/ggml-metal/kernels/reduce.metal +228 -0
  621. data/vendor/ggml/src/ggml-metal/kernels/rope.metal +333 -0
  622. data/vendor/ggml/src/ggml-metal/kernels/softmax.metal +223 -0
  623. data/vendor/ggml/src/ggml-metal/kernels/solve_tri.metal +75 -0
  624. data/vendor/ggml/src/ggml-metal/kernels/ssm.metal +473 -0
  625. data/vendor/ggml/src/ggml-metal/kernels/tri.metal +69 -0
  626. data/vendor/ggml/src/ggml-metal/kernels/unary.metal +400 -0
  627. data/vendor/ggml/src/ggml-metal/kernels/upscale.metal +179 -0
  628. data/vendor/ggml/src/ggml-metal/kernels/wkv.metal +179 -0
  629. data/vendor/ggml/src/ggml-musa/CMakeLists.txt +123 -0
  630. data/vendor/ggml/src/ggml-musa/mudnn.cu +112 -0
  631. data/vendor/ggml/src/ggml-musa/mudnn.cuh +12 -0
  632. data/vendor/ggml/src/ggml-opencl/CMakeLists.txt +246 -0
  633. data/vendor/ggml/src/ggml-opencl/cl-program-cache.cpp +453 -0
  634. data/vendor/ggml/src/ggml-opencl/cl-program-cache.h +75 -0
  635. data/vendor/ggml/src/ggml-opencl/fa_tune.h +92 -0
  636. data/vendor/ggml/src/ggml-opencl/ggml-opencl.cpp +26982 -0
  637. data/vendor/ggml/src/ggml-opencl/kernels/abs.cl +113 -0
  638. data/vendor/ggml/src/ggml-opencl/kernels/add.cl +190 -0
  639. data/vendor/ggml/src/ggml-opencl/kernels/add_id.cl +42 -0
  640. data/vendor/ggml/src/ggml-opencl/kernels/argsort.cl +86 -0
  641. data/vendor/ggml/src/ggml-opencl/kernels/clamp.cl +20 -0
  642. data/vendor/ggml/src/ggml-opencl/kernels/concat.cl +118 -0
  643. data/vendor/ggml/src/ggml-opencl/kernels/conv2d.cl +185 -0
  644. data/vendor/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl +176 -0
  645. data/vendor/ggml/src/ggml-opencl/kernels/cpy.cl +288 -0
  646. data/vendor/ggml/src/ggml-opencl/kernels/cumsum.cl +139 -0
  647. data/vendor/ggml/src/ggml-opencl/kernels/cvt.cl +2663 -0
  648. data/vendor/ggml/src/ggml-opencl/kernels/diag.cl +27 -0
  649. data/vendor/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl +58 -0
  650. data/vendor/ggml/src/ggml-opencl/kernels/div.cl +138 -0
  651. data/vendor/ggml/src/ggml-opencl/kernels/embed_kernel.py +26 -0
  652. data/vendor/ggml/src/ggml-opencl/kernels/exp.cl +125 -0
  653. data/vendor/ggml/src/ggml-opencl/kernels/expm1.cl +113 -0
  654. data/vendor/ggml/src/ggml-opencl/kernels/fill.cl +17 -0
  655. data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl +421 -0
  656. data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +422 -0
  657. data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl +2656 -0
  658. data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32_q4_0.cl +2011 -0
  659. data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_f32_q8_0.cl +1840 -0
  660. data/vendor/ggml/src/ggml-opencl/kernels/flash_attn_pre_f16.cl +156 -0
  661. data/vendor/ggml/src/ggml-opencl/kernels/gated_delta_net.cl +249 -0
  662. data/vendor/ggml/src/ggml-opencl/kernels/gelu.cl +89 -0
  663. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32.cl +162 -0
  664. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl +376 -0
  665. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_q8_1_dp4a.cl +190 -0
  666. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl +324 -0
  667. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_q8_1_dp4a.cl +169 -0
  668. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl +326 -0
  669. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl +348 -0
  670. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_q8_1_dp4a.cl +209 -0
  671. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl +328 -0
  672. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl +330 -0
  673. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl +356 -0
  674. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl +335 -0
  675. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_q8_1_dp4a.cl +200 -0
  676. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q8_0_f32_ns.cl +221 -0
  677. data/vendor/ggml/src/ggml-opencl/kernels/gemm_moe_q8_1_dp4a.cl +226 -0
  678. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_f32.cl +150 -0
  679. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_q8_1_dp4a.cl +143 -0
  680. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q1_0_f32.cl +94 -0
  681. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_0_f32.cl +139 -0
  682. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_0_q8_1_dp4a.cl +127 -0
  683. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_1_f32.cl +132 -0
  684. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl +490 -0
  685. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_q8_1_dp4a.cl +281 -0
  686. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl +131 -0
  687. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_q8_1_dp4a.cl +235 -0
  688. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl +134 -0
  689. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl +176 -0
  690. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_q8_1_dp4a.cl +164 -0
  691. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl +245 -0
  692. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32_tiled.cl +136 -0
  693. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_q8_1_dp4a.cl +144 -0
  694. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q8_0_f32.cl +129 -0
  695. data/vendor/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q8_0_q8_1_dp4a.cl +212 -0
  696. data/vendor/ggml/src/ggml-opencl/kernels/gemm_xmem_f16_f32_os8.cl +233 -0
  697. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl +156 -0
  698. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl +257 -0
  699. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl +120 -0
  700. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl +123 -0
  701. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl +266 -0
  702. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl +123 -0
  703. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl +125 -0
  704. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl +160 -0
  705. data/vendor/ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl +141 -0
  706. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +307 -0
  707. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q1_0_f32.cl +125 -0
  708. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32.cl +383 -0
  709. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_0_f32_spec.cl +272 -0
  710. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl +384 -0
  711. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +823 -0
  712. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32_o4.cl +349 -0
  713. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32_tiled.cl +118 -0
  714. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +296 -0
  715. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +299 -0
  716. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +453 -0
  717. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +409 -0
  718. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32_o4.cl +372 -0
  719. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32_tiled.cl +196 -0
  720. data/vendor/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl +280 -0
  721. data/vendor/ggml/src/ggml-opencl/kernels/get_rows.cl +193 -0
  722. data/vendor/ggml/src/ggml-opencl/kernels/glu.cl +443 -0
  723. data/vendor/ggml/src/ggml-opencl/kernels/group_norm.cl +121 -0
  724. data/vendor/ggml/src/ggml-opencl/kernels/im2col_f16.cl +57 -0
  725. data/vendor/ggml/src/ggml-opencl/kernels/im2col_f32.cl +57 -0
  726. data/vendor/ggml/src/ggml-opencl/kernels/l2_norm.cl +71 -0
  727. data/vendor/ggml/src/ggml-opencl/kernels/mean.cl +140 -0
  728. data/vendor/ggml/src/ggml-opencl/kernels/moe_add_id_glu.cl +76 -0
  729. data/vendor/ggml/src/ggml-opencl/kernels/moe_combine.cl +79 -0
  730. data/vendor/ggml/src/ggml-opencl/kernels/moe_reorder_b.cl +30 -0
  731. data/vendor/ggml/src/ggml-opencl/kernels/moe_reorder_quant_a_q8_1.cl +64 -0
  732. data/vendor/ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl +155 -0
  733. data/vendor/ggml/src/ggml-opencl/kernels/mul.cl +152 -0
  734. data/vendor/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl +130 -0
  735. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_kq_kqv.cl +273 -0
  736. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl +146 -0
  737. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl +196 -0
  738. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_iq4_nl_f32_l4_lm.cl +171 -0
  739. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q1_0_f32_l4_lm.cl +156 -0
  740. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
  741. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
  742. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl +189 -0
  743. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q5_0_f32_l4_lm.cl +173 -0
  744. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q5_1_f32_l4_lm.cl +175 -0
  745. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q5_k_f32_l4_lm.cl +202 -0
  746. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
  747. data/vendor/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl +154 -0
  748. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl +125 -0
  749. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl +125 -0
  750. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl +101 -0
  751. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl +1237 -0
  752. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_mrow.cl +306 -0
  753. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl +118 -0
  754. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl +189 -0
  755. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl +176 -0
  756. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl +283 -0
  757. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl +140 -0
  758. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl +222 -0
  759. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32.cl +164 -0
  760. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32_flat.cl +202 -0
  761. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl +144 -0
  762. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl +167 -0
  763. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q1_0_f32.cl +141 -0
  764. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q1_0_f32_flat.cl +190 -0
  765. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl +192 -0
  766. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl +307 -0
  767. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl +265 -0
  768. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl +272 -0
  769. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl +254 -0
  770. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
  771. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
  772. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +182 -0
  773. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl +205 -0
  774. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32.cl +241 -0
  775. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32_flat.cl +243 -0
  776. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32.cl +243 -0
  777. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32_flat.cl +247 -0
  778. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32.cl +187 -0
  779. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32_flat.cl +244 -0
  780. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl +194 -0
  781. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +272 -0
  782. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl +125 -0
  783. data/vendor/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl +202 -0
  784. data/vendor/ggml/src/ggml-opencl/kernels/neg.cl +125 -0
  785. data/vendor/ggml/src/ggml-opencl/kernels/norm.cl +164 -0
  786. data/vendor/ggml/src/ggml-opencl/kernels/pad.cl +39 -0
  787. data/vendor/ggml/src/ggml-opencl/kernels/quant_a_q8_1.cl +42 -0
  788. data/vendor/ggml/src/ggml-opencl/kernels/relu.cl +16 -0
  789. data/vendor/ggml/src/ggml-opencl/kernels/repeat.cl +38 -0
  790. data/vendor/ggml/src/ggml-opencl/kernels/rms_norm.cl +369 -0
  791. data/vendor/ggml/src/ggml-opencl/kernels/rope.cl +759 -0
  792. data/vendor/ggml/src/ggml-opencl/kernels/scale.cl +27 -0
  793. data/vendor/ggml/src/ggml-opencl/kernels/set_rows.cl +708 -0
  794. data/vendor/ggml/src/ggml-opencl/kernels/sigmoid.cl +29 -0
  795. data/vendor/ggml/src/ggml-opencl/kernels/silu.cl +30 -0
  796. data/vendor/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +108 -0
  797. data/vendor/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +108 -0
  798. data/vendor/ggml/src/ggml-opencl/kernels/softmax_f16.cl +107 -0
  799. data/vendor/ggml/src/ggml-opencl/kernels/softmax_f32.cl +107 -0
  800. data/vendor/ggml/src/ggml-opencl/kernels/softplus.cl +116 -0
  801. data/vendor/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
  802. data/vendor/ggml/src/ggml-opencl/kernels/sqr.cl +53 -0
  803. data/vendor/ggml/src/ggml-opencl/kernels/sqrt.cl +53 -0
  804. data/vendor/ggml/src/ggml-opencl/kernels/ssm_conv.cl +77 -0
  805. data/vendor/ggml/src/ggml-opencl/kernels/ssm_scan.cl +216 -0
  806. data/vendor/ggml/src/ggml-opencl/kernels/sub.cl +138 -0
  807. data/vendor/ggml/src/ggml-opencl/kernels/sum_rows.cl +140 -0
  808. data/vendor/ggml/src/ggml-opencl/kernels/tanh.cl +109 -0
  809. data/vendor/ggml/src/ggml-opencl/kernels/transpose.cl +143 -0
  810. data/vendor/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
  811. data/vendor/ggml/src/ggml-opencl/kernels/tsembd.cl +48 -0
  812. data/vendor/ggml/src/ggml-opencl/kernels/upscale.cl +120 -0
  813. data/vendor/ggml/src/ggml-opencl/libdl.h +79 -0
  814. data/vendor/ggml/src/ggml-openvino/CMakeLists.txt +22 -0
  815. data/vendor/ggml/src/ggml-openvino/ggml-decoder.cpp +2057 -0
  816. data/vendor/ggml/src/ggml-openvino/ggml-decoder.h +417 -0
  817. data/vendor/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +493 -0
  818. data/vendor/ggml/src/ggml-openvino/ggml-openvino-extra.h +216 -0
  819. data/vendor/ggml/src/ggml-openvino/ggml-openvino.cpp +1566 -0
  820. data/vendor/ggml/src/ggml-openvino/ggml-quants.cpp +1273 -0
  821. data/vendor/ggml/src/ggml-openvino/ggml-quants.h +186 -0
  822. data/vendor/ggml/src/ggml-openvino/model-cache.cpp +272 -0
  823. data/vendor/ggml/src/ggml-openvino/model-cache.h +56 -0
  824. data/vendor/ggml/src/ggml-openvino/openvino/decoder.h +133 -0
  825. data/vendor/ggml/src/ggml-openvino/openvino/frontend.cpp +27 -0
  826. data/vendor/ggml/src/ggml-openvino/openvino/frontend.h +23 -0
  827. data/vendor/ggml/src/ggml-openvino/openvino/input_model.cpp +17 -0
  828. data/vendor/ggml/src/ggml-openvino/openvino/input_model.h +29 -0
  829. data/vendor/ggml/src/ggml-openvino/openvino/node_context.h +171 -0
  830. data/vendor/ggml/src/ggml-openvino/openvino/op/add.cpp +45 -0
  831. data/vendor/ggml/src/ggml-openvino/openvino/op/add_id.cpp +76 -0
  832. data/vendor/ggml/src/ggml-openvino/openvino/op/argsort.cpp +47 -0
  833. data/vendor/ggml/src/ggml-openvino/openvino/op/clamp.cpp +33 -0
  834. data/vendor/ggml/src/ggml-openvino/openvino/op/concat.cpp +48 -0
  835. data/vendor/ggml/src/ggml-openvino/openvino/op/cont.cpp +40 -0
  836. data/vendor/ggml/src/ggml-openvino/openvino/op/cpy.cpp +295 -0
  837. data/vendor/ggml/src/ggml-openvino/openvino/op/cumsum.cpp +29 -0
  838. data/vendor/ggml/src/ggml-openvino/openvino/op/diag.cpp +58 -0
  839. data/vendor/ggml/src/ggml-openvino/openvino/op/div.cpp +146 -0
  840. data/vendor/ggml/src/ggml-openvino/openvino/op/fill.cpp +34 -0
  841. data/vendor/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +241 -0
  842. data/vendor/ggml/src/ggml-openvino/openvino/op/gated_delta_net.cpp +327 -0
  843. data/vendor/ggml/src/ggml-openvino/openvino/op/gated_delta_net.hpp +65 -0
  844. data/vendor/ggml/src/ggml-openvino/openvino/op/gather_matmul.hpp +43 -0
  845. data/vendor/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +142 -0
  846. data/vendor/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +75 -0
  847. data/vendor/ggml/src/ggml-openvino/openvino/op/glu_geglu_quick.cpp +64 -0
  848. data/vendor/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +110 -0
  849. data/vendor/ggml/src/ggml-openvino/openvino/op/im2col.cpp +120 -0
  850. data/vendor/ggml/src/ggml-openvino/openvino/op/l2_norm.cpp +61 -0
  851. data/vendor/ggml/src/ggml-openvino/openvino/op/mul_mat_id.cpp +283 -0
  852. data/vendor/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +100 -0
  853. data/vendor/ggml/src/ggml-openvino/openvino/op/norm.cpp +58 -0
  854. data/vendor/ggml/src/ggml-openvino/openvino/op/pad.cpp +95 -0
  855. data/vendor/ggml/src/ggml-openvino/openvino/op/permute.cpp +147 -0
  856. data/vendor/ggml/src/ggml-openvino/openvino/op/pool_2d.cpp +53 -0
  857. data/vendor/ggml/src/ggml-openvino/openvino/op/repeat.cpp +48 -0
  858. data/vendor/ggml/src/ggml-openvino/openvino/op/reshape.cpp +113 -0
  859. data/vendor/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +81 -0
  860. data/vendor/ggml/src/ggml-openvino/openvino/op/roll.cpp +36 -0
  861. data/vendor/ggml/src/ggml-openvino/openvino/op/rope.cpp +303 -0
  862. data/vendor/ggml/src/ggml-openvino/openvino/op/scale.cpp +86 -0
  863. data/vendor/ggml/src/ggml-openvino/openvino/op/set.cpp +76 -0
  864. data/vendor/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +95 -0
  865. data/vendor/ggml/src/ggml-openvino/openvino/op/softmax.cpp +166 -0
  866. data/vendor/ggml/src/ggml-openvino/openvino/op/solve_tri.cpp +108 -0
  867. data/vendor/ggml/src/ggml-openvino/openvino/op/sqr.cpp +35 -0
  868. data/vendor/ggml/src/ggml-openvino/openvino/op/ssm_conv.cpp +61 -0
  869. data/vendor/ggml/src/ggml-openvino/openvino/op/sum_rows.cpp +27 -0
  870. data/vendor/ggml/src/ggml-openvino/openvino/op/transpose.cpp +54 -0
  871. data/vendor/ggml/src/ggml-openvino/openvino/op/tri.cpp +82 -0
  872. data/vendor/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +27 -0
  873. data/vendor/ggml/src/ggml-openvino/openvino/op/unary_softplus.cpp +38 -0
  874. data/vendor/ggml/src/ggml-openvino/openvino/op/view.cpp +245 -0
  875. data/vendor/ggml/src/ggml-openvino/openvino/op_table.cpp +88 -0
  876. data/vendor/ggml/src/ggml-openvino/openvino/op_table.h +67 -0
  877. data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.cpp +212 -0
  878. data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_conv.h +17 -0
  879. data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.cpp +60 -0
  880. data/vendor/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.h +17 -0
  881. data/vendor/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +29 -0
  882. data/vendor/ggml/src/ggml-openvino/openvino/pass/mark_dequantization_subgraph.h +44 -0
  883. data/vendor/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp +58 -0
  884. data/vendor/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.h +17 -0
  885. data/vendor/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp +41 -0
  886. data/vendor/ggml/src/ggml-openvino/openvino/translate_session.cpp +432 -0
  887. data/vendor/ggml/src/ggml-openvino/openvino/translate_session.h +29 -0
  888. data/vendor/ggml/src/ggml-openvino/openvino/utils.cpp +886 -0
  889. data/vendor/ggml/src/ggml-openvino/openvino/utils.h +88 -0
  890. data/vendor/ggml/src/ggml-openvino/utils.cpp +1496 -0
  891. data/vendor/ggml/src/ggml-openvino/utils.h +195 -0
  892. data/vendor/ggml/src/ggml-opt.cpp +1094 -0
  893. data/vendor/ggml/src/ggml-quants.c +5667 -0
  894. data/vendor/ggml/src/ggml-quants.h +115 -0
  895. data/vendor/ggml/src/ggml-rpc/CMakeLists.txt +49 -0
  896. data/vendor/ggml/src/ggml-rpc/ggml-rpc.cpp +2351 -0
  897. data/vendor/ggml/src/ggml-rpc/transport-apple.cpp +481 -0
  898. data/vendor/ggml/src/ggml-rpc/transport-apple.h +27 -0
  899. data/vendor/ggml/src/ggml-rpc/transport.cpp +739 -0
  900. data/vendor/ggml/src/ggml-rpc/transport.h +38 -0
  901. data/vendor/ggml/src/ggml-sycl/CMakeLists.txt +218 -0
  902. data/vendor/ggml/src/ggml-sycl/add-id.cpp +81 -0
  903. data/vendor/ggml/src/ggml-sycl/add-id.hpp +8 -0
  904. data/vendor/ggml/src/ggml-sycl/backend.hpp +52 -0
  905. data/vendor/ggml/src/ggml-sycl/base.hpp +36 -0
  906. data/vendor/ggml/src/ggml-sycl/binbcast.cpp +650 -0
  907. data/vendor/ggml/src/ggml-sycl/binbcast.hpp +69 -0
  908. data/vendor/ggml/src/ggml-sycl/col2im-1d.cpp +102 -0
  909. data/vendor/ggml/src/ggml-sycl/col2im-1d.hpp +8 -0
  910. data/vendor/ggml/src/ggml-sycl/common.cpp +153 -0
  911. data/vendor/ggml/src/ggml-sycl/common.hpp +1029 -0
  912. data/vendor/ggml/src/ggml-sycl/concat.cpp +509 -0
  913. data/vendor/ggml/src/ggml-sycl/concat.hpp +20 -0
  914. data/vendor/ggml/src/ggml-sycl/conv.cpp +101 -0
  915. data/vendor/ggml/src/ggml-sycl/conv.hpp +20 -0
  916. data/vendor/ggml/src/ggml-sycl/conv2d-dw.cpp +170 -0
  917. data/vendor/ggml/src/ggml-sycl/conv2d-dw.hpp +10 -0
  918. data/vendor/ggml/src/ggml-sycl/conv2d-transpose.cpp +125 -0
  919. data/vendor/ggml/src/ggml-sycl/conv2d-transpose.hpp +10 -0
  920. data/vendor/ggml/src/ggml-sycl/conv2d.cpp +150 -0
  921. data/vendor/ggml/src/ggml-sycl/conv2d.hpp +10 -0
  922. data/vendor/ggml/src/ggml-sycl/conv3d.cpp +224 -0
  923. data/vendor/ggml/src/ggml-sycl/conv3d.hpp +8 -0
  924. data/vendor/ggml/src/ggml-sycl/convert.cpp +877 -0
  925. data/vendor/ggml/src/ggml-sycl/convert.hpp +64 -0
  926. data/vendor/ggml/src/ggml-sycl/count-equal.cpp +79 -0
  927. data/vendor/ggml/src/ggml-sycl/count-equal.hpp +9 -0
  928. data/vendor/ggml/src/ggml-sycl/cpy.cpp +1433 -0
  929. data/vendor/ggml/src/ggml-sycl/cpy.hpp +537 -0
  930. data/vendor/ggml/src/ggml-sycl/cross_entropy_loss.cpp +255 -0
  931. data/vendor/ggml/src/ggml-sycl/cross_entropy_loss.hpp +7 -0
  932. data/vendor/ggml/src/ggml-sycl/cumsum.cpp +148 -0
  933. data/vendor/ggml/src/ggml-sycl/cumsum.hpp +5 -0
  934. data/vendor/ggml/src/ggml-sycl/dequantize.hpp +1681 -0
  935. data/vendor/ggml/src/ggml-sycl/diag.cpp +67 -0
  936. data/vendor/ggml/src/ggml-sycl/diag.hpp +5 -0
  937. data/vendor/ggml/src/ggml-sycl/dmmv.cpp +2227 -0
  938. data/vendor/ggml/src/ggml-sycl/dmmv.hpp +27 -0
  939. data/vendor/ggml/src/ggml-sycl/dpct/helper.hpp +3782 -0
  940. data/vendor/ggml/src/ggml-sycl/dsv4-hc.cpp +280 -0
  941. data/vendor/ggml/src/ggml-sycl/dsv4-hc.hpp +10 -0
  942. data/vendor/ggml/src/ggml-sycl/element_wise.cpp +1432 -0
  943. data/vendor/ggml/src/ggml-sycl/element_wise.hpp +135 -0
  944. data/vendor/ggml/src/ggml-sycl/esimd.hpp +589 -0
  945. data/vendor/ggml/src/ggml-sycl/fattn-buffers.cpp +56 -0
  946. data/vendor/ggml/src/ggml-sycl/fattn-buffers.hpp +63 -0
  947. data/vendor/ggml/src/ggml-sycl/fattn-common.hpp +1185 -0
  948. data/vendor/ggml/src/ggml-sycl/fattn-mkl.cpp +689 -0
  949. data/vendor/ggml/src/ggml-sycl/fattn-onednn.cpp +450 -0
  950. data/vendor/ggml/src/ggml-sycl/fattn-onednn.hpp +18 -0
  951. data/vendor/ggml/src/ggml-sycl/fattn-tile.cpp +59 -0
  952. data/vendor/ggml/src/ggml-sycl/fattn-tile.hpp +1246 -0
  953. data/vendor/ggml/src/ggml-sycl/fattn-vec.hpp +684 -0
  954. data/vendor/ggml/src/ggml-sycl/fattn.cpp +452 -0
  955. data/vendor/ggml/src/ggml-sycl/fattn.hpp +42 -0
  956. data/vendor/ggml/src/ggml-sycl/fill.cpp +55 -0
  957. data/vendor/ggml/src/ggml-sycl/fill.hpp +5 -0
  958. data/vendor/ggml/src/ggml-sycl/fusion.cpp +212 -0
  959. data/vendor/ggml/src/ggml-sycl/fusion.hpp +17 -0
  960. data/vendor/ggml/src/ggml-sycl/fwht.cpp +119 -0
  961. data/vendor/ggml/src/ggml-sycl/fwht.hpp +12 -0
  962. data/vendor/ggml/src/ggml-sycl/gated_delta_net.cpp +366 -0
  963. data/vendor/ggml/src/ggml-sycl/gated_delta_net.hpp +19 -0
  964. data/vendor/ggml/src/ggml-sycl/gemm.hpp +93 -0
  965. data/vendor/ggml/src/ggml-sycl/getrows.cpp +368 -0
  966. data/vendor/ggml/src/ggml-sycl/getrows.hpp +20 -0
  967. data/vendor/ggml/src/ggml-sycl/ggml-sycl.cpp +7028 -0
  968. data/vendor/ggml/src/ggml-sycl/gla.cpp +106 -0
  969. data/vendor/ggml/src/ggml-sycl/gla.hpp +8 -0
  970. data/vendor/ggml/src/ggml-sycl/im2col.cpp +400 -0
  971. data/vendor/ggml/src/ggml-sycl/im2col.hpp +23 -0
  972. data/vendor/ggml/src/ggml-sycl/lightning-indexer.cpp +197 -0
  973. data/vendor/ggml/src/ggml-sycl/lightning-indexer.hpp +8 -0
  974. data/vendor/ggml/src/ggml-sycl/mem.cpp +162 -0
  975. data/vendor/ggml/src/ggml-sycl/mem.hpp +16 -0
  976. data/vendor/ggml/src/ggml-sycl/mmq.cpp +3030 -0
  977. data/vendor/ggml/src/ggml-sycl/mmq.hpp +33 -0
  978. data/vendor/ggml/src/ggml-sycl/mmvq.cpp +3033 -0
  979. data/vendor/ggml/src/ggml-sycl/mmvq.hpp +76 -0
  980. data/vendor/ggml/src/ggml-sycl/norm.cpp +963 -0
  981. data/vendor/ggml/src/ggml-sycl/norm.hpp +32 -0
  982. data/vendor/ggml/src/ggml-sycl/opt-step.cpp +131 -0
  983. data/vendor/ggml/src/ggml-sycl/opt-step.hpp +6 -0
  984. data/vendor/ggml/src/ggml-sycl/outprod.cpp +83 -0
  985. data/vendor/ggml/src/ggml-sycl/outprod.hpp +10 -0
  986. data/vendor/ggml/src/ggml-sycl/pad.cpp +97 -0
  987. data/vendor/ggml/src/ggml-sycl/pad.hpp +24 -0
  988. data/vendor/ggml/src/ggml-sycl/pad_reflect_1d.cpp +100 -0
  989. data/vendor/ggml/src/ggml-sycl/pad_reflect_1d.hpp +10 -0
  990. data/vendor/ggml/src/ggml-sycl/pool.cpp +185 -0
  991. data/vendor/ggml/src/ggml-sycl/pool.hpp +22 -0
  992. data/vendor/ggml/src/ggml-sycl/presets.hpp +79 -0
  993. data/vendor/ggml/src/ggml-sycl/quantize.hpp +133 -0
  994. data/vendor/ggml/src/ggml-sycl/quants.hpp +204 -0
  995. data/vendor/ggml/src/ggml-sycl/repeat_back.cpp +76 -0
  996. data/vendor/ggml/src/ggml-sycl/repeat_back.hpp +8 -0
  997. data/vendor/ggml/src/ggml-sycl/roll.cpp +122 -0
  998. data/vendor/ggml/src/ggml-sycl/roll.hpp +20 -0
  999. data/vendor/ggml/src/ggml-sycl/rope.cpp +651 -0
  1000. data/vendor/ggml/src/ggml-sycl/rope.hpp +26 -0
  1001. data/vendor/ggml/src/ggml-sycl/set.cpp +73 -0
  1002. data/vendor/ggml/src/ggml-sycl/set.hpp +5 -0
  1003. data/vendor/ggml/src/ggml-sycl/set_rows.cpp +577 -0
  1004. data/vendor/ggml/src/ggml-sycl/set_rows.hpp +8 -0
  1005. data/vendor/ggml/src/ggml-sycl/softmax.cpp +425 -0
  1006. data/vendor/ggml/src/ggml-sycl/softmax.hpp +24 -0
  1007. data/vendor/ggml/src/ggml-sycl/solve_tri.cpp +172 -0
  1008. data/vendor/ggml/src/ggml-sycl/solve_tri.hpp +8 -0
  1009. data/vendor/ggml/src/ggml-sycl/ssm_conv.cpp +136 -0
  1010. data/vendor/ggml/src/ggml-sycl/ssm_conv.hpp +5 -0
  1011. data/vendor/ggml/src/ggml-sycl/ssm_scan.cpp +170 -0
  1012. data/vendor/ggml/src/ggml-sycl/ssm_scan.hpp +5 -0
  1013. data/vendor/ggml/src/ggml-sycl/sycl_hw.cpp +67 -0
  1014. data/vendor/ggml/src/ggml-sycl/sycl_hw.hpp +38 -0
  1015. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq112-dv112.cpp +5 -0
  1016. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq128-dv128.cpp +5 -0
  1017. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq256-dv256.cpp +5 -0
  1018. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq40-dv40.cpp +5 -0
  1019. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp +6 -0
  1020. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq576-dv512.cpp +5 -0
  1021. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq64-dv64.cpp +5 -0
  1022. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq72-dv72.cpp +5 -0
  1023. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq80-dv80.cpp +5 -0
  1024. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq96-dv96.cpp +5 -0
  1025. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +8 -0
  1026. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +8 -0
  1027. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +8 -0
  1028. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +8 -0
  1029. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +8 -0
  1030. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +8 -0
  1031. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +8 -0
  1032. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +8 -0
  1033. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +8 -0
  1034. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +8 -0
  1035. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +8 -0
  1036. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +8 -0
  1037. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +8 -0
  1038. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +8 -0
  1039. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +8 -0
  1040. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +8 -0
  1041. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +8 -0
  1042. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +8 -0
  1043. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +8 -0
  1044. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +8 -0
  1045. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +8 -0
  1046. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +8 -0
  1047. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +8 -0
  1048. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +8 -0
  1049. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +8 -0
  1050. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +8 -0
  1051. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +8 -0
  1052. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +8 -0
  1053. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +8 -0
  1054. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +8 -0
  1055. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +8 -0
  1056. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +8 -0
  1057. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +8 -0
  1058. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +8 -0
  1059. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +8 -0
  1060. data/vendor/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +8 -0
  1061. data/vendor/ggml/src/ggml-sycl/topk-moe.cpp +620 -0
  1062. data/vendor/ggml/src/ggml-sycl/topk-moe.hpp +12 -0
  1063. data/vendor/ggml/src/ggml-sycl/tsembd.cpp +73 -0
  1064. data/vendor/ggml/src/ggml-sycl/tsembd.hpp +20 -0
  1065. data/vendor/ggml/src/ggml-sycl/type.hpp +112 -0
  1066. data/vendor/ggml/src/ggml-sycl/upscale.cpp +410 -0
  1067. data/vendor/ggml/src/ggml-sycl/upscale.hpp +9 -0
  1068. data/vendor/ggml/src/ggml-sycl/vecdotq.hpp +1726 -0
  1069. data/vendor/ggml/src/ggml-sycl/wkv.cpp +293 -0
  1070. data/vendor/ggml/src/ggml-sycl/wkv.hpp +10 -0
  1071. data/vendor/ggml/src/ggml-threading.cpp +12 -0
  1072. data/vendor/ggml/src/ggml-threading.h +14 -0
  1073. data/vendor/ggml/src/ggml-version.h.in +4 -0
  1074. data/vendor/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
  1075. data/vendor/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
  1076. data/vendor/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
  1077. data/vendor/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
  1078. data/vendor/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
  1079. data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
  1080. data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
  1081. data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
  1082. data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +149 -0
  1083. data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
  1084. data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
  1085. data/vendor/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
  1086. data/vendor/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
  1087. data/vendor/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
  1088. data/vendor/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
  1089. data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
  1090. data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
  1091. data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
  1092. data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
  1093. data/vendor/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
  1094. data/vendor/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
  1095. data/vendor/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +123 -0
  1096. data/vendor/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +160 -0
  1097. data/vendor/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
  1098. data/vendor/ggml/src/ggml-virtgpu/ggml-backend.cpp +72 -0
  1099. data/vendor/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
  1100. data/vendor/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
  1101. data/vendor/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
  1102. data/vendor/ggml/src/ggml-virtgpu/regenerate_remoting.py +333 -0
  1103. data/vendor/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
  1104. data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
  1105. data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
  1106. data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
  1107. data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +194 -0
  1108. data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
  1109. data/vendor/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +54 -0
  1110. data/vendor/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +99 -0
  1111. data/vendor/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
  1112. data/vendor/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
  1113. data/vendor/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
  1114. data/vendor/ggml/src/ggml-virtgpu/virtgpu.cpp +545 -0
  1115. data/vendor/ggml/src/ggml-virtgpu/virtgpu.h +115 -0
  1116. data/vendor/ggml/src/ggml-vulkan/CMakeLists.txt +254 -0
  1117. data/vendor/ggml/src/ggml-vulkan/cmake/host-toolchain.cmake.in +15 -0
  1118. data/vendor/ggml/src/ggml-vulkan/ggml-vulkan.cpp +20349 -0
  1119. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +43 -0
  1120. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +37 -0
  1121. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/add.comp +69 -0
  1122. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/add1.comp +28 -0
  1123. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp +42 -0
  1124. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/arange.comp +20 -0
  1125. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp +60 -0
  1126. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp +86 -0
  1127. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/argsort_large.comp +114 -0
  1128. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/col2im_1d.comp +61 -0
  1129. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/concat.comp +41 -0
  1130. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +53 -0
  1131. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp +105 -0
  1132. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +473 -0
  1133. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv3d_mm.comp +424 -0
  1134. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp +98 -0
  1135. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +25 -0
  1136. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +51 -0
  1137. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +350 -0
  1138. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_transpose.comp +67 -0
  1139. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/copy_transpose_02.comp +61 -0
  1140. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/count_equal.comp +31 -0
  1141. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/count_experts.comp +135 -0
  1142. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss.comp +78 -0
  1143. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cross_entropy_loss_back.comp +75 -0
  1144. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cumsum.comp +83 -0
  1145. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cumsum_multipass1.comp +60 -0
  1146. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/cumsum_multipass2.comp +66 -0
  1147. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_f32.comp +20 -0
  1148. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +741 -0
  1149. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl +1466 -0
  1150. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_head.glsl +13 -0
  1151. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp +42 -0
  1152. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp +35 -0
  1153. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp +44 -0
  1154. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xs.comp +43 -0
  1155. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp +49 -0
  1156. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp +40 -0
  1157. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp +51 -0
  1158. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp +32 -0
  1159. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_xs.comp +34 -0
  1160. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp +32 -0
  1161. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp +32 -0
  1162. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q1_0.comp +29 -0
  1163. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_0.comp +29 -0
  1164. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp +34 -0
  1165. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp +42 -0
  1166. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_0.comp +30 -0
  1167. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_1.comp +32 -0
  1168. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp +68 -0
  1169. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_0.comp +34 -0
  1170. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_1.comp +35 -0
  1171. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp +70 -0
  1172. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp +33 -0
  1173. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q8_0.comp +42 -0
  1174. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dequant_tq2_0.comp +31 -0
  1175. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +28 -0
  1176. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp +34 -0
  1177. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/div.comp +27 -0
  1178. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl +27 -0
  1179. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/fa_types.glsl +55 -0
  1180. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/bfloat16.comp +7 -0
  1181. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat.comp +7 -0
  1182. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2.comp +7 -0
  1183. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2_decode_vector.comp +7 -0
  1184. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/float_e2m1.comp +7 -0
  1185. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/float_e4m3.comp +7 -0
  1186. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/integer_dot.comp +7 -0
  1187. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/fill.comp +19 -0
  1188. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +760 -0
  1189. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +223 -0
  1190. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +647 -0
  1191. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +487 -0
  1192. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_dequant.glsl +146 -0
  1193. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mask_opt.comp +162 -0
  1194. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mmq_funcs.glsl +203 -0
  1195. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +121 -0
  1196. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp +115 -0
  1197. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +189 -0
  1198. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/geglu.comp +13 -0
  1199. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/geglu_erf.comp +27 -0
  1200. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/geglu_quick.comp +11 -0
  1201. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl +65 -0
  1202. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/generic_head.glsl +11 -0
  1203. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.glsl +75 -0
  1204. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +42 -0
  1205. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_back.comp +25 -0
  1206. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +51 -0
  1207. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/gla.comp +82 -0
  1208. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl +43 -0
  1209. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl +35 -0
  1210. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp +66 -0
  1211. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +138 -0
  1212. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +124 -0
  1213. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +41 -0
  1214. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/lightning_indexer.comp +151 -0
  1215. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +17 -0
  1216. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul.comp +27 -0
  1217. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp +48 -0
  1218. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +264 -0
  1219. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl +230 -0
  1220. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +35 -0
  1221. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp +132 -0
  1222. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp +95 -0
  1223. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp +90 -0
  1224. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp +105 -0
  1225. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp +87 -0
  1226. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp +93 -0
  1227. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp +88 -0
  1228. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +124 -0
  1229. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp +156 -0
  1230. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +128 -0
  1231. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp +132 -0
  1232. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +134 -0
  1233. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +165 -0
  1234. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp +130 -0
  1235. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_tq2_0.comp +102 -0
  1236. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +141 -0
  1237. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +564 -0
  1238. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +475 -0
  1239. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +690 -0
  1240. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +670 -0
  1241. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +89 -0
  1242. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +317 -0
  1243. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +488 -0
  1244. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +99 -0
  1245. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +194 -0
  1246. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/norm.comp +44 -0
  1247. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_adamw.comp +42 -0
  1248. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp +22 -0
  1249. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/out_prod.comp +59 -0
  1250. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pad.comp +64 -0
  1251. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pad_reflect_1d.comp +43 -0
  1252. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pool1d.comp +65 -0
  1253. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/pool2d.comp +74 -0
  1254. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp +127 -0
  1255. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/reglu.comp +9 -0
  1256. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/repeat.comp +26 -0
  1257. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp +37 -0
  1258. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +150 -0
  1259. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_back.comp +55 -0
  1260. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_partials.comp +65 -0
  1261. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/roll.comp +46 -0
  1262. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +218 -0
  1263. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +19 -0
  1264. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +17 -0
  1265. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +17 -0
  1266. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +17 -0
  1267. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +35 -0
  1268. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +17 -0
  1269. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/scale.comp +24 -0
  1270. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/silu_back.comp +26 -0
  1271. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/snake.comp +49 -0
  1272. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +195 -0
  1273. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp +54 -0
  1274. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large1.comp +62 -0
  1275. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large2.comp +79 -0
  1276. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large3.comp +65 -0
  1277. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large_common.glsl +53 -0
  1278. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/solve_tri.comp +81 -0
  1279. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +60 -0
  1280. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/ssm_scan.comp +134 -0
  1281. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/sub.comp +29 -0
  1282. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp +47 -0
  1283. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.glsl +19 -0
  1284. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/swiglu.comp +9 -0
  1285. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_clamp.comp +12 -0
  1286. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp +14 -0
  1287. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp +42 -0
  1288. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_argsort.comp +118 -0
  1289. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_moe.comp +221 -0
  1290. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_nary_search.comp +246 -0
  1291. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/topk_radix_select.comp +144 -0
  1292. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +42 -0
  1293. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +1938 -0
  1294. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/unary.comp +168 -0
  1295. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +178 -0
  1296. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/utils.glsl +37 -0
  1297. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +1367 -0
  1298. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/wkv6.comp +87 -0
  1299. data/vendor/ggml/src/ggml-vulkan/vulkan-shaders/wkv7.comp +91 -0
  1300. data/vendor/ggml/src/ggml-webgpu/CMakeLists.txt +83 -0
  1301. data/vendor/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +3457 -0
  1302. data/vendor/ggml/src/ggml-webgpu/ggml-webgpu.cpp +4790 -0
  1303. data/vendor/ggml/src/ggml-webgpu/pre_wgsl.hpp +808 -0
  1304. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/add_id.wgsl +64 -0
  1305. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
  1306. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +102 -0
  1307. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
  1308. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +142 -0
  1309. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +967 -0
  1310. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +92 -0
  1311. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/conv2d.wgsl +127 -0
  1312. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/conv2d_dw.wgsl +100 -0
  1313. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl +84 -0
  1314. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
  1315. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +89 -0
  1316. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +408 -0
  1317. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_decls.tmpl +134 -0
  1318. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_staging.tmpl +136 -0
  1319. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl +240 -0
  1320. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_blk.wgsl +101 -0
  1321. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_reduce.wgsl +84 -0
  1322. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl +503 -0
  1323. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl +149 -0
  1324. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl +794 -0
  1325. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl +178 -0
  1326. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/im2col.wgsl +77 -0
  1327. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl +40 -0
  1328. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +1052 -0
  1329. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id.wgsl +195 -0
  1330. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_gather.wgsl +52 -0
  1331. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl +154 -0
  1332. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +159 -0
  1333. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.wgsl +210 -0
  1334. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +167 -0
  1335. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl +1560 -0
  1336. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl +297 -0
  1337. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
  1338. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl +21 -0
  1339. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl +178 -0
  1340. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl +70 -0
  1341. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl +151 -0
  1342. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/rope.wgsl +228 -0
  1343. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/row_norm.wgsl +152 -0
  1344. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/scale.wgsl +65 -0
  1345. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/set.wgsl +109 -0
  1346. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +108 -0
  1347. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/set_rows_quant.wgsl +224 -0
  1348. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/soft_max.wgsl +210 -0
  1349. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/solve_tri.wgsl +120 -0
  1350. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/ssm_conv.wgsl +65 -0
  1351. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl +247 -0
  1352. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
  1353. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +213 -0
  1354. data/vendor/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
  1355. data/vendor/ggml/src/ggml-zdnn/CMakeLists.txt +36 -0
  1356. data/vendor/ggml/src/ggml-zdnn/common.hpp +59 -0
  1357. data/vendor/ggml/src/ggml-zdnn/ggml-zdnn.cpp +638 -0
  1358. data/vendor/ggml/src/ggml-zdnn/mmf.cpp +80 -0
  1359. data/vendor/ggml/src/ggml-zdnn/mmf.hpp +12 -0
  1360. data/vendor/ggml/src/ggml-zdnn/utils.cpp +79 -0
  1361. data/vendor/ggml/src/ggml-zdnn/utils.hpp +19 -0
  1362. data/vendor/ggml/src/ggml-zendnn/CMakeLists.txt +91 -0
  1363. data/vendor/ggml/src/ggml-zendnn/ggml-zendnn.cpp +838 -0
  1364. data/vendor/ggml/src/ggml.c +8088 -0
  1365. data/vendor/ggml/src/ggml.cpp +26 -0
  1366. data/vendor/ggml/src/gguf.cpp +1706 -0
  1367. metadata +1459 -0
@@ -0,0 +1,1919 @@
1
+ // SPDX-FileCopyrightText: Copyright 2025-2026 Arm Limited and/or its affiliates <open-source-office@arm.com>
2
+ // SPDX-License-Identifier: MIT
3
+ //
4
+ #include <arm_neon.h>
5
+ #include <cassert>
6
+ #include <cstdio>
7
+ #include <cstdlib>
8
+ #include <atomic>
9
+ #include <cfloat>
10
+ #include <cctype>
11
+ #include <algorithm>
12
+ #include <cmath>
13
+ #include <stdexcept>
14
+ #include <stdint.h>
15
+ #include <string.h>
16
+ #include <string>
17
+ #include <vector>
18
+ #include <array>
19
+ #include <cstddef>
20
+ #include <cstdint>
21
+ #include <fstream>
22
+ #include <map>
23
+ #include <iostream>
24
+ #include <climits>
25
+ #include <charconv>
26
+ #include <system_error>
27
+ #if defined(__linux__)
28
+ #include <asm/hwcap.h>
29
+ #include <dirent.h>
30
+ #include <sys/auxv.h>
31
+ #include <sys/types.h>
32
+ #include <sys/stat.h>
33
+ #include <unistd.h>
34
+ #elif defined(__APPLE__)
35
+ #include <sys/sysctl.h>
36
+ #include <sys/types.h>
37
+ #endif
38
+
39
+ #include "kleidiai.h"
40
+
41
+ #include "ggml-cpu.h"
42
+ #include "ggml-cpu-impl.h"
43
+ #include "ggml-impl.h"
44
+ #include "ggml-feats.h"
45
+ #include "ggml-backend-impl.h"
46
+ #include "ggml-threading.h"
47
+ #include "traits.h"
48
+
49
+ #include "kernels.h"
50
+
51
+ #include "kai/kai_common.h"
52
+
53
+ #define GGML_COMMON_DECL_CPP
54
+ #include "ggml-common.h"
55
+
56
+ static constexpr int GGML_KLEIDIAI_MAX_KERNEL_SLOTS = 2;
57
+ static constexpr uint32_t GGML_KLEIDIAI_PACK_MAGIC = 0x4b4c4149; // "KLAI"
58
+ static constexpr uint16_t GGML_KLEIDIAI_PACK_VERSION = 1;
59
+ static constexpr size_t GGML_KLEIDIAI_PACK_ALIGN = 64;
60
+
61
+ struct ggml_kleidiai_context {
62
+ cpu_feature features;
63
+ ggml_kleidiai_kernels * kernels_q4;
64
+ ggml_kleidiai_kernels * kernels_q8;
65
+ ggml_kleidiai_kernels * kernels_f32;
66
+ int sme_thread_cap; // <= 0 means "SME disabled/unknown"
67
+ int thread_hint; // <= 0 means "no hint"
68
+ int chunk_multiplier;
69
+ } static ctx = { CPU_FEATURE_NONE, nullptr, nullptr, nullptr, 0, -1, 4 };
70
+
71
+ static inline bool is_sme_family(cpu_feature f) {
72
+ return (f & (CPU_FEATURE_SME | CPU_FEATURE_SME2)) != CPU_FEATURE_NONE;
73
+ }
74
+
75
+ static const char* cpu_feature_to_string(cpu_feature f) {
76
+ if (f == CPU_FEATURE_NONE) {
77
+ return "NONE";
78
+ } else if ((f & CPU_FEATURE_SME2) == CPU_FEATURE_SME2) {
79
+ return "SME2";
80
+ } else if ((f & CPU_FEATURE_SME) == CPU_FEATURE_SME) {
81
+ return "SME";
82
+ } else if ((f & CPU_FEATURE_SVE) == CPU_FEATURE_SVE) {
83
+ return "SVE";
84
+ }
85
+ else if ((f & CPU_FEATURE_I8MM) == CPU_FEATURE_I8MM) {
86
+ return "I8MM";
87
+ } else if ((f & CPU_FEATURE_DOTPROD) == CPU_FEATURE_DOTPROD) {
88
+ return "DOTPROD";
89
+ }
90
+ else {
91
+ return "UNKNOWN";
92
+ }
93
+ }
94
+
95
+ #if defined(__linux__) && defined(__aarch64__)
96
+ static bool parse_cpu_dir_name(const char* name, size_t* cpu) {
97
+ if (strncmp(name, "cpu", 3) != 0 ||
98
+ name[3] < '0' || name[3] > '9') {
99
+ return false;
100
+ }
101
+
102
+ const char* first = name + 3;
103
+ const char* last = name + strlen(name);
104
+
105
+ size_t value = 0;
106
+ const auto [end, ec] = std::from_chars(first, last, value, 10);
107
+
108
+ if (ec != std::errc{} || end != last) {
109
+ return false;
110
+ }
111
+
112
+ *cpu = value;
113
+ return true;
114
+ }
115
+
116
+ static std::vector<size_t> detect_cpu_ids() {
117
+ std::vector<size_t> cpus;
118
+
119
+ DIR * dir = opendir("/sys/devices/system/cpu");
120
+ if (dir == nullptr) {
121
+ return cpus;
122
+ }
123
+
124
+ while (dirent * entry = readdir(dir)) {
125
+ size_t cpu = 0;
126
+ if (parse_cpu_dir_name(entry->d_name, &cpu)) {
127
+ cpus.push_back(cpu);
128
+ }
129
+ }
130
+ closedir(dir);
131
+
132
+ std::sort(cpus.begin(), cpus.end());
133
+ cpus.erase(std::unique(cpus.begin(), cpus.end()), cpus.end());
134
+ return cpus;
135
+ }
136
+ #endif
137
+
138
+ #if defined(__APPLE__) && defined(__aarch64__)
139
+ static bool apple_sme_counted_perf_level(std::string name) {
140
+ for (std::string::size_type i = 0; i < name.size(); ++i) {
141
+ name[i] = (char) std::tolower((unsigned char) name[i]);
142
+ }
143
+
144
+ // Conservative ceiling: only count perf-level names observed to provide full SME throughput.
145
+ // Future names should be calibrated here before they raise the automatic SME thread cap.
146
+ return name.find("super") != std::string::npos ||
147
+ name.find("performance") != std::string::npos;
148
+ }
149
+ #endif
150
+
151
+ static void add_smcus_from_smidr(uint64_t smidr, size_t & num_private, std::map<uint32_t, size_t> & shared_counts) {
152
+ // Arm ARM: SMIDR_EL1. SH==0 is implementation-defined; keep the existing
153
+ // conservative policy and only treat zero affinity as private.
154
+ const uint32_t sh = (uint32_t)((smidr >> 13) & 0x3);
155
+ const uint32_t nsmc = (uint32_t)((smidr >> 56) & 0xF);
156
+ const size_t shared_count = nsmc == 0xF ? 1 : (size_t)nsmc + 1;
157
+ const uint32_t affinity = (uint32_t)(smidr & 0xFFFu);
158
+ const uint32_t affinity2 = (uint32_t)((smidr >> 32) & 0xFFFFFu);
159
+ const uint32_t id = (affinity2 << 12) | affinity;
160
+
161
+ if (nsmc == 0xF) {
162
+ GGML_LOG_WARN("kleidiai: NSMC detected as 0xF indicating reseved value, setting min safe shared SMCU count to 1");
163
+ }
164
+
165
+ switch (sh) {
166
+ case 2: // private SMCU
167
+ ++num_private;
168
+ break;
169
+ case 3: // shared SMCU
170
+ if (shared_counts[id] < shared_count) {
171
+ shared_counts[id] = shared_count;
172
+ }
173
+ break;
174
+ case 0:
175
+ if (id == 0) {
176
+ ++num_private;
177
+ } else if (shared_counts[id] < shared_count) {
178
+ shared_counts[id] = shared_count;
179
+ }
180
+ break;
181
+ default:
182
+ break;
183
+ }
184
+ }
185
+
186
+ static size_t detect_num_smcus() {
187
+ const auto runtime_feat = ggml_feats_get_arch64_runtime();
188
+ if (!runtime_feat.has_sme) {
189
+ return 0;
190
+ }
191
+
192
+ #if defined(__linux__) && defined(__aarch64__)
193
+ // Linux/aarch64: Best-effort count of Streaming Mode Compute Units (SMCUs) via SMIDR_EL1 sysfs.
194
+ size_t num_private = 0;
195
+ std::map<uint32_t, size_t> shared_counts;
196
+
197
+ const std::vector<size_t> cpus = detect_cpu_ids();
198
+ for (const size_t cpu : cpus) {
199
+ const std::string path =
200
+ "/sys/devices/system/cpu/cpu" + std::to_string(cpu) +
201
+ "/regs/identification/smidr_el1";
202
+
203
+ std::ifstream file(path);
204
+ if (!file.is_open()) {
205
+ continue;
206
+ }
207
+
208
+ uint64_t smidr = 0;
209
+ if (!(file >> std::hex >> smidr)) {
210
+ continue;
211
+ }
212
+
213
+ add_smcus_from_smidr(smidr, num_private, shared_counts);
214
+ }
215
+
216
+ size_t total = num_private;
217
+ for (const auto & entry : shared_counts) {
218
+ total += entry.second;
219
+ }
220
+ return total;
221
+
222
+ #elif defined(__APPLE__) && defined(__aarch64__)
223
+ int perf_levels = 0;
224
+ size_t size = sizeof(perf_levels);
225
+ if (sysctlbyname("hw.nperflevels", &perf_levels, &size, nullptr, 0) != 0 ||
226
+ size != sizeof(perf_levels) || perf_levels <= 0) {
227
+ return 0;
228
+ }
229
+
230
+ size_t units = 0;
231
+ for (int i = 0; i < perf_levels; ++i) {
232
+ char key[64] = {};
233
+ int physical_cpus = 0;
234
+ int cpus_per_l2 = 0;
235
+
236
+ snprintf(key, sizeof(key), "hw.perflevel%d.physicalcpu", i);
237
+ size = sizeof(physical_cpus);
238
+ if (sysctlbyname(key, &physical_cpus, &size, nullptr, 0) != 0 ||
239
+ size != sizeof(physical_cpus) || physical_cpus <= 0) {
240
+ continue;
241
+ }
242
+
243
+ snprintf(key, sizeof(key), "hw.perflevel%d.cpusperl2", i);
244
+ size = sizeof(cpus_per_l2);
245
+ if (sysctlbyname(key, &cpus_per_l2, &size, nullptr, 0) != 0 ||
246
+ size != sizeof(cpus_per_l2) || cpus_per_l2 <= 0) {
247
+ continue;
248
+ }
249
+
250
+ snprintf(key, sizeof(key), "hw.perflevel%d.name", i);
251
+ size = 0;
252
+ if (sysctlbyname(key, nullptr, &size, nullptr, 0) != 0 || size == 0) {
253
+ continue;
254
+ }
255
+
256
+ std::string name(size, '\0');
257
+ if (sysctlbyname(key, &name[0], &size, nullptr, 0) != 0) {
258
+ continue;
259
+ }
260
+ name.resize(size);
261
+ while (!name.empty() && name.back() == '\0') {
262
+ name.pop_back();
263
+ }
264
+
265
+ if (apple_sme_counted_perf_level(name)) {
266
+ units += (size_t) ((physical_cpus + cpus_per_l2 - 1) / cpus_per_l2);
267
+ }
268
+ }
269
+
270
+ return units;
271
+
272
+ #elif defined(_WIN32) && (defined(_M_ARM64) || defined(__aarch64__))
273
+ // No verified Windows arm64 SMCU detection path yet. Return unknown and use
274
+ // GGML_KLEIDIAI_SME=N as a diagnostics/debug override for SME thread cap
275
+ // calibration until a detection mechanism is verified on real hardware.
276
+ return 0;
277
+
278
+ #else
279
+ return 0;
280
+ #endif
281
+ }
282
+
283
+ static int parse_uint_env(const char *s, const char *name, bool *ok) {
284
+ if (!s) { *ok = false; return 0; }
285
+ char *end = nullptr;
286
+ long v = strtol(s, &end, 10);
287
+ if (end == s || *end != '\0') {
288
+ GGML_LOG_WARN("kleidiai: invalid %s='%s' (expected integer)\n", name, s);
289
+ *ok = false;
290
+ return 0;
291
+ }
292
+ if (v < 0 || v > INT_MAX) {
293
+ GGML_LOG_WARN("kleidiai: out-of-range %s='%s'\n", name, s);
294
+ *ok = false;
295
+ return 0;
296
+ }
297
+ *ok = true;
298
+ return (int)v;
299
+ }
300
+
301
+ static void init_kleidiai_context(void) {
302
+ ggml_critical_section_start();
303
+ static bool initialized = false;
304
+
305
+ if (!initialized) {
306
+ initialized = true;
307
+
308
+ // Optional diagnostics/debug overrides; production defaults come from runtime detection.
309
+ const char *env_sme = getenv("GGML_KLEIDIAI_SME");
310
+ const char *env_threads = getenv("GGML_TOTAL_THREADS");
311
+ const char *env_chunk_mult = getenv("GGML_KLEIDIAI_CHUNK_MULTIPLIER");
312
+
313
+ const auto runtime_feat = ggml_feats_get_arch64_runtime();
314
+
315
+ size_t detected_smcus = 0;
316
+
317
+ ctx.features = (runtime_feat.has_dotprod ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) |
318
+ (runtime_feat.has_i8mm ? CPU_FEATURE_I8MM : CPU_FEATURE_NONE) |
319
+ (runtime_feat.has_fp16 ? CPU_FEATURE_FP16 : CPU_FEATURE_NONE) |
320
+ (runtime_feat.sve_cnt == QK8_0 ? CPU_FEATURE_SVE : CPU_FEATURE_NONE);
321
+
322
+ if (env_threads) {
323
+ bool ok = false;
324
+ int hint = parse_uint_env(env_threads, "GGML_TOTAL_THREADS", &ok);
325
+ if (ok && hint > 0) {
326
+ ctx.thread_hint = hint;
327
+ }
328
+ }
329
+
330
+ if (env_chunk_mult) {
331
+ bool ok = false;
332
+ int multiplier = parse_uint_env(env_chunk_mult, "GGML_KLEIDIAI_CHUNK_MULTIPLIER", &ok);
333
+ if (ok && multiplier > 0) {
334
+ ctx.chunk_multiplier = multiplier;
335
+ }
336
+ }
337
+
338
+ int sme_cores = 0;
339
+ bool sme_env_ok = false;
340
+ bool sme_env_set = (env_sme != nullptr);
341
+
342
+ const bool has_supported_sme_family = runtime_feat.has_sme;
343
+ bool sme_cap_detected = false;
344
+
345
+ if (has_supported_sme_family) {
346
+ detected_smcus = detect_num_smcus();
347
+ sme_cap_detected = detected_smcus > 0;
348
+ // Some platforms expose SME without exposing a calibrated SMCU count.
349
+ // Use one SME thread as the conservative default; add platform SMCU detection to raise it.
350
+ sme_cores = sme_cap_detected ? (int)detected_smcus : 1;
351
+
352
+ if (!sme_env_set && !sme_cap_detected) {
353
+ GGML_LOG_INFO("kleidiai: SME detected; SMCU count unavailable, using conservative SME thread cap=1\n");
354
+ }
355
+ }
356
+
357
+ // Runtime-detect SME support and available SMCUs first. The detected SMCU
358
+ // count is used as the SME thread cap, and GGML_KLEIDIAI_SME can debug-override that:
359
+ // - unset: use runtime detection.
360
+ // - 0: disable SME-family kernels.
361
+ // - N > 0: use N as the SME thread cap, if an SME-family kernel is selectable.
362
+ if (sme_env_set) {
363
+ bool ok = false;
364
+ int v = parse_uint_env(env_sme, "GGML_KLEIDIAI_SME", &ok);
365
+ sme_env_ok = ok;
366
+
367
+ if (ok) {
368
+ if (has_supported_sme_family) {
369
+ sme_cores = v;
370
+ } else {
371
+ if (v > 0) {
372
+ GGML_LOG_WARN("kleidiai: GGML_KLEIDIAI_SME=%d but SME is not supported on this CPU; disabling SME-family kernels\n", v);
373
+ }
374
+ sme_cores = 0;
375
+ }
376
+ } else {
377
+ GGML_LOG_WARN("kleidiai: GGML_KLEIDIAI_SME set but parsing failed; using automatic SME thread cap\n");
378
+ }
379
+ }
380
+
381
+ if (sme_cores > 0 && has_supported_sme_family) {
382
+ ctx.features |= CPU_FEATURE_SME;
383
+ if (runtime_feat.has_sme2) {
384
+ ctx.features |= CPU_FEATURE_SME2;
385
+ }
386
+ }
387
+
388
+ // Kernel selection
389
+ ctx.kernels_q4 = ggml_kleidiai_select_kernels_q4_0(ctx.features);
390
+ ctx.kernels_q8 = ggml_kleidiai_select_kernels_q8_0(ctx.features);
391
+ ctx.kernels_f32 = ggml_kleidiai_select_kernels_f32(ctx.features);
392
+
393
+ if (!ctx.kernels_q4) {
394
+ GGML_LOG_INFO("kleidiai: no compatible q4 kernels found for CPU features mask %d\n", (int)ctx.features);
395
+ } else {
396
+ GGML_LOG_INFO("kleidiai: primary q4 kernel feature %s\n", cpu_feature_to_string(ctx.kernels_q4->required_cpu));
397
+ }
398
+
399
+ if (!ctx.kernels_q8) {
400
+ GGML_LOG_INFO("kleidiai: no compatible q8 kernels found for CPU features mask %d\n", (int)ctx.features);
401
+ } else {
402
+ GGML_LOG_INFO("kleidiai: primary q8 kernel feature %s\n", cpu_feature_to_string(ctx.kernels_q8->required_cpu));
403
+ }
404
+
405
+ if (!ctx.kernels_f32) {
406
+ GGML_LOG_INFO("kleidiai: no compatible f32 kernels found for CPU features mask %d\n", (int)ctx.features);
407
+ } else {
408
+ GGML_LOG_INFO("kleidiai: primary f32 kernel feature %s\n", cpu_feature_to_string(ctx.kernels_f32->required_cpu));
409
+ }
410
+
411
+ const bool has_selected_sme_family_kernel =
412
+ (ctx.kernels_q4 && is_sme_family(ctx.kernels_q4->required_cpu)) ||
413
+ (ctx.kernels_q8 && is_sme_family(ctx.kernels_q8->required_cpu)) ||
414
+ (ctx.kernels_f32 && is_sme_family(ctx.kernels_f32->required_cpu));
415
+ ctx.sme_thread_cap = has_selected_sme_family_kernel ? sme_cores : 0;
416
+
417
+ if (has_selected_sme_family_kernel) {
418
+ if (sme_env_set && sme_env_ok && sme_cores > 0) {
419
+ GGML_LOG_INFO("kleidiai: SME enabled (GGML_KLEIDIAI_SME=%d debug override)\n", sme_cores);
420
+ } else if (sme_cap_detected) {
421
+ GGML_LOG_INFO("kleidiai: SME enabled (runtime-detected SME thread cap=%d)\n", sme_cores);
422
+ } else {
423
+ GGML_LOG_INFO("kleidiai: SME enabled (runtime SME detected, conservative thread cap=%d)\n", sme_cores);
424
+ }
425
+ } else {
426
+ GGML_LOG_INFO("kleidiai: SME disabled\n");
427
+ }
428
+ }
429
+
430
+ ggml_critical_section_end();
431
+ }
432
+
433
+ static inline int kleidiai_sme_thread_cap() {
434
+ return ctx.sme_thread_cap;
435
+ }
436
+
437
+ static inline size_t align_up(size_t value, size_t alignment) {
438
+ if (alignment == 0) {
439
+ return value;
440
+ }
441
+ const size_t remainder = value % alignment;
442
+ return remainder == 0 ? value : value + (alignment - remainder);
443
+ }
444
+
445
+ static inline size_t gcd_size(size_t a, size_t b) {
446
+ while (b != 0) {
447
+ const size_t t = a % b;
448
+ a = b;
449
+ b = t;
450
+ }
451
+ return a;
452
+ }
453
+
454
+ static inline bool lcm_size(size_t a, size_t b, size_t & result) {
455
+ if (a == 0 || b == 0) {
456
+ result = 0;
457
+ return false;
458
+ }
459
+ const size_t g = gcd_size(a, b);
460
+ const size_t q = a / g;
461
+ if (q > SIZE_MAX / b) {
462
+ return false;
463
+ }
464
+ result = q * b;
465
+ return true;
466
+ }
467
+
468
+ static inline size_t ceil_div_size(size_t a, size_t b) {
469
+ return b == 0 ? 0 : (a + b - 1) / b;
470
+ }
471
+
472
+ static inline size_t kleidiai_chunk_cols(size_t n, int nth_total, bool disable_chunking, size_t n_step) {
473
+ const size_t multiplier = (nth_total == 1 || disable_chunking) ? 1 : std::max<size_t>(1, (size_t) ctx.chunk_multiplier);
474
+ const size_t divisor = std::max<size_t>(1, (size_t) nth_total * multiplier);
475
+ const size_t chunk_cols = align_up(std::max<size_t>(1, ceil_div_size(n, divisor)), n_step);
476
+ return chunk_cols ? chunk_cols : n_step;
477
+ }
478
+
479
+ struct kleidiai_block_args {
480
+ size_t lhs_bl;
481
+ size_t rhs_bl;
482
+ size_t pack_bl;
483
+ };
484
+
485
+ static inline kleidiai_block_args kleidiai_get_block_args(ggml_type rhs_type) {
486
+ switch (rhs_type) {
487
+ case GGML_TYPE_Q4_0:
488
+ return { QK4_0, QK4_0, QK4_0 };
489
+ case GGML_TYPE_Q8_0:
490
+ return { 0, 0, QK8_0 };
491
+ default:
492
+ return { 0, 0, 0 };
493
+ }
494
+ }
495
+
496
+ static inline bool kleidiai_pack_fallback_allowed() {
497
+ if (ctx.sme_thread_cap <= 0) {
498
+ return false;
499
+ }
500
+ if (ctx.thread_hint <= 0) {
501
+ return true;
502
+ }
503
+ return ctx.thread_hint > ctx.sme_thread_cap;
504
+ }
505
+
506
+ struct kleidiai_weight_header {
507
+ uint32_t magic;
508
+ uint16_t version;
509
+ uint16_t slot_count;
510
+ uint64_t offsets[GGML_KLEIDIAI_MAX_KERNEL_SLOTS];
511
+ uint64_t sizes[GGML_KLEIDIAI_MAX_KERNEL_SLOTS];
512
+ };
513
+
514
+ static inline kleidiai_weight_header * kleidiai_weight_header_from_ptr(void * data) {
515
+ return reinterpret_cast<kleidiai_weight_header *>(data);
516
+ }
517
+
518
+ static inline const kleidiai_weight_header * kleidiai_weight_header_from_ptr(const void * data) {
519
+ return reinterpret_cast<const kleidiai_weight_header *>(data);
520
+ }
521
+
522
+ static inline bool kleidiai_is_weight_header_valid(const kleidiai_weight_header * header) {
523
+ if (!header) {
524
+ return false;
525
+ }
526
+ if (header->magic != GGML_KLEIDIAI_PACK_MAGIC || header->version != GGML_KLEIDIAI_PACK_VERSION) {
527
+ return false;
528
+ }
529
+ if (header->slot_count == 0 || header->slot_count > GGML_KLEIDIAI_MAX_KERNEL_SLOTS) {
530
+ return false;
531
+ }
532
+ return true;
533
+ }
534
+
535
+ static inline uint8_t * kleidiai_weight_slot_ptr(kleidiai_weight_header * header, int slot) {
536
+ if (!kleidiai_is_weight_header_valid(header)) {
537
+ return nullptr;
538
+ }
539
+ if (slot < 0 || slot >= header->slot_count) {
540
+ return nullptr;
541
+ }
542
+ return reinterpret_cast<uint8_t *>(header) + header->offsets[slot];
543
+ }
544
+
545
+ static inline const uint8_t * kleidiai_weight_slot_ptr(const kleidiai_weight_header * header, int slot) {
546
+ if (!kleidiai_is_weight_header_valid(header)) {
547
+ return nullptr;
548
+ }
549
+ if (slot < 0 || slot >= header->slot_count) {
550
+ return nullptr;
551
+ }
552
+ return reinterpret_cast<const uint8_t *>(header) + header->offsets[slot];
553
+ }
554
+
555
+ static inline ggml_kleidiai_kernels * kleidiai_primary_kernel_q4() {
556
+ return ctx.kernels_q4;
557
+ }
558
+
559
+ static inline ggml_kleidiai_kernels * kleidiai_primary_kernel_q8() {
560
+ return ctx.kernels_q8;
561
+ }
562
+
563
+ static inline ggml_kleidiai_kernels * kleidiai_primary_kernel_f32() {
564
+ return ctx.kernels_f32;
565
+ }
566
+
567
+ template <typename SelectFallback>
568
+ static int kleidiai_collect_kernel_chain_common(
569
+ ggml_kleidiai_kernels * primary,
570
+ cpu_feature features,
571
+ std::array<ggml_kleidiai_kernels *, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> & out,
572
+ SelectFallback select_fallback) {
573
+ int count = 0;
574
+ if (!primary) {
575
+ return 0;
576
+ }
577
+ out[count++] = primary;
578
+
579
+ if (primary->rhs_info.repack_mode == RHS_REPACK_SINGLE_ONLY) {
580
+ return count;
581
+ }
582
+
583
+ if (is_sme_family(primary->required_cpu)) {
584
+ const cpu_feature fallback_mask = static_cast<cpu_feature>(features & ~(CPU_FEATURE_SME | CPU_FEATURE_SME2));
585
+ if (fallback_mask != CPU_FEATURE_NONE) {
586
+ ggml_kleidiai_kernels * fallback = select_fallback(fallback_mask);
587
+ if (fallback && fallback != primary &&
588
+ fallback->rhs_info.repack_mode != RHS_REPACK_SINGLE_ONLY &&
589
+ fallback->lhs_type == primary->lhs_type &&
590
+ fallback->rhs_type == primary->rhs_type &&
591
+ fallback->op_type == primary->op_type) {
592
+ out[count++] = fallback;
593
+ }
594
+ }
595
+ }
596
+
597
+ return count;
598
+ }
599
+
600
+ static int kleidiai_collect_kernel_chain(const struct ggml_tensor * op,
601
+ std::array<ggml_kleidiai_kernels *, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> & out) {
602
+ ggml_kleidiai_kernels * primary = ggml_kleidiai_select_kernels(ctx.features, op);
603
+ return kleidiai_collect_kernel_chain_common(primary, ctx.features, out,
604
+ [&](cpu_feature mask) { return ggml_kleidiai_select_kernels(mask, op); });
605
+ }
606
+
607
+ static int kleidiai_collect_q4_chain(std::array<ggml_kleidiai_kernels *, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> & out) {
608
+ ggml_kleidiai_kernels * primary = kleidiai_primary_kernel_q4();
609
+ return kleidiai_collect_kernel_chain_common(primary, ctx.features, out,
610
+ [&](cpu_feature mask) { return ggml_kleidiai_select_kernels_q4_0(mask); });
611
+ }
612
+
613
+ static int kleidiai_collect_q8_chain(std::array<ggml_kleidiai_kernels *, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> & out) {
614
+ ggml_kleidiai_kernels * primary = kleidiai_primary_kernel_q8();
615
+ return kleidiai_collect_kernel_chain_common(primary, ctx.features, out,
616
+ [&](cpu_feature mask) { return ggml_kleidiai_select_kernels_q8_0(mask); });
617
+ }
618
+
619
+ static int kleidiai_collect_f32_chain(std::array<ggml_kleidiai_kernels *, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> & out) {
620
+ ggml_kleidiai_kernels * primary = kleidiai_primary_kernel_f32();
621
+ return kleidiai_collect_kernel_chain_common(primary, ctx.features, out,
622
+ [&](cpu_feature mask) { return ggml_kleidiai_select_kernels_f32(mask); });
623
+ }
624
+
625
+ static inline int64_t ggml_ne(const ggml_tensor * tensor, int dim) {
626
+ GGML_ASSERT(dim >= 0 && dim < GGML_MAX_DIMS);
627
+ return tensor->ne[dim];
628
+ }
629
+
630
+ namespace ggml::cpu::kleidiai {
631
+
632
+ static size_t round_down(size_t x, size_t y) {
633
+ return y == 0 ? x : x - (x % y);
634
+ }
635
+
636
+ static void transpose_f32kxn_f16nxk(size_t n, size_t k, float * dst, const uint16_t * src, size_t rhs_stride) {
637
+ size_t src_stride = rhs_stride / sizeof(uint16_t);
638
+ size_t dst_stride = n;
639
+
640
+ for (size_t k_idx = 0; k_idx < k; ++k_idx) {
641
+ for (size_t n_idx = 0; n_idx < n; ++n_idx) {
642
+ uint16_t v = *(src + k_idx + n_idx * src_stride);
643
+ *(dst + n_idx + k_idx * dst_stride) = kai_cast_f32_f16(v);
644
+ }
645
+ }
646
+ }
647
+
648
+ class tensor_traits : public ggml::cpu::tensor_traits {
649
+ bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
650
+ if (op->op != GGML_OP_MUL_MAT) {
651
+ return false;
652
+ }
653
+
654
+ std::array<ggml_kleidiai_kernels *, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> kernel_chain;
655
+ const int slot_count = kleidiai_collect_kernel_chain(op, kernel_chain);
656
+ if (slot_count == 0) {
657
+ return false;
658
+ }
659
+
660
+ const bool is_gemv = op->src[1]->ne[1] == 1;
661
+ const size_t k = op->src[0]->ne[0];
662
+ const size_t n = op->src[0]->ne[1];
663
+ const size_t m = op->src[1]->ne[1];
664
+
665
+ if (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_Q8_0) {
666
+ const size_t qk = (op->src[0]->type == GGML_TYPE_Q4_0) ? QK4_0 : QK8_0;
667
+
668
+ size_t cursor = 0;
669
+ bool any_slot = false;
670
+
671
+ for (int slot = 0; slot < slot_count; ++slot) {
672
+ ggml_kleidiai_kernels * kernels = kernel_chain[slot];
673
+ lhs_packing_info * lhs_info = is_gemv ? &kernels->gemv_lhs_info : &kernels->gemm_lhs_info;
674
+ kernel_info * kernel = is_gemv ? &kernels->gemv : &kernels->gemm;
675
+
676
+ if (!lhs_info || !lhs_info->packed_size_ex || !kernel) {
677
+ return false;
678
+ }
679
+
680
+ const size_t mr = kernel->get_mr();
681
+ const size_t kr = kernel->get_kr();
682
+ const size_t sr = kernel->get_sr();
683
+
684
+ const size_t packed = lhs_info->packed_size_ex(m, k, qk, mr, kr, sr);
685
+
686
+ cursor = align_up(cursor, GGML_KLEIDIAI_PACK_ALIGN);
687
+ cursor += packed;
688
+ any_slot = true;
689
+ }
690
+
691
+ if (!any_slot) {
692
+ return false;
693
+ }
694
+
695
+ size = cursor;
696
+ return true;
697
+ }
698
+
699
+ if (op->src[0]->type == GGML_TYPE_F32) {
700
+ ggml_kleidiai_kernels * primary = kernel_chain[0];
701
+ kernel_info * gemv_kernel = primary ? &primary->gemv : nullptr;
702
+ if (is_gemv && op->src[1]->nb[0] == (int64_t) sizeof(float) && gemv_kernel &&
703
+ gemv_kernel->get_lhs_offset_ex && gemv_kernel->get_rhs_packed_offset_ex &&
704
+ gemv_kernel->run_kernel_ex && gemv_kernel->get_dst_offset) {
705
+ size = 0;
706
+ return true;
707
+ }
708
+
709
+ size_t cursor = 0;
710
+ bool any_slot = false;
711
+
712
+ for (int slot = 0; slot < slot_count; ++slot) {
713
+ ggml_kleidiai_kernels * kernels = kernel_chain[slot];
714
+ lhs_packing_info * lhs_info = &kernels->gemm_lhs_info;
715
+ kernel_info * kernel = &kernels->gemm;
716
+
717
+ if (!lhs_info || !lhs_info->packed_size_ex || !kernel) {
718
+ return false;
719
+ }
720
+
721
+ const size_t mr = kernel->get_mr();
722
+ const size_t kr = kernel->get_kr();
723
+ const size_t sr = kernel->get_sr();
724
+
725
+ cursor = align_up(cursor, GGML_KLEIDIAI_PACK_ALIGN);
726
+ cursor += lhs_info->packed_size_ex(m, k, 0, mr, kr, sr);
727
+ any_slot = true;
728
+ }
729
+
730
+ if (!any_slot) {
731
+ return false;
732
+ }
733
+
734
+ size = cursor;
735
+ return true;
736
+ }
737
+
738
+ if (op->src[0]->type == GGML_TYPE_F16) {
739
+ const int64_t lhs_batch_size0 = op->src[1]->ne[2];
740
+ const int64_t rhs_batch_size0 = op->src[0]->ne[2];
741
+ GGML_ASSERT(rhs_batch_size0 > 0);
742
+ const int64_t r = lhs_batch_size0 / rhs_batch_size0;
743
+
744
+ size_t cursor = 0;
745
+ bool any_slot = false;
746
+
747
+ for (int slot = 0; slot < slot_count; ++slot) {
748
+ ggml_kleidiai_kernels * kernels = kernel_chain[slot];
749
+ lhs_packing_info * lhs_info = is_gemv ? &kernels->gemv_lhs_info : &kernels->gemm_lhs_info;
750
+ kernel_info * kernel = is_gemv ? &kernels->gemv : &kernels->gemm;
751
+ if (!lhs_info || !lhs_info->packed_size_ex || !kernels->rhs_info.packed_size_ex || !kernel) {
752
+ return false;
753
+ }
754
+
755
+ const size_t mr = kernel->get_mr();
756
+ const size_t kr = kernel->get_kr();
757
+ const size_t sr = kernel->get_sr();
758
+
759
+ cursor = align_up(cursor, GGML_KLEIDIAI_PACK_ALIGN);
760
+ cursor += lhs_info->packed_size_ex(m * r, k, 0, mr, kr, sr);
761
+ any_slot = true;
762
+ }
763
+
764
+ for (int slot = 0; slot < slot_count; ++slot) {
765
+ ggml_kleidiai_kernels * kernels = kernel_chain[slot];
766
+ kernel_info * kernel = is_gemv ? &kernels->gemv : &kernels->gemm;
767
+ if (!kernel || !kernels->rhs_info.packed_size_ex) {
768
+ return false;
769
+ }
770
+ cursor = align_up(cursor, GGML_KLEIDIAI_PACK_ALIGN);
771
+ cursor += kernels->rhs_info.packed_size_ex(n, k, kernel->get_nr(), kernel->get_kr(), 0);
772
+ }
773
+
774
+ cursor = align_up(cursor, GGML_KLEIDIAI_PACK_ALIGN);
775
+ cursor += k * n * sizeof(float);
776
+ cursor = align_up(cursor, GGML_KLEIDIAI_PACK_ALIGN);
777
+ cursor += n * sizeof(float);
778
+
779
+ if (!any_slot) {
780
+ return false;
781
+ }
782
+
783
+ size = cursor;
784
+ return true;
785
+ }
786
+
787
+ return false;
788
+ }
789
+
790
+ bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * dst) override {
791
+ if (dst->op == GGML_OP_MUL_MAT) {
792
+ if (dst->src[0]->type == GGML_TYPE_Q4_0 || dst->src[0]->type == GGML_TYPE_Q8_0) {
793
+ return compute_forward_qx(params, dst);
794
+ } else if (dst->src[0]->type == GGML_TYPE_F32) {
795
+ return compute_forward_f32(params, dst);
796
+ } else if (dst->src[0]->type == GGML_TYPE_F16) {
797
+ return compute_forward_fp16(params, dst);
798
+ }
799
+ } else if (dst->op == GGML_OP_GET_ROWS) {
800
+ if (dst->src[0]->type == GGML_TYPE_Q4_0 || dst->src[0]->type == GGML_TYPE_Q8_0) {
801
+ return compute_forward_get_rows(params, dst);
802
+ }
803
+ }
804
+ return false;
805
+ }
806
+
807
+ bool compute_forward_f32(ggml_compute_params * params, struct ggml_tensor * dst) {
808
+ GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
809
+
810
+ const ggml_tensor * src0 = dst->src[0];
811
+ const ggml_tensor * src1 = dst->src[1];
812
+
813
+ GGML_TENSOR_BINARY_OP_LOCALS
814
+
815
+ if (src1->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
816
+ return false;
817
+ }
818
+
819
+ ggml_kleidiai_kernels * kernels = kleidiai_primary_kernel_f32();
820
+ if (!kernels) {
821
+ return false;
822
+ }
823
+
824
+ const size_t k = ne00;
825
+ const size_t m = ne11;
826
+ const size_t n = ne01;
827
+ const bool use_gemv = m == 1 && src1->nb[0] == (int64_t) sizeof(float) &&
828
+ kernels->gemv.get_lhs_offset_ex &&
829
+ kernels->gemv.get_rhs_packed_offset_ex &&
830
+ kernels->gemv.run_kernel_ex &&
831
+ kernels->gemv.get_dst_offset;
832
+
833
+ kernel_info * kernel = use_gemv ? &kernels->gemv : &kernels->gemm;
834
+ lhs_packing_info * lhs_info = &kernels->gemm_lhs_info;
835
+
836
+ if (!kernel || !kernel->get_lhs_offset_ex ||
837
+ !kernel->get_rhs_packed_offset_ex || !kernel->run_kernel_ex || !kernel->get_dst_offset) {
838
+ return false;
839
+ }
840
+
841
+ if (!use_gemv && (!lhs_info || !lhs_info->get_offset || !lhs_info->get_packed_offset_ex ||
842
+ !lhs_info->packed_size_ex || !lhs_info->pack_func_ex)) {
843
+ return false;
844
+ }
845
+
846
+ const kleidiai_weight_header * header = kleidiai_weight_header_from_ptr(src0->data);
847
+ const bool has_header = kleidiai_is_weight_header_valid(header);
848
+
849
+ const uint8_t * rhs_base = has_header ? kleidiai_weight_slot_ptr(header, 0)
850
+ : static_cast<const uint8_t *>(src0->data);
851
+ if (!rhs_base) {
852
+ return false;
853
+ }
854
+
855
+ const int nth = params->nth > 0 ? params->nth : 1;
856
+ const int ith = params->ith;
857
+
858
+ const size_t mr = kernel->get_mr();
859
+ const size_t kr = kernel->get_kr();
860
+ const size_t sr = kernel->get_sr();
861
+
862
+ const size_t lhs_packed_size = use_gemv ? 0 : lhs_info->packed_size_ex(m, k, 0, mr, kr, sr);
863
+ if (!use_gemv) {
864
+ GGML_ASSERT(lhs_packed_size <= params->wsize);
865
+ }
866
+
867
+ uint8_t * lhs_packed = static_cast<uint8_t *>(params->wdata);
868
+ const size_t dst_stride = dst->nb[1];
869
+ const size_t n_step = kernel->get_n_step() ? kernel->get_n_step() : 1;
870
+ const bool disable_chunking = ggml_is_numa();
871
+ GGML_ASSERT(n <= (size_t) INT_MAX);
872
+
873
+ for (int64_t batch_idx = 0; batch_idx < ne12; ++batch_idx) {
874
+ const uint8_t * lhs_batch_base = static_cast<const uint8_t *>(src1->data) + batch_idx * src1->nb[2];
875
+ uint8_t * dst_batch_base = static_cast<uint8_t *>(dst->data) + batch_idx * dst->nb[2];
876
+
877
+ if (!use_gemv) {
878
+ const int64_t m_roundup_mr = kai_roundup((int64_t)m, (int64_t)mr);
879
+ int64_t max_threads = mr ? (m_roundup_mr / (int64_t)mr) : nth;
880
+ max_threads = std::max<int64_t>(1, max_threads);
881
+ const int64_t use_threads = std::min<int64_t>(nth, max_threads);
882
+
883
+ if (ith < use_threads) {
884
+ const int64_t num_m_per_thread0 = round_down((size_t)(m_roundup_mr / use_threads), mr);
885
+ const int64_t num_m_per_threadN_1 = (int64_t)m - (use_threads - 1) * num_m_per_thread0;
886
+
887
+ const int64_t m_start = (int64_t)ith * num_m_per_thread0;
888
+ const int64_t m_count = (ith == use_threads - 1) ? num_m_per_threadN_1 : num_m_per_thread0;
889
+
890
+ const size_t base_packed_off = lhs_info->get_packed_offset_ex(m_start, k, 0, mr, kr, sr);
891
+ const size_t next_block_off = lhs_info->get_packed_offset_ex(m_start + mr, k, 0, mr, kr, sr);
892
+ const size_t row_stride_bytes = mr ? (next_block_off - base_packed_off) / mr : 0;
893
+
894
+ int64_t remaining = m_count;
895
+ int64_t cur = m_start;
896
+
897
+ while (remaining > 0) {
898
+ const int64_t take = std::min<int64_t>((int64_t)m - cur, remaining);
899
+ const size_t src_off = lhs_info->get_offset(cur, src1->nb[1]);
900
+ const void * src_ptr = lhs_batch_base + src_off;
901
+ const size_t dst_off = base_packed_off + (size_t)(cur - m_start) * row_stride_bytes;
902
+ void * dst_ptr = lhs_packed + dst_off;
903
+
904
+ lhs_info->pack_func_ex(take, k, 0, mr, kr, sr, 0, src_ptr, src1->nb[1], dst_ptr);
905
+
906
+ cur += take;
907
+ remaining -= take;
908
+ }
909
+ }
910
+ }
911
+
912
+ if (ith == 0) {
913
+ ggml_threadpool_chunk_set(params->threadpool, 0);
914
+ }
915
+
916
+ ggml_barrier(params->threadpool);
917
+
918
+ const size_t chunk_cols = kleidiai_chunk_cols(n, nth, disable_chunking, n_step);
919
+ GGML_ASSERT(chunk_cols <= (size_t) INT_MAX);
920
+
921
+ int current_col = ggml_threadpool_chunk_add(params->threadpool, (int) chunk_cols);
922
+ while ((size_t) current_col < n) {
923
+ const size_t n_start = (size_t) current_col;
924
+ const size_t n_to_process = std::min(chunk_cols, n - n_start);
925
+
926
+ if (n_to_process > 0) {
927
+ const size_t lhs_offset = use_gemv ? kernel->get_lhs_offset_ex(0, k, 0)
928
+ : lhs_info->get_packed_offset_ex(0, k, 0, mr, kr, sr);
929
+ const size_t rhs_packed_offset = kernel->get_rhs_packed_offset_ex(n_start, k, 0);
930
+ const size_t dst_offset = kernel->get_dst_offset(0, n_start, dst_stride);
931
+
932
+ const void * lhs_ptr = use_gemv ? lhs_batch_base + lhs_offset
933
+ : lhs_packed + lhs_offset;
934
+ const void * rhs_ptr = rhs_base + rhs_packed_offset;
935
+ float * dst_ptr = reinterpret_cast<float *>(dst_batch_base + dst_offset);
936
+
937
+ kernel->run_kernel_ex(m, n_to_process, k, use_gemv ? src1->nb[1] : 0,
938
+ lhs_ptr,
939
+ rhs_ptr,
940
+ dst_ptr,
941
+ dst_stride,
942
+ sizeof(float),
943
+ -FLT_MAX,
944
+ FLT_MAX);
945
+ }
946
+
947
+ current_col = ggml_threadpool_chunk_add(params->threadpool, (int) chunk_cols);
948
+ }
949
+
950
+ if (batch_idx != ne12 - 1) {
951
+ ggml_barrier(params->threadpool);
952
+ }
953
+ }
954
+
955
+ return true;
956
+ }
957
+
958
+ bool compute_forward_fp16(ggml_compute_params * params, struct ggml_tensor * dst) {
959
+ const ggml_tensor * src0 = dst->src[0];
960
+ const ggml_tensor * src1 = dst->src[1];
961
+
962
+ GGML_TENSOR_BINARY_OP_LOCALS
963
+
964
+ ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
965
+ if (!kernels) {
966
+ return false;
967
+ }
968
+
969
+ const bool is_gemv = src1->ne[1] == 1;
970
+ kernel_info * kernel = is_gemv ? &kernels->gemv : &kernels->gemm;
971
+ lhs_packing_info * lhs_info = is_gemv ? &kernels->gemv_lhs_info : &kernels->gemm_lhs_info;
972
+ GGML_ASSERT(kernel);
973
+ if (!kernels->rhs_info.pack_func_ex ||
974
+ !kernel->get_lhs_offset_ex || !kernel->get_rhs_packed_offset_ex || !kernel->run_kernel_ex) {
975
+ return false;
976
+ }
977
+
978
+ const int nth = params->nth;
979
+ const int ith = params->ith;
980
+
981
+ const int64_t lhs_batch_size0 = ne12;
982
+ const int64_t rhs_batch_size0 = ne02;
983
+ const int64_t batch_size = lhs_batch_size0;
984
+
985
+ GGML_ASSERT(rhs_batch_size0 > 0);
986
+ GGML_ASSERT(lhs_batch_size0 % rhs_batch_size0 == 0);
987
+ const int64_t r = lhs_batch_size0 / rhs_batch_size0;
988
+
989
+ const int64_t m_group = ne11;
990
+ const int64_t m = m_group;
991
+ const int64_t n = ne01;
992
+ const int64_t k = ne00;
993
+
994
+ const size_t lhs_stride = src1->nb[1];
995
+ const size_t rhs_stride = src0->nb[1];
996
+ const size_t dst_stride = dst->nb[1];
997
+
998
+ const int64_t mr = (int64_t) kernel->get_mr();
999
+ const int64_t nr = (int64_t) kernel->get_nr();
1000
+ const int64_t kr = (int64_t) kernel->get_kr();
1001
+ const int64_t sr = (int64_t) kernel->get_sr();
1002
+
1003
+ const size_t lhs_packed_size = lhs_info->packed_size_ex(m, k, 0, mr, kr, sr);
1004
+ const size_t rhs_packed_size = kernels->rhs_info.packed_size_ex(n, k, nr, kr, 0);
1005
+ const size_t kxn_size = k * n * sizeof(float);
1006
+ const size_t bias_size = n * sizeof(float);
1007
+
1008
+ const size_t wsize_required = lhs_packed_size + rhs_packed_size + kxn_size + bias_size;
1009
+ GGML_ASSERT(wsize_required <= params->wsize);
1010
+
1011
+ uint8_t * lhs_packed = static_cast<uint8_t *>(params->wdata);
1012
+ uint8_t * rhs_packed = lhs_packed + lhs_packed_size;
1013
+ uint8_t * rhs_kxn = rhs_packed + rhs_packed_size;
1014
+ uint8_t * bias = rhs_kxn + kxn_size;
1015
+
1016
+ for (int64_t batch_idx = 0; batch_idx < batch_size; ++batch_idx) {
1017
+ const int64_t rhs_batch_idx = batch_idx / r;
1018
+ const uint8_t * rhs_batch_base = static_cast<const uint8_t *>(src0->data) + rhs_batch_idx * src0->nb[2];
1019
+ uint8_t * dst_batch_base = static_cast<uint8_t *>(dst->data) + batch_idx * dst->nb[2];
1020
+
1021
+ // LHS packing (threaded over m, honoring mr alignment and KV groups)
1022
+ {
1023
+ const int64_t m_roundup_mr = kai_roundup(m, mr);
1024
+ const int64_t num_threads = KAI_MIN(m_roundup_mr / mr, nth);
1025
+
1026
+ if (ith < num_threads) {
1027
+ const int64_t num_m_per_thread0 = round_down((size_t)(m_roundup_mr / num_threads), (size_t)mr);
1028
+ const int64_t num_m_per_threadN_1 = m - (num_threads - 1) * num_m_per_thread0;
1029
+
1030
+ const int64_t m_start = ith * num_m_per_thread0;
1031
+ const int64_t m_count = (ith == num_threads - 1) ? num_m_per_threadN_1 : num_m_per_thread0;
1032
+
1033
+ // Base packed offset (aligned) and per-row stride in bytes
1034
+ const size_t base_packed_off = lhs_info->get_packed_offset_ex(m_start, k, 0, mr, kr, sr);
1035
+ const size_t next_block_off = lhs_info->get_packed_offset_ex(m_start + mr, k, 0, mr, kr, sr);
1036
+ const size_t row_stride_bytes = (next_block_off - base_packed_off) / (size_t)mr;
1037
+
1038
+ int64_t remaining = m_count;
1039
+ int64_t cur = m_start;
1040
+
1041
+ while (remaining > 0) {
1042
+ const int64_t row_in_group = cur;
1043
+ const int64_t avail = m_group - row_in_group;
1044
+ const int64_t take = std::min(avail, remaining);
1045
+
1046
+ const uint8_t * lhs_batch_base = static_cast<const uint8_t *>(src1->data) + batch_idx * src1->nb[2];
1047
+ const void * src_ptr = lhs_batch_base + (size_t)row_in_group * lhs_stride;
1048
+ const size_t dst_off = base_packed_off + (size_t)(cur - m_start) * row_stride_bytes;
1049
+ void * dst_ptr = lhs_packed + dst_off;
1050
+
1051
+ lhs_info->pack_func_ex(take, k, 0, mr, kr, sr, 0, src_ptr, lhs_stride, dst_ptr);
1052
+
1053
+ cur += take;
1054
+ remaining -= take;
1055
+ }
1056
+ }
1057
+ }
1058
+
1059
+ // RHS packing (single thread), then synchronize
1060
+ if (ith == 0) {
1061
+ memset(bias, 0, (size_t)n * sizeof(float));
1062
+ transpose_f32kxn_f16nxk((size_t)n, (size_t)k,
1063
+ reinterpret_cast<float *>(rhs_kxn),
1064
+ reinterpret_cast<const uint16_t *>(rhs_batch_base),
1065
+ rhs_stride);
1066
+
1067
+ kernels->rhs_info.pack_func_ex(1, n, k, nr, kr, sr, 0, n * sizeof(float),
1068
+ rhs_kxn, bias, nullptr, rhs_packed, 0, nullptr);
1069
+ }
1070
+
1071
+ ggml_barrier(params->threadpool);
1072
+
1073
+ // Matmul (threaded over n)
1074
+ {
1075
+ const int64_t n_step = (int64_t) kernel->get_n_step();
1076
+ int64_t num_threads_n = KAI_MIN(n / n_step, nth);
1077
+ if (num_threads_n <= 0) {
1078
+ num_threads_n = 1;
1079
+ }
1080
+
1081
+ if (ith < num_threads_n) {
1082
+ const int64_t num_n_per_thread0 = round_down((size_t)(n / num_threads_n), (size_t)n_step);
1083
+ const int64_t num_n_per_threadN_1 = n - (num_threads_n - 1) * num_n_per_thread0;
1084
+
1085
+ const int64_t n_start = ith * num_n_per_thread0;
1086
+ const int64_t n_to_process = (ith == num_threads_n - 1) ? num_n_per_threadN_1 : num_n_per_thread0;
1087
+
1088
+ // LHS packed base at row 0 (consistent with packing above)
1089
+ const size_t lhs_packed_offset0 = lhs_info->get_packed_offset_ex(0, k, 0, mr, kr, sr);
1090
+ const size_t rhs_packed_offset = kernel->get_rhs_packed_offset_ex(n_start, k, 0);
1091
+ const size_t dst_offset = kernel->get_dst_offset((size_t)0, (size_t)n_start, dst_stride);
1092
+
1093
+ const void * lhs_ptr = lhs_packed + lhs_packed_offset0;
1094
+ const void * rhs_ptr = rhs_packed + rhs_packed_offset;
1095
+ float * dst_ptr = reinterpret_cast<float *>(dst_batch_base + dst_offset);
1096
+
1097
+ kernel->run_kernel_ex(m, n_to_process, k, 0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride, sizeof(float), -FLT_MAX, FLT_MAX);
1098
+ }
1099
+ }
1100
+
1101
+ if (batch_idx != batch_size - 1) {
1102
+ ggml_barrier(params->threadpool);
1103
+ }
1104
+ }
1105
+
1106
+ return true;
1107
+ }
1108
+
1109
+ bool compute_forward_qx(struct ggml_compute_params * params, struct ggml_tensor * dst) {
1110
+ GGML_ASSERT(dst->src[0]->type == GGML_TYPE_Q4_0 || dst->src[0]->type == GGML_TYPE_Q8_0);
1111
+
1112
+ const ggml_tensor * src0 = dst->src[0];
1113
+ const ggml_tensor * src1 = dst->src[1];
1114
+
1115
+ GGML_TENSOR_BINARY_OP_LOCALS
1116
+
1117
+ const kleidiai_weight_header * header = kleidiai_weight_header_from_ptr(src0->data);
1118
+ const bool has_header = kleidiai_is_weight_header_valid(header);
1119
+ const bool is_gemv = src1->ne[1] == 1;
1120
+ std::array<ggml_kleidiai_kernels *, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> kernel_chain;
1121
+ const int slot_total = kleidiai_collect_kernel_chain(dst, kernel_chain);
1122
+
1123
+ auto weight_for_slot = [&](int slot_index, size_t & size_out) -> const uint8_t * {
1124
+ if (slot_index < 0 || slot_index >= slot_total) {
1125
+ return nullptr;
1126
+ }
1127
+ if (has_header) {
1128
+ if (slot_index < header->slot_count) {
1129
+ size_out = static_cast<size_t>(header->sizes[slot_index]);
1130
+ return kleidiai_weight_slot_ptr(header, slot_index);
1131
+ }
1132
+ return nullptr;
1133
+ }
1134
+ if (slot_index == 0) {
1135
+ size_out = ggml_nbytes(src0);
1136
+ return static_cast<const uint8_t *>(src0->data);
1137
+ }
1138
+ return nullptr;
1139
+ };
1140
+
1141
+ struct runtime_slot {
1142
+ int slot_index;
1143
+ ggml_kleidiai_kernels * kernels;
1144
+ kernel_info * kernel;
1145
+ lhs_packing_info * lhs_info;
1146
+ size_t mr;
1147
+ size_t nr;
1148
+ size_t kr;
1149
+ size_t sr;
1150
+ size_t n_step;
1151
+ size_t lhs_packed_size;
1152
+ size_t lhs_offset;
1153
+ size_t lhs_bl;
1154
+ size_t rhs_bl;
1155
+ size_t pack_bl;
1156
+ size_t lhs_packed_offset0;
1157
+ int assigned_threads;
1158
+ int thread_begin;
1159
+ int thread_end;
1160
+ const uint8_t * rhs_base;
1161
+ };
1162
+
1163
+ std::array<runtime_slot, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> runtime{};
1164
+ int runtime_count = 0;
1165
+
1166
+ for (int slot = 0; slot < slot_total && runtime_count < GGML_KLEIDIAI_MAX_KERNEL_SLOTS; ++slot) {
1167
+ ggml_kleidiai_kernels * kernels = kernel_chain[slot];
1168
+ kernel_info * kinfo = is_gemv ? &kernels->gemv : &kernels->gemm;
1169
+ lhs_packing_info * linfo = is_gemv ? &kernels->gemv_lhs_info : &kernels->gemm_lhs_info;
1170
+ if (!kinfo || !linfo || !linfo->packed_size_ex || !linfo->pack_func_ex || !linfo->get_offset ||
1171
+ !kinfo->get_rhs_packed_offset_ex || !kinfo->run_kernel_ex || !kinfo->get_dst_offset) {
1172
+ continue;
1173
+ }
1174
+
1175
+ size_t rhs_size = 0;
1176
+ const uint8_t * rhs_ptr = weight_for_slot(slot, rhs_size);
1177
+ if (!rhs_ptr || rhs_size == 0) {
1178
+ continue;
1179
+ }
1180
+
1181
+ const kleidiai_block_args block_args = kleidiai_get_block_args(kernels->rhs_type);
1182
+
1183
+ runtime[runtime_count] = {
1184
+ slot,
1185
+ kernels,
1186
+ kinfo,
1187
+ linfo,
1188
+ kinfo->get_mr(),
1189
+ kinfo->get_nr(),
1190
+ kinfo->get_kr(),
1191
+ kinfo->get_sr(),
1192
+ kinfo->get_n_step(),
1193
+ 0,
1194
+ 0,
1195
+ block_args.lhs_bl,
1196
+ block_args.rhs_bl,
1197
+ block_args.pack_bl,
1198
+ 0,
1199
+ 0,
1200
+ 0,
1201
+ 0,
1202
+ rhs_ptr
1203
+ };
1204
+ ++runtime_count;
1205
+ }
1206
+
1207
+ if (runtime_count == 0) {
1208
+ GGML_LOG_WARN("kleidiai: no runtime kernel slot available for supported op %s\n", dst->name);
1209
+ return false;
1210
+ }
1211
+
1212
+ const int nth_total = params->nth > 0 ? params->nth : 1;
1213
+ const int ith_total = params->ith;
1214
+
1215
+ int sme_slot = -1;
1216
+ int non_sme_slot = -1;
1217
+ for (int i = 0; i < runtime_count; ++i) {
1218
+ if (is_sme_family(runtime[i].kernels->required_cpu)) {
1219
+ sme_slot = i;
1220
+ break;
1221
+ }
1222
+ }
1223
+
1224
+ for (int i = 0; i < runtime_count; ++i) {
1225
+ if (!is_sme_family(runtime[i].kernels->required_cpu)) {
1226
+ non_sme_slot = i;
1227
+ break;
1228
+ }
1229
+ }
1230
+
1231
+ const int sme_cap_limit = ctx.sme_thread_cap;
1232
+ const bool use_hybrid = sme_cap_limit > 0 &&
1233
+ runtime_count > 1 &&
1234
+ nth_total > sme_cap_limit;
1235
+ // Heuristic: disable hybrid for very small workloads where per-slot overhead dominates.
1236
+ // If rows are small or average columns per thread are small, keep single-slot.
1237
+ size_t min_cols_per_thread = 0;
1238
+ if (runtime_count > 0 && nth_total > 0) {
1239
+ min_cols_per_thread = (size_t) std::max<int64_t>(1, (int64_t)ne01 / (int64_t)nth_total);
1240
+ }
1241
+ const bool too_small_for_hybrid = (min_cols_per_thread < 2) || (ne11 < 128);
1242
+
1243
+ const bool hybrid_enabled = use_hybrid && !too_small_for_hybrid;
1244
+
1245
+ if (!hybrid_enabled) {
1246
+ int chosen_slot = 0;
1247
+ if (too_small_for_hybrid && sme_slot != -1) {
1248
+ chosen_slot = nth_total > sme_cap_limit && non_sme_slot != -1 ? non_sme_slot : sme_slot;
1249
+ } else if (runtime_count > 1 && ctx.sme_thread_cap > 0 && nth_total > ctx.sme_thread_cap) {
1250
+ chosen_slot = 1;
1251
+ }
1252
+ if (chosen_slot != 0 && chosen_slot < runtime_count) {
1253
+ runtime[0] = runtime[chosen_slot];
1254
+ runtime[0].assigned_threads = 0;
1255
+ runtime[0].thread_begin = 0;
1256
+ runtime[0].thread_end = 0;
1257
+ }
1258
+ runtime_count = runtime_count > 0 ? 1 : 0;
1259
+
1260
+ // Recompute SME slot based on the collapsed runtime[0]
1261
+ sme_slot = -1;
1262
+ if (runtime_count > 0 &&
1263
+ is_sme_family(runtime[0].kernels->required_cpu)) {
1264
+ sme_slot = 0;
1265
+ }
1266
+ }
1267
+
1268
+ int sme_cap = kleidiai_sme_thread_cap();
1269
+ if (sme_cap < 0) {
1270
+ sme_cap = nth_total;
1271
+ }
1272
+ sme_cap = std::min(sme_cap, nth_total);
1273
+
1274
+ int threads_remaining = nth_total;
1275
+ if (sme_slot != -1) {
1276
+ int sme_threads = std::min(std::max(sme_cap, 0), threads_remaining);
1277
+ runtime[sme_slot].assigned_threads = sme_threads;
1278
+ threads_remaining -= sme_threads;
1279
+ }
1280
+
1281
+ int fallback_indices[GGML_KLEIDIAI_MAX_KERNEL_SLOTS];
1282
+ int fallback_count = 0;
1283
+ // The current hybrid chain is bounded to SME + one non-SME fallback slot.
1284
+ GGML_ASSERT(GGML_KLEIDIAI_MAX_KERNEL_SLOTS == 2);
1285
+ for (int i = 0; i < runtime_count; ++i) {
1286
+ if (i == sme_slot) {
1287
+ continue;
1288
+ }
1289
+ fallback_indices[fallback_count++] = i;
1290
+ }
1291
+
1292
+ for (int fi = 0; fi < fallback_count; ++fi) {
1293
+ if (threads_remaining <= 0) {
1294
+ break;
1295
+ }
1296
+ const int slot_index = fallback_indices[fi];
1297
+ const int slots_left = fallback_count - fi;
1298
+ int share = (threads_remaining + slots_left - 1) / slots_left;
1299
+ share = std::min(share, threads_remaining);
1300
+ runtime[slot_index].assigned_threads = share;
1301
+ threads_remaining -= share;
1302
+ }
1303
+
1304
+ if (threads_remaining > 0) {
1305
+ const int fallback_slot = (sme_slot != -1) ? sme_slot : 0;
1306
+ runtime[fallback_slot].assigned_threads += threads_remaining;
1307
+ threads_remaining = 0;
1308
+ }
1309
+
1310
+ int thread_cursor = 0;
1311
+ for (int i = 0; i < runtime_count; ++i) {
1312
+ runtime[i].thread_begin = thread_cursor;
1313
+ thread_cursor += runtime[i].assigned_threads;
1314
+ runtime[i].thread_end = thread_cursor;
1315
+ }
1316
+
1317
+ if (thread_cursor < nth_total && runtime_count > 0) {
1318
+ runtime[runtime_count - 1].assigned_threads += nth_total - thread_cursor;
1319
+ runtime[runtime_count - 1].thread_end = nth_total;
1320
+ }
1321
+
1322
+ int local_slot = -1;
1323
+ int local_ith = 0;
1324
+ for (int i = 0; i < runtime_count; ++i) {
1325
+ if (ith_total >= runtime[i].thread_begin && ith_total < runtime[i].thread_end) {
1326
+ local_slot = i;
1327
+ local_ith = ith_total - runtime[i].thread_begin;
1328
+ break;
1329
+ }
1330
+ }
1331
+ if (local_slot == -1) {
1332
+ return false;
1333
+ }
1334
+
1335
+ const size_t k = ne00;
1336
+ const size_t m = ne11;
1337
+ const size_t n = ne01;
1338
+
1339
+ size_t cursor = 0;
1340
+ for (int i = 0; i < runtime_count; ++i) {
1341
+ runtime[i].lhs_packed_size = runtime[i].lhs_info->packed_size_ex(m, k, runtime[i].pack_bl, runtime[i].mr, runtime[i].kr, runtime[i].sr);
1342
+ cursor = align_up(cursor, GGML_KLEIDIAI_PACK_ALIGN);
1343
+ runtime[i].lhs_offset = cursor;
1344
+ runtime[i].lhs_packed_offset0 = runtime[i].lhs_info->get_packed_offset_ex(0, k, runtime[i].lhs_bl, runtime[i].mr, runtime[i].kr, runtime[i].sr);
1345
+ cursor += runtime[i].lhs_packed_size;
1346
+ }
1347
+
1348
+ GGML_ASSERT(cursor <= params->wsize);
1349
+ uint8_t * scratch = static_cast<uint8_t *>(params->wdata);
1350
+
1351
+ size_t common_step = 1;
1352
+ for (int i = 0; i < runtime_count; ++i) {
1353
+ if (runtime[i].assigned_threads == 0) {
1354
+ continue;
1355
+ }
1356
+ size_t next_step = 0;
1357
+ if (!lcm_size(common_step, runtime[i].n_step ? runtime[i].n_step : 1, next_step)) {
1358
+ return false;
1359
+ }
1360
+ common_step = next_step;
1361
+ }
1362
+ GGML_ASSERT(common_step > 0);
1363
+
1364
+ const bool disable_chunking = ggml_is_numa();
1365
+ const size_t chunk_multiplier = std::max(1, ctx.chunk_multiplier);
1366
+ const size_t chunk_divisor = (nth_total == 1 || disable_chunking) ? (size_t)nth_total : (size_t)nth_total * chunk_multiplier;
1367
+ size_t chunk_cols = align_up(std::max<size_t>(1, ceil_div_size(n, chunk_divisor)), common_step);
1368
+ if (chunk_cols == 0) {
1369
+ chunk_cols = common_step;
1370
+ }
1371
+ // If common_step is larger than n, the loop below runs one valid tail chunk
1372
+ // with cols == n.
1373
+ const size_t nchunk_size = std::max<size_t>(1, ceil_div_size(n, chunk_cols));
1374
+ GGML_ASSERT(nchunk_size <= (size_t)INT_MAX);
1375
+ const int nchunk = (int)nchunk_size;
1376
+ const size_t dst_stride = dst->nb[1];
1377
+
1378
+ auto run_chunk = [&](runtime_slot & slot, size_t global_start, size_t cols, uint8_t * dst_batch_base) {
1379
+ const size_t rhs_packed_offset = slot.kernel->get_rhs_packed_offset_ex(global_start, k, slot.rhs_bl);
1380
+ const size_t dst_offset = slot.kernel->get_dst_offset(0, global_start, dst_stride);
1381
+
1382
+ const uint8_t * lhs_ptr = scratch + slot.lhs_offset + slot.lhs_packed_offset0;
1383
+ const uint8_t * rhs_ptr = slot.rhs_base + rhs_packed_offset;
1384
+ float * dst_ptr = reinterpret_cast<float *>(dst_batch_base + dst_offset);
1385
+
1386
+ slot.kernel->run_kernel_ex(m, cols, k, slot.rhs_bl,
1387
+ lhs_ptr,
1388
+ rhs_ptr,
1389
+ dst_ptr,
1390
+ dst_stride,
1391
+ sizeof(float),
1392
+ -FLT_MAX,
1393
+ FLT_MAX);
1394
+ };
1395
+
1396
+ for (int64_t batch_idx = 0; batch_idx < ne12; ++batch_idx) {
1397
+ const uint8_t * lhs_batch_base = static_cast<const uint8_t *>(src1->data) + batch_idx * src1->nb[2];
1398
+ uint8_t * dst_batch_base = static_cast<uint8_t *>(dst->data) + batch_idx * dst->nb[2];
1399
+
1400
+ if (runtime[local_slot].assigned_threads > 0) {
1401
+ runtime_slot & slot = runtime[local_slot];
1402
+ const int64_t m_roundup_mr = kai_roundup((int64_t)m, (int64_t)slot.mr);
1403
+ int64_t max_threads = slot.mr ? (m_roundup_mr / (int64_t)slot.mr) : slot.assigned_threads;
1404
+ max_threads = std::max<int64_t>(1, max_threads);
1405
+ const int64_t use_threads = std::min<int64_t>(slot.assigned_threads, max_threads);
1406
+
1407
+ if (local_ith < use_threads) {
1408
+ const int64_t num_m_per_thread0 = round_down((size_t)(m_roundup_mr / use_threads), slot.mr);
1409
+ const int64_t num_m_per_threadN_1 = (int64_t)m - (use_threads - 1) * num_m_per_thread0;
1410
+
1411
+ const int64_t m_start = (int64_t)local_ith * num_m_per_thread0;
1412
+ const int64_t m_count = (local_ith == use_threads - 1) ? num_m_per_threadN_1 : num_m_per_thread0;
1413
+
1414
+ const size_t base_packed_off = slot.lhs_info->get_packed_offset_ex(m_start, k, slot.lhs_bl, slot.mr, slot.kr, slot.sr);
1415
+ const size_t next_block_off = slot.lhs_info->get_packed_offset_ex(m_start + slot.mr, k, slot.lhs_bl, slot.mr, slot.kr, slot.sr);
1416
+ const size_t row_stride_bytes = slot.mr ? (next_block_off - base_packed_off) / slot.mr : 0;
1417
+
1418
+ int64_t remaining = m_count;
1419
+ int64_t cur = m_start;
1420
+
1421
+ uint8_t * lhs_packed = scratch + slot.lhs_offset;
1422
+ while (remaining > 0) {
1423
+ const int64_t row_in_group = cur;
1424
+ const int64_t avail = (int64_t)m - row_in_group;
1425
+ const int64_t take = std::min(avail, remaining);
1426
+
1427
+ const size_t src_off = slot.lhs_info->get_offset(row_in_group, src1->nb[1]);
1428
+ const void * src_ptr = lhs_batch_base + src_off;
1429
+ const size_t dst_off = base_packed_off + (size_t)(cur - m_start) * row_stride_bytes;
1430
+ void * dst_ptr = lhs_packed + dst_off;
1431
+
1432
+ slot.lhs_info->pack_func_ex(take, k, slot.lhs_bl, slot.mr, slot.kr, slot.sr, 0, src_ptr, src1->nb[1], dst_ptr);
1433
+
1434
+ cur += take;
1435
+ remaining -= take;
1436
+ }
1437
+ }
1438
+ }
1439
+
1440
+ if (ith_total == 0) {
1441
+ ggml_threadpool_chunk_set(params->threadpool, nth_total);
1442
+ }
1443
+
1444
+ // Publishes both LHS packing and the initialized dynamic chunk queue.
1445
+ ggml_barrier(params->threadpool);
1446
+
1447
+ runtime_slot & slot = runtime[local_slot];
1448
+ int current_chunk = ith_total;
1449
+ while (current_chunk < nchunk) {
1450
+ const size_t global_start = (size_t)current_chunk * chunk_cols;
1451
+ if (global_start >= n) {
1452
+ break;
1453
+ }
1454
+
1455
+ const size_t cols = std::min(chunk_cols, n - global_start);
1456
+ if (cols > 0) {
1457
+ // KleidiAI GEMM/GEMV kernels accept arbitrary final tail widths;
1458
+ // only non-tail chunks are guaranteed to be n_step-aligned.
1459
+ run_chunk(slot, global_start, cols, dst_batch_base);
1460
+ }
1461
+
1462
+ current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
1463
+ }
1464
+
1465
+ if (batch_idx != ne12 - 1) {
1466
+ ggml_barrier(params->threadpool);
1467
+ }
1468
+ }
1469
+
1470
+ return true;
1471
+ }
1472
+
1473
+ bool compute_forward_get_rows(struct ggml_compute_params * params, struct ggml_tensor * dst) {
1474
+ GGML_ASSERT(dst->src[0]->type == GGML_TYPE_Q4_0 || dst->src[0]->type == GGML_TYPE_Q8_0);
1475
+ const ggml_tensor * src0 = dst->src[0];
1476
+ const ggml_tensor * src1 = dst->src[1];
1477
+
1478
+ GGML_TENSOR_BINARY_OP_LOCALS
1479
+
1480
+ const kleidiai_weight_header * header = kleidiai_weight_header_from_ptr(src0->data);
1481
+ const bool has_header = kleidiai_is_weight_header_valid(header);
1482
+
1483
+ std::array<ggml_kleidiai_kernels *, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> kernel_chain;
1484
+ const bool want_q8 = src0->type == GGML_TYPE_Q8_0;
1485
+ const int chain_count = want_q8 ? kleidiai_collect_q8_chain(kernel_chain)
1486
+ : kleidiai_collect_q4_chain(kernel_chain);
1487
+
1488
+ ggml_kleidiai_kernels * kernels = nullptr;
1489
+ const uint8_t * packed_base = static_cast<const uint8_t *>(src0->data);
1490
+
1491
+ if (has_header && chain_count > 0) {
1492
+ int select_slot = 0;
1493
+ if (select_slot >= header->slot_count) {
1494
+ select_slot = header->slot_count - 1;
1495
+ }
1496
+ if (select_slot >= 0 && select_slot < chain_count) {
1497
+ kernels = kernel_chain[select_slot];
1498
+ const uint8_t * slot_ptr = kleidiai_weight_slot_ptr(header, select_slot);
1499
+ if (slot_ptr) {
1500
+ packed_base = slot_ptr;
1501
+ }
1502
+ }
1503
+ }
1504
+
1505
+ if (!kernels && chain_count > 0) {
1506
+ kernels = kernel_chain[0];
1507
+ if (has_header) {
1508
+ const uint8_t * slot_ptr = kleidiai_weight_slot_ptr(header, 0);
1509
+ if (slot_ptr) {
1510
+ packed_base = slot_ptr;
1511
+ }
1512
+ }
1513
+ }
1514
+
1515
+ if (!kernels) {
1516
+ return false;
1517
+ }
1518
+
1519
+ rhs_packing_info * rhs_info = &kernels->rhs_info;
1520
+ kernel_info * kernel = &kernels->gemm;
1521
+ if (!rhs_info->to_float || !kernel->get_nr) {
1522
+ return false;
1523
+ }
1524
+
1525
+ const int64_t nc = ne00;
1526
+ const int64_t nr = ggml_nelements(src1);
1527
+
1528
+ const ggml_type rhs_type = kernels->rhs_type;
1529
+ size_t block_len = 0;
1530
+ size_t num_bytes_multiplier = 0;
1531
+ if (rhs_type == GGML_TYPE_Q4_0) {
1532
+ block_len = QK4_0;
1533
+ num_bytes_multiplier = sizeof(uint16_t);
1534
+ } else if (rhs_type == GGML_TYPE_Q8_0) {
1535
+ block_len = QK8_0;
1536
+ num_bytes_multiplier = sizeof(float);
1537
+ } else {
1538
+ return false;
1539
+ }
1540
+
1541
+ const size_t block_rows = kernel->get_nr();
1542
+ const size_t kr = kernel->get_kr();
1543
+
1544
+ const size_t packed_stride = rhs_info->packed_stride(nc, block_rows, kr, block_len);
1545
+
1546
+ const int ith = params->ith;
1547
+ const int nth = params->nth;
1548
+
1549
+ const int dr = (nr + nth - 1) / nth;
1550
+ const int ir0 = dr * ith;
1551
+ const int ir1 = MIN(ir0 + dr, nr);
1552
+
1553
+ for (int64_t i = ir0; i < ir1; ++i) {
1554
+ GGML_ASSERT(src1->type == GGML_TYPE_I32);
1555
+ int64_t row_idx = ((const int32_t *)src1->data)[i];
1556
+ GGML_ASSERT(row_idx >= 0 && row_idx < src0->ne[1]);
1557
+
1558
+ float *out = (float *)((char *)dst->data + i * nb1);
1559
+ rhs_info->to_float(packed_base, row_idx, nc, out, block_rows, packed_stride, kr, block_len, num_bytes_multiplier);
1560
+ }
1561
+
1562
+ return true;
1563
+ }
1564
+
1565
+ public:
1566
+ int repack(struct ggml_tensor * tensor, const void * data, size_t data_size) {
1567
+ GGML_ASSERT(tensor->type == GGML_TYPE_Q4_0 || tensor->type == GGML_TYPE_Q8_0 || tensor->type == GGML_TYPE_F32);
1568
+ const size_t n = tensor->ne[1];
1569
+ const size_t k = tensor->ne[0];
1570
+
1571
+ kleidiai_weight_header * header = kleidiai_weight_header_from_ptr(tensor->data);
1572
+ if (!header) {
1573
+ return -1;
1574
+ }
1575
+
1576
+ header->magic = GGML_KLEIDIAI_PACK_MAGIC;
1577
+ header->version = GGML_KLEIDIAI_PACK_VERSION;
1578
+ header->slot_count = 0;
1579
+
1580
+ uint8_t * base_ptr = static_cast<uint8_t *>(tensor->data);
1581
+ size_t cursor = sizeof(kleidiai_weight_header);
1582
+ cursor = align_up(cursor, GGML_KLEIDIAI_PACK_ALIGN);
1583
+
1584
+ std::array<ggml_kleidiai_kernels *, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> kernel_chain;
1585
+ const bool want_q8 = tensor->type == GGML_TYPE_Q8_0;
1586
+ const bool want_f32 = tensor->type == GGML_TYPE_F32;
1587
+ const int slot_total = want_f32 ? kleidiai_collect_f32_chain(kernel_chain)
1588
+ : want_q8 ? kleidiai_collect_q8_chain(kernel_chain)
1589
+ : kleidiai_collect_q4_chain(kernel_chain);
1590
+ const bool allow_fallback = kleidiai_pack_fallback_allowed();
1591
+
1592
+ std::vector<int8_t> qdata;
1593
+ std::vector<float> scales;
1594
+ std::vector<float> bias;
1595
+
1596
+ if (want_q8 && slot_total > 0) {
1597
+ qdata.resize(n * k, 0);
1598
+ scales.resize(n, 0.0f);
1599
+
1600
+ const size_t row_stride = tensor->nb[1];
1601
+ const size_t k_blocks = (k + QK8_0 - 1) / QK8_0;
1602
+
1603
+ for (size_t row = 0; row < n; ++row) {
1604
+ const auto * row_blocks = reinterpret_cast<const block_q8_0 *>(
1605
+ static_cast<const uint8_t *>(data) + row * row_stride);
1606
+
1607
+ float max_abs = 0.0f;
1608
+ for (size_t block = 0; block < k_blocks; ++block) {
1609
+ const block_q8_0 & blk = row_blocks[block];
1610
+ const float d = GGML_FP16_TO_FP32(blk.d);
1611
+ for (size_t l = 0; l < QK8_0; ++l) {
1612
+ const size_t linear_idx = block * QK8_0 + l;
1613
+ if (linear_idx >= k) {
1614
+ break;
1615
+ }
1616
+ const float value = d * static_cast<float>(blk.qs[l]);
1617
+ max_abs = std::max(max_abs, std::fabs(value));
1618
+ }
1619
+ }
1620
+
1621
+ float scale = max_abs > 0.0f ? max_abs / 127.0f : 0.0f;
1622
+ scales[row] = scale;
1623
+ const float inv_scale = scale > 0.0f ? 1.0f / scale : 0.0f;
1624
+
1625
+ for (size_t block = 0; block < k_blocks; ++block) {
1626
+ const block_q8_0 & blk = row_blocks[block];
1627
+ const float d = GGML_FP16_TO_FP32(blk.d);
1628
+ for (size_t l = 0; l < QK8_0; ++l) {
1629
+ const size_t linear_idx = block * QK8_0 + l;
1630
+ if (linear_idx >= k) {
1631
+ break;
1632
+ }
1633
+ const float value = d * static_cast<float>(blk.qs[l]);
1634
+ int32_t q = scale > 0.0f ? static_cast<int32_t>(std::lround(value * inv_scale)) : 0;
1635
+ q = std::clamp(q, -127, 127);
1636
+ qdata[row * k + linear_idx] = static_cast<int8_t>(q);
1637
+ }
1638
+ }
1639
+ }
1640
+ }
1641
+
1642
+ if (want_f32 && slot_total > 0) {
1643
+ bias.resize(n, 0.0f);
1644
+ }
1645
+
1646
+ for (int slot = 0; slot < slot_total && slot < GGML_KLEIDIAI_MAX_KERNEL_SLOTS; ++slot) {
1647
+ if (!allow_fallback && slot > 0) {
1648
+ break;
1649
+ }
1650
+ ggml_kleidiai_kernels * kernels = kernel_chain[slot];
1651
+ kernel_info * kernel = &kernels->gemm;
1652
+ rhs_packing_info * rhs_info = &kernels->rhs_info;
1653
+ if (!rhs_info || !rhs_info->pack_func_ex || !rhs_info->packed_size_ex || !kernel) {
1654
+ continue;
1655
+ }
1656
+
1657
+ const size_t nr = kernel->get_nr();
1658
+ const size_t kr = kernel->get_kr();
1659
+ const size_t sr = kernel->get_sr();
1660
+ const ggml_type rhs_type = kernels->rhs_type;
1661
+ const size_t block_len = rhs_type == GGML_TYPE_Q8_0 ? QK8_0 :
1662
+ rhs_type == GGML_TYPE_Q4_0 ? QK4_0 :
1663
+ rhs_type == GGML_TYPE_F32 ? 0 : SIZE_MAX;
1664
+ if (block_len == SIZE_MAX) {
1665
+ continue;
1666
+ }
1667
+
1668
+ const size_t packed_size = rhs_info->packed_size_ex(n, k, nr, kr, block_len);
1669
+ const size_t aligned_cursor = align_up(cursor, GGML_KLEIDIAI_PACK_ALIGN);
1670
+
1671
+ uint8_t * dst_ptr = base_ptr + aligned_cursor;
1672
+
1673
+ if (rhs_type == GGML_TYPE_Q4_0) {
1674
+ struct kai_rhs_pack_qs4cxs1s0_param params;
1675
+ params.lhs_zero_point = 1;
1676
+ params.rhs_zero_point = 8;
1677
+ rhs_info->pack_func_ex(1, n, k, nr, kr, sr, QK4_0, 0,
1678
+ static_cast<const uint8_t *>(data), nullptr, nullptr,
1679
+ dst_ptr, 0, &params);
1680
+ } else if (rhs_type == GGML_TYPE_Q8_0) {
1681
+ struct kai_rhs_pack_qsi8cx_params params;
1682
+ params.lhs_zero_point = 1;
1683
+ params.scale_multiplier = 1.0f;
1684
+ rhs_info->pack_func_ex(1, n, k, nr, kr, sr, 0, 0,
1685
+ qdata.data(), nullptr, scales.data(),
1686
+ dst_ptr, 0, &params);
1687
+ } else if (rhs_type == GGML_TYPE_F32) {
1688
+ rhs_info->pack_func_ex(1, n, k, nr, kr, sr, 0, tensor->nb[1],
1689
+ data, bias.data(), nullptr,
1690
+ dst_ptr, 0, nullptr);
1691
+ } else {
1692
+ continue;
1693
+ }
1694
+
1695
+ header->offsets[header->slot_count] = aligned_cursor;
1696
+ header->sizes[header->slot_count] = packed_size;
1697
+ ++header->slot_count;
1698
+
1699
+ cursor = aligned_cursor + packed_size;
1700
+ }
1701
+
1702
+ if (header->slot_count == 0) {
1703
+ header->magic = 0;
1704
+ header->version = 0;
1705
+ memcpy(tensor->data, data, data_size);
1706
+ }
1707
+
1708
+ return 0;
1709
+ }
1710
+ };
1711
+
1712
+ static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struct ggml_tensor *) {
1713
+ static tensor_traits traits;
1714
+ return &traits;
1715
+ }
1716
+ } // namespace ggml::cpu::kleidiai
1717
+
1718
+ static enum ggml_status ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
1719
+ tensor->extra = (void *) ggml::cpu::kleidiai::get_tensor_traits(buffer, tensor);
1720
+
1721
+ return GGML_STATUS_SUCCESS;
1722
+ GGML_UNUSED(buffer);
1723
+ }
1724
+
1725
+ static void ggml_backend_cpu_kleidiai_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
1726
+ const void * data, size_t offset, size_t size) {
1727
+ GGML_ASSERT(offset == 0);
1728
+ GGML_ASSERT(size == ggml_nbytes(tensor));
1729
+
1730
+ auto tensor_traits = (ggml::cpu::kleidiai::tensor_traits *) tensor->extra;
1731
+ auto OK = tensor_traits->repack(tensor, data, size);
1732
+
1733
+ GGML_ASSERT(OK == 0);
1734
+ GGML_UNUSED(buffer);
1735
+ }
1736
+
1737
+ static const char * ggml_backend_cpu_kleidiai_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
1738
+ GGML_UNUSED(buft);
1739
+ return "CPU_KLEIDIAI";
1740
+ }
1741
+
1742
+ static ggml_backend_buffer_t ggml_backend_cpu_kleidiai_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
1743
+ ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
1744
+
1745
+ if (buffer == nullptr) {
1746
+ return nullptr;
1747
+ }
1748
+
1749
+ buffer->buft = buft;
1750
+ buffer->iface.init_tensor = ggml_backend_cpu_kleidiai_buffer_init_tensor;
1751
+ buffer->iface.set_tensor = ggml_backend_cpu_kleidiai_buffer_set_tensor;
1752
+ buffer->iface.get_tensor = nullptr;
1753
+ buffer->iface.cpy_tensor = nullptr;
1754
+ return buffer;
1755
+ }
1756
+
1757
+ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
1758
+ GGML_UNUSED(buft);
1759
+ return TENSOR_ALIGNMENT;
1760
+ }
1761
+
1762
+ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor) {
1763
+ GGML_UNUSED(buft);
1764
+
1765
+ if (tensor->type != GGML_TYPE_Q4_0 && tensor->type != GGML_TYPE_Q8_0 && tensor->type != GGML_TYPE_F32) {
1766
+ return ggml_nbytes(tensor);
1767
+ }
1768
+
1769
+ const size_t n = tensor->ne[1];
1770
+ const size_t k = tensor->ne[0];
1771
+
1772
+ size_t cursor = sizeof(kleidiai_weight_header);
1773
+ cursor = align_up(cursor, GGML_KLEIDIAI_PACK_ALIGN);
1774
+
1775
+ std::array<ggml_kleidiai_kernels *, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> kernel_chain;
1776
+ const bool want_q8 = tensor->type == GGML_TYPE_Q8_0;
1777
+ const bool want_f32 = tensor->type == GGML_TYPE_F32;
1778
+ const int slot_total = want_f32 ? kleidiai_collect_f32_chain(kernel_chain)
1779
+ : want_q8 ? kleidiai_collect_q8_chain(kernel_chain)
1780
+ : kleidiai_collect_q4_chain(kernel_chain);
1781
+ const bool allow_fallback = kleidiai_pack_fallback_allowed();
1782
+
1783
+ size_t slot_count = 0;
1784
+ for (int slot = 0; slot < slot_total; ++slot) {
1785
+ if (!allow_fallback && slot > 0) {
1786
+ break;
1787
+ }
1788
+ ggml_kleidiai_kernels * kernels = kernel_chain[slot];
1789
+ if (!kernels) {
1790
+ continue;
1791
+ }
1792
+ kernel_info * kernel = &kernels->gemm;
1793
+ rhs_packing_info * rhs_info = &kernels->rhs_info;
1794
+ if (!kernel || !rhs_info || !rhs_info->packed_size_ex) {
1795
+ continue;
1796
+ }
1797
+
1798
+ const ggml_type rhs_type = kernels->rhs_type;
1799
+ const size_t block_len = rhs_type == GGML_TYPE_Q4_0 ? QK4_0 :
1800
+ rhs_type == GGML_TYPE_Q8_0 ? QK8_0 :
1801
+ rhs_type == GGML_TYPE_F32 ? 0 : SIZE_MAX;
1802
+ if (block_len == SIZE_MAX) {
1803
+ continue;
1804
+ }
1805
+
1806
+ cursor = align_up(cursor, GGML_KLEIDIAI_PACK_ALIGN);
1807
+ cursor += rhs_info->packed_size_ex(n, k, kernel->get_nr(), kernel->get_kr(), block_len);
1808
+ ++slot_count;
1809
+ }
1810
+
1811
+ if (slot_count == 0) {
1812
+ return ggml_nbytes(tensor);
1813
+ }
1814
+
1815
+ return std::max(cursor, ggml_nbytes(tensor));
1816
+ }
1817
+
1818
+ namespace ggml::cpu::kleidiai {
1819
+ class extra_buffer_type : ggml::cpu::extra_buffer_type {
1820
+ bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
1821
+ std::array<ggml_kleidiai_kernels *, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> kernel_chain;
1822
+ const int slot_total = kleidiai_collect_kernel_chain(op, kernel_chain);
1823
+ const bool src0_is_kleidiai =
1824
+ op->src[0]->buffer &&
1825
+ (ggml_n_dims(op->src[0]) == 2) &&
1826
+ op->src[0]->buffer->buft->context == this &&
1827
+ slot_total > 0;
1828
+
1829
+ if ((op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) &&
1830
+ (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_Q8_0 || op->src[0]->type == GGML_TYPE_F32) &&
1831
+ src0_is_kleidiai) {
1832
+ if (op->src[0]->type == GGML_TYPE_Q4_0 && ctx.kernels_q4 == nullptr) {
1833
+ return false;
1834
+ }
1835
+ if (op->src[0]->type == GGML_TYPE_Q8_0 && ctx.kernels_q8 == nullptr) {
1836
+ return false;
1837
+ }
1838
+ if (op->src[0]->type == GGML_TYPE_F32 && ctx.kernels_f32 == nullptr) {
1839
+ return false;
1840
+ }
1841
+ if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
1842
+ return false;
1843
+ }
1844
+
1845
+ if (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_Q8_0) {
1846
+ if ((op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_I32) &&
1847
+ ggml_ne(op->src[1], 3) == 1) {
1848
+ return true;
1849
+ }
1850
+ return false;
1851
+ }
1852
+
1853
+ if (op->op != GGML_OP_MUL_MAT || op->src[1]->type != GGML_TYPE_F32 || op->type != GGML_TYPE_F32) {
1854
+ return false;
1855
+ }
1856
+
1857
+ return true;
1858
+ }
1859
+
1860
+ return false;
1861
+ }
1862
+
1863
+ ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
1864
+ if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) {
1865
+ if (op->src[0]->buffer && op->src[0]->buffer->buft->context == this) {
1866
+ return (ggml::cpu::tensor_traits *) op->src[0]->extra;
1867
+ } else {
1868
+ // KleidiAI only has kernels for Q4_0 and Q8_0. For a quantized weight of any
1869
+ // other type (K-quants, IQ) it declines the op and returns nullptr below, so
1870
+ // KleidiAI does not accelerate it. Another CPU backend may still take the op,
1871
+ // and this can run during graph planning, so the message says what KleidiAI
1872
+ // did rather than what ends up executing. Warn once per process.
1873
+ if (ggml_is_quantized(op->src[0]->type) &&
1874
+ op->src[0]->type != GGML_TYPE_Q4_0 && op->src[0]->type != GGML_TYPE_Q8_0) {
1875
+ static std::atomic<bool> warned(false);
1876
+ if (!warned.exchange(true)) {
1877
+ GGML_LOG_WARN("kleidiai: no kernel for tensor type %s, not accelerated by KleidiAI "
1878
+ "(kernels available for Q4_0 and Q8_0)\n",
1879
+ ggml_type_name(op->src[0]->type));
1880
+ }
1881
+ }
1882
+ if (op->src[0]->type != GGML_TYPE_F16) {
1883
+ return nullptr;
1884
+ }
1885
+ std::array<ggml_kleidiai_kernels *, GGML_KLEIDIAI_MAX_KERNEL_SLOTS> kernel_chain;
1886
+ const int slot_total = kleidiai_collect_kernel_chain(op, kernel_chain);
1887
+ if (slot_total > 0 && op->src[1]->ne[1] > 1) {
1888
+ if ((op->src[0]->nb[1] * op->src[0]->ne[1] != op->src[0]->nb[2]) ||
1889
+ (op->src[1]->nb[1] * op->src[1]->ne[1] != op->src[1]->nb[2])) {
1890
+ return nullptr;
1891
+ }
1892
+ return ggml::cpu::kleidiai::get_tensor_traits(NULL, NULL);
1893
+ }
1894
+ }
1895
+ }
1896
+ return nullptr;
1897
+ }
1898
+ };
1899
+ } // namespace ggml::cpu::kleidiai
1900
+
1901
+ ggml_backend_buffer_type_t ggml_backend_cpu_kleidiai_buffer_type(void) {
1902
+ static ggml::cpu::kleidiai::extra_buffer_type ctx;
1903
+ static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_kleidiai = {
1904
+ /* .iface = */ {
1905
+ /* .get_name = */ ggml_backend_cpu_kleidiai_buffer_type_get_name,
1906
+ /* .alloc_buffer = */ ggml_backend_cpu_kleidiai_buffer_type_alloc_buffer,
1907
+ /* .get_alignment = */ ggml_backend_cpu_kleidiai_buffer_type_get_alignment,
1908
+ /* .get_max_size = */ nullptr, // defaults to SIZE_MAX
1909
+ /* .get_alloc_size = */ ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size,
1910
+ /* .is_host = */ nullptr,
1911
+ },
1912
+ /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
1913
+ /* .context = */ &ctx,
1914
+ };
1915
+
1916
+ init_kleidiai_context();
1917
+
1918
+ return &ggml_backend_cpu_buffer_type_kleidiai;
1919
+ }