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 +1,1795 @@
1
+ #define GGML_COMMON_IMPL_CPP
2
+ #define GGML_COMMON_DECL_CPP
3
+
4
+ #include "repack.h"
5
+
6
+ #include "ggml-common.h"
7
+ #include "ggml-cpu.h"
8
+ #include "ggml-impl.h"
9
+ #include "ime_kernels.h"
10
+
11
+ #include <algorithm>
12
+ #include <cassert>
13
+ #include <cmath>
14
+ #include <cstring>
15
+
16
+ // clang-format off
17
+ #if defined(__riscv)
18
+
19
+ #if !defined(__riscv_v) || !defined(__riscv_v_intrinsic)
20
+ #error "riscv v extension or v_intrinsic not enabled"
21
+ #else
22
+ #include <riscv_vector.h>
23
+ #endif
24
+
25
+ #if !defined(__riscv_zfh)
26
+ #error "riscv zfh extension not enabled"
27
+ #endif
28
+
29
+ #else
30
+ #error "riscv not enabled in this build"
31
+ #endif
32
+
33
+ #if defined(__GNUC__)
34
+ #pragma GCC diagnostic ignored "-Wcast-qual"
35
+ #pragma GCC diagnostic ignored "-Wunused-parameter"
36
+ #endif
37
+
38
+ // clang-format on
39
+
40
+ template <int K> constexpr int QK_0() {
41
+ if constexpr (K == 4) {
42
+ return QK4_0;
43
+ }
44
+ if constexpr (K == 8) {
45
+ return QK8_0;
46
+ }
47
+ return -1;
48
+ }
49
+
50
+ template <int K, int N> struct block {
51
+ ggml_half d[N]; // deltas for N qK_0 blocks
52
+ uint8_t qs[(QK_0<K>() * N * K) / 8]; // quants for N qK_0 blocks
53
+ };
54
+
55
+ template <int K, int N> struct block_with_zp {
56
+ ggml_half d[N]; // deltas for N qK_1 blocks
57
+ uint8_t zp[N]; // zero points for N qK_1 blocks
58
+ uint8_t qs[(QK_0<K>() * N * K) / 8]; // quants for N qK_1 blocks
59
+ };
60
+
61
+ // control size
62
+ static_assert(sizeof(block<4, 16>) == 16 * sizeof(ggml_half) + QK4_0 * 8, "wrong block<4,16> size/padding");
63
+ static_assert(sizeof(block_with_zp<4, 16>) == 16 * sizeof(ggml_half) + QK4_0 * 8 + 16 * sizeof(uint8_t),
64
+ "wrong block_with_zp<4,16> size/padding");
65
+
66
+ static_assert(sizeof(block<8, 16>) == 16 * sizeof(ggml_half) + QK4_0 * 16, "wrong block<8,16> size/padding");
67
+
68
+ static_assert(sizeof(block<4, 32>) == 32 * sizeof(ggml_half) + QK4_0 * 16, "wrong block<4,32> size/padding");
69
+ static_assert(sizeof(block_with_zp<4, 32>) == 32 * sizeof(ggml_half) + QK4_0 * 16 + 32 * sizeof(uint8_t),
70
+ "wrong block_with_zp<4,32> size/padding");
71
+
72
+ using block_q4_0x16 = block<4, 16>;
73
+ using block_q4_1x16 = block_with_zp<4, 16>;
74
+ using block_q8_0x16 = block<8, 16>;
75
+
76
+ using block_q4_0x32 = block<4, 32>;
77
+ using block_q4_1x32 = block_with_zp<4, 32>;
78
+ using block_q8_0x32 = block<8, 32>;
79
+
80
+ struct block_q4_0x32x256 {
81
+ block_q4_0x32 blocks[8]; // [f16 * 32 | i4 * 32 * 32] * 8
82
+ };
83
+
84
+ struct block_q4_1x32x256 {
85
+ block_q4_0x32 blocks[8];
86
+ uint8_t zps[32 * 8];
87
+ };
88
+
89
+ static block_q4_0x16 make_block_q4_0x16(block_q4_0 * in, unsigned int blck_size_interleave) {
90
+ block_q4_0x16 out;
91
+ GGML_ASSERT(QK4_0 / blck_size_interleave == 2);
92
+
93
+ for (int i = 0; i < 16; i++) {
94
+ out.d[i] = in[i].d;
95
+ }
96
+
97
+ for (int i = 0; i < 16; i++) {
98
+ // [0, 15], in.d & 0x0F
99
+ for (int j = 0; j < QK4_0 / 4; j++) {
100
+ //src [b0 b16] ......... [b8 b24] ......... [b15 b31]
101
+ //dst [b0 b8] ......... [b7 b15]
102
+ out.qs[i * QK4_0 / 4 + j] = (in[i].qs[j] & 0x0F) | ((in[i].qs[j + QK4_0 / 4] & 0x0F) << 4);
103
+ }
104
+ }
105
+
106
+ for (int i = 0; i < 16; i++) {
107
+ // [16, 31], in.d & 0xF0
108
+ for (int j = 0; j < QK4_0 / 4; j++) {
109
+ //src [b0 b16] ......... [b8 b24] ......... [b15 b31]
110
+ //dst [b16 b24] ......... [b23 b31]
111
+ out.qs[4 * QK4_0 + i * QK4_0 / 4 + j] = ((in[i].qs[j] & 0xF0) >> 4) | (in[i].qs[j + QK4_0 / 4] & 0xF0);
112
+ }
113
+ }
114
+
115
+ return out;
116
+ }
117
+
118
+ static block_q4_1x16 make_block_q4_1x16(block_q4_1 * in, unsigned int blck_size_interleave) {
119
+ block_q4_1x16 out;
120
+ GGML_ASSERT(QK4_1 / blck_size_interleave == 2);
121
+
122
+ for (int i = 0; i < 16; i++) {
123
+ float d = GGML_FP16_TO_FP32(in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d);
124
+ float m = GGML_FP16_TO_FP32(in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.m);
125
+ float mid = -std::nearbyintf(m / d);
126
+ mid = std::min(15.0f, std::max(0.0f, mid));
127
+ out.d[i] = GGML_FP32_TO_FP16(d);
128
+ out.zp[i] = static_cast<uint8_t>(mid);
129
+ }
130
+
131
+ for (int i = 0; i < 16; i++) {
132
+ // [0, 15], in.d & 0x0F
133
+ for (int j = 0; j < QK4_1 / 4; j++) {
134
+ //src [b0 b16] ......... [b8 b24] ......... [b15 b31]
135
+ //dst [b0 b8] ......... [b7 b15]
136
+ out.qs[i * QK4_1 / 4 + j] = (in[i].qs[j] & 0x0F) | ((in[i].qs[j + QK4_1 / 4] & 0x0F) << 4);
137
+ }
138
+ }
139
+
140
+ for (int i = 0; i < 16; i++) {
141
+ // [16, 31], in.d & 0xF0
142
+ for (int j = 0; j < QK4_1 / 4; j++) {
143
+ //src [b0 b16] ......... [b8 b24] ......... [b15 b31]
144
+ //dst [b16 b24] ......... [b23 b31]
145
+ out.qs[4 * QK4_1 + i * QK4_1 / 4 + j] = ((in[i].qs[j] & 0xF0) >> 4) | (in[i].qs[j + QK4_1 / 4] & 0xF0);
146
+ }
147
+ }
148
+
149
+ return out;
150
+ }
151
+
152
+ static int repack_q4_0_to_q4_0_16_bl(ggml_tensor * t,
153
+ int interleave_block,
154
+ const void * GGML_RESTRICT data,
155
+ size_t data_size) {
156
+ GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
157
+ GGML_ASSERT(interleave_block == 16);
158
+
159
+ constexpr int nrows_interleaved = 16;
160
+
161
+ block_q4_0x16 * dst = (block_q4_0x16 *) t->data;
162
+ const block_q4_0 * src = (const block_q4_0 *) data;
163
+ block_q4_0 dst_tmp[16];
164
+ int nrow = ggml_nrows(t);
165
+ int nblocks = t->ne[0] / QK4_0;
166
+
167
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
168
+
169
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK4_0 != 0) {
170
+ return -1;
171
+ }
172
+
173
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
174
+ for (int64_t x = 0; x < nblocks; x++) {
175
+ for (int i = 0; i < nrows_interleaved; i++) {
176
+ dst_tmp[i] = src[x + i * nblocks];
177
+ }
178
+ *dst++ = make_block_q4_0x16(dst_tmp, interleave_block);
179
+ }
180
+ src += nrows_interleaved * nblocks;
181
+ }
182
+ return 0;
183
+
184
+ GGML_UNUSED(data_size);
185
+ }
186
+
187
+ static int repack_q4_1_to_q4_1_16_bl(ggml_tensor * t,
188
+ int interleave_block,
189
+ const void * GGML_RESTRICT data,
190
+ size_t data_size) {
191
+ GGML_ASSERT(t->type == GGML_TYPE_Q4_1);
192
+ GGML_ASSERT(interleave_block == 16);
193
+
194
+ constexpr int nrows_interleaved = 16;
195
+
196
+ block_q4_1x16 * dst = (block_q4_1x16 *) t->data;
197
+ const block_q4_1 * src = (const block_q4_1 *) data;
198
+ block_q4_1 dst_tmp[16];
199
+ int nrow = ggml_nrows(t);
200
+ int nblocks = t->ne[0] / QK4_1;
201
+
202
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_1));
203
+
204
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK4_1 != 0) {
205
+ return -1;
206
+ }
207
+
208
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
209
+ for (int64_t x = 0; x < nblocks; x++) {
210
+ for (int i = 0; i < nrows_interleaved; i++) {
211
+ dst_tmp[i] = src[x + i * nblocks];
212
+ }
213
+ *dst++ = make_block_q4_1x16(dst_tmp, interleave_block);
214
+ }
215
+ src += nrows_interleaved * nblocks;
216
+ }
217
+ return 0;
218
+
219
+ GGML_UNUSED(data_size);
220
+ }
221
+
222
+ static inline void get_scale_min_k4(int j,
223
+ const uint8_t * GGML_RESTRICT q,
224
+ uint8_t * GGML_RESTRICT d,
225
+ uint8_t * GGML_RESTRICT m) {
226
+ if (j < 4) {
227
+ *d = q[j] & 63;
228
+ *m = q[j + 4] & 63;
229
+ } else {
230
+ *d = (q[j + 4] & 0xF) | ((q[j - 4] >> 6) << 4);
231
+ *m = (q[j + 4] >> 4) | ((q[j - 0] >> 6) << 4);
232
+ }
233
+ }
234
+
235
+ static int repack_q4_k_to_q4_1_16_bl(ggml_tensor * t,
236
+ int interleave_block,
237
+ const void * GGML_RESTRICT data,
238
+ size_t data_size) {
239
+ GGML_ASSERT(t->type == GGML_TYPE_Q4_K);
240
+ GGML_ASSERT(interleave_block == 16);
241
+ GGML_ASSERT(QK_K / QK4_1 == 8);
242
+
243
+ constexpr int nrows_interleaved = 16;
244
+
245
+ block_q4_1x16 * dst = (block_q4_1x16 *) t->data;
246
+ const block_q4_K * src = (const block_q4_K *) data;
247
+ block_q4_1 dst_tmp[16];
248
+ int nrow = ggml_nrows(t);
249
+ int nblocks = t->ne[0] / QK_K;
250
+
251
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK_K != 0) {
252
+ return -1;
253
+ }
254
+
255
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
256
+ for (int64_t x = 0; x < nblocks; x++) {
257
+ for (int j = 0; j < 8; j++) {
258
+ for (int i = 0; i < nrows_interleaved; i++) {
259
+ uint8_t sc, m;
260
+ const float d = GGML_FP16_TO_FP32(src[x + i * nblocks].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d);
261
+ const float min =
262
+ GGML_FP16_TO_FP32(src[x + i * nblocks].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin);
263
+ get_scale_min_k4(j, src[x + i * nblocks].scales, &sc, &m);
264
+ const float d1 = d * sc;
265
+ const float m1 = min * m;
266
+
267
+ dst_tmp[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d = GGML_FP32_TO_FP16(d1);
268
+ dst_tmp[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.m = GGML_FP32_TO_FP16(-m1);
269
+ // src -> [b0, b32] [b1, b33] ... [b31, b63]
270
+ // dst -> [b0, b16] [b1, b17] ... [b15, b31] [b32, b48] [b33, b49] ... [b47, b63]
271
+ const uint8_t * q = src[x + i * nblocks].qs + (j / 2) * QK4_1;
272
+ if (j % 2 == 0) {
273
+ for (int ii = 0; ii < 16; ii++) {
274
+ dst_tmp[i].qs[ii] = (q[ii] & 0x0F) | ((q[ii + 16] & 0x0F) << 4);
275
+ }
276
+ } else {
277
+ for (int ii = 0; ii < 16; ii++) {
278
+ dst_tmp[i].qs[ii] = ((q[ii] & 0xF0) >> 4) | (q[ii + 16] & 0xF0);
279
+ }
280
+ }
281
+ }
282
+ *dst++ = make_block_q4_1x16(dst_tmp, interleave_block);
283
+ }
284
+ }
285
+ src += nrows_interleaved * nblocks;
286
+ }
287
+ return 0;
288
+
289
+ GGML_UNUSED(data_size);
290
+ }
291
+
292
+ static block_q4_0x32 make_block_q4_0x32(block_q4_0 * in, unsigned int blck_size_interleave) {
293
+ block_q4_0x32 out;
294
+ assert(QK4_0 / blck_size_interleave == 1);
295
+ GGML_UNUSED(blck_size_interleave);
296
+
297
+ for (int i = 0; i < 32; i++) {
298
+ out.d[i] = in[i].d;
299
+ }
300
+
301
+ for (int i = 0; i < 32; i++) {
302
+ // [0, 15], in.d & 0x0F
303
+ for (int j = 0; j < QK4_0 / 4; j++) {
304
+ //src [b0 b16] ......... [b8 b24] ......... [b15 b31]
305
+ //dst [b0 b1] ......... [b14 b15]
306
+ out.qs[i * QK4_0 / 2 + j] = (in[i].qs[j * 2] & 0x0F) | ((in[i].qs[j * 2 + 1] & 0x0F) << 4);
307
+ }
308
+ }
309
+
310
+ for (int i = 0; i < 32; i++) {
311
+ // [16, 31], in.d & 0xF0
312
+ for (int j = 0; j < QK4_0 / 4; j++) {
313
+ //src [b0 b16] ......... [b8 b24] ......... [b15 b31]
314
+ //dst [b16 b17] ......... [b30 b31]
315
+ out.qs[i * QK4_0 / 2 + QK4_0 / 4 + j] = ((in[i].qs[j * 2] & 0xF0) >> 4) | (in[i].qs[j * 2 + 1] & 0xF0);
316
+ }
317
+ }
318
+
319
+ return out;
320
+ }
321
+
322
+ static block_q4_1x32 make_block_q4_1x32(block_q4_1 * in, unsigned int blck_size_interleave) {
323
+ block_q4_1x32 out;
324
+ GGML_ASSERT(QK4_1 / blck_size_interleave == 1);
325
+ GGML_UNUSED(blck_size_interleave);
326
+
327
+ for (int i = 0; i < 32; i++) {
328
+ float d = GGML_FP16_TO_FP32(in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d);
329
+ float m = GGML_FP16_TO_FP32(in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.m);
330
+ float mid = -std::nearbyintf(m / d);
331
+ mid = std::min(15.0f, std::max(0.0f, mid));
332
+ out.d[i] = GGML_FP32_TO_FP16(d);
333
+ out.zp[i] = static_cast<uint8_t>(mid);
334
+ }
335
+
336
+ for (int i = 0; i < 32; i++) {
337
+ // [0, 15], in.d & 0x0F
338
+ for (int j = 0; j < QK4_1 / 4; j++) {
339
+ //src [b0 b16] ......... [b8 b24] ......... [b15 b31]
340
+ //dst [b0 b1] ......... [b14 b15]
341
+ out.qs[i * QK4_1 / 2 + j] = (in[i].qs[j * 2] & 0x0F) | ((in[i].qs[j * 2 + 1] & 0x0F) << 4);
342
+ }
343
+ }
344
+
345
+ for (int i = 0; i < 32; i++) {
346
+ // [16, 31], in.d & 0xF0
347
+ for (int j = 0; j < QK4_1 / 4; j++) {
348
+ //src [b0 b16] ......... [b8 b24] ......... [b15 b31]
349
+ //dst [b16 b24] ......... [b23 b31]
350
+ out.qs[i * QK4_1 / 2 + QK4_1 / 4 + j] = ((in[i].qs[j * 2] & 0xF0) >> 4) | (in[i].qs[j * 2 + 1] & 0xF0);
351
+ }
352
+ }
353
+
354
+ return out;
355
+ }
356
+
357
+ static block_q8_0x32 make_block_q8_0x32(block_q8_0 * in, unsigned int blck_size_interleave) {
358
+ block_q8_0x32 out;
359
+ GGML_ASSERT(QK8_0 / blck_size_interleave == 1);
360
+ GGML_UNUSED(blck_size_interleave);
361
+
362
+ for (int i = 0; i < 32; i++) {
363
+ out.d[i] = in[i].d;
364
+ }
365
+
366
+ for (int i = 0; i < 32; i++) {
367
+ memcpy(out.qs + i * QK8_0, in[i].qs, QK8_0);
368
+ }
369
+
370
+ return out;
371
+ }
372
+
373
+ static int repack_q2_k_to_q2_k_32_bl(ggml_tensor * t,
374
+ int interleave_block,
375
+ const void * GGML_RESTRICT data,
376
+ size_t data_size) {
377
+ GGML_ASSERT(t->type == GGML_TYPE_Q2_K);
378
+ GGML_ASSERT(interleave_block == 32);
379
+ GGML_ASSERT(QK_K == 256);
380
+
381
+ constexpr int nrows_interleaved = 32;
382
+
383
+ const block_q2_K * src = (const block_q2_K *) data;
384
+
385
+ auto * dst = (spacemit_kernels::nrow_block_q2_k<32> *) t->data;
386
+
387
+ int nrow = ggml_nrows(t);
388
+ int nblocks = t->ne[0] / QK_K;
389
+
390
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q2_K));
391
+
392
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK_K != 0) {
393
+ return -1;
394
+ }
395
+
396
+ uint8_t qs_aux[256] = { 0 };
397
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
398
+ for (int64_t x = 0; x < nblocks; x++) {
399
+ for (int i = 0; i < nrows_interleaved; i++) {
400
+ const block_q2_K * src_block = &src[(b + i) * nblocks + x];
401
+
402
+ // scale for [16, N]
403
+ for (int j = 0; j < 16; j++) {
404
+ auto zp_aux = (dst->scales[j * nrows_interleaved + i]) & 0xF0;
405
+
406
+ dst->scales[j * nrows_interleaved + i] = (src_block->scales[j] & 0x0F) | zp_aux;
407
+ }
408
+
409
+ // zp for [N, 16]
410
+ for (int j = 0; j < 16; j++) {
411
+ auto scale_aux = (dst->scales[16 * i + j]) & 0x0F;
412
+
413
+ dst->scales[16 * i + j] = (src_block->scales[j] & 0xF0) | scale_aux;
414
+ }
415
+
416
+ for (int k = 0; k < 4; k++) {
417
+ for (int j = 0; j < 32; j++) {
418
+ qs_aux[k * 32 + j] = (src_block->qs[j] >> (2 * k)) & 0x03;
419
+ }
420
+ }
421
+
422
+ for (int k = 0; k < 4; k++) {
423
+ for (int j = 0; j < 32; j++) {
424
+ qs_aux[k * 32 + j + 128] = (src_block->qs[j + 32] >> (2 * k)) & 0x03;
425
+ }
426
+ }
427
+
428
+ // from nrows_interleaved * [2 * 32byte]
429
+ // to 4 * [nrows_interleaved * 16byte]
430
+ for (int k = 0; k < 4; k++) {
431
+ for (int j = 0; j < 16; j++) {
432
+ uint8_t qs0 = qs_aux[j + k * 64];
433
+ uint8_t qs16 = qs_aux[j + 16 + k * 64];
434
+ uint8_t qs32 = qs_aux[j + 32 + k * 64];
435
+ uint8_t qs48 = qs_aux[j + 48 + k * 64];
436
+
437
+ dst->qs[(k * nrows_interleaved + i) * 16 + j] =
438
+ (qs0 & 0x03) | ((qs16 & 0x03) << 2) | ((qs32 & 0x03) << 4) | ((qs48 & 0x03) << 6);
439
+ }
440
+ }
441
+
442
+ dst->scales16[i] = src_block->GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d;
443
+ dst->zeros16[i] = src_block->GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin;
444
+ }
445
+ dst++;
446
+ }
447
+ }
448
+
449
+ return 0;
450
+ }
451
+
452
+ static int repack_q3_k_to_q3_k_32_bl(ggml_tensor * t,
453
+ int interleave_block,
454
+ const void * GGML_RESTRICT data,
455
+ size_t data_size) {
456
+ GGML_ASSERT(t->type == GGML_TYPE_Q3_K);
457
+ GGML_ASSERT(interleave_block == 32);
458
+ GGML_ASSERT(QK_K == 256);
459
+
460
+ constexpr int nrows_interleaved = 32;
461
+
462
+ const uint32_t kmask1 = 0x03030303;
463
+ const uint32_t kmask2 = 0x0f0f0f0f;
464
+
465
+ const block_q3_K * src = (const block_q3_K *) data;
466
+
467
+ auto * dst = (spacemit_kernels::nrow_block_q3_k<32> *) t->data;
468
+
469
+ int nrow = ggml_nrows(t);
470
+ int nblocks = t->ne[0] / QK_K;
471
+
472
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q3_K));
473
+
474
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK_K != 0) {
475
+ return -1;
476
+ }
477
+
478
+ uint32_t b_scale_aux[4] = { 0 };
479
+ uint8_t qs_aux[256] = { 0 };
480
+
481
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
482
+ for (int64_t x = 0; x < nblocks; x++) {
483
+ for (int i = 0; i < nrows_interleaved; i++) {
484
+ const block_q3_K * src_block = &src[(b + i) * nblocks + x];
485
+
486
+ uint32_t * auxs = b_scale_aux;
487
+ int8_t * scale = (int8_t *) auxs;
488
+ memcpy(auxs, src_block->scales, 12);
489
+
490
+ uint32_t tmp = auxs[2];
491
+ auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
492
+ auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
493
+ auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
494
+ auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
495
+
496
+ for (int j = 0; j < 16; j++) {
497
+ dst->scales[j * nrows_interleaved + i] = scale[j] - 32;
498
+ }
499
+
500
+ for (int k = 0; k < 4; k++) {
501
+ for (int j = 0; j < 32; j++) {
502
+ qs_aux[k * 32 + j] = (src_block->qs[j] >> (2 * k)) & 0x03;
503
+ }
504
+ }
505
+
506
+ for (int k = 0; k < 4; k++) {
507
+ for (int j = 0; j < 32; j++) {
508
+ qs_aux[k * 32 + j + 128] = (src_block->qs[j + 32] >> (2 * k)) & 0x03;
509
+ }
510
+ }
511
+
512
+ // from nrows_interleaved * [2 * 32byte]
513
+ // to 4 * [nrows_interleaved * 16byte]
514
+ for (int k = 0; k < 4; k++) {
515
+ for (int j = 0; j < 16; j++) {
516
+ uint8_t qs0 = qs_aux[j + k * 64];
517
+ uint8_t qs16 = qs_aux[j + 16 + k * 64];
518
+ uint8_t qs32 = qs_aux[j + 32 + k * 64];
519
+ uint8_t qs48 = qs_aux[j + 48 + k * 64];
520
+
521
+ dst->qs[(k * nrows_interleaved + i) * 16 + j] =
522
+ (qs0 & 0x03) | ((qs16 & 0x03) << 2) | ((qs32 & 0x03) << 4) | ((qs48 & 0x03) << 6);
523
+ }
524
+ }
525
+
526
+ //memcpy(dst->hmask + i * 32, src_block->hmask, 32);
527
+
528
+ // from nrows_interleaved * [32byte]
529
+ // to 16 * [nrows_interleaved * uint16_t]
530
+ uint16_t * dst_mask = ((uint16_t *) dst->hmask) + i;
531
+ for (int j = 0; j < 16; j++, dst_mask += nrows_interleaved) {
532
+ uint8_t b_shift = j / 2;
533
+ uint8_t * b_mask_col = (uint8_t *) (src_block->hmask + (j % 2) * 16);
534
+ // b0 - b15
535
+ uint16_t msk_out_0 = 0;
536
+
537
+ for (int k = 0; k < 8; k++) {
538
+ msk_out_0 |= (uint16_t) ((b_mask_col[k] >> b_shift) & 0x01) << k;
539
+ }
540
+ for (int k = 8; k < 16; k++) {
541
+ msk_out_0 |= (uint16_t) ((b_mask_col[k] >> b_shift) & 0x01) << k;
542
+ }
543
+
544
+ dst_mask[0] = msk_out_0;
545
+ }
546
+
547
+ dst->scales16[i] = src_block->d;
548
+ }
549
+
550
+ dst++;
551
+ }
552
+ }
553
+
554
+ return 0;
555
+ }
556
+
557
+ static int repack_q4_0_to_q4_0_32_bl_ref(ggml_tensor * t,
558
+ int interleave_block,
559
+ const void * GGML_RESTRICT data,
560
+ size_t data_size) {
561
+ GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
562
+ GGML_ASSERT(interleave_block == 32); // unused
563
+
564
+ constexpr int nrows_interleaved = 32;
565
+
566
+ block_q4_0x32 * dst = (block_q4_0x32 *) t->data;
567
+ const block_q4_0 * src = (const block_q4_0 *) data;
568
+ block_q4_0 dst_tmp[32];
569
+ int nrow = ggml_nrows(t);
570
+ int nblocks = t->ne[0] / QK4_0;
571
+
572
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
573
+
574
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK4_0 != 0) {
575
+ return -1;
576
+ }
577
+
578
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
579
+ for (int64_t x = 0; x < nblocks; x++) {
580
+ for (int i = 0; i < nrows_interleaved; i++) {
581
+ dst_tmp[i] = src[x + i * nblocks];
582
+ }
583
+ *dst++ = make_block_q4_0x32(dst_tmp, interleave_block);
584
+ }
585
+ src += nrows_interleaved * nblocks;
586
+ }
587
+ return 0;
588
+
589
+ GGML_UNUSED(data_size);
590
+ }
591
+
592
+ static int repack_q4_0_to_q4_0_256_32_bl_ref(ggml_tensor * t,
593
+ int interleave_block,
594
+ const void * GGML_RESTRICT data,
595
+ size_t data_size) {
596
+ GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
597
+ GGML_ASSERT(interleave_block == 32); // unused
598
+
599
+ constexpr int nrows_interleaved = 32;
600
+
601
+ block_q4_0x32x256 * dst = (block_q4_0x32x256 *) t->data;
602
+ const block_q4_0 * src = (const block_q4_0 *) data;
603
+ block_q4_0 dst_tmp[32];
604
+ int nrow = ggml_nrows(t);
605
+ int nblocks = t->ne[0] / QK4_0;
606
+
607
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
608
+ GGML_ASSERT(nblocks % 8 == 0); // for 256-block interleaving
609
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK4_0 != 0) {
610
+ return -1;
611
+ }
612
+
613
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
614
+ for (int64_t x = 0; x < nblocks; x += 8) {
615
+ for (int j = 0; j < 8; j++) {
616
+ for (int i = 0; i < nrows_interleaved; i++) {
617
+ dst_tmp[i] = src[x + j + i * nblocks];
618
+ }
619
+ dst->blocks[j] = make_block_q4_0x32(dst_tmp, interleave_block);
620
+ }
621
+ dst++;
622
+ }
623
+ src += nrows_interleaved * nblocks;
624
+ }
625
+ return 0;
626
+
627
+ GGML_UNUSED(data_size);
628
+ }
629
+
630
+ static int repack_q4_0_to_q4_1_256_32_bl_ref(ggml_tensor * t,
631
+ int interleave_block,
632
+ const void * GGML_RESTRICT data,
633
+ size_t data_size) {
634
+ GGML_ASSERT(t->type == GGML_TYPE_Q4_1);
635
+ GGML_ASSERT(interleave_block == 32); // unused
636
+
637
+ constexpr int nrows_interleaved = 32;
638
+
639
+ block_q4_1x32x256 * dst = (block_q4_1x32x256 *) t->data;
640
+ const block_q4_1 * src = (const block_q4_1 *) data;
641
+ block_q4_1 dst_tmp[32];
642
+ int nrow = ggml_nrows(t);
643
+ int nblocks = t->ne[0] / QK4_0;
644
+
645
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_1));
646
+ GGML_ASSERT(nblocks % 8 == 0); // for 256-block interleaving
647
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK4_0 != 0) {
648
+ return -1;
649
+ }
650
+
651
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
652
+ for (int64_t x = 0; x < nblocks; x += 8) {
653
+ for (int j = 0; j < 8; j++) {
654
+ for (int i = 0; i < nrows_interleaved; i++) {
655
+ dst_tmp[i] = src[x + j + i * nblocks];
656
+ }
657
+
658
+ block_q4_0x32 * dst_block = &dst->blocks[j];
659
+ uint8_t * dst_zp = dst->zps + j * nrows_interleaved;
660
+
661
+ for (int i = 0; i < nrows_interleaved; i++) {
662
+ float d = GGML_FP16_TO_FP32(dst_tmp[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d);
663
+ float m = GGML_FP16_TO_FP32(dst_tmp[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.m);
664
+ float mid = -std::nearbyintf(m / d);
665
+ mid = std::min(15.0f, std::max(0.0f, mid));
666
+
667
+ dst_block->d[i] = GGML_FP32_TO_FP16(d);
668
+ dst_zp[i] = static_cast<uint8_t>(mid);
669
+ }
670
+
671
+ for (int i = 0; i < nrows_interleaved; i++) {
672
+ for (int k = 0; k < QK4_1 / 4; k++) {
673
+ dst_block->qs[i * QK4_1 / 2 + k] =
674
+ (dst_tmp[i].qs[k * 2] & 0x0F) | ((dst_tmp[i].qs[k * 2 + 1] & 0x0F) << 4);
675
+ }
676
+ }
677
+
678
+ for (int i = 0; i < nrows_interleaved; i++) {
679
+ for (int k = 0; k < QK4_1 / 4; k++) {
680
+ dst_block->qs[i * QK4_1 / 2 + QK4_1 / 4 + k] =
681
+ ((dst_tmp[i].qs[k * 2] & 0xF0) >> 4) | (dst_tmp[i].qs[k * 2 + 1] & 0xF0);
682
+ }
683
+ }
684
+ }
685
+ dst++;
686
+ }
687
+ src += nrows_interleaved * nblocks;
688
+ }
689
+ return 0;
690
+
691
+ GGML_UNUSED(data_size);
692
+ }
693
+
694
+ // RVV optimized version of repack_q4_0_to_q4_0_32_bl
695
+ // Eliminates the intermediate dst_tmp buffer and vectorizes nibble repack.
696
+ static int repack_q4_0_to_q4_0_32_bl(ggml_tensor * t,
697
+ int interleave_block,
698
+ const void * GGML_RESTRICT data,
699
+ size_t data_size) {
700
+ GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
701
+ GGML_ASSERT(interleave_block == 32);
702
+
703
+ constexpr int nrows_interleaved = 32;
704
+ constexpr int qs_bytes = QK4_0 / 2; // 16
705
+
706
+ block_q4_0x32 * dst = (block_q4_0x32 *) t->data;
707
+ const block_q4_0 * src = (const block_q4_0 *) data;
708
+ int nrow = ggml_nrows(t);
709
+ int nblocks = t->ne[0] / QK4_0;
710
+
711
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
712
+
713
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK4_0 != 0) {
714
+ return -1;
715
+ }
716
+
717
+ const ptrdiff_t row_stride = (ptrdiff_t) nblocks * sizeof(block_q4_0);
718
+
719
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
720
+ for (int64_t x = 0; x < nblocks; x++) {
721
+ const block_q4_0 * col_src = src + x;
722
+
723
+ // --- 1) Gather 32 scale values (ggml_half d) with stride load ---
724
+ // d is at offset 0 of each block_q4_0, stride between rows = row_stride
725
+ {
726
+ const uint8_t * d_base = (const uint8_t *) &col_src->d;
727
+ ggml_half * d_dst = dst->d;
728
+ size_t remaining = 32;
729
+ size_t offset = 0;
730
+ while (remaining > 0) {
731
+ size_t vl = __riscv_vsetvl_e16m1(remaining);
732
+ vuint16m1_t vd =
733
+ __riscv_vlse16_v_u16m1((const uint16_t *) (d_base + offset * row_stride), row_stride, vl);
734
+ __riscv_vse16_v_u16m1((uint16_t *) (d_dst + offset), vd, vl);
735
+ offset += vl;
736
+ remaining -= vl;
737
+ }
738
+ }
739
+
740
+ // --- 2) Nibble repack qs for each of the 32 rows ---
741
+ // For each row i:
742
+ // src qs[16]: [b0|b16] [b1|b17] ... [b15|b31] (lo nibble = b_j, hi nibble = b_{j+16})
743
+ // dst qs low 8B: (qs[2j] & 0x0F) | ((qs[2j+1] & 0x0F) << 4) for j=0..7
744
+ // dst qs high 8B: ((qs[2j] >> 4)) | (qs[2j+1] & 0xF0) for j=0..7
745
+ {
746
+ const size_t vl8 = __riscv_vsetvl_e8m1(8);
747
+ for (int i = 0; i < 32; i++) {
748
+ const uint8_t * sq = col_src[i * nblocks].qs;
749
+ uint8_t * dq = dst->qs + i * qs_bytes;
750
+
751
+ // stride-2 load to separate even/odd bytes
752
+ vuint8m1_t v_even = __riscv_vlse8_v_u8m1(sq, 2, vl8); // qs[0], qs[2], ..., qs[14]
753
+ vuint8m1_t v_odd = __riscv_vlse8_v_u8m1(sq + 1, 2, vl8); // qs[1], qs[3], ..., qs[15]
754
+
755
+ // low nibble part: (even & 0x0F) | ((odd & 0x0F) << 4)
756
+ vuint8m1_t v_even_lo = __riscv_vand_vx_u8m1(v_even, 0x0F, vl8);
757
+ vuint8m1_t v_odd_lo = __riscv_vand_vx_u8m1(v_odd, 0x0F, vl8);
758
+ vuint8m1_t v_lo = __riscv_vor_vv_u8m1(v_even_lo, __riscv_vsll_vx_u8m1(v_odd_lo, 4, vl8), vl8);
759
+
760
+ // high nibble part: (even >> 4) | (odd & 0xF0)
761
+ vuint8m1_t v_even_hi = __riscv_vsrl_vx_u8m1(v_even, 4, vl8);
762
+ vuint8m1_t v_odd_hi = __riscv_vand_vx_u8m1(v_odd, 0xF0, vl8);
763
+ vuint8m1_t v_hi = __riscv_vor_vv_u8m1(v_even_hi, v_odd_hi, vl8);
764
+
765
+ __riscv_vse8_v_u8m1(dq, v_lo, vl8);
766
+ __riscv_vse8_v_u8m1(dq + 8, v_hi, vl8);
767
+ }
768
+ }
769
+
770
+ dst++;
771
+ }
772
+ src += nrows_interleaved * nblocks;
773
+ }
774
+ return 0;
775
+
776
+ GGML_UNUSED(data_size);
777
+ }
778
+
779
+ static int repack_q4_1_to_q4_1_32_bl_ref(ggml_tensor * t,
780
+ int interleave_block,
781
+ const void * GGML_RESTRICT data,
782
+ size_t data_size) {
783
+ GGML_ASSERT(t->type == GGML_TYPE_Q4_1);
784
+ GGML_ASSERT(interleave_block == 32); // unused
785
+
786
+ constexpr int nrows_interleaved = 32;
787
+
788
+ block_q4_1x32 * dst = (block_q4_1x32 *) t->data;
789
+ const block_q4_1 * src = (const block_q4_1 *) data;
790
+ block_q4_1 dst_tmp[32];
791
+ int nrow = ggml_nrows(t);
792
+ int nblocks = t->ne[0] / QK4_1;
793
+
794
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_1));
795
+
796
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK4_1 != 0) {
797
+ return -1;
798
+ }
799
+
800
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
801
+ for (int64_t x = 0; x < nblocks; x++) {
802
+ for (int i = 0; i < nrows_interleaved; i++) {
803
+ dst_tmp[i] = src[x + i * nblocks];
804
+ }
805
+ *dst++ = make_block_q4_1x32(dst_tmp, interleave_block);
806
+ }
807
+ src += nrows_interleaved * nblocks;
808
+ }
809
+ return 0;
810
+
811
+ GGML_UNUSED(data_size);
812
+ }
813
+
814
+ // RVV optimized version of repack_q4_1_to_q4_1_32_bl
815
+ // Eliminates the intermediate dst_tmp buffer and vectorizes nibble repack + zp computation.
816
+ static int repack_q4_1_to_q4_1_32_bl(ggml_tensor * t,
817
+ int interleave_block,
818
+ const void * GGML_RESTRICT data,
819
+ size_t data_size) {
820
+ GGML_ASSERT(t->type == GGML_TYPE_Q4_1);
821
+ GGML_ASSERT(interleave_block == 32);
822
+
823
+ constexpr int nrows_interleaved = 32;
824
+ constexpr int qs_bytes = QK4_1 / 2; // 16
825
+
826
+ block_q4_1x32 * dst = (block_q4_1x32 *) t->data;
827
+ const block_q4_1 * src = (const block_q4_1 *) data;
828
+ int nrow = ggml_nrows(t);
829
+ int nblocks = t->ne[0] / QK4_1;
830
+
831
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_1));
832
+
833
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK4_1 != 0) {
834
+ return -1;
835
+ }
836
+
837
+ const ptrdiff_t row_stride = (ptrdiff_t) nblocks * sizeof(block_q4_1);
838
+
839
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
840
+ for (int64_t x = 0; x < nblocks; x++) {
841
+ const block_q4_1 * col_src = src + x;
842
+
843
+ // --- 1) Gather d and m, compute zp = clamp(nearbyint(-m/d), 0, 15) ---
844
+ // block_q4_1 layout: [d(f16), m(f16), qs[16]]
845
+ // d is at byte offset 0, m is at byte offset 2 from each block start
846
+ {
847
+ const uint8_t * dm_base = (const uint8_t *) &col_src->GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d;
848
+ ggml_half * d_dst = dst->d;
849
+ uint8_t * zp_dst = dst->zp;
850
+ size_t remaining = 32;
851
+ size_t offset = 0;
852
+ while (remaining > 0) {
853
+ size_t vl = __riscv_vsetvl_e16m1(remaining);
854
+
855
+ // stride load d (f16) from each row
856
+ vuint16m1_t vd_raw =
857
+ __riscv_vlse16_v_u16m1((const uint16_t *) (dm_base + offset * row_stride), row_stride, vl);
858
+ __riscv_vse16_v_u16m1((uint16_t *) (d_dst + offset), vd_raw, vl);
859
+
860
+ // stride load m (f16) from each row (offset +2 bytes from d)
861
+ vuint16m1_t vm_raw =
862
+ __riscv_vlse16_v_u16m1((const uint16_t *) (dm_base + 2 + offset * row_stride), row_stride, vl);
863
+
864
+ // convert to f32 for zp computation: zp = nearbyint(-m / d)
865
+ vfloat16m1_t vd_f16 = __riscv_vreinterpret_v_u16m1_f16m1(vd_raw);
866
+ vfloat16m1_t vm_f16 = __riscv_vreinterpret_v_u16m1_f16m1(vm_raw);
867
+
868
+ // -m / d in f16 directly (SpaceMIT X60 supports f16 arithmetic)
869
+ vfloat16m1_t v_neg_m = __riscv_vfneg_v_f16m1(vm_f16, vl);
870
+ vfloat16m1_t v_ratio = __riscv_vfdiv_vv_f16m1(v_neg_m, vd_f16, vl);
871
+
872
+ // Convert to f32 for nearbyint, then clamp
873
+ vfloat32m2_t v_ratio_f32 = __riscv_vfwcvt_f_f_v_f32m2(v_ratio, vl);
874
+
875
+ // Use integer rounding: convert f32 -> int (rounds to nearest)
876
+ vint32m2_t v_zp_i32 = __riscv_vfcvt_x_f_v_i32m2(v_ratio_f32, vl);
877
+
878
+ // clamp to [0, 15]
879
+ v_zp_i32 = __riscv_vmax_vx_i32m2(v_zp_i32, 0, vl);
880
+ v_zp_i32 = __riscv_vmin_vx_i32m2(v_zp_i32, 15, vl);
881
+
882
+ // narrow i32 -> u8
883
+ vint16m1_t v_zp_i16 = __riscv_vncvt_x_x_w_i16m1(v_zp_i32, vl);
884
+ vint8mf2_t v_zp_i8 = __riscv_vncvt_x_x_w_i8mf2(v_zp_i16, vl);
885
+ vuint8mf2_t v_zp_u8 = __riscv_vreinterpret_v_i8mf2_u8mf2(v_zp_i8);
886
+ __riscv_vse8_v_u8mf2(zp_dst + offset, v_zp_u8, vl);
887
+
888
+ offset += vl;
889
+ remaining -= vl;
890
+ }
891
+ }
892
+
893
+ // --- 2) Nibble repack qs for each of the 32 rows ---
894
+ {
895
+ const size_t vl8 = __riscv_vsetvl_e8m1(8);
896
+ for (int i = 0; i < 32; i++) {
897
+ const uint8_t * sq = col_src[i * nblocks].qs;
898
+ uint8_t * dq = dst->qs + i * qs_bytes;
899
+
900
+ // stride-2 load to separate even/odd bytes
901
+ vuint8m1_t v_even = __riscv_vlse8_v_u8m1(sq, 2, vl8);
902
+ vuint8m1_t v_odd = __riscv_vlse8_v_u8m1(sq + 1, 2, vl8);
903
+
904
+ // low nibble part: (even & 0x0F) | ((odd & 0x0F) << 4)
905
+ vuint8m1_t v_even_lo = __riscv_vand_vx_u8m1(v_even, 0x0F, vl8);
906
+ vuint8m1_t v_odd_lo = __riscv_vand_vx_u8m1(v_odd, 0x0F, vl8);
907
+ vuint8m1_t v_lo = __riscv_vor_vv_u8m1(v_even_lo, __riscv_vsll_vx_u8m1(v_odd_lo, 4, vl8), vl8);
908
+
909
+ // high nibble part: (even >> 4) | (odd & 0xF0)
910
+ vuint8m1_t v_even_hi = __riscv_vsrl_vx_u8m1(v_even, 4, vl8);
911
+ vuint8m1_t v_odd_hi = __riscv_vand_vx_u8m1(v_odd, 0xF0, vl8);
912
+ vuint8m1_t v_hi = __riscv_vor_vv_u8m1(v_even_hi, v_odd_hi, vl8);
913
+
914
+ __riscv_vse8_v_u8m1(dq, v_lo, vl8);
915
+ __riscv_vse8_v_u8m1(dq + 8, v_hi, vl8);
916
+ }
917
+ }
918
+
919
+ dst++;
920
+ }
921
+ src += nrows_interleaved * nblocks;
922
+ }
923
+ return 0;
924
+
925
+ GGML_UNUSED(data_size);
926
+ }
927
+
928
+ static int repack_q4_k_to_q4_1_32_bl(ggml_tensor * t,
929
+ int interleave_block,
930
+ const void * GGML_RESTRICT data,
931
+ size_t data_size) {
932
+ GGML_ASSERT(t->type == GGML_TYPE_Q4_K);
933
+ GGML_ASSERT(interleave_block == 32);
934
+ GGML_ASSERT(QK_K / QK4_1 == 8);
935
+
936
+ constexpr int nrows_interleaved = 32;
937
+
938
+ block_q4_1x32 * dst = (block_q4_1x32 *) t->data;
939
+ const block_q4_K * src = (const block_q4_K *) data;
940
+ block_q4_1 dst_tmp[32];
941
+ int nrow = ggml_nrows(t);
942
+ int nblocks = t->ne[0] / QK_K;
943
+
944
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK_K != 0) {
945
+ return -1;
946
+ }
947
+
948
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
949
+ for (int64_t x = 0; x < nblocks; x++) {
950
+ for (int j = 0; j < 8; j++) {
951
+ for (int i = 0; i < nrows_interleaved; i++) {
952
+ uint8_t sc, m;
953
+ const float d = GGML_FP16_TO_FP32(src[x + i * nblocks].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d);
954
+ const float min =
955
+ GGML_FP16_TO_FP32(src[x + i * nblocks].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin);
956
+ get_scale_min_k4(j, src[x + i * nblocks].scales, &sc, &m);
957
+ const float d1 = d * sc;
958
+ const float m1 = min * m;
959
+
960
+ dst_tmp[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d = GGML_FP32_TO_FP16(d1);
961
+ dst_tmp[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.m = GGML_FP32_TO_FP16(-m1);
962
+ // src -> [b0, b32] [b1, b33] ... [b31, b63]
963
+ // dst -> [b0, b16] [b1, b17] ... [b15, b31] [b32, b48] [b33, b49] ... [b47, b63]
964
+ const uint8_t * q = src[x + i * nblocks].qs + (j / 2) * QK4_1;
965
+ if (j % 2 == 0) {
966
+ for (int ii = 0; ii < 16; ii++) {
967
+ dst_tmp[i].qs[ii] = (q[ii] & 0x0F) | ((q[ii + 16] & 0x0F) << 4);
968
+ }
969
+ } else {
970
+ for (int ii = 0; ii < 16; ii++) {
971
+ dst_tmp[i].qs[ii] = ((q[ii] & 0xF0) >> 4) | (q[ii + 16] & 0xF0);
972
+ }
973
+ }
974
+ }
975
+ *dst++ = make_block_q4_1x32(dst_tmp, interleave_block);
976
+ }
977
+ }
978
+ src += nrows_interleaved * nblocks;
979
+ }
980
+ return 0;
981
+
982
+ GGML_UNUSED(data_size);
983
+ }
984
+
985
+ static int repack_q6_k_to_q8_0_32_bl_ref(ggml_tensor * t,
986
+ int interleave_block,
987
+ const void * GGML_RESTRICT data,
988
+ size_t data_size) {
989
+ GGML_ASSERT(t->type == GGML_TYPE_Q6_K);
990
+ GGML_ASSERT(interleave_block == 32);
991
+ GGML_ASSERT(QK_K / QK4_1 == 8);
992
+
993
+ constexpr int nrows_interleaved = 32;
994
+
995
+ block_q8_0x32 * dst = (block_q8_0x32 *) t->data;
996
+ const block_q6_K * src = (const block_q6_K *) data;
997
+ block_q8_0 dst_tmp[32];
998
+ int8_t aux8[QK4_1];
999
+ int nrow = ggml_nrows(t);
1000
+ int nblocks = t->ne[0] / QK_K;
1001
+
1002
+ if (t->ne[0] % QK_K != 0) {
1003
+ return -1;
1004
+ }
1005
+
1006
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
1007
+ int64_t nrow_real = std::min((int64_t) nrow - b, (int64_t) nrows_interleaved);
1008
+ for (int64_t x = 0; x < nblocks; x++) {
1009
+ for (int bi = 0; bi < 8; bi++) {
1010
+ int i = 0;
1011
+ for (; i < nrow_real; i++) {
1012
+ const uint8_t * q4 = src[x + i * nblocks].ql;
1013
+ const uint8_t * qh = src[x + i * nblocks].qh;
1014
+ const int8_t * scales = src[x + i * nblocks].scales;
1015
+ float d = GGML_FP16_TO_FP32(src[x + i * nblocks].d);
1016
+
1017
+ q4 += 64 * (bi / 4);
1018
+ qh += 32 * (bi / 4);
1019
+ int8_t * GGML_RESTRICT a = aux8;
1020
+
1021
+ int8_t bi_idx = bi % 4;
1022
+
1023
+ if (bi_idx == 0) {
1024
+ for (int l = 0; l < 32; ++l) {
1025
+ a[l] = (int8_t) ((q4[l] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
1026
+ }
1027
+ } else if (bi_idx == 1) {
1028
+ for (int l = 0; l < 32; ++l) {
1029
+ a[l] = (int8_t) ((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
1030
+ }
1031
+ } else if (bi_idx == 2) {
1032
+ for (int l = 0; l < 32; ++l) {
1033
+ a[l] = (int8_t) ((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
1034
+ }
1035
+ } else if (bi_idx == 3) {
1036
+ for (int l = 0; l < 32; ++l) {
1037
+ a[l] = (int8_t) ((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
1038
+ }
1039
+ }
1040
+ a = aux8;
1041
+
1042
+ float a_max_abs = 0.0f;
1043
+ float scale_0 = scales[bi * 2 + 0] * d;
1044
+ float scale_1 = scales[bi * 2 + 1] * d;
1045
+ for (int l = 0; l < 16; ++l) {
1046
+ a_max_abs = std::max(a_max_abs, std::abs(a[l] * scale_0));
1047
+ }
1048
+
1049
+ for (int l = 16; l < 32; ++l) {
1050
+ a_max_abs = std::max(a_max_abs, std::abs(a[l] * scale_1));
1051
+ }
1052
+
1053
+ float reflect_scale = a_max_abs / ((1 << 7) - 1);
1054
+ float reflect_scale_0 = scale_0 / reflect_scale;
1055
+ float reflect_scale_1 = scale_1 / reflect_scale;
1056
+
1057
+ for (int l = 0; l < 16; ++l) {
1058
+ float a_temp = std::clamp(std::nearbyintf(a[l] * reflect_scale_0), -128.0f, 127.0f);
1059
+ a[l] = (int8_t) (a_temp);
1060
+ }
1061
+
1062
+ for (int l = 16; l < 32; ++l) {
1063
+ float a_temp = std::clamp(std::nearbyintf(a[l] * reflect_scale_1), -128.0f, 127.0f);
1064
+ a[l] = (int8_t) (a_temp);
1065
+ }
1066
+
1067
+ dst_tmp[i].d = GGML_FP32_TO_FP16(reflect_scale);
1068
+
1069
+ memcpy(dst_tmp[i].qs, a, 32 * sizeof(int8_t));
1070
+ }
1071
+
1072
+ for (; i < nrows_interleaved; i++) {
1073
+ memset(&dst_tmp[i], 0, sizeof(block_q8_0));
1074
+ }
1075
+
1076
+ *dst++ = make_block_q8_0x32(dst_tmp, interleave_block);
1077
+ }
1078
+ }
1079
+ src += nrows_interleaved * nblocks;
1080
+ }
1081
+ return 0;
1082
+
1083
+ GGML_UNUSED(data_size);
1084
+ }
1085
+
1086
+ // RVV optimized version of repack_q6_k_to_q8_0_32_bl
1087
+ // Vectorizes the Q6_K dequant -> requant pipeline using RVV intrinsics.
1088
+ // For each sub-block (bi), dequant 32 Q6_K values to int6 -> apply two sub-block scales ->
1089
+ // find max abs -> compute reflect_scale -> requant to int8 -> gather d with stride load.
1090
+ static int repack_q6_k_to_q8_0_32_bl(ggml_tensor * t,
1091
+ int interleave_block,
1092
+ const void * GGML_RESTRICT data,
1093
+ size_t data_size) {
1094
+ GGML_ASSERT(t->type == GGML_TYPE_Q6_K);
1095
+ GGML_ASSERT(interleave_block == 32);
1096
+ GGML_ASSERT(QK_K / QK4_1 == 8);
1097
+
1098
+ constexpr int nrows_interleaved = 32;
1099
+
1100
+ block_q8_0x32 * dst = (block_q8_0x32 *) t->data;
1101
+ const block_q6_K * src = (const block_q6_K *) data;
1102
+ int nrow = ggml_nrows(t);
1103
+ int nblocks = t->ne[0] / QK_K;
1104
+
1105
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK_K != 0) {
1106
+ return -1;
1107
+ }
1108
+
1109
+ const ptrdiff_t row_stride = (ptrdiff_t) nblocks * sizeof(block_q6_K);
1110
+
1111
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
1112
+ for (int64_t x = 0; x < nblocks; x++) {
1113
+ for (int bi = 0; bi < 8; bi++) {
1114
+ // --- 1) Gather 32 d values with stride load ---
1115
+ // We need to compute reflect_scale per row first, so gather d later.
1116
+ // Process each row: dequant Q6_K sub-block -> requant to Q8_0
1117
+ for (int i = 0; i < nrows_interleaved; i++) {
1118
+ const block_q6_K * src_blk = &src[x + i * nblocks];
1119
+ const uint8_t * q4 = src_blk->ql + 64 * (bi / 4);
1120
+ const uint8_t * qh = src_blk->qh + 32 * (bi / 4);
1121
+ const int8_t * scales = src_blk->scales;
1122
+ float d = GGML_FP16_TO_FP32(src_blk->d);
1123
+
1124
+ int8_t bi_idx = bi % 4;
1125
+
1126
+ // --- Dequant 32 Q6_K values to int6 (range [-32, 31]) using RVV ---
1127
+ // vl = 32 for e8m2 (VLEN=256) or loop for smaller VLEN
1128
+ const size_t vl16 = __riscv_vsetvl_e8m1(16);
1129
+
1130
+ vint8m1_t va_lo, va_hi; // 16 elements each
1131
+
1132
+ if (bi_idx == 0) {
1133
+ // a[l] = (q4[l] & 0xF) | (((qh[l] >> 0) & 3) << 4) - 32
1134
+ vuint8m1_t vq4_lo = __riscv_vle8_v_u8m1(q4, vl16);
1135
+ vuint8m1_t vq4_hi = __riscv_vle8_v_u8m1(q4 + 16, vl16);
1136
+ vuint8m1_t vqh_lo = __riscv_vle8_v_u8m1(qh, vl16);
1137
+ vuint8m1_t vqh_hi = __riscv_vle8_v_u8m1(qh + 16, vl16);
1138
+
1139
+ vuint8m1_t vlo4_lo = __riscv_vand_vx_u8m1(vq4_lo, 0x0F, vl16);
1140
+ vuint8m1_t vlo4_hi = __riscv_vand_vx_u8m1(vq4_hi, 0x0F, vl16);
1141
+ vuint8m1_t vh_lo = __riscv_vsll_vx_u8m1(__riscv_vand_vx_u8m1(vqh_lo, 0x03, vl16), 4, vl16);
1142
+ vuint8m1_t vh_hi = __riscv_vsll_vx_u8m1(__riscv_vand_vx_u8m1(vqh_hi, 0x03, vl16), 4, vl16);
1143
+
1144
+ vuint8m1_t vcomb_lo = __riscv_vor_vv_u8m1(vlo4_lo, vh_lo, vl16);
1145
+ vuint8m1_t vcomb_hi = __riscv_vor_vv_u8m1(vlo4_hi, vh_hi, vl16);
1146
+
1147
+ va_lo = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(vcomb_lo), 32, vl16);
1148
+ va_hi = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(vcomb_hi), 32, vl16);
1149
+ } else if (bi_idx == 1) {
1150
+ // a[l] = (q4[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4) - 32
1151
+ vuint8m1_t vq4_lo = __riscv_vle8_v_u8m1(q4 + 32, vl16);
1152
+ vuint8m1_t vq4_hi = __riscv_vle8_v_u8m1(q4 + 48, vl16);
1153
+ vuint8m1_t vqh_lo = __riscv_vle8_v_u8m1(qh, vl16);
1154
+ vuint8m1_t vqh_hi = __riscv_vle8_v_u8m1(qh + 16, vl16);
1155
+
1156
+ vuint8m1_t vlo4_lo = __riscv_vand_vx_u8m1(vq4_lo, 0x0F, vl16);
1157
+ vuint8m1_t vlo4_hi = __riscv_vand_vx_u8m1(vq4_hi, 0x0F, vl16);
1158
+ vuint8m1_t vh_lo = __riscv_vsll_vx_u8m1(
1159
+ __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(vqh_lo, 2, vl16), 0x03, vl16), 4, vl16);
1160
+ vuint8m1_t vh_hi = __riscv_vsll_vx_u8m1(
1161
+ __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(vqh_hi, 2, vl16), 0x03, vl16), 4, vl16);
1162
+
1163
+ vuint8m1_t vcomb_lo = __riscv_vor_vv_u8m1(vlo4_lo, vh_lo, vl16);
1164
+ vuint8m1_t vcomb_hi = __riscv_vor_vv_u8m1(vlo4_hi, vh_hi, vl16);
1165
+
1166
+ va_lo = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(vcomb_lo), 32, vl16);
1167
+ va_hi = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(vcomb_hi), 32, vl16);
1168
+ } else if (bi_idx == 2) {
1169
+ // a[l] = (q4[l] >> 4) | (((qh[l] >> 4) & 3) << 4) - 32
1170
+ vuint8m1_t vq4_lo = __riscv_vle8_v_u8m1(q4, vl16);
1171
+ vuint8m1_t vq4_hi = __riscv_vle8_v_u8m1(q4 + 16, vl16);
1172
+ vuint8m1_t vqh_lo = __riscv_vle8_v_u8m1(qh, vl16);
1173
+ vuint8m1_t vqh_hi = __riscv_vle8_v_u8m1(qh + 16, vl16);
1174
+
1175
+ vuint8m1_t vhi4_lo = __riscv_vsrl_vx_u8m1(vq4_lo, 4, vl16);
1176
+ vuint8m1_t vhi4_hi = __riscv_vsrl_vx_u8m1(vq4_hi, 4, vl16);
1177
+ vuint8m1_t vh_lo = __riscv_vsll_vx_u8m1(
1178
+ __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(vqh_lo, 4, vl16), 0x03, vl16), 4, vl16);
1179
+ vuint8m1_t vh_hi = __riscv_vsll_vx_u8m1(
1180
+ __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(vqh_hi, 4, vl16), 0x03, vl16), 4, vl16);
1181
+
1182
+ vuint8m1_t vcomb_lo = __riscv_vor_vv_u8m1(vhi4_lo, vh_lo, vl16);
1183
+ vuint8m1_t vcomb_hi = __riscv_vor_vv_u8m1(vhi4_hi, vh_hi, vl16);
1184
+
1185
+ va_lo = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(vcomb_lo), 32, vl16);
1186
+ va_hi = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(vcomb_hi), 32, vl16);
1187
+ } else { // bi_idx == 3
1188
+ // a[l] = (q4[l+32] >> 4) | (((qh[l] >> 6) & 3) << 4) - 32
1189
+ vuint8m1_t vq4_lo = __riscv_vle8_v_u8m1(q4 + 32, vl16);
1190
+ vuint8m1_t vq4_hi = __riscv_vle8_v_u8m1(q4 + 48, vl16);
1191
+ vuint8m1_t vqh_lo = __riscv_vle8_v_u8m1(qh, vl16);
1192
+ vuint8m1_t vqh_hi = __riscv_vle8_v_u8m1(qh + 16, vl16);
1193
+
1194
+ vuint8m1_t vhi4_lo = __riscv_vsrl_vx_u8m1(vq4_lo, 4, vl16);
1195
+ vuint8m1_t vhi4_hi = __riscv_vsrl_vx_u8m1(vq4_hi, 4, vl16);
1196
+ vuint8m1_t vh_lo = __riscv_vsll_vx_u8m1(
1197
+ __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(vqh_lo, 6, vl16), 0x03, vl16), 4, vl16);
1198
+ vuint8m1_t vh_hi = __riscv_vsll_vx_u8m1(
1199
+ __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(vqh_hi, 6, vl16), 0x03, vl16), 4, vl16);
1200
+
1201
+ vuint8m1_t vcomb_lo = __riscv_vor_vv_u8m1(vhi4_lo, vh_lo, vl16);
1202
+ vuint8m1_t vcomb_hi = __riscv_vor_vv_u8m1(vhi4_hi, vh_hi, vl16);
1203
+
1204
+ va_lo = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(vcomb_lo), 32, vl16);
1205
+ va_hi = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(vcomb_hi), 32, vl16);
1206
+ }
1207
+
1208
+ // --- Widen to i16 for scaled abs computation ---
1209
+ float scale_0 = scales[bi * 2 + 0] * d;
1210
+ float scale_1 = scales[bi * 2 + 1] * d;
1211
+
1212
+ // Widen i8 -> i16 -> f32 for abs*scale computation
1213
+ vint16m2_t va_lo_w = __riscv_vsext_vf2_i16m2(va_lo, vl16);
1214
+ vint16m2_t va_hi_w = __riscv_vsext_vf2_i16m2(va_hi, vl16);
1215
+
1216
+ // Compute |a[l] * scale_0| for lo half, |a[l] * scale_1| for hi half
1217
+ vfloat32m4_t vf_lo = __riscv_vfcvt_f_x_v_f32m4(__riscv_vsext_vf2_i32m4(va_lo_w, vl16), vl16);
1218
+ vfloat32m4_t vf_hi = __riscv_vfcvt_f_x_v_f32m4(__riscv_vsext_vf2_i32m4(va_hi_w, vl16), vl16);
1219
+
1220
+ vfloat32m4_t vabs_lo = __riscv_vfabs_v_f32m4(__riscv_vfmul_vf_f32m4(vf_lo, scale_0, vl16), vl16);
1221
+ vfloat32m4_t vabs_hi = __riscv_vfabs_v_f32m4(__riscv_vfmul_vf_f32m4(vf_hi, scale_1, vl16), vl16);
1222
+
1223
+ // Find max abs across both halves
1224
+ vfloat32m4_t vabs_max = __riscv_vfmax_vv_f32m4(vabs_lo, vabs_hi, vl16);
1225
+
1226
+ // Reduce to scalar max
1227
+ vfloat32m1_t vzero = __riscv_vfmv_v_f_f32m1(0.0f, 1);
1228
+ vfloat32m1_t vmax_red = __riscv_vfredmax_vs_f32m4_f32m1(vabs_max, vzero, vl16);
1229
+ float a_max_abs = __riscv_vfmv_f_s_f32m1_f32(vmax_red);
1230
+
1231
+ float reflect_scale = a_max_abs / 127.0f;
1232
+ float reflect_scale_0 = scale_0 / reflect_scale;
1233
+ float reflect_scale_1 = scale_1 / reflect_scale;
1234
+
1235
+ // --- Requant: a[l] = clamp(nearbyint(a[l] * reflect_scale_x), -128, 127) ---
1236
+ vfloat32m4_t vscaled_lo = __riscv_vfmul_vf_f32m4(vf_lo, reflect_scale_0, vl16);
1237
+ vfloat32m4_t vscaled_hi = __riscv_vfmul_vf_f32m4(vf_hi, reflect_scale_1, vl16);
1238
+
1239
+ // fcvt.x rounds to nearest (using current rounding mode)
1240
+ vint32m4_t vi_lo = __riscv_vfcvt_x_f_v_i32m4(vscaled_lo, vl16);
1241
+ vint32m4_t vi_hi = __riscv_vfcvt_x_f_v_i32m4(vscaled_hi, vl16);
1242
+
1243
+ // Clamp to [-128, 127]
1244
+ vi_lo = __riscv_vmax_vx_i32m4(vi_lo, -128, vl16);
1245
+ vi_lo = __riscv_vmin_vx_i32m4(vi_lo, 127, vl16);
1246
+ vi_hi = __riscv_vmax_vx_i32m4(vi_hi, -128, vl16);
1247
+ vi_hi = __riscv_vmin_vx_i32m4(vi_hi, 127, vl16);
1248
+
1249
+ // Narrow i32 -> i16 -> i8
1250
+ vint16m2_t vi16_lo = __riscv_vncvt_x_x_w_i16m2(vi_lo, vl16);
1251
+ vint16m2_t vi16_hi = __riscv_vncvt_x_x_w_i16m2(vi_hi, vl16);
1252
+ vint8m1_t vi8_lo = __riscv_vncvt_x_x_w_i8m1(vi16_lo, vl16);
1253
+ vint8m1_t vi8_hi = __riscv_vncvt_x_x_w_i8m1(vi16_hi, vl16);
1254
+
1255
+ // Store d and qs directly into dst block
1256
+ dst->d[i] = GGML_FP32_TO_FP16(reflect_scale);
1257
+ int8_t * dq = (int8_t *) dst->qs + i * QK8_0;
1258
+ __riscv_vse8_v_i8m1(dq, vi8_lo, vl16);
1259
+ __riscv_vse8_v_i8m1(dq + 16, vi8_hi, vl16);
1260
+ }
1261
+ dst++;
1262
+ }
1263
+ }
1264
+ src += nrows_interleaved * nblocks;
1265
+ }
1266
+ return 0;
1267
+
1268
+ GGML_UNUSED(data_size);
1269
+ }
1270
+
1271
+ static int repack_q8_0_to_q8_0_32_bl_ref(ggml_tensor * t,
1272
+ int interleave_block,
1273
+ const void * GGML_RESTRICT data,
1274
+ size_t data_size) {
1275
+ GGML_ASSERT(t->type == GGML_TYPE_Q8_0);
1276
+ GGML_ASSERT(interleave_block == 32); // unused
1277
+
1278
+ constexpr int nrows_interleaved = 32;
1279
+
1280
+ block_q8_0x32 * dst = (block_q8_0x32 *) t->data;
1281
+ const block_q8_0 * src = (const block_q8_0 *) data;
1282
+ block_q8_0 dst_tmp[32];
1283
+ int nrow = ggml_nrows(t);
1284
+ int nblocks = t->ne[0] / QK8_0;
1285
+
1286
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q8_0));
1287
+
1288
+ if (t->ne[0] % QK8_0 != 0) {
1289
+ return -1;
1290
+ }
1291
+
1292
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
1293
+ int64_t nrows_real = std::min((int64_t) nrow - b, (int64_t) nrows_interleaved);
1294
+ for (int64_t x = 0; x < nblocks; x++) {
1295
+ int i = 0;
1296
+ for (; i < nrows_real; i++) {
1297
+ dst_tmp[i] = src[x + i * nblocks];
1298
+ }
1299
+ for (; i < nrows_interleaved; i++) {
1300
+ memset(&dst_tmp[i], 0, sizeof(block_q8_0));
1301
+ }
1302
+ *dst++ = make_block_q8_0x32(dst_tmp, interleave_block);
1303
+ }
1304
+ src += nrows_interleaved * nblocks;
1305
+ }
1306
+ return 0;
1307
+
1308
+ GGML_UNUSED(data_size);
1309
+ }
1310
+
1311
+ // RVV optimized version of repack_q8_0_to_q8_0_32_bl
1312
+ // Eliminates the intermediate dst_tmp buffer and vectorizes scale gather + qs copy.
1313
+ static int repack_q8_0_to_q8_0_32_bl(ggml_tensor * t,
1314
+ int interleave_block,
1315
+ const void * GGML_RESTRICT data,
1316
+ size_t data_size) {
1317
+ GGML_ASSERT(t->type == GGML_TYPE_Q8_0);
1318
+ GGML_ASSERT(interleave_block == 32);
1319
+
1320
+ constexpr int nrows_interleaved = 32;
1321
+
1322
+ block_q8_0x32 * dst = (block_q8_0x32 *) t->data;
1323
+ const block_q8_0 * src = (const block_q8_0 *) data;
1324
+ int nrow = ggml_nrows(t);
1325
+ int nblocks = t->ne[0] / QK8_0;
1326
+
1327
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q8_0));
1328
+
1329
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK8_0 != 0) {
1330
+ return -1;
1331
+ }
1332
+
1333
+ const ptrdiff_t row_stride = (ptrdiff_t) nblocks * sizeof(block_q8_0);
1334
+
1335
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
1336
+ for (int64_t x = 0; x < nblocks; x++) {
1337
+ const block_q8_0 * col_src = src + x;
1338
+
1339
+ // --- 1) Gather 32 scale values (ggml_half d) with stride load ---
1340
+ {
1341
+ const uint8_t * d_base = (const uint8_t *) &col_src->d;
1342
+ ggml_half * d_dst = dst->d;
1343
+ size_t remaining = 32;
1344
+ size_t offset = 0;
1345
+ while (remaining > 0) {
1346
+ size_t vl = __riscv_vsetvl_e16m1(remaining);
1347
+ vuint16m1_t vd =
1348
+ __riscv_vlse16_v_u16m1((const uint16_t *) (d_base + offset * row_stride), row_stride, vl);
1349
+ __riscv_vse16_v_u16m1((uint16_t *) (d_dst + offset), vd, vl);
1350
+ offset += vl;
1351
+ remaining -= vl;
1352
+ }
1353
+ }
1354
+
1355
+ // --- 2) Copy qs for each of the 32 rows (32 bytes per row) ---
1356
+ {
1357
+ for (int i = 0; i < 32; i++) {
1358
+ const int8_t * sq = col_src[i * nblocks].qs;
1359
+ int8_t * dq = (int8_t *) dst->qs + i * QK8_0;
1360
+
1361
+ size_t len = QK8_0;
1362
+ size_t idx = 0;
1363
+ while (len > 0) {
1364
+ size_t vl = __riscv_vsetvl_e8m2(len);
1365
+ vint8m2_t vs = __riscv_vle8_v_i8m2(sq + idx, vl);
1366
+ __riscv_vse8_v_i8m2(dq + idx, vs, vl);
1367
+ idx += vl;
1368
+ len -= vl;
1369
+ }
1370
+ }
1371
+ }
1372
+
1373
+ dst++;
1374
+ }
1375
+ src += nrows_interleaved * nblocks;
1376
+ }
1377
+ return 0;
1378
+
1379
+ GGML_UNUSED(data_size);
1380
+ }
1381
+
1382
+ static void convert_mxfp4_to_5bit(const block_mxfp4 & src, spacemit_kernels::nrow_block_mxfp4<1> & dst) {
1383
+ dst.e[0] = src.e;
1384
+
1385
+ // Decode all 32 mxfp4 values to signed integers via kvalues_mxfp4
1386
+ int8_t vals[32];
1387
+ for (int j = 0; j < QK_MXFP4 / 2; j++) {
1388
+ vals[j] = kvalues_mxfp4[src.qs[j] & 0xF];
1389
+ vals[j + QK_MXFP4 / 2] = kvalues_mxfp4[src.qs[j] >> 4];
1390
+ }
1391
+
1392
+ // vals [b0, b1, b2, b3, ..., b30, b31]
1393
+ // Pack abs into qs with reorder: [b0,b1]..[b14,b15]..[b30,b31]
1394
+ for (int j = 0; j < QK_MXFP4 / 2; j++) {
1395
+ uint8_t lo0 = static_cast<uint8_t>(std::abs(vals[j * 2]));
1396
+ uint8_t lo1 = static_cast<uint8_t>(std::abs(vals[j * 2 + 1]));
1397
+ dst.qs[j] = (lo0 & 0x0F) | ((lo1 & 0x0F) << 4);
1398
+ }
1399
+
1400
+ // Pack sign bits into qh[4] (32 bits total, 1 bit per weight)
1401
+ // reorder: [0,1,2,...,15,16,17,...,31] after the qs reorder above
1402
+ uint32_t sign_bits = 0;
1403
+ for (int j = 0; j < 32; j++) {
1404
+ if (vals[j] < 0) {
1405
+ sign_bits |= (1u << j);
1406
+ }
1407
+ }
1408
+ memcpy(dst.qh, &sign_bits, 4);
1409
+ }
1410
+
1411
+ static spacemit_kernels::nrow_block_mxfp4<32> make_block_mxfp4x32(spacemit_kernels::nrow_block_mxfp4<1> * in,
1412
+ unsigned int blck_size_interleave) {
1413
+ spacemit_kernels::nrow_block_mxfp4<32> out;
1414
+ GGML_ASSERT(QK_MXFP4 / blck_size_interleave == 1);
1415
+ GGML_UNUSED(blck_size_interleave);
1416
+
1417
+ for (int i = 0; i < 32; i++) {
1418
+ out.e[i] = in[i].e[0];
1419
+ }
1420
+
1421
+ // qs: copy per-row 16 bytes
1422
+ for (int i = 0; i < 32; i++) {
1423
+ memcpy(out.qs + i * 16, in[i].qs, 16);
1424
+ }
1425
+
1426
+ // qh: copy per-row 4 bytes
1427
+ for (int i = 0; i < 32; i++) {
1428
+ memcpy(out.qh + i * 4, in[i].qh, 4);
1429
+ }
1430
+
1431
+ return out;
1432
+ }
1433
+
1434
+ static int repack_mxfp4_to_mxfp4_32_bl(ggml_tensor * t,
1435
+ int interleave_block,
1436
+ const void * GGML_RESTRICT data,
1437
+ size_t data_size) {
1438
+ GGML_ASSERT(t->type == GGML_TYPE_MXFP4);
1439
+ GGML_ASSERT(interleave_block == 32);
1440
+
1441
+ constexpr int nrows_interleaved = 32;
1442
+
1443
+ spacemit_kernels::nrow_block_mxfp4<32> * dst = (spacemit_kernels::nrow_block_mxfp4<32> *) t->data;
1444
+ const block_mxfp4 * src = (const block_mxfp4 *) data;
1445
+ spacemit_kernels::nrow_block_mxfp4<1> dst_tmp[32];
1446
+ int nrow = ggml_nrows(t);
1447
+ int nblocks = t->ne[0] / QK_MXFP4;
1448
+
1449
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_mxfp4));
1450
+
1451
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK_MXFP4 != 0) {
1452
+ return -1;
1453
+ }
1454
+
1455
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
1456
+ for (int64_t x = 0; x < nblocks; x++) {
1457
+ for (int i = 0; i < nrows_interleaved; i++) {
1458
+ convert_mxfp4_to_5bit(src[x + i * nblocks], dst_tmp[i]);
1459
+ }
1460
+ *dst++ = make_block_mxfp4x32(dst_tmp, interleave_block);
1461
+ }
1462
+ src += nrows_interleaved * nblocks;
1463
+ }
1464
+ return 0;
1465
+ }
1466
+
1467
+ static spacemit_kernels::nrow_block_q5_1<32> make_block_q5_1x32(spacemit_kernels::nrow_block_q5_1<1> * in,
1468
+ unsigned int blck_size_interleave) {
1469
+ spacemit_kernels::nrow_block_q5_1<32> out;
1470
+ GGML_ASSERT(QK5_1 / blck_size_interleave == 1);
1471
+ GGML_UNUSED(blck_size_interleave);
1472
+
1473
+ for (int i = 0; i < 32; i++) {
1474
+ out.scales16[i] = in[i].scales16[0];
1475
+ out.zp[i] = in[i].zp[0];
1476
+ }
1477
+
1478
+ // qs: low 4 bits, reorder from [b0,b16],[b1,b17]... to [b0,b1]...[b14,b15] and [b16,b17]...[b30,b31]
1479
+ for (int i = 0; i < 32; i++) {
1480
+ // low half [0..15]
1481
+ for (int j = 0; j < QK5_1 / 4; j++) {
1482
+ out.qs[i * QK5_1 / 2 + j] = (in[i].qs[j * 2] & 0x0F) | ((in[i].qs[j * 2 + 1] & 0x0F) << 4);
1483
+ }
1484
+ // high half [16..31]
1485
+ for (int j = 0; j < QK5_1 / 4; j++) {
1486
+ out.qs[i * QK5_1 / 2 + QK5_1 / 4 + j] = ((in[i].qs[j * 2] & 0xF0) >> 4) | (in[i].qs[j * 2 + 1] & 0xF0);
1487
+ }
1488
+ }
1489
+
1490
+ // qh: 5th bit, copy directly
1491
+ for (int i = 0; i < 32; i++) {
1492
+ for (int j = 0; j < 4; j++) {
1493
+ out.qh[i * 4 + j] = in[i].qh[j];
1494
+ }
1495
+ }
1496
+
1497
+ return out;
1498
+ }
1499
+
1500
+ static spacemit_kernels::nrow_block_q5_0<32> make_block_q5_0x32(spacemit_kernels::nrow_block_q5_0<1> * in,
1501
+ unsigned int blck_size_interleave) {
1502
+ spacemit_kernels::nrow_block_q5_0<32> out;
1503
+ GGML_ASSERT(QK5_0 / blck_size_interleave == 1);
1504
+ GGML_UNUSED(blck_size_interleave);
1505
+
1506
+ for (int i = 0; i < 32; i++) {
1507
+ out.scales16[i] = in[i].scales16[0];
1508
+ }
1509
+
1510
+ // qs: low 4 bits, reorder from [b0,b16],[b1,b17]... to [b0,b1]...[b14,b15] and [b16,b17]...[b30,b31]
1511
+ for (int i = 0; i < 32; i++) {
1512
+ // low half [0..15]
1513
+ for (int j = 0; j < QK5_0 / 4; j++) {
1514
+ out.qs[i * QK5_0 / 2 + j] = (in[i].qs[j * 2] & 0x0F) | ((in[i].qs[j * 2 + 1] & 0x0F) << 4);
1515
+ }
1516
+ // high half [16..31]
1517
+ for (int j = 0; j < QK5_0 / 4; j++) {
1518
+ out.qs[i * QK5_0 / 2 + QK5_0 / 4 + j] = ((in[i].qs[j * 2] & 0xF0) >> 4) | (in[i].qs[j * 2 + 1] & 0xF0);
1519
+ }
1520
+ }
1521
+
1522
+ // qh: 5th bit, copy directly
1523
+ for (int i = 0; i < 32; i++) {
1524
+ for (int j = 0; j < 4; j++) {
1525
+ out.qh[i * 4 + j] = in[i].qh[j];
1526
+ }
1527
+ }
1528
+
1529
+ return out;
1530
+ }
1531
+
1532
+ static int repack_q5_0_to_q5_0_32_bl(ggml_tensor * t,
1533
+ int interleave_block,
1534
+ const void * GGML_RESTRICT data,
1535
+ size_t data_size) {
1536
+ GGML_ASSERT(t->type == GGML_TYPE_Q5_0);
1537
+ GGML_ASSERT(interleave_block == 32); // unused
1538
+
1539
+ constexpr int nrows_interleaved = 32;
1540
+
1541
+ spacemit_kernels::nrow_block_q5_0<32> * dst = (spacemit_kernels::nrow_block_q5_0<32> *) t->data;
1542
+ const block_q5_0 * src = (const block_q5_0 *) data;
1543
+ spacemit_kernels::nrow_block_q5_0<1> dst_tmp[32];
1544
+ int nrow = ggml_nrows(t);
1545
+ int nblocks = t->ne[0] / QK5_0;
1546
+
1547
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q5_0));
1548
+
1549
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK5_0 != 0) {
1550
+ return -1;
1551
+ }
1552
+
1553
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
1554
+ for (int64_t x = 0; x < nblocks; x++) {
1555
+ for (int i = 0; i < nrows_interleaved; i++) {
1556
+ const block_q5_0 & s = src[x + i * nblocks];
1557
+
1558
+ dst_tmp[i].scales16[0] = s.d;
1559
+ memcpy(dst_tmp[i].qs, s.qs, sizeof(dst_tmp[i].qs));
1560
+ memcpy(dst_tmp[i].qh, s.qh, sizeof(dst_tmp[i].qh));
1561
+ }
1562
+ *dst++ = make_block_q5_0x32(dst_tmp, interleave_block);
1563
+ }
1564
+ src += nrows_interleaved * nblocks;
1565
+ }
1566
+ return 0;
1567
+ }
1568
+
1569
+ static int repack_q5_1_to_q5_1_32_bl(ggml_tensor * t,
1570
+ int interleave_block,
1571
+ const void * GGML_RESTRICT data,
1572
+ size_t data_size) {
1573
+ GGML_ASSERT(t->type == GGML_TYPE_Q5_1);
1574
+ GGML_ASSERT(interleave_block == 32); // unused
1575
+
1576
+ constexpr int nrows_interleaved = 32;
1577
+
1578
+ spacemit_kernels::nrow_block_q5_1<32> * dst = (spacemit_kernels::nrow_block_q5_1<32> *) t->data;
1579
+ const block_q5_1 * src = (const block_q5_1 *) data;
1580
+ spacemit_kernels::nrow_block_q5_1<1> dst_tmp[32];
1581
+ int nrow = ggml_nrows(t);
1582
+ int nblocks = t->ne[0] / QK5_1;
1583
+
1584
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q5_1));
1585
+
1586
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK5_1 != 0) {
1587
+ return -1;
1588
+ }
1589
+
1590
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
1591
+ for (int64_t x = 0; x < nblocks; x++) {
1592
+ for (int i = 0; i < nrows_interleaved; i++) {
1593
+ const block_q5_1 & s = src[x + i * nblocks];
1594
+
1595
+ float d = GGML_FP16_TO_FP32(s.GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d);
1596
+ float m = GGML_FP16_TO_FP32(s.GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.m);
1597
+
1598
+ if (d == 0.0f) {
1599
+ dst_tmp[i].scales16[0] = GGML_FP32_TO_FP16(std::fabs(m));
1600
+ dst_tmp[i].zp[0] = m < 0.0f ? 1 : 0;
1601
+ memset(dst_tmp[i].qh, 0, sizeof(dst_tmp[i].qh));
1602
+ memset(dst_tmp[i].qs, m > 0.0f ? 0x11 : 0x00, sizeof(dst_tmp[i].qs));
1603
+ continue;
1604
+ }
1605
+
1606
+ float mid = std::nearbyintf(-m / d);
1607
+ mid = std::min(31.0f, std::max(0.0f, mid));
1608
+
1609
+ dst_tmp[i].scales16[0] = GGML_FP32_TO_FP16(d);
1610
+ dst_tmp[i].zp[0] = static_cast<uint8_t>(mid);
1611
+
1612
+ // qs: copy low 4 bits directly (same nibble packing)
1613
+ memcpy(dst_tmp[i].qs, s.qs, QK5_1 / 2);
1614
+
1615
+ // qh: copy 5th bit directly
1616
+ memcpy(dst_tmp[i].qh, s.qh, 4);
1617
+ }
1618
+ *dst++ = make_block_q5_1x32(dst_tmp, interleave_block);
1619
+ }
1620
+ src += nrows_interleaved * nblocks;
1621
+ }
1622
+ return 0;
1623
+ }
1624
+
1625
+ static int repack_q5_k_to_q5_1_32_bl(ggml_tensor * t,
1626
+ int interleave_block,
1627
+ const void * GGML_RESTRICT data,
1628
+ size_t data_size) {
1629
+ GGML_ASSERT(t->type == GGML_TYPE_Q5_K);
1630
+ GGML_ASSERT(interleave_block == 32);
1631
+ GGML_ASSERT(QK_K / QK5_1 == 8);
1632
+
1633
+ constexpr int nrows_interleaved = 32;
1634
+
1635
+ spacemit_kernels::nrow_block_q5_1<32> * dst = (spacemit_kernels::nrow_block_q5_1<32> *) t->data;
1636
+ const block_q5_K * src = (const block_q5_K *) data;
1637
+ spacemit_kernels::nrow_block_q5_1<1> dst_tmp[32];
1638
+ int nrow = ggml_nrows(t);
1639
+ int nblocks = t->ne[0] / QK_K;
1640
+
1641
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % QK_K != 0) {
1642
+ return -1;
1643
+ }
1644
+
1645
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
1646
+ for (int64_t x = 0; x < nblocks; x++) {
1647
+ for (int j = 0; j < 8; j++) {
1648
+ for (int i = 0; i < nrows_interleaved; i++) {
1649
+ uint8_t sc, m;
1650
+ const float d = GGML_FP16_TO_FP32(src[x + i * nblocks].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d);
1651
+ const float min =
1652
+ GGML_FP16_TO_FP32(src[x + i * nblocks].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin);
1653
+ get_scale_min_k4(j, src[x + i * nblocks].scales, &sc, &m);
1654
+
1655
+ float d1 = d * sc;
1656
+ float m1 = min * m;
1657
+
1658
+ float mid = std::nearbyintf(m1 / d1);
1659
+ mid = std::min(31.0f, std::max(0.0f, mid));
1660
+ dst_tmp[i].scales16[0] = GGML_FP32_TO_FP16(d1);
1661
+ dst_tmp[i].zp[0] = static_cast<uint8_t>(mid);
1662
+
1663
+ // src -> [b0, b32] [b1, b33] ... [b31, b63]
1664
+ // dst -> [b0, b16] [b1, b17] ... [b15, b31] [b32, b48] [b33, b49] ... [b47, b63]
1665
+ const uint8_t * q = src[x + i * nblocks].qs + (j / 2) * QK5_1;
1666
+ if (j % 2 == 0) {
1667
+ for (int ii = 0; ii < 16; ii++) {
1668
+ dst_tmp[i].qs[ii] = (q[ii] & 0x0F) | ((q[ii + 16] & 0x0F) << 4);
1669
+ }
1670
+ } else {
1671
+ for (int ii = 0; ii < 16; ii++) {
1672
+ dst_tmp[i].qs[ii] = ((q[ii] & 0xF0) >> 4) | (q[ii + 16] & 0xF0);
1673
+ }
1674
+ }
1675
+
1676
+ // Extract the 5th bit (qh) for this sub-block
1677
+ // block_q5_K.qh[32]: for sub-block j, the 5th bit is at bit position j in qh[l]
1678
+ // qs was reordered: dst_qs maps to src weights [0,16,1,17,...,15,31]
1679
+ // So qh must follow the same reorder to stay aligned with qs
1680
+ // dst qh[4] = 32 bits for 32 weights in the reordered layout:
1681
+ // byte 0: weights 0..7 (from src_qh[0..7])
1682
+ // byte 1: weights 8..15 (from src_qh[8..15])
1683
+ // byte 2: weights 16..23 (from src_qh[16..23])
1684
+ // byte 3: weights 24..31 (from src_qh[24..31])
1685
+ const uint8_t * src_qh = src[x + i * nblocks].qh;
1686
+ for (int bi = 0; bi < 4; bi++) {
1687
+ uint8_t qh_byte = 0;
1688
+ for (int k = 0; k < 8; k++) {
1689
+ int src_idx = bi * 8 + k;
1690
+ qh_byte |= ((src_qh[src_idx] >> j) & 1) << k;
1691
+ }
1692
+ dst_tmp[i].qh[bi] = qh_byte;
1693
+ }
1694
+ }
1695
+ *dst++ = make_block_q5_1x32(dst_tmp, interleave_block);
1696
+ }
1697
+ }
1698
+ src += nrows_interleaved * nblocks;
1699
+ }
1700
+ return 0;
1701
+ }
1702
+
1703
+ namespace ggml::cpu::riscv64_spacemit {
1704
+
1705
+ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> int repack(ggml_tensor *, const void *, size_t);
1706
+
1707
+ template <> int repack<block_q4_0, 32, 16>(ggml_tensor * t, const void * data, size_t data_size) {
1708
+ return repack_q4_0_to_q4_0_16_bl(t, 16, data, data_size);
1709
+ }
1710
+
1711
+ template <> int repack<block_q4_1, 32, 16>(ggml_tensor * t, const void * data, size_t data_size) {
1712
+ return repack_q4_1_to_q4_1_16_bl(t, 16, data, data_size);
1713
+ }
1714
+
1715
+ template <> int repack<block_q4_K, 32, 16>(ggml_tensor * t, const void * data, size_t data_size) {
1716
+ return repack_q4_k_to_q4_1_16_bl(t, 16, data, data_size);
1717
+ }
1718
+
1719
+ template <> int repack<block_q2_K, 256, 32>(ggml_tensor * t, const void * data, size_t data_size) {
1720
+ return repack_q2_k_to_q2_k_32_bl(t, 32, data, data_size);
1721
+ }
1722
+
1723
+ template <> int repack<block_q3_K, 256, 32>(ggml_tensor * t, const void * data, size_t data_size) {
1724
+ return repack_q3_k_to_q3_k_32_bl(t, 32, data, data_size);
1725
+ }
1726
+
1727
+ template <> int repack<block_q4_0, 32, 32>(ggml_tensor * t, const void * data, size_t data_size) {
1728
+ #if 0
1729
+ return repack_q4_0_to_q4_0_32_bl_ref(t, 32, data, data_size);
1730
+ #else
1731
+ return repack_q4_0_to_q4_0_32_bl(t, 32, data, data_size);
1732
+ #endif
1733
+ }
1734
+
1735
+ template <> int repack<block_q4_0, 256, 32>(ggml_tensor * t, const void * data, size_t data_size) {
1736
+ #if 1
1737
+ return repack_q4_0_to_q4_0_256_32_bl_ref(t, 32, data, data_size);
1738
+ #else
1739
+ //return repack_q4_0_to_q4_0_256_32_bl(t, 32, data, data_size);
1740
+ #endif
1741
+ }
1742
+
1743
+ template <> int repack<block_q4_1, 32, 32>(ggml_tensor * t, const void * data, size_t data_size) {
1744
+ #if 0
1745
+ return repack_q4_1_to_q4_1_32_bl_ref(t, 32, data, data_size);
1746
+ #else
1747
+ return repack_q4_1_to_q4_1_32_bl(t, 32, data, data_size);
1748
+ #endif
1749
+ }
1750
+
1751
+ template <> int repack<block_q4_1, 256, 32>(ggml_tensor * t, const void * data, size_t data_size) {
1752
+ #if 1
1753
+ return repack_q4_0_to_q4_1_256_32_bl_ref(t, 32, data, data_size);
1754
+ #else
1755
+ return repack_q4_1_to_q4_1_256_32_bl(t, 32, data, data_size);
1756
+ #endif
1757
+ }
1758
+
1759
+ template <> int repack<block_q4_K, 32, 32>(ggml_tensor * t, const void * data, size_t data_size) {
1760
+ return repack_q4_k_to_q4_1_32_bl(t, 32, data, data_size);
1761
+ }
1762
+
1763
+ template <> int repack<block_q6_K, 32, 32>(ggml_tensor * t, const void * data, size_t data_size) {
1764
+ #if 1
1765
+ return repack_q6_k_to_q8_0_32_bl_ref(t, 32, data, data_size);
1766
+ #else
1767
+ return repack_q6_k_to_q8_0_32_bl(t, 32, data, data_size);
1768
+ #endif
1769
+ }
1770
+
1771
+ template <> int repack<block_q8_0, 32, 32>(ggml_tensor * t, const void * data, size_t data_size) {
1772
+ #if 1
1773
+ return repack_q8_0_to_q8_0_32_bl_ref(t, 32, data, data_size);
1774
+ #else
1775
+ return repack_q8_0_to_q8_0_32_bl(t, 32, data, data_size);
1776
+ #endif
1777
+ }
1778
+
1779
+ template <> int repack<block_mxfp4, 32, 32>(ggml_tensor * t, const void * data, size_t data_size) {
1780
+ return repack_mxfp4_to_mxfp4_32_bl(t, 32, data, data_size);
1781
+ }
1782
+
1783
+ template <> int repack<block_q5_0, 32, 32>(ggml_tensor * t, const void * data, size_t data_size) {
1784
+ return repack_q5_0_to_q5_0_32_bl(t, 32, data, data_size);
1785
+ }
1786
+
1787
+ template <> int repack<block_q5_1, 32, 32>(ggml_tensor * t, const void * data, size_t data_size) {
1788
+ return repack_q5_1_to_q5_1_32_bl(t, 32, data, data_size);
1789
+ }
1790
+
1791
+ template <> int repack<block_q5_K, 32, 32>(ggml_tensor * t, const void * data, size_t data_size) {
1792
+ return repack_q5_k_to_q5_1_32_bl(t, 32, data, data_size);
1793
+ }
1794
+
1795
+ } // namespace ggml::cpu::riscv64_spacemit