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 +1,1273 @@
1
+ #include "ggml-quants.h"
2
+
3
+ #include "ggml-common.h"
4
+ #include "ggml-impl.h"
5
+ #include "ggml-openvino-extra.h"
6
+ #include "ggml.h"
7
+
8
+ #include <algorithm>
9
+ #include <cassert>
10
+ #include <cmath>
11
+ #include <cstddef>
12
+ #include <cstdint>
13
+ #include <limits>
14
+ #include <memory>
15
+ #include <openvino/core/except.hpp>
16
+ #include <openvino/core/node.hpp>
17
+ #include <openvino/core/node_output.hpp>
18
+ #include <openvino/core/parallel.hpp>
19
+ #include <openvino/core/shape.hpp>
20
+ #include <openvino/core/type/element_type.hpp>
21
+ #include <openvino/core/type/element_type_traits.hpp>
22
+ #include <openvino/core/type/float16.hpp>
23
+ #include <openvino/core/type/float4_e2m1.hpp>
24
+ #include <openvino/core/type/float8_e8m0.hpp>
25
+ #include <openvino/op/add.hpp>
26
+ #include <openvino/op/constant.hpp>
27
+ #include <openvino/op/convert.hpp>
28
+ #include <openvino/op/multiply.hpp>
29
+ #include <openvino/op/reshape.hpp>
30
+ #include <openvino/op/subtract.hpp>
31
+ #include <openvino/op/util/attr_types.hpp>
32
+ #include <openvino/pass/constant_folding.hpp>
33
+ #include <openvino/runtime/tensor.hpp>
34
+ #include <string>
35
+ #include <vector>
36
+
37
+ void unpack_32_4(const uint8_t * data, uint8_t * dst) {
38
+ std::fill_n(dst, 16, 0);
39
+ for (int j = 0; j < 16; ++j) {
40
+ uint8_t x = (data[j] & 0x0F);
41
+ uint8_t y = (data[j] >> 4);
42
+ if (j % 2 != 0) {
43
+ x <<= 4;
44
+ y <<= 4;
45
+ }
46
+ dst[j / 2] |= x;
47
+ dst[8 + j / 2] |= y; // Last 16 weights are in the higher bits
48
+ }
49
+ }
50
+
51
+ static constexpr size_t MXFP4_BLOCK_SIZE = 32;
52
+ static constexpr size_t MXFP4_BLOCK_QS_SIZE = MXFP4_BLOCK_SIZE / 2;
53
+ static constexpr size_t MXFP4_BLOCK_BYTES = sizeof(uint8_t) + MXFP4_BLOCK_QS_SIZE;
54
+
55
+ static void pack_32_mxfp4_for_openvino(const uint8_t * data, uint8_t * dst) {
56
+ for (int j = 0; j < static_cast<int>(MXFP4_BLOCK_QS_SIZE); j += 2) {
57
+ const uint8_t v0 = data[j] & 0x0F;
58
+ const uint8_t v1 = (data[j + 1] & 0x0F) << 4;
59
+ const uint8_t v16 = data[j] >> 4;
60
+ const uint8_t v17 = data[j + 1] & 0xF0;
61
+ dst[j / 2] = v0 | v1;
62
+ dst[MXFP4_BLOCK_SIZE / 4 + j / 2] = v16 | v17;
63
+ }
64
+ }
65
+
66
+ void extract_mxfp4_data(const ggml_tensor * tensor, ov::Tensor & weights_arr, ov::Tensor & scales_arr) {
67
+ GGML_ASSERT(tensor->type == GGML_TYPE_MXFP4);
68
+ GGML_ASSERT(weights_arr.get_element_type() == ov::element::f4e2m1);
69
+ GGML_ASSERT(scales_arr.get_element_type() == ov::element::f8e8m0);
70
+
71
+ const auto * data = static_cast<const uint8_t *>(tensor->data);
72
+ auto * weights = static_cast<uint8_t *>(weights_arr.data());
73
+ auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f8e8m0>::value_type>();
74
+ const size_t n_blocks = scales_arr.get_size();
75
+
76
+ ov::parallel_for(n_blocks, [&](size_t i) {
77
+ const uint8_t * block = data + i * MXFP4_BLOCK_BYTES;
78
+ pack_32_mxfp4_for_openvino(block + sizeof(uint8_t), weights + i * MXFP4_BLOCK_QS_SIZE);
79
+ scales[i] = ov::float8_e8m0::from_bits(block[0]);
80
+ });
81
+ }
82
+
83
+ // Extracts (weight, scales, zp) from Q4_0 tensors.
84
+ // Data layout is: |16 bit scale|32 x 4bit weights|.
85
+ // When zp_arr is empty (symmetric), weights are stored as signed i4 (value - 8).
86
+ void extract_q4_0_data(const ggml_tensor * tensor,
87
+ ov::Tensor & weights_arr,
88
+ ov::Tensor & scales_arr,
89
+ ov::Tensor & zp_arr) {
90
+ const uint64_t bytes_per_block = 18; // 2 bytes scale, 32x0.5 byte weights
91
+
92
+ auto * data = static_cast<uint8_t *>(tensor->data);
93
+ auto * weights = static_cast<uint8_t *>(weights_arr.data());
94
+ auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
95
+
96
+ bool is_symmetric = (weights_arr.get_element_type() == ov::element::i4); // Signed i4 path
97
+
98
+ if (!is_symmetric) {
99
+ auto * zp = static_cast<uint8_t *>(zp_arr.data());
100
+ ov::parallel_for(scales_arr.get_size(), [&](size_t i) {
101
+ scales[i] = ov::float16::from_bits(*((uint16_t *) (data + i * bytes_per_block)));
102
+ // Pack two 4-bit zero points per byte
103
+ if (i % 2 == 0) {
104
+ zp[i / 2] = 8; // Lower nibble
105
+ } else {
106
+ zp[i / 2] |= (8 << 4); // Upper nibble
107
+ }
108
+ unpack_32_4(data + i * bytes_per_block + 2, weights + i * 16);
109
+ });
110
+ } else {
111
+ // Symmetric: unpack as u4 then convert to i4 by subtracting 8 (XOR each nibble)
112
+ ov::parallel_for(scales_arr.get_size(), [&](size_t i) {
113
+ scales[i] = ov::float16::from_bits(*((uint16_t *) (data + i * bytes_per_block)));
114
+ unpack_32_4(data + i * bytes_per_block + 2, weights + i * 16);
115
+ // Convert u4 to i4: subtract 8 from each nibble. XOR 0x88 flips each nibble by 8.
116
+ for (int j = 0; j < 16; ++j) {
117
+ weights[i * 16 + j] ^= 0x88;
118
+ }
119
+ });
120
+ }
121
+ }
122
+
123
+ // Extracts (weight, scales, zp) from Q4_1 tensors.
124
+ // Data layout is: |16 bit scale|16 bit min|32 x 4bit weights|.
125
+ void extract_q4_1_data(const ggml_tensor * tensor,
126
+ ov::Tensor & weights_arr,
127
+ ov::Tensor & scales_arr,
128
+ ov::Tensor & zp_arr,
129
+ bool use_bias) {
130
+ const uint64_t bytes_per_block = 20; // 2 bytes scale, 2 bytes min, 32x0.5 byte weights
131
+
132
+ auto * data = static_cast<uint8_t *>(tensor->data);
133
+ auto * weights = static_cast<uint8_t *>(weights_arr.data());
134
+ auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
135
+
136
+ if (use_bias) {
137
+ // Store bias (min) directly as f16 instead of computing u4 zero points
138
+ auto * bias = zp_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
139
+ ov::parallel_for(scales_arr.get_size(), [&](size_t i) {
140
+ float scale = static_cast<float>(ov::float16::from_bits(*((uint16_t *) (data + i * bytes_per_block))));
141
+ float min = static_cast<float>(ov::float16::from_bits(*((uint16_t *) (data + i * bytes_per_block + 2))));
142
+ scales[i] = ov::float16(scale);
143
+ bias[i] = ov::float16(min); // bias = min, dequant: w*s + bias
144
+ unpack_32_4(data + i * bytes_per_block + 4, weights + i * 16);
145
+ });
146
+ } else {
147
+ auto * zp = static_cast<uint8_t *>(zp_arr.data());
148
+ ov::parallel_for(scales_arr.get_size(), [&](size_t i) {
149
+ float scale = static_cast<float>(ov::float16::from_bits(*((uint16_t *) (data + i * bytes_per_block))));
150
+ float min = static_cast<float>(ov::float16::from_bits(*((uint16_t *) (data + i * bytes_per_block + 2))));
151
+ scales[i] = ov::float16(scale);
152
+ // zp = -min / scale (bias = min, so zp = -bias/scale)
153
+ uint8_t zp_val = (scale != 0.0f) ? (uint8_t) std::round(-min / scale) : 0;
154
+ // Pack two 4-bit zero points per byte
155
+ if (i % 2 == 0) {
156
+ zp[i / 2] = zp_val & 0x0F; // Lower nibble
157
+ } else {
158
+ zp[i / 2] |= (zp_val << 4); // Upper nibble
159
+ }
160
+ unpack_32_4(data + i * bytes_per_block + 4, weights + i * 16);
161
+ });
162
+ }
163
+ }
164
+
165
+ // Extracts (weight, scales, zp) from Q5_1 tensors.
166
+ // Data layout is: |16 bit scale|16 bit min|32 bit qh (5th bits)|32 x 4bit low nibbles|.
167
+ // Reconstructed quant q in [0,31]: q = (low nibble) | (qh_bit << 4). Dequant: w*d + m.
168
+ // Weights are stored as u8 (5-bit values do not fit u4), matching make_int8_weights.
169
+ void extract_q5_1_data(const ggml_tensor * tensor,
170
+ ov::Tensor & weights_arr,
171
+ ov::Tensor & scales_arr,
172
+ ov::Tensor & zp_arr,
173
+ bool use_bias) {
174
+ const uint64_t bytes_per_block = 24; // 2 scale + 2 min + 4 qh + 16 (32x0.5) weights
175
+ const int qk = 32;
176
+
177
+ auto * data = static_cast<uint8_t *>(tensor->data);
178
+ auto * weights = static_cast<uint8_t *>(weights_arr.data()); // u8 weights, one byte per weight
179
+ auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
180
+
181
+ // Read a 16-bit little-endian value without aliasing/const-qual violations.
182
+ auto read_u16 = [](const uint8_t * p) {
183
+ uint16_t v;
184
+ memcpy(&v, p, sizeof(v));
185
+ return v;
186
+ };
187
+
188
+ auto unpack_block = [&](const uint8_t * block, uint8_t * dst) {
189
+ uint32_t qh;
190
+ memcpy(&qh, block + 4, sizeof(uint32_t));
191
+ const uint8_t * qs = block + 8;
192
+ for (int j = 0; j < qk / 2; ++j) {
193
+ const uint8_t lo = qs[j] & 0x0F;
194
+ const uint8_t hi = qs[j] >> 4;
195
+ const uint8_t bit_lo = (qh >> j) & 1;
196
+ const uint8_t bit_hi = (qh >> (j + qk / 2)) & 1;
197
+ dst[j] = lo | (bit_lo << 4); // first 16 weights
198
+ dst[j + qk / 2] = hi | (bit_hi << 4); // last 16 weights
199
+ }
200
+ };
201
+
202
+ if (use_bias) {
203
+ // Store bias (min) directly as f16: dequant w*d + m
204
+ auto * bias = zp_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
205
+ ov::parallel_for(scales_arr.get_size(), [&](size_t i) {
206
+ const uint8_t * block = data + i * bytes_per_block;
207
+ float scale = static_cast<float>(ov::float16::from_bits(read_u16(block)));
208
+ float min = static_cast<float>(ov::float16::from_bits(read_u16(block + 2)));
209
+ scales[i] = ov::float16(scale);
210
+ bias[i] = ov::float16(min);
211
+ unpack_block(block, weights + i * qk);
212
+ });
213
+ } else {
214
+ auto * zp = static_cast<uint8_t *>(zp_arr.data()); // u8 zero points
215
+ ov::parallel_for(scales_arr.get_size(), [&](size_t i) {
216
+ const uint8_t * block = data + i * bytes_per_block;
217
+ float scale = static_cast<float>(ov::float16::from_bits(read_u16(block)));
218
+ float min = static_cast<float>(ov::float16::from_bits(read_u16(block + 2)));
219
+ scales[i] = ov::float16(scale);
220
+ // zp = -min / scale (dequant: (w - zp) * s == w*s + min)
221
+ zp[i] = (scale != 0.0f) ? (uint8_t) std::lround(-min / scale) : 0;
222
+ unpack_block(block, weights + i * qk);
223
+ });
224
+ }
225
+ }
226
+
227
+ // Extracts (weight, scales, zp) from Q8_0 tensors.
228
+ // Data layout is: |16 bit scale|32 x 8bit weights|.
229
+ // When zp_arr is empty (symmetric), weights are stored as signed i8 directly.
230
+ void extract_q8_0_data(const ggml_tensor * tensor,
231
+ ov::Tensor & weights_arr,
232
+ ov::Tensor & scales_arr,
233
+ ov::Tensor & zp_arr) {
234
+ const uint64_t weights_per_block = 32;
235
+ const uint64_t bytes_per_block = 34; // 2 bytes scale, 32x1 byte weights
236
+
237
+ auto * data = static_cast<uint8_t *>(tensor->data);
238
+ auto * weights = static_cast<uint8_t *>(weights_arr.data());
239
+ auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
240
+
241
+ bool is_symmetric = (weights_arr.get_element_type() == ov::element::i8); // Signed i8 path
242
+
243
+ if (!is_symmetric) {
244
+ auto * zp = static_cast<uint8_t *>(zp_arr.data());
245
+ ov::parallel_for(scales_arr.get_size(), [&](size_t i) {
246
+ uint8_t * block_data = data + i * bytes_per_block;
247
+ scales[i] = ov::float16::from_bits(*(uint16_t *) block_data);
248
+ zp[i] = 128;
249
+ for (size_t j = 0; j < weights_per_block; ++j) {
250
+ uint8_t x = block_data[j + 2];
251
+ x ^= 1 << 7; // Convert int8 to uint8 by flipping sign bit
252
+ weights[i * weights_per_block + j] = x;
253
+ }
254
+ });
255
+ } else {
256
+ // Symmetric: store original int8 values directly (no unsigned bias)
257
+ ov::parallel_for(scales_arr.get_size(), [&](size_t i) {
258
+ uint8_t * block_data = data + i * bytes_per_block;
259
+ scales[i] = ov::float16::from_bits(*(uint16_t *) block_data);
260
+ // Copy int8 weights as-is (the tensor element type is i8)
261
+ memcpy(weights + i * weights_per_block, block_data + 2, weights_per_block);
262
+ });
263
+ }
264
+ }
265
+
266
+ void unpack_256_4(const uint8_t * data, uint8_t * dst) {
267
+ // Initialize the output array with zeros
268
+ std::fill_n(dst, 128, 0);
269
+
270
+ for (size_t i = 0; i < 4; ++i) {
271
+ for (int j = 0; j < 32; ++j) {
272
+ uint8_t x = (data[i * 32 + j] & 0x0F);
273
+ uint8_t y = (data[i * 32 + j] >> 4);
274
+ if (j % 2 != 0) {
275
+ x <<= 4;
276
+ y <<= 4;
277
+ }
278
+ dst[i * 32 + j / 2] |= x;
279
+ dst[i * 32 + 16 + j / 2] |= y; // Last 16 weights are in the higher bits
280
+ }
281
+ }
282
+ }
283
+
284
+ void extract_q4_k_data(const ggml_tensor * tensor,
285
+ ov::Tensor & weights_arr,
286
+ ov::Tensor & scales_arr,
287
+ ov::Tensor & zp_arr,
288
+ bool use_bias) {
289
+ const uint64_t bytes_per_block = 2 + 2 + 12 + 128;
290
+ const uint64_t n_super_block = tensor->nb[3] / bytes_per_block;
291
+
292
+ auto * data = static_cast<uint8_t *>(tensor->data);
293
+ auto * weights = static_cast<uint8_t *>(weights_arr.data());
294
+ auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
295
+
296
+ // For bias path, zp_arr holds f16 bias values; for zp path, it holds packed u4 zero points
297
+ auto * zp_u4 = use_bias ? nullptr : static_cast<uint8_t *>(zp_arr.data());
298
+ auto * bias_f16 = use_bias ? zp_arr.data<ov::element_type_traits<ov::element::f16>::value_type>() : nullptr;
299
+
300
+ ov::parallel_for(n_super_block, [&](size_t i) {
301
+ uint8_t * block_data = data + i * bytes_per_block;
302
+
303
+ // Extract scale factors and offsets
304
+ float scale_scales = static_cast<float>(ov::float16::from_bits(*((uint16_t *) block_data)));
305
+ float scale_mins = static_cast<float>(ov::float16::from_bits(*((uint16_t *) block_data + 1)));
306
+
307
+ // Extract qs1 and qs2
308
+ uint8_t * qs1 = block_data + 4;
309
+
310
+ // Calculate scales
311
+ float scale_vals[8];
312
+ scale_vals[0] = scale_scales * static_cast<float>((*(qs1) & 0b111111));
313
+ scale_vals[1] = scale_scales * static_cast<float>((*(qs1 + 1) & 0b111111));
314
+ scale_vals[2] = scale_scales * static_cast<float>((*(qs1 + 2) & 0b111111));
315
+ scale_vals[3] = scale_scales * static_cast<float>((*(qs1 + 3) & 0b111111));
316
+ scale_vals[4] = scale_scales * static_cast<float>((*(qs1 + 8) & 0b00001111) | ((*(qs1) >> 6) << 4));
317
+ scale_vals[5] = scale_scales * static_cast<float>((*(qs1 + 9) & 0b00001111) | ((*(qs1 + 1) >> 6) << 4));
318
+ scale_vals[6] = scale_scales * static_cast<float>((*(qs1 + 10) & 0b00001111) | ((*(qs1 + 2) >> 6) << 4));
319
+ scale_vals[7] = scale_scales * static_cast<float>((*(qs1 + 11) & 0b00001111) | ((*(qs1 + 3) >> 6) << 4));
320
+
321
+ // Calculate min values (bias = -min)
322
+ float min_vals[8];
323
+ min_vals[0] = scale_mins * static_cast<float>((*(qs1 + 4) & 0b111111));
324
+ min_vals[1] = scale_mins * static_cast<float>((*(qs1 + 5) & 0b111111));
325
+ min_vals[2] = scale_mins * static_cast<float>((*(qs1 + 6) & 0b111111));
326
+ min_vals[3] = scale_mins * static_cast<float>((*(qs1 + 7) & 0b111111));
327
+ min_vals[4] = scale_mins * static_cast<float>((*(qs1 + 8) >> 4) | ((*(qs1 + 4) >> 6) << 4));
328
+ min_vals[5] = scale_mins * static_cast<float>((*(qs1 + 9) >> 4) | ((*(qs1 + 5) >> 6) << 4));
329
+ min_vals[6] = scale_mins * static_cast<float>((*(qs1 + 10) >> 4) | ((*(qs1 + 6) >> 6) << 4));
330
+ min_vals[7] = scale_mins * static_cast<float>((*(qs1 + 11) >> 4) | ((*(qs1 + 7) >> 6) << 4));
331
+
332
+ // Store scales and compute zero points or bias
333
+ for (int j = 0; j < 8; j++) {
334
+ scales[i * 8 + j] = ov::float16(scale_vals[j]);
335
+ if (use_bias) {
336
+ // Store bias = -min directly as f16, dequant: w*s + bias
337
+ bias_f16[i * 8 + j] = ov::float16(-min_vals[j]);
338
+ } else {
339
+ // zp = min / scale (since bias = -min and zp = -bias/scale)
340
+ uint8_t zp_val = (scale_vals[j] != 0.0f) ? (uint8_t) std::round(min_vals[j] / scale_vals[j]) : 0;
341
+ // Pack two 4-bit zero points per byte
342
+ size_t idx = i * 8 + j;
343
+ if (idx % 2 == 0) {
344
+ zp_u4[idx / 2] = zp_val & 0x0F;
345
+ } else {
346
+ zp_u4[idx / 2] |= (zp_val << 4);
347
+ }
348
+ }
349
+ }
350
+ unpack_256_4(block_data + 16, weights + i * 128);
351
+ });
352
+ }
353
+
354
+ void extract_q6_k_data(const ggml_tensor * tensor,
355
+ ov::Tensor & weights_arr,
356
+ ov::Tensor & scales_arr,
357
+ ov::Tensor & zp_arr) {
358
+ const uint64_t bytes_per_block = 128 + 64 + 16 + 2;
359
+ const uint64_t n_super_block = tensor->nb[3] / bytes_per_block;
360
+
361
+ auto * data = static_cast<uint8_t *>(tensor->data);
362
+ auto * weights = static_cast<uint8_t *>(weights_arr.data());
363
+ auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
364
+
365
+ bool is_symmetric = (weights_arr.get_element_type() == ov::element::i8); // Signed i8 path
366
+
367
+ if (!is_symmetric) {
368
+ auto * zp = static_cast<uint8_t *>(zp_arr.data());
369
+ ov::parallel_for(n_super_block, [&](size_t i) {
370
+ uint8_t * block_data = data + i * bytes_per_block;
371
+ float scale_factor = static_cast<float>(ov::float16::from_bits(*((uint16_t *) block_data + 104)));
372
+ for (size_t j = 0; j < 16; j++) {
373
+ scales[j + i * 16] =
374
+ ov::float16(scale_factor * static_cast<float>(*((int8_t *) (block_data + 128 + 64 + j))));
375
+ zp[j + i * 16] = 32;
376
+ }
377
+ uint8_t * ql = block_data;
378
+ uint8_t * qh = block_data + 128;
379
+ for (int64_t j = 0; j < 32; ++j) {
380
+ weights[i * 256 + j] = (ql[j] & 0xF) | (((qh[j] >> 0) & 3) << 4);
381
+ weights[i * 256 + j + 32] = (ql[32 + j] & 0xF) | (((qh[j] >> 2) & 3) << 4);
382
+ weights[i * 256 + j + 64] = (ql[j] >> 4) | (((qh[j] >> 4) & 3) << 4);
383
+ weights[i * 256 + j + 96] = (ql[32 + j] >> 4) | (((qh[j] >> 6) & 3) << 4);
384
+ weights[i * 256 + j + 128] = (ql[64 + j] & 0xF) | (((qh[32 + j] >> 0) & 3) << 4);
385
+ weights[i * 256 + j + 160] = (ql[96 + j] & 0xF) | (((qh[32 + j] >> 2) & 3) << 4);
386
+ weights[i * 256 + j + 192] = (ql[64 + j] >> 4) | (((qh[32 + j] >> 4) & 3) << 4);
387
+ weights[i * 256 + j + 224] = (ql[96 + j] >> 4) | (((qh[32 + j] >> 6) & 3) << 4);
388
+ }
389
+ });
390
+ } else {
391
+ // Symmetric: subtract 32 from each weight to store as signed i8
392
+ ov::parallel_for(n_super_block, [&](size_t i) {
393
+ uint8_t * block_data = data + i * bytes_per_block;
394
+ float scale_factor = static_cast<float>(ov::float16::from_bits(*((uint16_t *) block_data + 104)));
395
+ for (size_t j = 0; j < 16; j++) {
396
+ scales[j + i * 16] =
397
+ ov::float16(scale_factor * static_cast<float>(*((int8_t *) (block_data + 128 + 64 + j))));
398
+ }
399
+ uint8_t * ql = block_data;
400
+ uint8_t * qh = block_data + 128;
401
+ auto * signed_weights = reinterpret_cast<int8_t *>(weights);
402
+ for (int64_t j = 0; j < 32; ++j) {
403
+ signed_weights[i * 256 + j] = static_cast<int8_t>((ql[j] & 0xF) | (((qh[j] >> 0) & 3) << 4)) - 32;
404
+ signed_weights[i * 256 + j + 32] =
405
+ static_cast<int8_t>((ql[32 + j] & 0xF) | (((qh[j] >> 2) & 3) << 4)) - 32;
406
+ signed_weights[i * 256 + j + 64] = static_cast<int8_t>((ql[j] >> 4) | (((qh[j] >> 4) & 3) << 4)) - 32;
407
+ signed_weights[i * 256 + j + 96] =
408
+ static_cast<int8_t>((ql[32 + j] >> 4) | (((qh[j] >> 6) & 3) << 4)) - 32;
409
+ signed_weights[i * 256 + j + 128] =
410
+ static_cast<int8_t>((ql[64 + j] & 0xF) | (((qh[32 + j] >> 0) & 3) << 4)) - 32;
411
+ signed_weights[i * 256 + j + 160] =
412
+ static_cast<int8_t>((ql[96 + j] & 0xF) | (((qh[32 + j] >> 2) & 3) << 4)) - 32;
413
+ signed_weights[i * 256 + j + 192] =
414
+ static_cast<int8_t>((ql[64 + j] >> 4) | (((qh[32 + j] >> 4) & 3) << 4)) - 32;
415
+ signed_weights[i * 256 + j + 224] =
416
+ static_cast<int8_t>((ql[96 + j] >> 4) | (((qh[32 + j] >> 6) & 3) << 4)) - 32;
417
+ }
418
+ });
419
+ }
420
+ }
421
+
422
+ static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t * d, uint8_t * m) {
423
+ if (j < 4) {
424
+ *d = q[j] & 63;
425
+ *m = q[j + 4] & 63;
426
+ } else {
427
+ *d = (q[j + 4] & 0xF) | ((q[j - 4] >> 6) << 4);
428
+ *m = (q[j + 4] >> 4) | ((q[j - 0] >> 6) << 4);
429
+ }
430
+ }
431
+
432
+ void extract_q5_k_data(const ggml_tensor * tensor,
433
+ ov::Tensor & weights_arr,
434
+ ov::Tensor & scales_arr,
435
+ ov::Tensor & zp_arr,
436
+ bool use_bias) {
437
+ const uint64_t bytes_per_block = 4 + 12 + 32 + 128;
438
+ const uint64_t n_super_block = tensor->nb[3] / bytes_per_block;
439
+
440
+ auto * data = static_cast<uint8_t *>(tensor->data);
441
+ auto * weights = static_cast<uint8_t *>(weights_arr.data());
442
+ auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
443
+
444
+ // For bias path, zp_arr holds f16 bias values; for zp path, it holds u8 zero points
445
+ auto * zp_u8 = use_bias ? nullptr : static_cast<uint8_t *>(zp_arr.data());
446
+ auto * bias_f16 = use_bias ? zp_arr.data<ov::element_type_traits<ov::element::f16>::value_type>() : nullptr;
447
+
448
+ ov::parallel_for(n_super_block, [&](size_t i) {
449
+ uint8_t * block_data = data + i * bytes_per_block;
450
+
451
+ const float d = static_cast<float>(ov::float16::from_bits(*((uint16_t *) block_data)));
452
+ const float min_factor = static_cast<float>(ov::float16::from_bits(*((uint16_t *) block_data + 1)));
453
+
454
+ const uint8_t * scales_data = block_data + 4; // 12 bytes of scales
455
+ const uint8_t * qh = block_data + 4 + 12; // 32 bytes of high bits
456
+ const uint8_t * ql = block_data + 4 + 12 + 32; // 128 bytes of low bits
457
+
458
+ int is = 0;
459
+ uint8_t u1 = 1;
460
+ uint8_t u2 = 2;
461
+
462
+ // Process 2 blocks in one iteration
463
+ for (int j = 0; j < 256; j += 64) { // 256 = QK_K, so 4 iterations of 64
464
+ uint8_t sc;
465
+ uint8_t m;
466
+
467
+ // Get scale and min for first 32 elements
468
+ get_scale_min_k4(is + 0, scales_data, &sc, &m);
469
+ const float d1 = d * sc;
470
+ const float m1 = min_factor * m;
471
+
472
+ // Get scale and min for second 32 elements
473
+ get_scale_min_k4(is + 1, scales_data, &sc, &m);
474
+ const float d2 = d * sc;
475
+ const float m2 = min_factor * m;
476
+
477
+ scales[i * 8 + is] = ov::float16(d1);
478
+ scales[i * 8 + is + 1] = ov::float16(d2);
479
+ if (use_bias) {
480
+ // Store bias = -min directly as f16, dequant: w*s + bias
481
+ bias_f16[i * 8 + is] = ov::float16(-m1);
482
+ bias_f16[i * 8 + is + 1] = ov::float16(-m2);
483
+ } else {
484
+ // zp = min / scale (since bias = -min and zp = -bias/scale)
485
+ zp_u8[i * 8 + is] = (d1 != 0.0f) ? (uint8_t) std::round(m1 / d1) : 0;
486
+ zp_u8[i * 8 + is + 1] = (d2 != 0.0f) ? (uint8_t) std::round(m2 / d2) : 0;
487
+ }
488
+
489
+ // Extract weights for first 32 elements (matching deq formula exactly)
490
+ for (int l = 0; l < 32; ++l) {
491
+ weights[i * 256 + j + l] = (ql[l] & 0xF) + ((qh[l] & u1) ? 16 : 0);
492
+ }
493
+
494
+ // Extract weights for second 32 elements
495
+ for (int l = 0; l < 32; ++l) {
496
+ weights[i * 256 + j + l + 32] = (ql[l] >> 4) + ((qh[l] & u2) ? 16 : 0);
497
+ }
498
+
499
+ ql += 32;
500
+ is += 2;
501
+ u1 <<= 2;
502
+ u2 <<= 2;
503
+ }
504
+ });
505
+ }
506
+
507
+ // TODO Reorder for make_intX_weights
508
+
509
+ // If for_gather_matmul is true, weight may be N-D (e.g. 3D MoE expert weights [n_expert, rows, cols]).
510
+ // The dequantization chain below is built as usual but left in f16 (no final Convert to f32) --
511
+ // ov::pass::MarkDequantization (registered in translate_session.cpp) marks the chain so it survives
512
+ // model-build-time ConstantFolding. mul_mat_id.cpp constructs ov::op::internal::GatherMatmul directly
513
+ // on top of the resulting f16 chain.
514
+ ov::Output<ov::Node> make_int8_weights(ov::Tensor & weight,
515
+ ov::Tensor & scales,
516
+ ov::Tensor & zp,
517
+ size_t group_size,
518
+ bool use_bias,
519
+ bool for_gather_matmul) {
520
+ ov::Shape orig_shape = weight.get_shape();
521
+ bool is_signed = (weight.get_element_type() == ov::element::i8); // Symmetric: signed weights, no ZP
522
+
523
+ // Expand dimensions for scales and zp/bias
524
+ auto scale_shape = scales.get_shape();
525
+
526
+ // Group the innermost (last) dimension. For 2D weights [rows, cols] this yields
527
+ // [rows, cols/group_size, group_size]; for 3D MoE experts [n_expert, rows, cols] this yields
528
+ // [n_expert, rows, cols/group_size, group_size].
529
+ ov::Shape packed_shape = orig_shape;
530
+ packed_shape.back() /= group_size;
531
+ packed_shape.push_back(group_size);
532
+ const size_t group_dim = packed_shape.size() - 2;
533
+
534
+ if (packed_shape[group_dim] == 1) {
535
+ // Requantized channel-wise case
536
+ packed_shape.erase(packed_shape.begin() + group_dim);
537
+ } else {
538
+ scale_shape.push_back(1);
539
+ scales.set_shape(scale_shape);
540
+ if (!is_signed && zp.get_size() > 0) {
541
+ auto zp_shape = zp.get_shape();
542
+ zp_shape.push_back(1);
543
+ zp.set_shape(zp_shape);
544
+ }
545
+ }
546
+
547
+ auto scales_f16 = std::make_shared<ov::op::v0::Constant>(scales);
548
+
549
+ ov::Output<ov::Node> result;
550
+ if (is_signed) {
551
+ // Signed path: q * s (no zero point subtraction needed)
552
+ auto weights_node = std::make_shared<ov::op::v0::Constant>(ov::element::i8, packed_shape,
553
+ static_cast<uint8_t *>(weight.data()), nullptr);
554
+ weights_node->get_rt_info()["__gguf_tensor_holder"] = weight;
555
+ auto weights_f16 = std::make_shared<ov::op::v0::Convert>(weights_node, ov::element::f16);
556
+ auto mul = std::make_shared<ov::op::v1::Multiply>(weights_f16, scales_f16, ov::op::AutoBroadcastType::NUMPY);
557
+ result = mul;
558
+ } else {
559
+ // Unsigned path
560
+ auto weights_node = std::make_shared<ov::op::v0::Constant>(ov::element::u8, packed_shape,
561
+ static_cast<uint8_t *>(weight.data()), nullptr);
562
+ weights_node->get_rt_info()["__gguf_tensor_holder"] = weight;
563
+ auto weights_f16 = std::make_shared<ov::op::v0::Convert>(weights_node, ov::element::f16);
564
+
565
+ if (use_bias && zp.get_size() > 0) {
566
+ // Accurate dequant in the FUSABLE zero-point form: (w - zp) * s, where the zero
567
+ // point is an exact f16 value zp = -bias/scale (the zp tensor holds bias values
568
+ // coming in). Algebraically equal to w*s + bias, but unlike an Add(bias) graph this
569
+ // matches CompressedWeightsBlock's pattern (Constant->Convert->Subtract->Multiply),
570
+ // so for_gather_matmul weights still fuse into GatherMatmulCompressed. Also avoids
571
+ // the round(min/scale) error of an integer zero point. Convert bias -> zero-point IN
572
+ // PLACE in the (possibly buffer-backed) zp tensor to avoid a duplicate allocation.
573
+ auto * bias_zp_data = zp.data<ov::float16>();
574
+ const auto * scale_data = scales.data<ov::float16>();
575
+ const size_t n = zp.get_size();
576
+ for (size_t i = 0; i < n; i++) {
577
+ float s = static_cast<float>(scale_data[i]);
578
+ float b = static_cast<float>(bias_zp_data[i]);
579
+ bias_zp_data[i] = ov::float16(s != 0.0f ? -b / s : 0.0f);
580
+ }
581
+ auto zero_point_f16 = std::make_shared<ov::op::v0::Constant>(zp);
582
+ auto w_zp =
583
+ std::make_shared<ov::op::v1::Subtract>(weights_f16, zero_point_f16, ov::op::AutoBroadcastType::NUMPY);
584
+ result = std::make_shared<ov::op::v1::Multiply>(w_zp, scales_f16, ov::op::AutoBroadcastType::NUMPY);
585
+ } else {
586
+ // Zero point path: (w - zp) * s
587
+ auto zero_point = std::make_shared<ov::op::v0::Constant>(zp);
588
+ float zp_value;
589
+ if (ov::op::util::get_single_value(zero_point, zp_value)) {
590
+ zero_point = ov::op::v0::Constant::create(zero_point->get_element_type(), {}, {zp_value});
591
+ }
592
+ auto zero_point_f16 = std::make_shared<ov::op::v0::Convert>(zero_point, ov::element::f16);
593
+ auto w_zp =
594
+ std::make_shared<ov::op::v1::Subtract>(weights_f16, zero_point_f16, ov::op::AutoBroadcastType::NUMPY);
595
+ auto mul = std::make_shared<ov::op::v1::Multiply>(w_zp, scales_f16, ov::op::AutoBroadcastType::NUMPY);
596
+ result = mul;
597
+ }
598
+ }
599
+
600
+ if (packed_shape.size() != orig_shape.size()) {
601
+ // If not requantized channel-wise case, reshape back to original shape
602
+ auto final_shape =
603
+ std::make_shared<ov::op::v0::Constant>(ov::element::i64, ov::Shape{orig_shape.size()}, orig_shape);
604
+ auto reshaped = std::make_shared<ov::op::v1::Reshape>(result, final_shape, false);
605
+ result = reshaped;
606
+ }
607
+
608
+ if (for_gather_matmul) {
609
+ return result;
610
+ }
611
+ return std::make_shared<ov::op::v0::Convert>(result, ov::element::f32);
612
+ }
613
+
614
+ // See make_int8_weights for the meaning of for_gather_matmul.
615
+ ov::Output<ov::Node> make_int4_weights(ov::Tensor & weight,
616
+ ov::Tensor & scales,
617
+ ov::Tensor & zp,
618
+ size_t group_size,
619
+ bool use_bias,
620
+ bool for_gather_matmul) {
621
+ ov::Shape orig_weight_shape = weight.get_shape();
622
+ bool is_signed = (weight.get_element_type() == ov::element::i4); // Symmetric: signed weights, no ZP
623
+
624
+ // Expand dimensions for scales and zp/bias
625
+ ov::Shape scale_shape = scales.get_shape();
626
+
627
+ // Create INT4 weight tensor. Group the innermost (last) dimension: for 2D weights
628
+ // [rows, cols] this yields [rows, cols/group_size, group_size]; for 3D MoE experts
629
+ // [n_expert, rows, cols] this yields [n_expert, rows, cols/group_size, group_size].
630
+ ov::Shape packed_shape = orig_weight_shape;
631
+ packed_shape.back() /= group_size;
632
+ packed_shape.push_back(group_size);
633
+ const size_t group_dim = packed_shape.size() - 2;
634
+
635
+ if (packed_shape[group_dim] == 1) {
636
+ // Requantized channel-wise case
637
+ packed_shape.erase(packed_shape.begin() + group_dim);
638
+ } else {
639
+ scale_shape.push_back(1);
640
+ scales.set_shape(scale_shape);
641
+ if (!is_signed && zp.get_size() > 0) {
642
+ auto zp_shape = zp.get_shape();
643
+ zp_shape.push_back(1);
644
+ zp.set_shape(zp_shape);
645
+ }
646
+ }
647
+
648
+ auto scales_f16 = std::make_shared<ov::op::v0::Constant>(scales);
649
+
650
+ ov::Output<ov::Node> result;
651
+ if (is_signed) {
652
+ // Signed path: q * s (no zero point subtraction needed)
653
+ auto weights_node = std::make_shared<ov::op::v0::Constant>(ov::element::i4, packed_shape,
654
+ static_cast<uint8_t *>(weight.data()), nullptr);
655
+ weights_node->get_rt_info()["__gguf_tensor_holder"] = weight;
656
+ auto weights_f16 = std::make_shared<ov::op::v0::Convert>(weights_node, ov::element::f16);
657
+ auto mul = std::make_shared<ov::op::v1::Multiply>(weights_f16, scales_f16, ov::op::AutoBroadcastType::NUMPY);
658
+ result = mul;
659
+ } else {
660
+ // Unsigned path
661
+ auto weights_node = std::make_shared<ov::op::v0::Constant>(ov::element::u4, packed_shape,
662
+ static_cast<uint8_t *>(weight.data()), nullptr);
663
+ weights_node->get_rt_info()["__gguf_tensor_holder"] = weight;
664
+ auto weights_f16 = std::make_shared<ov::op::v0::Convert>(weights_node, ov::element::f16);
665
+
666
+ if (use_bias && zp.get_size() > 0) {
667
+ // Accurate dequant in the FUSABLE zero-point form: (w - zp) * s with an exact f16
668
+ // zp = -bias/scale. Equivalent to w*s + bias but matches CompressedWeightsBlock's
669
+ // pattern so for_gather_matmul weights still fuse into GatherMatmulCompressed, and
670
+ // avoids the round(min/scale) error of an integer zp. Convert bias -> zero-point IN
671
+ // PLACE in the (possibly buffer-backed) zp tensor to avoid a duplicate allocation.
672
+ auto * bias_zp_data = zp.data<ov::float16>();
673
+ const auto * scale_data = scales.data<ov::float16>();
674
+ const size_t n = zp.get_size();
675
+ for (size_t i = 0; i < n; i++) {
676
+ float s = static_cast<float>(scale_data[i]);
677
+ float b = static_cast<float>(bias_zp_data[i]);
678
+ bias_zp_data[i] = ov::float16(s != 0.0f ? -b / s : 0.0f);
679
+ }
680
+ auto zero_points_f16 = std::make_shared<ov::op::v0::Constant>(zp);
681
+ auto w_zp =
682
+ std::make_shared<ov::op::v1::Subtract>(weights_f16, zero_points_f16, ov::op::AutoBroadcastType::NUMPY);
683
+ result = std::make_shared<ov::op::v1::Multiply>(w_zp, scales_f16, ov::op::AutoBroadcastType::NUMPY);
684
+ } else {
685
+ // Zero point path: (w - zp) * s
686
+ auto zero_points_node = std::make_shared<ov::op::v0::Constant>(zp);
687
+ float zp_value;
688
+ if (ov::op::util::get_single_value(zero_points_node, zp_value)) {
689
+ zero_points_node = ov::op::v0::Constant::create(zero_points_node->get_element_type(), {}, {zp_value});
690
+ }
691
+ auto zero_points_f16 = std::make_shared<ov::op::v0::Convert>(zero_points_node, ov::element::f16);
692
+ auto w_zp =
693
+ std::make_shared<ov::op::v1::Subtract>(weights_f16, zero_points_f16, ov::op::AutoBroadcastType::NUMPY);
694
+ auto mul = std::make_shared<ov::op::v1::Multiply>(w_zp, scales_f16, ov::op::AutoBroadcastType::NUMPY);
695
+ result = mul;
696
+ }
697
+ }
698
+
699
+ if (packed_shape.size() != orig_weight_shape.size()) {
700
+ // If not requantized channel-wise case, reshape back to original shape
701
+ auto final_shape = std::make_shared<ov::op::v0::Constant>(ov::element::i64, ov::Shape{orig_weight_shape.size()},
702
+ orig_weight_shape);
703
+ auto reshaped = std::make_shared<ov::op::v1::Reshape>(result, final_shape, false);
704
+ result = reshaped;
705
+ }
706
+
707
+ if (for_gather_matmul) {
708
+ return result;
709
+ }
710
+ return std::make_shared<ov::op::v0::Convert>(result, ov::element::f32);
711
+ }
712
+
713
+ ov::Output<ov::Node> make_mxfp4_weights(ov::Tensor & weight, ov::Tensor & scales) {
714
+ const ov::Shape final_shape = weight.get_shape();
715
+ GGML_ASSERT(!final_shape.empty());
716
+ GGML_ASSERT(final_shape.back() % MXFP4_BLOCK_SIZE == 0);
717
+
718
+ ov::Shape packed_shape = final_shape;
719
+ packed_shape.back() /= MXFP4_BLOCK_SIZE;
720
+ packed_shape.push_back(MXFP4_BLOCK_SIZE);
721
+
722
+ ov::Shape scale_shape = packed_shape;
723
+ scale_shape.back() = 1;
724
+ scales.set_shape(scale_shape);
725
+
726
+ auto weights_node = std::make_shared<ov::op::v0::Constant>(ov::element::f4e2m1, packed_shape,
727
+ static_cast<uint8_t *>(weight.data()), nullptr);
728
+ weights_node->get_rt_info()["__gguf_tensor_holder"] = weight;
729
+ auto weights_f32 = std::make_shared<ov::op::v0::Convert>(weights_node, ov::element::f32);
730
+
731
+ auto scales_node = std::make_shared<ov::op::v0::Constant>(scales);
732
+ auto scales_f32 = std::make_shared<ov::op::v0::Convert>(scales_node, ov::element::f32);
733
+ ov::Output<ov::Node> result =
734
+ std::make_shared<ov::op::v1::Multiply>(weights_f32, scales_f32, ov::op::AutoBroadcastType::NUMPY);
735
+
736
+ auto final_shape_node =
737
+ std::make_shared<ov::op::v0::Constant>(ov::element::i64, ov::Shape{final_shape.size()}, final_shape);
738
+ return std::make_shared<ov::op::v1::Reshape>(result, final_shape_node, false);
739
+ }
740
+
741
+ ov::Output<ov::Node> make_mxfp4_moe_packed_weights(ov::Tensor & weight) {
742
+ auto weights_node = std::make_shared<ov::op::v0::Constant>(ov::element::u8, weight.get_shape(),
743
+ static_cast<uint8_t *>(weight.data()), nullptr);
744
+ weights_node->get_rt_info()["__gguf_tensor_holder"] = weight;
745
+ weights_node->get_rt_info()["__ggml_openvino_mxfp4_moe_packed"] = true;
746
+ return weights_node;
747
+ }
748
+
749
+ // Extract quantized weights from tensor and create weight subgraph
750
+ std::shared_ptr<ov::Node> extract_quantized_weights(const ggml_tensor * tensor,
751
+ const void * data,
752
+ ov::Tensor & weights,
753
+ ov::Tensor & scales,
754
+ ov::Tensor & zp,
755
+ bool use_bias) {
756
+ // Create a temporary tensor for extraction functions that read from tensor->data
757
+ ggml_tensor temp_tensor = *tensor;
758
+ temp_tensor.data = const_cast<void *>(data);
759
+
760
+ if (tensor->type == GGML_TYPE_MXFP4) {
761
+ extract_mxfp4_data(&temp_tensor, weights, scales);
762
+ auto result = make_mxfp4_weights(weights, scales).get_node_shared_ptr();
763
+ result->set_friendly_name(tensor->name);
764
+ return result;
765
+ }
766
+
767
+ // Determine block size based on tensor type
768
+ int64_t weights_per_block;
769
+ bool is_u4;
770
+ switch (tensor->type) {
771
+ case GGML_TYPE_Q4_0:
772
+ case GGML_TYPE_Q4_1:
773
+ case GGML_TYPE_Q4_K:
774
+ is_u4 = true;
775
+ weights_per_block = 32;
776
+ break;
777
+ case GGML_TYPE_Q8_0:
778
+ case GGML_TYPE_Q5_1:
779
+ case GGML_TYPE_Q5_K:
780
+ is_u4 = false;
781
+ weights_per_block = 32;
782
+ break;
783
+ case GGML_TYPE_Q6_K:
784
+ is_u4 = false;
785
+ weights_per_block = 16;
786
+ break;
787
+ default:
788
+ throw std::runtime_error("Unsupported quantized type for extraction: " +
789
+ std::string(ggml_type_name(tensor->type)));
790
+ }
791
+
792
+ // 3D MoE expert weights (for_gather_matmul) always use the exact f16 zero-point extraction
793
+ // (see make_int8_weights/make_int4_weights) rather than the rounded integer zero point --
794
+ // round(min/scale) error is what corrupts Q4_K/Q5_1 experts, and the f16-zp form still fuses
795
+ // into GatherMatmulCompressed since it stays a Subtract, not an Add.
796
+ const bool for_gather_matmul = tensor->ne[2] > 1;
797
+ use_bias = use_bias || for_gather_matmul;
798
+
799
+ // Extract quantized data
800
+ switch (tensor->type) {
801
+ case GGML_TYPE_Q4_0:
802
+ extract_q4_0_data(&temp_tensor, weights, scales, zp);
803
+ break;
804
+ case GGML_TYPE_Q4_1:
805
+ extract_q4_1_data(&temp_tensor, weights, scales, zp, use_bias);
806
+ break;
807
+ case GGML_TYPE_Q4_K:
808
+ extract_q4_k_data(&temp_tensor, weights, scales, zp, use_bias);
809
+ break;
810
+ case GGML_TYPE_Q5_1:
811
+ extract_q5_1_data(&temp_tensor, weights, scales, zp, use_bias);
812
+ break;
813
+ case GGML_TYPE_Q8_0:
814
+ extract_q8_0_data(&temp_tensor, weights, scales, zp);
815
+ break;
816
+ case GGML_TYPE_Q6_K:
817
+ extract_q6_k_data(&temp_tensor, weights, scales, zp);
818
+ break;
819
+ case GGML_TYPE_Q5_K:
820
+ extract_q5_k_data(&temp_tensor, weights, scales, zp, use_bias);
821
+ break;
822
+ default:
823
+ throw std::runtime_error("Unsupported quantized type: " + std::string(ggml_type_name(tensor->type)));
824
+ }
825
+
826
+ // Create the OpenVINO weight subgraph. 3D expert weights (MoE) are routed through the
827
+ // GatherMatmul-oriented path: dequantized in f16, with constant folding disabled on the chain.
828
+ ov::Output<ov::Node> weight_node;
829
+ if (is_u4) {
830
+ weight_node = make_int4_weights(weights, scales, zp, weights_per_block, use_bias, for_gather_matmul);
831
+ } else {
832
+ weight_node = make_int8_weights(weights, scales, zp, weights_per_block, use_bias, for_gather_matmul);
833
+ }
834
+
835
+ auto result = weight_node.get_node_shared_ptr();
836
+ result->set_friendly_name(tensor->name);
837
+ return result;
838
+ }
839
+
840
+ // Requantize weights to target format, writing to provided buffers
841
+ std::shared_ptr<ov::Node> requantize_to_buffers(const ggml_tensor * tensor,
842
+ const void * data,
843
+ ExtraQuantType requant_type,
844
+ int64_t block_size,
845
+ ov::Tensor & weights,
846
+ ov::Tensor & scales,
847
+ ov::Tensor & zp) {
848
+ int64_t n_elements = ggml_nelements(tensor);
849
+ const int64_t ne0 = tensor->ne[0]; // elements per row
850
+ const int64_t n_rows = n_elements / ne0;
851
+ const auto * type_traits = ggml_get_type_traits(tensor->type);
852
+ const size_t src_row_bytes = ggml_row_size(tensor->type, ne0);
853
+
854
+ bool is_u4 = (requant_type == ExtraQuantType::Q4_0_C || requant_type == ExtraQuantType::Q4_0_128);
855
+
856
+ // Streaming dequant (opt-in via GGML_OPENVINO_REDUCE_COMPILE_MEM or
857
+ // GGML_OPENVINO_MEMORY_OPTIMIZE): instead of
858
+ // materializing the full n_elements F32 array (e.g. ~1 GB for token_embd), dequantize
859
+ // a chunk of complete rows into a small scratch and quantize/convert it straight into
860
+ // the output buffers, capping the transient F32 footprint at CHUNK_ROWS*ne0 floats.
861
+ //
862
+ // Only valid (and only used) for the Q8_0_C / Q8_1_C / F16 targets whose block size
863
+ // divides a row (channel-wise _C uses block_size == ne0) so no target block straddles
864
+ // a row boundary, and Q8/F16 have no cross-block packing. The u4 (Q4_0) path packs two
865
+ // weights per byte with running zp ORs that assume a single whole-array call, so it is
866
+ // never streamed. When the flag is off, behavior is identical to the original
867
+ // full-materialization path.
868
+ const bool stream_requant = ggml_openvino_reduce_compile_mem_enabled() && !is_u4 &&
869
+ !(block_size > 0 && ne0 % block_size != 0);
870
+
871
+ if (!stream_requant) {
872
+ // Full materialization (original behavior): dequantize the whole tensor to F32,
873
+ // then convert/quantize in one call.
874
+ std::vector<float> weights_f32(n_elements);
875
+ type_traits->to_float(data, weights_f32.data(), n_elements);
876
+ if (requant_type == ExtraQuantType::F16) {
877
+ ggml_get_type_traits(GGML_TYPE_F16)->from_float_ref(weights_f32.data(), weights.data(), n_elements);
878
+ auto result = std::make_shared<ov::op::v0::Constant>(weights);
879
+ result->set_friendly_name(tensor->name);
880
+ return result;
881
+ }
882
+ if (is_u4) {
883
+ quantize_q4_0(weights_f32.data(), weights, scales, zp, n_elements, block_size);
884
+ } else if (requant_type == ExtraQuantType::Q8_1_C) {
885
+ quantize_q8_1(weights_f32.data(), weights, scales, zp, n_elements, block_size);
886
+ } else {
887
+ quantize_q8_0(weights_f32.data(), weights, scales, zp, n_elements, block_size);
888
+ }
889
+ } else {
890
+ // Streaming path for Q8_0_C / Q8_1_C / F16 (covers token_embd, output.weight,
891
+ // and per-layer Q6_K/Q5_K requant — the large transient cases).
892
+ const int64_t CHUNK_ROWS = std::min<int64_t>(n_rows, 256);
893
+ std::vector<float> scratch(CHUNK_ROWS * ne0);
894
+ // F16 destination: 2 bytes/element, advanced per chunk by r0*ne0 elements.
895
+ auto * f16_base = static_cast<uint8_t *>(weights.data());
896
+ for (int64_t r0 = 0; r0 < n_rows; r0 += CHUNK_ROWS) {
897
+ const int64_t rows = std::min(CHUNK_ROWS, n_rows - r0);
898
+ const int64_t elems = rows * ne0;
899
+ const auto * src = static_cast<const uint8_t *>(data) + r0 * src_row_bytes;
900
+ type_traits->to_float(src, scratch.data(), elems);
901
+
902
+ if (requant_type == ExtraQuantType::F16) {
903
+ ggml_get_type_traits(GGML_TYPE_F16)
904
+ ->from_float_ref(scratch.data(), f16_base + (r0 * ne0) * sizeof(uint16_t), elems);
905
+ } else {
906
+ const int64_t block_offset = (r0 * ne0) / block_size;
907
+ if (requant_type == ExtraQuantType::Q8_1_C) {
908
+ quantize_q8_1(scratch.data(), weights, scales, zp, elems, block_size, block_offset);
909
+ } else {
910
+ quantize_q8_0(scratch.data(), weights, scales, zp, elems, block_size, block_offset);
911
+ }
912
+ }
913
+ }
914
+ if (requant_type == ExtraQuantType::F16) {
915
+ auto result = std::make_shared<ov::op::v0::Constant>(weights);
916
+ result->set_friendly_name(tensor->name);
917
+ return result;
918
+ }
919
+ }
920
+
921
+ // Create the OpenVINO weight subgraph
922
+ ov::Output<ov::Node> weight_node;
923
+ if (is_u4) {
924
+ weight_node = make_int4_weights(weights, scales, zp, block_size);
925
+ } else {
926
+ weight_node = make_int8_weights(weights, scales, zp, block_size);
927
+ }
928
+
929
+ auto result = weight_node.get_node_shared_ptr();
930
+ result->set_friendly_name(tensor->name);
931
+ return result;
932
+ }
933
+
934
+ OvWeight process_weight_tensor(const ggml_tensor * tensor, const void * data, void * output_base_ptr, bool use_bias) {
935
+ GGML_ASSERT(tensor != nullptr);
936
+ GGML_ASSERT(data != nullptr);
937
+
938
+ OvWeight result;
939
+
940
+ // Get shape for weights: [rows, cols], or [n_expert, rows, cols] for 3D MoE expert weights.
941
+ ov::Shape node_shape = (tensor->ne[2] > 1) ?
942
+ ov::Shape{static_cast<size_t>(tensor->ne[2]), static_cast<size_t>(tensor->ne[1]),
943
+ static_cast<size_t>(tensor->ne[0])} :
944
+ ov::Shape{static_cast<size_t>(tensor->ne[1]), static_cast<size_t>(tensor->ne[0])};
945
+
946
+ // Handle F16/F32/BF16 weights
947
+ if (tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_F16 || tensor->type == GGML_TYPE_BF16) {
948
+ ov::element::Type element_type;
949
+ switch (tensor->type) {
950
+ case GGML_TYPE_F32:
951
+ element_type = ov::element::f32;
952
+ break;
953
+ case GGML_TYPE_F16:
954
+ element_type = ov::element::f16;
955
+ break;
956
+ case GGML_TYPE_BF16:
957
+ element_type = ov::element::bf16;
958
+ break;
959
+ default:
960
+ OPENVINO_THROW("Unexpected tensor type in F16/F32/BF16 path");
961
+ }
962
+
963
+ if (output_base_ptr && output_base_ptr != data) {
964
+ // Using external buffer - copy data and create shared-memory constant
965
+ size_t tensor_bytes = ggml_nbytes(tensor);
966
+ memcpy(output_base_ptr, data, tensor_bytes);
967
+ result.weights = ov::Tensor(element_type, node_shape, output_base_ptr);
968
+ } else {
969
+ result.weights = ov::Tensor(element_type, node_shape, data);
970
+ }
971
+ result.weight_node = std::make_shared<ov::op::v0::Constant>(result.weights);
972
+ return result;
973
+ }
974
+
975
+ // Handle quantized weights
976
+ if (!ggml_is_quantized(tensor->type)) {
977
+ OPENVINO_THROW("Unsupported weight tensor type: ", ggml_type_name(tensor->type));
978
+ }
979
+
980
+ result.layout = ggml_openvino_get_extracted_layout(tensor, use_bias);
981
+ const auto & layout = result.layout;
982
+ if (layout.total_size == 0) {
983
+ OPENVINO_THROW("Unsupported quantized type: ", ggml_type_name(tensor->type));
984
+ }
985
+
986
+ // 3D MoE expert weights (for_gather_matmul) always use the exact f16 zero-point path (see
987
+ // extract_quantized_weights) -- must be kept in sync with the "use_bias || for_gather_matmul"
988
+ // check in ggml_openvino_get_extracted_layout, which sizes/offsets the zp slot accordingly.
989
+ // Requantized tensors (layout.is_requant) are handled by requantize_to_buffers instead, whose
990
+ // zp sizing/type is unaffected by for_gather_matmul, so they are excluded here.
991
+ const bool for_gather_matmul = tensor->ne[2] > 1;
992
+ const bool zp_is_f16 = !layout.is_requant && (use_bias || for_gather_matmul);
993
+
994
+ const bool is_3d_mxfp4_moe = tensor->type == GGML_TYPE_MXFP4 && (tensor->ne[2] > 1 || tensor->ne[3] > 1);
995
+ if (is_3d_mxfp4_moe) {
996
+ ov::Shape packed_shape = {static_cast<size_t>(tensor->ne[3]),
997
+ static_cast<size_t>(tensor->ne[2]),
998
+ static_cast<size_t>(tensor->ne[1]),
999
+ static_cast<size_t>(tensor->ne[0] / MXFP4_BLOCK_SIZE),
1000
+ MXFP4_BLOCK_BYTES};
1001
+ const size_t tensor_bytes = ggml_nbytes(tensor);
1002
+ if (output_base_ptr) {
1003
+ auto * buf_base = static_cast<uint8_t *>(output_base_ptr);
1004
+ memcpy(buf_base + layout.weights_offset, data, tensor_bytes);
1005
+ result.weights = ov::Tensor(ov::element::u8, packed_shape, buf_base + layout.weights_offset);
1006
+ } else {
1007
+ result.weights = ov::Tensor(ov::element::u8, packed_shape);
1008
+ memcpy(result.weights.data(), data, tensor_bytes);
1009
+ }
1010
+ result.weight_node = make_mxfp4_moe_packed_weights(result.weights).get_node_shared_ptr();
1011
+ result.weight_node->set_friendly_name(tensor->name);
1012
+ return result;
1013
+ }
1014
+
1015
+ if (use_bias) {
1016
+ OPENVINO_ASSERT(!layout.is_requant,
1017
+ "use_bias is only used for test-backend-ops, which should not have requantization");
1018
+ // bias node will be created on the fly and not use backend buffer
1019
+ output_base_ptr = nullptr;
1020
+ }
1021
+
1022
+ // F16 requant path - no separate scales/zp needed in result
1023
+ if (layout.is_requant && layout.requant_type.has_value() && layout.requant_type.value() == ExtraQuantType::F16) {
1024
+ if (output_base_ptr) {
1025
+ result.weights = ov::Tensor(ov::element::f16, node_shape,
1026
+ static_cast<uint8_t *>(output_base_ptr) + layout.weights_offset);
1027
+ } else {
1028
+ result.weights = ov::Tensor(ov::element::f16, node_shape);
1029
+ }
1030
+ ov::Tensor dummy_scales, dummy_zp; // Not used for F16
1031
+ result.weight_node =
1032
+ requantize_to_buffers(tensor, data, ExtraQuantType::F16, 0, result.weights, dummy_scales, dummy_zp);
1033
+ return result;
1034
+ }
1035
+
1036
+ // Quantized path (normal extraction or quantized requant)
1037
+ // Create weight/scale/zp tensors - shared between both paths
1038
+ // For symmetric quantization, use signed types (i4/i8) and no ZP tensor
1039
+ ov::element::Type weight_type = tensor->type == GGML_TYPE_MXFP4 ?
1040
+ ov::element::f4e2m1 :
1041
+ (layout.is_symmetric ? (layout.is_u4 ? ov::element::i4 : ov::element::i8) :
1042
+ (layout.is_u4 ? ov::element::u4 : ov::element::u8));
1043
+ ov::Shape scale_shape = node_shape;
1044
+ scale_shape.back() /= layout.weights_per_block;
1045
+
1046
+ if (tensor->type == GGML_TYPE_MXFP4) {
1047
+ if (tensor->ne[2] == 1 && tensor->ne[3] == 1) {
1048
+ node_shape = {static_cast<size_t>(tensor->ne[1]), static_cast<size_t>(tensor->ne[0])};
1049
+ } else {
1050
+ node_shape.clear();
1051
+ for (int i = GGML_MAX_DIMS - 1; i >= 0; --i) {
1052
+ node_shape.push_back(static_cast<size_t>(tensor->ne[i]));
1053
+ }
1054
+ }
1055
+
1056
+ scale_shape = node_shape;
1057
+ scale_shape.back() /= layout.weights_per_block;
1058
+ }
1059
+
1060
+ if (output_base_ptr) {
1061
+ uint8_t * buf_base = static_cast<uint8_t *>(output_base_ptr);
1062
+ result.weights = ov::Tensor(weight_type, node_shape, buf_base + layout.weights_offset);
1063
+ const ov::element::Type scale_type = tensor->type == GGML_TYPE_MXFP4 ? ov::element::f8e8m0 : ov::element::f16;
1064
+ result.scales = ov::Tensor(scale_type, scale_shape, buf_base + layout.scales_offset);
1065
+ if (!layout.is_symmetric) {
1066
+ ov::element::Type zp_type =
1067
+ zp_is_f16 ? ov::element::f16 : (layout.is_u4 ? ov::element::u4 : ov::element::u8);
1068
+ result.zp = ov::Tensor(zp_type, scale_shape, buf_base + layout.zp_offset);
1069
+ }
1070
+ // else: result.zp remains default-constructed (empty) for symmetric
1071
+ } else {
1072
+ result.weights = ov::Tensor(weight_type, node_shape);
1073
+ const ov::element::Type scale_type = tensor->type == GGML_TYPE_MXFP4 ? ov::element::f8e8m0 : ov::element::f16;
1074
+ result.scales = ov::Tensor(scale_type, scale_shape);
1075
+ if (!layout.is_symmetric) {
1076
+ if (zp_is_f16) {
1077
+ result.zp = ov::Tensor(ov::element::f16, scale_shape);
1078
+ } else {
1079
+ ov::element::Type zp_type = layout.is_u4 ? ov::element::u4 : ov::element::u8;
1080
+ result.zp = ov::Tensor(zp_type, scale_shape);
1081
+ }
1082
+ }
1083
+ // else: result.zp remains default-constructed (empty) for symmetric
1084
+ }
1085
+
1086
+ if (layout.is_requant && layout.requant_type.has_value()) {
1087
+ result.weight_node = requantize_to_buffers(tensor, data, layout.requant_type.value(), layout.weights_per_block,
1088
+ result.weights, result.scales, result.zp);
1089
+ } else {
1090
+ result.weight_node =
1091
+ extract_quantized_weights(tensor, data, result.weights, result.scales, result.zp, use_bias);
1092
+ }
1093
+
1094
+ return result;
1095
+ }
1096
+
1097
+ void quantize_q4_0(const float * x,
1098
+ ov::Tensor & weights_arr,
1099
+ ov::Tensor & scales_arr,
1100
+ ov::Tensor & zp_arr,
1101
+ int64_t k,
1102
+ int64_t qk) {
1103
+ assert(k % qk == 0);
1104
+ const int nb = k / qk;
1105
+
1106
+ auto * weights = static_cast<uint8_t *>(weights_arr.data());
1107
+ auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>();
1108
+ bool is_symmetric = (weights_arr.get_element_type() == ov::element::i4); // Signed i4 path
1109
+
1110
+ if (!is_symmetric) {
1111
+ auto * zp = static_cast<uint8_t *>(zp_arr.data());
1112
+ for (int i = 0; i < nb; i++) {
1113
+ float amax = 0.0f;
1114
+ float max = 0.0f;
1115
+ for (int j = 0; j < qk; j++) {
1116
+ const float v = x[i * qk + j];
1117
+ if (amax < fabsf(v)) {
1118
+ amax = fabsf(v);
1119
+ max = v;
1120
+ }
1121
+ }
1122
+ const float d = max / -8;
1123
+ if (d == 0) {
1124
+ scales[i] = ov::float16(1.0f);
1125
+ if (i % 2 == 0) {
1126
+ zp[i / 2] = 8;
1127
+ } else {
1128
+ zp[i / 2] |= (8 << 4);
1129
+ }
1130
+ memset(weights + i * qk / 2, 8 | (8 << 4), qk / 2);
1131
+ continue;
1132
+ }
1133
+ const float id = 1.0f / d;
1134
+ scales[i] = ov::float16(d);
1135
+ if (i % 2 == 0) {
1136
+ zp[i / 2] = 8;
1137
+ } else {
1138
+ zp[i / 2] |= (8 << 4);
1139
+ }
1140
+ for (int j = 0; j < qk / 2; ++j) {
1141
+ const float x0 = x[i * qk + 2 * j] * id;
1142
+ const float x1 = x[i * qk + 2 * j + 1] * id;
1143
+ const uint8_t xi0 = MIN(15, (int8_t) (x0 + 8.5f));
1144
+ const uint8_t xi1 = MIN(15, (int8_t) (x1 + 8.5f));
1145
+ weights[i * qk / 2 + j] = xi0 | (xi1 << 4);
1146
+ }
1147
+ }
1148
+ } else {
1149
+ // Symmetric: produce signed i4 values in [-8, 7]
1150
+ for (int i = 0; i < nb; i++) {
1151
+ float amax = 0.0f;
1152
+ float max = 0.0f;
1153
+ for (int j = 0; j < qk; j++) {
1154
+ const float v = x[i * qk + j];
1155
+ if (amax < fabsf(v)) {
1156
+ amax = fabsf(v);
1157
+ max = v;
1158
+ }
1159
+ }
1160
+ const float d = max / -8;
1161
+ if (d == 0) {
1162
+ scales[i] = ov::float16(1.0f);
1163
+ // i4 value 0 packed: 0x00
1164
+ memset(weights + i * qk / 2, 0, qk / 2);
1165
+ continue;
1166
+ }
1167
+ const float id = 1.0f / d;
1168
+ scales[i] = ov::float16(d);
1169
+ for (int j = 0; j < qk / 2; ++j) {
1170
+ const float x0 = x[i * qk + 2 * j] * id;
1171
+ const float x1 = x[i * qk + 2 * j + 1] * id;
1172
+ // Signed i4: range [-8, 7]. Quantize as round(x*id), then pack as 4-bit two's complement.
1173
+ int8_t si0 = (int8_t) std::max(-8, std::min(7, (int) roundf(x0)));
1174
+ int8_t si1 = (int8_t) std::max(-8, std::min(7, (int) roundf(x1)));
1175
+ weights[i * qk / 2 + j] = (si0 & 0x0F) | ((si1 & 0x0F) << 4);
1176
+ }
1177
+ }
1178
+ }
1179
+ }
1180
+
1181
+ void quantize_q8_0(const float * x,
1182
+ ov::Tensor & weights_arr,
1183
+ ov::Tensor & scales_arr,
1184
+ ov::Tensor & zp_arr,
1185
+ int64_t k,
1186
+ int64_t qk,
1187
+ int64_t block_offset) {
1188
+ assert(k % qk == 0);
1189
+ const int nb = k / qk;
1190
+
1191
+ // block_offset lets a caller quantize a chunk of blocks into the right place in the
1192
+ // output buffers (used for streaming requant). x points at this chunk's first block;
1193
+ // outputs are advanced by block_offset blocks. Q8 has one scale/zp per block (no
1194
+ // nibble packing), so any block boundary is safe.
1195
+ auto * weights = static_cast<uint8_t *>(weights_arr.data()) + block_offset * qk;
1196
+ auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>() + block_offset;
1197
+ bool is_symmetric = (weights_arr.get_element_type() == ov::element::i8); // Signed i8 path
1198
+
1199
+ if (!is_symmetric) {
1200
+ auto * zp = static_cast<uint8_t *>(zp_arr.data()) + block_offset;
1201
+ for (int i = 0; i < nb; i++) {
1202
+ float amax = 0.0f;
1203
+ for (int j = 0; j < qk; j++) {
1204
+ const float v = x[i * qk + j];
1205
+ amax = std::max(amax, fabsf(v));
1206
+ }
1207
+ const float d = amax / 127.0f;
1208
+ const float id = d ? 1.0f / d : 0.0f;
1209
+ scales[i] = ov::float16(d);
1210
+ zp[i] = 128;
1211
+ for (int j = 0; j < qk; ++j) {
1212
+ const float x0 = x[i * qk + j] * id;
1213
+ const int8_t xi0 = roundf(x0);
1214
+ weights[i * qk + j] = (uint8_t) (xi0 + 128);
1215
+ }
1216
+ }
1217
+ } else {
1218
+ // Symmetric: store signed int8 values directly
1219
+ auto * signed_weights = reinterpret_cast<int8_t *>(weights);
1220
+ for (int i = 0; i < nb; i++) {
1221
+ float amax = 0.0f;
1222
+ for (int j = 0; j < qk; j++) {
1223
+ const float v = x[i * qk + j];
1224
+ amax = std::max(amax, fabsf(v));
1225
+ }
1226
+ const float d = amax / 127.0f;
1227
+ const float id = d ? 1.0f / d : 0.0f;
1228
+ scales[i] = ov::float16(d);
1229
+ for (int j = 0; j < qk; ++j) {
1230
+ const float x0 = x[i * qk + j] * id;
1231
+ signed_weights[i * qk + j] = (int8_t) roundf(x0);
1232
+ }
1233
+ }
1234
+ }
1235
+ }
1236
+
1237
+ void quantize_q8_1(const float * x,
1238
+ ov::Tensor & weights_arr,
1239
+ ov::Tensor & scales_arr,
1240
+ ov::Tensor & zp_arr,
1241
+ int64_t k,
1242
+ int64_t qk,
1243
+ int64_t block_offset) {
1244
+ assert(k % qk == 0);
1245
+ const int nb = k / qk;
1246
+
1247
+ // See quantize_q8_0: block_offset places this chunk's output at the right block.
1248
+ auto * weights = static_cast<uint8_t *>(weights_arr.data()) + block_offset * qk;
1249
+ auto * scales = scales_arr.data<ov::element_type_traits<ov::element::f16>::value_type>() + block_offset;
1250
+ auto * zp = static_cast<uint8_t *>(zp_arr.data()) + block_offset;
1251
+ for (int i = 0; i < nb; i++) {
1252
+ float min = std::numeric_limits<float>::max();
1253
+ float max = std::numeric_limits<float>::lowest();
1254
+
1255
+ for (int j = 0; j < qk; j++) {
1256
+ const float v = x[i * qk + j];
1257
+ min = std::min(v, min);
1258
+ max = std::max(v, max);
1259
+ }
1260
+
1261
+ const float d = (max - min) / ((1 << 8) - 1);
1262
+ const float id = d ? 1.0f / d : 0.0f;
1263
+ scales[i] = ov::float16(d);
1264
+ // zp = -min / scale (Q8_1 is asymmetric)
1265
+ zp[i] = (d != 0.0f) ? (uint8_t) std::round(-min / d) : 0;
1266
+
1267
+ for (int j = 0; j < qk; ++j) {
1268
+ const float x0 = (x[i * qk + j] - min) * id;
1269
+ const uint8_t xi0 = roundf(x0);
1270
+ weights[i * qk + j] = xi0;
1271
+ }
1272
+ }
1273
+ }