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 +1,1560 @@
1
+ #ifndef SRC0
2
+ #define SRC0 src0
3
+ #endif
4
+ #ifndef SRC1
5
+ #define SRC1 src1
6
+ #endif
7
+
8
+ #ifdef U32_DEQUANT_HELPERS
9
+ #define SRC0_TYPE u32
10
+
11
+ fn byte_of(v: u32, b: u32) -> u32 {
12
+ return (v >> (b * 8u)) & 0xFFu;
13
+ }
14
+
15
+ fn sbyte_of(v: u32, b: u32) -> i32 {
16
+ let raw = i32((v >> (b * 8u)) & 0xFFu);
17
+ return select(raw, raw - 256, raw >= 128);
18
+ }
19
+ #endif
20
+
21
+ #ifdef VEC
22
+ #define VEC_SIZE 4u
23
+ #define SRC0_TYPE vec4<SRC0_INNER_TYPE>
24
+ #define SRC1_TYPE vec4<SRC1_INNER_TYPE>
25
+
26
+ fn inner_dot(src0_val: SRC0_TYPE, src1_val: SRC1_TYPE) -> f32 {
27
+ return f32(dot(SRC1_TYPE(src0_val), src1_val));
28
+ }
29
+ #endif
30
+
31
+ #ifdef SCALAR
32
+ #define VEC_SIZE 1u
33
+ #define SRC0_TYPE SRC0_INNER_TYPE
34
+ #define SRC1_TYPE SRC1_INNER_TYPE
35
+
36
+ fn inner_dot(src0_val: SRC0_TYPE, src1_val: SRC1_TYPE) -> f32 {
37
+ return f32(src0_val) * f32(src1_val);
38
+ }
39
+ #endif
40
+
41
+ #ifdef MUL_ACC_FLOAT
42
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
43
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
44
+
45
+ let k_vec = params.k / VEC_SIZE;
46
+ let src1_idx_base_vec = src1_idx_base / VEC_SIZE;
47
+
48
+ // Each thread walks K, loads from the vector, and updates
49
+ // a small block of output rows held in registers.
50
+ for (var k = thread_id; k < k_vec; k += WG_SIZE) {
51
+ var x_vals: array<SRC1_TYPE, NUM_COLS>;
52
+ for (var col = 0u;col < NUM_COLS;col += 1) {
53
+ x_vals[col] = SRC1[src1_idx_base_vec + col * (params.stride_11 / VEC_SIZE) + k];
54
+ }
55
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
56
+ let output_row = row_base + row;
57
+ if (output_row < params.m) {
58
+ let src0_idx = (src0_batch_offset + output_row * params.stride_01) / VEC_SIZE + k;
59
+ let w = SRC0[src0_idx];
60
+ for (var col = 0u;col < NUM_COLS;col += 1) {
61
+ acc[col][row] += inner_dot(w, x_vals[col]);
62
+ }
63
+ }
64
+ }
65
+ }
66
+
67
+ return acc;
68
+ }
69
+ #endif
70
+
71
+ #ifdef MUL_ACC_Q1_0
72
+ #define BLOCK_SIZE 128
73
+ #define BLOCK_SIZE_BYTES 18
74
+ #define THREADS_PER_BLOCK 16
75
+ #define ELEMS_PER_THREAD (BLOCK_SIZE/THREADS_PER_BLOCK)
76
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
77
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
78
+
79
+ let num_blocks = params.k / BLOCK_SIZE;
80
+ let thread_within_block = thread_id % THREADS_PER_BLOCK;
81
+ for (var block = thread_id / THREADS_PER_BLOCK; block < num_blocks; block += WG_SIZE / THREADS_PER_BLOCK) {
82
+ let x_base = src1_idx_base + block * BLOCK_SIZE + thread_within_block * ELEMS_PER_THREAD;
83
+ var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
84
+ for (var col = 0u; col < NUM_COLS;col += 1) {
85
+ for (var i = 0u; i < ELEMS_PER_THREAD; i++) {
86
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
87
+ }
88
+ }
89
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
90
+ let output_row = row_base + row;
91
+ if (output_row < params.m) {
92
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
93
+ let d = f32(load_f16_at_src0(block_byte_base));
94
+ let q_byte = load_u32_at_src0(block_byte_base + 2u + thread_within_block) & 0xFFu;
95
+ for (var col = 0u;col < NUM_COLS;col += 1) {
96
+ var row_sum = 0.0;
97
+ for (var bit = 0u; bit < 8u; bit++) {
98
+ let w = select(-d, d, ((q_byte >> bit) & 1u) != 0u);
99
+ row_sum += w * x_block[col][bit];
100
+ }
101
+ acc[col][row] += row_sum;
102
+ }
103
+ }
104
+ }
105
+ }
106
+
107
+ return acc;
108
+ }
109
+ #endif
110
+
111
+ #ifdef MUL_ACC_Q4_0
112
+ #define BLOCK_SIZE 32
113
+ #define BLOCK_SIZE_BYTES 18
114
+ #define THREADS_PER_BLOCK 4
115
+ #define ELEMS_PER_THREAD (BLOCK_SIZE/THREADS_PER_BLOCK)
116
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
117
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
118
+
119
+ let num_blocks = params.k / BLOCK_SIZE;
120
+ let thread_within_block = thread_id % 4;
121
+ for (var block = thread_id/THREADS_PER_BLOCK; block < num_blocks; block += WG_SIZE/THREADS_PER_BLOCK) {
122
+ let x_base = src1_idx_base + block * BLOCK_SIZE + thread_within_block * 4;
123
+ var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
124
+ for (var col = 0u; col < NUM_COLS;col += 1) {
125
+ for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
126
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
127
+ x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
128
+ }
129
+ }
130
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
131
+ let output_row = row_base + row;
132
+ if (output_row < params.m) {
133
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
134
+ let d = f32(load_f16_at_src0(block_byte_base));
135
+ let q_packed = load_u32_at_src0(block_byte_base + 2u + 4u * thread_within_block);
136
+ for (var col = 0u;col < NUM_COLS;col += 1) {
137
+ var row_sum = 0.0;
138
+ for (var byte_idx = 0u; byte_idx < 4u; byte_idx++) {
139
+ let q_byte = get_byte(q_packed, byte_idx);
140
+ let q_lo = (f32(q_byte & 0xFu) - 8.0) * d;
141
+ let q_hi = (f32((q_byte >> 4u) & 0xFu) - 8.0) * d;
142
+ row_sum += q_lo * x_block[col][byte_idx];
143
+ row_sum += q_hi * x_block[col][byte_idx + 4u];
144
+ }
145
+ acc[col][row] += row_sum;
146
+ }
147
+ }
148
+ }
149
+ }
150
+
151
+ return acc;
152
+ }
153
+ #endif
154
+
155
+ #ifdef MUL_ACC_Q4_1
156
+ #define BLOCK_SIZE 32
157
+ #define BLOCK_SIZE_BYTES 20
158
+ #define THREADS_PER_BLOCK 4
159
+ #define ELEMS_PER_THREAD (BLOCK_SIZE/THREADS_PER_BLOCK)
160
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
161
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
162
+
163
+ let num_blocks = params.k / BLOCK_SIZE;
164
+ let thread_within_block = thread_id % THREADS_PER_BLOCK;
165
+ for (var block = thread_id / THREADS_PER_BLOCK; block < num_blocks; block += WG_SIZE / THREADS_PER_BLOCK) {
166
+ let x_base = src1_idx_base + block * BLOCK_SIZE + thread_within_block * 4;
167
+ var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
168
+ for (var col = 0u; col < NUM_COLS;col += 1) {
169
+ for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
170
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
171
+ x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
172
+ }
173
+ }
174
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
175
+ let output_row = row_base + row;
176
+ if (output_row < params.m) {
177
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
178
+ let d = f32(load_f16_at_src0(block_byte_base));
179
+ let m = f32(load_f16_at_src0(block_byte_base + 2u));
180
+ let q_packed = load_u32_at_src0(block_byte_base + 4u + 4u * thread_within_block);
181
+ for (var col = 0u;col < NUM_COLS;col += 1) {
182
+ var row_sum = 0.0;
183
+ for (var byte_idx = 0u; byte_idx < 4u; byte_idx++) {
184
+ let q_byte = get_byte(q_packed, byte_idx);
185
+ let q_lo = f32(q_byte & 0xFu) * d + m;
186
+ let q_hi = f32((q_byte >> 4u) & 0xFu) * d + m;
187
+ row_sum += q_lo * x_block[col][byte_idx];
188
+ row_sum += q_hi * x_block[col][byte_idx + 4u];
189
+ }
190
+ acc[col][row] += row_sum;
191
+ }
192
+ }
193
+ }
194
+ }
195
+
196
+ return acc;
197
+ }
198
+ #endif
199
+
200
+ #ifdef MUL_ACC_Q5_0
201
+ #define BLOCK_SIZE 32
202
+ #define BLOCK_SIZE_BYTES 22
203
+ #define THREADS_PER_BLOCK 4
204
+ #define ELEMS_PER_THREAD (BLOCK_SIZE/THREADS_PER_BLOCK)
205
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
206
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
207
+
208
+ let num_blocks = params.k / BLOCK_SIZE;
209
+ let thread_within_block = thread_id % THREADS_PER_BLOCK;
210
+ for (var block = thread_id / THREADS_PER_BLOCK; block < num_blocks; block += WG_SIZE / THREADS_PER_BLOCK) {
211
+ let x_base = src1_idx_base + block * BLOCK_SIZE + thread_within_block * 4;
212
+ var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
213
+ for (var col = 0u; col < NUM_COLS;col += 1) {
214
+ for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
215
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
216
+ x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
217
+ }
218
+ }
219
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
220
+ let output_row = row_base + row;
221
+ if (output_row < params.m) {
222
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
223
+ let d = f32(load_f16_at_src0(block_byte_base));
224
+ let qh_packed = load_u32_at_src0(block_byte_base + 2u);
225
+ let q_packed = load_u32_at_src0(block_byte_base + 6u + 4u * thread_within_block);
226
+ let qh_shift = thread_within_block * 4u;
227
+ for (var col = 0u;col < NUM_COLS;col += 1) {
228
+ var row_sum = 0.0;
229
+ for (var byte_idx = 0u; byte_idx < 4u; byte_idx++) {
230
+ let q_byte = get_byte(q_packed, byte_idx);
231
+ let qh_lo = ((qh_packed >> (qh_shift + byte_idx)) << 4u) & 0x10u;
232
+ let qh_hi = (qh_packed >> (qh_shift + byte_idx + 12u)) & 0x10u;
233
+ let q_lo = (f32((q_byte & 0xFu) | qh_lo) - 16.0) * d;
234
+ let q_hi = (f32(((q_byte >> 4u) & 0xFu) | qh_hi) - 16.0) * d;
235
+ row_sum += q_lo * x_block[col][byte_idx];
236
+ row_sum += q_hi * x_block[col][byte_idx + 4u];
237
+ }
238
+ acc[col][row] += row_sum;
239
+ }
240
+ }
241
+ }
242
+ }
243
+
244
+ return acc;
245
+ }
246
+ #endif
247
+
248
+ #ifdef MUL_ACC_Q5_1
249
+ #define BLOCK_SIZE 32
250
+ #define BLOCK_SIZE_BYTES 24
251
+ #define THREADS_PER_BLOCK 4
252
+ #define ELEMS_PER_THREAD (BLOCK_SIZE/THREADS_PER_BLOCK)
253
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
254
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
255
+
256
+ let num_blocks = params.k / BLOCK_SIZE;
257
+ let thread_within_block = thread_id % THREADS_PER_BLOCK;
258
+ for (var block = thread_id / THREADS_PER_BLOCK; block < num_blocks; block += WG_SIZE / THREADS_PER_BLOCK) {
259
+ let x_base = src1_idx_base + block * BLOCK_SIZE + thread_within_block * 4;
260
+ var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
261
+ for (var col = 0u; col < NUM_COLS;col += 1) {
262
+ for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
263
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
264
+ x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
265
+ }
266
+ }
267
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
268
+ let output_row = row_base + row;
269
+ if (output_row < params.m) {
270
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
271
+ let d = f32(load_f16_at_src0(block_byte_base));
272
+ let m = f32(load_f16_at_src0(block_byte_base + 2u));
273
+ let qh_packed = load_u32_at_src0(block_byte_base + 4u);
274
+ let q_packed = load_u32_at_src0(block_byte_base + 8u + 4u * thread_within_block);
275
+ let qh_shift = thread_within_block * 4u;
276
+ for (var col = 0u;col < NUM_COLS;col += 1) {
277
+ var row_sum = 0.0;
278
+ for (var byte_idx = 0u; byte_idx < 4u; byte_idx++) {
279
+ let q_byte = get_byte(q_packed, byte_idx);
280
+ let qh_lo = ((qh_packed >> (qh_shift + byte_idx)) << 4u) & 0x10u;
281
+ let qh_hi = (qh_packed >> (qh_shift + byte_idx + 12u)) & 0x10u;
282
+ let q_lo = f32((q_byte & 0xFu) | qh_lo) * d + m;
283
+ let q_hi = f32(((q_byte >> 4u) & 0xFu) | qh_hi) * d + m;
284
+ row_sum += q_lo * x_block[col][byte_idx];
285
+ row_sum += q_hi * x_block[col][byte_idx + 4u];
286
+ }
287
+ acc[col][row] += row_sum;
288
+ }
289
+ }
290
+ }
291
+ }
292
+
293
+ return acc;
294
+ }
295
+ #endif
296
+
297
+ #ifdef MUL_ACC_Q8_0
298
+ #define BLOCK_SIZE 32
299
+ #define BLOCK_SIZE_BYTES 34
300
+ #define THREADS_PER_BLOCK 4
301
+ #define ELEMS_PER_THREAD (BLOCK_SIZE/THREADS_PER_BLOCK)
302
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
303
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
304
+
305
+ let num_blocks = params.k / BLOCK_SIZE;
306
+ let thread_within_block = thread_id % THREADS_PER_BLOCK;
307
+ for (var block = thread_id / THREADS_PER_BLOCK; block < num_blocks; block += WG_SIZE / THREADS_PER_BLOCK) {
308
+ let x_base = src1_idx_base + block * BLOCK_SIZE + thread_within_block * ELEMS_PER_THREAD;
309
+ var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
310
+ for (var col = 0u; col < NUM_COLS;col += 1) {
311
+ for (var i = 0u; i < ELEMS_PER_THREAD; i++) {
312
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
313
+ }
314
+ }
315
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
316
+ let output_row = row_base + row;
317
+ if (output_row < params.m) {
318
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
319
+ let d = f32(load_f16_at_src0(block_byte_base));
320
+ var q_packed: array<u32, ELEMS_PER_THREAD / 4u>;
321
+ for (var packed_idx = 0u; packed_idx < ELEMS_PER_THREAD / 4u; packed_idx++) {
322
+ q_packed[packed_idx] = load_u32_at_src0(block_byte_base + 2u + 4u * (thread_within_block * 2u + packed_idx));
323
+ }
324
+ for (var col = 0u;col < NUM_COLS;col += 1) {
325
+ var row_sum = 0.0;
326
+ for (var packed_idx = 0u; packed_idx < ELEMS_PER_THREAD / 4u; packed_idx++) {
327
+ for (var byte_idx = 0u; byte_idx < 4u; byte_idx++) {
328
+ let q_val = f32(get_byte_i32(q_packed[packed_idx], byte_idx)) * d;
329
+ row_sum += q_val * x_block[col][packed_idx * 4u + byte_idx];
330
+ }
331
+ }
332
+ acc[col][row] += row_sum;
333
+ }
334
+ }
335
+ }
336
+ }
337
+
338
+ return acc;
339
+ }
340
+ #endif
341
+
342
+ #ifdef MUL_ACC_Q8_1
343
+ #define BLOCK_SIZE 32
344
+ #define BLOCK_SIZE_BYTES 36
345
+ #define THREADS_PER_BLOCK 4
346
+ #define ELEMS_PER_THREAD (BLOCK_SIZE/THREADS_PER_BLOCK)
347
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
348
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
349
+
350
+ let num_blocks = params.k / BLOCK_SIZE;
351
+ let thread_within_block = thread_id % THREADS_PER_BLOCK;
352
+ for (var block = thread_id / THREADS_PER_BLOCK; block < num_blocks; block += WG_SIZE / THREADS_PER_BLOCK) {
353
+ let x_base = src1_idx_base + block * BLOCK_SIZE + thread_within_block * ELEMS_PER_THREAD;
354
+ var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
355
+ for (var col = 0u; col < NUM_COLS;col += 1) {
356
+ for (var i = 0u; i < ELEMS_PER_THREAD; i++) {
357
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
358
+ }
359
+ }
360
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
361
+ let output_row = row_base + row;
362
+ if (output_row < params.m) {
363
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
364
+ let d = f32(load_f16_at_src0(block_byte_base));
365
+ let m = f32(load_f16_at_src0(block_byte_base + 2u));
366
+ var q_packed: array<u32, ELEMS_PER_THREAD / 4u>;
367
+ for (var packed_idx = 0u; packed_idx < ELEMS_PER_THREAD / 4u; packed_idx++) {
368
+ q_packed[packed_idx] = load_u32_at_src0(block_byte_base + 4u + 4u * (thread_within_block * 2u + packed_idx));
369
+ }
370
+ for (var col = 0u;col < NUM_COLS;col += 1) {
371
+ var row_sum = 0.0;
372
+ for (var packed_idx = 0u; packed_idx < ELEMS_PER_THREAD / 4u; packed_idx++) {
373
+ for (var byte_idx = 0u; byte_idx < 4u; byte_idx++) {
374
+ let q_val = f32(get_byte_i32(q_packed[packed_idx], byte_idx)) * d + m;
375
+ row_sum += q_val * x_block[col][packed_idx * 4u + byte_idx];
376
+ }
377
+ }
378
+ acc[col][row] += row_sum;
379
+ }
380
+ }
381
+ }
382
+ }
383
+
384
+ return acc;
385
+ }
386
+ #endif
387
+
388
+ #ifdef MUL_ACC_Q2_K
389
+ #define BLOCK_SIZE 256
390
+ #define BLOCK_SIZE_BYTES 84
391
+ #define THREADS_PER_BLOCK 16
392
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
393
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
394
+
395
+ let tid = thread_id % THREADS_PER_BLOCK;
396
+ let block_group = thread_id / THREADS_PER_BLOCK;
397
+ let num_block_groups: u32 = WG_SIZE / THREADS_PER_BLOCK;
398
+
399
+ let lane = tid / 2u;
400
+ let phase = tid % 2u;
401
+ let iq = lane / 4u;
402
+ let ir = lane % 4u;
403
+ let is = ir / 2u;
404
+
405
+ let y_offset = 128u * iq + 8u * ir + 4u * phase;
406
+ let sc0_byte = 8u * iq + is;
407
+ let sc2_byte = 8u * iq + is + 2u;
408
+ let sc4_byte = 8u * iq + is + 4u;
409
+ let sc6_byte = 8u * iq + is + 6u;
410
+ let qs_byte = 16u + (16u * iq + 4u * ir) * 2u + 4u * phase;
411
+
412
+ let num_blocks = params.k / BLOCK_SIZE;
413
+
414
+ for (var block = block_group; block < num_blocks; block += num_block_groups) {
415
+ let x_base = src1_idx_base + block * BLOCK_SIZE + y_offset;
416
+ var x_block: array<array<f32, 16>, NUM_COLS>;
417
+ for (var col = 0u; col < NUM_COLS;col += 1) {
418
+ for (var i = 0u; i < 4u; i++) {
419
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
420
+ x_block[col][i + 4u] = f32(SRC1[x_base + col * params.stride_11 + 32u + i]);
421
+ x_block[col][i + 8u] = f32(SRC1[x_base + col * params.stride_11 + 64u + i]);
422
+ x_block[col][i + 12u] = f32(SRC1[x_base + col * params.stride_11 + 96u + i]);
423
+ }
424
+ }
425
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
426
+ let output_row = row_base + row;
427
+ if (output_row < params.m) {
428
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
429
+
430
+ let dall = f32(load_f16_at_src0(block_byte_base + 80u));
431
+ let dmin = f32(load_f16_at_src0(block_byte_base + 82u)) * (1.0 / 16.0);
432
+
433
+ let sc0 = byte_of(load_u32_at_src0_aligned(block_byte_base + sc0_byte), sc0_byte & 3u);
434
+ let sc2 = byte_of(load_u32_at_src0_aligned(block_byte_base + sc2_byte), sc2_byte & 3u);
435
+ let sc4 = byte_of(load_u32_at_src0_aligned(block_byte_base + sc4_byte), sc4_byte & 3u);
436
+ let sc6 = byte_of(load_u32_at_src0_aligned(block_byte_base + sc6_byte), sc6_byte & 3u);
437
+
438
+ let q_u32 = load_u32_at_src0_aligned(block_byte_base + qs_byte);
439
+ let qs0 = q_u32 & 0xFFFFu;
440
+ let qs1 = q_u32 >> 16u;
441
+
442
+ for (var col = 0u;col < NUM_COLS;col += 1) {
443
+ var sumy = vec4<f32>(0.0, 0.0, 0.0, 0.0);
444
+ var acc1 = vec4<f32>(0.0, 0.0, 0.0, 0.0);
445
+ var acc2 = vec4<f32>(0.0, 0.0, 0.0, 0.0);
446
+
447
+ sumy[0] = x_block[col][0] + x_block[col][1] + x_block[col][2] + x_block[col][3];
448
+ sumy[1] = x_block[col][4] + x_block[col][5] + x_block[col][6] + x_block[col][7];
449
+ sumy[2] = x_block[col][8] + x_block[col][9] + x_block[col][10] + x_block[col][11];
450
+ sumy[3] = x_block[col][12] + x_block[col][13] + x_block[col][14] + x_block[col][15];
451
+
452
+ acc1[0] = x_block[col][0] * f32(qs0 & 0x0003u) + x_block[col][2] * f32(qs1 & 0x0003u);
453
+ acc2[0] = x_block[col][1] * f32(qs0 & 0x0300u) + x_block[col][3] * f32(qs1 & 0x0300u);
454
+ acc1[1] = x_block[col][4] * f32(qs0 & 0x000Cu) + x_block[col][6] * f32(qs1 & 0x000Cu);
455
+ acc2[1] = x_block[col][5] * f32(qs0 & 0x0C00u) + x_block[col][7] * f32(qs1 & 0x0C00u);
456
+ acc1[2] = x_block[col][8] * f32(qs0 & 0x0030u) + x_block[col][10] * f32(qs1 & 0x0030u);
457
+ acc2[2] = x_block[col][9] * f32(qs0 & 0x3000u) + x_block[col][11] * f32(qs1 & 0x3000u);
458
+ acc1[3] = x_block[col][12] * f32(qs0 & 0x00C0u) + x_block[col][14] * f32(qs1 & 0x00C0u);
459
+ acc2[3] = x_block[col][13] * f32(qs0 & 0xC000u) + x_block[col][15] * f32(qs1 & 0xC000u);
460
+
461
+ acc[col][row] += dall * ((acc1[0] + (1.0/256.0) * acc2[0]) * f32(sc0 & 0xFu) +
462
+ (acc1[1] + (1.0/256.0) * acc2[1]) * f32(sc2 & 0xFu) / 4.0 +
463
+ (acc1[2] + (1.0/256.0) * acc2[2]) * f32(sc4 & 0xFu) / 16.0 +
464
+ (acc1[3] + (1.0/256.0) * acc2[3]) * f32(sc6 & 0xFu) / 64.0)
465
+ - dmin * (sumy[0] * f32(sc0 & 0xF0u) + sumy[1] * f32(sc2 & 0xF0u) +
466
+ sumy[2] * f32(sc4 & 0xF0u) + sumy[3] * f32(sc6 & 0xF0u));
467
+ }
468
+ }
469
+ }
470
+ }
471
+
472
+ return acc;
473
+ }
474
+ #endif
475
+
476
+ #ifdef MUL_ACC_Q3_K
477
+ #define BLOCK_SIZE 256
478
+ #define BLOCK_SIZE_BYTES 110
479
+ #define THREADS_PER_BLOCK 16
480
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
481
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
482
+
483
+ let tid = thread_id % THREADS_PER_BLOCK;
484
+ let block_group = thread_id / THREADS_PER_BLOCK;
485
+ let num_block_groups: u32 = WG_SIZE / THREADS_PER_BLOCK;
486
+
487
+ let lane = tid / 2u;
488
+ let phase = tid % 2u;
489
+ let ip = lane / 4u;
490
+ let il = 2u * ((lane % 4u) / 2u);
491
+ let ir = lane % 2u;
492
+ let l0 = 8u * ir;
493
+
494
+ let q_byte = 32u + 32u * ip + l0 + 16u * phase;
495
+ let h_byte = l0 + 16u * phase;
496
+ let y_offset = 128u * ip + 32u * il + l0 + 16u * phase;
497
+
498
+ let s_shift1 = 4u * ip;
499
+ let s_shift2 = s_shift1 + il;
500
+
501
+ let v1 = select(64.0, 4.0, il == 0u);
502
+ let v2 = 4.0 * v1;
503
+ let shift = 2u * il;
504
+
505
+ var qm0: u32; var qm1: u32; var qm2: u32; var qm3: u32;
506
+ if (il == 0u) {
507
+ qm0 = 0x0003u; qm1 = 0x0300u; qm2 = 0x000Cu; qm3 = 0x0C00u;
508
+ } else {
509
+ qm0 = 0x0030u; qm1 = 0x3000u; qm2 = 0x00C0u; qm3 = 0xC000u;
510
+ }
511
+
512
+ let mm_idx = 2u * ip + il / 2u;
513
+ var hm0: u32; var hm1: u32; var hm2: u32; var hm3: u32;
514
+ switch (mm_idx) {
515
+ case 0u: { hm0=0x0001u; hm1=0x0100u; hm2=0x0002u; hm3=0x0200u; }
516
+ case 1u: { hm0=0x0004u; hm1=0x0400u; hm2=0x0008u; hm3=0x0800u; }
517
+ case 2u: { hm0=0x0010u; hm1=0x1000u; hm2=0x0020u; hm3=0x2000u; }
518
+ default: { hm0=0x0040u; hm1=0x4000u; hm2=0x0080u; hm3=0x8000u; }
519
+ }
520
+
521
+ let num_blocks = params.k / BLOCK_SIZE;
522
+
523
+ for (var block = block_group; block < num_blocks; block += num_block_groups) {
524
+ let x_base = src1_idx_base + block * BLOCK_SIZE + y_offset;
525
+ var x_block: array<array<f32, 16>, NUM_COLS>;
526
+ for (var col = 0u; col < NUM_COLS;col += 1) {
527
+ for (var i = 0u; i < 8u; i++) {
528
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
529
+ x_block[col][i + 8u] = f32(SRC1[x_base + col * params.stride_11 + 32u + i]);
530
+ }
531
+ }
532
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
533
+ let output_row = row_base + row;
534
+ if (output_row < params.m) {
535
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
536
+
537
+ let d = f32(load_f16_at_src0(block_byte_base + 108u));
538
+ let a_base = 96u;
539
+ let a_il0 = load_u16_at_src0(block_byte_base + a_base + il * 2u);
540
+ let a_il1 = load_u16_at_src0(block_byte_base + a_base + (il + 1u) * 2u);
541
+ let a_4 = load_u16_at_src0(block_byte_base + a_base + 8u);
542
+ let a_5 = load_u16_at_src0(block_byte_base + a_base + 10u);
543
+
544
+ var scales32 = a_4 | (a_5 << 16u);
545
+ let aux32 = ((scales32 >> s_shift2) << 4u) & 0x30303030u;
546
+ scales32 = a_il0 | (a_il1 << 16u);
547
+ scales32 = ((scales32 >> s_shift1) & 0x0F0F0F0Fu) | aux32;
548
+
549
+ let scale0 = f32(i32(byte_of(scales32, phase + 0u)) - 32);
550
+ let scale1 = f32(i32(byte_of(scales32, phase + 2u)) - 32);
551
+
552
+ let q_u32_0 = load_u32_at_src0(block_byte_base + q_byte + 0u);
553
+ let q_u32_1 = load_u32_at_src0(block_byte_base + q_byte + 4u);
554
+ let h_u32_0 = load_u32_at_src0(block_byte_base + h_byte + 0u);
555
+ let h_u32_1 = load_u32_at_src0(block_byte_base + h_byte + 4u);
556
+
557
+ for (var col = 0u;col < NUM_COLS;col += 1) {
558
+ var s1 = 0.0; var s2 = 0.0; var s3 = 0.0;
559
+ var s4 = 0.0; var s5 = 0.0; var s6 = 0.0;
560
+
561
+ for (var l = 0u; l < 8u; l += 2u) {
562
+ let q_u32 = select(q_u32_0, q_u32_1, l >= 4u);
563
+ let qs = select(q_u32 & 0xFFFFu, q_u32 >> 16u, (l & 2u) != 0u);
564
+ let h_u32 = select(h_u32_0, h_u32_1, l >= 4u);
565
+ let hv = select(h_u32 & 0xFFFFu, h_u32 >> 16u, (l & 2u) != 0u);
566
+
567
+ s1 += x_block[col][l + 0u] * f32(qs & qm0);
568
+ s2 += x_block[col][l + 1u] * f32(qs & qm1);
569
+ s3 += select(0.0, x_block[col][l + 0u], (hv & hm0) == 0u) +
570
+ select(0.0, x_block[col][l + 1u], (hv & hm1) == 0u);
571
+ s4 += x_block[col][l + 8u] * f32(qs & qm2);
572
+ s5 += x_block[col][l + 9u] * f32(qs & qm3);
573
+ s6 += select(0.0, x_block[col][l + 8u], (hv & hm2) == 0u) +
574
+ select(0.0, x_block[col][l + 9u], (hv & hm3) == 0u);
575
+ }
576
+
577
+ let d1 = d * (s1 + (1.0/256.0) * s2 - s3 * v1);
578
+ let d2 = d * (s4 + (1.0/256.0) * s5 - s6 * v2);
579
+ acc[col][row] += (d1 * scale0 + 0.25 * d2 * scale1) / f32(1u << shift);
580
+ }
581
+ }
582
+ }
583
+ }
584
+
585
+ return acc;
586
+ }
587
+ #endif
588
+
589
+ #ifdef MUL_ACC_Q4_K
590
+ #define BLOCK_SIZE 256
591
+ #define BLOCK_SIZE_BYTES 144
592
+ #define THREADS_PER_BLOCK 16
593
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
594
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
595
+
596
+ let tid = thread_id % THREADS_PER_BLOCK;
597
+ let block_group = thread_id / THREADS_PER_BLOCK;
598
+ let num_block_groups: u32 = WG_SIZE / THREADS_PER_BLOCK;
599
+
600
+ let il = tid / 4u;
601
+ let ir = tid % 4u;
602
+ let im = il / 2u;
603
+ let in = il % 2u;
604
+ let l0 = 4u * (2u * ir + in);
605
+
606
+ let y_offset = 64u * im + l0;
607
+ let q_offset = 32u * im + l0;
608
+ let sc0_byte = 4u + im * 2u;
609
+ let sc2_byte = 4u + (im + 2u) * 2u;
610
+ let sc4_byte = 4u + (im + 4u) * 2u;
611
+
612
+ let num_blocks = params.k / BLOCK_SIZE;
613
+
614
+ for (var block = block_group; block < num_blocks; block += num_block_groups) {
615
+ let x_base = src1_idx_base + block * BLOCK_SIZE + y_offset;
616
+ var x_block: array<array<f32, 16>, NUM_COLS>;
617
+ for (var col = 0u; col < NUM_COLS;col += 1) {
618
+ let col_base = x_base + col * params.stride_11;
619
+ for (var i = 0u; i < 4u; i++) {
620
+ x_block[col][i] = f32(SRC1[col_base + i]);
621
+ x_block[col][i + 4u] = f32(SRC1[col_base + 32u + i]);
622
+ x_block[col][i + 8u] = f32(SRC1[col_base + 128u + i]);
623
+ x_block[col][i + 12u] = f32(SRC1[col_base + 160u + i]);
624
+ }
625
+ }
626
+
627
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
628
+ let output_row = row_base + row;
629
+ if (output_row < params.m) {
630
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
631
+
632
+ let d = f32(load_f16_at_src0(block_byte_base + 0u));
633
+ let dmin = f32(load_f16_at_src0(block_byte_base + 2u));
634
+
635
+ let sc0_u32 = load_u32_at_src0_aligned(block_byte_base + sc0_byte);
636
+ let sc0 = select(sc0_u32 & 0xFFFFu, sc0_u32 >> 16u, (sc0_byte & 2u) != 0u);
637
+ let sc2_u32 = load_u32_at_src0_aligned(block_byte_base + sc2_byte);
638
+ let sc2 = select(sc2_u32 & 0xFFFFu, sc2_u32 >> 16u, (sc2_byte & 2u) != 0u);
639
+ let sc4_u32 = load_u32_at_src0_aligned(block_byte_base + sc4_byte);
640
+ let sc4 = select(sc4_u32 & 0xFFFFu, sc4_u32 >> 16u, (sc4_byte & 2u) != 0u);
641
+
642
+ let sc16_0 = sc0 & 0x3F3Fu;
643
+ let sc16_1 = sc2 & 0x3F3Fu;
644
+ let sc16_2 = (sc4 & 0x0F0Fu) | ((sc0 & 0xC0C0u) >> 2u);
645
+ let sc16_3 = ((sc4 >> 4u) & 0x0F0Fu) | ((sc2 & 0xC0C0u) >> 2u);
646
+
647
+ let scale0 = f32(sc16_0 & 0xFFu);
648
+ let scale1 = f32((sc16_0 >> 8u) & 0xFFu);
649
+ let min0 = f32(sc16_1 & 0xFFu);
650
+ let min1 = f32((sc16_1 >> 8u) & 0xFFu);
651
+ let scale2 = f32(sc16_2 & 0xFFu);
652
+ let scale3 = f32((sc16_2 >> 8u) & 0xFFu);
653
+ let min2 = f32(sc16_3 & 0xFFu);
654
+ let min3 = f32((sc16_3 >> 8u) & 0xFFu);
655
+
656
+ let q1_u32 = load_u32_at_src0_aligned(block_byte_base + 16u + q_offset);
657
+ let q2_u32 = load_u32_at_src0_aligned(block_byte_base + 80u + q_offset);
658
+
659
+ for (var col = 0u;col < NUM_COLS;col += 1) {
660
+ var dot = vec4<f32>(0.0, 0.0, 0.0, 0.0);
661
+ var sumx = vec4<f32>(0.0, 0.0, 0.0, 0.0);
662
+ for (var i = 0u; i < 4u; i++) {
663
+ let q1b = byte_of(q1_u32, i);
664
+ let q2b = byte_of(q2_u32, i);
665
+ dot[0] += x_block[col][i] * f32(q1b & 0x0Fu);
666
+ dot[1] += x_block[col][i + 4u] * f32(q1b >> 4u);
667
+ dot[2] += x_block[col][i + 8u] * f32(q2b & 0x0Fu);
668
+ dot[3] += x_block[col][i + 12u] * f32(q2b >> 4u);
669
+ sumx[0] += x_block[col][i];
670
+ sumx[1] += x_block[col][i + 4u];
671
+ sumx[2] += x_block[col][i + 8u];
672
+ sumx[3] += x_block[col][i + 12u];
673
+ }
674
+
675
+ acc[col][row] += d * (dot[0] * scale0 + dot[1] * scale1 + dot[2] * scale2 + dot[3] * scale3)
676
+ - dmin * (sumx[0] * min0 + sumx[1] * min1 + sumx[2] * min2 + sumx[3] * min3);
677
+ }
678
+ }
679
+ }
680
+ }
681
+
682
+ return acc;
683
+ }
684
+ #endif
685
+
686
+ #ifdef MUL_ACC_Q5_K
687
+ #define BLOCK_SIZE 256
688
+ #define BLOCK_SIZE_BYTES 176
689
+ #define THREADS_PER_BLOCK 16
690
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
691
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
692
+
693
+ let tid = thread_id % THREADS_PER_BLOCK;
694
+ let block_group = thread_id / THREADS_PER_BLOCK;
695
+ let num_block_groups: u32 = WG_SIZE / THREADS_PER_BLOCK;
696
+
697
+ let il = tid / 4u;
698
+ let ir = tid % 4u;
699
+ let im = il / 2u;
700
+ let in = il % 2u;
701
+ let l0 = 4u * (2u * ir + in);
702
+
703
+ let y_offset = 64u * im + l0;
704
+ let q_offset = 48u + 32u * im + l0;
705
+ let qh_offset = 16u + 8u * ir + 4u * in;
706
+ let sc0_byte = 4u + im * 2u;
707
+ let sc2_byte = 4u + (im + 2u) * 2u;
708
+ let sc4_byte = 4u + (im + 4u) * 2u;
709
+
710
+ let hm1 = 1u << (2u * im);
711
+ let hm2 = hm1 << 1u;
712
+ let hm3 = hm1 << 4u;
713
+ let hm4 = hm2 << 4u;
714
+
715
+ let num_blocks = params.k / BLOCK_SIZE;
716
+
717
+ for (var block = block_group; block < num_blocks; block += num_block_groups) {
718
+ let x_base = src1_idx_base + block * BLOCK_SIZE + y_offset;
719
+ var x_block: array<array<f32, 16>, NUM_COLS>;
720
+ for (var col = 0u; col < NUM_COLS;col += 1) {
721
+ let col_base = x_base + col * params.stride_11;
722
+ for (var i = 0u; i < 4u; i++) {
723
+ x_block[col][i] = f32(SRC1[col_base + i]);
724
+ x_block[col][i + 4u] = f32(SRC1[col_base + 32u + i]);
725
+ x_block[col][i + 8u] = f32(SRC1[col_base + 128u + i]);
726
+ x_block[col][i + 12u] = f32(SRC1[col_base + 160u + i]);
727
+ }
728
+ }
729
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
730
+ let output_row = row_base + row;
731
+ if (output_row < params.m) {
732
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
733
+
734
+ let d = f32(load_f16_at_src0(block_byte_base + 0u));
735
+ let dmin = f32(load_f16_at_src0(block_byte_base + 2u));
736
+
737
+ let sc0_u32 = load_u32_at_src0_aligned(block_byte_base + sc0_byte);
738
+ let sc0 = select(sc0_u32 & 0xFFFFu, sc0_u32 >> 16u, (sc0_byte & 2u) != 0u);
739
+ let sc2_u32 = load_u32_at_src0_aligned(block_byte_base + sc2_byte);
740
+ let sc2 = select(sc2_u32 & 0xFFFFu, sc2_u32 >> 16u, (sc2_byte & 2u) != 0u);
741
+ let sc4_u32 = load_u32_at_src0_aligned(block_byte_base + sc4_byte);
742
+ let sc4 = select(sc4_u32 & 0xFFFFu, sc4_u32 >> 16u, (sc4_byte & 2u) != 0u);
743
+
744
+ let sc16_0 = sc0 & 0x3F3Fu;
745
+ let sc16_1 = sc2 & 0x3F3Fu;
746
+ let sc16_2 = (sc4 & 0x0F0Fu) | ((sc0 & 0xC0C0u) >> 2u);
747
+ let sc16_3 = ((sc4 >> 4u) & 0x0F0Fu) | ((sc2 & 0xC0C0u) >> 2u);
748
+
749
+ let f0 = f32(sc16_0 & 0xFFu);
750
+ let f1 = f32((sc16_0 >> 8u) & 0xFFu);
751
+ let m0 = f32(sc16_1 & 0xFFu);
752
+ let m1 = f32((sc16_1 >> 8u) & 0xFFu);
753
+ let f4 = f32(sc16_2 & 0xFFu);
754
+ let f5 = f32((sc16_2 >> 8u) & 0xFFu);
755
+ let m4 = f32(sc16_3 & 0xFFu);
756
+ let m5 = f32((sc16_3 >> 8u) & 0xFFu);
757
+
758
+ let q1_u32 = load_u32_at_src0_aligned(block_byte_base + q_offset);
759
+ let q2_u32 = load_u32_at_src0_aligned(block_byte_base + q_offset + 64u);
760
+ let qh_u32 = load_u32_at_src0_aligned(block_byte_base + qh_offset);
761
+
762
+ for (var col = 0u;col < NUM_COLS;col += 1) {
763
+ var vals = vec4<f32>(0.0, 0.0, 0.0, 0.0);
764
+ var sumy = vec4<f32>(0.0, 0.0, 0.0, 0.0);
765
+ for (var i = 0u; i < 4u; i++) {
766
+ let q1b = byte_of(q1_u32, i);
767
+ let q2b = byte_of(q2_u32, i);
768
+ let qhb = byte_of(qh_u32, i);
769
+
770
+ let yl0 = x_block[col][i];
771
+ let yl8 = x_block[col][i + 4u];
772
+ let yh0 = x_block[col][i + 8u];
773
+ let yh8 = x_block[col][i + 12u];
774
+
775
+ sumy[0] += yl0;
776
+ sumy[1] += yl8;
777
+ sumy[2] += yh0;
778
+ sumy[3] += yh8;
779
+
780
+ let q0 = f32((q1b & 0x0Fu) | select(0u, 0x10u, (qhb & hm1) != 0u));
781
+ let q1 = f32((q1b >> 4u) | select(0u, 0x10u, (qhb & hm2) != 0u));
782
+ let q2 = f32((q2b & 0x0Fu) | select(0u, 0x10u, (qhb & hm3) != 0u));
783
+ let q3 = f32((q2b >> 4u) | select(0u, 0x10u, (qhb & hm4) != 0u));
784
+
785
+ vals[0] += yl0 * q0;
786
+ vals[1] += yl8 * q1;
787
+ vals[2] += yh0 * q2;
788
+ vals[3] += yh8 * q3;
789
+ }
790
+
791
+ acc[col][row] += d * (f0 * vals[0] + f1 * vals[1] + f4 * vals[2] + f5 * vals[3])
792
+ - dmin * (sumy[0] * m0 + sumy[1] * m1 +
793
+ sumy[2] * m4 + sumy[3] * m5);
794
+ }
795
+ }
796
+ }
797
+ }
798
+
799
+ return acc;
800
+ }
801
+ #endif
802
+
803
+ #ifdef MUL_ACC_Q6_K
804
+ #define BLOCK_SIZE 256
805
+ #define BLOCK_SIZE_BYTES 210
806
+ #define THREADS_PER_BLOCK 16
807
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
808
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
809
+
810
+ let tid = thread_id % THREADS_PER_BLOCK;
811
+ let block_group = thread_id / THREADS_PER_BLOCK;
812
+ let num_block_groups: u32 = WG_SIZE / THREADS_PER_BLOCK;
813
+
814
+ let ip = tid / 8u;
815
+ let il = tid % 8u;
816
+ let l0 = 4u * il;
817
+ let is = 8u * ip + l0 / 16u;
818
+
819
+ let y_offset = 128u * ip + l0;
820
+ let q_offset_l = 64u * ip + l0;
821
+ let q_offset_h = 32u * ip + l0;
822
+
823
+ let num_blocks = params.k / BLOCK_SIZE;
824
+ let sc_base_byte = 192u + (is & ~3u);
825
+ let sc_byte_pos = is & 3u;
826
+
827
+ for (var block = block_group; block < num_blocks; block += num_block_groups) {
828
+ let x_base = src1_idx_base + block * BLOCK_SIZE + y_offset;
829
+ var x_block: array<array<f32, 16>, NUM_COLS>;
830
+ for (var col = 0u; col < NUM_COLS;col += 1) {
831
+ let col_base = x_base + col * params.stride_11;
832
+ for (var l = 0u; l < 4u; l++) {
833
+ x_block[col][l] = f32(SRC1[col_base + l]);
834
+ x_block[col][l + 4u] = f32(SRC1[col_base + 32u + l]);
835
+ x_block[col][l + 8u] = f32(SRC1[col_base + 64u + l]);
836
+ x_block[col][l + 12u] = f32(SRC1[col_base + 96u + l]);
837
+ }
838
+ }
839
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
840
+ let output_row = row_base + row;
841
+ if (output_row < params.m) {
842
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
843
+
844
+ let d = f32(load_f16_at_src0(block_byte_base + 208u));
845
+ let ql1_u32 = load_u32_at_src0(block_byte_base + q_offset_l);
846
+ let ql2_u32 = load_u32_at_src0(block_byte_base + q_offset_l + 32u);
847
+ let qh_u32 = load_u32_at_src0(block_byte_base + 128u + q_offset_h);
848
+ let sc_u32_0 = load_u32_at_src0(block_byte_base + sc_base_byte);
849
+ let sc_u32_1 = load_u32_at_src0(block_byte_base + sc_base_byte + 4u);
850
+
851
+ let sc0 = sbyte_of(sc_u32_0, sc_byte_pos);
852
+ let sc2 = sbyte_of(sc_u32_0, sc_byte_pos + 2u);
853
+ let sc4 = sbyte_of(sc_u32_1, sc_byte_pos);
854
+ let sc6 = sbyte_of(sc_u32_1, sc_byte_pos + 2u);
855
+
856
+ for (var col = 0u;col < NUM_COLS;col += 1) {
857
+ var sums = vec4<f32>(0.0, 0.0, 0.0, 0.0);
858
+
859
+ for (var l = 0u; l < 4u; l++) {
860
+ let q1b = byte_of(ql1_u32, l);
861
+ let q2b = byte_of(ql2_u32, l);
862
+ let qhb = byte_of(qh_u32, l);
863
+
864
+ let dq0 = f32(i32((q1b & 0x0Fu) | ((qhb & 0x03u) << 4u)) - 32);
865
+ let dq1 = f32(i32((q2b & 0x0Fu) | ((qhb & 0x0Cu) << 2u)) - 32);
866
+ let dq2 = f32(i32((q1b >> 4u) | (qhb & 0x30u)) - 32);
867
+ let dq3 = f32(i32((q2b >> 4u) | ((qhb & 0xC0u) >> 2u)) - 32);
868
+
869
+ sums[0] += x_block[col][l] * dq0;
870
+ sums[1] += x_block[col][l + 4u] * dq1;
871
+ sums[2] += x_block[col][l + 8u] * dq2;
872
+ sums[3] += x_block[col][l + 12u] * dq3;
873
+ }
874
+
875
+ acc[col][row] += d * (sums[0] * f32(sc0) + sums[1] * f32(sc2) +
876
+ sums[2] * f32(sc4) + sums[3] * f32(sc6));
877
+ }
878
+ }
879
+ }
880
+ }
881
+
882
+ return acc;
883
+ }
884
+ #endif
885
+
886
+ #ifdef MUL_ACC_IQ1_S
887
+ #define BLOCK_SIZE 256
888
+ #define BLOCK_SIZE_BYTES 50
889
+ #define THREADS_PER_BLOCK 16
890
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
891
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
892
+
893
+ let tid = thread_id % THREADS_PER_BLOCK;
894
+ let block_group = thread_id / THREADS_PER_BLOCK;
895
+ let num_block_groups: u32 = WG_SIZE / THREADS_PER_BLOCK;
896
+
897
+ let sub_blk = tid / 2u;
898
+ let half = tid % 2u;
899
+ let slot0 = half * 2u;
900
+ let y_offset = sub_blk * 32u + slot0 * 8u;
901
+
902
+ let num_blocks = params.k / BLOCK_SIZE;
903
+
904
+ for (var block = block_group; block < num_blocks; block += num_block_groups) {
905
+ let x_base = src1_idx_base + block * BLOCK_SIZE + y_offset;
906
+ var x_block: array<array<f32, 16>, NUM_COLS>;
907
+ for (var col = 0u; col < NUM_COLS;col += 1) {
908
+ for (var i = 0u; i < 16u; i++) {
909
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
910
+ }
911
+ }
912
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
913
+ let output_row = row_base + row;
914
+ if (output_row < params.m) {
915
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
916
+
917
+ let d = f32(load_f16_at_src0(block_byte_base));
918
+ let qh = load_u32_at_src0(block_byte_base + 34u + sub_blk * 2u) & 0xFFFFu;
919
+ let dl = d * f32(2u * ((qh >> 12u) & 7u) + 1u);
920
+ let delta = select(IQ1_DELTA, -IQ1_DELTA, (qh & 0x8000u) != 0u);
921
+ let qs_w = load_u32_at_src0(block_byte_base + 2u + sub_blk * 4u);
922
+
923
+ for (var col = 0u;col < NUM_COLS;col += 1) {
924
+ var row_sum = 0.0;
925
+ for (var ll = 0u; ll < 2u; ll++) {
926
+ let l = slot0 + ll;
927
+ let qs_byte = get_byte(qs_w, l);
928
+ let ig = (qs_byte | (((qh >> (3u * l)) & 7u) << 8u)) * 8u;
929
+ let gw = iq1_grid[ig / 16u];
930
+ let bit_base = (ig % 16u) * 2u;
931
+ for (var j = 0u; j < 8u; j++) {
932
+ let g = (gw >> (bit_base + j * 2u)) & 3u;
933
+ let gs = select(f32(g), f32(g) - 4.0, (g & 2u) != 0u);
934
+ row_sum += dl * (gs + delta) * x_block[col][ll * 8u + j];
935
+ }
936
+ }
937
+ acc[col][row] += row_sum;
938
+ }
939
+ }
940
+ }
941
+ }
942
+
943
+ return acc;
944
+ }
945
+ #endif
946
+
947
+ #ifdef MUL_ACC_IQ1_M
948
+ #define BLOCK_SIZE 256
949
+ #define BLOCK_SIZE_BYTES 56
950
+ #define THREADS_PER_BLOCK 16
951
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
952
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
953
+
954
+ let tid = thread_id % THREADS_PER_BLOCK;
955
+ let block_group = thread_id / THREADS_PER_BLOCK;
956
+ let num_block_groups: u32 = WG_SIZE / THREADS_PER_BLOCK;
957
+
958
+ let sub_blk = tid / 2u;
959
+ let half = tid % 2u;
960
+ let slot0 = half * 2u;
961
+ let y_offset = sub_blk * 32u + slot0 * 8u;
962
+
963
+ let num_blocks = params.k / BLOCK_SIZE;
964
+
965
+ for (var block = block_group; block < num_blocks; block += num_block_groups) {
966
+ let x_base = src1_idx_base + block * BLOCK_SIZE + y_offset;
967
+ var x_block: array<array<f32, 16>, NUM_COLS>;
968
+ for (var col = 0u; col < NUM_COLS;col += 1) {
969
+ for (var i = 0u; i < 16u; i++) {
970
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
971
+ }
972
+ }
973
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
974
+ let output_row = row_base + row;
975
+ if (output_row < params.m) {
976
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
977
+
978
+ let sc_lo = load_u32_at_src0(block_byte_base + 48u);
979
+ let sc_hi = load_u32_at_src0(block_byte_base + 52u);
980
+ let sc0 = sc_lo & 0xFFFFu;
981
+ let sc1 = (sc_lo >> 16u) & 0xFFFFu;
982
+ let sc2 = sc_hi & 0xFFFFu;
983
+ let sc3 = (sc_hi >> 16u) & 0xFFFFu;
984
+ let d_bits = (sc0 >> 12u) | ((sc1 >> 8u) & 0xF0u) | ((sc2 >> 4u) & 0xF00u) | (sc3 & 0xF000u);
985
+ let d = f32(bitcast<vec2<f16>>(d_bits)[0]);
986
+
987
+ let sc_u16 = select(select(sc2, sc3, sub_blk >= 6u),
988
+ select(sc0, sc1, sub_blk >= 2u),
989
+ sub_blk < 4u);
990
+
991
+ let qs_w = load_u32_at_src0(block_byte_base + sub_blk * 4u);
992
+ let qh = load_u32_at_src0(block_byte_base + 32u + sub_blk * 2u) & 0xFFFFu;
993
+ let qh_lo = qh & 0xFFu;
994
+ let qh_hi = (qh >> 8u) & 0xFFu;
995
+
996
+ for (var col = 0u;col < NUM_COLS;col += 1) {
997
+ var row_sum = 0.0;
998
+ for (var ll = 0u; ll < 2u; ll++) {
999
+ let l = slot0 + ll;
1000
+ let bit_off = 6u * (sub_blk % 2u) + 3u * (l / 2u);
1001
+ let sub_scale = (sc_u16 >> bit_off) & 0x7u;
1002
+ let dl = d * f32(2u * sub_scale + 1u);
1003
+ let qh_byte = select(qh_lo, qh_hi, l >= 2u);
1004
+ let ll2 = l % 2u;
1005
+ let grid_idx = get_byte(qs_w, l) | (((qh_byte >> (4u * ll2)) & 7u) << 8u);
1006
+ let delta = select(IQ1_DELTA, -IQ1_DELTA, ((qh_byte >> (3u + 4u * ll2)) & 1u) != 0u);
1007
+ let ig = grid_idx * 8u;
1008
+ let gw = iq1_grid[ig / 16u];
1009
+ let bit_base = (ig % 16u) * 2u;
1010
+ for (var j = 0u; j < 8u; j++) {
1011
+ let g = (gw >> (bit_base + j * 2u)) & 3u;
1012
+ let gs = select(f32(g), f32(g) - 4.0, (g & 2u) != 0u);
1013
+ row_sum += dl * (gs + delta) * x_block[col][ll * 8u + j];
1014
+ }
1015
+ }
1016
+ acc[col][row] += row_sum;
1017
+ }
1018
+ }
1019
+ }
1020
+ }
1021
+
1022
+ return acc;
1023
+ }
1024
+ #endif
1025
+
1026
+ #ifdef MUL_ACC_IQ2_XXS
1027
+ #define BLOCK_SIZE 256
1028
+ #define BLOCK_SIZE_BYTES 66
1029
+ #define THREADS_PER_BLOCK 16
1030
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
1031
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
1032
+
1033
+ let tid = thread_id % THREADS_PER_BLOCK;
1034
+ let block_group = thread_id / THREADS_PER_BLOCK;
1035
+ let num_block_groups: u32 = WG_SIZE / THREADS_PER_BLOCK;
1036
+
1037
+ let sub_blk = tid / 2u;
1038
+ let half = tid % 2u;
1039
+ let slot0 = half * 2u;
1040
+ let y_offset = sub_blk * 32u + slot0 * 8u;
1041
+
1042
+ let num_blocks = params.k / BLOCK_SIZE;
1043
+
1044
+ for (var block = block_group; block < num_blocks; block += num_block_groups) {
1045
+ let x_base = src1_idx_base + block * BLOCK_SIZE + y_offset;
1046
+ var x_block: array<array<f32, 16>, NUM_COLS>;
1047
+ for (var col = 0u; col < NUM_COLS;col += 1) {
1048
+ for (var i = 0u; i < 16u; i++) {
1049
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
1050
+ }
1051
+ }
1052
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
1053
+ let output_row = row_base + row;
1054
+ if (output_row < params.m) {
1055
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
1056
+ let d = f32(load_f16_at_src0(block_byte_base));
1057
+ let aux_lo = load_u32_at_src0(block_byte_base + 2u + sub_blk * 8u);
1058
+ let aux_hi = load_u32_at_src0(block_byte_base + 2u + sub_blk * 8u + 4u);
1059
+ let ls = aux_hi >> 28u;
1060
+ let db = d * (0.5 + f32(ls)) * 0.25;
1061
+
1062
+ for (var col = 0u;col < NUM_COLS;col += 1) {
1063
+ var row_sum = 0.0;
1064
+ for (var ll = 0u; ll < 2u; ll++) {
1065
+ let l = slot0 + ll;
1066
+ let grid_idx = (aux_lo >> (8u * l)) & 0xFFu;
1067
+ let signs_idx = (aux_hi >> (7u * l)) & 0x7Fu;
1068
+ let signs = (ksigns_iq2xs[signs_idx / 4u] >> ((signs_idx % 4u) * 8u)) & 0xFFu;
1069
+ let gw_lo = iq2xxs_grid[grid_idx * 2u];
1070
+ let gw_hi = iq2xxs_grid[grid_idx * 2u + 1u];
1071
+ for (var j = 0u; j < 8u; j++) {
1072
+ let gw = select(gw_hi, gw_lo, j < 4u);
1073
+ let b = f32((gw >> ((j & 3u) * 8u)) & 0xFFu);
1074
+ let s = select(1.0, -1.0, ((signs >> j) & 1u) != 0u);
1075
+ row_sum += db * b * s * x_block[col][ll * 8u + j];
1076
+ }
1077
+ }
1078
+ acc[col][row] += row_sum;
1079
+ }
1080
+ }
1081
+ }
1082
+ }
1083
+
1084
+ return acc;
1085
+ }
1086
+ #endif
1087
+
1088
+ #ifdef MUL_ACC_IQ2_XS
1089
+ #define BLOCK_SIZE 256
1090
+ #define BLOCK_SIZE_BYTES 74
1091
+ #define THREADS_PER_BLOCK 16
1092
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
1093
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
1094
+
1095
+ let tid = thread_id % THREADS_PER_BLOCK;
1096
+ let block_group = thread_id / THREADS_PER_BLOCK;
1097
+ let num_block_groups: u32 = WG_SIZE / THREADS_PER_BLOCK;
1098
+
1099
+ let sub_blk = tid / 2u;
1100
+ let half = tid % 2u;
1101
+ let slot0 = half * 2u;
1102
+ let y_offset = sub_blk * 32u + slot0 * 8u;
1103
+
1104
+ let num_blocks = params.k / BLOCK_SIZE;
1105
+
1106
+ for (var block = block_group; block < num_blocks; block += num_block_groups) {
1107
+ let x_base = src1_idx_base + block * BLOCK_SIZE + y_offset;
1108
+ var x_block: array<array<f32, 16>, NUM_COLS>;
1109
+ for (var col = 0u; col < NUM_COLS;col += 1) {
1110
+ for (var i = 0u; i < 16u; i++) {
1111
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
1112
+ }
1113
+ }
1114
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
1115
+ let output_row = row_base + row;
1116
+ if (output_row < params.m) {
1117
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
1118
+ let d = f32(load_f16_at_src0(block_byte_base));
1119
+ let qs_lo = load_u32_at_src0(block_byte_base + 2u + sub_blk * 8u);
1120
+ let qs_hi = load_u32_at_src0(block_byte_base + 2u + sub_blk * 8u + 4u);
1121
+ let scales_word = load_u32_at_src0(block_byte_base + 66u + (sub_blk / 4u) * 4u);
1122
+ let scales_byte = get_byte(scales_word, sub_blk % 4u);
1123
+
1124
+ for (var col = 0u;col < NUM_COLS;col += 1) {
1125
+ var row_sum = 0.0;
1126
+ for (var ll = 0u; ll < 2u; ll++) {
1127
+ let l = slot0 + ll;
1128
+ let qs_word = select(qs_hi, qs_lo, l < 2u);
1129
+ let half2 = (l % 2u) * 16u;
1130
+ let qs_val = (qs_word >> half2) & 0xFFFFu;
1131
+ let grid_idx = qs_val & 0x1FFu;
1132
+ let signs_idx = (qs_val >> 9u) & 0x7Fu;
1133
+ let sub_scale = (scales_byte >> (4u * (l / 2u))) & 0xFu;
1134
+ let db = d * (0.5 + f32(sub_scale)) * 0.25;
1135
+ let signs = (ksigns_iq2xs[signs_idx / 4u] >> ((signs_idx % 4u) * 8u)) & 0xFFu;
1136
+ let gw_lo = iq2xs_grid[grid_idx * 2u];
1137
+ let gw_hi = iq2xs_grid[grid_idx * 2u + 1u];
1138
+ for (var j = 0u; j < 8u; j++) {
1139
+ let gw = select(gw_hi, gw_lo, j < 4u);
1140
+ let b = f32((gw >> ((j & 3u) * 8u)) & 0xFFu);
1141
+ let s = select(1.0, -1.0, ((signs >> j) & 1u) != 0u);
1142
+ row_sum += db * b * s * x_block[col][ll * 8u + j];
1143
+ }
1144
+ }
1145
+ acc[col][row] += row_sum;
1146
+ }
1147
+ }
1148
+ }
1149
+ }
1150
+
1151
+ return acc;
1152
+ }
1153
+ #endif
1154
+
1155
+ #ifdef MUL_ACC_IQ2_S
1156
+ #define BLOCK_SIZE 256
1157
+ #define BLOCK_SIZE_BYTES 82
1158
+ #define THREADS_PER_BLOCK 16
1159
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
1160
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
1161
+
1162
+ let tid = thread_id % THREADS_PER_BLOCK;
1163
+ let block_group = thread_id / THREADS_PER_BLOCK;
1164
+ let num_block_groups: u32 = WG_SIZE / THREADS_PER_BLOCK;
1165
+
1166
+ let sub_blk = tid / 2u;
1167
+ let half = tid % 2u;
1168
+ let slot0 = half * 2u;
1169
+ let y_offset = sub_blk * 32u + slot0 * 8u;
1170
+
1171
+ let num_blocks = params.k / BLOCK_SIZE;
1172
+
1173
+ for (var block = block_group; block < num_blocks; block += num_block_groups) {
1174
+ let x_base = src1_idx_base + block * BLOCK_SIZE + y_offset;
1175
+ var x_block: array<array<f32, 16>, NUM_COLS>;
1176
+ for (var col = 0u; col < NUM_COLS;col += 1) {
1177
+ for (var i = 0u; i < 16u; i++) {
1178
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
1179
+ }
1180
+ }
1181
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
1182
+ let output_row = row_base + row;
1183
+ if (output_row < params.m) {
1184
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
1185
+ let d = f32(load_f16_at_src0(block_byte_base));
1186
+ let qs_w = load_u32_at_src0(block_byte_base + 2u + sub_blk * 4u);
1187
+ let sg_w = load_u32_at_src0(block_byte_base + 34u + sub_blk * 4u);
1188
+ let qh_word = load_u32_at_src0(block_byte_base + 66u + (sub_blk / 4u) * 4u);
1189
+ let qh_byte = get_byte(qh_word, sub_blk % 4u);
1190
+ let sc_word = load_u32_at_src0(block_byte_base + 74u + (sub_blk / 4u) * 4u);
1191
+ let scales_byte = get_byte(sc_word, sub_blk % 4u);
1192
+
1193
+ for (var col = 0u;col < NUM_COLS;col += 1) {
1194
+ var row_sum = 0.0;
1195
+ for (var ll = 0u; ll < 2u; ll++) {
1196
+ let l = slot0 + ll;
1197
+ let qs_byte = get_byte(qs_w, l);
1198
+ let sign_byte = get_byte(sg_w, l);
1199
+ let grid_idx = qs_byte | (((qh_byte >> (2u * l)) & 3u) << 8u);
1200
+ let sub_scale = (scales_byte >> (4u * (l / 2u))) & 0xFu;
1201
+ let db = d * (0.5 + f32(sub_scale)) * 0.25;
1202
+ let gw_lo = iq2s_grid[grid_idx * 2u];
1203
+ let gw_hi = iq2s_grid[grid_idx * 2u + 1u];
1204
+ for (var j = 0u; j < 8u; j++) {
1205
+ let gw = select(gw_hi, gw_lo, j < 4u);
1206
+ let b = f32((gw >> ((j & 3u) * 8u)) & 0xFFu);
1207
+ let s = select(1.0, -1.0, ((sign_byte >> j) & 1u) != 0u);
1208
+ row_sum += db * b * s * x_block[col][ll * 8u + j];
1209
+ }
1210
+ }
1211
+ acc[col][row] += row_sum;
1212
+ }
1213
+ }
1214
+ }
1215
+ }
1216
+
1217
+ return acc;
1218
+ }
1219
+ #endif
1220
+
1221
+ #ifdef MUL_ACC_IQ3_XXS
1222
+ #define BLOCK_SIZE 256
1223
+ #define BLOCK_SIZE_BYTES 98
1224
+ #define THREADS_PER_BLOCK 16
1225
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
1226
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
1227
+
1228
+ let tid = thread_id % THREADS_PER_BLOCK;
1229
+ let block_group = thread_id / THREADS_PER_BLOCK;
1230
+ let num_block_groups: u32 = WG_SIZE / THREADS_PER_BLOCK;
1231
+
1232
+ let sub_blk = tid / 2u;
1233
+ let half = tid % 2u;
1234
+ let slot0 = half * 2u;
1235
+ let y_offset = sub_blk * 32u + slot0 * 8u;
1236
+
1237
+ let num_blocks = params.k / BLOCK_SIZE;
1238
+
1239
+ for (var block = block_group; block < num_blocks; block += num_block_groups) {
1240
+ let x_base = src1_idx_base + block * BLOCK_SIZE + y_offset;
1241
+ var x_block: array<array<f32, 16>, NUM_COLS>;
1242
+ for (var col = 0u; col < NUM_COLS;col += 1) {
1243
+ for (var i = 0u; i < 16u; i++) {
1244
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
1245
+ }
1246
+ }
1247
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
1248
+ let output_row = row_base + row;
1249
+ if (output_row < params.m) {
1250
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
1251
+ let d = f32(load_f16_at_src0(block_byte_base));
1252
+ let qs_lo = load_u32_at_src0(block_byte_base + 2u + sub_blk * 8u);
1253
+ let qs_hi = load_u32_at_src0(block_byte_base + 2u + sub_blk * 8u + 4u);
1254
+ let aux = load_u32_at_src0(block_byte_base + 66u + sub_blk * 4u);
1255
+ let ls = aux >> 28u;
1256
+ let db = d * (0.5 + f32(ls)) * 0.5;
1257
+
1258
+ for (var col = 0u;col < NUM_COLS;col += 1) {
1259
+ var row_sum = 0.0;
1260
+ for (var ll = 0u; ll < 2u; ll++) {
1261
+ let l = slot0 + ll;
1262
+ let qs_word = select(qs_hi, qs_lo, l < 2u);
1263
+ let byte_pos = (l % 2u) * 2u;
1264
+ let grid_idx_0 = (qs_word >> (byte_pos * 8u)) & 0xFFu;
1265
+ let grid_idx_1 = (qs_word >> ((byte_pos + 1u) * 8u)) & 0xFFu;
1266
+ let signs_idx = (aux >> (7u * l)) & 0x7Fu;
1267
+ let signs = (ksigns_iq2xs[signs_idx / 4u] >> ((signs_idx % 4u) * 8u)) & 0xFFu;
1268
+ let grid1 = iq3xxs_grid[grid_idx_0];
1269
+ let grid2 = iq3xxs_grid[grid_idx_1];
1270
+ for (var j = 0u; j < 4u; j++) {
1271
+ let b1 = f32((grid1 >> (j * 8u)) & 0xFFu);
1272
+ let b2 = f32((grid2 >> (j * 8u)) & 0xFFu);
1273
+ let s1 = select(1.0, -1.0, ((signs >> j) & 1u) != 0u);
1274
+ let s2 = select(1.0, -1.0, ((signs >> (j + 4u)) & 1u) != 0u);
1275
+ row_sum += db * b1 * s1 * x_block[col][ll * 8u + j];
1276
+ row_sum += db * b2 * s2 * x_block[col][ll * 8u + j + 4u];
1277
+ }
1278
+ }
1279
+ acc[col][row] += row_sum;
1280
+ }
1281
+ }
1282
+ }
1283
+ }
1284
+
1285
+ return acc;
1286
+ }
1287
+ #endif
1288
+
1289
+ #ifdef MUL_ACC_IQ3_S
1290
+ #define BLOCK_SIZE 256
1291
+ #define BLOCK_SIZE_BYTES 110
1292
+ #define THREADS_PER_BLOCK 16
1293
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
1294
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
1295
+
1296
+ let tid = thread_id % THREADS_PER_BLOCK;
1297
+ let block_group = thread_id / THREADS_PER_BLOCK;
1298
+ let num_block_groups: u32 = WG_SIZE / THREADS_PER_BLOCK;
1299
+
1300
+ let sub_blk = tid / 2u;
1301
+ let half = tid % 2u;
1302
+ let slot0 = half * 2u;
1303
+ let y_offset = sub_blk * 32u + slot0 * 8u;
1304
+
1305
+ let num_blocks = params.k / BLOCK_SIZE;
1306
+
1307
+ for (var block = block_group; block < num_blocks; block += num_block_groups) {
1308
+ let x_base = src1_idx_base + block * BLOCK_SIZE + y_offset;
1309
+ var x_block: array<array<f32, 16>, NUM_COLS>;
1310
+ for (var col = 0u; col < NUM_COLS;col += 1) {
1311
+ for (var i = 0u; i < 16u; i++) {
1312
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
1313
+ }
1314
+ }
1315
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
1316
+ let output_row = row_base + row;
1317
+ if (output_row < params.m) {
1318
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
1319
+ let d = f32(load_f16_at_src0(block_byte_base));
1320
+ let qs_lo = load_u32_at_src0(block_byte_base + 2u + sub_blk * 8u);
1321
+ let qs_hi = load_u32_at_src0(block_byte_base + 2u + sub_blk * 8u + 4u);
1322
+ let qh_word = load_u32_at_src0(block_byte_base + 66u + (sub_blk / 4u) * 4u);
1323
+ let qh_byte = get_byte(qh_word, sub_blk % 4u);
1324
+ let sg_w = load_u32_at_src0(block_byte_base + 74u + sub_blk * 4u);
1325
+ let sc_word = load_u32_at_src0(block_byte_base + 106u);
1326
+ let scales_byte = get_byte(sc_word, sub_blk / 2u);
1327
+ let sub_scale = (scales_byte >> (4u * (sub_blk % 2u))) & 0xFu;
1328
+ let db = d * (1.0 + 2.0 * f32(sub_scale));
1329
+
1330
+ for (var col = 0u;col < NUM_COLS;col += 1) {
1331
+ var row_sum = 0.0;
1332
+ for (var ll = 0u; ll < 2u; ll++) {
1333
+ let l = slot0 + ll;
1334
+ let qs_word = select(qs_hi, qs_lo, l < 2u);
1335
+ let byte_pos = (l % 2u) * 2u;
1336
+ let qs0 = (qs_word >> (byte_pos * 8u)) & 0xFFu;
1337
+ let qs1 = (qs_word >> ((byte_pos + 1u) * 8u)) & 0xFFu;
1338
+ let grid_idx_1 = qs0 | (((qh_byte >> (2u * l)) & 1u) << 8u);
1339
+ let grid_idx_2 = qs1 | (((qh_byte >> (2u * l + 1u)) & 1u) << 8u);
1340
+ let sign_byte = get_byte(sg_w, l);
1341
+ let grid1 = iq3s_grid[grid_idx_1];
1342
+ let grid2 = iq3s_grid[grid_idx_2];
1343
+ for (var j = 0u; j < 4u; j++) {
1344
+ let b1 = f32((grid1 >> (j * 8u)) & 0xFFu);
1345
+ let b2 = f32((grid2 >> (j * 8u)) & 0xFFu);
1346
+ let s1 = select(1.0, -1.0, ((sign_byte >> j) & 1u) != 0u);
1347
+ let s2 = select(1.0, -1.0, ((sign_byte >> (j + 4u)) & 1u) != 0u);
1348
+ row_sum += db * b1 * s1 * x_block[col][ll * 8u + j];
1349
+ row_sum += db * b2 * s2 * x_block[col][ll * 8u + j + 4u];
1350
+ }
1351
+ }
1352
+ acc[col][row] += row_sum;
1353
+ }
1354
+ }
1355
+ }
1356
+ }
1357
+
1358
+ return acc;
1359
+ }
1360
+ #endif
1361
+
1362
+ #ifdef MUL_ACC_IQ4_NL
1363
+ #define BLOCK_SIZE 32
1364
+ #define BLOCK_SIZE_BYTES 18
1365
+ #define THREADS_PER_BLOCK 4
1366
+ #define ELEMS_PER_THREAD (BLOCK_SIZE/THREADS_PER_BLOCK)
1367
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
1368
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
1369
+
1370
+ let num_blocks = params.k / BLOCK_SIZE;
1371
+ let thread_within_block = thread_id % THREADS_PER_BLOCK;
1372
+ for (var block = thread_id / THREADS_PER_BLOCK; block < num_blocks; block += WG_SIZE / THREADS_PER_BLOCK) {
1373
+ let x_base = src1_idx_base + block * BLOCK_SIZE + thread_within_block * 4u;
1374
+ var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
1375
+ for (var col = 0u; col < NUM_COLS;col += 1) {
1376
+ for (var i = 0u; i < ELEMS_PER_THREAD / 2u; i++) {
1377
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
1378
+ x_block[col][i + 4u] = f32(SRC1[x_base + col * params.stride_11 + i + 16u]);
1379
+ }
1380
+ }
1381
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
1382
+ let output_row = row_base + row;
1383
+ if (output_row < params.m) {
1384
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
1385
+ let d = f32(load_f16_at_src0(block_byte_base));
1386
+ let q_packed = load_u32_at_src0(block_byte_base + 2u + 4u * thread_within_block);
1387
+ for (var col = 0u;col < NUM_COLS;col += 1) {
1388
+ var row_sum = 0.0;
1389
+ for (var byte_idx = 0u; byte_idx < 4u; byte_idx++) {
1390
+ let q_byte = get_byte(q_packed, byte_idx);
1391
+ let q_lo = f32(kvalues_iq4nl[q_byte & 0xFu]) * d;
1392
+ let q_hi = f32(kvalues_iq4nl[(q_byte >> 4u) & 0xFu]) * d;
1393
+ row_sum += q_lo * x_block[col][byte_idx];
1394
+ row_sum += q_hi * x_block[col][byte_idx + 4u];
1395
+ }
1396
+ acc[col][row] += row_sum;
1397
+ }
1398
+ }
1399
+ }
1400
+ }
1401
+
1402
+ return acc;
1403
+ }
1404
+ #endif
1405
+
1406
+ #ifdef MUL_ACC_IQ4_XS
1407
+ #define BLOCK_SIZE 256
1408
+ #define BLOCK_SIZE_BYTES 136
1409
+ #define THREADS_PER_BLOCK 16
1410
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
1411
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
1412
+
1413
+ let tid = thread_id % THREADS_PER_BLOCK;
1414
+ let block_group = thread_id / THREADS_PER_BLOCK;
1415
+ let num_block_groups: u32 = WG_SIZE / THREADS_PER_BLOCK;
1416
+
1417
+ let sub_blk = tid / 2u;
1418
+ let half = tid % 2u;
1419
+ let y_offset = sub_blk * 32u + half * 16u;
1420
+
1421
+ let num_blocks = params.k / BLOCK_SIZE;
1422
+
1423
+ for (var block = block_group; block < num_blocks; block += num_block_groups) {
1424
+ let x_base = src1_idx_base + block * BLOCK_SIZE + y_offset;
1425
+ var x_block: array<array<f32, 16>, NUM_COLS>;
1426
+ for (var col = 0u; col < NUM_COLS;col += 1) {
1427
+ for (var i = 0u; i < 16u; i++) {
1428
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
1429
+ }
1430
+ }
1431
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
1432
+ let output_row = row_base + row;
1433
+ if (output_row < params.m) {
1434
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
1435
+ let d = f32(load_f16_at_src0(block_byte_base));
1436
+ let scales_h = load_u16_at_src0(block_byte_base + 2u);
1437
+ let scales_l_word = load_u32_at_src0(block_byte_base + 4u);
1438
+ let sl_byte = get_byte(scales_l_word, sub_blk / 2u);
1439
+ let sl = (sl_byte >> (4u * (sub_blk % 2u))) & 0xFu;
1440
+ let sh_bits = (scales_h >> (2u * sub_blk)) & 3u;
1441
+ let ls = i32(sl | (sh_bits << 4u));
1442
+ let dl = d * f32(ls - 32);
1443
+
1444
+ let qs_byte_off = 8u + sub_blk * 16u;
1445
+ let q_w0 = load_u32_at_src0(block_byte_base + qs_byte_off);
1446
+ let q_w1 = load_u32_at_src0(block_byte_base + qs_byte_off + 4u);
1447
+ let q_w2 = load_u32_at_src0(block_byte_base + qs_byte_off + 8u);
1448
+ let q_w3 = load_u32_at_src0(block_byte_base + qs_byte_off + 12u);
1449
+
1450
+ for (var col = 0u;col < NUM_COLS;col += 1) {
1451
+ var row_sum = 0.0;
1452
+ for (var i = 0u; i < 16u; i++) {
1453
+ let q_word = select(
1454
+ select(q_w0, q_w1, i >= 4u),
1455
+ select(q_w2, q_w3, i >= 12u),
1456
+ i >= 8u);
1457
+ let q_byte = get_byte(q_word, i % 4u);
1458
+ let nib = select(q_byte & 0xFu, (q_byte >> 4u) & 0xFu, half == 1u);
1459
+ row_sum += f32(kvalues_iq4nl[nib]) * dl * x_block[col][i];
1460
+ }
1461
+ acc[col][row] += row_sum;
1462
+ }
1463
+ }
1464
+ }
1465
+ }
1466
+
1467
+ return acc;
1468
+ }
1469
+ #endif
1470
+
1471
+ #ifdef MUL_ACC_MXFP4
1472
+ #define BLOCK_SIZE 32
1473
+ #define BLOCK_SIZE_BYTES 17
1474
+ #define THREADS_PER_BLOCK 4
1475
+ #define ELEMS_PER_THREAD (BLOCK_SIZE/THREADS_PER_BLOCK)
1476
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
1477
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
1478
+
1479
+ let num_blocks = params.k / BLOCK_SIZE;
1480
+ let thread_within_block = thread_id % 4;
1481
+ for (var block = thread_id/THREADS_PER_BLOCK; block < num_blocks; block += WG_SIZE/THREADS_PER_BLOCK) {
1482
+ let x_base = src1_idx_base + block * BLOCK_SIZE + thread_within_block * 4;
1483
+ var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
1484
+ for (var col = 0u; col < NUM_COLS;col += 1) {
1485
+ for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
1486
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
1487
+ x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
1488
+ }
1489
+ }
1490
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
1491
+ let output_row = row_base + row;
1492
+ if (output_row < params.m) {
1493
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
1494
+ let eu8 = get_byte(load_u32_at_src0(block_byte_base), 0);
1495
+ let e = ldexp(1.0, i32(eu8) - 128);
1496
+ let q_packed = load_u32_at_src0(block_byte_base + 1u + 4u * thread_within_block);
1497
+ for (var col = 0u;col < NUM_COLS;col += 1) {
1498
+ var row_sum = 0.0;
1499
+ for (var byte_idx = 0u; byte_idx < 4u; byte_idx++) {
1500
+ let q_byte = get_byte(q_packed, byte_idx);
1501
+ let q_lo = f32(kvalues_mxfp4[q_byte & 0xFu]) * e;
1502
+ let q_hi = f32(kvalues_mxfp4[(q_byte >> 4u) & 0xFu]) * e;
1503
+ row_sum += q_lo * x_block[col][byte_idx];
1504
+ row_sum += q_hi * x_block[col][byte_idx + 4u];
1505
+ }
1506
+ acc[col][row] += row_sum;
1507
+ }
1508
+ }
1509
+ }
1510
+ }
1511
+
1512
+ return acc;
1513
+ }
1514
+ #endif
1515
+
1516
+ #ifdef MUL_ACC_NVFP4
1517
+ #define BLOCK_SIZE 64
1518
+ #define BLOCK_SIZE_BYTES 36
1519
+ #define THREADS_PER_BLOCK 4
1520
+ #define ELEMS_PER_THREAD (BLOCK_SIZE/THREADS_PER_BLOCK)
1521
+ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src1_idx_base: u32) -> array<array<f32, OUTPUTS_PER_WG>, NUM_COLS> {
1522
+ var acc: array<array<f32, OUTPUTS_PER_WG>, NUM_COLS>;
1523
+
1524
+ let num_blocks = params.k / BLOCK_SIZE;
1525
+ let sub = thread_id % THREADS_PER_BLOCK;
1526
+ for (var block = thread_id/THREADS_PER_BLOCK; block < num_blocks; block += WG_SIZE/THREADS_PER_BLOCK) {
1527
+ let x_base = src1_idx_base + block * BLOCK_SIZE + sub * ELEMS_PER_THREAD;
1528
+ var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
1529
+ for (var col = 0u; col < NUM_COLS;col += 1) {
1530
+ for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
1531
+ x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
1532
+ x_block[col][i + 8] = f32(SRC1[x_base + col * params.stride_11 + i + 8]);
1533
+ }
1534
+ }
1535
+ for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
1536
+ let output_row = row_base + row;
1537
+ if (output_row < params.m) {
1538
+ let block_byte_base = (src0_batch_offset + output_row * params.stride_01 + block) * BLOCK_SIZE_BYTES;
1539
+ let d = ue4m3_to_fp32(get_byte(load_u32_at_src0_aligned(block_byte_base), sub)) * 0.5;
1540
+ let q_w0 = load_u32_at_src0_aligned(block_byte_base + 4u + 8u * sub);
1541
+ let q_w1 = load_u32_at_src0_aligned(block_byte_base + 8u + 8u * sub);
1542
+ for (var col = 0u;col < NUM_COLS;col += 1) {
1543
+ var row_sum = 0.0;
1544
+ for (var l = 0u; l < 8u; l++) {
1545
+ let q_word = select(q_w0, q_w1, l >= 4u);
1546
+ let q_byte = get_byte(q_word, l % 4u);
1547
+ let q_lo = f32(kvalues_mxfp4[q_byte & 0xFu]) * d;
1548
+ let q_hi = f32(kvalues_mxfp4[(q_byte >> 4u) & 0xFu]) * d;
1549
+ row_sum += q_lo * x_block[col][l];
1550
+ row_sum += q_hi * x_block[col][l + 8u];
1551
+ }
1552
+ acc[col][row] += row_sum;
1553
+ }
1554
+ }
1555
+ }
1556
+ }
1557
+
1558
+ return acc;
1559
+ }
1560
+ #endif