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 +1,1355 @@
1
+ #include "common.cuh"
2
+ #include "fattn-common.cuh"
3
+
4
+ // nbatch_fa == number of KQ rows to process per iteration
5
+ // nbatch_K == number of K columns to load in parallel for KQ calculation
6
+
7
+ // TODO optimize kernel parameters for FP16 NVIDIA (P100)
8
+ // TODO optimize kernel parameters for head sizes 40, 72, 80, 96, 112
9
+
10
+ // The ROCm compiler cannot handle templating in __launch_bounds__.
11
+ // As a workaround, define a macro to package the kernel parameters as uint32_t:
12
+ #define GGML_CUDA_FATTN_TILE_CONFIG_CASE(DKQ_, DV_, ncols_, nthreads, occupancy, nbatch_fa, nbatch_K) \
13
+ if (DKQ == (DKQ_) && DV == (DV_) && ncols == (ncols_)) { \
14
+ static_assert((nthreads) <= 512, "bad nthreads"); \
15
+ static_assert((occupancy) <= 8, "bad occupancy"); \
16
+ static_assert((nbatch_fa) <= 256, "bad nbatch_fa"); \
17
+ static_assert((nbatch_K) <= 256, "bad nbatch_K"); \
18
+ return ((nthreads) << 0) | ((occupancy) << 10) | ((nbatch_fa) << 14) | ((nbatch_K) << 23); \
19
+ } \
20
+
21
+ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nvidia_fp16(const int DKQ, const int DV, const int ncols) {
22
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 2, 64, 2, 64, 40)
23
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 4, 128, 2, 64, 40)
24
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 8, 256, 2, 64, 40)
25
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 16, 256, 2, 64, 40)
26
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 32, 256, 2, 64, 40)
27
+
28
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 2, 64, 2, 64, 64)
29
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 4, 128, 2, 64, 64)
30
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 8, 256, 2, 64, 64)
31
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 16, 256, 2, 64, 64)
32
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 32, 256, 2, 64, 64)
33
+
34
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 2, 64, 2, 64, 72)
35
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 4, 128, 2, 64, 72)
36
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 8, 256, 2, 64, 72)
37
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 16, 256, 2, 64, 72)
38
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 32, 256, 2, 64, 72)
39
+
40
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 2, 64, 2, 64, 40)
41
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 4, 128, 2, 64, 40)
42
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 8, 256, 2, 64, 40)
43
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 16, 256, 2, 64, 40)
44
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 32, 256, 2, 64, 40)
45
+
46
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 2, 64, 2, 64, 48)
47
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 4, 128, 2, 64, 48)
48
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 8, 256, 2, 64, 48)
49
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 16, 256, 2, 64, 48)
50
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 32, 256, 2, 64, 48)
51
+
52
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 2, 64, 2, 64, 56)
53
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 4, 128, 2, 64, 56)
54
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 8, 256, 2, 64, 56)
55
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 16, 256, 2, 64, 56)
56
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 32, 256, 2, 64, 56)
57
+
58
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 2, 64, 2, 64, 64)
59
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 4, 128, 2, 64, 64)
60
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 8, 256, 2, 64, 64)
61
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 16, 256, 2, 64, 64)
62
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 2, 64, 64)
63
+
64
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 2, 64, 2, 64, 64)
65
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 4, 128, 2, 64, 64)
66
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 8, 256, 2, 64, 64)
67
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 16, 256, 2, 64, 64)
68
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 32, 256, 2, 64, 64)
69
+
70
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 2, 64, 2, 64, 64)
71
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 4, 128, 2, 64, 64)
72
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 8, 256, 2, 64, 64)
73
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 64, 64)
74
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 64, 64)
75
+
76
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 16, 256, 2, 64, 64)
77
+
78
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 64, 64)
79
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 64, 64)
80
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
81
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 64, 64)
82
+
83
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 64, 64)
84
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 64, 64)
85
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 64, 64)
86
+
87
+ return 0;
88
+ }
89
+
90
+ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nvidia_fp32(const int DKQ, const int DV, const int ncols) {
91
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 2, 64, 2, 32, 40)
92
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 4, 128, 2, 32, 40)
93
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 8, 256, 2, 32, 40)
94
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 16, 256, 2, 32, 40)
95
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 32, 256, 2, 32, 40)
96
+
97
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 2, 128, 3, 64, 64)
98
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 4, 128, 3, 32, 64)
99
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 8, 128, 3, 32, 64)
100
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 16, 128, 3, 64, 64)
101
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 32, 256, 2, 64, 64)
102
+
103
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 2, 64, 2, 32, 72)
104
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 4, 128, 2, 32, 72)
105
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 8, 256, 2, 32, 72)
106
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 16, 256, 2, 32, 72)
107
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 32, 256, 2, 32, 72)
108
+
109
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 2, 64, 2, 32, 40)
110
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 4, 128, 2, 32, 40)
111
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 8, 256, 2, 32, 40)
112
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 16, 256, 2, 32, 40)
113
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 32, 256, 2, 32, 40)
114
+
115
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 2, 64, 2, 32, 48)
116
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 4, 128, 2, 32, 48)
117
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 8, 256, 2, 32, 48)
118
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 16, 256, 2, 32, 48)
119
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 32, 256, 2, 32, 48)
120
+
121
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 2, 64, 2, 32, 56)
122
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 4, 128, 2, 32, 56)
123
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 8, 256, 2, 32, 56)
124
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 16, 256, 2, 32, 56)
125
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 32, 256, 2, 32, 56)
126
+
127
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 2, 128, 3, 64, 64)
128
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 4, 128, 3, 32, 128)
129
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 8, 128, 3, 64, 128)
130
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 16, 128, 3, 32, 128)
131
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 2, 64, 64)
132
+
133
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 2, 128, 3, 64, 64)
134
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 4, 128, 3, 32, 64)
135
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 8, 256, 2, 32, 64)
136
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 16, 256, 2, 32, 64)
137
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 32, 256, 2, 32, 64)
138
+
139
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 2, 128, 3, 64, 64)
140
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 4, 128, 3, 32, 64)
141
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 8, 256, 2, 32, 256)
142
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 32, 128)
143
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 32, 64)
144
+
145
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 16, 256, 2, 32, 64)
146
+
147
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 32, 64)
148
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 32, 64)
149
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 32, 64)
150
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 32, 64)
151
+
152
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 32, 64)
153
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 32, 64)
154
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 32, 64)
155
+
156
+ return 0;
157
+ }
158
+
159
+ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_amd(const int DKQ, const int DV, const int ncols) {
160
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 2, 64, 2, 32, 40)
161
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 4, 128, 2, 32, 40)
162
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 8, 256, 2, 32, 40)
163
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 16, 256, 2, 32, 40)
164
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 32, 256, 2, 32, 40)
165
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 64, 256, 2, 32, 40)
166
+
167
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 2, 64, 3, 32, 64)
168
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 4, 128, 3, 64, 64)
169
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 8, 128, 2, 32, 64)
170
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 16, 256, 2, 128, 64)
171
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 32, 256, 2, 64, 64)
172
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 64, 256, 2, 64, 64)
173
+
174
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 2, 64, 2, 32, 72)
175
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 4, 128, 2, 32, 72)
176
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 8, 256, 2, 32, 72)
177
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 16, 256, 2, 32, 72)
178
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 32, 256, 2, 32, 72)
179
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 64, 256, 2, 32, 72)
180
+
181
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 2, 64, 2, 32, 40)
182
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 4, 128, 2, 32, 40)
183
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 8, 256, 2, 32, 40)
184
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 16, 256, 2, 32, 40)
185
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 32, 256, 2, 32, 40)
186
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 64, 256, 2, 32, 40)
187
+
188
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 2, 64, 2, 32, 48)
189
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 4, 128, 2, 32, 48)
190
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 8, 256, 2, 32, 48)
191
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 16, 256, 2, 32, 48)
192
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 32, 256, 2, 32, 48)
193
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 64, 256, 2, 32, 48)
194
+
195
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 2, 64, 2, 32, 56)
196
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 4, 128, 2, 32, 56)
197
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 8, 256, 2, 32, 56)
198
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 16, 256, 2, 32, 56)
199
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 32, 256, 2, 32, 56)
200
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 64, 256, 2, 32, 56)
201
+
202
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 2, 256, 2, 128, 64)
203
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 4, 128, 2, 64, 128)
204
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 8, 256, 2, 64, 128)
205
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 16, 256, 2, 64, 128)
206
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 2, 64, 64)
207
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 64, 256, 2, 64, 32)
208
+
209
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 2, 256, 2, 128, 64)
210
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 4, 256, 2, 64, 64)
211
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 8, 256, 2, 64, 64)
212
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 16, 256, 2, 32, 64)
213
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 32, 256, 2, 32, 64)
214
+
215
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 2, 256, 2, 128, 64)
216
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 4, 256, 2, 64, 128)
217
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 8, 256, 2, 64, 128)
218
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 32, 128)
219
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 32, 128)
220
+
221
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 32, 512, 1, 128, 64)
222
+
223
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 64, 64)
224
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 64, 64)
225
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
226
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 64, 64)
227
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 32, 512, 1, 128, 64)
228
+
229
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 64, 64)
230
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 64, 64)
231
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 64, 64)
232
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 512, 1, 128, 64)
233
+
234
+ return 0;
235
+ }
236
+
237
+ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_amd_rdna(const int DKQ, const int DV, const int ncols) {
238
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 2, 64, 2, 32, 40)
239
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 4, 128, 2, 32, 40)
240
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 8, 256, 2, 32, 40)
241
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 16, 256, 2, 32, 40)
242
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 32, 256, 2, 32, 40)
243
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 40, 40, 64, 256, 2, 32, 40)
244
+
245
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 2, 64, 8, 32, 64)
246
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 4, 64, 8, 32, 64)
247
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 8, 128, 5, 128, 64)
248
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 16, 128, 5, 128, 64)
249
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 32, 128, 4, 64, 64)
250
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64, 64, 64, 128, 5, 64, 64)
251
+
252
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 2, 64, 2, 32, 72)
253
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 4, 128, 2, 32, 72)
254
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 8, 256, 2, 32, 72)
255
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 16, 256, 2, 32, 72)
256
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 32, 256, 2, 32, 72)
257
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72, 72, 64, 256, 2, 32, 72)
258
+
259
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 2, 64, 2, 32, 40)
260
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 4, 128, 2, 32, 40)
261
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 8, 256, 2, 32, 40)
262
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 16, 256, 2, 32, 40)
263
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 32, 256, 2, 32, 40)
264
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80, 80, 64, 256, 2, 32, 40)
265
+
266
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 2, 64, 2, 32, 48)
267
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 4, 128, 2, 32, 48)
268
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 8, 256, 2, 32, 48)
269
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 16, 256, 2, 32, 48)
270
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 32, 256, 2, 32, 48)
271
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE( 96, 96, 64, 256, 2, 32, 48)
272
+
273
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 2, 64, 2, 32, 56)
274
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 4, 128, 2, 32, 56)
275
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 8, 256, 2, 32, 56)
276
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 16, 256, 2, 32, 56)
277
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 32, 256, 2, 32, 56)
278
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(112, 112, 64, 256, 2, 32, 56)
279
+
280
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 2, 64, 8, 32, 64)
281
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 4, 128, 8, 64, 64)
282
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 8, 128, 8, 64, 64)
283
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 16, 256, 3, 128, 128)
284
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 32, 256, 3, 128, 64)
285
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(128, 128, 64, 256, 3, 64, 64)
286
+
287
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 2, 64, 8, 32, 64)
288
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 4, 128, 6, 32, 64)
289
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 8, 128, 6, 32, 64)
290
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 16, 256, 5, 32, 64)
291
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(192, 128, 32, 256, 3, 64, 64)
292
+
293
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 2, 64, 8, 32, 64)
294
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 4, 128, 6, 32, 256)
295
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 8, 128, 6, 32, 256)
296
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5, 32, 256)
297
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3, 64, 128)
298
+
299
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 32, 256, 2, 128, 64)
300
+
301
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 64, 64)
302
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 64, 64)
303
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
304
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 4, 64, 64)
305
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 32, 256, 2, 128, 64)
306
+
307
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 64, 64)
308
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 64, 64)
309
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4, 64, 64)
310
+ GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2, 128, 64)
311
+
312
+ return 0;
313
+ }
314
+
315
+ static __host__ uint32_t ggml_cuda_fattn_tile_get_config(const int DKQ, const int DV, const int ncols, const int cc) {
316
+ if (GGML_CUDA_CC_IS_AMD(cc)) {
317
+ if (GGML_CUDA_CC_IS_RDNA(cc)) {
318
+ return ggml_cuda_fattn_tile_get_config_amd_rdna(DKQ, DV, ncols);
319
+ }
320
+ return ggml_cuda_fattn_tile_get_config_amd(DKQ, DV, ncols);
321
+ }
322
+ if (fast_fp16_available(cc)) {
323
+ return ggml_cuda_fattn_tile_get_config_nvidia_fp16(DKQ, DV, ncols);
324
+ }
325
+ return ggml_cuda_fattn_tile_get_config_nvidia_fp32(DKQ, DV, ncols);
326
+ }
327
+
328
+ static constexpr __device__ uint32_t ggml_cuda_fattn_tile_get_config(const int DKQ, const int DV, const int ncols) {
329
+ #ifdef GGML_USE_HIP
330
+ #ifdef RDNA
331
+ return ggml_cuda_fattn_tile_get_config_amd_rdna(DKQ, DV, ncols);
332
+ #else
333
+ return ggml_cuda_fattn_tile_get_config_amd(DKQ, DV, ncols);
334
+ #endif // RDNA
335
+ #else
336
+ #ifdef FAST_FP16_AVAILABLE
337
+ return ggml_cuda_fattn_tile_get_config_nvidia_fp16(DKQ, DV, ncols);
338
+ #else
339
+ return ggml_cuda_fattn_tile_get_config_nvidia_fp32(DKQ, DV, ncols);
340
+ #endif // FAST_FP16_AVAILABLE
341
+ #endif // GGML_USE_HIP
342
+ }
343
+
344
+ static __host__ int ggml_cuda_fattn_tile_get_nthreads(const int DKQ, const int DV, const int ncols, const int cc) {
345
+ return (ggml_cuda_fattn_tile_get_config(DKQ, DV, ncols, cc) >> 0) & ((1 << 10) - 1);
346
+ }
347
+
348
+ static constexpr __device__ int ggml_cuda_fattn_tile_get_nthreads(const int DKQ, const int DV, const int ncols) {
349
+ return (ggml_cuda_fattn_tile_get_config(DKQ, DV, ncols) >> 0) & ((1 << 10) - 1);
350
+ }
351
+
352
+ static __host__ int ggml_cuda_fattn_tile_get_occupancy(const int DKQ, const int DV, const int ncols, const int cc) {
353
+ return (ggml_cuda_fattn_tile_get_config(DKQ, DV, ncols, cc) >> 10) & ((1 << 4) - 1);
354
+ }
355
+
356
+ static constexpr __device__ int ggml_cuda_fattn_tile_get_occupancy(const int DKQ, const int DV, const int ncols) {
357
+ return (ggml_cuda_fattn_tile_get_config(DKQ, DV, ncols) >> 10) & ((1 << 4) - 1);
358
+ }
359
+
360
+ static __host__ int ggml_cuda_fattn_tile_get_nbatch_fa(const int DKQ, const int DV, const int ncols, const int cc) {
361
+ return (ggml_cuda_fattn_tile_get_config(DKQ, DV, ncols, cc) >> 14) & ((1 << 9) - 1);
362
+ }
363
+
364
+ static constexpr __device__ int ggml_cuda_fattn_tile_get_nbatch_fa(const int DKQ, const int DV, const int ncols) {
365
+ return (ggml_cuda_fattn_tile_get_config(DKQ, DV, ncols) >> 14) & ((1 << 9) - 1);
366
+ }
367
+
368
+ static __host__ int ggml_cuda_fattn_tile_get_nbatch_K(const int DKQ, const int DV, const int ncols, const int cc) {
369
+ return (ggml_cuda_fattn_tile_get_config(DKQ, DV, ncols, cc) >> 23) & ((1 << 9) - 1);
370
+ }
371
+
372
+ static constexpr __device__ int ggml_cuda_fattn_tile_get_nbatch_K(const int DKQ, const int DV, const int ncols) {
373
+ return (ggml_cuda_fattn_tile_get_config(DKQ, DV, ncols) >> 23) & ((1 << 9) - 1);
374
+ }
375
+
376
+ // TODO: deduplicate with mma-f16
377
+ template<int warp_size, int nwarps, int I, int J, int J_padding, bool oob_check>
378
+ static __device__ __forceinline__ void flash_attn_tile_load_tile(
379
+ const half2 * const __restrict__ KV, half2 * const __restrict__ tile_KV, const int stride_KV, const int i_sup) {
380
+ constexpr int cpy_nb = ggml_cuda_get_max_cpy_bytes();
381
+ constexpr int cpy_ne = cpy_nb / 4;
382
+
383
+ auto load = [&] __device__ (const int n) {
384
+ const int stride_j = warp_size >> n;
385
+
386
+ if (stride_j == 0) {
387
+ return;
388
+ }
389
+
390
+ const int j0_start = stride_j == warp_size ? 0 : ((J/2)/cpy_ne) - ((J/2)/cpy_ne) % (2*stride_j);
391
+ const int j0_stop = ((J/2)/cpy_ne) - ((J/2)/cpy_ne) % (1*stride_j);
392
+ const int stride_i = warp_size / stride_j;
393
+
394
+ if (j0_start == j0_stop) {
395
+ return;
396
+ }
397
+
398
+ #pragma unroll
399
+ for (int i0 = 0; i0 < I; i0 += nwarps*stride_i) {
400
+ const int i = i0 + threadIdx.y*stride_i + (stride_j == warp_size ? 0 : threadIdx.x / stride_j);
401
+
402
+ if (i0 + nwarps*stride_i <= I || i < I) {
403
+ #pragma unroll
404
+ for (int j0 = j0_start; j0 < j0_stop; j0 += stride_j) {
405
+ const int j = j0*cpy_ne + (stride_j == warp_size ? threadIdx.x : threadIdx.x % stride_j)*cpy_ne;
406
+
407
+ const __align__(16) half2 zero[cpy_ne] = {{0.0f, 0.0f}};
408
+ ggml_cuda_memcpy_1<cpy_nb>(
409
+ tile_KV + i*(J/2 + J_padding) + j,
410
+ !oob_check || i < i_sup ? KV + i*stride_KV + j : zero);
411
+ }
412
+ }
413
+ }
414
+ };
415
+ // 1: max 64*16=512 bytes, 512 half
416
+ // 2: max 32*16=512 bytes, 256 half
417
+ // 3: max 16*16=256 bytes, 128 half
418
+ // 4: max 8*16=128 bytes, 64 half
419
+ // 5: max 4*16= 64 bytes, 32 half
420
+ // 6: max 2*16= 32 bytes, 16 half
421
+ // 7: max 1*16= 16 bytes, 8 half
422
+ static_assert(J % 8 == 0, "bad J");
423
+ static_assert((J/2) % cpy_ne == 0, "bad J");
424
+ ggml_cuda_unroll<7>{}(load);
425
+ }
426
+
427
+ template<int warp_size, int nwarps, int I, int J, int J_padding, bool oob_check>
428
+ static __device__ __forceinline__ void flash_attn_tile_load_tile(
429
+ const half2 * const __restrict__ KV, float * const __restrict__ tile_KV, const int stride_KV, const int i_sup) {
430
+ constexpr int cpy_nb = ggml_cuda_get_max_cpy_bytes();
431
+ constexpr int cpy_ne = cpy_nb / 4;
432
+
433
+ auto load = [&] __device__ (const int n) {
434
+ const int stride_j = warp_size >> n;
435
+
436
+ if (stride_j == 0) {
437
+ return;
438
+ }
439
+
440
+ const int j0_start = stride_j == warp_size ? 0 : (J/cpy_ne) - (J/cpy_ne) % (2*stride_j);
441
+ const int j0_stop = (J/cpy_ne) - (J/cpy_ne) % (1*stride_j);
442
+ const int stride_i = warp_size / stride_j;
443
+
444
+ if (j0_start == j0_stop) {
445
+ return;
446
+ }
447
+
448
+ #pragma unroll
449
+ for (int i0 = 0; i0 < I; i0 += nwarps*stride_i) {
450
+ const int i = i0 + threadIdx.y*stride_i + (stride_j == warp_size ? 0 : threadIdx.x / stride_j);
451
+
452
+ if (i0 + nwarps*stride_i <= I || i < I) {
453
+ #pragma unroll
454
+ for (int j0 = j0_start; j0 < j0_stop; j0 += stride_j) {
455
+ const int j = j0*(cpy_ne/2) + (stride_j == warp_size ? threadIdx.x : threadIdx.x % stride_j)*(cpy_ne/2);
456
+
457
+ const half2 zero[cpy_ne/2] = {{0.0f, 0.0f}};
458
+ __align__(16) half2 tmp_h2[cpy_ne/2];
459
+ ggml_cuda_memcpy_1<sizeof(tmp_h2)>(
460
+ tmp_h2, !oob_check || i < i_sup ? KV + i*stride_KV + j : zero);
461
+
462
+ __align__(16) float2 tmp_f2[cpy_ne/2];
463
+ #pragma unroll
464
+ for (int l = 0; l < cpy_ne/2; ++l) {
465
+ tmp_f2[l] = __half22float2(tmp_h2[l]);
466
+ }
467
+ ggml_cuda_memcpy_1<sizeof(tmp_f2)>(tile_KV + i*(J + J_padding) + 2*j, tmp_f2);
468
+ }
469
+ }
470
+ }
471
+ };
472
+ // 1: max 32*16=512 bytes, 128 float
473
+ // 2: max 16*16=256 bytes, 64 float
474
+ // 3: max 8*16=128 bytes, 32 float
475
+ // 4: max 4*16= 64 bytes, 16 float
476
+ // 5: max 2*16= 32 bytes, 8 float
477
+ static_assert(J % 8 == 0, "bad J");
478
+ static_assert(J % cpy_ne == 0, "bad J");
479
+ ggml_cuda_unroll<5>{}(load);
480
+ }
481
+
482
+ // Function that performs a single iteration in for the KQ matrix multiplication:
483
+ template <int warp_size, int nwarps, int ncols1, int ncols2, int DKQ, int nbatch_fa, int nbatch_K,
484
+ bool use_logit_softcap, bool oob_check, typename T_vec_dot>
485
+ static __device__ __forceinline__ void flash_attn_tile_iter_KQ(
486
+ T_vec_dot * const Q_tmp,
487
+ const half2 * const __restrict__ K_h2,
488
+ T_vec_dot * const KV_tmp,
489
+ const int stride_K2,
490
+ const int k_VKQ_0,
491
+ const int k_VKQ_sup,
492
+ const int k_KQ_0,
493
+ float * KQ_acc) {
494
+ constexpr int cpy_nb = ggml_cuda_get_max_cpy_bytes();
495
+ constexpr int cpy_ne = cpy_nb / 4;
496
+
497
+ constexpr int ncols = ncols1*ncols2;
498
+ constexpr int cpw = ncols > nwarps ? ncols/nwarps : 1; // Q columns per warp
499
+ constexpr int np = nwarps > ncols ? nwarps/ncols : 1; // number of parallel warps per Q column
500
+
501
+ flash_attn_tile_load_tile<warp_size, nwarps, nbatch_fa, nbatch_K, cpy_ne, oob_check>
502
+ (K_h2 + int64_t(k_VKQ_0)*stride_K2 + k_KQ_0/2, KV_tmp, stride_K2, k_VKQ_sup);
503
+ __syncthreads();
504
+
505
+ #ifdef FAST_FP16_AVAILABLE
506
+ static_assert((nbatch_K/2) % cpy_ne == 0, "bad nbatch_K");
507
+ #pragma unroll
508
+ for (int k_KQ_1 = 0; k_KQ_1 < nbatch_K/2; k_KQ_1 += cpy_ne) {
509
+ __align__(16) half2 K_k[nbatch_fa/(np*warp_size)][cpy_ne];
510
+ __align__(16) half2 Q_k[cpw][cpy_ne];
511
+ #else
512
+ static_assert(nbatch_K % cpy_ne == 0, "bad nbatch_K");
513
+ #pragma unroll
514
+ for (int k_KQ_1 = 0; k_KQ_1 < nbatch_K; k_KQ_1 += cpy_ne) {
515
+ __align__(16) float K_k[nbatch_fa/(np*warp_size)][cpy_ne];
516
+ __align__(16) float Q_k[cpw][cpy_ne];
517
+ #endif // FAST_FP16_AVAILABLE
518
+
519
+ #pragma unroll
520
+ for (int i_KQ_0 = 0; i_KQ_0 < nbatch_fa; i_KQ_0 += np*warp_size) {
521
+ const int i_KQ = i_KQ_0 + (threadIdx.y % np)*warp_size + threadIdx.x;
522
+
523
+ #ifdef FAST_FP16_AVAILABLE
524
+ ggml_cuda_memcpy_1<cpy_nb>(&K_k[i_KQ_0/(np*warp_size)], &KV_tmp[i_KQ*(nbatch_K/2 + cpy_ne) + k_KQ_1]);
525
+ #else
526
+ ggml_cuda_memcpy_1<cpy_nb>(&K_k[i_KQ_0/(np*warp_size)], &KV_tmp[i_KQ*(nbatch_K + cpy_ne) + k_KQ_1]);
527
+ #endif // FAST_FP16_AVAILABLE
528
+ }
529
+ #pragma unroll
530
+ for (int jc0 = 0; jc0 < cpw; ++jc0) {
531
+ const int jc = jc0 + (threadIdx.y / np)*cpw;
532
+
533
+ #ifdef FAST_FP16_AVAILABLE
534
+ ggml_cuda_memcpy_1<cpy_nb>(&Q_k[jc0], &Q_tmp[jc*(DKQ/2) + k_KQ_0/2 + k_KQ_1]);
535
+ #else
536
+ ggml_cuda_memcpy_1<cpy_nb>(&Q_k[jc0], &Q_tmp[jc* DKQ + k_KQ_0 + k_KQ_1]);
537
+ #endif // FAST_FP16_AVAILABLE
538
+ }
539
+
540
+ #pragma unroll
541
+ for (int i_KQ_0 = 0; i_KQ_0 < nbatch_fa; i_KQ_0 += np*warp_size) {
542
+ #pragma unroll
543
+ for (int jc0 = 0; jc0 < cpw; ++jc0) {
544
+ #pragma unroll
545
+ for (int k = 0; k < cpy_ne; ++k) {
546
+ ggml_cuda_mad(KQ_acc[i_KQ_0/(np*warp_size)*cpw + jc0], K_k[i_KQ_0/(np*warp_size)][k], Q_k[jc0][k]);
547
+ }
548
+ }
549
+ }
550
+ }
551
+
552
+ if (k_KQ_0 + nbatch_K < DKQ) {
553
+ __syncthreads(); // Sync not needed on last iteration.
554
+ }
555
+ }
556
+
557
+ // Function that performs a single iteration of the main loop over up to nbatch_fa tokens.
558
+ template <int warp_size, int nwarps, int ncols1, int ncols2, int DKQ, int DV, int nbatch_fa, int nbatch_K,
559
+ bool use_logit_softcap, bool oob_check, typename T_vec_dot, typename T_KQ, typename T_acc>
560
+ static __device__ __forceinline__ void flash_attn_tile_iter(
561
+ T_vec_dot * const Q_tmp,
562
+ const half2 * const __restrict__ K_h2,
563
+ const half2 * const __restrict__ V_h2,
564
+ const half * const __restrict__ mask,
565
+ const uint3 ne01,
566
+ const float logit_softcap,
567
+ const float slope,
568
+ T_KQ * const KQ,
569
+ T_vec_dot * const KV_tmp,
570
+ const int stride_K2,
571
+ const int stride_V2,
572
+ const int stride_mask,
573
+ float * const KQ_max,
574
+ float * const KQ_sum,
575
+ T_acc * const VKQ,
576
+ const int k_VKQ_0,
577
+ const int k_VKQ_max,
578
+ const int col_Q_0) {
579
+ constexpr int cpy_nb = ggml_cuda_get_max_cpy_bytes();
580
+ constexpr int cpy_ne = cpy_nb / 4;
581
+
582
+ constexpr int ncols = ncols1*ncols2;
583
+ constexpr int cpw = ncols > nwarps ? ncols/nwarps : 1; // Q columns per warp
584
+ constexpr int np = nwarps > ncols ? nwarps/ncols : 1; // number of parallel warps per Q column
585
+
586
+ constexpr int DVp = (DV + 2*warp_size - 1) & ~(2*warp_size - 1); // DV padded to multiple of 2*warp_size.
587
+
588
+ // KQ_cs == KQ chunk size, number of KQ values in j direction to store as one contiguous chunk in memory.
589
+ // KQ is originally 2D but uses a Z-shaped 3D memory pattern like KQ[ncols/KQ_cs][DVp][KQ_cs].
590
+ #ifdef FAST_FP16_AVAILABLE
591
+ constexpr int KQ_cs = cpw < 2*cpy_ne ? cpw : 2*cpy_ne;
592
+ #else
593
+ constexpr int KQ_cs = cpw < 1*cpy_ne ? cpw : 1*cpy_ne;
594
+ #endif // FAST_FP16_AVAILABLE
595
+ static_assert(cpw % KQ_cs == 0, "bad KQ_cs");
596
+ const int k_VKQ_sup = k_VKQ_max - k_VKQ_0; // k supremum, only smaller k values have valid KV data
597
+
598
+ float KQ_max_new[cpw];
599
+ #pragma unroll
600
+ for (int jc0 = 0; jc0 < cpw; ++jc0) {
601
+ KQ_max_new[jc0] = KQ_max[jc0];
602
+ }
603
+
604
+ float KQ_acc[nbatch_fa/(np*warp_size) * cpw] = {0.0f}; // Accumulators for KQ matrix multiplication.
605
+
606
+ // KQ = K @ Q matrix multiplication:
607
+ constexpr int nbatch_K_last = DKQ % nbatch_K;
608
+ #pragma unroll
609
+ for (int k_KQ_0 = 0; k_KQ_0 < DKQ - nbatch_K_last; k_KQ_0 += nbatch_K) {
610
+ flash_attn_tile_iter_KQ<warp_size, nwarps, ncols1, ncols2, DKQ, nbatch_fa, nbatch_K, use_logit_softcap, oob_check>(
611
+ Q_tmp, K_h2, KV_tmp, stride_K2, k_VKQ_0, k_VKQ_sup, k_KQ_0, KQ_acc);
612
+ }
613
+ if (nbatch_K_last > 0) {
614
+ constexpr int k_KQ_0 = DKQ - nbatch_K_last;
615
+ flash_attn_tile_iter_KQ<warp_size, nwarps, ncols1, ncols2, DKQ, nbatch_fa, nbatch_K_last, use_logit_softcap, oob_check>(
616
+ Q_tmp, K_h2, KV_tmp, stride_K2, k_VKQ_0, k_VKQ_sup, k_KQ_0, KQ_acc);
617
+ }
618
+
619
+ // Apply logit softcap + mask, update KQ_max:
620
+ #pragma unroll
621
+ for (int jc0 = 0; jc0 < cpw; ++jc0) {
622
+ const int j = fastmodulo(col_Q_0 + (jc0 + (threadIdx.y / np)*cpw)/ncols2, ne01);
623
+
624
+ #pragma unroll
625
+ for (int i_KQ_0 = 0; i_KQ_0 < nbatch_fa; i_KQ_0 += np*warp_size) {
626
+ const int i_KQ = i_KQ_0 + (threadIdx.y % np)*warp_size + threadIdx.x;
627
+
628
+ #if defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
629
+ // Without the v_dot2_f32_f16 instruction there is a higher risk of numerical overflow in the KQ calculation.
630
+ // Therefore, scale down Q values and apply the inverse scale the FP32 KQ values afterwards again.
631
+ KQ_acc[i_KQ_0/(np*warp_size)*cpw + jc0] *= 4.0f;
632
+ #endif // defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
633
+
634
+ if (use_logit_softcap) {
635
+ KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] = logit_softcap * tanhf(KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0]);
636
+ }
637
+
638
+ if (!oob_check || i_KQ < k_VKQ_sup) {
639
+ KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] += (ncols2 > 1 || mask) ?
640
+ slope*__half2float(mask[j*stride_mask + k_VKQ_0 + i_KQ]) : 0.0f;
641
+
642
+ KQ_max_new[jc0] = fmaxf(KQ_max_new[jc0], KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] + FATTN_KQ_MAX_OFFSET);
643
+ }
644
+ }
645
+
646
+ KQ_max_new[jc0] = warp_reduce_max<warp_size>(KQ_max_new[jc0]);
647
+ }
648
+
649
+ if constexpr (np == 1) {
650
+ __syncthreads();
651
+ } else {
652
+ static_assert(cpw == 1, "bad cpw");
653
+ __shared__ float KQ_max_new_shared[nwarps];
654
+ if (threadIdx.x == 0) {
655
+ KQ_max_new_shared[threadIdx.y] = KQ_max_new[0];
656
+ }
657
+ __syncthreads();
658
+ KQ_max_new[0] = KQ_max_new_shared[(threadIdx.y & ~(np-1)) + threadIdx.x % np];
659
+ KQ_max_new[0] = warp_reduce_max<np>(KQ_max_new[0]);
660
+ }
661
+
662
+ // Calculate KQ softmax, write to shared KQ buffer, re-scale VKQ accumulators:
663
+ #pragma unroll
664
+ for (int jc0 = 0; jc0 < cpw; jc0 += KQ_cs) {
665
+ #ifdef FAST_FP16_AVAILABLE
666
+ __align__(16) half tmp[nbatch_fa/(np*warp_size)][KQ_cs];
667
+ #else
668
+ __align__(16) float tmp[nbatch_fa/(np*warp_size)][KQ_cs];
669
+ #endif // FAST_FP16_AVAILABLE
670
+
671
+ #pragma unroll
672
+ for (int jc1 = 0; jc1 < KQ_cs; ++jc1) {
673
+ const int jc = jc0 + jc1;
674
+
675
+ const float KQ_max_scale = expf(KQ_max[jc] - KQ_max_new[jc]);
676
+ KQ_max[jc] = KQ_max_new[jc];
677
+
678
+ float KQ_sum_add = 0.0f;
679
+ #pragma unroll
680
+ for (int i0 = 0; i0 < nbatch_fa; i0 += np*warp_size) {
681
+ const float val = !oob_check || i0 + (threadIdx.y % np)*warp_size + threadIdx.x < static_cast<uint32_t>(k_VKQ_sup) ?
682
+ expf(KQ_acc[(i0/(np*warp_size))*cpw + jc] - KQ_max[jc]) : 0.0f;
683
+ KQ_sum_add += val;
684
+ tmp[i0/(np*warp_size)][jc1] = val;
685
+ }
686
+ KQ_sum[jc] = KQ_sum[jc]*KQ_max_scale + KQ_sum_add;
687
+
688
+ #ifdef FAST_FP16_AVAILABLE
689
+ const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
690
+ #pragma unroll
691
+ for (int i0 = 0; i0 < DVp/2; i0 += warp_size) {
692
+ VKQ[jc*((DVp/2)/warp_size) + i0/warp_size] *= KQ_max_scale_h2;
693
+ }
694
+ #else
695
+ #pragma unroll
696
+ for (int i0 = 0; i0 < DVp/2; i0 += warp_size) {
697
+ VKQ[jc*((DVp/2)/warp_size) + i0/warp_size].x *= KQ_max_scale;
698
+ VKQ[jc*((DVp/2)/warp_size) + i0/warp_size].y *= KQ_max_scale;
699
+ }
700
+ #endif // FAST_FP16_AVAILABLE
701
+ }
702
+
703
+ #pragma unroll
704
+ for (int i0 = 0; i0 < nbatch_fa; i0 += np*warp_size) {
705
+ const int i = i0 + (threadIdx.y % np)*warp_size + threadIdx.x;
706
+
707
+ ggml_cuda_memcpy_1<sizeof(tmp[0])>(
708
+ KQ + (jc0/KQ_cs + (threadIdx.y / np)*(cpw/KQ_cs))*(nbatch_fa*KQ_cs) + i*KQ_cs,
709
+ tmp[i0/(np*warp_size)]);
710
+ }
711
+ }
712
+
713
+ // VKQ = V @ KQ matrix multiplication:
714
+ static_assert(DV <= DKQ, "bad DV");
715
+ static_assert(DV % nbatch_K == 0 || (nbatch_K % 3 == 0 && DV % (nbatch_K*2/3) == 0), "bad nbatch_K");
716
+ constexpr int nbatch_V = (DV % nbatch_K == 0 ? nbatch_K : nbatch_K*2/3) * nbatch_fa / DV; // Number of V columns that fit in SRAM for K.
717
+ static_assert(nbatch_fa % nbatch_V == 0, "bad nbatch_V");
718
+ static_assert(nbatch_V % np == 0, "bad nbatch_V");
719
+ #pragma unroll
720
+ for (int k0 = 0; k0 < nbatch_fa; k0 += nbatch_V) {
721
+ flash_attn_tile_load_tile<warp_size, nwarps, nbatch_V, DV, 0, oob_check>
722
+ (V_h2 + int64_t(k_VKQ_0 + k0)*stride_V2, KV_tmp, stride_V2, k_VKQ_sup - k0);
723
+ __syncthreads();
724
+
725
+ #ifdef FAST_FP16_AVAILABLE
726
+ #pragma unroll
727
+ for (int k1 = 0; k1 < nbatch_V; k1 += np) {
728
+ __align__(16) half2 V_k[(DVp/2)/warp_size];
729
+ __align__(16) half2 KQ_k[cpw];
730
+
731
+ constexpr int cpy_ne_D = cpy_ne/2 < (DVp/2)/warp_size ? cpy_ne/2 : (DVp/2)/warp_size;
732
+ #pragma unroll
733
+ for (int i0 = 0; i0 < DVp/2; i0 += warp_size*cpy_ne_D) {
734
+ ggml_cuda_memcpy_1<cpy_ne_D*4>(&V_k[i0/warp_size], &KV_tmp[(k1 + threadIdx.y % np)*(DV/2) + i0 + threadIdx.x*cpy_ne_D]);
735
+ }
736
+ #pragma unroll
737
+ for (int jc_VKQ_0 = 0; jc_VKQ_0 < cpw; jc_VKQ_0 += KQ_cs) {
738
+ const int jc_KQ = jc_VKQ_0/KQ_cs + (threadIdx.y / np)*(cpw/KQ_cs);
739
+
740
+ __align__(16) half tmp[KQ_cs];
741
+ ggml_cuda_memcpy_1<KQ_cs*sizeof(half)>(
742
+ &tmp, KQ + jc_KQ*(nbatch_fa*KQ_cs) + (k0 + k1 + threadIdx.y % np)*KQ_cs);
743
+ #pragma unroll
744
+ for (int jc_VKQ_1 = 0; jc_VKQ_1 < KQ_cs; ++jc_VKQ_1) {
745
+ KQ_k[jc_VKQ_0+jc_VKQ_1] = __half2half2(tmp[jc_VKQ_1]);
746
+ }
747
+ }
748
+
749
+ #pragma unroll
750
+ for (int i0 = 0; i0 < DVp/2; i0 += warp_size) {
751
+ #pragma unroll
752
+ for (int jc_VKQ_0 = 0; jc_VKQ_0 < cpw; ++jc_VKQ_0) {
753
+ VKQ[jc_VKQ_0*((DVp/2)/warp_size) + i0/warp_size] += V_k[i0/warp_size]*KQ_k[jc_VKQ_0];
754
+ }
755
+ }
756
+ }
757
+ #else
758
+ #pragma unroll
759
+ for (int k1 = 0; k1 < nbatch_V; k1 += np) {
760
+ __align__(16) float2 V_k[(DVp/2)/warp_size];
761
+ __align__(16) float KQ_k[cpw];
762
+
763
+ constexpr int cpy_ne_D = cpy_ne < DVp/warp_size ? cpy_ne : DVp/warp_size;
764
+ #pragma unroll
765
+ for (int i0 = 0; i0 < DVp; i0 += warp_size*cpy_ne_D) {
766
+ ggml_cuda_memcpy_1<cpy_ne_D*4>(&V_k[i0/(2*warp_size)], &KV_tmp[(k1 + threadIdx.y % np)*DV + i0 + threadIdx.x*cpy_ne_D]);
767
+ }
768
+ #pragma unroll
769
+ for (int jc_VKQ_0 = 0; jc_VKQ_0 < cpw; jc_VKQ_0 += KQ_cs) {
770
+ const int jc_KQ = jc_VKQ_0/KQ_cs + (threadIdx.y / np)*(cpw/KQ_cs);
771
+
772
+ ggml_cuda_memcpy_1<KQ_cs*sizeof(float)>(
773
+ &KQ_k[jc_VKQ_0], KQ + jc_KQ*(nbatch_fa*KQ_cs) + (k0 + k1 + threadIdx.y % np)*KQ_cs);
774
+ }
775
+
776
+ #pragma unroll
777
+ for (int i0 = 0; i0 < DVp/2; i0 += warp_size) {
778
+ #pragma unroll
779
+ for (int jc_VKQ_0 = 0; jc_VKQ_0 < cpw; ++jc_VKQ_0) {
780
+ VKQ[jc_VKQ_0*((DVp/2)/warp_size) + i0/warp_size].x += V_k[i0/warp_size].x*KQ_k[jc_VKQ_0];
781
+ VKQ[jc_VKQ_0*((DVp/2)/warp_size) + i0/warp_size].y += V_k[i0/warp_size].y*KQ_k[jc_VKQ_0];
782
+ }
783
+ }
784
+ }
785
+ #endif // FAST_FP16_AVAILABLE
786
+
787
+ __syncthreads();
788
+ }
789
+ }
790
+
791
+ template<int DKQ, int DV, int ncols1, int ncols2, bool use_logit_softcap> // D == head size
792
+ __launch_bounds__(ggml_cuda_fattn_tile_get_nthreads(DKQ, DV, ncols1*ncols2), ggml_cuda_fattn_tile_get_occupancy(DKQ, DV, ncols1*ncols2))
793
+ static __global__ void flash_attn_tile(
794
+ const char * Q_ptr,
795
+ const char * K_ptr,
796
+ const char * V_ptr,
797
+ const char * mask_ptr,
798
+ const char * sinks_ptr,
799
+ const int * KV_max_ptr,
800
+ float * dst_ptr,
801
+ float2 * dst_meta_ptr,
802
+ const float scale,
803
+ const float max_bias,
804
+ const float m0,
805
+ const float m1,
806
+ const uint32_t n_head_log2,
807
+ const float logit_softcap,
808
+ const int32_t ne00, const uint3 ne01, const int32_t ne02, const int32_t ne03,
809
+ const int32_t nb01, const int32_t nb02, const int32_t nb03,
810
+ const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
811
+ const int32_t nb11, const int32_t nb12, const int64_t nb13,
812
+ const int32_t nb21, const int32_t nb22, const int64_t nb23,
813
+ const int32_t ne31, const int32_t ne32, const int32_t ne33,
814
+ const int32_t nb31, const int32_t nb32, const int64_t nb33) {
815
+ #ifdef FLASH_ATTN_AVAILABLE
816
+ const char * GGML_CUDA_RESTRICT Q = Q_ptr;
817
+ const char * GGML_CUDA_RESTRICT K = K_ptr;
818
+ const char * GGML_CUDA_RESTRICT V = V_ptr;
819
+ const char * GGML_CUDA_RESTRICT mask = mask_ptr;
820
+ const char * GGML_CUDA_RESTRICT sinks = sinks_ptr;
821
+ const int * GGML_CUDA_RESTRICT KV_max = KV_max_ptr;
822
+ float * GGML_CUDA_RESTRICT dst = dst_ptr;
823
+ float2 * GGML_CUDA_RESTRICT dst_meta = dst_meta_ptr;
824
+
825
+ // Skip unused kernel variants for faster compilation:
826
+
827
+ if ((use_logit_softcap && !(DV == 128 || DV == 256 || DV == 512))) {
828
+ GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
829
+ max_bias, m0, m1, n_head_log2, logit_softcap,
830
+ ne00, ne01, ne02, ne03,
831
+ nb01, nb02, nb03,
832
+ ne10, ne11, ne12, ne13,
833
+ nb11, nb12, nb13,
834
+ nb21, nb22, nb23,
835
+ ne31, ne32, ne33,
836
+ nb31, nb32, nb33);
837
+ NO_DEVICE_CODE;
838
+ return;
839
+ }
840
+
841
+ static_assert(ggml_cuda_fattn_tile_get_config(DKQ, DV, ncols1*ncols2) != 0, "kernel config not defined");
842
+
843
+ constexpr int ncols = ncols1*ncols2;
844
+ constexpr int warp_size = 32;
845
+ constexpr int nwarps = ggml_cuda_fattn_tile_get_nthreads (DKQ, DV, ncols1*ncols2) / warp_size;
846
+ constexpr int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, ncols1*ncols2);
847
+ constexpr int nbatch_K = ggml_cuda_fattn_tile_get_nbatch_K (DKQ, DV, ncols1*ncols2);
848
+
849
+ // In this kernel Q, K, V are matrices while i, j, k are matrix indices.
850
+
851
+ const int col_Q_0 = blockIdx.x * ncols1; // Index of the first Q column for this CUDA block to work on.
852
+
853
+ const int sequence = blockIdx.z / (ne02/ncols2);
854
+ const int head0 = blockIdx.z*ncols2 - sequence*ne02; // == blockIdx.z % (ne02/ncols2)
855
+ const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
856
+ const float * Q_f = (const float *) (Q + nb03*sequence + nb02* head0);
857
+ const half2 * K_h2 = (const half2 *) (K + nb13*sequence + nb12*(head0 / gqa_ratio));
858
+ const half2 * V_h2 = (const half2 *) (V + nb23*sequence + nb22*(head0 / gqa_ratio)); // K and V have same shape
859
+
860
+ const half * maskh = mask ? (const half *) (mask + nb33*(sequence % ne33)) : nullptr;
861
+
862
+ const int stride_K2 = nb11 / sizeof(half2);
863
+ const int stride_V2 = nb21 / sizeof(half2);
864
+ const int stride_mask = nb31 / sizeof(half);
865
+
866
+ const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head0, n_head_log2, m0, m1) : 1.0f;
867
+
868
+ constexpr int cpy_nb = ggml_cuda_get_max_cpy_bytes();
869
+ constexpr int cpy_ne = cpy_nb / 4;
870
+
871
+ constexpr int cpw = ncols > nwarps ? ncols/nwarps : 1; // Q columns per warp.
872
+ constexpr int np = nwarps > ncols ? nwarps/ncols : 1; // Number of parallel warps per Q column.
873
+ static_assert(cpw == 1 || np == 1, "bad cpw / np");
874
+ static_assert(nbatch_fa % (np*warp_size) == 0, "nbatch_fa % (np*warp_size) != 0");
875
+
876
+ constexpr int DKQp = (DKQ + 2*warp_size - 1) & ~(2*warp_size - 1); // DKQ padded to multiple of 2*warp_size.
877
+ constexpr int DVp = (DV + 2*warp_size - 1) & ~(2*warp_size - 1); // DV padded to multiple of 2*warp_size.
878
+
879
+ // Q_tmp == SRAM buffer to hold Q data for the entire lifetime of the kernel.
880
+ // KV_tmp == SRAM buffer to hold fragments of K/V data while iterating over ne11.
881
+ // KV_tmp is padded to avoid memory conflicts for K (cpy_ne) and OOB accesses for V (DVp-DV).
882
+ // KQ == SRAM buffer to hold KQ fragments between KQ and VKQ matrix multiplications.
883
+ // VKQ == Accumulators in registers for the final VKQ result.
884
+ #ifdef FAST_FP16_AVAILABLE
885
+ __shared__ half2 Q_tmp[ncols * DKQ/2];
886
+ __shared__ half2 KV_tmp[nbatch_fa * (nbatch_K/2 + cpy_ne) + DVp-DV];
887
+ __shared__ half KQ[ncols * nbatch_fa];
888
+ __align__(16) half2 VKQ[cpw * ((DVp/2)/warp_size)] = {{0.0f, 0.0f}};
889
+ #else
890
+ __shared__ float Q_tmp[ncols * DKQ];
891
+ __shared__ float KV_tmp[nbatch_fa * (nbatch_K + cpy_ne) + DVp-DV];
892
+ __shared__ float KQ[ncols * nbatch_fa];
893
+ __align__(16) float2 VKQ[cpw * ((DVp/2)/warp_size)] = {{0.0f, 0.0f}};
894
+ #endif // FAST_FP16_AVAILABLE
895
+
896
+ float KQ_max[cpw];
897
+ #pragma unroll
898
+ for (int j0 = 0; j0 < ncols; j0 += nwarps) {
899
+ KQ_max[j0/nwarps] = -FLT_MAX/2.0f;
900
+ }
901
+ float KQ_sum[cpw] = {0.0f};
902
+
903
+ ggml_cuda_pdl_sync();
904
+
905
+ // Load Q data, convert to FP16 if fast:
906
+ #pragma unroll
907
+ for (int jc0 = 0; jc0 < cpw; ++jc0) {
908
+ const int jc = jc0 + (threadIdx.y / np)*cpw;
909
+
910
+ const int j = jc / ncols2;
911
+ const int c = jc % ncols2;
912
+
913
+ constexpr int cpy_ne_D = cpy_ne < DKQp/warp_size ? cpy_ne : DKQp/warp_size;
914
+
915
+ #pragma unroll
916
+ for (int i0 = 0; i0 < DKQp; i0 += np*warp_size*cpy_ne_D) {
917
+ if (i0 + np*warp_size*cpy_ne_D <= DKQ || i0 + (threadIdx.y % np)*(warp_size*cpy_ne_D) + threadIdx.x*cpy_ne_D < DKQ) {
918
+ __align__(16) float tmp_f[cpy_ne_D] = {0.0f};
919
+ ggml_cuda_memcpy_1<sizeof(tmp_f)>
920
+ (tmp_f, &Q_f[c*(nb02/sizeof(float)) + fastmodulo(col_Q_0 + j, ne01)*(nb01/sizeof(float))
921
+ + i0 + (threadIdx.y % np)*(warp_size*cpy_ne_D) + threadIdx.x*cpy_ne_D]);
922
+
923
+ #pragma unroll
924
+ for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
925
+ tmp_f[i1] *= scale;
926
+ }
927
+
928
+ #ifdef FAST_FP16_AVAILABLE
929
+ __align__(16) half2 tmp_h2[cpy_ne_D/2];
930
+ #pragma unroll
931
+ for (int i1 = 0; i1 < cpy_ne_D; i1 += 2) {
932
+ tmp_h2[i1/2] = make_half2(tmp_f[i1 + 0], tmp_f[i1 + 1]);
933
+ #if defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
934
+ // Without the v_dot2_f32_f16 instruction there is a higher risk of numerical overflow in the KQ calculation.
935
+ // Therefore, scale down Q values and apply the inverse scale the FP32 KQ values afterwards again.
936
+ tmp_h2[i1/2] *= make_half2(0.25f, 0.25f);
937
+ #endif // defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
938
+ }
939
+ ggml_cuda_memcpy_1<sizeof(tmp_h2)>(
940
+ &Q_tmp[jc*(DKQ/2) + i0/2 + (threadIdx.y % np)*(warp_size*cpy_ne_D/2) + threadIdx.x*(cpy_ne_D/2)],
941
+ tmp_h2);
942
+ #else
943
+ ggml_cuda_memcpy_1<sizeof(tmp_f)>(
944
+ &Q_tmp[jc* DKQ + i0 + (threadIdx.y % np)*(warp_size*cpy_ne_D) + threadIdx.x* cpy_ne_D],
945
+ tmp_f);
946
+ #endif // FAST_FP16_AVAILABLE
947
+ }
948
+ }
949
+ }
950
+
951
+ __syncthreads();
952
+
953
+ // Main loop over KV cache:
954
+ const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
955
+ if (ncols2 == 1) {
956
+ // Branch with out-of-bounds checks.
957
+ int k_VKQ_0 = blockIdx.y*nbatch_fa;
958
+ while (k_VKQ_0 < k_VKQ_max - nbatch_fa) {
959
+ constexpr bool oob_check = false;
960
+ flash_attn_tile_iter<warp_size, nwarps, ncols1, ncols2, DKQ, DV, nbatch_fa, nbatch_K, use_logit_softcap, oob_check>
961
+ (Q_tmp, K_h2, V_h2, maskh, ne01, logit_softcap, slope, KQ, KV_tmp,
962
+ stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max, col_Q_0);
963
+ k_VKQ_0 += gridDim.y*nbatch_fa;
964
+ }
965
+ if (k_VKQ_0 < k_VKQ_max) {
966
+ constexpr bool oob_check = true;
967
+ flash_attn_tile_iter<warp_size, nwarps, ncols1, ncols2, DKQ, DV, nbatch_fa, nbatch_K, use_logit_softcap, oob_check>
968
+ (Q_tmp, K_h2, V_h2, maskh, ne01, logit_softcap, slope, KQ, KV_tmp,
969
+ stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max, col_Q_0);
970
+ }
971
+ } else {
972
+ // Branch without out-of-bounds checks.
973
+ for (int k_VKQ_0 = blockIdx.y*nbatch_fa; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*nbatch_fa) {
974
+ constexpr bool oob_check = false;
975
+ flash_attn_tile_iter<warp_size, nwarps, ncols1, ncols2, DKQ, DV, nbatch_fa, nbatch_K, use_logit_softcap, oob_check>
976
+ (Q_tmp, K_h2, V_h2, maskh, ne01, logit_softcap, slope, KQ, KV_tmp,
977
+ stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max, col_Q_0);
978
+ }
979
+ }
980
+
981
+ #pragma unroll
982
+ for (int jc0 = 0; jc0 < cpw; ++jc0) {
983
+ KQ_sum[jc0] = warp_reduce_sum<warp_size>(KQ_sum[jc0]);
984
+ }
985
+
986
+ if constexpr (np > 1) {
987
+ static_assert(cpw == 1, "bad cpw");
988
+ static_assert(nbatch_fa*nbatch_K >= nwarps*DVp, "KV_tmp too small");
989
+
990
+ #ifdef FAST_FP16_AVAILABLE
991
+ half2 * VKQ_combine = (half2 *) KV_tmp;
992
+ #else
993
+ float * VKQ_combine = (float *) KV_tmp;
994
+ #endif // FAST_FP16_AVAILABLE
995
+ float * KQ_sum_combine = (float *) Q_tmp;
996
+
997
+ if (threadIdx.y % np != 0) {
998
+ #ifdef FAST_FP16_AVAILABLE
999
+ constexpr int cpy_ne_D = cpy_ne < (DVp/2)/warp_size ? cpy_ne : (DVp/2)/warp_size;
1000
+ #pragma unroll
1001
+ for (int i0 = 0; i0 < DVp/2; i0 += warp_size*cpy_ne_D) {
1002
+ ggml_cuda_memcpy_1<cpy_ne_D*4>(&VKQ_combine[threadIdx.y*(DVp/2) + i0 + threadIdx.x*cpy_ne_D], &VKQ[i0/warp_size]);
1003
+ }
1004
+ #else
1005
+ constexpr int cpy_ne_D = cpy_ne < DVp/warp_size ? cpy_ne : DVp/warp_size;
1006
+ #pragma unroll
1007
+ for (int i0 = 0; i0 < DVp; i0 += warp_size*cpy_ne_D) {
1008
+ ggml_cuda_memcpy_1<cpy_ne_D*4>(
1009
+ &VKQ_combine[threadIdx.y*DVp + i0 + threadIdx.x*cpy_ne_D], ((const float *) VKQ) + i0/warp_size);
1010
+ }
1011
+ #endif // FAST_FP16_AVAILABLE
1012
+
1013
+ if (threadIdx.x == 0) {
1014
+ KQ_sum_combine[threadIdx.y] = KQ_sum[0];
1015
+ }
1016
+
1017
+ return;
1018
+ }
1019
+
1020
+ __syncthreads();
1021
+
1022
+ #pragma unroll
1023
+ for (int ip = 1; ip < np; ++ip) {
1024
+ #ifdef FAST_FP16_AVAILABLE
1025
+ constexpr int cpy_ne_D = cpy_ne < (DVp/2)/warp_size ? cpy_ne : (DVp/2)/warp_size;
1026
+ #pragma unroll
1027
+ for (int i0 = 0; i0 < DVp/2; i0 += warp_size*cpy_ne_D) {
1028
+ __align__(16) half2 tmp[cpy_ne_D];
1029
+ ggml_cuda_memcpy_1<cpy_ne_D*4>(tmp, &VKQ_combine[(threadIdx.y + ip)*(DVp/2) + i0 + threadIdx.x*cpy_ne_D]);
1030
+ #pragma unroll
1031
+ for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
1032
+ VKQ[i0/warp_size + i1] += tmp[i1];
1033
+ }
1034
+ }
1035
+ #else
1036
+ constexpr int cpy_ne_D = cpy_ne < DVp/warp_size ? cpy_ne : DVp/warp_size;
1037
+ #pragma unroll
1038
+ for (int i0 = 0; i0 < DVp; i0 += warp_size*cpy_ne_D) {
1039
+ __align__(16) float tmp[cpy_ne_D];
1040
+ ggml_cuda_memcpy_1<cpy_ne_D*4>(tmp, &VKQ_combine[(threadIdx.y + ip)*DVp + i0 + threadIdx.x*cpy_ne_D]);
1041
+ #pragma unroll
1042
+ for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
1043
+ ((float *)VKQ)[i0/warp_size + i1] += tmp[i1];
1044
+ }
1045
+ }
1046
+ #endif // FAST_FP16_AVAILABLE
1047
+
1048
+ KQ_sum[0] += KQ_sum_combine[threadIdx.y + ip];
1049
+ }
1050
+ }
1051
+
1052
+ // Attention sink: adjust KQ max and sum only for the first of all parallel blocks:
1053
+ if (sinks && blockIdx.y == 0) {
1054
+ #pragma unroll
1055
+ for (int jc0 = 0; jc0 < cpw; ++jc0) {
1056
+ const int jc = jc0 + (threadIdx.y/np)*cpw;
1057
+ const float sink = ((const float *) sinks)[head0 + jc % ncols2];
1058
+
1059
+ float KQ_max_new_j = fmaxf(KQ_max[jc0], sink);
1060
+ const float KQ_max_scale = expf(KQ_max[jc0] - KQ_max_new_j);
1061
+ KQ_max[jc0] = KQ_max_new_j;
1062
+
1063
+ const float val = expf(sink - KQ_max[jc0]);
1064
+ KQ_sum[jc0] = KQ_sum[jc0]*KQ_max_scale + val;
1065
+
1066
+ #ifdef FAST_FP16_AVAILABLE
1067
+ const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
1068
+ #pragma unroll
1069
+ for (int i0 = 0; i0 < DVp/2; i0 += warp_size) {
1070
+ VKQ[jc0*((DVp/2)/warp_size) + i0/warp_size] *= KQ_max_scale_h2;
1071
+ }
1072
+ #else
1073
+ #pragma unroll
1074
+ for (int i0 = 0; i0 < DVp/2; i0 += warp_size) {
1075
+ VKQ[jc0*((DVp/2)/warp_size) + i0/warp_size].x *= KQ_max_scale;
1076
+ VKQ[jc0*((DVp/2)/warp_size) + i0/warp_size].y *= KQ_max_scale;
1077
+ }
1078
+ #endif // FAST_FP16_AVAILABLE
1079
+ }
1080
+ }
1081
+
1082
+ // Write back results:
1083
+ #pragma unroll
1084
+ for (int jc0 = 0; jc0 < cpw; ++jc0) {
1085
+ const int jc = jc0 + (threadIdx.y/np)*cpw;
1086
+
1087
+ const int j = jc / ncols2;
1088
+ const int c = jc % ncols2;
1089
+
1090
+ if (ncols1 > 1 && col_Q_0 + j >= int(ne01.z)) {
1091
+ return;
1092
+ }
1093
+
1094
+ const float scale = gridDim.y == 1 ? 1.0f/KQ_sum[jc0] : 1.0f;
1095
+
1096
+ const int j_dst_unrolled = ((sequence*int(ne01.z) + col_Q_0 + j)*ne02 + head0 + c)*gridDim.y + blockIdx.y;
1097
+
1098
+ #ifdef FAST_FP16_AVAILABLE
1099
+ constexpr int cpy_ne_D = cpy_ne/2 < (DVp/2)/warp_size ? cpy_ne/2 : (DVp/2)/warp_size;
1100
+ #pragma unroll
1101
+ for (int i0 = 0; i0 < DVp/2; i0 += warp_size*cpy_ne_D) {
1102
+ __align__(16) float2 tmp[cpy_ne_D];
1103
+ #pragma unroll
1104
+ for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
1105
+ tmp[i1] = __half22float2(VKQ[jc0*((DVp/2)/warp_size) + i0/warp_size + i1]);
1106
+ tmp[i1].x *= scale;
1107
+ tmp[i1].y *= scale;
1108
+ }
1109
+ if (i0 + warp_size*cpy_ne_D <= DV/2 || i0 + threadIdx.x*cpy_ne_D < DV/2) {
1110
+ ggml_cuda_memcpy_1<sizeof(tmp)>(&dst[j_dst_unrolled*DV + 2*i0 + threadIdx.x*(2*cpy_ne_D)], tmp);
1111
+ }
1112
+ }
1113
+ #else
1114
+ constexpr int cpy_ne_D = cpy_ne < DVp/warp_size ? cpy_ne : DVp/warp_size;
1115
+ #pragma unroll
1116
+ for (int i0 = 0; i0 < DVp; i0 += warp_size*cpy_ne_D) {
1117
+ if (i0 + warp_size*cpy_ne_D <= DV || i0 + threadIdx.x*cpy_ne_D < DV) {
1118
+ #pragma unroll
1119
+ for (int i1 = 0; i1 < cpy_ne_D/2; ++i1) {
1120
+ VKQ[jc0*((DVp/2)/warp_size) + i0/(2*warp_size) + i1].x *= scale;
1121
+ VKQ[jc0*((DVp/2)/warp_size) + i0/(2*warp_size) + i1].y *= scale;
1122
+ }
1123
+ ggml_cuda_memcpy_1<cpy_ne_D*4>(
1124
+ &dst[j_dst_unrolled*DV + i0 + threadIdx.x*cpy_ne_D],
1125
+ &VKQ[jc0*((DVp/2)/warp_size) + i0/(2*warp_size)]);
1126
+ }
1127
+ }
1128
+ #endif // FAST_FP16_AVAILABLE
1129
+
1130
+ if (gridDim.y != 1 && threadIdx.x == 0) {
1131
+ dst_meta[j_dst_unrolled] = make_float2(KQ_max[jc0], KQ_sum[jc0]);
1132
+ }
1133
+ }
1134
+ #else
1135
+ GGML_UNUSED_VARS(Q_ptr, K_ptr, V_ptr, mask_ptr, sinks_ptr, KV_max_ptr, dst_ptr, dst_meta_ptr, scale,
1136
+ max_bias, m0, m1, n_head_log2, logit_softcap,
1137
+ ne00, ne01, ne02, ne03,
1138
+ nb01, nb02, nb03,
1139
+ ne10, ne11, ne12, ne13,
1140
+ nb11, nb12, nb13,
1141
+ nb21, nb22, nb23,
1142
+ ne31, ne32, ne33,
1143
+ nb31, nb32, nb33);
1144
+ NO_DEVICE_CODE;
1145
+ #endif // FLASH_ATTN_AVAILABLE
1146
+ }
1147
+
1148
+ template <int DKQ, int DV, int ncols2, bool use_logit_softcap>
1149
+ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
1150
+ const ggml_tensor * Q = dst->src[0];
1151
+
1152
+ const int id = ggml_cuda_get_device();
1153
+ const int cc = ggml_cuda_info().devices[id].cc;
1154
+ const int warp_size = 32;
1155
+
1156
+ constexpr size_t nbytes_shared = 0;
1157
+
1158
+ #ifdef GGML_USE_HIP
1159
+ if constexpr (DKQ <= 128) {
1160
+ if (Q->ne[1] > 32/ncols2) {
1161
+ constexpr int cols_per_block = 64;
1162
+ const int nwarps = ggml_cuda_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
1163
+ const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
1164
+ fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
1165
+ launch_fattn<DV, cols_per_block/ncols2, ncols2>
1166
+ (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, false, warp_size);
1167
+ return;
1168
+ }
1169
+ }
1170
+ #endif // GGML_USE_HIP
1171
+
1172
+ #ifndef GGML_USE_HIP
1173
+ if constexpr (DKQ <= 256)
1174
+ #endif // GGML_USE_HIP
1175
+ {
1176
+ if (Q->ne[1] > 16/ncols2) {
1177
+ constexpr int cols_per_block = 32;
1178
+ const int nwarps = ggml_cuda_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
1179
+ const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
1180
+ fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
1181
+ launch_fattn<DV, cols_per_block/ncols2, ncols2>
1182
+ (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, false, warp_size);
1183
+ return;
1184
+ }
1185
+ }
1186
+
1187
+ if constexpr (ncols2 <= 16) {
1188
+ if (Q->ne[1] > 8/ncols2) {
1189
+ constexpr int cols_per_block = 16;
1190
+ const int nwarps = ggml_cuda_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
1191
+ const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
1192
+ fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
1193
+ launch_fattn<DV, cols_per_block/ncols2, ncols2>
1194
+ (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, false, warp_size);
1195
+ return;
1196
+ }
1197
+ }
1198
+
1199
+ if constexpr (ncols2 <= 8) {
1200
+ if (Q->ne[1] > 4/ncols2) {
1201
+ constexpr int cols_per_block = 8;
1202
+ const int nwarps = ggml_cuda_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
1203
+ const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
1204
+ fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
1205
+ launch_fattn<DV, cols_per_block/ncols2, ncols2>
1206
+ (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, false, warp_size);
1207
+ return;
1208
+ }
1209
+ }
1210
+
1211
+ if constexpr (ncols2 <= 4) {
1212
+ if (Q->ne[1] > 2/ncols2) {
1213
+ constexpr int cols_per_block = 4;
1214
+ const int nwarps = ggml_cuda_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
1215
+ const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
1216
+ fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
1217
+ launch_fattn<DV, cols_per_block/ncols2, ncols2>
1218
+ (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, false, warp_size);
1219
+ return;
1220
+ }
1221
+ }
1222
+
1223
+ if constexpr (ncols2 <= 2) {
1224
+ constexpr int cols_per_block = 2;
1225
+ const int nwarps = ggml_cuda_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
1226
+ const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
1227
+ fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
1228
+ launch_fattn<DV, cols_per_block/ncols2, ncols2>
1229
+ (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, false, warp_size);
1230
+ return;
1231
+ }
1232
+
1233
+ GGML_ABORT("fatal error");
1234
+ }
1235
+
1236
+ template <int DKQ, int DV, bool use_logit_softcap>
1237
+ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
1238
+ const ggml_tensor * KQV = dst;
1239
+ const ggml_tensor * Q = dst->src[0];
1240
+ const ggml_tensor * K = dst->src[1];
1241
+ const ggml_tensor * mask = dst->src[3];
1242
+
1243
+ float max_bias = 0.0f;
1244
+ memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
1245
+
1246
+ GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
1247
+ const int gqa_ratio = Q->ne[2] / K->ne[2];
1248
+
1249
+ // On NVIDIA (Pascal and older) the GQA optimizations seem to be detrimental in some cases.
1250
+ // However, for DKQ == 576, DV == 512 only the kernel variant with GQA optimizations is implemented.
1251
+ const bool nvidia = GGML_CUDA_CC_IS_NVIDIA(ggml_cuda_info().devices[ggml_cuda_get_device()].cc);
1252
+ const int gqa_limit = nvidia && gqa_ratio <= 4 && DV <= 256 ? 16 : INT_MAX;
1253
+ const bool use_gqa_opt = mask && max_bias == 0.0f && Q->ne[1] <= gqa_limit && K->ne[1] % FATTN_KQ_STRIDE == 0;
1254
+
1255
+ if constexpr (DKQ == 320) {
1256
+ // This branch is only used for Mistral Small 4 which has a GQA ratio of 32.
1257
+ // On AMD, simply use that GQA ratio with 32 columns / block since we always have enough SRAM.
1258
+ // On NVIDIA however, the tile kernel is only used for GPUs that can't use the mma kernel (Pascal and older).
1259
+ // Therefore, use a GQA ratio of 16 with 16 columns / block to stay below 48 kiB of SRAM / block.
1260
+ #ifdef GGML_USE_HIP
1261
+ if (use_gqa_opt && gqa_ratio % 32 == 0) {
1262
+ launch_fattn_tile_switch_ncols1<DKQ, DV, 32, use_logit_softcap>(ctx, dst);
1263
+ return;
1264
+ }
1265
+ #else
1266
+ if (use_gqa_opt && gqa_ratio % 16 == 0) {
1267
+ launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
1268
+ return;
1269
+ }
1270
+ #endif // GGML_USE_HIP
1271
+ GGML_ABORT("flash-attn tile (320/256): expected GQA ratio multiple of 32");
1272
+ }
1273
+
1274
+ if constexpr (DKQ == 576) {
1275
+ if (use_gqa_opt && gqa_ratio % 16 == 0) {
1276
+ launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
1277
+ return;
1278
+ }
1279
+ if (use_gqa_opt && gqa_ratio % 4 == 0) {
1280
+ launch_fattn_tile_switch_ncols1<DKQ, DV, 4, use_logit_softcap>(ctx, dst);
1281
+ return;
1282
+ }
1283
+ }
1284
+
1285
+ if constexpr (DKQ == 192) {
1286
+ // MiMo-V2.5 / V2.5-Pro / V2-Flash: gqa_ratio is 8 (SWA) or 16 (full attn)
1287
+ if (use_gqa_opt && gqa_ratio % 16 == 0) {
1288
+ launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
1289
+ return;
1290
+ }
1291
+ if (use_gqa_opt && gqa_ratio % 8 == 0) {
1292
+ launch_fattn_tile_switch_ncols1<DKQ, DV, 8, use_logit_softcap>(ctx, dst);
1293
+ return;
1294
+ }
1295
+ GGML_ABORT("flash-attn tile (192/128): expected GQA ratio multiple of 8");
1296
+ }
1297
+
1298
+ if constexpr (DKQ <= 512 && DKQ != 320 && DKQ != 192) {
1299
+ if (use_gqa_opt && gqa_ratio % 8 == 0) {
1300
+ launch_fattn_tile_switch_ncols1<DKQ, DV, 8, use_logit_softcap>(ctx, dst);
1301
+ return;
1302
+ }
1303
+
1304
+ if (use_gqa_opt && gqa_ratio % 4 == 0) {
1305
+ launch_fattn_tile_switch_ncols1<DKQ, DV, 4, use_logit_softcap>(ctx, dst);
1306
+ return;
1307
+ }
1308
+
1309
+ if (use_gqa_opt && gqa_ratio % 2 == 0) {
1310
+ launch_fattn_tile_switch_ncols1<DKQ, DV, 2, use_logit_softcap>(ctx, dst);
1311
+ return;
1312
+ }
1313
+
1314
+ if constexpr (DV <= 256) {
1315
+ launch_fattn_tile_switch_ncols1<DKQ, DV, 1, use_logit_softcap>(ctx, dst);
1316
+ return;
1317
+ }
1318
+ }
1319
+ GGML_ABORT("fatal error");
1320
+ }
1321
+
1322
+ template <int DKQ, int DV>
1323
+ void ggml_cuda_flash_attn_ext_tile_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
1324
+ const ggml_tensor * KQV = dst;
1325
+
1326
+ float logit_softcap;
1327
+ memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
1328
+
1329
+ if (logit_softcap == 0.0f) {
1330
+ constexpr bool use_logit_softcap = false;
1331
+ launch_fattn_tile_switch_ncols2<DKQ, DV, use_logit_softcap>(ctx, dst);
1332
+ } else {
1333
+ constexpr bool use_logit_softcap = true;
1334
+ launch_fattn_tile_switch_ncols2<DKQ, DV, use_logit_softcap>(ctx, dst);
1335
+ }
1336
+ }
1337
+
1338
+ void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
1339
+
1340
+ #define DECL_FATTN_TILE_CASE(DKQ, DV) \
1341
+ template void ggml_cuda_flash_attn_ext_tile_case \
1342
+ <DKQ, DV>(ggml_backend_cuda_context & ctx, ggml_tensor * dst) \
1343
+
1344
+ extern DECL_FATTN_TILE_CASE( 40, 40);
1345
+ extern DECL_FATTN_TILE_CASE( 64, 64);
1346
+ extern DECL_FATTN_TILE_CASE( 72, 72);
1347
+ extern DECL_FATTN_TILE_CASE( 80, 80);
1348
+ extern DECL_FATTN_TILE_CASE( 96, 96);
1349
+ extern DECL_FATTN_TILE_CASE(112, 112);
1350
+ extern DECL_FATTN_TILE_CASE(128, 128);
1351
+ extern DECL_FATTN_TILE_CASE(192, 128);
1352
+ extern DECL_FATTN_TILE_CASE(256, 256);
1353
+ extern DECL_FATTN_TILE_CASE(320, 256);
1354
+ extern DECL_FATTN_TILE_CASE(512, 512);
1355
+ extern DECL_FATTN_TILE_CASE(576, 512);