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 +1,1768 @@
1
+ #pragma once
2
+
3
+ #include "vecdotq.cuh"
4
+
5
+ #include "mmq.cuh"
6
+
7
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q1_0(
8
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
9
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
10
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
11
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
12
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
13
+
14
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
15
+ int * x_qs = (int *) x_tile;
16
+ float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K);
17
+ #else
18
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, I);
19
+ int * x_qs = (int *) x_tile;
20
+ float * x_df = (float *) (x_qs + txs.qs);
21
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
22
+
23
+ constexpr int blocks_per_iter = MMQ_ITER_K / QK1_0;
24
+ constexpr int threads_per_row = blocks_per_iter * QI1_0;
25
+ constexpr int nrows = warp_size / threads_per_row;
26
+ constexpr int scale_entries_per_block = QK1_0 / QK8_1;
27
+ constexpr int scale_entries_per_row = blocks_per_iter * scale_entries_per_block;
28
+
29
+ const int txi = threadIdx.x % threads_per_row;
30
+ const int kbx = txi / QI1_0;
31
+ const int kqsx = txi % QI1_0;
32
+
33
+ #pragma unroll
34
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
35
+ int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row;
36
+
37
+ if (fallback) {
38
+ i = min(i, i_max);
39
+ }
40
+
41
+ const block_q1_0 * bxi = (const block_q1_0 *) x + kbx0 + i*stride + kbx;
42
+ const int16_t * qxi = (const int16_t *) bxi->qs + kqsx * 2;
43
+
44
+ const int dst_offset = kbx*(scale_entries_per_block*QI8_0) + kqsx*QI8_0;
45
+ #pragma unroll
46
+ for (int j = 0; j < 2; ++j) {
47
+ const int q = qxi[j];
48
+
49
+ // unpack crumbs into nibble indices
50
+ const int n0 = __byte_perm(0x11100100, 0x11100100, q >> 0); // [0, 1, 4, 5] [ 8, 9, 12, 13]
51
+ const int n1 = __byte_perm(0x11100100, 0x11100100, q >> 2); // [2, 3, 6, 7] [10, 11, 14, 15]
52
+ // unpack nibbles into byte values
53
+ const int s0 = __byte_perm(0x01FF, 0x01FF, n0 >> 0);
54
+ const int s1 = __byte_perm(0x01FF, 0x01FF, n1 >> 0);
55
+ const int s2 = __byte_perm(0x01FF, 0x01FF, n0 >> 16);
56
+ const int s3 = __byte_perm(0x01FF, 0x01FF, n1 >> 16);
57
+ // unshuffle values
58
+ const int v0 = __byte_perm(s0, s1, 0x5410);
59
+ const int v1 = __byte_perm(s0, s1, 0x7632);
60
+ const int v2 = __byte_perm(s2, s3, 0x5410);
61
+ const int v3 = __byte_perm(s2, s3, 0x7632);
62
+
63
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
64
+ x_qs[i*sram_stride + dst_offset + j*4+0] = v0;
65
+ x_qs[i*sram_stride + dst_offset + j*4+1] = v1;
66
+ x_qs[i*sram_stride + dst_offset + j*4+2] = v2;
67
+ x_qs[i*sram_stride + dst_offset + j*4+3] = v3;
68
+ #else
69
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j*4+0] = v0;
70
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j*4+1] = v1;
71
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j*4+2] = v2;
72
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j*4+3] = v3;
73
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
74
+ }
75
+ }
76
+
77
+ const int ksx = threadIdx.x % scale_entries_per_row;
78
+ const int scale_block = ksx / scale_entries_per_block;
79
+
80
+ #pragma unroll
81
+ for (int i0 = 0; i0 < I; i0 += nwarps) {
82
+ int i = i0 + threadIdx.y;
83
+
84
+ if (fallback) {
85
+ i = min(i, i_max);
86
+ }
87
+
88
+ const block_q1_0 * bxi = (const block_q1_0 *) x + kbx0 + i*stride + scale_block;
89
+
90
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
91
+ x_df[i*sram_stride + ksx] = bxi->d;
92
+ #else
93
+ x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + ksx] = bxi->d;
94
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
95
+ }
96
+ }
97
+
98
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q2_0(
99
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
100
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
101
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
102
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
103
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
104
+
105
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
106
+ int * x_qs = (int *) x_tile;
107
+ float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K);
108
+ #else
109
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, I);
110
+ int * x_qs = (int *) x_tile;
111
+ float * x_df = (float *) (x_qs + txs.qs);
112
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
113
+
114
+ constexpr int blocks_per_iter = MMQ_ITER_K / QK2_0;
115
+ constexpr int threads_per_row = blocks_per_iter * QI2_0;
116
+ constexpr int nrows = warp_size / threads_per_row;
117
+ constexpr int scale_entries_per_block = QK2_0 / QK8_1;
118
+ constexpr int scale_entries_per_row = blocks_per_iter * scale_entries_per_block;
119
+
120
+ const int txi = threadIdx.x % threads_per_row;
121
+ const int kbx = txi / QI2_0;
122
+ const int kqsx = txi % QI2_0;
123
+
124
+ #pragma unroll
125
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
126
+ int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row;
127
+
128
+ if (fallback) {
129
+ i = min(i, i_max);
130
+ }
131
+
132
+ const block_q2_0 * bxi = (const block_q2_0 *) x + kbx0 + i*stride + kbx;
133
+ const int16_t * qxi = (const int16_t *) bxi->qs + kqsx * 4;
134
+
135
+ const int dst_offset = kbx*(scale_entries_per_block*QI8_0) + kqsx*QI8_0;
136
+
137
+ #pragma unroll
138
+ for (int j = 0; j < 4; ++j) {
139
+ const int q = qxi[j];
140
+
141
+ #if defined(GGML_USE_HIP)
142
+ const uint32_t qx_indices = (q & 0x03) | ((q & 0x0C) << 6) | ((q & 0x30) << 12) | ((q & 0xC0) << 18);
143
+ const uint32_t qy_bits = q >> 8;
144
+ const uint32_t qy_indices = (qy_bits & 0x03) | ((qy_bits & 0x0C) << 6) | ((qy_bits & 0x30) << 12) | ((qy_bits & 0xC0) << 18);
145
+ const int qx = __builtin_amdgcn_perm(0x020100FF, 0x020100FF, qx_indices);
146
+ const int qy = __builtin_amdgcn_perm(0x020100FF, 0x020100FF, qy_indices);
147
+ #else
148
+ // unpack even and odd crumbs into byte values
149
+ const int qe = __byte_perm(0x020100FF, 0x020100FF, q >> 0);
150
+ const int qo = __byte_perm(0x020100FF, 0x020100FF, q >> 2);
151
+ // unshuffle values
152
+ const int qx = __byte_perm(qe, qo, 0x5140);
153
+ const int qy = __byte_perm(qe, qo, 0x7362);
154
+ #endif // defined(GGML_USE_HIP)
155
+
156
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
157
+ x_qs[i*sram_stride + dst_offset + j*2+0] = qx;
158
+ x_qs[i*sram_stride + dst_offset + j*2+1] = qy;
159
+ #else
160
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j*2+0] = qx;
161
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j*2+1] = qy;
162
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
163
+ }
164
+ }
165
+
166
+ const int ksx = threadIdx.x % scale_entries_per_row;
167
+ const int scale_block = ksx / scale_entries_per_block;
168
+
169
+ #pragma unroll
170
+ for (int i0 = 0; i0 < I; i0 += nwarps) {
171
+ int i = i0 + threadIdx.y;
172
+
173
+ if (fallback) {
174
+ i = min(i, i_max);
175
+ }
176
+
177
+ const block_q2_0 * bxi = (const block_q2_0 *) x + kbx0 + i*stride + scale_block;
178
+
179
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
180
+ x_df[i*sram_stride + ksx] = bxi->d;
181
+ #else
182
+ x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + ksx] = bxi->d;
183
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
184
+ }
185
+ }
186
+
187
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q4_0(
188
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
189
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
190
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
191
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
192
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
193
+
194
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
195
+ int * x_qs = (int *) x_tile;
196
+ float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K);
197
+ #else
198
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, I);
199
+ int * x_qs = (int *) x_tile;
200
+ float * x_df = (float *) (x_qs + txs.qs);
201
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
202
+
203
+ constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_0);
204
+ constexpr int nrows = warp_size / threads_per_row;
205
+ const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x;
206
+ const int kbx = txi / QI4_0;
207
+ const int kqsx = txi % QI4_0;
208
+
209
+ #pragma unroll
210
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
211
+ int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row);
212
+
213
+ if (fallback) {
214
+ i = min(i, i_max);
215
+ }
216
+
217
+ const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbx;
218
+ const int qs0 = get_int_b2(bxi->qs, kqsx);
219
+
220
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
221
+ x_qs[i*sram_stride + kbx*(2*QI4_0) + kqsx + 0] = __vsubss4((qs0 >> 0) & 0x0F0F0F0F, 0x08080808);
222
+ x_qs[i*sram_stride + kbx*(2*QI4_0) + kqsx + QI4_0] = __vsubss4((qs0 >> 4) & 0x0F0F0F0F, 0x08080808);
223
+ #else
224
+ x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0;
225
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
226
+ }
227
+
228
+ constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_0;
229
+ constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row;
230
+ const int kbxd = threadIdx.x % blocks_per_tile_x_row;
231
+
232
+ #pragma unroll
233
+ for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) {
234
+ int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row;
235
+
236
+ if (fallback) {
237
+ i = min(i, i_max);
238
+ }
239
+
240
+ const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbxd;
241
+
242
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
243
+ x_df[i*sram_stride + kbxd] = bxi->d;
244
+ #else
245
+ x_df[i*(MMQ_TILE_NE_K/QI4_0) + i/QI4_0 + kbxd] = bxi->d;
246
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
247
+ }
248
+ }
249
+
250
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q4_1(
251
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
252
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
253
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
254
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
255
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
256
+
257
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
258
+ int * x_qs = (int *) x_tile;
259
+ half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);
260
+ #else
261
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, I);
262
+ int * x_qs = (int *) x_tile;
263
+ half2 * x_dm = (half2 *) (x_qs + txs.qs);
264
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
265
+
266
+ constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_1);
267
+ constexpr int nrows = warp_size / threads_per_row;
268
+ const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x;
269
+ const int kbx = txi / QI4_1;
270
+ const int kqsx = txi % QI4_1;
271
+
272
+ #pragma unroll
273
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
274
+ int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row);
275
+
276
+ if (fallback) {
277
+ i = min(i, i_max);
278
+ }
279
+
280
+ const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbx;
281
+ const int qs0 = get_int_b4(bxi->qs, kqsx);
282
+
283
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
284
+ x_qs[i*sram_stride + kbx*(2*QI4_1) + kqsx + 0] = (qs0 >> 0) & 0x0F0F0F0F;
285
+ x_qs[i*sram_stride + kbx*(2*QI4_1) + kqsx + QI4_1] = (qs0 >> 4) & 0x0F0F0F0F;
286
+ #else
287
+ x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0;
288
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
289
+ }
290
+
291
+ constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_1;
292
+ constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row;
293
+ const int kbxd = threadIdx.x % blocks_per_tile_x_row;
294
+
295
+ #pragma unroll
296
+ for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) {
297
+ int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row;
298
+
299
+ if (fallback) {
300
+ i = min(i, i_max);
301
+ }
302
+
303
+ const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbxd;
304
+
305
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
306
+ x_dm[i*sram_stride + kbxd] = bxi->dm;
307
+ #else
308
+ x_dm[i*(MMQ_TILE_NE_K/QI4_1) + i/QI4_1 + kbxd] = bxi->dm;
309
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
310
+ }
311
+ }
312
+
313
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q5_0(
314
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
315
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
316
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
317
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
318
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
319
+
320
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
321
+ int * x_qs = (int *) x_tile;
322
+ float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
323
+ #else
324
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_0, I);
325
+ int * x_qs = (int *) x_tile;
326
+ float * x_df = (float *) (x_qs + txs.qs);
327
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
328
+
329
+ constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_0);
330
+ constexpr int nrows = warp_size / threads_per_row;
331
+ const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x;
332
+ const int kbx = txi / QI5_0;
333
+ const int kqsx = txi % QI5_0;
334
+
335
+ #pragma unroll
336
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
337
+ int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row);
338
+
339
+ if (fallback) {
340
+ i = min(i, i_max);
341
+ }
342
+
343
+ const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbx;
344
+
345
+ const int ql = get_int_b2(bxi->qs, kqsx);
346
+ const int qh = get_int_b2(bxi->qh, 0) >> (4 * kqsx);
347
+
348
+ int qs0 = (ql >> 0) & 0x0F0F0F0F;
349
+ qs0 |= (qh << 4) & 0x00000010; // 0 -> 4
350
+ qs0 |= (qh << 11) & 0x00001000; // 1 -> 12
351
+ qs0 |= (qh << 18) & 0x00100000; // 2 -> 20
352
+ qs0 |= (qh << 25) & 0x10000000; // 3 -> 28
353
+ qs0 = __vsubss4(qs0, 0x10101010); // subtract 16
354
+
355
+ int qs1 = (ql >> 4) & 0x0F0F0F0F;
356
+ qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4
357
+ qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12
358
+ qs1 |= (qh << 2) & 0x00100000; // 18 -> 20
359
+ qs1 |= (qh << 9) & 0x10000000; // 19 -> 28
360
+ qs1 = __vsubss4(qs1, 0x10101010); // subtract 16
361
+
362
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
363
+ x_qs[i*sram_stride + kbx*(2*QI5_0) + kqsx + 0] = qs0;
364
+ x_qs[i*sram_stride + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1;
365
+ #else
366
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + 0] = qs0;
367
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1;
368
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
369
+ }
370
+
371
+ constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_0;
372
+ constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row;
373
+ const int kbxd = threadIdx.x % blocks_per_tile_x_row;
374
+
375
+ #pragma unroll
376
+ for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) {
377
+ int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row;
378
+
379
+ if (fallback) {
380
+ i = min(i, i_max);
381
+ }
382
+
383
+ const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbxd;
384
+
385
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
386
+ x_df[i*sram_stride + kbxd] = bxi->d;
387
+ #else
388
+ x_df[i*(MMQ_TILE_NE_K/QI5_0) + i/QI5_0 + kbxd] = bxi->d;
389
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
390
+ }
391
+ }
392
+
393
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q5_1(
394
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
395
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
396
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
397
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
398
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
399
+
400
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
401
+ int * x_qs = (int *) x_tile;
402
+ half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);
403
+ #else
404
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, I);
405
+ int * x_qs = (int *) x_tile;
406
+ half2 * x_dm = (half2 *) (x_qs + txs.qs);
407
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
408
+
409
+ constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_1);
410
+ constexpr int nrows = warp_size / threads_per_row;
411
+ const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x;
412
+ const int kbx = txi / QI5_1;
413
+ const int kqsx = txi % QI5_1;
414
+
415
+ #pragma unroll
416
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
417
+ int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row);
418
+
419
+ if (fallback) {
420
+ i = min(i, i_max);
421
+ }
422
+
423
+ const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbx;
424
+
425
+ const int ql = get_int_b4(bxi->qs, kqsx);
426
+ const int qh = get_int_b4(bxi->qh, 0) >> (4 * kqsx);
427
+
428
+ int qs0 = (ql >> 0) & 0x0F0F0F0F;
429
+ qs0 |= (qh << 4) & 0x00000010; // 0 -> 4
430
+ qs0 |= (qh << 11) & 0x00001000; // 1 -> 12
431
+ qs0 |= (qh << 18) & 0x00100000; // 2 -> 20
432
+ qs0 |= (qh << 25) & 0x10000000; // 3 -> 28
433
+
434
+ int qs1 = (ql >> 4) & 0x0F0F0F0F;
435
+ qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4
436
+ qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12
437
+ qs1 |= (qh << 2) & 0x00100000; // 18 -> 20
438
+ qs1 |= (qh << 9) & 0x10000000; // 19 -> 28
439
+
440
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
441
+ x_qs[i*sram_stride + kbx*(2*QI5_1) + kqsx + 0] = qs0;
442
+ x_qs[i*sram_stride + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1;
443
+ #else
444
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + 0] = qs0;
445
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1;
446
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
447
+ }
448
+
449
+ constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_1;
450
+ constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row;
451
+ const int kbxd = threadIdx.x % blocks_per_tile_x_row;
452
+
453
+ #pragma unroll
454
+ for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) {
455
+ int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row;
456
+
457
+ if (fallback) {
458
+ i = min(i, i_max);
459
+ }
460
+
461
+ const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbxd;
462
+
463
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
464
+ x_dm[i*sram_stride + kbxd] = bxi->dm;
465
+ #else
466
+ x_dm[i*(MMQ_TILE_NE_K/QI5_1) + i/QI5_1 + kbxd] = bxi->dm;
467
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
468
+ }
469
+ }
470
+
471
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q8_0(
472
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
473
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
474
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
475
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
476
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
477
+
478
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
479
+ int * x_qs = (int *) x_tile;
480
+ float * x_df = (float *) (x_tile + 2*MMQ_TILE_NE_K);
481
+ #else
482
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, I);
483
+ int * x_qs = (int *) x_tile;
484
+ float * x_df = (float *) (x_qs + txs.qs);
485
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
486
+
487
+ // MMQ_ITER_K / (4 * QR8_0) == 64 required. but NV has only 32 threads per warp
488
+ constexpr int threads_per_row = 32;
489
+ constexpr int nrows = warp_size / threads_per_row;
490
+ const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x;
491
+ const int kbx = txi / QI8_0;
492
+ const int kqsx = txi % QI8_0;
493
+
494
+ #pragma unroll
495
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
496
+ int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row);
497
+
498
+ if (fallback) {
499
+ i = min(i, i_max);
500
+ }
501
+
502
+ const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbx;
503
+
504
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
505
+ x_qs[i*sram_stride + 0 + txi] = get_int_b2(bxi[0].qs, kqsx);
506
+ x_qs[i*sram_stride + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx);
507
+ #else
508
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + 0 + txi] = get_int_b2(bxi[0].qs, kqsx);
509
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx);
510
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
511
+ }
512
+
513
+ constexpr int blocks_per_tile_x_row = 2*MMQ_TILE_NE_K / QI8_0;
514
+ constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row;
515
+ const int kbxd = threadIdx.x % blocks_per_tile_x_row;
516
+
517
+ #pragma unroll
518
+ for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) {
519
+ int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row;
520
+
521
+ if (fallback) {
522
+ i = min(i, i_max);
523
+ }
524
+
525
+ const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbxd;
526
+
527
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
528
+ x_df[i*sram_stride + kbxd] = bxi->d;
529
+ #else
530
+ x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + kbxd] = bxi->d;
531
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
532
+ }
533
+ }
534
+
535
+ // ---------------------------------------------------------------------------------------------
536
+
537
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q2_K(
538
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
539
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
540
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
541
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
542
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
543
+
544
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
545
+ int * x_qs = (int *) x_tile;
546
+ half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);
547
+ #else
548
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, I);
549
+ int * x_qs = (int *) x_tile;
550
+ half2 * x_dm = (half2 *) (x_qs + txs.qs);
551
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
552
+
553
+ constexpr int threads_per_row = MMQ_ITER_K / (4 * QR2_K);
554
+ constexpr int nrows = ggml_cuda_get_physical_warp_size() / threads_per_row;
555
+ const int kqsx = threadIdx.x % threads_per_row;
556
+
557
+ #pragma unroll
558
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
559
+ int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row;
560
+
561
+ if (fallback) {
562
+ i = min(i, i_max);
563
+ }
564
+
565
+ const block_q2_K * bxi = (const block_q2_K *) x + kbx0 + i*stride;
566
+
567
+ const int x_ql_0 = get_int_b2(bxi->qs, kqsx);
568
+
569
+ #pragma unroll
570
+ for (int l = 0; l < QR2_K; ++l) {
571
+ const int k = (kqsx/8)*32 + l*8 + kqsx % 8;
572
+
573
+ const int x_qs_k = (x_ql_0 >> (2*l)) & 0x03030303;
574
+
575
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
576
+ x_qs[i*sram_stride + k] = x_qs_k;
577
+ #else
578
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k;
579
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
580
+ }
581
+
582
+ const int sc_m = bxi->scales[kqsx];
583
+ #ifdef FAST_FP16_AVAILABLE
584
+ const half2 x_dm_ik = __hmul2(bxi->dm, make_half2(sc_m & 0x0F, sc_m >> 4));
585
+ #else
586
+ const float2 bxi_dmf = __half22float2(bxi->dm);
587
+ const half2 x_dm_ik = make_half2(bxi_dmf.x*(sc_m & 0x0F), bxi_dmf.y*(sc_m >> 4));
588
+ #endif // FAST_FP16_AVAILABLE
589
+
590
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
591
+ x_dm[i*sram_stride + kqsx] = x_dm_ik;
592
+ #else
593
+ x_dm[i*(MMQ_TILE_NE_K + 1) + kqsx] = x_dm_ik;
594
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
595
+ }
596
+ }
597
+
598
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q3_K(
599
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
600
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
601
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
602
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
603
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
604
+
605
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
606
+ int * x_qs = (int *) x_tile;
607
+ float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
608
+ #else
609
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, I);
610
+ int * x_qs = (int *) x_tile;
611
+ float * x_df = (float *) (x_qs + txs.qs);
612
+ int * x_sc = (int *) (x_df + txs.dm);
613
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
614
+
615
+ constexpr int threads_per_row = MMQ_ITER_K / (4 * QR3_K);
616
+ constexpr int nrows = warp_size / threads_per_row;
617
+ const int kqsx = threadIdx.x % threads_per_row;
618
+
619
+ #pragma unroll
620
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
621
+ int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row;
622
+
623
+ if (fallback) {
624
+ i = min(i, i_max);
625
+ }
626
+
627
+ const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride;
628
+
629
+ const int x_ql_0 = get_int_b2(bxi->qs, kqsx);
630
+ const int x_qh_0 = get_int_b2(bxi->hmask, kqsx % (QI3_K/2)) >> (4 * (kqsx / (QI3_K/2)));
631
+
632
+ #pragma unroll
633
+ for (int l = 0; l < QR3_K; ++l) {
634
+ const int k = (kqsx/8)*32 + l*8 + kqsx % 8;
635
+
636
+ const int x_ql_k = (x_ql_0 >> (2*l)) & 0x03030303;
637
+ const int x_qh_k = ((x_qh_0 >> l) << 2) & 0x04040404;
638
+
639
+ const int x_qs_k = __vsubss4(x_ql_k | x_qh_k, 0x04040404);
640
+
641
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
642
+ x_qs[i*sram_stride + k] = x_qs_k;
643
+ #else
644
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k;
645
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
646
+ }
647
+ }
648
+
649
+ constexpr int rows_per_warp = warp_size / 4;
650
+ #pragma unroll
651
+ for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) {
652
+ int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/4;
653
+
654
+ if (fallback) {
655
+ i = min(i, i_max);
656
+ }
657
+
658
+ const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride;
659
+
660
+ const int ksc = threadIdx.x % 4;
661
+
662
+ const int ksc_low = ksc % (QI3_K/8);
663
+ const int shift_low = 4 * (ksc / (QI3_K/8));
664
+ const int sc_low = (get_int_b2(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F;
665
+
666
+ const int ksc_high = QI3_K/8;
667
+ const int shift_high = 2 * ksc;
668
+ const int sc_high = ((get_int_b2(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030;
669
+
670
+ const int sc = __vsubss4(sc_low | sc_high, 0x20202020);
671
+
672
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
673
+ const int8_t * sc8 = (const int8_t *) &sc;
674
+ const float d = bxi->d;
675
+
676
+ #pragma unroll
677
+ for (int l = 0; l < int(sizeof(int)); ++l) {
678
+ x_df[i*sram_stride + sizeof(int)*ksc + l] = d*sc8[l];
679
+ }
680
+ #else
681
+ x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = sc;
682
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
683
+ }
684
+
685
+ #if !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE))
686
+ #pragma unroll
687
+ for (int i0 = 0; i0 < I; i0 += nwarps*warp_size) {
688
+ int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % I;
689
+
690
+ if (fallback) {
691
+ i = min(i, i_max);
692
+ }
693
+
694
+ const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride;
695
+
696
+ x_df[i] = bxi->d;
697
+ }
698
+ #endif // !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)) || defined(AMD_WMMA_AVAILABLE)
699
+ }
700
+
701
+ static __device__ __forceinline__ int unpack_scales_q45_K(const int * scales, const int ksc) {
702
+ // scale arrangement after the following two lines:
703
+ // - ksc == 0: sc0, sc1, sc2, sc3
704
+ // - ksc == 1: sc4, sc5, sc6, sc7
705
+ // - ksc == 2: m0, m1, m2, m3
706
+ // - ksc == 3: m4, m5, m6, m7
707
+ return ((scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F) | // lower 4 bits
708
+ ((scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030); // upper 2 bits
709
+ }
710
+
711
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q4_K(
712
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
713
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
714
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
715
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
716
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
717
+
718
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
719
+ int * x_qs = (int *) x_tile;
720
+ half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K);
721
+ #else
722
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, I);
723
+ int * x_qs = (int *) x_tile;
724
+ half2 * x_dm = (half2 *) (x_qs + txs.qs);
725
+ int * x_sc = (int *) (x_dm + txs.dm);
726
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
727
+
728
+ constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_K);
729
+ constexpr int nrows = warp_size / threads_per_row;
730
+ const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x;
731
+
732
+ #pragma unroll
733
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
734
+ int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row);
735
+
736
+ if (fallback) {
737
+ i = min(i, i_max);
738
+ }
739
+
740
+ const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride;
741
+ const int qs0 = get_int_b4(bxi->qs, txi);
742
+
743
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
744
+ x_qs[i*sram_stride + 16*(txi/8) + txi % 8 + 0] = (qs0 >> 0) & 0x0F0F0F0F;
745
+ x_qs[i*sram_stride + 16*(txi/8) + txi % 8 + 8] = (qs0 >> 4) & 0x0F0F0F0F;
746
+ #else
747
+ x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0;
748
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
749
+ }
750
+
751
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
752
+ constexpr int rows_per_warp = warp_size / 2;
753
+ #pragma unroll
754
+ for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) {
755
+ #if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
756
+ // Need if on AMD instead of % because warp_size == 64
757
+ // This causes double work and throughput loss (MI300X)
758
+ // H100 loses about 100 t/s with 'if' condition over '%'
759
+ int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/2;
760
+ if (i < I) {
761
+ #else
762
+ int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/2) % I;
763
+ {
764
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
765
+ if (fallback) {
766
+ i = min(i, i_max);
767
+ }
768
+
769
+ const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride;
770
+
771
+ const int * scales = (const int *) bxi->scales;
772
+ const int ksc = threadIdx.x % 2;
773
+
774
+ const int sc32 = unpack_scales_q45_K(scales, ksc + 0);
775
+ const int m32 = unpack_scales_q45_K(scales, ksc + 2);
776
+
777
+ const uint8_t * sc8 = (const uint8_t *) &sc32;
778
+ const uint8_t * m8 = (const uint8_t *) &m32;
779
+
780
+ const half2 dm = bxi->dm * make_half2(1.0f, -1.0f);
781
+
782
+ #pragma unroll
783
+ for (int l = 0; l < sizeof(int); ++l) {
784
+ x_dm[i*sram_stride + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]);
785
+ }
786
+ }
787
+ }
788
+ #else
789
+ #pragma unroll
790
+ for (int i0 = 0; i0 < I; i0 += nwarps*warp_size) {
791
+ int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % I;
792
+
793
+ if (fallback) {
794
+ i = min(i, i_max);
795
+ }
796
+
797
+ const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride;
798
+
799
+ x_dm[i] = bxi->dm;
800
+ }
801
+ constexpr int rows_per_warp = warp_size / 4;
802
+ #pragma unroll
803
+ for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) {
804
+ int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % I;
805
+
806
+ if (fallback) {
807
+ i = min(i, i_max);
808
+ }
809
+
810
+ const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / (QI4_K/8);
811
+
812
+ const int * scales = (const int *) bxi->scales;
813
+
814
+ const int ksc = threadIdx.x % (MMQ_TILE_NE_K/8);
815
+ const int scales8 = unpack_scales_q45_K(scales, ksc);
816
+
817
+ x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8;
818
+ }
819
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
820
+ }
821
+
822
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q5_K(
823
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
824
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
825
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
826
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
827
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
828
+
829
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
830
+ int * x_qs = (int *) x_tile;
831
+ half2 * x_dm = (half2 *) (x_qs + MMQ_TILE_NE_K*2);
832
+ #else
833
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, I);
834
+ int * x_qs = (int *) x_tile;
835
+ half2 * x_dm = (half2 *) (x_qs + txs.qs);
836
+ int * x_sc = (int *) (x_dm + txs.dm);
837
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
838
+
839
+ constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_K);
840
+ constexpr int nrows = warp_size / threads_per_row;
841
+ const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x;
842
+
843
+ #pragma unroll
844
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
845
+ int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row);
846
+
847
+ if (fallback) {
848
+ i = min(i, i_max);
849
+ }
850
+
851
+ const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
852
+ const int ky = QR5_K*txi;
853
+
854
+ const int ql = get_int_b4(bxi->qs, txi);
855
+ const int ql0 = (ql >> 0) & 0x0F0F0F0F;
856
+ const int ql1 = (ql >> 4) & 0x0F0F0F0F;
857
+
858
+ const int qh = get_int_b4(bxi->qh, txi % (QI5_K/4));
859
+ const int qh0 = ((qh >> (2 * (txi / (QI5_K/4)) + 0)) << 4) & 0x10101010;
860
+ const int qh1 = ((qh >> (2 * (txi / (QI5_K/4)) + 1)) << 4) & 0x10101010;
861
+
862
+ const int kq0 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + 0;
863
+ const int kq1 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + QI5_K/4;
864
+
865
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
866
+ x_qs[i*sram_stride + kq0] = ql0 | qh0;
867
+ x_qs[i*sram_stride + kq1] = ql1 | qh1;
868
+ #else
869
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = ql0 | qh0;
870
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = ql1 | qh1;
871
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
872
+ }
873
+
874
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
875
+ constexpr int rows_per_warp = warp_size / 2;
876
+ #pragma unroll
877
+ for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) {
878
+ #if defined(AMD_MFMA_AVAILABLE)
879
+ // Need if on AMD instead of % because warp_size == 64
880
+ // This causes double work and throughput loss (MI300X)
881
+ // H100 loses about 100 t/s with 'if' condition over '%'
882
+ int i = i0 + threadIdx.y*rows_per_warp + threadIdx.x/2;
883
+ if (i < I) {
884
+ #else
885
+ int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/2) % I;
886
+ {
887
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
888
+ if (fallback) {
889
+ i = min(i, i_max);
890
+ }
891
+
892
+ const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
893
+
894
+ const int * scales = (const int *) bxi->scales;
895
+ const int ksc = threadIdx.x % 2;
896
+
897
+ const int sc32 = unpack_scales_q45_K(scales, ksc + 0);
898
+ const int m32 = unpack_scales_q45_K(scales, ksc + 2);
899
+
900
+ const uint8_t * sc8 = (const uint8_t *) &sc32;
901
+ const uint8_t * m8 = (const uint8_t *) &m32;
902
+
903
+ const half2 dm = bxi->dm * make_half2(1.0f, -1.0f);
904
+
905
+ #pragma unroll
906
+ for (int l = 0; l < int(sizeof(int)); ++l) {
907
+ x_dm[i*sram_stride + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]);
908
+ }
909
+ }
910
+ }
911
+ #else
912
+ #pragma unroll
913
+ for (int i0 = 0; i0 < I; i0 += nwarps*warp_size) {
914
+ int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % I;
915
+
916
+ if (fallback) {
917
+ i = min(i, i_max);
918
+ }
919
+
920
+ const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
921
+
922
+ x_dm[i] = bxi->dm;
923
+ }
924
+
925
+ constexpr int rows_per_warp = warp_size / 4;
926
+ #pragma unroll
927
+ for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) {
928
+ int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % I;
929
+
930
+ if (fallback) {
931
+ i = min(i, i_max);
932
+ }
933
+
934
+ const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
935
+
936
+ const int * scales = (const int *) bxi->scales;
937
+
938
+ const int ksc = threadIdx.x % (MMQ_TILE_NE_K/8);
939
+ const int scales8 = unpack_scales_q45_K(scales, ksc);
940
+
941
+ x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8;
942
+ }
943
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
944
+ }
945
+
946
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q6_K(
947
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
948
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
949
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
950
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
951
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
952
+
953
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
954
+ int * x_qs = (int *) x_tile;
955
+ float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
956
+ int * x_sc = (int *) (x_df + MMQ_TILE_NE_K/QI6_K);
957
+ #else
958
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, I);
959
+ int * x_qs = (int *) x_tile;
960
+ float * x_df = (float *) (x_qs + txs.qs);
961
+ int * x_sc = (int *) (x_df + txs.dm);
962
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
963
+
964
+ constexpr int threads_per_row = MMQ_ITER_K / (4 * QR6_K);
965
+ constexpr int nrows = warp_size / threads_per_row;
966
+ const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x;
967
+
968
+ #pragma unroll
969
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
970
+ int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row);
971
+
972
+ if (fallback) {
973
+ i = min(i, i_max);
974
+ }
975
+
976
+ const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride;
977
+
978
+ const int ql = get_int_b2(bxi->ql, txi);
979
+ const int ql0 = (ql >> 0) & 0x0F0F0F0F;
980
+ const int ql1 = (ql >> 4) & 0x0F0F0F0F;
981
+
982
+ const int qh = get_int_b2(bxi->qh, (QI6_K/4) * (txi / (QI6_K/2)) + txi % (QI6_K/4));
983
+ const int qh0 = ((qh >> ((txi & 0x08) >> 2)) << 4) & 0x30303030;
984
+ const int qh1 = (qh >> ((txi & 0x08) >> 2)) & 0x30303030;
985
+
986
+ const int kq0 = 2*txi - txi % (QI6_K/2) + 0;
987
+ const int kq1 = 2*txi - txi % (QI6_K/2) + QI6_K/2;
988
+
989
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
990
+ x_qs[i*sram_stride + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
991
+ x_qs[i*sram_stride + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
992
+ #else
993
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
994
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
995
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
996
+ }
997
+
998
+ #pragma unroll
999
+ for (int i0 = 0; i0 < I; i0 += nwarps*warp_size) {
1000
+ int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % I;
1001
+
1002
+ if (fallback) {
1003
+ i = min(i, i_max);
1004
+ }
1005
+
1006
+ const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride;
1007
+
1008
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1009
+ x_df[i*sram_stride] = bxi->d;
1010
+ #else
1011
+ x_df[i*(MMQ_TILE_NE_K/QI6_K) + i/QI6_K] = bxi->d;
1012
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1013
+ }
1014
+
1015
+ constexpr int rows_per_warp = warp_size / 4;
1016
+ #pragma unroll
1017
+ for (int i0 = 0; i0 < I; i0 += nwarps*rows_per_warp) {
1018
+ int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/(MMQ_TILE_NE_K/8)) % I;
1019
+
1020
+ if (fallback) {
1021
+ i = min(i, i_max);
1022
+ }
1023
+
1024
+ const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / 4;
1025
+
1026
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1027
+ x_sc[i*sram_stride + threadIdx.x%4] = get_int_b2(bxi->scales, threadIdx.x % (MMQ_TILE_NE_K/8));
1028
+ #else
1029
+ x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + threadIdx.x%(MMQ_TILE_NE_K/8)] = get_int_b2(bxi->scales, threadIdx.x%(QI6_K/8));
1030
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1031
+ }
1032
+ }
1033
+
1034
+ // ---------------------------------------------------------------------------------------------
1035
+
1036
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq1_s(
1037
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
1038
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
1039
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
1040
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
1041
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
1042
+
1043
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1044
+ int * x_qs = (int *) x_tile;
1045
+ half2 * x_ds = (half2 *) (x_qs + MMQ_TILE_NE_K*2);
1046
+ #else
1047
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, I);
1048
+ int * x_qs = (int *) x_tile;
1049
+ half2 * x_ds = (half2 *) (x_qs + txs.qs);
1050
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1051
+
1052
+ constexpr int threads_per_row = MMQ_ITER_K / (4 * QR1_S);
1053
+ constexpr int nrows = warp_size / threads_per_row;
1054
+ const int kqsx = threadIdx.x % threads_per_row;
1055
+
1056
+ #pragma unroll
1057
+ for (int i0 = 0; i0 < I; i0 += nwarps * nrows) {
1058
+ int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row;
1059
+
1060
+ if (fallback) {
1061
+ i = min(i, i_max);
1062
+ }
1063
+
1064
+ const block_iq1_s * bxi = (const block_iq1_s *) x + kbx0 + i*stride;
1065
+
1066
+ const int qs_packed = get_int_b2(bxi->qs, kqsx);
1067
+ const uint8_t * qs = (const uint8_t *) &qs_packed;
1068
+
1069
+ const int qh = bxi->qh[kqsx];
1070
+
1071
+ #pragma unroll
1072
+ for (int l = 0; l < QR1_S/2; ++l) {
1073
+ const int grid = iq1s_grid_gpu[qs[l] | (((qh >> (3*l)) & 0x07) << 8)];
1074
+
1075
+ const int grid0 = (grid >> 0) & 0x0F0F0F0F;
1076
+ const int grid1 = (grid >> 4) & 0x0F0F0F0F;
1077
+
1078
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1079
+ x_qs[i*sram_stride + 8*kqsx + (2*l+0)] = grid0;
1080
+ x_qs[i*sram_stride + 8*kqsx + (2*l+1)] = grid1;
1081
+ #else
1082
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid0;
1083
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid1;
1084
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1085
+ }
1086
+
1087
+ const float d1q = __half2float(bxi->d) * (((qh >> 11) & 0x0E) + 1);
1088
+ const float delta = -1.0f + IQ1S_DELTA - (qh & 0x8000) * (2.0f*IQ1S_DELTA/0x8000);
1089
+
1090
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1091
+ x_ds[i*sram_stride + kqsx] = make_half2(d1q, d1q*delta);
1092
+ #else
1093
+ x_ds[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = make_half2(d1q, d1q*delta);
1094
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1095
+ }
1096
+ }
1097
+
1098
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq2_xxs(
1099
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
1100
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
1101
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
1102
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
1103
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
1104
+
1105
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1106
+ int * x_qs = (int *) x_tile;
1107
+ float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
1108
+ #else
1109
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_XXS, I);
1110
+ int * x_qs = (int *) x_tile;
1111
+ float * x_df = (float *) (x_qs + txs.qs);
1112
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1113
+
1114
+ constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XXS)) / 2;
1115
+ constexpr int nrows = warp_size / threads_per_row;
1116
+ const int kqsx = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x;
1117
+
1118
+ #pragma unroll
1119
+ for (int i0 = 0; i0 < I; i0 += nwarps * nrows) {
1120
+ int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row;
1121
+
1122
+ if (fallback) {
1123
+ i = min(i, i_max);
1124
+ }
1125
+
1126
+ const block_iq2_xxs * bxi = (const block_iq2_xxs *) x + kbx0 + i*stride;
1127
+
1128
+ const int q2 = get_int_b2(bxi->qs, 2*kqsx+0);
1129
+ const uint8_t * aux8 = (const uint8_t *) &q2;
1130
+ const uint32_t aux32 = get_int_b2(bxi->qs, 2*kqsx+1);
1131
+
1132
+ #pragma unroll
1133
+ for (int l = 0; l < QR2_XXS; ++l) {
1134
+ const uint2 grid_pos = ((const uint2*)iq2xxs_grid)[aux8[l]];
1135
+ const uint32_t signs = unpack_ksigns(aux32 >> (7 * l));
1136
+
1137
+ const int signs0 = __vcmpne4(signs & 0x08040201, 0);
1138
+ const int grid0 = __vsub4(grid_pos.x ^ signs0, signs0);
1139
+
1140
+ const int signs1 = __vcmpne4(signs & 0x80402010, 0);
1141
+ const int grid1 = __vsub4(grid_pos.y ^ signs1, signs1);
1142
+
1143
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1144
+ x_qs[i*sram_stride + 8*kqsx + (2*l + 0)] = grid0;
1145
+ x_qs[i*sram_stride + 8*kqsx + (2*l + 1)] = grid1;
1146
+ #else
1147
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid0;
1148
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid1;
1149
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1150
+ }
1151
+
1152
+ const int ls = aux32 >> 27 | 1; // (scale * 2 + 1)
1153
+ const float d = bxi->d;
1154
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1155
+ x_df[i*sram_stride + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4
1156
+ #else
1157
+ x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = d * ls / 8; // (d * scale + d / 2) / 4
1158
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1159
+ }
1160
+ }
1161
+
1162
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq2_xs(
1163
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
1164
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
1165
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
1166
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
1167
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
1168
+
1169
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1170
+ int * x_qs = (int *) x_tile;
1171
+ float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
1172
+ #else
1173
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_XS, I);
1174
+ int * x_qs = (int *) x_tile;
1175
+ float * x_df = (float *) (x_qs + txs.qs);
1176
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1177
+
1178
+ constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XS)) / 2;
1179
+ constexpr int nrows = warp_size / threads_per_row;
1180
+ const int kqsx = threadIdx.x % threads_per_row;
1181
+
1182
+ #pragma unroll
1183
+ for (int i0 = 0; i0 < I; i0 += nwarps * nrows) {
1184
+ int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row;
1185
+
1186
+ if (fallback) {
1187
+ i = min(i, i_max);
1188
+ }
1189
+
1190
+ const block_iq2_xs * bxi = (const block_iq2_xs *) x + kbx0 + i*stride;
1191
+
1192
+ const int2 q2_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1));
1193
+ const uint16_t * q2 = (const uint16_t *) &q2_packed;
1194
+
1195
+ #pragma unroll
1196
+ for (int l = 0; l < QR2_XS; ++l) {
1197
+ const uint2 grid_pos = ((const uint2*)iq2xs_grid)[q2[l] & 0x1FF];
1198
+ const uint32_t signs = unpack_ksigns(q2[l] >> 9);
1199
+
1200
+ const int signs0 = __vcmpne4(signs & 0x08040201, 0);
1201
+ const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
1202
+
1203
+ const int signs1 = __vcmpne4(signs & 0x80402010, 0);
1204
+ const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
1205
+
1206
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1207
+ x_qs[i*sram_stride + 8*kqsx + (2*l + 0)] = grid_l;
1208
+ x_qs[i*sram_stride + 8*kqsx + (2*l + 1)] = grid_h;
1209
+ #else
1210
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l;
1211
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h;
1212
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1213
+ }
1214
+
1215
+ const int ls = bxi->scales[kqsx];
1216
+ const float d = bxi->d;
1217
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1218
+ x_df[i*sram_stride + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4;
1219
+ x_df[i*sram_stride + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4;
1220
+ #else
1221
+ x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4;
1222
+ x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4;
1223
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1224
+ }
1225
+ }
1226
+
1227
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq2_s(
1228
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
1229
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
1230
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
1231
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
1232
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
1233
+
1234
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1235
+ int * x_qs = (int *) x_tile;
1236
+ float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
1237
+ #else
1238
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_S, I);
1239
+ int * x_qs = (int *) x_tile;
1240
+ float * x_df = (float *) (x_qs + txs.qs);
1241
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1242
+ constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_S)) / 2;
1243
+ constexpr int nrows = warp_size / threads_per_row;
1244
+ const int kqsx = threadIdx.x % threads_per_row;
1245
+
1246
+ #pragma unroll
1247
+ for (int i0 = 0; i0 < I; i0 += nwarps * nrows) {
1248
+ int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row;
1249
+
1250
+ if (fallback) {
1251
+ i = min(i, i_max);
1252
+ }
1253
+
1254
+ const block_iq2_s * bxi = (const block_iq2_s *) x + kbx0 + i*stride;
1255
+
1256
+ const int qs_packed = get_int_b2(bxi->qs, kqsx);
1257
+ const uint8_t * qs = (const uint8_t *) &qs_packed;
1258
+
1259
+ const int qh = bxi->qh[kqsx];
1260
+
1261
+ const int signs_packed_32 = get_int_b2(bxi->qs, QK_K/32 + kqsx);
1262
+ const uint8_t * signs_packed_8 = (const uint8_t *) &signs_packed_32;
1263
+
1264
+ #pragma unroll
1265
+ for (int l = 0; l < QR2_S; ++l) {
1266
+ const int * grid_pos = (const int *)(iq2s_grid + (qs[l] | ((qh << (8-2*l)) & 0x300)));
1267
+
1268
+ const int signs0 = __vcmpne4(((signs_packed_8[l] & 0x03) << 7) | ((signs_packed_8[l] & 0x0C) << 21), 0x00000000);
1269
+ const int signs1 = __vcmpne4(((signs_packed_8[l] & 0x30) << 3) | ((signs_packed_8[l] & 0xC0) << 17), 0x00000000);
1270
+
1271
+ const int grid_l = __vsub4(grid_pos[0] ^ signs0, signs0);
1272
+ const int grid_h = __vsub4(grid_pos[1] ^ signs1, signs1);
1273
+
1274
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1275
+ x_qs[i*sram_stride + 8*kqsx + (2*l + 0)] = grid_l;
1276
+ x_qs[i*sram_stride + 8*kqsx + (2*l + 1)] = grid_h;
1277
+ #else
1278
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l;
1279
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h;
1280
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1281
+ }
1282
+
1283
+ const int ls = bxi->scales[kqsx];
1284
+ const float d = bxi->d;
1285
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1286
+ x_df[i*sram_stride + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4;
1287
+ x_df[i*sram_stride + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4;
1288
+ #else
1289
+ x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4;
1290
+ x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4;
1291
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1292
+ }
1293
+ }
1294
+
1295
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq3_xxs(
1296
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
1297
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
1298
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
1299
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
1300
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
1301
+
1302
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1303
+ int * x_qs = (int *) x_tile;
1304
+ float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
1305
+ #else
1306
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_XXS, I);
1307
+ int * x_qs = (int *) x_tile;
1308
+ float * x_df = (float *) (x_qs + txs.qs);
1309
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1310
+
1311
+ constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_XXS)) / 2;
1312
+ constexpr int nrows = warp_size / threads_per_row;
1313
+ const int kqsx = threadIdx.x % threads_per_row;
1314
+
1315
+ #pragma unroll
1316
+ for (int i0 = 0; i0 < I; i0 += nwarps * nrows) {
1317
+ int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row;
1318
+
1319
+ if (fallback) {
1320
+ i = min(i, i_max);
1321
+ }
1322
+
1323
+ const block_iq3_xxs * bxi = (const block_iq3_xxs *) x + kbx0 + i*stride;
1324
+
1325
+ const int2 q3_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1));
1326
+ const uint8_t * q3 = (const uint8_t *) &q3_packed;
1327
+ const uint32_t aux32 = get_int_b2(bxi->qs, QK_K/16 + kqsx);
1328
+
1329
+ #pragma unroll
1330
+ for (int l = 0; l < QR3_XXS; ++l) {
1331
+ const int2 grid_pos = make_int2(iq3xxs_grid[q3[2*l+0]], iq3xxs_grid[q3[2*l+1]]);
1332
+ const uint32_t signs = unpack_ksigns(aux32 >> (7*l));
1333
+
1334
+ const int signs0 = __vcmpne4(signs & 0x08040201, 0);
1335
+ const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
1336
+
1337
+ const int signs1 = __vcmpne4(signs & 0x80402010, 0);
1338
+ const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
1339
+
1340
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1341
+ x_qs[i*sram_stride + 8*kqsx + (2*l + 0)] = grid_l;
1342
+ x_qs[i*sram_stride + 8*kqsx + (2*l + 1)] = grid_h;
1343
+ #else
1344
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l;
1345
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h;
1346
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1347
+ }
1348
+
1349
+ const int ls = aux32 >> 28;
1350
+ const float d = bxi->d;
1351
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1352
+ x_df[i*sram_stride + kqsx] = (ls*d + d/2)/2;
1353
+ #else
1354
+ x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = (ls*d + d/2)/2;
1355
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1356
+ }
1357
+ }
1358
+
1359
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq3_s(
1360
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
1361
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
1362
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
1363
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
1364
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
1365
+
1366
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1367
+ int * x_qs = (int *) x_tile;
1368
+ float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
1369
+ #else
1370
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, I);
1371
+ int * x_qs = (int *) x_tile;
1372
+ float * x_df = (float *) (x_qs + txs.qs);
1373
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1374
+
1375
+ constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_S)) / 2;
1376
+ constexpr int nrows = warp_size / threads_per_row;
1377
+ const int kqsx = threadIdx.x % threads_per_row;
1378
+
1379
+ #pragma unroll
1380
+ for (int i0 = 0; i0 < I; i0 += nwarps * nrows) {
1381
+ int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row;
1382
+
1383
+ if (fallback) {
1384
+ i = min(i, i_max);
1385
+ }
1386
+
1387
+ const block_iq3_s * bxi = (const block_iq3_s *) x + kbx0 + i*stride;
1388
+
1389
+ const int2 qs_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1));
1390
+ const uint8_t * qs = (const uint8_t *) &qs_packed;
1391
+
1392
+ const int qh = bxi->qh[kqsx];
1393
+
1394
+ const int signs_packed_32 = get_int_b2(bxi->signs, kqsx);
1395
+ const uint8_t * signs_packed_8 = (const uint8_t *) &signs_packed_32;
1396
+
1397
+ #pragma unroll
1398
+ for (int l = 0; l < QR3_S; ++l) {
1399
+ const int2 grid_pos = make_int2(
1400
+ iq3s_grid[qs[2*l+0] | ((qh << (8 - 2*l)) & 0x100)],
1401
+ iq3s_grid[qs[2*l+1] | ((qh << (7 - 2*l)) & 0x100)]);
1402
+
1403
+ const int signs0 = __vcmpne4(((signs_packed_8[l] & 0x03) << 7) | ((signs_packed_8[l] & 0x0C) << 21), 0x00000000);
1404
+ const int signs1 = __vcmpne4(((signs_packed_8[l] & 0x30) << 3) | ((signs_packed_8[l] & 0xC0) << 17), 0x00000000);
1405
+
1406
+ const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
1407
+ const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
1408
+
1409
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1410
+ x_qs[i*sram_stride + 8*kqsx + (2*l+0)] = grid_l;
1411
+ x_qs[i*sram_stride + 8*kqsx + (2*l+1)] = grid_h;
1412
+ #else
1413
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid_l;
1414
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid_h;
1415
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1416
+ }
1417
+
1418
+ const int ls = 1 + 2*((bxi->scales[kqsx/2] >> (((2*kqsx) << 1) & 0x04)) & 0x0F);
1419
+ const float d = bxi->d;
1420
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1421
+ x_df[i*sram_stride + kqsx] = ls*d;
1422
+ #else
1423
+ x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = ls*d;
1424
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1425
+ }
1426
+ }
1427
+
1428
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq4_xs(
1429
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
1430
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
1431
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
1432
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
1433
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
1434
+
1435
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1436
+ int * x_qs = (int *) x_tile;
1437
+ float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
1438
+ #else
1439
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_XS, I);
1440
+ int * x_qs = (int *) x_tile;
1441
+ float * x_df = (float *) (x_qs + txs.qs);
1442
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1443
+
1444
+ constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_XS);
1445
+ constexpr int nrows = warp_size / threads_per_row;
1446
+ const int kqsx = threadIdx.x % threads_per_row;
1447
+
1448
+ #pragma unroll
1449
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
1450
+ int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row);
1451
+
1452
+ if (fallback) {
1453
+ i = min(i, i_max);
1454
+ }
1455
+
1456
+ const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride;
1457
+
1458
+ const int aux_q4 = get_int_b4(bxi->qs, kqsx);
1459
+ const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl);
1460
+ const int k0 = 8 * (kqsx / 4) + kqsx % 4;
1461
+
1462
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1463
+ x_qs[i*sram_stride + k0 + 0] = v.x;
1464
+ x_qs[i*sram_stride + k0 + 4] = v.y;
1465
+ #else
1466
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x;
1467
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 4] = v.y;
1468
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1469
+ }
1470
+
1471
+ constexpr int rows_per_warp = warp_size / 8;
1472
+ #pragma unroll
1473
+ for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) {
1474
+ int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / (MMQ_TILE_NE_K/4);
1475
+
1476
+ if (fallback) {
1477
+ i = min(i, i_max);
1478
+ }
1479
+
1480
+ const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride;
1481
+
1482
+ const float d = __half2float(bxi->d);
1483
+
1484
+ const int ls = ((bxi->scales_l[(threadIdx.x % 8)/2] >> (4*(threadIdx.x % 2))) & 0x0F)
1485
+ | (((bxi->scales_h >> (2*(threadIdx.x % 8))) & 0x03) << 4);
1486
+
1487
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1488
+ x_df[i*sram_stride + threadIdx.x % 8] = d * (ls - 32);
1489
+ #else
1490
+ x_df[i*(MMQ_TILE_NE_K/4) + i/4 + threadIdx.x % 8] = d * (ls - 32);
1491
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1492
+ }
1493
+ }
1494
+
1495
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq4_nl(
1496
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
1497
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
1498
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
1499
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
1500
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
1501
+
1502
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1503
+ int * x_qs = (int *) x_tile;
1504
+ float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
1505
+ #else
1506
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_NL, I);
1507
+ int * x_qs = (int *) x_tile;
1508
+ float * x_df = (float *) (x_qs + txs.qs);
1509
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1510
+
1511
+ constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_NL);
1512
+ constexpr int nrows = warp_size / threads_per_row;
1513
+ const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x;
1514
+ const int kbx = txi / QI4_NL;
1515
+ const int kqsx = txi % QI4_NL;
1516
+
1517
+ #pragma unroll
1518
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
1519
+ int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row);
1520
+
1521
+ if (fallback) {
1522
+ i = min(i, i_max);
1523
+ }
1524
+
1525
+ const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbx;
1526
+
1527
+ const int aux_q4 = get_int_b2(bxi->qs, kqsx);
1528
+ const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl);
1529
+ const int k0 = kbx * (2 * QI4_NL) + kqsx;
1530
+
1531
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1532
+ x_qs[i*sram_stride + k0 + 0] = v.x;
1533
+ x_qs[i*sram_stride + k0 + QI4_NL] = v.y;
1534
+ #else
1535
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x;
1536
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI4_NL] = v.y;
1537
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1538
+ }
1539
+
1540
+ constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_NL;
1541
+ constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row;
1542
+ const int kbxd = threadIdx.x % blocks_per_tile_x_row;
1543
+
1544
+ #pragma unroll
1545
+ for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) {
1546
+ int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row;
1547
+
1548
+ if (fallback) {
1549
+ i = min(i, i_max);
1550
+ }
1551
+
1552
+ const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbxd;
1553
+
1554
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1555
+ x_df[i*sram_stride + kbxd] = __half2float(bxi->d);
1556
+ #else
1557
+ x_df[i*(MMQ_TILE_NE_K/QI4_NL) + i/QI4_NL + kbxd] = __half2float(bxi->d);
1558
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1559
+ }
1560
+ }
1561
+
1562
+ // ---------------------------------------------------------------------------------------------
1563
+
1564
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_mxfp4(
1565
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
1566
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
1567
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
1568
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
1569
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
1570
+
1571
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1572
+ int * x_qs = (int *) x_tile;
1573
+ float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
1574
+ #else
1575
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_MXFP4, I);
1576
+ int * x_qs = (int *) x_tile;
1577
+ float * x_df = (float *) (x_qs + txs.qs);
1578
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1579
+
1580
+ constexpr int threads_per_row = MMQ_ITER_K / (4 * QR_MXFP4);
1581
+ constexpr int nrows = warp_size / threads_per_row;
1582
+ const int txi = warp_size > threads_per_row ? threadIdx.x % threads_per_row : threadIdx.x;
1583
+ const int kbx = txi / QI_MXFP4;
1584
+ const int kqsx = txi % QI_MXFP4;
1585
+
1586
+ #pragma unroll
1587
+ for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
1588
+ int i = i0 + (nrows == 1 ? threadIdx.y : threadIdx.y*nrows + threadIdx.x/threads_per_row);
1589
+
1590
+ if (fallback) {
1591
+ i = min(i, i_max);
1592
+ }
1593
+
1594
+ const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbx;
1595
+
1596
+ const int aux_q4 = get_int_b1(bxi->qs, kqsx);
1597
+ const int2 v = get_int_from_table_16(aux_q4, kvalues_mxfp4);
1598
+ const int k0 = kbx * (2 * QI_MXFP4) + kqsx;
1599
+
1600
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1601
+ x_qs[i*sram_stride + k0 + 0] = v.x;
1602
+ x_qs[i*sram_stride + k0 + QI_MXFP4] = v.y;
1603
+ #else
1604
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x;
1605
+ x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI_MXFP4] = v.y;
1606
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1607
+ }
1608
+
1609
+ constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI_MXFP4;
1610
+ constexpr int rows_per_warp = warp_size / blocks_per_tile_x_row;
1611
+ const int kbxd = threadIdx.x % blocks_per_tile_x_row;
1612
+
1613
+ #pragma unroll
1614
+ for (int i0 = 0; i0 < I; i0 += nwarps * rows_per_warp) {
1615
+ int i = i0 + threadIdx.y * rows_per_warp + threadIdx.x / blocks_per_tile_x_row;
1616
+
1617
+ if (fallback) {
1618
+ i = min(i, i_max);
1619
+ }
1620
+
1621
+ const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbxd;
1622
+
1623
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1624
+ x_df[i*sram_stride + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f;
1625
+ #else
1626
+ x_df[i*(MMQ_TILE_NE_K/QI_MXFP4) + i/QI_MXFP4 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f;
1627
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1628
+ }
1629
+ }
1630
+
1631
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_mxfp4_fp4(
1632
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
1633
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
1634
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
1635
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
1636
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
1637
+
1638
+ int * x_qs = (int *) x_tile;
1639
+ uint32_t * x_sc = (uint32_t *) (x_qs + 2 * MMQ_TILE_NE_K);
1640
+
1641
+ const int txi = threadIdx.x;
1642
+
1643
+ constexpr int iter_k = ggml_cuda_mmq_get_K_vram(type, J, fallback);
1644
+
1645
+ constexpr int threads_per_row = iter_k / QK_MXFP4; // each thread processes 1 block
1646
+ constexpr int rows_per_warp = warp_size / threads_per_row;
1647
+ const int kbx = txi % threads_per_row;
1648
+ const int row_in_warp = txi / threads_per_row;
1649
+
1650
+ #pragma unroll
1651
+ for (int i0 = 0; i0 < I; i0 += rows_per_warp * nwarps) {
1652
+ int i = i0 + threadIdx.y * rows_per_warp + row_in_warp;
1653
+
1654
+ if constexpr (fallback) {
1655
+ i = min(i, i_max);
1656
+ }
1657
+
1658
+ const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i * stride + kbx;
1659
+
1660
+ // quantize_mxfp4_mmq permutes nibbles to match the quantized format
1661
+ const int k0 = kbx * 4;
1662
+ memcpy(x_qs + i*sram_stride + k0, bxi->qs, 16);
1663
+
1664
+ // Load E8M0 scales: pack 2 consecutive scales into one uint32
1665
+ if (kbx % 2 == 0) {
1666
+ uint32_t e = bxi->e;
1667
+ e |= ((bxi + 1)->e << 8);
1668
+ x_sc[i*sram_stride + kbx / 2] = e;
1669
+ }
1670
+ }
1671
+ }
1672
+
1673
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_nvfp4(
1674
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kb0, const int i_max, const int stride) {
1675
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
1676
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
1677
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
1678
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
1679
+
1680
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1681
+ int * x_qs = (int *) x_tile;
1682
+ float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
1683
+ #else
1684
+ constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_NVFP4, I);
1685
+ int * x_qs = (int *) x_tile;
1686
+ float * x_df = (float *) (x_qs + txs.qs);
1687
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1688
+
1689
+ constexpr int threads_per_row = MMQ_ITER_K / QK_NVFP4;
1690
+ constexpr int rows_per_warp = warp_size / threads_per_row;
1691
+ const int kbx = threadIdx.x % threads_per_row;
1692
+ const int row_in_warp = threadIdx.x / threads_per_row;
1693
+
1694
+ #pragma unroll
1695
+ for (int i0 = 0; i0 < I; i0 += rows_per_warp * nwarps) {
1696
+ int i = i0 + threadIdx.y * rows_per_warp + row_in_warp;
1697
+
1698
+ if constexpr (fallback) {
1699
+ i = min(i, i_max);
1700
+ }
1701
+
1702
+ const block_nvfp4 * bxi = (const block_nvfp4 *) x + kb0 + i * stride + kbx;
1703
+ const uint32_t * __restrict__ src_qs = reinterpret_cast<const uint32_t *>(bxi->qs);
1704
+ const int kqs = 16 * kbx;
1705
+ const int ksc = 4 * kbx;
1706
+
1707
+ #pragma unroll
1708
+ for (int sub = 0; sub < QK_NVFP4 / QK_NVFP4_SUB; ++sub) {
1709
+ const int2 q0 = get_int_from_table_16(src_qs[2 * sub + 0], kvalues_mxfp4);
1710
+ const int2 q1 = get_int_from_table_16(src_qs[2 * sub + 1], kvalues_mxfp4);
1711
+
1712
+ #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1713
+ x_qs[i*sram_stride + kqs + 4 * sub + 0] = q0.x;
1714
+ x_qs[i*sram_stride + kqs + 4 * sub + 1] = q1.x;
1715
+ x_qs[i*sram_stride + kqs + 4 * sub + 2] = q0.y;
1716
+ x_qs[i*sram_stride + kqs + 4 * sub + 3] = q1.y;
1717
+ x_df[i*sram_stride + ksc + sub] = ggml_cuda_ue4m3_to_fp32(bxi->d[sub]);
1718
+ #else
1719
+ x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 0] = q0.x;
1720
+ x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 1] = q1.x;
1721
+ x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 2] = q0.y;
1722
+ x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 3] = q1.y;
1723
+ x_df[i * (2 * MMQ_TILE_NE_K * 2 / QI_NVFP4) + i / (QK_NVFP4_SUB / QI_NVFP4) + ksc + sub] = ggml_cuda_ue4m3_to_fp32(bxi->d[sub]);
1724
+ #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
1725
+ }
1726
+ }
1727
+ }
1728
+
1729
+ template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_nvfp4_nvfp4(
1730
+ const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
1731
+ constexpr int warp_size = ggml_cuda_get_physical_warp_size();
1732
+ constexpr int nwarps = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
1733
+ constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
1734
+ constexpr int iter_k = ggml_cuda_mmq_get_K_vram(type, J, fallback);
1735
+ constexpr int threads_per_row = iter_k / QK_NVFP4; // each thread processes 1 block
1736
+ constexpr int rows_per_warp = warp_size / threads_per_row;
1737
+ constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
1738
+
1739
+ uint32_t * x_u32 = (uint32_t *) x_tile;
1740
+
1741
+ const int txi = threadIdx.x;
1742
+ const int kbx = txi % threads_per_row;
1743
+ const int row_in_warp = txi / threads_per_row;
1744
+
1745
+ const block_nvfp4 * bxi_base = (const block_nvfp4 *) x + kbx0 + kbx;
1746
+ uint32_t * x_u32_scale = x_u32 + 64 + kbx;
1747
+
1748
+ #pragma unroll
1749
+ for (int i0 = 0; i0 < I; i0 += rows_per_warp * nwarps) {
1750
+ int i = i0 + threadIdx.y * rows_per_warp + row_in_warp;
1751
+
1752
+ if constexpr (fallback) {
1753
+ i = min(i, i_max);
1754
+ }
1755
+
1756
+ const block_nvfp4 * bxi = bxi_base + i * stride;
1757
+
1758
+ const uint32_t * src_qs = reinterpret_cast<const uint32_t *>(bxi->qs);
1759
+
1760
+ #pragma unroll
1761
+ for (int sub = 0; sub < QK_NVFP4 / QK_NVFP4_SUB; ++sub) {
1762
+ x_u32[i*sram_stride + 8*kbx + 2 * sub + 0] = src_qs[2 * sub + 0];
1763
+ x_u32[i*sram_stride + 8*kbx + 2 * sub + 1] = src_qs[2 * sub + 1];
1764
+ }
1765
+
1766
+ x_u32_scale[i*sram_stride] = get_int_b4(bxi->d, 0);
1767
+ }
1768
+ }