whispercpp 1.3.4 → 1.3.5

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 (630) hide show
  1. checksums.yaml +4 -4
  2. data/README.md +60 -43
  3. data/ext/extconf.rb +2 -2
  4. data/ext/ruby_whisper.c +14 -2
  5. data/ext/ruby_whisper.h +39 -0
  6. data/ext/ruby_whisper_context.c +22 -22
  7. data/ext/ruby_whisper_model.c +12 -12
  8. data/ext/ruby_whisper_params.c +47 -23
  9. data/ext/ruby_whisper_segment.c +84 -19
  10. data/ext/ruby_whisper_token.c +351 -0
  11. data/ext/ruby_whisper_transcribe.cpp +1 -1
  12. data/ext/ruby_whisper_vad_context.c +75 -0
  13. data/ext/ruby_whisper_vad_context_detect.cpp +50 -0
  14. data/ext/ruby_whisper_vad_segment.c +139 -0
  15. data/ext/ruby_whisper_vad_segments.c +106 -0
  16. data/ext/sources/CMakeLists.txt +4 -1
  17. data/ext/sources/bindings/javascript/package.json +1 -1
  18. data/ext/sources/cmake/arm64-apple-clang.cmake +16 -0
  19. data/ext/sources/cmake/arm64-windows-llvm.cmake +16 -0
  20. data/ext/sources/cmake/riscv64-spacemit-linux-gnu-gcc.cmake +29 -0
  21. data/ext/sources/cmake/x64-windows-llvm.cmake +5 -0
  22. data/ext/sources/examples/addon.node/vad-example.js +2 -2
  23. data/ext/sources/examples/cli/cli.cpp +121 -112
  24. data/ext/sources/examples/lsp/CMakeLists.txt +2 -1
  25. data/ext/sources/examples/quantize/CMakeLists.txt +2 -1
  26. data/ext/sources/examples/server/server.cpp +10 -11
  27. data/ext/sources/examples/talk-llama/CMakeLists.txt +5 -1
  28. data/ext/sources/examples/talk-llama/llama-adapter.cpp +12 -3
  29. data/ext/sources/examples/talk-llama/llama-adapter.h +7 -1
  30. data/ext/sources/examples/talk-llama/llama-arch.cpp +2046 -1974
  31. data/ext/sources/examples/talk-llama/llama-arch.h +67 -2
  32. data/ext/sources/examples/talk-llama/llama-batch.cpp +75 -33
  33. data/ext/sources/examples/talk-llama/llama-batch.h +17 -4
  34. data/ext/sources/examples/talk-llama/llama-chat.cpp +79 -3
  35. data/ext/sources/examples/talk-llama/llama-chat.h +4 -0
  36. data/ext/sources/examples/talk-llama/llama-context.cpp +775 -78
  37. data/ext/sources/examples/talk-llama/llama-context.h +57 -9
  38. data/ext/sources/examples/talk-llama/llama-cparams.h +1 -0
  39. data/ext/sources/examples/talk-llama/llama-grammar.cpp +288 -53
  40. data/ext/sources/examples/talk-llama/llama-grammar.h +22 -1
  41. data/ext/sources/examples/talk-llama/llama-graph.cpp +381 -64
  42. data/ext/sources/examples/talk-llama/llama-graph.h +103 -13
  43. data/ext/sources/examples/talk-llama/llama-hparams.cpp +26 -2
  44. data/ext/sources/examples/talk-llama/llama-hparams.h +41 -10
  45. data/ext/sources/examples/talk-llama/llama-impl.cpp +7 -3
  46. data/ext/sources/examples/talk-llama/llama-impl.h +1 -1
  47. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +5 -3
  48. data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +145 -65
  49. data/ext/sources/examples/talk-llama/llama-kv-cache.h +22 -7
  50. data/ext/sources/examples/talk-llama/llama-kv-cells.h +44 -2
  51. data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +12 -10
  52. data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +32 -19
  53. data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +2 -2
  54. data/ext/sources/examples/talk-llama/llama-mmap.cpp +172 -37
  55. data/ext/sources/examples/talk-llama/llama-mmap.h +8 -3
  56. data/ext/sources/examples/talk-llama/llama-model-loader.cpp +91 -9
  57. data/ext/sources/examples/talk-llama/llama-model-loader.h +6 -0
  58. data/ext/sources/examples/talk-llama/llama-model-saver.cpp +3 -0
  59. data/ext/sources/examples/talk-llama/llama-model.cpp +1529 -13134
  60. data/ext/sources/examples/talk-llama/llama-model.h +44 -3
  61. data/ext/sources/examples/talk-llama/llama-quant.cpp +8 -23
  62. data/ext/sources/examples/talk-llama/llama-sampling.cpp +1294 -198
  63. data/ext/sources/examples/talk-llama/llama-sampling.h +19 -7
  64. data/ext/sources/examples/talk-llama/llama-vocab.cpp +133 -37
  65. data/ext/sources/examples/talk-llama/llama-vocab.h +45 -40
  66. data/ext/sources/examples/talk-llama/llama.cpp +729 -2
  67. data/ext/sources/examples/talk-llama/llama.h +152 -14
  68. data/ext/sources/examples/talk-llama/models/afmoe.cpp +191 -0
  69. data/ext/sources/examples/talk-llama/models/apertus.cpp +125 -0
  70. data/ext/sources/examples/talk-llama/models/arcee.cpp +135 -0
  71. data/ext/sources/examples/talk-llama/models/arctic.cpp +138 -0
  72. data/ext/sources/examples/talk-llama/models/arwkv7.cpp +86 -0
  73. data/ext/sources/examples/talk-llama/models/baichuan.cpp +122 -0
  74. data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +144 -0
  75. data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +135 -0
  76. data/ext/sources/examples/talk-llama/models/bert.cpp +178 -0
  77. data/ext/sources/examples/talk-llama/models/bitnet.cpp +160 -0
  78. data/ext/sources/examples/talk-llama/models/bloom.cpp +101 -0
  79. data/ext/sources/examples/talk-llama/models/chameleon.cpp +178 -0
  80. data/ext/sources/examples/talk-llama/models/chatglm.cpp +132 -0
  81. data/ext/sources/examples/talk-llama/models/codeshell.cpp +111 -0
  82. data/ext/sources/examples/talk-llama/models/cogvlm.cpp +102 -0
  83. data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +134 -0
  84. data/ext/sources/examples/talk-llama/models/command-r.cpp +122 -0
  85. data/ext/sources/examples/talk-llama/models/dbrx.cpp +123 -0
  86. data/ext/sources/examples/talk-llama/models/deci.cpp +135 -0
  87. data/ext/sources/examples/talk-llama/models/deepseek.cpp +144 -0
  88. data/ext/sources/examples/talk-llama/models/deepseek2.cpp +259 -0
  89. data/ext/sources/examples/talk-llama/models/dots1.cpp +134 -0
  90. data/ext/sources/examples/talk-llama/models/dream.cpp +105 -0
  91. data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +150 -0
  92. data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +110 -0
  93. data/ext/sources/examples/talk-llama/models/exaone.cpp +114 -0
  94. data/ext/sources/examples/talk-llama/models/exaone4.cpp +123 -0
  95. data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +113 -0
  96. data/ext/sources/examples/talk-llama/models/falcon.cpp +120 -0
  97. data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +116 -0
  98. data/ext/sources/examples/talk-llama/models/gemma.cpp +112 -0
  99. data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +128 -0
  100. data/ext/sources/examples/talk-llama/models/gemma3.cpp +155 -0
  101. data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +384 -0
  102. data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +170 -0
  103. data/ext/sources/examples/talk-llama/models/glm4.cpp +150 -0
  104. data/ext/sources/examples/talk-llama/models/gpt2.cpp +105 -0
  105. data/ext/sources/examples/talk-llama/models/gptneox.cpp +144 -0
  106. data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +196 -0
  107. data/ext/sources/examples/talk-llama/models/granite.cpp +211 -0
  108. data/ext/sources/examples/talk-llama/models/graph-context-mamba.cpp +283 -0
  109. data/ext/sources/examples/talk-llama/models/grok.cpp +159 -0
  110. data/ext/sources/examples/talk-llama/models/grovemoe.cpp +141 -0
  111. data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +132 -0
  112. data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +154 -0
  113. data/ext/sources/examples/talk-llama/models/internlm2.cpp +120 -0
  114. data/ext/sources/examples/talk-llama/models/jais.cpp +86 -0
  115. data/ext/sources/examples/talk-llama/models/jamba.cpp +106 -0
  116. data/ext/sources/examples/talk-llama/models/lfm2.cpp +175 -0
  117. data/ext/sources/examples/talk-llama/models/llada-moe.cpp +122 -0
  118. data/ext/sources/examples/talk-llama/models/llada.cpp +99 -0
  119. data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +178 -0
  120. data/ext/sources/examples/talk-llama/models/llama.cpp +168 -0
  121. data/ext/sources/examples/talk-llama/models/maincoder.cpp +117 -0
  122. data/ext/sources/examples/talk-llama/models/mamba.cpp +55 -0
  123. data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +123 -0
  124. data/ext/sources/examples/talk-llama/models/minicpm3.cpp +199 -0
  125. data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +124 -0
  126. data/ext/sources/examples/talk-llama/models/mistral3.cpp +160 -0
  127. data/ext/sources/examples/talk-llama/models/models.h +569 -0
  128. data/ext/sources/examples/talk-llama/models/modern-bert.cpp +116 -0
  129. data/ext/sources/examples/talk-llama/models/mpt.cpp +126 -0
  130. data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +150 -0
  131. data/ext/sources/examples/talk-llama/models/nemotron.cpp +122 -0
  132. data/ext/sources/examples/talk-llama/models/neo-bert.cpp +104 -0
  133. data/ext/sources/examples/talk-llama/models/olmo.cpp +121 -0
  134. data/ext/sources/examples/talk-llama/models/olmo2.cpp +150 -0
  135. data/ext/sources/examples/talk-llama/models/olmoe.cpp +124 -0
  136. data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +127 -0
  137. data/ext/sources/examples/talk-llama/models/openelm.cpp +124 -0
  138. data/ext/sources/examples/talk-llama/models/orion.cpp +123 -0
  139. data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +121 -0
  140. data/ext/sources/examples/talk-llama/models/phi2.cpp +121 -0
  141. data/ext/sources/examples/talk-llama/models/phi3.cpp +152 -0
  142. data/ext/sources/examples/talk-llama/models/plamo.cpp +110 -0
  143. data/ext/sources/examples/talk-llama/models/plamo2.cpp +316 -0
  144. data/ext/sources/examples/talk-llama/models/plamo3.cpp +128 -0
  145. data/ext/sources/examples/talk-llama/models/plm.cpp +168 -0
  146. data/ext/sources/examples/talk-llama/models/qwen.cpp +108 -0
  147. data/ext/sources/examples/talk-llama/models/qwen2.cpp +126 -0
  148. data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +151 -0
  149. data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +117 -0
  150. data/ext/sources/examples/talk-llama/models/qwen3.cpp +117 -0
  151. data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +124 -0
  152. data/ext/sources/examples/talk-llama/models/qwen3next.cpp +873 -0
  153. data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +149 -0
  154. data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +141 -0
  155. data/ext/sources/examples/talk-llama/models/refact.cpp +94 -0
  156. data/ext/sources/examples/talk-llama/models/rnd1.cpp +126 -0
  157. data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +162 -0
  158. data/ext/sources/examples/talk-llama/models/rwkv6.cpp +94 -0
  159. data/ext/sources/examples/talk-llama/models/rwkv6qwen2.cpp +86 -0
  160. data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +135 -0
  161. data/ext/sources/examples/talk-llama/models/rwkv7.cpp +90 -0
  162. data/ext/sources/examples/talk-llama/models/seed-oss.cpp +124 -0
  163. data/ext/sources/examples/talk-llama/models/smallthinker.cpp +126 -0
  164. data/ext/sources/examples/talk-llama/models/smollm3.cpp +128 -0
  165. data/ext/sources/examples/talk-llama/models/stablelm.cpp +146 -0
  166. data/ext/sources/examples/talk-llama/models/starcoder.cpp +100 -0
  167. data/ext/sources/examples/talk-llama/models/starcoder2.cpp +121 -0
  168. data/ext/sources/examples/talk-llama/models/t5-dec.cpp +166 -0
  169. data/ext/sources/examples/talk-llama/models/t5-enc.cpp +96 -0
  170. data/ext/sources/examples/talk-llama/models/wavtokenizer-dec.cpp +149 -0
  171. data/ext/sources/examples/talk-llama/models/xverse.cpp +108 -0
  172. data/ext/sources/examples/talk-llama/unicode.cpp +102 -16
  173. data/ext/sources/examples/vad-speech-segments/CMakeLists.txt +1 -1
  174. data/ext/sources/examples/whisper.wasm/index-tmpl.html +1 -1
  175. data/ext/sources/ggml/CMakeLists.txt +82 -54
  176. data/ext/sources/ggml/include/ggml-alloc.h +9 -0
  177. data/ext/sources/ggml/include/ggml-backend.h +4 -1
  178. data/ext/sources/ggml/include/ggml-cpu.h +1 -0
  179. data/ext/sources/ggml/include/ggml-hexagon.h +19 -0
  180. data/ext/sources/ggml/include/ggml-rpc.h +8 -11
  181. data/ext/sources/ggml/include/ggml-zendnn.h +22 -0
  182. data/ext/sources/ggml/include/ggml.h +190 -12
  183. data/ext/sources/ggml/src/CMakeLists.txt +82 -11
  184. data/ext/sources/ggml/src/ggml-alloc.c +124 -41
  185. data/ext/sources/ggml/src/ggml-backend-impl.h +1 -4
  186. data/ext/sources/ggml/src/ggml-backend-reg.cpp +27 -3
  187. data/ext/sources/ggml/src/ggml-backend.cpp +71 -21
  188. data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +17 -3
  189. data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +5 -9
  190. data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +57 -45
  191. data/ext/sources/ggml/src/ggml-cann/acl_tensor.h +138 -47
  192. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +2179 -1696
  193. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +238 -317
  194. data/ext/sources/ggml/src/ggml-cann/common.h +283 -208
  195. data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +626 -776
  196. data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +156 -86
  197. data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +1 -0
  198. data/ext/sources/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp +4 -0
  199. data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +428 -26
  200. data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +1004 -0
  201. data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +4 -5
  202. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/cpu-feats.cpp +38 -0
  203. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +108 -49
  204. data/ext/sources/ggml/src/ggml-cpu/arch/s390/cpu-feats.cpp +50 -0
  205. data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +6 -6
  206. data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +50 -2
  207. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +5 -3
  208. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +195 -71
  209. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +4 -0
  210. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +573 -106
  211. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.h +33 -44
  212. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +298 -112
  213. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm-ppc.h +333 -0
  214. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +819 -125
  215. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.h +6 -0
  216. data/ext/sources/ggml/src/ggml-cpu/ops.cpp +708 -431
  217. data/ext/sources/ggml/src/ggml-cpu/ops.h +5 -4
  218. data/ext/sources/ggml/src/ggml-cpu/repack.cpp +671 -31
  219. data/ext/sources/ggml/src/ggml-cpu/repack.h +14 -0
  220. data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +41 -43
  221. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +3 -2
  222. data/ext/sources/ggml/src/ggml-cpu/unary-ops.cpp +151 -0
  223. data/ext/sources/ggml/src/ggml-cpu/unary-ops.h +7 -0
  224. data/ext/sources/ggml/src/ggml-cpu/vec.cpp +124 -1
  225. data/ext/sources/ggml/src/ggml-cpu/vec.h +261 -146
  226. data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +72 -1
  227. data/ext/sources/ggml/src/ggml-cuda/argmax.cu +2 -2
  228. data/ext/sources/ggml/src/ggml-cuda/argsort.cu +123 -6
  229. data/ext/sources/ggml/src/ggml-cuda/argsort.cuh +16 -0
  230. data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +1 -1
  231. data/ext/sources/ggml/src/ggml-cuda/common.cuh +353 -80
  232. data/ext/sources/ggml/src/ggml-cuda/convert.cuh +10 -0
  233. data/ext/sources/ggml/src/ggml-cuda/cpy-utils.cuh +1 -1
  234. data/ext/sources/ggml/src/ggml-cuda/cpy.cu +339 -246
  235. data/ext/sources/ggml/src/ggml-cuda/cpy.cuh +1 -5
  236. data/ext/sources/ggml/src/ggml-cuda/cumsum.cu +307 -0
  237. data/ext/sources/ggml/src/ggml-cuda/cumsum.cuh +5 -0
  238. data/ext/sources/ggml/src/ggml-cuda/diag.cu +77 -0
  239. data/ext/sources/ggml/src/ggml-cuda/diag.cuh +5 -0
  240. data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +31 -21
  241. data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +663 -596
  242. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +35 -741
  243. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +1241 -0
  244. data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +30 -37
  245. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +14 -13
  246. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cuh +48 -0
  247. data/ext/sources/ggml/src/ggml-cuda/fattn.cu +83 -37
  248. data/ext/sources/ggml/src/ggml-cuda/fill.cu +37 -0
  249. data/ext/sources/ggml/src/ggml-cuda/fill.cuh +3 -0
  250. data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +1155 -164
  251. data/ext/sources/ggml/src/ggml-cuda/mean.cu +5 -4
  252. data/ext/sources/ggml/src/ggml-cuda/mma.cuh +741 -48
  253. data/ext/sources/ggml/src/ggml-cuda/mmf.cu +60 -12
  254. data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +381 -42
  255. data/ext/sources/ggml/src/ggml-cuda/mmid.cu +164 -0
  256. data/ext/sources/ggml/src/ggml-cuda/mmid.cuh +5 -0
  257. data/ext/sources/ggml/src/ggml-cuda/mmq.cu +69 -176
  258. data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +498 -171
  259. data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +375 -79
  260. data/ext/sources/ggml/src/ggml-cuda/mmvf.cuh +3 -2
  261. data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +241 -95
  262. data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +1 -1
  263. data/ext/sources/ggml/src/ggml-cuda/pad.cu +64 -33
  264. data/ext/sources/ggml/src/ggml-cuda/quantize.cu +151 -0
  265. data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +14 -0
  266. data/ext/sources/ggml/src/ggml-cuda/rope.cu +192 -77
  267. data/ext/sources/ggml/src/ggml-cuda/rope.cuh +2 -0
  268. data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +101 -47
  269. data/ext/sources/ggml/src/ggml-cuda/set.cu +39 -0
  270. data/ext/sources/ggml/src/ggml-cuda/set.cuh +7 -0
  271. data/ext/sources/ggml/src/ggml-cuda/softmax.cu +203 -6
  272. data/ext/sources/ggml/src/ggml-cuda/solve_tri.cu +275 -0
  273. data/ext/sources/ggml/src/ggml-cuda/solve_tri.cuh +3 -0
  274. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +14 -20
  275. data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +49 -84
  276. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq112-dv112.cu +5 -0
  277. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq128-dv128.cu +5 -0
  278. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq256-dv256.cu +5 -0
  279. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq40-dv40.cu +5 -0
  280. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq576-dv512.cu +5 -0
  281. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq64-dv64.cu +5 -0
  282. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq72-dv72.cu +5 -0
  283. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq80-dv80.cu +5 -0
  284. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq96-dv96.cu +5 -0
  285. data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +19 -1
  286. data/ext/sources/ggml/src/ggml-cuda/top-k.cu +96 -0
  287. data/ext/sources/ggml/src/ggml-cuda/top-k.cuh +3 -0
  288. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +168 -76
  289. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cuh +11 -4
  290. data/ext/sources/ggml/src/ggml-cuda/tri.cu +136 -0
  291. data/ext/sources/ggml/src/ggml-cuda/tri.cuh +5 -0
  292. data/ext/sources/ggml/src/ggml-cuda/unary.cu +105 -11
  293. data/ext/sources/ggml/src/ggml-cuda/unary.cuh +36 -0
  294. data/ext/sources/ggml/src/ggml-cuda/upscale.cu +163 -7
  295. data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +4 -0
  296. data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +12 -1
  297. data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +6 -0
  298. data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +80 -0
  299. data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +3151 -0
  300. data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +44 -0
  301. data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +682 -0
  302. data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +360 -0
  303. data/ext/sources/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +157 -0
  304. data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +566 -0
  305. data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +112 -0
  306. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +35 -0
  307. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-dma.c +63 -0
  308. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-dma.h +157 -0
  309. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +165 -0
  310. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +92 -0
  311. data/ext/sources/ggml/src/ggml-hexagon/htp/htp_iface.idl +16 -0
  312. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.c +94 -0
  313. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.c +72 -0
  314. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.c +49 -0
  315. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.c +1020 -0
  316. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +1353 -0
  317. data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +1001 -0
  318. data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +2503 -0
  319. data/ext/sources/ggml/src/ggml-hexagon/htp/ops-utils.h +149 -0
  320. data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +487 -0
  321. data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +168 -0
  322. data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +402 -0
  323. data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +287 -0
  324. data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.c +297 -0
  325. data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.h +57 -0
  326. data/ext/sources/ggml/src/ggml-hexagon/htp-utils.c +454 -0
  327. data/ext/sources/ggml/src/ggml-hexagon/htp-utils.h +221 -0
  328. data/ext/sources/ggml/src/ggml-hexagon/op-desc.h +153 -0
  329. data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +8 -13
  330. data/ext/sources/ggml/src/ggml-impl.h +67 -6
  331. data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp +2 -2
  332. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.m +29 -20
  333. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +652 -285
  334. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +103 -56
  335. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +496 -118
  336. data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +231 -9
  337. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +1227 -224
  338. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +12 -0
  339. data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +14 -8
  340. data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +1972 -704
  341. data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +3 -1
  342. data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +11 -0
  343. data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +1430 -120
  344. data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +63 -0
  345. data/ext/sources/ggml/src/ggml-opencl/kernels/expm1.cl +82 -0
  346. data/ext/sources/ggml/src/ggml-opencl/kernels/fill.cl +17 -0
  347. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +4 -3
  348. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32.cl +162 -0
  349. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl +156 -0
  350. data/ext/sources/ggml/src/ggml-opencl/kernels/get_rows.cl +36 -12
  351. data/ext/sources/ggml/src/ggml-opencl/kernels/mean.cl +39 -0
  352. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_kq_kqv.cl +273 -0
  353. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl +24 -10
  354. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl +24 -10
  355. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl +154 -0
  356. data/ext/sources/ggml/src/ggml-opencl/kernels/pad.cl +29 -20
  357. data/ext/sources/ggml/src/ggml-opencl/kernels/rms_norm.cl +25 -10
  358. data/ext/sources/ggml/src/ggml-opencl/kernels/rope.cl +50 -24
  359. data/ext/sources/ggml/src/ggml-opencl/kernels/set_rows.cl +35 -16
  360. data/ext/sources/ggml/src/ggml-opencl/kernels/softplus.cl +88 -0
  361. data/ext/sources/ggml/src/ggml-opencl/kernels/sqr.cl +53 -0
  362. data/ext/sources/ggml/src/ggml-opencl/kernels/sqrt.cl +53 -0
  363. data/ext/sources/ggml/src/ggml-opencl/kernels/ssm_conv.cl +77 -0
  364. data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +13 -0
  365. data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +438 -156
  366. data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +48 -3
  367. data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +77 -0
  368. data/ext/sources/ggml/src/ggml-sycl/add-id.hpp +8 -0
  369. data/ext/sources/ggml/src/ggml-sycl/backend.hpp +6 -0
  370. data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +0 -9
  371. data/ext/sources/ggml/src/ggml-sycl/binbcast.hpp +0 -6
  372. data/ext/sources/ggml/src/ggml-sycl/common.hpp +117 -15
  373. data/ext/sources/ggml/src/ggml-sycl/concat.cpp +55 -44
  374. data/ext/sources/ggml/src/ggml-sycl/convert.cpp +34 -0
  375. data/ext/sources/ggml/src/ggml-sycl/count-equal.cpp +79 -0
  376. data/ext/sources/ggml/src/ggml-sycl/count-equal.hpp +9 -0
  377. data/ext/sources/ggml/src/ggml-sycl/cpy.cpp +0 -3
  378. data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +18 -0
  379. data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +76 -3
  380. data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +333 -300
  381. data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +10 -2
  382. data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +335 -110
  383. data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +22 -0
  384. data/ext/sources/ggml/src/ggml-sycl/norm.cpp +156 -0
  385. data/ext/sources/ggml/src/ggml-sycl/norm.hpp +2 -0
  386. data/ext/sources/ggml/src/ggml-sycl/pad.cpp +97 -0
  387. data/ext/sources/ggml/src/ggml-sycl/pad.hpp +24 -0
  388. data/ext/sources/ggml/src/ggml-sycl/pad_reflect_1d.cpp +100 -0
  389. data/ext/sources/ggml/src/ggml-sycl/pad_reflect_1d.hpp +10 -0
  390. data/ext/sources/ggml/src/ggml-sycl/presets.hpp +2 -0
  391. data/ext/sources/ggml/src/ggml-sycl/repeat_back.cpp +76 -0
  392. data/ext/sources/ggml/src/ggml-sycl/repeat_back.hpp +8 -0
  393. data/ext/sources/ggml/src/ggml-sycl/roll.cpp +122 -0
  394. data/ext/sources/ggml/src/ggml-sycl/roll.hpp +20 -0
  395. data/ext/sources/ggml/src/ggml-sycl/rope.cpp +30 -17
  396. data/ext/sources/ggml/src/ggml-sycl/set.cpp +73 -0
  397. data/ext/sources/ggml/src/ggml-sycl/set.hpp +5 -0
  398. data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +327 -162
  399. data/ext/sources/ggml/src/ggml-sycl/softmax.hpp +4 -0
  400. data/ext/sources/ggml/src/ggml-sycl/ssm_conv.cpp +127 -0
  401. data/ext/sources/ggml/src/ggml-sycl/ssm_conv.hpp +5 -0
  402. data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +58 -0
  403. data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +38 -18
  404. data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +5013 -2859
  405. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/abs.comp +21 -0
  406. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +2 -2
  407. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add.comp +2 -2
  408. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add1.comp +28 -0
  409. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp +1 -1
  410. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/arange.comp +20 -0
  411. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp +2 -2
  412. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp +33 -26
  413. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argsort_large.comp +114 -0
  414. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ceil.comp +22 -0
  415. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/clamp.comp +2 -2
  416. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/concat.comp +2 -2
  417. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +2 -2
  418. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp +1 -1
  419. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +47 -49
  420. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp +1 -1
  421. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +2 -2
  422. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +3 -3
  423. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +4 -4
  424. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_transpose.comp +67 -0
  425. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/cos.comp +2 -2
  426. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/count_equal.comp +2 -2
  427. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/count_experts.comp +51 -0
  428. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/cumsum.comp +83 -0
  429. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/cumsum_multipass1.comp +60 -0
  430. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/cumsum_multipass2.comp +66 -0
  431. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_f32.comp +1 -1
  432. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{dequant_funcs.comp → dequant_funcs.glsl} +9 -21
  433. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{dequant_funcs_cm2.comp → dequant_funcs_cm2.glsl} +18 -4
  434. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{dequant_head.comp → dequant_head.glsl} +1 -1
  435. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp +1 -1
  436. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp +1 -1
  437. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp +1 -1
  438. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xs.comp +1 -1
  439. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp +1 -1
  440. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp +1 -1
  441. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp +1 -1
  442. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp +1 -1
  443. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_xs.comp +1 -1
  444. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp +3 -3
  445. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp +3 -3
  446. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp +1 -1
  447. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_0.comp +1 -1
  448. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_1.comp +1 -1
  449. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp +3 -3
  450. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_0.comp +1 -1
  451. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_1.comp +1 -1
  452. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp +3 -3
  453. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp +1 -1
  454. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q8_0.comp +1 -1
  455. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +29 -0
  456. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp +1 -1
  457. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/div.comp +2 -2
  458. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +3 -3
  459. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/fill.comp +19 -0
  460. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +39 -17
  461. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{flash_attn_base.comp → flash_attn_base.glsl} +19 -1
  462. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +45 -7
  463. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +50 -12
  464. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +1 -1
  465. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/floor.comp +22 -0
  466. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/geglu.comp +2 -2
  467. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/geglu_erf.comp +2 -2
  468. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/geglu_quick.comp +2 -2
  469. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu.comp +2 -2
  470. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_erf.comp +2 -2
  471. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_quick.comp +2 -2
  472. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{generic_binary_head.comp → generic_binary_head.glsl} +17 -2
  473. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{generic_head.comp → generic_head.glsl} +2 -0
  474. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{generic_unary_head.comp → generic_unary_head.glsl} +7 -0
  475. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +4 -4
  476. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +3 -3
  477. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{glu_head.comp → glu_head.glsl} +1 -1
  478. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp +2 -2
  479. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardsigmoid.comp +2 -2
  480. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardswish.comp +2 -2
  481. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +19 -7
  482. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +2 -3
  483. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +2 -2
  484. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/leaky_relu.comp +2 -2
  485. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +18 -0
  486. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul.comp +2 -2
  487. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +2 -2
  488. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{mul_mat_vec_base.comp → mul_mat_vec_base.glsl} +70 -25
  489. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +35 -0
  490. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp +71 -21
  491. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp +41 -25
  492. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp +2 -2
  493. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp +44 -26
  494. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp +2 -2
  495. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp +2 -2
  496. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp +2 -2
  497. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +9 -7
  498. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp +9 -7
  499. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +4 -6
  500. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp +2 -2
  501. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +4 -6
  502. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +4 -6
  503. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp +2 -2
  504. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +39 -36
  505. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +494 -0
  506. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +78 -103
  507. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +34 -23
  508. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{mul_mm_funcs.comp → mul_mm_funcs.glsl} +69 -59
  509. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +72 -0
  510. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +88 -228
  511. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +454 -0
  512. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +78 -0
  513. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +97 -13
  514. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/neg.comp +20 -0
  515. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/norm.comp +2 -2
  516. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_adamw.comp +2 -2
  517. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp +1 -1
  518. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/pad.comp +21 -6
  519. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/pool2d.comp +1 -1
  520. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp +10 -10
  521. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/reglu.comp +2 -2
  522. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/relu.comp +2 -2
  523. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/repeat.comp +2 -2
  524. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp +2 -2
  525. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +50 -4
  526. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_back.comp +2 -2
  527. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_partials.comp +2 -2
  528. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/roll.comp +2 -2
  529. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +234 -0
  530. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +20 -0
  531. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +6 -50
  532. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +6 -33
  533. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +6 -33
  534. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +28 -0
  535. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +6 -39
  536. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/round.comp +29 -0
  537. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/scale.comp +2 -2
  538. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp +2 -2
  539. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/silu.comp +2 -2
  540. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/silu_back.comp +2 -2
  541. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sin.comp +2 -2
  542. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +1 -1
  543. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp +2 -2
  544. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large1.comp +62 -0
  545. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large2.comp +79 -0
  546. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large3.comp +65 -0
  547. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large_common.glsl +53 -0
  548. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/softplus.comp +23 -0
  549. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/solve_tri.comp +81 -0
  550. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp +2 -2
  551. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/square.comp +2 -2
  552. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +44 -0
  553. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_scan.comp +124 -0
  554. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/step.comp +22 -0
  555. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sub.comp +2 -2
  556. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp +2 -25
  557. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.glsl +25 -0
  558. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/swiglu.comp +2 -2
  559. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp +2 -2
  560. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp +2 -2
  561. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp +1 -1
  562. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/topk_argsort.comp +118 -0
  563. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/topk_moe.comp +213 -0
  564. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/topk_nary_search.comp +246 -0
  565. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +43 -0
  566. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/trunc.comp +22 -0
  567. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{types.comp → types.glsl} +345 -26
  568. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +90 -12
  569. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +335 -151
  570. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/xielu.comp +35 -0
  571. data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +28 -2
  572. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +169 -0
  573. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +1964 -435
  574. data/ext/sources/ggml/src/ggml-webgpu/pre_wgsl.hpp +778 -0
  575. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/bin_op.tmpl.wgsl +188 -0
  576. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cpy.tmpl.wgsl +101 -0
  577. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +33 -10
  578. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +591 -0
  579. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.tmpl.wgsl +1 -1
  580. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.tmpl.wgsl +323 -0
  581. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl +6 -6
  582. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +97 -0
  583. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.tmpl.wgsl +247 -0
  584. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.tmpl.wgsl +302 -0
  585. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.tmpl.wgsl +267 -0
  586. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +83 -17
  587. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rope.tmpl.wgsl +295 -0
  588. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/scale.tmpl.wgsl +90 -0
  589. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.tmpl.wgsl +112 -0
  590. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/soft_max.tmpl.wgsl +345 -0
  591. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary_op.wgsl +483 -0
  592. data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +92 -0
  593. data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +466 -0
  594. data/ext/sources/ggml/src/ggml.c +425 -33
  595. data/ext/sources/include/whisper.h +1 -0
  596. data/ext/sources/src/CMakeLists.txt +3 -1
  597. data/ext/sources/src/whisper.cpp +101 -35
  598. data/ext/sources/tests/CMakeLists.txt +2 -2
  599. data/ext/sources/tests/test-vad-full.cpp +4 -2
  600. data/ext/sources/tests/test-vad.cpp +1 -1
  601. data/extsources.rb +1 -0
  602. data/lib/whisper/model/uri.rb +17 -18
  603. data/sig/whisper.rbs +119 -2
  604. data/test/test_params.rb +16 -8
  605. data/test/test_segment.rb +0 -1
  606. data/test/test_token.rb +70 -0
  607. data/test/test_vad.rb +1 -1
  608. data/test/test_vad_context.rb +50 -0
  609. data/test/test_vad_segment.rb +19 -0
  610. data/test/test_vad_segments.rb +16 -0
  611. data/test/test_whisper.rb +7 -0
  612. data/whispercpp.gemspec +1 -1
  613. metadata +287 -34
  614. data/ext/sources/build-xcframework.sh +0 -571
  615. data/ext/sources/ggml/src/ggml-cann/Doxyfile +0 -2579
  616. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp +0 -105
  617. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp +0 -55
  618. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add.tmpl.wgsl +0 -44
  619. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_in_place.tmpl.wgsl +0 -41
  620. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl +0 -60
  621. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul.tmpl.wgsl +0 -44
  622. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_in_place.tmpl.wgsl +0 -41
  623. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_in_place.wgsl +0 -48
  624. /data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{test_bfloat16_support.comp → feature-tests/bfloat16.comp} +0 -0
  625. /data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{test_coopmat_support.comp → feature-tests/coopmat.comp} +0 -0
  626. /data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{test_coopmat2_support.comp → feature-tests/coopmat2.comp} +0 -0
  627. /data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{test_integer_dot_support.comp → feature-tests/integer_dot.comp} +0 -0
  628. /data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{glu_main.comp → glu_main.glsl} +0 -0
  629. /data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{rte.comp → rte.glsl} +0 -0
  630. /data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/{utils.comp → utils.glsl} +0 -0
@@ -126,36 +126,48 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
126
126
  )
127
127
  if (NOT ARM_MCPU_RESULT)
128
128
  string(REGEX MATCH "-mcpu=[^ ']+" ARM_MCPU_FLAG "${ARM_MCPU}")
129
+ string(REGEX MATCH "-march=[^ ']+" ARM_MARCH_FLAG "${ARM_MCPU}")
130
+
131
+ # on some old GCC we need to read -march=
132
+ if (ARM_MARCH_FLAG AND NOT "${ARM_MARCH_FLAG}" STREQUAL "-march=native")
133
+ set(ARM_NATIVE_FLAG "${ARM_MARCH_FLAG}")
134
+ elseif(ARM_MCPU_FLAG AND NOT "${ARM_MCPU_FLAG}" STREQUAL "-mcpu=native")
135
+ set(ARM_NATIVE_FLAG "${ARM_MCPU_FLAG}")
136
+ endif()
129
137
  endif()
130
- if ("${ARM_MCPU_FLAG}" STREQUAL "")
131
- set(ARM_MCPU_FLAG -mcpu=native)
132
- message(STATUS "ARM -mcpu not found, -mcpu=native will be used")
138
+
139
+ if ("${ARM_NATIVE_FLAG}" STREQUAL "")
140
+ set(ARM_NATIVE_FLAG -mcpu=native)
141
+ message(WARNING "ARM -march/-mcpu not found, -mcpu=native will be used")
142
+ else()
143
+ message(STATUS "ARM detected flags: ${ARM_NATIVE_FLAG}")
133
144
  endif()
134
145
 
135
146
  include(CheckCXXSourceRuns)
136
147
 
137
- function(check_arm_feature tag code)
148
+ macro(check_arm_feature tag feature code)
138
149
  set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
139
- set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+${tag}")
150
+ set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+${tag}")
140
151
  check_cxx_source_runs("${code}" GGML_MACHINE_SUPPORTS_${tag})
141
152
  if (GGML_MACHINE_SUPPORTS_${tag})
142
- set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+${tag}" PARENT_SCOPE)
153
+ set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+${tag}")
143
154
  else()
144
- set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+no${tag}")
155
+ set(CMAKE_REQUIRED_FLAGS "${ARM_NATIVE_FLAG}+no${tag}")
145
156
  check_cxx_source_compiles("int main() { return 0; }" GGML_MACHINE_SUPPORTS_no${tag})
146
157
  if (GGML_MACHINE_SUPPORTS_no${tag})
147
- set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE)
158
+ set(ARM_NATIVE_FLAG_FIX "${ARM_NATIVE_FLAG_FIX}+no${tag}")
159
+ list(APPEND ARCH_FLAGS -U__ARM_FEATURE_${feature})
148
160
  endif()
149
161
  endif()
150
162
  set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
151
- endfunction()
163
+ endmacro()
152
164
 
153
- check_arm_feature(dotprod "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }")
154
- check_arm_feature(i8mm "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }")
155
- check_arm_feature(sve "#include <arm_sve.h>\nint main() { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }")
156
- check_arm_feature(sme "#include <arm_sme.h>\n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }")
165
+ check_arm_feature(dotprod DOTPROD "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }")
166
+ check_arm_feature(i8mm MATMUL_INT8 "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }")
167
+ check_arm_feature(sve SVE "#include <arm_sve.h>\nint main() { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }")
168
+ check_arm_feature(sme SME "#include <arm_sme.h>\n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }")
157
169
 
158
- list(APPEND ARCH_FLAGS "${ARM_MCPU_FLAG}${ARM_MCPU_FLAG_FIX}")
170
+ list(APPEND ARCH_FLAGS "${ARM_NATIVE_FLAG}${ARM_NATIVE_FLAG_FIX}")
159
171
  else()
160
172
  if (GGML_CPU_ARM_ARCH)
161
173
  list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
@@ -205,35 +217,28 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
205
217
  endif()
206
218
  endif()
207
219
 
208
- # show enabled features
209
- if (CMAKE_HOST_SYSTEM_NAME STREQUAL "Windows")
210
- set(FEAT_INPUT_FILE "NUL")
211
- else()
212
- set(FEAT_INPUT_FILE "/dev/null")
213
- endif()
220
+ message(STATUS "Checking for ARM features using flags:")
221
+ foreach(flag IN LISTS ARCH_FLAGS)
222
+ message(STATUS " ${flag}")
223
+ endforeach()
214
224
 
215
- execute_process(
216
- COMMAND ${CMAKE_C_COMPILER} ${ARCH_FLAGS} -dM -E -
217
- INPUT_FILE ${FEAT_INPUT_FILE}
218
- OUTPUT_VARIABLE ARM_FEATURE
219
- RESULT_VARIABLE ARM_FEATURE_RESULT
220
- )
221
- if (ARM_FEATURE_RESULT)
222
- message(WARNING "Failed to get ARM features")
223
- else()
224
- foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
225
- string(FIND "${ARM_FEATURE}" "__ARM_FEATURE_${feature} 1" feature_pos)
226
- if (NOT ${feature_pos} EQUAL -1)
227
- # Special handling for MATMUL_INT8 when machine doesn't support i8mm
228
- if ("${feature}" STREQUAL "MATMUL_INT8" AND GGML_MACHINE_SUPPORTS_noi8mm)
229
- message(STATUS "ARM feature ${feature} detected but unsetting due to machine not supporting i8mm")
230
- list(APPEND ARCH_FLAGS -U__ARM_FEATURE_MATMUL_INT8)
231
- else()
232
- message(STATUS "ARM feature ${feature} enabled")
233
- endif()
234
- endif()
235
- endforeach()
236
- endif()
225
+ include(CheckCXXSourceCompiles)
226
+ set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
227
+ string(REPLACE ";" " " ARCH_FLAGS_STR "${ARCH_FLAGS}")
228
+ set(CMAKE_REQUIRED_FLAGS "${ARCH_FLAGS_STR}")
229
+ foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
230
+ set(ARM_FEATURE "HAVE_${feature}")
231
+ check_cxx_source_compiles(
232
+ "
233
+ #if !defined(__ARM_FEATURE_${feature})
234
+ # error \"Feature ${feature} is not defined\"
235
+ #endif
236
+ int main() { return 0; }
237
+ "
238
+ ${ARM_FEATURE}
239
+ )
240
+ endforeach()
241
+ set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
237
242
  endif()
238
243
  elseif (GGML_SYSTEM_ARCH STREQUAL "x86")
239
244
  message(STATUS "x86 detected")
@@ -388,9 +393,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
388
393
  string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
389
394
 
390
395
  if (EXTRACTED_NUMBER GREATER_EQUAL 10)
391
- list(APPEND ARCH_FLAGS -mcpu=power10 -mpowerpc64)
396
+ list(APPEND ARCH_FLAGS -mcpu=power10)
392
397
  elseif (EXTRACTED_NUMBER EQUAL 9)
393
- list(APPEND ARCH_FLAGS -mcpu=power9 -mpowerpc64)
398
+ list(APPEND ARCH_FLAGS -mcpu=power9)
394
399
  elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
395
400
  list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native)
396
401
  else()
@@ -448,51 +453,94 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
448
453
  ggml-cpu/spacemit/ime_kernels.h
449
454
  )
450
455
  endif()
451
- set(MARCH_STR "rv64gc")
452
- if (GGML_RV_ZFH)
453
- string(APPEND MARCH_STR "_zfh")
454
- endif()
455
- if (GGML_XTHEADVECTOR)
456
- string(APPEND MARCH_STR "_xtheadvector")
457
- elseif (GGML_RVV)
458
- string(APPEND MARCH_STR "_v")
459
- if (GGML_RV_ZVFH)
460
- string(APPEND MARCH_STR "_zvfh")
456
+ if(NOT GGML_CPU_ALL_VARIANTS)
457
+ set(MARCH_STR "rv64gc")
458
+ if (GGML_RV_ZFH)
459
+ string(APPEND MARCH_STR "_zfh")
461
460
  endif()
461
+
462
+ if (GGML_XTHEADVECTOR)
463
+ string(APPEND MARCH_STR "_xtheadvector")
464
+ elseif (GGML_RVV)
465
+ string(APPEND MARCH_STR "_v")
466
+ if (GGML_RV_ZVFH)
467
+ string(APPEND MARCH_STR "_zvfh")
468
+ endif()
469
+ if (GGML_RV_ZVFBFWMA)
470
+ string(APPEND MARCH_STR "_zvfbfwma")
471
+ endif()
472
+ endif()
473
+ if (GGML_RV_ZICBOP)
474
+ string(APPEND MARCH_STR "_zicbop")
475
+ endif()
476
+ if (GGML_RV_ZIHINTPAUSE)
477
+ string(APPEND MARCH_STR "_zihintpause")
478
+ endif()
479
+ list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
480
+ else()
481
+ # Begin with the lowest baseline
482
+ set(ARCH_DEFINITIONS "")
483
+
484
+ if (GGML_INTERNAL_RVV)
485
+ message(STATUS "RVV enabled")
486
+ list(APPEND ARCH_DEFINITIONS GGML_USE_RVV)
487
+ list(APPEND ARCH_FLAGS -march=rv64gc_v -mabi=lp64d)
488
+ endif()
489
+
490
+ ggml_add_cpu_backend_features(${GGML_CPU_NAME} riscv ${ARCH_DEFINITIONS})
462
491
  endif()
463
- if (GGML_RV_ZICBOP)
464
- string(APPEND MARCH_STR "_zicbop")
465
- endif()
466
- list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
467
492
  elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
468
493
  message(STATUS "s390x detected")
469
- list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/s390/quants.c)
470
- file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
471
- string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
472
-
473
- # TODO: Separation to determine activation of VX/VXE/VXE2
474
- if (${S390X_M} MATCHES "8561|8562")
475
- message(STATUS "z15 target")
476
- list(APPEND ARCH_FLAGS -march=z15)
477
- elseif (${S390X_M} MATCHES "3931")
478
- message(STATUS "z16 target")
479
- list(APPEND ARCH_FLAGS -march=z16)
480
- elseif (${S390X_M} MATCHES "9175|9176")
481
- # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
482
- # binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
483
- message(STATUS "z17 target")
484
- list(APPEND ARCH_FLAGS -march=arch15)
485
- else()
486
- message(STATUS "Unknown target")
487
- message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
488
- list(APPEND ARCH_FLAGS -march=native -mtune=native)
494
+ list(APPEND GGML_CPU_SOURCES
495
+ ggml-cpu/arch/s390/quants.c)
496
+
497
+ # for native compilation
498
+ if (GGML_NATIVE)
499
+ # check machine level to determine target
500
+ file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
501
+ string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
502
+
503
+ # TODO: Separation to determine activation of VX/VXE/VXE2
504
+ if (${S390X_M} MATCHES "8561|8562")
505
+ message(STATUS "z15 target")
506
+ list(APPEND ARCH_FLAGS -march=z15)
507
+ elseif (${S390X_M} MATCHES "3931")
508
+ message(STATUS "z16 target")
509
+ list(APPEND ARCH_FLAGS -march=z16)
510
+ elseif (${S390X_M} MATCHES "9175|9176")
511
+ # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
512
+ # binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
513
+ message(STATUS "z17 target")
514
+ list(APPEND ARCH_FLAGS -march=arch15)
515
+ else()
516
+ message(STATUS "Unknown target")
517
+ message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")
518
+ list(APPEND ARCH_FLAGS -march=native -mtune=native)
519
+ endif()
520
+ # for cross-compilation
521
+ elseif(GGML_CPU_ALL_VARIANTS)
522
+ # range through IBM z15 to z17
523
+ # NOTE: update when a new hardware level is released
524
+ foreach (ZHW RANGE 15 17)
525
+ if(DEFINED GGML_INTERNAL_Z${ZHW})
526
+ message(STATUS "z${ZHW} cross-compile target")
527
+ list(APPEND ARCH_FLAGS -march=z${ZHW})
528
+ endif()
529
+ endforeach()
489
530
  endif()
490
531
 
491
- if (GGML_VXE)
492
- message(STATUS "VX/VXE/VXE2 enabled")
532
+ if (GGML_VXE OR GGML_INTERNAL_VXE2)
533
+ message(STATUS "VXE2 enabled")
493
534
  list(APPEND ARCH_FLAGS -mvx -mzvector)
494
- list(APPEND ARCH_DEFINITIONS GGML_VXE)
535
+ list(APPEND ARCH_DEFINITIONS GGML_USE_VXE2)
536
+ endif()
537
+
538
+ if (GGML_INTERNAL_NNPA)
539
+ message(STATUS "NNPA enabled")
540
+ list(APPEND ARCH_DEFINITIONS GGML_USE_NNPA)
495
541
  endif()
542
+
543
+ ggml_add_cpu_backend_features(${GGML_CPU_NAME} s390 ${ARCH_DEFINITIONS})
496
544
  elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
497
545
  message(STATUS "Wasm detected")
498
546
  list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)
@@ -513,9 +561,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
513
561
 
514
562
  # Fetch KleidiAI sources:
515
563
  include(FetchContent)
516
- set(KLEIDIAI_COMMIT_TAG "v1.14.0")
564
+ set(KLEIDIAI_COMMIT_TAG "v1.16.0")
517
565
  set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
518
- set(KLEIDIAI_ARCHIVE_MD5 "45e110675d93f99f82c23a1afcca76bc")
566
+ set(KLEIDIAI_ARCHIVE_MD5 "0a9e9008adb6031f9e8cf70dff4a3321")
519
567
 
520
568
  if (POLICY CMP0135)
521
569
  cmake_policy(SET CMP0135 NEW)
@@ -556,6 +604,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
556
604
  ${KLEIDIAI_SRC}/kai/ukernels/
557
605
  ${KLEIDIAI_SRC}/kai/ukernels/matmul/
558
606
  ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
607
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/
559
608
  ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
560
609
  ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)
561
610
 
@@ -566,6 +615,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
566
615
  string(FIND "${ARCH_FLAGS_TEMP}" "+dotprod" DOTPROD_ENABLED)
567
616
  string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED)
568
617
  string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED)
618
+ string(FIND "${ARCH_FLAGS_TEMP}" "+sve" SVE_ENABLED)
569
619
 
570
620
  set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS_TEMP})
571
621
 
@@ -574,23 +624,34 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
574
624
  ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c
575
625
  ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c
576
626
  ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c
577
- ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)
627
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c
628
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c
629
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c)
578
630
 
579
631
  if (NOT DOTPROD_ENABLED MATCHES -1)
580
632
  list(APPEND GGML_KLEIDIAI_SOURCES
581
633
  ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c
582
634
  ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c
583
- ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c)
635
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c
636
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c
637
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c
638
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c)
584
639
  endif()
585
640
 
586
641
  if (NOT I8MM_ENABLED MATCHES -1)
587
- list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c)
642
+ list(APPEND GGML_KLEIDIAI_SOURCES
643
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c
644
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c)
588
645
  endif()
589
646
 
590
647
  if (NOT SME_ENABLED MATCHES -1)
591
648
  list(APPEND GGML_KLEIDIAI_SOURCES
592
649
  ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c
593
650
  ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
651
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c
652
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa_asm.S
653
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c
654
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot_asm.S
594
655
  ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
595
656
  ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S
596
657
  ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
@@ -599,6 +660,15 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
599
660
  set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2")
600
661
  endif()
601
662
 
663
+ if (NOT SVE_ENABLED MATCHES -1)
664
+ list(APPEND GGML_KLEIDIAI_SOURCES
665
+ ${KLEIDIAI_SRC}/kai/kai_common_sve_asm.S
666
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod_asm.S
667
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.c
668
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm_asm.S
669
+ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm.c)
670
+ endif()
671
+
602
672
  set_source_files_properties(${GGML_KLEIDIAI_SOURCES} PROPERTIES COMPILE_OPTIONS "${PRIVATE_ARCH_FLAGS}")
603
673
  list(APPEND GGML_CPU_SOURCES ${GGML_KLEIDIAI_SOURCES})
604
674
  endif()
@@ -149,6 +149,7 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
149
149
  if (op->op == GGML_OP_MUL_MAT && is_contiguous_2d(op->src[0]) && // src0 must be contiguous
150
150
  is_contiguous_2d(op->src[1]) && // src1 must be contiguous
151
151
  op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_amx_buffer_type() &&
152
+ op->src[0]->ne[0] % (TILE_K * 2 * 32) == 0 && // TODO: not sure if correct (https://github.com/ggml-org/llama.cpp/pull/16315)
152
153
  op->ne[0] % (TILE_N * 2) == 0 && // out_features is 32x
153
154
  (qtype_has_amx_kernels(op->src[0]->type) || (op->src[0]->type == GGML_TYPE_F16))) {
154
155
  // src1 must be host buffer
@@ -8,6 +8,10 @@
8
8
  #include <sys/sysctl.h>
9
9
  #endif
10
10
 
11
+ #if !defined(HWCAP2_SVE2)
12
+ #define HWCAP2_SVE2 (1 << 1)
13
+ #endif
14
+
11
15
  #if !defined(HWCAP2_I8MM)
12
16
  #define HWCAP2_I8MM (1 << 13)
13
17
  #endif