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
@@ -66,6 +66,7 @@ struct block_q4_0_packed16
66
66
  #define QUANT_AUXF 1
67
67
  #define A_TYPE block_q4_0
68
68
  #define A_TYPE_PACKED16 block_q4_0_packed16
69
+ #define DATA_A_QUANT_LEGACY
69
70
  #endif
70
71
 
71
72
  #define QUANT_K_Q4_1 32
@@ -98,6 +99,7 @@ struct block_q4_1_packed32
98
99
  #define A_TYPE block_q4_1
99
100
  #define A_TYPE_PACKED16 block_q4_1_packed16
100
101
  #define A_TYPE_PACKED32 block_q4_1_packed32
102
+ #define DATA_A_QUANT_LEGACY
101
103
  #endif
102
104
 
103
105
  #define QUANT_K_Q5_0 32
@@ -123,6 +125,7 @@ struct block_q5_0_packed16
123
125
  #define QUANT_AUXF 1
124
126
  #define A_TYPE block_q5_0
125
127
  #define A_TYPE_PACKED16 block_q5_0_packed16
128
+ #define DATA_A_QUANT_LEGACY
126
129
  #endif
127
130
 
128
131
  #define QUANT_K_Q5_1 32
@@ -158,6 +161,7 @@ struct block_q5_1_packed32
158
161
  #define A_TYPE block_q5_1
159
162
  #define A_TYPE_PACKED16 block_q5_1_packed16
160
163
  #define A_TYPE_PACKED32 block_q5_1_packed32
164
+ #define DATA_A_QUANT_LEGACY
161
165
  #endif
162
166
 
163
167
  #define QUANT_K_Q8_0 32
@@ -168,16 +172,12 @@ struct block_q8_0
168
172
  float16_t d;
169
173
  int8_t qs[32];
170
174
  };
175
+
171
176
  struct block_q8_0_packed16
172
177
  {
173
178
  float16_t d;
174
179
  int16_t qs[32/2];
175
180
  };
176
- struct block_q8_0_packed32
177
- {
178
- float16_t d;
179
- int32_t qs[32/4];
180
- };
181
181
 
182
182
  #if defined(DATA_A_Q8_0)
183
183
  #define QUANT_K QUANT_K_Q8_0
@@ -185,7 +185,7 @@ struct block_q8_0_packed32
185
185
  #define QUANT_AUXF 1
186
186
  #define A_TYPE block_q8_0
187
187
  #define A_TYPE_PACKED16 block_q8_0_packed16
188
- #define A_TYPE_PACKED32 block_q8_0_packed32
188
+ #define DATA_A_QUANT_LEGACY
189
189
  #endif
190
190
 
191
191
  #define QUANT_K_Q8_1 32
@@ -196,11 +196,13 @@ struct block_q8_1
196
196
  f16vec2 ds;
197
197
  int8_t qs[32];
198
198
  };
199
+
199
200
  struct block_q8_1_packed16
200
201
  {
201
202
  f16vec2 ds;
202
203
  int16_t qs[16];
203
204
  };
205
+
204
206
  struct block_q8_1_packed32
205
207
  {
206
208
  f16vec2 ds;
@@ -213,6 +215,7 @@ struct block_q8_1_x4
213
215
  f16vec2 ds[4];
214
216
  int32_t qs[32];
215
217
  };
218
+
216
219
  struct block_q8_1_x4_packed128
217
220
  {
218
221
  f16vec2 ds[4];
@@ -226,21 +229,21 @@ struct block_q2_K
226
229
  {
227
230
  uint8_t scales[QUANT_K_Q2_K/16];
228
231
  uint8_t qs[QUANT_K_Q2_K/4];
229
- f16vec2 d;
232
+ f16vec2 dm;
230
233
  };
231
234
 
232
235
  struct block_q2_K_packed16
233
236
  {
234
237
  uint16_t scales[QUANT_K_Q2_K/16/2];
235
238
  uint16_t qs[QUANT_K_Q2_K/4/2];
236
- f16vec2 d;
239
+ f16vec2 dm;
237
240
  };
238
241
 
239
242
  struct block_q2_K_packed32
240
243
  {
241
244
  uint32_t scales[QUANT_K_Q2_K/16/4];
242
245
  uint32_t qs[QUANT_K_Q2_K/4/4];
243
- f16vec2 d;
246
+ f16vec2 dm;
244
247
  };
245
248
 
246
249
  #if defined(DATA_A_Q2_K)
@@ -249,6 +252,8 @@ struct block_q2_K_packed32
249
252
  #define A_TYPE block_q2_K
250
253
  #define A_TYPE_PACKED16 block_q2_K_packed16
251
254
  #define A_TYPE_PACKED32 block_q2_K_packed32
255
+ #define SCALES_PER_32 2
256
+ #define DATA_A_QUANT_K
252
257
  #endif
253
258
 
254
259
  #define QUANT_K_Q3_K 256
@@ -274,27 +279,28 @@ struct block_q3_K_packed16
274
279
  #define QUANT_R 1
275
280
  #define A_TYPE block_q3_K
276
281
  #define A_TYPE_PACKED16 block_q3_K_packed16
282
+ #define DATA_A_QUANT_K
277
283
  #endif
278
284
 
279
285
  #define QUANT_K_Q4_K 256
280
286
 
281
287
  struct block_q4_K
282
288
  {
283
- f16vec2 d;
289
+ f16vec2 dm;
284
290
  uint8_t scales[3*QUANT_K_Q4_K/64];
285
291
  uint8_t qs[QUANT_K_Q4_K/2];
286
292
  };
287
293
 
288
294
  struct block_q4_K_packed16
289
295
  {
290
- f16vec2 d;
296
+ f16vec2 dm;
291
297
  uint16_t scales[3*QUANT_K_Q4_K/64/2];
292
298
  uint16_t qs[QUANT_K_Q4_K/2/2];
293
299
  };
294
300
 
295
301
  struct block_q4_K_packed32
296
302
  {
297
- f16vec2 d;
303
+ f16vec2 dm;
298
304
  uint32_t scales[3*QUANT_K_Q4_K/64/4];
299
305
  uint32_t qs[QUANT_K_Q4_K/2/4];
300
306
  };
@@ -310,13 +316,14 @@ struct block_q4_K_packed128
310
316
  #define A_TYPE block_q4_K
311
317
  #define A_TYPE_PACKED16 block_q4_K_packed16
312
318
  #define A_TYPE_PACKED32 block_q4_K_packed32
319
+ #define DATA_A_QUANT_K
313
320
  #endif
314
321
 
315
322
  #define QUANT_K_Q5_K 256
316
323
 
317
324
  struct block_q5_K
318
325
  {
319
- f16vec2 d;
326
+ f16vec2 dm;
320
327
  uint8_t scales[12];
321
328
  uint8_t qh[QUANT_K_Q5_K/8];
322
329
  uint8_t qs[QUANT_K_Q5_K/2];
@@ -324,12 +331,20 @@ struct block_q5_K
324
331
 
325
332
  struct block_q5_K_packed16
326
333
  {
327
- f16vec2 d;
334
+ f16vec2 dm;
328
335
  uint16_t scales[12/2];
329
336
  uint16_t qh[QUANT_K_Q5_K/8/2];
330
337
  uint16_t qs[QUANT_K_Q5_K/2/2];
331
338
  };
332
339
 
340
+ struct block_q5_K_packed32
341
+ {
342
+ f16vec2 dm;
343
+ uint32_t scales[12/4];
344
+ uint32_t qh[QUANT_K_Q5_K/8/4];
345
+ uint32_t qs[QUANT_K_Q5_K/2/4];
346
+ };
347
+
333
348
  struct block_q5_K_packed128
334
349
  {
335
350
  uvec4 q5k[11];
@@ -340,6 +355,8 @@ struct block_q5_K_packed128
340
355
  #define QUANT_R 1
341
356
  #define A_TYPE block_q5_K
342
357
  #define A_TYPE_PACKED16 block_q5_K_packed16
358
+ #define A_TYPE_PACKED32 block_q5_K_packed32
359
+ #define DATA_A_QUANT_K
343
360
  #endif
344
361
 
345
362
  #define QUANT_K_Q6_K 256
@@ -356,7 +373,7 @@ struct block_q6_K_packed16
356
373
  {
357
374
  uint16_t ql[QUANT_K_Q6_K/2/2];
358
375
  uint16_t qh[QUANT_K_Q6_K/4/2];
359
- int8_t scales[QUANT_K_Q6_K/16];
376
+ int16_t scales[QUANT_K_Q6_K/16/2];
360
377
  float16_t d;
361
378
  };
362
379
 
@@ -365,6 +382,7 @@ struct block_q6_K_packed16
365
382
  #define QUANT_R 1
366
383
  #define A_TYPE block_q6_K
367
384
  #define A_TYPE_PACKED16 block_q6_K_packed16
385
+ #define DATA_A_QUANT_K
368
386
  #endif
369
387
 
370
388
  // IQuants
@@ -378,6 +396,12 @@ struct block_iq1_s {
378
396
  uint16_t qh[QUANT_K_IQ1_S/32];
379
397
  };
380
398
 
399
+ struct block_iq1_s_packed16 {
400
+ float16_t d;
401
+ uint16_t qs[QUANT_K_IQ1_S/8/2];
402
+ uint16_t qh[QUANT_K_IQ1_S/32];
403
+ };
404
+
381
405
  #define QUANT_K_IQ1_M 256
382
406
  #define QUANT_R_IQ1_M 1
383
407
 
@@ -387,6 +411,18 @@ struct block_iq1_m {
387
411
  uint16_t scales[QUANT_K_IQ1_M/64];
388
412
  };
389
413
 
414
+ struct block_iq1_m_packed16 {
415
+ uint16_t qs[QUANT_K_IQ1_M/8/2];
416
+ uint16_t qh[QUANT_K_IQ1_M/16/2];
417
+ uint16_t scales[QUANT_K_IQ1_M/64];
418
+ };
419
+
420
+ struct block_iq1_m_packed32 {
421
+ uint32_t qs[QUANT_K_IQ1_M/8/4];
422
+ uint32_t qh[QUANT_K_IQ1_M/16/4];
423
+ uint32_t scales[QUANT_K_IQ1_M/64/2];
424
+ };
425
+
390
426
  struct block_iq1_m_packed64 {
391
427
  uint64_t qs[QUANT_K_IQ1_M/8/8];
392
428
  uint64_t qh[QUANT_K_IQ1_M/16/8];
@@ -397,12 +433,15 @@ struct block_iq1_m_packed64 {
397
433
  #define QUANT_K QUANT_K_IQ1_S
398
434
  #define QUANT_R QUANT_R_IQ1_S
399
435
  #define A_TYPE block_iq1_s
436
+ #define A_TYPE_PACKED16 block_iq1_s_packed16
400
437
  #endif
401
438
 
402
439
  #if defined(DATA_A_IQ1_M)
403
440
  #define QUANT_K QUANT_K_IQ1_M
404
441
  #define QUANT_R QUANT_R_IQ1_M
405
442
  #define A_TYPE block_iq1_m
443
+ #define A_TYPE_PACKED16 block_iq1_m_packed16
444
+ #define A_TYPE_PACKED32 block_iq1_m_packed32
406
445
  #endif
407
446
 
408
447
  #if defined(DATA_A_IQ1_S) || defined(DATA_A_IQ1_M)
@@ -541,7 +580,270 @@ const uint[1024] iq1s_grid_const = {
541
580
  0x55dd55df, 0x55d555d7, 0x5503550c, 0x557f5501, 0x5577557d, 0x55405575, 0x555d555f, 0x55555557
542
581
  };
543
582
 
583
+ // Same content as iq1s_grid_const except each 2-bit value is expanded to 4-bit
584
+ // and has 1 added to it (allows packed values to be extracted with & 0x0F0F0F0F
585
+ // and 0xF0F0F0F0).
586
+ const uint32_t[2048] iq1s_grid_gpu_const = {
587
+ 0x00000000, 0x00000002, 0x00000101, 0x00000200, 0x00000202, 0x00010001, 0x00010101, 0x00020000,
588
+ 0x00020002, 0x00020200, 0x00020202, 0x01000101, 0x01010001, 0x01010100, 0x01010102, 0x01020101,
589
+ 0x02000000, 0x02000002, 0x02000200, 0x02000202, 0x02010101, 0x02020000, 0x02020002, 0x02020200,
590
+ 0x02020202, 0x00000110, 0x00000111, 0x00010011, 0x00010110, 0x00010112, 0x00010211, 0x00010212,
591
+ 0x00020111, 0x01000011, 0x01000112, 0x01000211, 0x01010012, 0x01010111, 0x01010212, 0x01020011,
592
+ 0x01020110, 0x01020112, 0x01020210, 0x02000111, 0x02010011, 0x02010110, 0x02010112, 0x02020111,
593
+ 0x00000020, 0x00000022, 0x00000220, 0x00000222, 0x00010121, 0x00020020, 0x00020022, 0x00020220,
594
+ 0x00020222, 0x01000121, 0x01010021, 0x01010221, 0x01020120, 0x01020221, 0x02000020, 0x02000022,
595
+ 0x02000220, 0x02000222, 0x02010021, 0x02010121, 0x02010221, 0x02020020, 0x02020022, 0x02020220,
596
+ 0x02020222, 0x00011001, 0x00011100, 0x00011102, 0x00021101, 0x01001001, 0x01001201, 0x01011101,
597
+ 0x01011202, 0x01021100, 0x01021101, 0x02011001, 0x02011201, 0x02021101, 0x00001011, 0x00001110,
598
+ 0x00001111, 0x00001112, 0x00011111, 0x00011210, 0x00011212, 0x00021211, 0x01001010, 0x01001111,
599
+ 0x01001212, 0x01011010, 0x01011011, 0x01011110, 0x01011111, 0x01011112, 0x01011211, 0x01021010,
600
+ 0x01021012, 0x01021111, 0x01021210, 0x01021212, 0x02001011, 0x02011011, 0x02011111, 0x02011210,
601
+ 0x02011212, 0x02021011, 0x02021110, 0x02021111, 0x02021112, 0x02021211, 0x00011120, 0x00011221,
602
+ 0x01001021, 0x01001120, 0x01011020, 0x01011022, 0x01011121, 0x01011220, 0x01021020, 0x01021021,
603
+ 0x01021122, 0x01021221, 0x02001121, 0x02011021, 0x02011120, 0x02011221, 0x00002000, 0x00002002,
604
+ 0x00002200, 0x00002202, 0x00012101, 0x00022000, 0x00022002, 0x00022200, 0x00022202, 0x01002101,
605
+ 0x01012001, 0x01012102, 0x01022101, 0x02002000, 0x02002002, 0x02002200, 0x02002202, 0x02012101,
606
+ 0x02022000, 0x02022002, 0x02022200, 0x02022202, 0x00002111, 0x00012011, 0x00012110, 0x00012211,
607
+ 0x00022110, 0x00022111, 0x01002011, 0x01012010, 0x01012011, 0x01012111, 0x01022011, 0x01022110,
608
+ 0x01022211, 0x02012011, 0x02012110, 0x02012112, 0x02012211, 0x02022111, 0x00002020, 0x00002022,
609
+ 0x00002220, 0x00002222, 0x00012121, 0x00022020, 0x00022022, 0x00022220, 0x00022222, 0x01002121,
610
+ 0x01012021, 0x01012221, 0x01022021, 0x01022121, 0x02002020, 0x02002022, 0x02002121, 0x02002220,
611
+ 0x02002222, 0x02012121, 0x02022020, 0x02022022, 0x02022220, 0x02022222, 0x00110000, 0x00110001,
612
+ 0x00110100, 0x00110201, 0x00120100, 0x00120101, 0x01100001, 0x01100100, 0x01110000, 0x01110101,
613
+ 0x01110200, 0x01120001, 0x01120100, 0x01120101, 0x01120201, 0x02110001, 0x02110100, 0x02110102,
614
+ 0x02120001, 0x02120101, 0x00100011, 0x00100110, 0x00100112, 0x00100211, 0x00110010, 0x00110012,
615
+ 0x00110111, 0x00110210, 0x00120011, 0x00120110, 0x00120211, 0x01100111, 0x01100212, 0x01110010,
616
+ 0x01110011, 0x01110012, 0x01110110, 0x01110111, 0x01110112, 0x01110211, 0x01120010, 0x01120111,
617
+ 0x02100110, 0x02110012, 0x02110111, 0x02120011, 0x02120110, 0x00110021, 0x00110120, 0x00110122,
618
+ 0x00120121, 0x01100020, 0x01100122, 0x01100221, 0x01110022, 0x01110121, 0x01110220, 0x01110222,
619
+ 0x01120120, 0x01120122, 0x02100121, 0x02110021, 0x02110120, 0x02110122, 0x02120121, 0x00101001,
620
+ 0x00101102, 0x00101201, 0x00111100, 0x00111101, 0x00111200, 0x00111201, 0x00121001, 0x00121102,
621
+ 0x01101001, 0x01101101, 0x01101102, 0x01101200, 0x01101202, 0x01111001, 0x01111100, 0x01111101,
622
+ 0x01111102, 0x01111201, 0x01121002, 0x01121101, 0x01121200, 0x02101100, 0x02101201, 0x02111000,
623
+ 0x02111100, 0x02111101, 0x02111200, 0x02111201, 0x02111202, 0x02121001, 0x02121100, 0x02121101,
624
+ 0x02121201, 0x00101012, 0x00101111, 0x00101212, 0x00111011, 0x00111110, 0x00111111, 0x00111112,
625
+ 0x00111211, 0x00121010, 0x00121012, 0x00121111, 0x00121210, 0x00121212, 0x01101011, 0x01101110,
626
+ 0x01101111, 0x01101112, 0x01111011, 0x01111012, 0x01111110, 0x01111111, 0x01111112, 0x01111211,
627
+ 0x01111212, 0x01121011, 0x01121110, 0x01121111, 0x01121112, 0x01121211, 0x02101010, 0x02101012,
628
+ 0x02101110, 0x02101111, 0x02101210, 0x02101212, 0x02111010, 0x02111011, 0x02111110, 0x02111111,
629
+ 0x02111112, 0x02111211, 0x02111212, 0x02121010, 0x02121012, 0x02121111, 0x00101021, 0x00101120,
630
+ 0x00101121, 0x00101122, 0x00111121, 0x00111122, 0x00111220, 0x00111222, 0x00121021, 0x00121122,
631
+ 0x01101020, 0x01101022, 0x01101120, 0x01101121, 0x01101220, 0x01101222, 0x01111021, 0x01111121,
632
+ 0x01111122, 0x01111220, 0x01111221, 0x01121021, 0x01121120, 0x01121121, 0x01121220, 0x01121221,
633
+ 0x01121222, 0x02101122, 0x02101222, 0x02111022, 0x02111121, 0x02121120, 0x02121221, 0x00112001,
634
+ 0x00112102, 0x00122101, 0x01102001, 0x01102100, 0x01102102, 0x01102201, 0x01112000, 0x01112101,
635
+ 0x01112200, 0x01112202, 0x01122000, 0x01122001, 0x01122100, 0x01122102, 0x01122201, 0x02102101,
636
+ 0x02112001, 0x02112100, 0x02122101, 0x00112010, 0x00112012, 0x00112111, 0x00112212, 0x00122011,
637
+ 0x00122111, 0x01102012, 0x01102110, 0x01102111, 0x01102210, 0x01112011, 0x01112110, 0x01112111,
638
+ 0x01112112, 0x01112211, 0x01112212, 0x01122010, 0x01122111, 0x01122212, 0x02102211, 0x02112011,
639
+ 0x02112012, 0x02112111, 0x02112210, 0x02122011, 0x02122112, 0x02122211, 0x00102221, 0x00112122,
640
+ 0x00122120, 0x00122122, 0x01102120, 0x01102122, 0x01102221, 0x01112020, 0x01112022, 0x01112121,
641
+ 0x01112220, 0x01122021, 0x01122122, 0x01122221, 0x02102121, 0x02112021, 0x02112122, 0x02112222,
642
+ 0x00200000, 0x00200002, 0x00200200, 0x00200202, 0x00210101, 0x00220000, 0x00220002, 0x00220101,
643
+ 0x00220200, 0x00220202, 0x01200101, 0x01210001, 0x01210201, 0x01220001, 0x01220101, 0x02200000,
644
+ 0x02200002, 0x02200200, 0x02200202, 0x02210101, 0x02220000, 0x02220002, 0x02220101, 0x02220200,
645
+ 0x02220202, 0x00200111, 0x00210011, 0x00210110, 0x00210211, 0x00220111, 0x01200012, 0x01200110,
646
+ 0x01200211, 0x01210111, 0x01210210, 0x01210212, 0x01220011, 0x01220110, 0x01220111, 0x01220112,
647
+ 0x02200111, 0x02210010, 0x02210112, 0x02210211, 0x02220111, 0x00200021, 0x00200220, 0x00200222,
648
+ 0x00210021, 0x00210121, 0x00220020, 0x00220022, 0x00220220, 0x00220222, 0x01200121, 0x01210021,
649
+ 0x01210122, 0x01210221, 0x01220121, 0x02200021, 0x02200220, 0x02200222, 0x02210021, 0x02210121,
650
+ 0x02220020, 0x02220022, 0x02220220, 0x02220222, 0x00201101, 0x00211100, 0x00211102, 0x00211201,
651
+ 0x00221101, 0x01201100, 0x01201101, 0x01201102, 0x01201201, 0x01211002, 0x01211101, 0x01211200,
652
+ 0x01211202, 0x01221102, 0x02201101, 0x02211001, 0x02211100, 0x02211201, 0x02221001, 0x02221101,
653
+ 0x00201211, 0x00211111, 0x00221011, 0x00221211, 0x01201010, 0x01201111, 0x01201210, 0x01211011,
654
+ 0x01211110, 0x01211111, 0x01211211, 0x01221012, 0x01221111, 0x01221210, 0x02201211, 0x02211010,
655
+ 0x02211110, 0x02211111, 0x02211210, 0x02211212, 0x02221011, 0x02221110, 0x02221112, 0x02221211,
656
+ 0x00201121, 0x00211020, 0x00211022, 0x00211221, 0x00221121, 0x01201021, 0x01201221, 0x01211121,
657
+ 0x01221020, 0x01221021, 0x01221221, 0x02201120, 0x02201122, 0x02211020, 0x02211222, 0x00202000,
658
+ 0x00202002, 0x00202200, 0x00202202, 0x00212101, 0x00222000, 0x00222002, 0x00222200, 0x00222202,
659
+ 0x01202101, 0x01212001, 0x01212100, 0x01222101, 0x02202000, 0x02202002, 0x02202200, 0x02202202,
660
+ 0x02222000, 0x02222002, 0x02222200, 0x02222202, 0x00202211, 0x00212011, 0x00212110, 0x00212211,
661
+ 0x00222111, 0x01202112, 0x01202211, 0x01212012, 0x01212111, 0x01222011, 0x01222110, 0x01222112,
662
+ 0x01222211, 0x02202111, 0x02212010, 0x02212112, 0x02212211, 0x02222110, 0x02222111, 0x00202020,
663
+ 0x00202022, 0x00202220, 0x00202222, 0x00222020, 0x00222022, 0x00222220, 0x00222222, 0x01202121,
664
+ 0x01212021, 0x01212122, 0x01212221, 0x01222121, 0x02202020, 0x02202022, 0x02202220, 0x02202222,
665
+ 0x02212121, 0x02222020, 0x02222022, 0x02222220, 0x02222222, 0x10000101, 0x10010001, 0x10010102,
666
+ 0x10020101, 0x11000201, 0x11010002, 0x11010101, 0x11010200, 0x11010202, 0x11020001, 0x11020100,
667
+ 0x11020102, 0x12010100, 0x12010201, 0x12020001, 0x12020102, 0x10000010, 0x10000011, 0x10000110,
668
+ 0x10000112, 0x10000211, 0x10010012, 0x10010111, 0x10010112, 0x10010210, 0x10010212, 0x10020011,
669
+ 0x10020112, 0x10020211, 0x11000111, 0x11000210, 0x11000212, 0x11010011, 0x11010110, 0x11010111,
670
+ 0x11010112, 0x11010211, 0x11010212, 0x11020111, 0x11020210, 0x11020212, 0x12000011, 0x12000110,
671
+ 0x12000112, 0x12010010, 0x12010012, 0x12010111, 0x12020010, 0x12020011, 0x12020012, 0x10000121,
672
+ 0x10010021, 0x10010120, 0x10010122, 0x10020121, 0x11000021, 0x11010022, 0x11010121, 0x11010222,
673
+ 0x11020120, 0x11020221, 0x12000221, 0x12010120, 0x12020121, 0x10001001, 0x10011101, 0x10011201,
674
+ 0x10021201, 0x11001101, 0x11001200, 0x11001202, 0x11011001, 0x11011100, 0x11011101, 0x11011102,
675
+ 0x11021001, 0x11021002, 0x11021101, 0x11021200, 0x11021202, 0x12001001, 0x12001102, 0x12001201,
676
+ 0x12011000, 0x12011002, 0x12011101, 0x12021000, 0x12021001, 0x12021201, 0x10001011, 0x10001012,
677
+ 0x10001111, 0x10001212, 0x10011011, 0x10011110, 0x10011111, 0x10011112, 0x10011211, 0x10021010,
678
+ 0x10021111, 0x10021212, 0x11001011, 0x11001110, 0x11001111, 0x11001112, 0x11001211, 0x11011010,
679
+ 0x11011011, 0x11011110, 0x11011111, 0x11011112, 0x11011210, 0x11011211, 0x11021011, 0x11021110,
680
+ 0x11021111, 0x11021112, 0x11021211, 0x12001012, 0x12001110, 0x12001111, 0x12001210, 0x12011011,
681
+ 0x12011110, 0x12011111, 0x12011112, 0x12011211, 0x12011212, 0x12021111, 0x12021210, 0x12021212,
682
+ 0x10001021, 0x10001121, 0x10001221, 0x10011120, 0x10011121, 0x10011220, 0x10011222, 0x10021021,
683
+ 0x10021120, 0x10021221, 0x11001020, 0x11001022, 0x11001121, 0x11001220, 0x11011020, 0x11011021,
684
+ 0x11011022, 0x11011121, 0x11011122, 0x11011221, 0x11021022, 0x11021121, 0x11021220, 0x12001021,
685
+ 0x12001121, 0x12001222, 0x12011120, 0x12011121, 0x12021021, 0x12021120, 0x12021122, 0x10002101,
686
+ 0x10012001, 0x10012101, 0x10012202, 0x10022101, 0x11002002, 0x11002201, 0x11012000, 0x11012101,
687
+ 0x11012200, 0x11022001, 0x11022100, 0x11022102, 0x11022201, 0x12002101, 0x12012001, 0x12012100,
688
+ 0x12012102, 0x12012201, 0x12022101, 0x10002011, 0x10002111, 0x10002112, 0x10002212, 0x10012010,
689
+ 0x10012110, 0x10012111, 0x10012210, 0x10022011, 0x10022110, 0x10022112, 0x11002010, 0x11002111,
690
+ 0x11002212, 0x11012011, 0x11012012, 0x11012110, 0x11012111, 0x11012112, 0x11012211, 0x11022010,
691
+ 0x11022012, 0x11022111, 0x11022112, 0x11022212, 0x12002112, 0x12002211, 0x12012012, 0x12012111,
692
+ 0x12012112, 0x12012210, 0x12022011, 0x12022110, 0x12022112, 0x12022211, 0x10012122, 0x11002120,
693
+ 0x11002122, 0x11002221, 0x11012121, 0x11012220, 0x11012222, 0x11022120, 0x11022221, 0x12012120,
694
+ 0x12022121, 0x10100001, 0x10100100, 0x10100101, 0x10100102, 0x10100201, 0x10110002, 0x10110101,
695
+ 0x10110202, 0x10120001, 0x10120100, 0x10120201, 0x11100000, 0x11100101, 0x11100200, 0x11110001,
696
+ 0x11110100, 0x11110101, 0x11110102, 0x11110201, 0x11120101, 0x11120200, 0x12100102, 0x12100201,
697
+ 0x12110101, 0x12110200, 0x12120000, 0x12120001, 0x12120102, 0x12120201, 0x10100111, 0x10100210,
698
+ 0x10100211, 0x10100212, 0x10110011, 0x10110110, 0x10110111, 0x10110112, 0x10110210, 0x10110211,
699
+ 0x10120010, 0x10120111, 0x10120112, 0x10120210, 0x10120212, 0x11100011, 0x11100110, 0x11100111,
700
+ 0x11100112, 0x11100211, 0x11110010, 0x11110011, 0x11110012, 0x11110110, 0x11110111, 0x11110112,
701
+ 0x11110210, 0x11110211, 0x11110212, 0x11120011, 0x11120110, 0x11120111, 0x11120112, 0x11120211,
702
+ 0x12100012, 0x12100111, 0x12110011, 0x12110110, 0x12110111, 0x12110112, 0x12110211, 0x12120010,
703
+ 0x12120111, 0x12120212, 0x10100021, 0x10100122, 0x10110022, 0x10110121, 0x10110222, 0x10120021,
704
+ 0x10120120, 0x11100022, 0x11100121, 0x11100222, 0x11110021, 0x11110120, 0x11110121, 0x11110122,
705
+ 0x11110221, 0x11120022, 0x11120121, 0x12100121, 0x12110020, 0x12110022, 0x12110121, 0x12110221,
706
+ 0x12110222, 0x12120120, 0x10101100, 0x10101101, 0x10111001, 0x10111100, 0x10111101, 0x10111102,
707
+ 0x10111200, 0x10111201, 0x10121001, 0x10121101, 0x10121200, 0x10121202, 0x11101001, 0x11101100,
708
+ 0x11101101, 0x11101102, 0x11101201, 0x11101202, 0x11111000, 0x11111001, 0x11111100, 0x11111101,
709
+ 0x11111102, 0x11111200, 0x11111201, 0x11111202, 0x11121001, 0x11121002, 0x11121100, 0x11121101,
710
+ 0x11121102, 0x11121201, 0x12101000, 0x12101200, 0x12101202, 0x12111001, 0x12111100, 0x12111101,
711
+ 0x12111102, 0x12111201, 0x12121001, 0x12121100, 0x12121101, 0x12121202, 0x10101011, 0x10101012,
712
+ 0x10101110, 0x10101111, 0x10101112, 0x10101211, 0x10111010, 0x10111011, 0x10111012, 0x10111110,
713
+ 0x10111111, 0x10111112, 0x10111211, 0x10111212, 0x10121011, 0x10121110, 0x10121111, 0x10121112,
714
+ 0x10121211, 0x11101010, 0x11101011, 0x11101012, 0x11101110, 0x11101111, 0x11101112, 0x11101210,
715
+ 0x11101211, 0x11111010, 0x11111011, 0x11111012, 0x11111110, 0x11111111, 0x11111112, 0x11111210,
716
+ 0x11111211, 0x11111212, 0x11121010, 0x11121011, 0x11121110, 0x11121111, 0x11121112, 0x11121210,
717
+ 0x11121211, 0x11121212, 0x12101011, 0x12101110, 0x12101111, 0x12101211, 0x12101212, 0x12111010,
718
+ 0x12111011, 0x12111110, 0x12111111, 0x12111112, 0x12111210, 0x12111211, 0x12121011, 0x12121110,
719
+ 0x12121111, 0x12121112, 0x12121211, 0x10101020, 0x10101021, 0x10101022, 0x10101120, 0x10101122,
720
+ 0x10101220, 0x10101221, 0x10111021, 0x10111120, 0x10111121, 0x10111220, 0x10111221, 0x10121020,
721
+ 0x10121021, 0x10121022, 0x10121120, 0x10121121, 0x10121122, 0x10121220, 0x10121221, 0x11101021,
722
+ 0x11101121, 0x11101122, 0x11101220, 0x11101221, 0x11101222, 0x11111020, 0x11111021, 0x11111022,
723
+ 0x11111120, 0x11111121, 0x11111122, 0x11111220, 0x11111221, 0x11111222, 0x11121021, 0x11121120,
724
+ 0x11121121, 0x11121221, 0x12101022, 0x12101121, 0x12101122, 0x12101220, 0x12101221, 0x12101222,
725
+ 0x12111021, 0x12111121, 0x12111222, 0x12121022, 0x12121121, 0x12121122, 0x12121220, 0x12121221,
726
+ 0x10102100, 0x10102101, 0x10102102, 0x10102201, 0x10112000, 0x10112101, 0x10112200, 0x10122001,
727
+ 0x10122202, 0x11102101, 0x11102200, 0x11102202, 0x11112001, 0x11112100, 0x11112101, 0x11112102,
728
+ 0x11112200, 0x11112201, 0x11122000, 0x11122002, 0x11122100, 0x11122101, 0x12102002, 0x12102201,
729
+ 0x12112000, 0x12112002, 0x12112101, 0x12112200, 0x12122001, 0x12122201, 0x10102011, 0x10102012,
730
+ 0x10102111, 0x10102212, 0x10112011, 0x10112110, 0x10112111, 0x10112112, 0x10112211, 0x10122111,
731
+ 0x11102011, 0x11102110, 0x11102111, 0x11102112, 0x11102211, 0x11112010, 0x11112011, 0x11112012,
732
+ 0x11112110, 0x11112111, 0x11112112, 0x11112210, 0x11112211, 0x11112212, 0x11122011, 0x11122110,
733
+ 0x11122111, 0x11122112, 0x11122211, 0x12102011, 0x12102111, 0x12102211, 0x12112011, 0x12112110,
734
+ 0x12112111, 0x12112112, 0x12112210, 0x12112211, 0x12122111, 0x10102120, 0x10102220, 0x10112121,
735
+ 0x10112222, 0x10122020, 0x10122121, 0x10122122, 0x10122221, 0x11102121, 0x11102220, 0x11102221,
736
+ 0x11112021, 0x11112121, 0x11112122, 0x11112220, 0x11112221, 0x11122022, 0x11122121, 0x11122220,
737
+ 0x11122222, 0x12102021, 0x12102222, 0x12112022, 0x12112121, 0x12112122, 0x12112220, 0x12112222,
738
+ 0x12122021, 0x10200101, 0x10210100, 0x10210102, 0x10210201, 0x10220101, 0x11200100, 0x11210000,
739
+ 0x11210101, 0x11210102, 0x11210200, 0x11210202, 0x11220001, 0x11220100, 0x11220102, 0x11220201,
740
+ 0x12200001, 0x12210102, 0x12220101, 0x10200011, 0x10200110, 0x10200112, 0x10200211, 0x10210012,
741
+ 0x10210111, 0x10220011, 0x10220012, 0x10220112, 0x10220211, 0x11200111, 0x11200211, 0x11210011,
742
+ 0x11210111, 0x11210112, 0x11210211, 0x11220111, 0x11220112, 0x11220212, 0x12200110, 0x12200212,
743
+ 0x12210012, 0x12210111, 0x12220011, 0x12220112, 0x12220211, 0x10210021, 0x10210122, 0x10210221,
744
+ 0x11200020, 0x11200021, 0x11200122, 0x11210121, 0x11210122, 0x11210220, 0x11220020, 0x12200121,
745
+ 0x12210021, 0x12210122, 0x12220121, 0x10211001, 0x10211002, 0x10211101, 0x10211102, 0x10211202,
746
+ 0x10221001, 0x10221102, 0x10221201, 0x11201000, 0x11201002, 0x11201101, 0x11201200, 0x11201202,
747
+ 0x11211001, 0x11211100, 0x11211101, 0x11211102, 0x11211201, 0x11211202, 0x11221000, 0x11221002,
748
+ 0x11221101, 0x12201100, 0x12201101, 0x12201201, 0x12211000, 0x12211002, 0x12211100, 0x12211101,
749
+ 0x12211102, 0x12211200, 0x12211202, 0x12221001, 0x12221100, 0x12221201, 0x10201111, 0x10201210,
750
+ 0x10201212, 0x10211011, 0x10211111, 0x10211112, 0x10211211, 0x11201110, 0x11201111, 0x11201112,
751
+ 0x11201211, 0x11211010, 0x11211011, 0x11211110, 0x11211111, 0x11211112, 0x11211211, 0x11221011,
752
+ 0x11221110, 0x11221111, 0x11221112, 0x11221211, 0x12201112, 0x12201211, 0x12201212, 0x12211011,
753
+ 0x12211111, 0x12211112, 0x12211211, 0x12211212, 0x12221012, 0x12221111, 0x12221112, 0x12221210,
754
+ 0x10201022, 0x10201221, 0x10211121, 0x10221020, 0x10221122, 0x10221220, 0x10221221, 0x11201020,
755
+ 0x11201121, 0x11201220, 0x11201222, 0x11211021, 0x11211120, 0x11211121, 0x11211122, 0x11211220,
756
+ 0x11211222, 0x11221020, 0x11221121, 0x11221220, 0x12201020, 0x12201022, 0x12201121, 0x12201222,
757
+ 0x12211120, 0x12211122, 0x12211220, 0x12211221, 0x12221020, 0x12221120, 0x12221122, 0x12221222,
758
+ 0x10212102, 0x10212201, 0x10222101, 0x11202001, 0x11212002, 0x11212101, 0x11212202, 0x11222001,
759
+ 0x11222201, 0x12202101, 0x12212001, 0x12212200, 0x12222102, 0x10202011, 0x10202110, 0x10212010,
760
+ 0x10212111, 0x10222011, 0x10222110, 0x10222112, 0x10222211, 0x11202010, 0x11202011, 0x11202111,
761
+ 0x11202112, 0x11202210, 0x11212011, 0x11212110, 0x11212111, 0x11212112, 0x11212211, 0x11222010,
762
+ 0x11222111, 0x11222212, 0x12202012, 0x12202110, 0x12202212, 0x12212111, 0x12222011, 0x12222110,
763
+ 0x12222111, 0x12222211, 0x10212021, 0x10212122, 0x10212220, 0x11202021, 0x11202120, 0x11202221,
764
+ 0x11212020, 0x11212121, 0x11212220, 0x11212222, 0x11222120, 0x11222121, 0x11222221, 0x12202122,
765
+ 0x12212120, 0x12212220, 0x12212222, 0x12222122, 0x20000000, 0x20000002, 0x20000200, 0x20000202,
766
+ 0x20020000, 0x20020002, 0x20020200, 0x20020202, 0x21000101, 0x21010000, 0x21010001, 0x21010100,
767
+ 0x21010102, 0x21010201, 0x21020101, 0x22000000, 0x22000002, 0x22000200, 0x22000202, 0x22010101,
768
+ 0x22020000, 0x22020002, 0x22020200, 0x22020202, 0x20000111, 0x20010011, 0x20010110, 0x20010112,
769
+ 0x20010211, 0x20020111, 0x21000011, 0x21000110, 0x21000211, 0x21010010, 0x21010012, 0x21010111,
770
+ 0x21010112, 0x21010210, 0x21010211, 0x21020110, 0x21020112, 0x21020211, 0x22000111, 0x22000211,
771
+ 0x22010110, 0x22010112, 0x22010211, 0x22020111, 0x20000020, 0x20000022, 0x20000220, 0x20000222,
772
+ 0x20010121, 0x20020020, 0x20020022, 0x20020220, 0x20020222, 0x21010021, 0x21010120, 0x21010221,
773
+ 0x21020121, 0x22000020, 0x22000022, 0x22000220, 0x22000222, 0x22010121, 0x22020020, 0x22020022,
774
+ 0x22020220, 0x22020222, 0x20011100, 0x20011201, 0x21001001, 0x21001100, 0x21011001, 0x21011101,
775
+ 0x21011202, 0x21021001, 0x21021100, 0x21021201, 0x22011100, 0x22011201, 0x20001011, 0x20001211,
776
+ 0x20011012, 0x20011111, 0x20011212, 0x20021112, 0x20021211, 0x21001010, 0x21001011, 0x21001111,
777
+ 0x21001210, 0x21011011, 0x21011110, 0x21011111, 0x21011112, 0x21011211, 0x21011212, 0x21021111,
778
+ 0x21021112, 0x21021210, 0x21021212, 0x22001011, 0x22001110, 0x22001112, 0x22001211, 0x22011010,
779
+ 0x22011012, 0x22011111, 0x22011210, 0x22021112, 0x20011021, 0x20011122, 0x20011221, 0x20021121,
780
+ 0x21001021, 0x21001120, 0x21001221, 0x21001222, 0x21011020, 0x21011121, 0x21011221, 0x21011222,
781
+ 0x21021021, 0x21021122, 0x21021222, 0x22001121, 0x22011021, 0x22011222, 0x22021120, 0x20002000,
782
+ 0x20002002, 0x20002200, 0x20002202, 0x20012101, 0x20022000, 0x20022002, 0x20022200, 0x20022202,
783
+ 0x21002001, 0x21002101, 0x21012001, 0x21012100, 0x21012201, 0x21022101, 0x21022201, 0x22002000,
784
+ 0x22002002, 0x22002200, 0x22002202, 0x22012101, 0x22022000, 0x22022002, 0x22022200, 0x22022202,
785
+ 0x20002111, 0x20002112, 0x20012011, 0x20012110, 0x20012112, 0x20022111, 0x21002011, 0x21002110,
786
+ 0x21002112, 0x21002211, 0x21012010, 0x21012012, 0x21012111, 0x21012212, 0x21022011, 0x21022110,
787
+ 0x22002111, 0x22012112, 0x22012211, 0x22022111, 0x20002020, 0x20002022, 0x20002220, 0x20002222,
788
+ 0x20012121, 0x20022020, 0x20022022, 0x20022220, 0x20022222, 0x21002121, 0x21012021, 0x21012120,
789
+ 0x21012122, 0x22002020, 0x22002022, 0x22002220, 0x22002222, 0x22012121, 0x22022020, 0x22022022,
790
+ 0x22022220, 0x22022222, 0x20100101, 0x20110001, 0x20110102, 0x20110200, 0x20110201, 0x20120101,
791
+ 0x21100001, 0x21100102, 0x21100201, 0x21110101, 0x21110200, 0x21110202, 0x21120201, 0x21120202,
792
+ 0x22100101, 0x22110001, 0x22110100, 0x22110102, 0x22110201, 0x22120101, 0x20100011, 0x20100110,
793
+ 0x20100112, 0x20100211, 0x20110010, 0x20110111, 0x20110210, 0x20110212, 0x20120011, 0x20120110,
794
+ 0x20120112, 0x20120211, 0x21100010, 0x21100111, 0x21110010, 0x21110011, 0x21110110, 0x21110111,
795
+ 0x21110112, 0x21110211, 0x21120012, 0x21120111, 0x22100110, 0x22100112, 0x22110012, 0x22110111,
796
+ 0x22110210, 0x22120011, 0x22120110, 0x22120112, 0x22120211, 0x20100121, 0x20110021, 0x20110120,
797
+ 0x20110221, 0x20120121, 0x21100120, 0x21100122, 0x21100221, 0x21110020, 0x21110022, 0x21110121,
798
+ 0x21110220, 0x21120122, 0x21120221, 0x22100121, 0x22110120, 0x22110122, 0x22120221, 0x20101001,
799
+ 0x20101100, 0x20101102, 0x20111000, 0x20111101, 0x20111200, 0x20121102, 0x21101000, 0x21101202,
800
+ 0x21111001, 0x21111100, 0x21111101, 0x21111102, 0x21111200, 0x21111201, 0x21121000, 0x21121001,
801
+ 0x21121002, 0x21121101, 0x22101100, 0x22101102, 0x22111002, 0x22111100, 0x22111101, 0x22111200,
802
+ 0x22121001, 0x22121201, 0x20101010, 0x20101111, 0x20101210, 0x20101212, 0x20111010, 0x20111011,
803
+ 0x20111110, 0x20111111, 0x20111112, 0x20111211, 0x20121011, 0x20121111, 0x20121211, 0x20121212,
804
+ 0x21101011, 0x21101110, 0x21101111, 0x21101112, 0x21101211, 0x21111010, 0x21111011, 0x21111012,
805
+ 0x21111110, 0x21111111, 0x21111112, 0x21111210, 0x21111211, 0x21111212, 0x21121011, 0x21121110,
806
+ 0x21121111, 0x21121112, 0x21121211, 0x22101011, 0x22101111, 0x22101210, 0x22111011, 0x22111012,
807
+ 0x22111110, 0x22111111, 0x22111112, 0x22111211, 0x22111212, 0x22121010, 0x22121012, 0x22121111,
808
+ 0x22121210, 0x22121212, 0x20101021, 0x20101120, 0x20111020, 0x20111121, 0x20111221, 0x20121020,
809
+ 0x20121122, 0x20121221, 0x21101121, 0x21101220, 0x21101221, 0x21111021, 0x21111022, 0x21111121,
810
+ 0x21111122, 0x21111221, 0x21121121, 0x21121220, 0x22101022, 0x22101120, 0x22101221, 0x22101222,
811
+ 0x22111022, 0x22111120, 0x22111121, 0x22121120, 0x22121122, 0x22121221, 0x20102101, 0x20112102,
812
+ 0x20112201, 0x20122101, 0x21102001, 0x21102102, 0x21112000, 0x21112002, 0x21112101, 0x21112102,
813
+ 0x21112202, 0x21122100, 0x21122101, 0x22102101, 0x22112001, 0x22112102, 0x22112201, 0x22122101,
814
+ 0x20102110, 0x20102112, 0x20102211, 0x20112010, 0x20112012, 0x20112111, 0x20112210, 0x20112212,
815
+ 0x20122010, 0x20122011, 0x20122110, 0x20122112, 0x21102010, 0x21102012, 0x21102111, 0x21102210,
816
+ 0x21102212, 0x21112011, 0x21112110, 0x21112111, 0x21112112, 0x21112211, 0x21122012, 0x21122111,
817
+ 0x21122112, 0x21122212, 0x22102011, 0x22102110, 0x22112010, 0x22112012, 0x22112111, 0x22112212,
818
+ 0x22122011, 0x22122112, 0x20102121, 0x20112121, 0x20122121, 0x21102120, 0x21102122, 0x21102221,
819
+ 0x21112020, 0x21112121, 0x21112220, 0x21122021, 0x22102121, 0x22112021, 0x22112120, 0x22112121,
820
+ 0x22112122, 0x20200000, 0x20200002, 0x20200200, 0x20200202, 0x20210101, 0x20220000, 0x20220002,
821
+ 0x20220200, 0x20220202, 0x21200101, 0x21210001, 0x21210100, 0x21210102, 0x21210201, 0x22200000,
822
+ 0x22200002, 0x22200200, 0x22200202, 0x22210101, 0x22220000, 0x22220002, 0x22220200, 0x22220202,
823
+ 0x20200111, 0x20200211, 0x20210011, 0x20210110, 0x20210112, 0x20210211, 0x20210212, 0x21200112,
824
+ 0x21200211, 0x21210011, 0x21210111, 0x21210210, 0x21210212, 0x21220011, 0x21220110, 0x22200111,
825
+ 0x22210010, 0x22210012, 0x22210112, 0x22210211, 0x20200022, 0x20200220, 0x20200222, 0x20210020,
826
+ 0x20210221, 0x20220022, 0x20220220, 0x20220222, 0x21200121, 0x21210021, 0x21210122, 0x21210221,
827
+ 0x21220121, 0x22200020, 0x22200022, 0x22200220, 0x22200222, 0x22210121, 0x22220020, 0x22220022,
828
+ 0x22220220, 0x22220222, 0x20211201, 0x20221101, 0x21201001, 0x21201100, 0x21211000, 0x21211100,
829
+ 0x21211101, 0x21211200, 0x21211202, 0x21221001, 0x21221101, 0x21221102, 0x21221200, 0x21221201,
830
+ 0x22201101, 0x20201112, 0x20201211, 0x20211010, 0x20211012, 0x20211111, 0x20211210, 0x20221112,
831
+ 0x20221211, 0x21201012, 0x21201111, 0x21211011, 0x21211110, 0x21211111, 0x21211112, 0x21211211,
832
+ 0x21221111, 0x21221212, 0x22201011, 0x22201110, 0x22201111, 0x22201112, 0x22201211, 0x22211012,
833
+ 0x22211111, 0x22211210, 0x20201121, 0x20211021, 0x20211122, 0x20211222, 0x20221021, 0x20221121,
834
+ 0x21201120, 0x21201122, 0x21201222, 0x21211022, 0x21211121, 0x21211122, 0x21211220, 0x21221020,
835
+ 0x21221022, 0x22201122, 0x22211020, 0x22211121, 0x22211122, 0x22211221, 0x22221021, 0x22221120,
836
+ 0x22221122, 0x20202000, 0x20202002, 0x20202200, 0x20202202, 0x20222000, 0x20222002, 0x20222200,
837
+ 0x20222202, 0x21212001, 0x21212100, 0x21212102, 0x21212201, 0x22202000, 0x22202002, 0x22202200,
838
+ 0x22202202, 0x22212101, 0x22222000, 0x22222002, 0x22222200, 0x22222202, 0x20202111, 0x20212110,
839
+ 0x20212211, 0x20222011, 0x20222111, 0x21202011, 0x21212010, 0x21212111, 0x21212212, 0x21222011,
840
+ 0x21222112, 0x21222211, 0x22212010, 0x22212112, 0x20202020, 0x20202022, 0x20202220, 0x20202222,
841
+ 0x20222020, 0x20222022, 0x20222220, 0x20222222, 0x21212021, 0x21212120, 0x21212122, 0x22202020,
842
+ 0x22202022, 0x22202220, 0x22202222, 0x22212121, 0x22222020, 0x22222022, 0x22222220, 0x22222222,
843
+ };
844
+
544
845
  shared uint16_t iq1s_grid[2048];
846
+ shared uint32_t iq1s_grid_gpu[2048];
545
847
 
546
848
  #define NEEDS_INIT_IQ_SHMEM
547
849
  void init_iq_shmem(uvec3 wgsize)
@@ -555,6 +857,12 @@ void init_iq_shmem(uvec3 wgsize)
555
857
  iq1s_grid[2*idx+1] = g.y;
556
858
  }
557
859
  }
860
+ [[unroll]] for (uint i = 0; i < iq1s_grid_gpu_const.length(); i += wgsize.x) {
861
+ uint idx = i + gl_LocalInvocationIndex.x;
862
+ if (iq1s_grid_gpu_const.length() % wgsize.x == 0 || idx < iq1s_grid_gpu_const.length()) {
863
+ iq1s_grid_gpu[idx] = iq1s_grid_gpu_const[idx];
864
+ }
865
+ }
558
866
  barrier();
559
867
  }
560
868
  #endif
@@ -1326,10 +1634,28 @@ struct block_iq4_xs
1326
1634
  uint8_t qs[QUANT_K_IQ4_XS/2];
1327
1635
  };
1328
1636
 
1637
+ struct block_iq4_xs_packed16
1638
+ {
1639
+ float16_t d;
1640
+ uint16_t scales_h;
1641
+ uint16_t scales_l[QUANT_K_IQ4_XS/128];
1642
+ uint16_t qs[QUANT_K_IQ4_XS/4];
1643
+ };
1644
+
1645
+ struct block_iq4_xs_packed32
1646
+ {
1647
+ float16_t d;
1648
+ uint16_t scales_h;
1649
+ uint32_t scales_l;
1650
+ uint32_t qs[QUANT_K_IQ4_XS/8];
1651
+ };
1652
+
1329
1653
  #if defined(DATA_A_IQ4_XS)
1330
1654
  #define QUANT_K QUANT_K_IQ4_XS
1331
1655
  #define QUANT_R QUANT_R_IQ4_XS
1332
1656
  #define A_TYPE block_iq4_xs
1657
+ #define A_TYPE_PACKED16 block_iq4_xs_packed16
1658
+ #define A_TYPE_PACKED32 block_iq4_xs_packed32
1333
1659
  #endif
1334
1660
 
1335
1661
  #define QUANT_K_IQ4_NL 32
@@ -1363,18 +1689,11 @@ struct block_mxfp4
1363
1689
  uint8_t qs[QUANT_K_MXFP4/2];
1364
1690
  };
1365
1691
 
1366
- //struct block_mxfp4_packed16
1367
- //{
1368
- // uint8_t e;
1369
- // uint16_t qs[QUANT_K_MXFP4/2/2];
1370
- //};
1371
-
1372
1692
  #if defined(DATA_A_MXFP4)
1373
1693
  #define QUANT_K QUANT_K_MXFP4
1374
1694
  #define QUANT_R QUANT_R_MXFP4
1375
1695
  #define QUANT_AUXF 1
1376
1696
  #define A_TYPE block_mxfp4
1377
- //#define A_TYPE_PACKED16 block_mxfp4_packed16
1378
1697
  #endif
1379
1698
 
1380
1699
  #if defined(DATA_A_IQ4_NL) || defined(DATA_A_IQ4_XS)
@@ -1397,12 +1716,12 @@ void init_iq_shmem(uvec3 wgsize)
1397
1716
  #endif
1398
1717
 
1399
1718
  #if defined(DATA_A_MXFP4)
1400
- const FLOAT_TYPE kvalues_mxfp4_const[16] = {
1401
- FLOAT_TYPE(0.0f), FLOAT_TYPE(0.5f), FLOAT_TYPE(1.0f), FLOAT_TYPE(1.5f), FLOAT_TYPE(2.0f), FLOAT_TYPE(3.0f), FLOAT_TYPE(4.0f), FLOAT_TYPE(6.0f),
1402
- FLOAT_TYPE(-0.0f), FLOAT_TYPE(-0.5f), FLOAT_TYPE(-1.0f), FLOAT_TYPE(-1.5f), FLOAT_TYPE(-2.0f), FLOAT_TYPE(-3.0f), FLOAT_TYPE(-4.0f), FLOAT_TYPE(-6.0f)
1719
+ const int8_t kvalues_mxfp4_const[16] = {
1720
+ int8_t(0), int8_t(1), int8_t(2), int8_t(3), int8_t(4), int8_t(6), int8_t(8), int8_t(12),
1721
+ int8_t(0), int8_t(-1), int8_t(-2), int8_t(-3), int8_t(-4), int8_t(-6), int8_t(-8), int8_t(-12),
1403
1722
  };
1404
1723
 
1405
- shared FLOAT_TYPE kvalues_mxfp4[16];
1724
+ shared int8_t kvalues_mxfp4[16];
1406
1725
 
1407
1726
  #define NEEDS_INIT_IQ_SHMEM
1408
1727
  void init_iq_shmem(uvec3 wgsize)