whispercpp 1.3.5 → 1.3.6

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 (610) hide show
  1. checksums.yaml +4 -4
  2. data/LICENSE +1 -1
  3. data/README.md +99 -2
  4. data/ext/extconf.rb +1 -0
  5. data/ext/ruby_whisper.c +20 -4
  6. data/ext/ruby_whisper.h +30 -2
  7. data/ext/ruby_whisper_context.c +216 -124
  8. data/ext/ruby_whisper_context_params.c +163 -0
  9. data/ext/ruby_whisper_model.c +0 -1
  10. data/ext/ruby_whisper_params.c +0 -1
  11. data/ext/ruby_whisper_segment.c +0 -1
  12. data/ext/ruby_whisper_token.c +29 -9
  13. data/ext/ruby_whisper_transcribe.cpp +4 -1
  14. data/ext/ruby_whisper_vad_context.c +48 -1
  15. data/ext/ruby_whisper_vad_context_detect.cpp +6 -5
  16. data/ext/ruby_whisper_vad_params.c +0 -1
  17. data/ext/ruby_whisper_vad_segment.c +0 -1
  18. data/ext/ruby_whisper_vad_segments.c +0 -1
  19. data/ext/sources/CMakeLists.txt +1 -1
  20. data/ext/sources/bindings/javascript/package.json +1 -1
  21. data/ext/sources/cmake/whisper-config.cmake.in +5 -40
  22. data/ext/sources/examples/bench/bench.cpp +23 -18
  23. data/ext/sources/examples/cli/cli.cpp +8 -0
  24. data/ext/sources/examples/common-ggml.cpp +2 -0
  25. data/ext/sources/examples/miniaudio.h +4507 -2131
  26. data/ext/sources/examples/server/server.cpp +18 -4
  27. data/ext/sources/examples/talk-llama/CMakeLists.txt +3 -2
  28. data/ext/sources/examples/talk-llama/llama-adapter.cpp +7 -13
  29. data/ext/sources/examples/talk-llama/llama-adapter.h +4 -3
  30. data/ext/sources/examples/talk-llama/llama-arch.cpp +335 -17
  31. data/ext/sources/examples/talk-llama/llama-arch.h +42 -0
  32. data/ext/sources/examples/talk-llama/llama-batch.cpp +3 -1
  33. data/ext/sources/examples/talk-llama/llama-chat.cpp +21 -1
  34. data/ext/sources/examples/talk-llama/llama-chat.h +1 -0
  35. data/ext/sources/examples/talk-llama/llama-context.cpp +508 -520
  36. data/ext/sources/examples/talk-llama/llama-context.h +27 -28
  37. data/ext/sources/examples/talk-llama/llama-cparams.h +5 -0
  38. data/ext/sources/examples/talk-llama/llama-ext.h +12 -0
  39. data/ext/sources/examples/talk-llama/llama-grammar.cpp +8 -8
  40. data/ext/sources/examples/talk-llama/llama-graph.cpp +583 -130
  41. data/ext/sources/examples/talk-llama/llama-graph.h +131 -10
  42. data/ext/sources/examples/talk-llama/llama-hparams.cpp +57 -40
  43. data/ext/sources/examples/talk-llama/llama-hparams.h +79 -10
  44. data/ext/sources/examples/talk-llama/llama-impl.cpp +4 -4
  45. data/ext/sources/examples/talk-llama/llama-impl.h +13 -1
  46. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +3 -1
  47. data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +274 -89
  48. data/ext/sources/examples/talk-llama/llama-kv-cache.h +2 -3
  49. data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.cpp +275 -0
  50. data/ext/sources/examples/talk-llama/llama-memory-hybrid-iswa.h +140 -0
  51. data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +11 -13
  52. data/ext/sources/examples/talk-llama/llama-mmap.cpp +28 -11
  53. data/ext/sources/examples/talk-llama/llama-model-loader.cpp +527 -119
  54. data/ext/sources/examples/talk-llama/llama-model-loader.h +35 -5
  55. data/ext/sources/examples/talk-llama/llama-model-saver.cpp +60 -46
  56. data/ext/sources/examples/talk-llama/llama-model-saver.h +5 -2
  57. data/ext/sources/examples/talk-llama/llama-model.cpp +1365 -647
  58. data/ext/sources/examples/talk-llama/llama-model.h +72 -19
  59. data/ext/sources/examples/talk-llama/llama-quant.cpp +578 -346
  60. data/ext/sources/examples/talk-llama/{llama-sampling.cpp → llama-sampler.cpp} +190 -76
  61. data/ext/sources/examples/talk-llama/{llama-sampling.h → llama-sampler.h} +0 -2
  62. data/ext/sources/examples/talk-llama/llama-vocab.cpp +118 -48
  63. data/ext/sources/examples/talk-llama/llama-vocab.h +5 -0
  64. data/ext/sources/examples/talk-llama/llama.cpp +76 -22
  65. data/ext/sources/examples/talk-llama/llama.h +63 -30
  66. data/ext/sources/examples/talk-llama/models/afmoe.cpp +2 -3
  67. data/ext/sources/examples/talk-llama/models/apertus.cpp +3 -3
  68. data/ext/sources/examples/talk-llama/models/arcee.cpp +3 -3
  69. data/ext/sources/examples/talk-llama/models/arctic.cpp +4 -5
  70. data/ext/sources/examples/talk-llama/models/baichuan.cpp +4 -3
  71. data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +1 -2
  72. data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +3 -5
  73. data/ext/sources/examples/talk-llama/models/bert.cpp +13 -7
  74. data/ext/sources/examples/talk-llama/models/bitnet.cpp +9 -24
  75. data/ext/sources/examples/talk-llama/models/bloom.cpp +2 -2
  76. data/ext/sources/examples/talk-llama/models/chameleon.cpp +3 -3
  77. data/ext/sources/examples/talk-llama/models/chatglm.cpp +2 -2
  78. data/ext/sources/examples/talk-llama/models/codeshell.cpp +3 -3
  79. data/ext/sources/examples/talk-llama/models/cogvlm.cpp +3 -3
  80. data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +2 -2
  81. data/ext/sources/examples/talk-llama/models/command-r.cpp +2 -2
  82. data/ext/sources/examples/talk-llama/models/dbrx.cpp +4 -5
  83. data/ext/sources/examples/talk-llama/models/deci.cpp +3 -3
  84. data/ext/sources/examples/talk-llama/models/deepseek.cpp +4 -6
  85. data/ext/sources/examples/talk-llama/models/deepseek2.cpp +24 -21
  86. data/ext/sources/examples/talk-llama/models/delta-net-base.cpp +445 -0
  87. data/ext/sources/examples/talk-llama/models/dots1.cpp +4 -6
  88. data/ext/sources/examples/talk-llama/models/dream.cpp +3 -3
  89. data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +4 -6
  90. data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +3 -3
  91. data/ext/sources/examples/talk-llama/models/eurobert.cpp +97 -0
  92. data/ext/sources/examples/talk-llama/models/exaone-moe.cpp +145 -0
  93. data/ext/sources/examples/talk-llama/models/exaone.cpp +3 -3
  94. data/ext/sources/examples/talk-llama/models/exaone4.cpp +3 -3
  95. data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +2 -4
  96. data/ext/sources/examples/talk-llama/models/falcon.cpp +3 -3
  97. data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +1 -1
  98. data/ext/sources/examples/talk-llama/models/gemma.cpp +1 -1
  99. data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +1 -1
  100. data/ext/sources/examples/talk-llama/models/gemma3.cpp +1 -1
  101. data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +7 -7
  102. data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +3 -3
  103. data/ext/sources/examples/talk-llama/models/glm4.cpp +14 -7
  104. data/ext/sources/examples/talk-llama/models/gpt2.cpp +2 -2
  105. data/ext/sources/examples/talk-llama/models/gptneox.cpp +2 -2
  106. data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +4 -5
  107. data/ext/sources/examples/talk-llama/models/granite.cpp +4 -5
  108. data/ext/sources/examples/talk-llama/models/grok.cpp +4 -4
  109. data/ext/sources/examples/talk-llama/models/grovemoe.cpp +5 -7
  110. data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +3 -3
  111. data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +4 -5
  112. data/ext/sources/examples/talk-llama/models/internlm2.cpp +3 -3
  113. data/ext/sources/examples/talk-llama/models/jais.cpp +2 -2
  114. data/ext/sources/examples/talk-llama/models/jais2.cpp +123 -0
  115. data/ext/sources/examples/talk-llama/models/jamba.cpp +3 -3
  116. data/ext/sources/examples/talk-llama/models/kimi-linear.cpp +381 -0
  117. data/ext/sources/examples/talk-llama/models/lfm2.cpp +145 -124
  118. data/ext/sources/examples/talk-llama/models/llada-moe.cpp +4 -4
  119. data/ext/sources/examples/talk-llama/models/llada.cpp +3 -3
  120. data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +4 -4
  121. data/ext/sources/examples/talk-llama/models/llama.cpp +18 -11
  122. data/ext/sources/examples/talk-llama/models/maincoder.cpp +3 -3
  123. data/ext/sources/examples/talk-llama/models/{graph-context-mamba.cpp → mamba-base.cpp} +9 -3
  124. data/ext/sources/examples/talk-llama/models/mamba.cpp +1 -2
  125. data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +11 -5
  126. data/ext/sources/examples/talk-llama/models/minicpm3.cpp +14 -13
  127. data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +4 -5
  128. data/ext/sources/examples/talk-llama/models/mistral3.cpp +4 -4
  129. data/ext/sources/examples/talk-llama/models/models.h +181 -46
  130. data/ext/sources/examples/talk-llama/models/modern-bert.cpp +2 -9
  131. data/ext/sources/examples/talk-llama/models/mpt.cpp +2 -2
  132. data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +26 -14
  133. data/ext/sources/examples/talk-llama/models/nemotron.cpp +3 -3
  134. data/ext/sources/examples/talk-llama/models/neo-bert.cpp +2 -2
  135. data/ext/sources/examples/talk-llama/models/olmo.cpp +3 -3
  136. data/ext/sources/examples/talk-llama/models/olmo2.cpp +3 -3
  137. data/ext/sources/examples/talk-llama/models/olmoe.cpp +4 -4
  138. data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +1 -1
  139. data/ext/sources/examples/talk-llama/models/openelm.cpp +3 -3
  140. data/ext/sources/examples/talk-llama/models/orion.cpp +3 -3
  141. data/ext/sources/examples/talk-llama/models/paddleocr.cpp +122 -0
  142. data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +3 -3
  143. data/ext/sources/examples/talk-llama/models/phi2.cpp +2 -2
  144. data/ext/sources/examples/talk-llama/models/phi3.cpp +3 -3
  145. data/ext/sources/examples/talk-llama/models/plamo.cpp +3 -3
  146. data/ext/sources/examples/talk-llama/models/plamo2.cpp +9 -5
  147. data/ext/sources/examples/talk-llama/models/plamo3.cpp +2 -2
  148. data/ext/sources/examples/talk-llama/models/plm.cpp +15 -14
  149. data/ext/sources/examples/talk-llama/models/qwen.cpp +2 -2
  150. data/ext/sources/examples/talk-llama/models/qwen2.cpp +3 -3
  151. data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +4 -4
  152. data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +3 -3
  153. data/ext/sources/examples/talk-llama/models/qwen3.cpp +12 -9
  154. data/ext/sources/examples/talk-llama/models/qwen35.cpp +381 -0
  155. data/ext/sources/examples/talk-llama/models/qwen35moe.cpp +422 -0
  156. data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +15 -8
  157. data/ext/sources/examples/talk-llama/models/qwen3next.cpp +84 -432
  158. data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +9 -18
  159. data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +8 -17
  160. data/ext/sources/examples/talk-llama/models/refact.cpp +2 -2
  161. data/ext/sources/examples/talk-llama/models/rnd1.cpp +4 -4
  162. data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +2 -0
  163. data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +2 -0
  164. data/ext/sources/examples/talk-llama/models/seed-oss.cpp +3 -3
  165. data/ext/sources/examples/talk-llama/models/smallthinker.cpp +4 -4
  166. data/ext/sources/examples/talk-llama/models/smollm3.cpp +3 -3
  167. data/ext/sources/examples/talk-llama/models/stablelm.cpp +2 -2
  168. data/ext/sources/examples/talk-llama/models/starcoder.cpp +2 -2
  169. data/ext/sources/examples/talk-llama/models/starcoder2.cpp +3 -3
  170. data/ext/sources/examples/talk-llama/models/step35-iswa.cpp +165 -0
  171. data/ext/sources/examples/talk-llama/models/t5-dec.cpp +2 -2
  172. data/ext/sources/examples/talk-llama/models/t5-enc.cpp +2 -2
  173. data/ext/sources/examples/talk-llama/models/xverse.cpp +3 -3
  174. data/ext/sources/examples/talk-llama/unicode.cpp +21 -65
  175. data/ext/sources/ggml/CMakeLists.txt +9 -3
  176. data/ext/sources/ggml/include/ggml-backend.h +1 -1
  177. data/ext/sources/ggml/include/ggml-cann.h +1 -1
  178. data/ext/sources/ggml/include/ggml-cpu.h +5 -0
  179. data/ext/sources/ggml/include/ggml-openvino.h +37 -0
  180. data/ext/sources/ggml/include/ggml-opt.h +1 -1
  181. data/ext/sources/ggml/include/ggml-rpc.h +6 -1
  182. data/ext/sources/ggml/include/ggml-virtgpu.h +14 -0
  183. data/ext/sources/ggml/include/ggml.h +56 -9
  184. data/ext/sources/ggml/src/CMakeLists.txt +3 -0
  185. data/ext/sources/ggml/src/ggml-alloc.c +4 -9
  186. data/ext/sources/ggml/src/ggml-backend-dl.cpp +48 -0
  187. data/ext/sources/ggml/src/ggml-backend-dl.h +45 -0
  188. data/ext/sources/ggml/src/ggml-backend-reg.cpp +28 -86
  189. data/ext/sources/ggml/src/ggml-backend.cpp +5 -2
  190. data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +1 -1
  191. data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +6 -2
  192. data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +1 -1
  193. data/ext/sources/ggml/src/ggml-cann/acl_tensor.h +1 -1
  194. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +348 -189
  195. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +40 -85
  196. data/ext/sources/ggml/src/ggml-cann/common.h +3 -4
  197. data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +44 -62
  198. data/ext/sources/ggml/src/ggml-common.h +11 -0
  199. data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +16 -11
  200. data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +42 -19
  201. data/ext/sources/ggml/src/ggml-cpu/amx/common.h +34 -10
  202. data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +85 -85
  203. data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +85 -1
  204. data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +2744 -548
  205. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +1653 -0
  206. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +1391 -0
  207. data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +8 -10
  208. data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +9 -9
  209. data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +118 -18
  210. data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +107 -26
  211. data/ext/sources/ggml/src/ggml-cpu/binary-ops.cpp +2 -6
  212. data/ext/sources/ggml/src/ggml-cpu/common.h +8 -0
  213. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +3 -0
  214. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +59 -12
  215. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +15 -0
  216. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +21 -20
  217. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +965 -252
  218. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +584 -197
  219. data/ext/sources/ggml/src/ggml-cpu/ops.cpp +903 -188
  220. data/ext/sources/ggml/src/ggml-cpu/ops.h +1 -0
  221. data/ext/sources/ggml/src/ggml-cpu/quants.c +40 -0
  222. data/ext/sources/ggml/src/ggml-cpu/quants.h +3 -0
  223. data/ext/sources/ggml/src/ggml-cpu/repack.cpp +2890 -679
  224. data/ext/sources/ggml/src/ggml-cpu/repack.h +119 -8
  225. data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +136 -0
  226. data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +111 -3
  227. data/ext/sources/ggml/src/ggml-cpu/unary-ops.cpp +1 -1
  228. data/ext/sources/ggml/src/ggml-cpu/vec.cpp +17 -0
  229. data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +1 -1
  230. data/ext/sources/ggml/src/ggml-cuda/argsort.cu +19 -10
  231. data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +32 -30
  232. data/ext/sources/ggml/src/ggml-cuda/common.cuh +134 -18
  233. data/ext/sources/ggml/src/ggml-cuda/convert.cu +41 -27
  234. data/ext/sources/ggml/src/ggml-cuda/cpy.cu +6 -3
  235. data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +78 -64
  236. data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +384 -143
  237. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +36 -22
  238. data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +3 -3
  239. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +26 -5
  240. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cuh +1 -1
  241. data/ext/sources/ggml/src/ggml-cuda/fattn.cu +127 -12
  242. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cu +263 -0
  243. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cuh +4 -0
  244. data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +595 -200
  245. data/ext/sources/ggml/src/ggml-cuda/mean.cu +9 -8
  246. data/ext/sources/ggml/src/ggml-cuda/mma.cuh +173 -6
  247. data/ext/sources/ggml/src/ggml-cuda/mmf.cu +30 -10
  248. data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +158 -85
  249. data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +34 -22
  250. data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +127 -67
  251. data/ext/sources/ggml/src/ggml-cuda/mmvf.cuh +2 -0
  252. data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +157 -65
  253. data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +1 -0
  254. data/ext/sources/ggml/src/ggml-cuda/norm.cu +18 -76
  255. data/ext/sources/ggml/src/ggml-cuda/pad.cu +13 -10
  256. data/ext/sources/ggml/src/ggml-cuda/quantize.cu +1 -1
  257. data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +2 -16
  258. data/ext/sources/ggml/src/ggml-cuda/rope.cu +233 -133
  259. data/ext/sources/ggml/src/ggml-cuda/softmax.cu +8 -83
  260. data/ext/sources/ggml/src/ggml-cuda/solve_tri.cu +1 -1
  261. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +56 -32
  262. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +1 -1
  263. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +5 -0
  264. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +1 -0
  265. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +5 -0
  266. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +1 -0
  267. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +1 -0
  268. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +1 -0
  269. data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +3 -3
  270. data/ext/sources/ggml/src/ggml-cuda/top-k.cu +0 -1
  271. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +199 -135
  272. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cuh +20 -14
  273. data/ext/sources/ggml/src/ggml-cuda/unary.cu +55 -0
  274. data/ext/sources/ggml/src/ggml-cuda/unary.cuh +2 -0
  275. data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +31 -17
  276. data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +10 -0
  277. data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +82 -45
  278. data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +334 -160
  279. data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +7 -5
  280. data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +328 -197
  281. data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +281 -0
  282. data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +765 -234
  283. data/ext/sources/ggml/src/ggml-hexagon/htp/cpy-ops.c +252 -0
  284. data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +412 -265
  285. data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +23 -23
  286. data/ext/sources/ggml/src/ggml-hexagon/htp/{htp-dma.c → hex-dma.c} +1 -1
  287. data/ext/sources/ggml/src/ggml-hexagon/htp/{htp-dma.h → hex-dma.h} +28 -3
  288. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +77 -0
  289. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
  290. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +51 -0
  291. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +1 -1
  292. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +27 -37
  293. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +6 -35
  294. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-arith.h +443 -0
  295. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +240 -0
  296. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +245 -0
  297. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +251 -0
  298. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
  299. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +215 -0
  300. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
  301. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.h +210 -0
  302. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-reduce.h +296 -0
  303. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-scale.h +133 -0
  304. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +141 -0
  305. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +126 -0
  306. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
  307. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +20 -1347
  308. data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +211 -13
  309. data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +1119 -952
  310. data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +254 -244
  311. data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +36 -36
  312. data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +155 -138
  313. data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c +339 -0
  314. data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +128 -0
  315. data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +209 -114
  316. data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.c +1 -5
  317. data/ext/sources/ggml/src/ggml-hexagon/htp-drv.cpp +418 -0
  318. data/ext/sources/ggml/src/ggml-hexagon/htp-drv.h +121 -0
  319. data/ext/sources/ggml/src/ggml-hexagon/libdl.h +79 -0
  320. data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +38 -0
  321. data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +6 -0
  322. data/ext/sources/ggml/src/ggml-impl.h +62 -0
  323. data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +10 -10
  324. data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp +13 -2
  325. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.h +8 -0
  326. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.m +147 -17
  327. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +274 -73
  328. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +22 -4
  329. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +102 -36
  330. data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +174 -23
  331. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +580 -280
  332. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +5 -4
  333. data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +320 -107
  334. data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +1068 -825
  335. data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +19 -1
  336. data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +3108 -636
  337. data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +41 -99
  338. data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +45 -0
  339. data/ext/sources/ggml/src/ggml-opencl/kernels/cumsum.cl +139 -0
  340. data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +204 -0
  341. data/ext/sources/ggml/src/ggml-opencl/kernels/diag.cl +27 -0
  342. data/ext/sources/ggml/src/ggml-opencl/kernels/exp.cl +125 -0
  343. data/ext/sources/ggml/src/ggml-opencl/kernels/expm1.cl +87 -56
  344. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_1_f32.cl +132 -0
  345. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl +195 -0
  346. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl +283 -0
  347. data/ext/sources/ggml/src/ggml-opencl/kernels/l2_norm.cl +71 -0
  348. data/ext/sources/ggml/src/ggml-opencl/kernels/mean.cl +114 -13
  349. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
  350. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
  351. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
  352. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl +129 -0
  353. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
  354. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
  355. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +180 -0
  356. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mv_q6_k.cl → mul_mv_q6_k_f32.cl} +4 -0
  357. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +194 -0
  358. data/ext/sources/ggml/src/ggml-opencl/kernels/neg.cl +125 -0
  359. data/ext/sources/ggml/src/ggml-opencl/kernels/repeat.cl +31 -32
  360. data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +14 -4
  361. data/ext/sources/ggml/src/ggml-opencl/kernels/softplus.cl +88 -60
  362. data/ext/sources/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
  363. data/ext/sources/ggml/src/ggml-opencl/kernels/sum_rows.cl +114 -13
  364. data/ext/sources/ggml/src/ggml-opencl/kernels/tanh.cl +94 -48
  365. data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +26 -0
  366. data/ext/sources/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
  367. data/ext/sources/ggml/src/ggml-openvino/.clang-format +154 -0
  368. data/ext/sources/ggml/src/ggml-openvino/CMakeLists.txt +22 -0
  369. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +975 -0
  370. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.h +294 -0
  371. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +373 -0
  372. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.h +182 -0
  373. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +1110 -0
  374. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +884 -0
  375. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.h +153 -0
  376. data/ext/sources/ggml/src/ggml-openvino/openvino/decoder.h +74 -0
  377. data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.cpp +27 -0
  378. data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.h +23 -0
  379. data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.cpp +17 -0
  380. data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.h +29 -0
  381. data/ext/sources/ggml/src/ggml-openvino/openvino/node_context.h +112 -0
  382. data/ext/sources/ggml/src/ggml-openvino/openvino/op/cont.cpp +48 -0
  383. data/ext/sources/ggml/src/ggml-openvino/openvino/op/cpy.cpp +21 -0
  384. data/ext/sources/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +90 -0
  385. data/ext/sources/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +69 -0
  386. data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +61 -0
  387. data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +62 -0
  388. data/ext/sources/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +90 -0
  389. data/ext/sources/ggml/src/ggml-openvino/openvino/op/permute.cpp +102 -0
  390. data/ext/sources/ggml/src/ggml-openvino/openvino/op/reshape.cpp +83 -0
  391. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +46 -0
  392. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +123 -0
  393. data/ext/sources/ggml/src/ggml-openvino/openvino/op/scale.cpp +41 -0
  394. data/ext/sources/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +76 -0
  395. data/ext/sources/ggml/src/ggml-openvino/openvino/op/softmax.cpp +89 -0
  396. data/ext/sources/ggml/src/ggml-openvino/openvino/op/transpose.cpp +23 -0
  397. data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +27 -0
  398. data/ext/sources/ggml/src/ggml-openvino/openvino/op/view.cpp +53 -0
  399. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +46 -0
  400. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +39 -0
  401. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.cpp +123 -0
  402. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/eliminate_zp.h +17 -0
  403. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.cpp +60 -0
  404. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.h +17 -0
  405. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +29 -0
  406. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp +58 -0
  407. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.h +17 -0
  408. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.cpp +293 -0
  409. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.h +28 -0
  410. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +226 -0
  411. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +85 -0
  412. data/ext/sources/ggml/src/ggml-openvino/utils.cpp +823 -0
  413. data/ext/sources/ggml/src/ggml-openvino/utils.h +123 -0
  414. data/ext/sources/ggml/src/ggml-quants.c +96 -5
  415. data/ext/sources/ggml/src/ggml-quants.h +3 -0
  416. data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +15 -88
  417. data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +5 -1
  418. data/ext/sources/ggml/src/ggml-sycl/backend.hpp +1 -0
  419. data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +21 -20
  420. data/ext/sources/ggml/src/ggml-sycl/common.hpp +315 -10
  421. data/ext/sources/ggml/src/ggml-sycl/convert.cpp +69 -1
  422. data/ext/sources/ggml/src/ggml-sycl/convert.hpp +22 -1
  423. data/ext/sources/ggml/src/ggml-sycl/count-equal.cpp +1 -1
  424. data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +791 -47
  425. data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +78 -68
  426. data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -0
  427. data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp +1179 -0
  428. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +55 -0
  429. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +1338 -0
  430. data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +667 -0
  431. data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +225 -0
  432. data/ext/sources/ggml/src/ggml-sycl/fattn.hpp +22 -0
  433. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.cpp +309 -0
  434. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +8 -0
  435. data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +316 -51
  436. data/ext/sources/ggml/src/ggml-sycl/norm.cpp +65 -66
  437. data/ext/sources/ggml/src/ggml-sycl/outprod.cpp +3 -3
  438. data/ext/sources/ggml/src/ggml-sycl/presets.hpp +3 -0
  439. data/ext/sources/ggml/src/ggml-sycl/quants.hpp +1 -1
  440. data/ext/sources/ggml/src/ggml-sycl/rope.cpp +450 -287
  441. data/ext/sources/ggml/src/ggml-sycl/rope.hpp +6 -0
  442. data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +6 -6
  443. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq112-dv112.cpp +5 -0
  444. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq128-dv128.cpp +5 -0
  445. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq256-dv256.cpp +5 -0
  446. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq40-dv40.cpp +5 -0
  447. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq576-dv512.cpp +5 -0
  448. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq64-dv64.cpp +5 -0
  449. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq72-dv72.cpp +5 -0
  450. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq80-dv80.cpp +5 -0
  451. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq96-dv96.cpp +5 -0
  452. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +7 -0
  453. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +7 -0
  454. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +7 -0
  455. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +7 -0
  456. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +7 -0
  457. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +7 -0
  458. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +7 -0
  459. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +7 -0
  460. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +7 -0
  461. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +7 -0
  462. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +7 -0
  463. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +7 -0
  464. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +7 -0
  465. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +7 -0
  466. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +7 -0
  467. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +7 -0
  468. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +7 -0
  469. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +7 -0
  470. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +7 -0
  471. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +7 -0
  472. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +7 -0
  473. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +7 -0
  474. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +7 -0
  475. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +7 -0
  476. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +7 -0
  477. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +7 -0
  478. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +7 -0
  479. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +7 -0
  480. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +7 -0
  481. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +7 -0
  482. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +7 -0
  483. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +7 -0
  484. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +7 -0
  485. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +7 -0
  486. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +7 -0
  487. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +7 -0
  488. data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +13 -0
  489. data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +1 -1
  490. data/ext/sources/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
  491. data/ext/sources/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
  492. data/ext/sources/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
  493. data/ext/sources/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
  494. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
  495. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
  496. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
  497. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
  498. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +148 -0
  499. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
  500. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
  501. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
  502. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
  503. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
  504. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
  505. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
  506. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
  507. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
  508. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
  509. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
  510. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
  511. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +119 -0
  512. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +158 -0
  513. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
  514. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +69 -0
  515. data/ext/sources/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
  516. data/ext/sources/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
  517. data/ext/sources/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
  518. data/ext/sources/ggml/src/ggml-virtgpu/regenerate_remoting.py +333 -0
  519. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
  520. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
  521. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
  522. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
  523. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +192 -0
  524. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
  525. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +53 -0
  526. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +98 -0
  527. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
  528. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
  529. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
  530. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +544 -0
  531. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +117 -0
  532. data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +1 -1
  533. data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +1250 -465
  534. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +16 -8
  535. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/elu.comp +27 -0
  536. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +374 -170
  537. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +66 -22
  538. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +389 -201
  539. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +106 -58
  540. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mask_opt.comp +162 -0
  541. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +9 -8
  542. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +128 -0
  543. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +12 -9
  544. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl +20 -17
  545. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +11 -3
  546. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +8 -4
  547. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +3 -1
  548. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +5 -3
  549. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +3 -3
  550. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +2 -3
  551. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +36 -63
  552. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +7 -4
  553. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +7 -4
  554. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +7 -4
  555. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +10 -5
  556. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +7 -4
  557. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sgn.comp +21 -0
  558. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +16 -10
  559. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +55 -35
  560. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +1314 -109
  561. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +1660 -1371
  562. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
  563. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +106 -0
  564. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
  565. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +141 -0
  566. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +65 -72
  567. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +75 -0
  568. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cpy.tmpl.wgsl +6 -0
  569. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
  570. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +40 -5
  571. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +105 -60
  572. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{get_rows.tmpl.wgsl → get_rows.wgsl} +53 -259
  573. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat.tmpl.wgsl → mul_mat.wgsl} +68 -257
  574. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +692 -23
  575. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat_reg_tile.tmpl.wgsl → mul_mat_reg_tile.wgsl} +28 -128
  576. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat_subgroup_matrix.tmpl.wgsl → mul_mat_subgroup_matrix.wgsl} +31 -137
  577. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +480 -0
  578. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
  579. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl +67 -0
  580. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{scale.tmpl.wgsl → scale.wgsl} +9 -36
  581. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +40 -12
  582. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
  583. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +193 -0
  584. data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +6 -1
  585. data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +31 -32
  586. data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +9 -6
  587. data/ext/sources/ggml/src/ggml.c +167 -33
  588. data/ext/sources/ggml/src/gguf.cpp +229 -44
  589. data/ext/sources/src/whisper.cpp +6 -28
  590. data/sig/whisper.rbs +43 -2
  591. data/test/test_context_params.rb +82 -0
  592. data/test/test_token.rb +11 -0
  593. data/test/test_vad_context.rb +58 -8
  594. data/test/test_whisper.rb +20 -0
  595. data/whispercpp.gemspec +1 -1
  596. metadata +240 -28
  597. data/ext/sources/ggml/cmake/BuildTypes.cmake +0 -54
  598. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm-ppc.h +0 -333
  599. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.c +0 -94
  600. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.c +0 -72
  601. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.c +0 -49
  602. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.c +0 -1020
  603. data/ext/sources/ggml/src/ggml-hexagon/htp/ops-utils.h +0 -149
  604. data/ext/sources/ggml/src/ggml-hexagon/htp-utils.c +0 -454
  605. data/ext/sources/ggml/src/ggml-hexagon/htp-utils.h +0 -221
  606. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/bin_op.tmpl.wgsl +0 -188
  607. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary_head.tmpl +0 -45
  608. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.tmpl.wgsl +0 -267
  609. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.tmpl.wgsl +0 -112
  610. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary_op.wgsl +0 -483
@@ -1,222 +1,31 @@
1
- #define(VARIANTS)
2
-
3
- [
4
- {
5
- "SHADER_SUFFIX": "f32_vec",
6
- "REPLS": {
7
- "TYPE" : "vec4<f32>",
8
- "DST_TYPE": "vec4<f32>",
9
- "BLOCK_SIZE": 4
10
- },
11
- "DECLS": ["F32_VEC"]
12
- },
13
- {
14
- "REPLS": {
15
- "TYPE" : "f32",
16
- "DST_TYPE": "f32",
17
- "BLOCK_SIZE": 1
18
- },
19
- "DECLS": ["F32"]
20
- },
21
- {
22
- "REPLS": {
23
- "TYPE" : "f16",
24
- "DST_TYPE": "f32",
25
- "BLOCK_SIZE": 1
26
- },
27
- "DECLS": ["F16"]
28
- },
29
- {
30
- "REPLS": {
31
- "TYPE" : "i32",
32
- "DST_TYPE": "i32",
33
- "BLOCK_SIZE": 1
34
- },
35
- "DECLS": ["I32"]
36
- },
37
- {
38
- "REPLS": {
39
- "TYPE" : "q4_0",
40
- "DST_TYPE": "f32",
41
- "BLOCK_SIZE": 32
42
- },
43
- "DECLS": ["BYTE_HELPERS", "Q4_0_T", "Q4_0"]
44
- },
45
- {
46
- "REPLS": {
47
- "TYPE" : "q4_1",
48
- "DST_TYPE": "f32",
49
- "BLOCK_SIZE": 32
50
- },
51
- "DECLS": ["BYTE_HELPERS", "Q4_1_T", "Q4_1"]
52
- },
53
- {
54
- "REPLS": {
55
- "TYPE" : "q5_0",
56
- "DST_TYPE": "f32",
57
- "BLOCK_SIZE": 32
58
- },
59
- "DECLS": ["BYTE_HELPERS", "Q5_0_T", "Q5_0"]
60
- },
61
- {
62
- "REPLS": {
63
- "TYPE" : "q5_1",
64
- "DST_TYPE": "f32",
65
- "BLOCK_SIZE": 32
66
- },
67
- "DECLS": ["BYTE_HELPERS", "Q5_1_T", "Q5_1"]
68
- },
69
- {
70
- "REPLS": {
71
- "TYPE" : "q8_0",
72
- "DST_TYPE": "f32",
73
- "BLOCK_SIZE": 32
74
- },
75
- "DECLS": ["BYTE_HELPERS", "Q8_0_T", "Q8_0"]
76
- },
77
- {
78
- "REPLS": {
79
- "TYPE" : "q2_k",
80
- "DST_TYPE": "f32",
81
- "BLOCK_SIZE": 256
82
- },
83
- "DECLS": ["BYTE_HELPERS", "Q2_K_T", "Q2_K"]
84
- },
85
- {
86
- "REPLS": {
87
- "TYPE" : "q3_k",
88
- "DST_TYPE": "f32",
89
- "BLOCK_SIZE": 256
90
- },
91
- "DECLS": ["BYTE_HELPERS", "Q3_K_T", "Q3_K"]
92
- },
93
- {
94
- "REPLS": {
95
- "TYPE" : "q4_k",
96
- "DST_TYPE": "f32",
97
- "BLOCK_SIZE": 256
98
- },
99
- "DECLS": ["Q45_K_SCALE_MIN", "BYTE_HELPERS", "Q4_K_T", "Q4_K"]
100
- },
101
- {
102
- "REPLS": {
103
- "TYPE" : "q5_k",
104
- "DST_TYPE": "f32",
105
- "BLOCK_SIZE": 256
106
- },
107
- "DECLS": ["Q45_K_SCALE_MIN", "BYTE_HELPERS", "Q5_K_T", "Q5_K"]
108
- },
109
- {
110
- "REPLS": {
111
- "TYPE" : "q6_k",
112
- "DST_TYPE": "f32",
113
- "BLOCK_SIZE": 256
114
- },
115
- "DECLS": ["BYTE_HELPERS", "Q6_K_T", "Q6_K"]
116
- },
117
- {
118
- "REPLS": {
119
- "TYPE" : "iq2_xxs",
120
- "DST_TYPE": "f32",
121
- "BLOCK_SIZE": 256
122
- },
123
- "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_XXS_GRID", "IQ2_XXS_T", "IQ2_XXS"]
124
- },
125
- {
126
- "REPLS": {
127
- "TYPE" : "iq2_xs",
128
- "DST_TYPE": "f32",
129
- "BLOCK_SIZE": 256
130
- },
131
- "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_XS_GRID", "IQ2_XS_T", "IQ2_XS"]
132
- },
133
- {
134
- "REPLS": {
135
- "TYPE": "iq2_s",
136
- "DST_TYPE": "f32",
137
- "BLOCK_SIZE": 256
138
- },
139
- "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_S_GRID", "IQ2_S_T", "IQ2_S"]
140
- },
141
- {
142
- "REPLS": {
143
- "TYPE": "iq3_xxs",
144
- "DST_TYPE": "f32",
145
- "BLOCK_SIZE": 256
146
- },
147
- "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ3_XSS_GRID", "IQ3_XSS_T", "IQ3_XSS"]
148
- },
149
- {
150
- "REPLS": {
151
- "TYPE": "iq3_s",
152
- "DST_TYPE": "f32",
153
- "BLOCK_SIZE": 256
154
- },
155
- "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ3_S_GRID", "IQ3_S_T", "IQ3_S"]
156
- },
157
- {
158
- "REPLS": {
159
- "TYPE": "iq1_s",
160
- "DST_TYPE": "f32",
161
- "BLOCK_SIZE": 256
162
- },
163
- "DECLS": ["BYTE_HELPERS", "IQ1_GRID", "IQ1_S_T", "IQ1_S"]
164
- },
165
- {
166
- "REPLS": {
167
- "TYPE": "iq1_m",
168
- "DST_TYPE": "f32",
169
- "BLOCK_SIZE": 256
170
- },
171
- "DECLS": ["BYTE_HELPERS", "IQ1_GRID", "IQ1_M_T", "IQ1_M"]
172
- },
173
- {
174
- "REPLS": {
175
- "TYPE": "iq4_nl",
176
- "DST_TYPE": "f32",
177
- "BLOCK_SIZE": 32,
178
- },
179
- "DECLS": ["BYTE_HELPERS", "IQ4_GRID", "IQ4_NL_T", "IQ4_NL"]
180
- },
181
- {
182
- "REPLS": {
183
- "TYPE": "iq4_xs",
184
- "DST_TYPE": "f32",
185
- "BLOCK_SIZE": 256,
186
- },
187
- "DECLS": ["BYTE_HELPERS", "IQ4_GRID", "IQ4_XS_T", "IQ4_XS"]
188
- }
189
- ]
190
-
191
- #end(VARIANTS)
192
-
193
- #define(DECLS)
194
-
195
- #decl(F32_VEC)
1
+ enable f16;
2
+ #include "common_decls.tmpl"
3
+
4
+ #ifdef F32_VEC
196
5
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
197
6
  dst[(dst_base / 4) + offset] = src[(src_base / 4) + offset];
198
7
  }
199
- #enddecl(F32_VEC)
8
+ #endif
200
9
 
201
- #decl(F32)
10
+ #ifdef F32
202
11
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
203
12
  dst[dst_base + offset] = src[src_base + offset];
204
13
  }
205
- #enddecl(F32)
14
+ #endif
206
15
 
207
- #decl(F16)
16
+ #ifdef F16
208
17
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
209
18
  dst[dst_base + offset] = f32(src[src_base + offset]);
210
19
  }
211
- #enddecl(F16)
20
+ #endif
212
21
 
213
- #decl(I32)
22
+ #ifdef I32
214
23
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
215
24
  dst[dst_base + offset] = src[src_base + offset];
216
25
  }
217
- #enddecl(I32)
26
+ #endif
218
27
 
219
- #decl(Q4_0)
28
+ #ifdef Q4_0
220
29
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
221
30
  let block_q4_0 = src[src_base + offset];
222
31
  let d = f32(block_q4_0.d);
@@ -232,9 +41,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
232
41
  }
233
42
  }
234
43
  }
235
- #enddecl(Q4_0)
44
+ #endif
236
45
 
237
- #decl(Q4_1)
46
+ #ifdef Q4_1
238
47
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
239
48
  let block_q4_1 = src[src_base + offset];
240
49
  let d = f32(block_q4_1.d);
@@ -251,9 +60,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
251
60
  }
252
61
  }
253
62
  }
254
- #enddecl(Q4_1)
63
+ #endif
255
64
 
256
- #decl(Q5_0)
65
+ #ifdef Q5_0
257
66
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
258
67
  let block_q5_0 = src[src_base + offset];
259
68
  let d = f32(block_q5_0.d);
@@ -272,10 +81,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
272
81
  }
273
82
  }
274
83
  }
84
+ #endif
275
85
 
276
- #enddecl(Q5_0)
277
-
278
- #decl(Q5_1)
86
+ #ifdef Q5_1
279
87
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
280
88
  let block_q5_1 = src[src_base + offset];
281
89
  let d = f32(block_q5_1.d);
@@ -294,9 +102,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
294
102
  }
295
103
  }
296
104
  }
297
- #enddecl(Q5_1)
105
+ #endif
298
106
 
299
- #decl(Q8_0)
107
+ #ifdef Q8_0
300
108
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
301
109
  let block_q8_0 = src[src_base + offset];
302
110
  let d = f32(block_q8_0.d);
@@ -310,9 +118,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
310
118
  }
311
119
  }
312
120
  }
313
- #enddecl(Q8_0)
121
+ #endif
314
122
 
315
- #decl(Q2_K)
123
+ #ifdef Q2_K
316
124
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
317
125
  let block = src[src_base + offset];
318
126
  let d = f32(block.d);
@@ -340,9 +148,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
340
148
  }
341
149
  }
342
150
  }
343
- #enddecl(Q2_K)
151
+ #endif
344
152
 
345
- #decl(Q3_K)
153
+ #ifdef Q3_K
346
154
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
347
155
  let block = src[src_base + offset];
348
156
  let d = f32(block.d);
@@ -398,9 +206,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
398
206
  }
399
207
  }
400
208
  }
401
- #enddecl(Q3_K)
209
+ #endif
402
210
 
403
- #decl(Q4_K)
211
+ #ifdef Q4_K
404
212
  // 8 blocks of 32 elements each
405
213
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
406
214
  let block = src[src_base + offset];
@@ -425,9 +233,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
425
233
  }
426
234
  }
427
235
  }
428
- #enddecl(Q4_K)
236
+ #endif
429
237
 
430
- #decl(Q5_K)
238
+ #ifdef Q5_K
431
239
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
432
240
  let block = src[src_base + offset];
433
241
  let d = f32(block.d);
@@ -455,9 +263,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
455
263
  }
456
264
  }
457
265
  }
458
- #enddecl(Q5_K)
266
+ #endif
459
267
 
460
- #decl(Q6_K)
268
+ #ifdef Q6_K
461
269
  // 16 blocks of 16 elements each
462
270
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
463
271
  let block = src[src_base + offset];
@@ -511,10 +319,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
511
319
  sc_b_idx += 8;
512
320
  }
513
321
  }
322
+ #endif
514
323
 
515
- #enddecl(Q6_K)
516
-
517
- #decl(IQ2_XXS)
324
+ #ifdef IQ2_XXS
518
325
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
519
326
  let block = src[src_base + offset];
520
327
  let d = f32(block.d);
@@ -536,9 +343,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
536
343
  }
537
344
  }
538
345
  }
539
- #enddecl(IQ2_XXS)
346
+ #endif
540
347
 
541
- #decl(IQ2_XS)
348
+ #ifdef IQ2_XS
542
349
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
543
350
  let block = src[src_base + offset];
544
351
  let d = f32(block.d);
@@ -568,9 +375,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
568
375
  }
569
376
  }
570
377
  }
571
- #enddecl(IQ2_XS)
378
+ #endif
572
379
 
573
- #decl(IQ2_S)
380
+ #ifdef IQ2_S
574
381
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
575
382
  let block = src[src_base + offset];
576
383
  let d = f32(block.d);
@@ -608,10 +415,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
608
415
  }
609
416
  }
610
417
  }
418
+ #endif
611
419
 
612
- #enddecl(IQ2_S)
613
-
614
- #decl(IQ3_XSS)
420
+ #ifdef IQ3_XXS
615
421
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
616
422
  let block = src[src_base + offset];
617
423
  let d = f32(block.d);
@@ -638,9 +444,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
638
444
  }
639
445
  }
640
446
  }
641
- #enddecl(IQ3_XSS)
447
+ #endif
642
448
 
643
- #decl(IQ3_S)
449
+ #ifdef IQ3_S
644
450
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
645
451
  let block = src[src_base + offset];
646
452
  let d = f32(block.d);
@@ -683,9 +489,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
683
489
  }
684
490
  }
685
491
  }
686
- #enddecl(IQ3_S)
492
+ #endif
687
493
 
688
- #decl(IQ1_S)
494
+ #ifdef IQ1_S
689
495
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
690
496
  let block = src[src_base + offset];
691
497
  let d = f32(block.d);
@@ -707,10 +513,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
707
513
  }
708
514
  }
709
515
  }
516
+ #endif
710
517
 
711
- #enddecl(IQ1_S)
712
-
713
- #decl(IQ1_M)
518
+ #ifdef IQ1_M
714
519
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
715
520
  let block = src[src_base + offset];
716
521
 
@@ -751,10 +556,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
751
556
  }
752
557
  }
753
558
  }
559
+ #endif
754
560
 
755
- #enddecl(IQ1_M)
756
-
757
- #decl(IQ4_NL)
561
+ #ifdef IQ4_NL
758
562
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
759
563
  let block = src[src_base + offset];
760
564
  let d = f32(block.d);
@@ -770,9 +574,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
770
574
  dst_i++;
771
575
  }
772
576
  }
773
- #enddecl(IQ4_NL)
577
+ #endif
774
578
 
775
- #decl(IQ4_XS)
579
+ #ifdef IQ4_XS
776
580
  fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
777
581
  let block = src[src_base + offset];
778
582
  let d = f32(block.d);
@@ -791,24 +595,16 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
791
595
  dst_i += 16;
792
596
  }
793
597
  }
794
- #enddecl(IQ4_XS)
795
-
796
- #end(DECLS)
797
-
798
- #define(SHADER)
799
-
800
- enable f16;
801
-
802
- DECLS
598
+ #endif
803
599
 
804
600
  @group(0) @binding(0)
805
- var<storage, read_write> src: array<{{TYPE}}>;
601
+ var<storage, read_write> src: array<SRC_TYPE>;
806
602
 
807
603
  @group(0) @binding(1)
808
604
  var<storage, read_write> idx: array<i32>;
809
605
 
810
606
  @group(0) @binding(2)
811
- var<storage, read_write> dst: array<{{DST_TYPE}}>;
607
+ var<storage, read_write> dst: array<DST_TYPE>;
812
608
 
813
609
  struct Params {
814
610
  offset_src: u32, // in elements
@@ -842,8 +638,7 @@ struct Params {
842
638
  @group(0) @binding(3)
843
639
  var<uniform> params: Params;
844
640
 
845
- override wg_size: u32;
846
- @compute @workgroup_size(wg_size)
641
+ @compute @workgroup_size(WG_SIZE)
847
642
  fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
848
643
  if (gid.x >= params.n_rows * params.ne2 * params.ne3) {
849
644
  return;
@@ -866,9 +661,8 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
866
661
  let i_src_row = params.offset_src + idx_val * params.stride_src1 + i_dst2 * params.stride_src2 + i_dst3 * params.stride_src3;
867
662
  let i_dst_row = params.offset_dst + i_dst1 * params.stride_dst1 + i_dst2 * params.stride_dst2 + i_dst3 * params.stride_dst3;
868
663
 
869
- for (var i: u32 = 0; i < params.ne0/{{BLOCK_SIZE}}; i++) {
664
+ for (var i: u32 = 0; i < params.ne0/BLOCK_SIZE; i++) {
870
665
  copy_elements(i_src_row, i_dst_row, i);
871
666
  }
872
667
  }
873
668
 
874
- #end(SHADER)