local-llm-rn 1.0.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (626) hide show
  1. package/cpp/CMakeLists.txt +285 -0
  2. package/cpp/common/CMakeLists.txt +149 -0
  3. package/cpp/common/arg.cpp +3799 -0
  4. package/cpp/common/arg.h +131 -0
  5. package/cpp/common/base64.hpp +392 -0
  6. package/cpp/common/build-info.cpp.in +4 -0
  7. package/cpp/common/chat-parser-xml-toolcall.cpp +879 -0
  8. package/cpp/common/chat-parser-xml-toolcall.h +45 -0
  9. package/cpp/common/chat-parser.cpp +1649 -0
  10. package/cpp/common/chat-parser.h +133 -0
  11. package/cpp/common/chat-peg-parser.cpp +124 -0
  12. package/cpp/common/chat-peg-parser.h +105 -0
  13. package/cpp/common/chat.cpp +3355 -0
  14. package/cpp/common/chat.h +252 -0
  15. package/cpp/common/common.cpp +1824 -0
  16. package/cpp/common/common.h +930 -0
  17. package/cpp/common/console.cpp +1137 -0
  18. package/cpp/common/console.h +41 -0
  19. package/cpp/common/debug.cpp +167 -0
  20. package/cpp/common/debug.h +43 -0
  21. package/cpp/common/download.cpp +792 -0
  22. package/cpp/common/download.h +84 -0
  23. package/cpp/common/http.h +84 -0
  24. package/cpp/common/jinja/README.md +88 -0
  25. package/cpp/common/jinja/caps.cpp +285 -0
  26. package/cpp/common/jinja/caps.h +30 -0
  27. package/cpp/common/jinja/lexer.cpp +341 -0
  28. package/cpp/common/jinja/lexer.h +157 -0
  29. package/cpp/common/jinja/parser.cpp +591 -0
  30. package/cpp/common/jinja/parser.h +21 -0
  31. package/cpp/common/jinja/runtime.cpp +867 -0
  32. package/cpp/common/jinja/runtime.h +638 -0
  33. package/cpp/common/jinja/string.cpp +213 -0
  34. package/cpp/common/jinja/string.h +61 -0
  35. package/cpp/common/jinja/utils.h +149 -0
  36. package/cpp/common/jinja/value.cpp +1393 -0
  37. package/cpp/common/jinja/value.h +756 -0
  38. package/cpp/common/json-partial.cpp +324 -0
  39. package/cpp/common/json-partial.h +39 -0
  40. package/cpp/common/json-schema-to-grammar.cpp +1153 -0
  41. package/cpp/common/json-schema-to-grammar.h +43 -0
  42. package/cpp/common/llguidance.cpp +258 -0
  43. package/cpp/common/log.cpp +446 -0
  44. package/cpp/common/log.h +119 -0
  45. package/cpp/common/ngram-cache.cpp +285 -0
  46. package/cpp/common/ngram-cache.h +101 -0
  47. package/cpp/common/ngram-map.cpp +530 -0
  48. package/cpp/common/ngram-map.h +115 -0
  49. package/cpp/common/ngram-mod.cpp +60 -0
  50. package/cpp/common/ngram-mod.h +38 -0
  51. package/cpp/common/peg-parser.cpp +1712 -0
  52. package/cpp/common/peg-parser.h +459 -0
  53. package/cpp/common/preset.cpp +483 -0
  54. package/cpp/common/preset.h +83 -0
  55. package/cpp/common/regex-partial.cpp +204 -0
  56. package/cpp/common/regex-partial.h +56 -0
  57. package/cpp/common/sampling.cpp +745 -0
  58. package/cpp/common/sampling.h +119 -0
  59. package/cpp/common/speculative.cpp +1074 -0
  60. package/cpp/common/speculative.h +41 -0
  61. package/cpp/common/unicode.cpp +64 -0
  62. package/cpp/common/unicode.h +22 -0
  63. package/cpp/ggml/CMakeLists.txt +494 -0
  64. package/cpp/ggml/cmake/GitVars.cmake +22 -0
  65. package/cpp/ggml/cmake/common.cmake +50 -0
  66. package/cpp/ggml/cmake/ggml-config.cmake.in +191 -0
  67. package/cpp/ggml/include/ggml-alloc.h +85 -0
  68. package/cpp/ggml/include/ggml-backend.h +373 -0
  69. package/cpp/ggml/include/ggml-blas.h +25 -0
  70. package/cpp/ggml/include/ggml-cann.h +123 -0
  71. package/cpp/ggml/include/ggml-cpp.h +39 -0
  72. package/cpp/ggml/include/ggml-cpu.h +151 -0
  73. package/cpp/ggml/include/ggml-cuda.h +47 -0
  74. package/cpp/ggml/include/ggml-hexagon.h +19 -0
  75. package/cpp/ggml/include/ggml-metal.h +61 -0
  76. package/cpp/ggml/include/ggml-opencl.h +26 -0
  77. package/cpp/ggml/include/ggml-opt.h +256 -0
  78. package/cpp/ggml/include/ggml-rpc.h +30 -0
  79. package/cpp/ggml/include/ggml-sycl.h +49 -0
  80. package/cpp/ggml/include/ggml-virtgpu.h +14 -0
  81. package/cpp/ggml/include/ggml-vulkan.h +29 -0
  82. package/cpp/ggml/include/ggml-webgpu.h +19 -0
  83. package/cpp/ggml/include/ggml-zdnn.h +17 -0
  84. package/cpp/ggml/include/ggml-zendnn.h +22 -0
  85. package/cpp/ggml/include/ggml.h +2753 -0
  86. package/cpp/ggml/include/gguf.h +204 -0
  87. package/cpp/ggml/src/CMakeLists.txt +492 -0
  88. package/cpp/ggml/src/ggml-alloc.c +1244 -0
  89. package/cpp/ggml/src/ggml-backend-dl.cpp +48 -0
  90. package/cpp/ggml/src/ggml-backend-dl.h +45 -0
  91. package/cpp/ggml/src/ggml-backend-impl.h +255 -0
  92. package/cpp/ggml/src/ggml-backend-reg.cpp +566 -0
  93. package/cpp/ggml/src/ggml-backend.cpp +2270 -0
  94. package/cpp/ggml/src/ggml-blas/CMakeLists.txt +101 -0
  95. package/cpp/ggml/src/ggml-blas/ggml-blas.cpp +518 -0
  96. package/cpp/ggml/src/ggml-common.h +1878 -0
  97. package/cpp/ggml/src/ggml-cpu/CMakeLists.txt +691 -0
  98. package/cpp/ggml/src/ggml-cpu/amx/amx.cpp +247 -0
  99. package/cpp/ggml/src/ggml-cpu/amx/amx.h +8 -0
  100. package/cpp/ggml/src/ggml-cpu/amx/common.h +91 -0
  101. package/cpp/ggml/src/ggml-cpu/amx/mmq.cpp +2512 -0
  102. package/cpp/ggml/src/ggml-cpu/amx/mmq.h +10 -0
  103. package/cpp/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp +98 -0
  104. package/cpp/ggml/src/ggml-cpu/arch/arm/quants.c +4052 -0
  105. package/cpp/ggml/src/ggml-cpu/arch/arm/repack.cpp +4935 -0
  106. package/cpp/ggml/src/ggml-cpu/arch/loongarch/quants.c +2159 -0
  107. package/cpp/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp +82 -0
  108. package/cpp/ggml/src/ggml-cpu/arch/powerpc/quants.c +2305 -0
  109. package/cpp/ggml/src/ggml-cpu/arch/riscv/cpu-feats.cpp +38 -0
  110. package/cpp/ggml/src/ggml-cpu/arch/riscv/quants.c +2726 -0
  111. package/cpp/ggml/src/ggml-cpu/arch/riscv/repack.cpp +342 -0
  112. package/cpp/ggml/src/ggml-cpu/arch/s390/cpu-feats.cpp +50 -0
  113. package/cpp/ggml/src/ggml-cpu/arch/s390/quants.c +1468 -0
  114. package/cpp/ggml/src/ggml-cpu/arch/wasm/quants.c +1221 -0
  115. package/cpp/ggml/src/ggml-cpu/arch/x86/cpu-feats.cpp +327 -0
  116. package/cpp/ggml/src/ggml-cpu/arch/x86/quants.c +3820 -0
  117. package/cpp/ggml/src/ggml-cpu/arch/x86/repack.cpp +6307 -0
  118. package/cpp/ggml/src/ggml-cpu/arch-fallback.h +313 -0
  119. package/cpp/ggml/src/ggml-cpu/binary-ops.cpp +154 -0
  120. package/cpp/ggml/src/ggml-cpu/binary-ops.h +16 -0
  121. package/cpp/ggml/src/ggml-cpu/cmake/FindSIMD.cmake +100 -0
  122. package/cpp/ggml/src/ggml-cpu/common.h +95 -0
  123. package/cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h +529 -0
  124. package/cpp/ggml/src/ggml-cpu/ggml-cpu.c +3734 -0
  125. package/cpp/ggml/src/ggml-cpu/ggml-cpu.cpp +701 -0
  126. package/cpp/ggml/src/ggml-cpu/hbm.cpp +55 -0
  127. package/cpp/ggml/src/ggml-cpu/hbm.h +8 -0
  128. package/cpp/ggml/src/ggml-cpu/kleidiai/kernels.cpp +938 -0
  129. package/cpp/ggml/src/ggml-cpu/kleidiai/kernels.h +90 -0
  130. package/cpp/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +798 -0
  131. package/cpp/ggml/src/ggml-cpu/kleidiai/kleidiai.h +17 -0
  132. package/cpp/ggml/src/ggml-cpu/llamafile/sgemm.cpp +4033 -0
  133. package/cpp/ggml/src/ggml-cpu/llamafile/sgemm.h +25 -0
  134. package/cpp/ggml/src/ggml-cpu/ops.cpp +10978 -0
  135. package/cpp/ggml/src/ggml-cpu/ops.h +116 -0
  136. package/cpp/ggml/src/ggml-cpu/quants.c +1193 -0
  137. package/cpp/ggml/src/ggml-cpu/quants.h +97 -0
  138. package/cpp/ggml/src/ggml-cpu/repack.cpp +3316 -0
  139. package/cpp/ggml/src/ggml-cpu/repack.h +173 -0
  140. package/cpp/ggml/src/ggml-cpu/simd-gemm.h +136 -0
  141. package/cpp/ggml/src/ggml-cpu/simd-mappings.h +1279 -0
  142. package/cpp/ggml/src/ggml-cpu/spacemit/ime.cpp +1025 -0
  143. package/cpp/ggml/src/ggml-cpu/spacemit/ime.h +13 -0
  144. package/cpp/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +3196 -0
  145. package/cpp/ggml/src/ggml-cpu/spacemit/ime_kernels.h +26 -0
  146. package/cpp/ggml/src/ggml-cpu/traits.cpp +36 -0
  147. package/cpp/ggml/src/ggml-cpu/traits.h +38 -0
  148. package/cpp/ggml/src/ggml-cpu/unary-ops.cpp +337 -0
  149. package/cpp/ggml/src/ggml-cpu/unary-ops.h +35 -0
  150. package/cpp/ggml/src/ggml-cpu/vec.cpp +629 -0
  151. package/cpp/ggml/src/ggml-cpu/vec.h +1585 -0
  152. package/cpp/ggml/src/ggml-hexagon/CMakeLists.txt +117 -0
  153. package/cpp/ggml/src/ggml-hexagon/ggml-hexagon.cpp +3232 -0
  154. package/cpp/ggml/src/ggml-hexagon/htp/CMakeLists.txt +45 -0
  155. package/cpp/ggml/src/ggml-hexagon/htp/act-ops.c +815 -0
  156. package/cpp/ggml/src/ggml-hexagon/htp/argsort-ops.c +281 -0
  157. package/cpp/ggml/src/ggml-hexagon/htp/binary-ops.c +827 -0
  158. package/cpp/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +157 -0
  159. package/cpp/ggml/src/ggml-hexagon/htp/cpy-ops.c +251 -0
  160. package/cpp/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +666 -0
  161. package/cpp/ggml/src/ggml-hexagon/htp/get-rows-ops.c +111 -0
  162. package/cpp/ggml/src/ggml-hexagon/htp/hex-dma.c +63 -0
  163. package/cpp/ggml/src/ggml-hexagon/htp/hex-dma.h +182 -0
  164. package/cpp/ggml/src/ggml-hexagon/htp/hex-dump.h +77 -0
  165. package/cpp/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
  166. package/cpp/ggml/src/ggml-hexagon/htp/hex-utils.h +51 -0
  167. package/cpp/ggml/src/ggml-hexagon/htp/htp-ctx.h +35 -0
  168. package/cpp/ggml/src/ggml-hexagon/htp/htp-msg.h +154 -0
  169. package/cpp/ggml/src/ggml-hexagon/htp/htp-ops.h +65 -0
  170. package/cpp/ggml/src/ggml-hexagon/htp/htp_iface.idl +16 -0
  171. package/cpp/ggml/src/ggml-hexagon/htp/hvx-arith.h +470 -0
  172. package/cpp/ggml/src/ggml-hexagon/htp/hvx-base.h +173 -0
  173. package/cpp/ggml/src/ggml-hexagon/htp/hvx-copy.h +245 -0
  174. package/cpp/ggml/src/ggml-hexagon/htp/hvx-div.h +116 -0
  175. package/cpp/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
  176. package/cpp/ggml/src/ggml-hexagon/htp/hvx-exp.h +215 -0
  177. package/cpp/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
  178. package/cpp/ggml/src/ggml-hexagon/htp/hvx-inverse.h +176 -0
  179. package/cpp/ggml/src/ggml-hexagon/htp/hvx-reduce.h +266 -0
  180. package/cpp/ggml/src/ggml-hexagon/htp/hvx-scale.h +133 -0
  181. package/cpp/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +141 -0
  182. package/cpp/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +126 -0
  183. package/cpp/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
  184. package/cpp/ggml/src/ggml-hexagon/htp/hvx-utils.h +18 -0
  185. package/cpp/ggml/src/ggml-hexagon/htp/main.c +1150 -0
  186. package/cpp/ggml/src/ggml-hexagon/htp/matmul-ops.c +2595 -0
  187. package/cpp/ggml/src/ggml-hexagon/htp/rope-ops.c +498 -0
  188. package/cpp/ggml/src/ggml-hexagon/htp/set-rows-ops.c +167 -0
  189. package/cpp/ggml/src/ggml-hexagon/htp/softmax-ops.c +421 -0
  190. package/cpp/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +130 -0
  191. package/cpp/ggml/src/ggml-hexagon/htp/unary-ops.c +384 -0
  192. package/cpp/ggml/src/ggml-hexagon/htp/worker-pool.c +293 -0
  193. package/cpp/ggml/src/ggml-hexagon/htp/worker-pool.h +57 -0
  194. package/cpp/ggml/src/ggml-hexagon/htp-drv.cpp +418 -0
  195. package/cpp/ggml/src/ggml-hexagon/htp-drv.h +121 -0
  196. package/cpp/ggml/src/ggml-hexagon/libdl.h +79 -0
  197. package/cpp/ggml/src/ggml-hexagon/libggml-htp.inf +38 -0
  198. package/cpp/ggml/src/ggml-hexagon/op-desc.h +153 -0
  199. package/cpp/ggml/src/ggml-impl.h +724 -0
  200. package/cpp/ggml/src/ggml-metal/CMakeLists.txt +124 -0
  201. package/cpp/ggml/src/ggml-metal/ggml-metal-common.cpp +457 -0
  202. package/cpp/ggml/src/ggml-metal/ggml-metal-common.h +52 -0
  203. package/cpp/ggml/src/ggml-metal/ggml-metal-context.h +41 -0
  204. package/cpp/ggml/src/ggml-metal/ggml-metal-context.m +702 -0
  205. package/cpp/ggml/src/ggml-metal/ggml-metal-device.cpp +1890 -0
  206. package/cpp/ggml/src/ggml-metal/ggml-metal-device.h +290 -0
  207. package/cpp/ggml/src/ggml-metal/ggml-metal-device.m +1749 -0
  208. package/cpp/ggml/src/ggml-metal/ggml-metal-impl.h +1054 -0
  209. package/cpp/ggml/src/ggml-metal/ggml-metal-ops.cpp +4370 -0
  210. package/cpp/ggml/src/ggml-metal/ggml-metal-ops.h +94 -0
  211. package/cpp/ggml/src/ggml-metal/ggml-metal.cpp +937 -0
  212. package/cpp/ggml/src/ggml-metal/ggml-metal.metal +9819 -0
  213. package/cpp/ggml/src/ggml-musa/CMakeLists.txt +125 -0
  214. package/cpp/ggml/src/ggml-musa/mudnn.cu +112 -0
  215. package/cpp/ggml/src/ggml-musa/mudnn.cuh +12 -0
  216. package/cpp/ggml/src/ggml-opencl/CMakeLists.txt +150 -0
  217. package/cpp/ggml/src/ggml-opencl/ggml-opencl.cpp +11553 -0
  218. package/cpp/ggml/src/ggml-opencl/kernels/add.cl +190 -0
  219. package/cpp/ggml/src/ggml-opencl/kernels/add_id.cl +42 -0
  220. package/cpp/ggml/src/ggml-opencl/kernels/argsort.cl +86 -0
  221. package/cpp/ggml/src/ggml-opencl/kernels/clamp.cl +20 -0
  222. package/cpp/ggml/src/ggml-opencl/kernels/concat.cl +51 -0
  223. package/cpp/ggml/src/ggml-opencl/kernels/conv2d.cl +185 -0
  224. package/cpp/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl +176 -0
  225. package/cpp/ggml/src/ggml-opencl/kernels/cpy.cl +184 -0
  226. package/cpp/ggml/src/ggml-opencl/kernels/cvt.cl +417 -0
  227. package/cpp/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl +58 -0
  228. package/cpp/ggml/src/ggml-opencl/kernels/div.cl +138 -0
  229. package/cpp/ggml/src/ggml-opencl/kernels/embed_kernel.py +26 -0
  230. package/cpp/ggml/src/ggml-opencl/kernels/expm1.cl +113 -0
  231. package/cpp/ggml/src/ggml-opencl/kernels/fill.cl +17 -0
  232. package/cpp/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl +370 -0
  233. package/cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +371 -0
  234. package/cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl +373 -0
  235. package/cpp/ggml/src/ggml-opencl/kernels/gelu.cl +89 -0
  236. package/cpp/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32.cl +162 -0
  237. package/cpp/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl +156 -0
  238. package/cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle.cl +268 -0
  239. package/cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general.cl +274 -0
  240. package/cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl +195 -0
  241. package/cpp/ggml/src/ggml-opencl/kernels/get_rows.cl +187 -0
  242. package/cpp/ggml/src/ggml-opencl/kernels/glu.cl +378 -0
  243. package/cpp/ggml/src/ggml-opencl/kernels/group_norm.cl +121 -0
  244. package/cpp/ggml/src/ggml-opencl/kernels/im2col_f16.cl +57 -0
  245. package/cpp/ggml/src/ggml-opencl/kernels/im2col_f32.cl +57 -0
  246. package/cpp/ggml/src/ggml-opencl/kernels/mean.cl +140 -0
  247. package/cpp/ggml/src/ggml-opencl/kernels/mul.cl +152 -0
  248. package/cpp/ggml/src/ggml-opencl/kernels/mul_mat_Ab_Bi_8x4.cl +139 -0
  249. package/cpp/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl +130 -0
  250. package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_kq_kqv.cl +273 -0
  251. package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl +146 -0
  252. package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl +147 -0
  253. package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
  254. package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
  255. package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
  256. package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl +129 -0
  257. package/cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl +154 -0
  258. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl +118 -0
  259. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl +118 -0
  260. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl +94 -0
  261. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl +84 -0
  262. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl +118 -0
  263. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl +189 -0
  264. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl +176 -0
  265. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl +283 -0
  266. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl +140 -0
  267. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl +222 -0
  268. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl +144 -0
  269. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl +167 -0
  270. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl +192 -0
  271. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl +307 -0
  272. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl +265 -0
  273. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl +272 -0
  274. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl +254 -0
  275. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
  276. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
  277. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +180 -0
  278. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl +194 -0
  279. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +194 -0
  280. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl +125 -0
  281. package/cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl +202 -0
  282. package/cpp/ggml/src/ggml-opencl/kernels/norm.cl +161 -0
  283. package/cpp/ggml/src/ggml-opencl/kernels/pad.cl +39 -0
  284. package/cpp/ggml/src/ggml-opencl/kernels/relu.cl +16 -0
  285. package/cpp/ggml/src/ggml-opencl/kernels/repeat.cl +38 -0
  286. package/cpp/ggml/src/ggml-opencl/kernels/rms_norm.cl +190 -0
  287. package/cpp/ggml/src/ggml-opencl/kernels/rope.cl +747 -0
  288. package/cpp/ggml/src/ggml-opencl/kernels/scale.cl +27 -0
  289. package/cpp/ggml/src/ggml-opencl/kernels/set_rows.cl +208 -0
  290. package/cpp/ggml/src/ggml-opencl/kernels/sigmoid.cl +29 -0
  291. package/cpp/ggml/src/ggml-opencl/kernels/silu.cl +30 -0
  292. package/cpp/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +108 -0
  293. package/cpp/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +108 -0
  294. package/cpp/ggml/src/ggml-opencl/kernels/softmax_f16.cl +107 -0
  295. package/cpp/ggml/src/ggml-opencl/kernels/softmax_f32.cl +107 -0
  296. package/cpp/ggml/src/ggml-opencl/kernels/softplus.cl +116 -0
  297. package/cpp/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
  298. package/cpp/ggml/src/ggml-opencl/kernels/sqr.cl +53 -0
  299. package/cpp/ggml/src/ggml-opencl/kernels/sqrt.cl +53 -0
  300. package/cpp/ggml/src/ggml-opencl/kernels/ssm_conv.cl +77 -0
  301. package/cpp/ggml/src/ggml-opencl/kernels/sub.cl +138 -0
  302. package/cpp/ggml/src/ggml-opencl/kernels/sum_rows.cl +140 -0
  303. package/cpp/ggml/src/ggml-opencl/kernels/tanh.cl +109 -0
  304. package/cpp/ggml/src/ggml-opencl/kernels/transpose.cl +117 -0
  305. package/cpp/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
  306. package/cpp/ggml/src/ggml-opencl/kernels/tsembd.cl +48 -0
  307. package/cpp/ggml/src/ggml-opencl/kernels/upscale.cl +120 -0
  308. package/cpp/ggml/src/ggml-opt.cpp +1093 -0
  309. package/cpp/ggml/src/ggml-quants.c +5325 -0
  310. package/cpp/ggml/src/ggml-quants.h +106 -0
  311. package/cpp/ggml/src/ggml-rpc/CMakeLists.txt +9 -0
  312. package/cpp/ggml/src/ggml-rpc/ggml-rpc.cpp +2118 -0
  313. package/cpp/ggml/src/ggml-threading.cpp +12 -0
  314. package/cpp/ggml/src/ggml-threading.h +14 -0
  315. package/cpp/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
  316. package/cpp/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
  317. package/cpp/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
  318. package/cpp/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
  319. package/cpp/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
  320. package/cpp/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
  321. package/cpp/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
  322. package/cpp/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
  323. package/cpp/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +148 -0
  324. package/cpp/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
  325. package/cpp/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
  326. package/cpp/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
  327. package/cpp/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
  328. package/cpp/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
  329. package/cpp/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
  330. package/cpp/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
  331. package/cpp/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
  332. package/cpp/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
  333. package/cpp/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
  334. package/cpp/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
  335. package/cpp/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
  336. package/cpp/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +119 -0
  337. package/cpp/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +158 -0
  338. package/cpp/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
  339. package/cpp/ggml/src/ggml-virtgpu/ggml-backend.cpp +69 -0
  340. package/cpp/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
  341. package/cpp/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
  342. package/cpp/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
  343. package/cpp/ggml/src/ggml-virtgpu/regenerate_remoting.py +333 -0
  344. package/cpp/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
  345. package/cpp/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
  346. package/cpp/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
  347. package/cpp/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
  348. package/cpp/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +192 -0
  349. package/cpp/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
  350. package/cpp/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +53 -0
  351. package/cpp/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +98 -0
  352. package/cpp/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
  353. package/cpp/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
  354. package/cpp/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
  355. package/cpp/ggml/src/ggml-virtgpu/virtgpu.cpp +544 -0
  356. package/cpp/ggml/src/ggml-virtgpu/virtgpu.h +117 -0
  357. package/cpp/ggml/src/ggml-webgpu/CMakeLists.txt +80 -0
  358. package/cpp/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +1231 -0
  359. package/cpp/ggml/src/ggml-webgpu/ggml-webgpu.cpp +3150 -0
  360. package/cpp/ggml/src/ggml-webgpu/pre_wgsl.hpp +778 -0
  361. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
  362. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +106 -0
  363. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
  364. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +107 -0
  365. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +923 -0
  366. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/cpy.tmpl.wgsl +107 -0
  367. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
  368. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +182 -0
  369. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +636 -0
  370. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.wgsl +668 -0
  371. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/glu.tmpl.wgsl +323 -0
  372. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl +40 -0
  373. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl +713 -0
  374. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +103 -0
  375. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +138 -0
  376. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.wgsl +188 -0
  377. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +194 -0
  378. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
  379. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +123 -0
  380. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/rope.tmpl.wgsl +295 -0
  381. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/scale.wgsl +63 -0
  382. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +109 -0
  383. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/soft_max.tmpl.wgsl +345 -0
  384. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
  385. package/cpp/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +193 -0
  386. package/cpp/ggml/src/ggml-zdnn/CMakeLists.txt +36 -0
  387. package/cpp/ggml/src/ggml-zdnn/common.hpp +59 -0
  388. package/cpp/ggml/src/ggml-zdnn/ggml-zdnn.cpp +633 -0
  389. package/cpp/ggml/src/ggml-zdnn/mmf.cpp +80 -0
  390. package/cpp/ggml/src/ggml-zdnn/mmf.hpp +12 -0
  391. package/cpp/ggml/src/ggml-zdnn/utils.cpp +79 -0
  392. package/cpp/ggml/src/ggml-zdnn/utils.hpp +19 -0
  393. package/cpp/ggml/src/ggml-zendnn/CMakeLists.txt +92 -0
  394. package/cpp/ggml/src/ggml-zendnn/ggml-zendnn.cpp +469 -0
  395. package/cpp/ggml/src/ggml.c +7669 -0
  396. package/cpp/ggml/src/ggml.cpp +26 -0
  397. package/cpp/ggml/src/gguf.cpp +1699 -0
  398. package/cpp/include/llama-cpp.h +32 -0
  399. package/cpp/include/llama.h +1568 -0
  400. package/cpp/mtmd/CMakeLists.txt +98 -0
  401. package/cpp/mtmd/README.md +63 -0
  402. package/cpp/mtmd/clip-graph.h +117 -0
  403. package/cpp/mtmd/clip-impl.h +586 -0
  404. package/cpp/mtmd/clip-model.h +390 -0
  405. package/cpp/mtmd/clip.cpp +4154 -0
  406. package/cpp/mtmd/clip.h +121 -0
  407. package/cpp/mtmd/deprecation-warning.cpp +22 -0
  408. package/cpp/mtmd/legacy-models/convert_image_encoder_to_gguf.py +412 -0
  409. package/cpp/mtmd/legacy-models/glmedge-convert-image-encoder-to-gguf.py +280 -0
  410. package/cpp/mtmd/legacy-models/glmedge-surgery.py +33 -0
  411. package/cpp/mtmd/legacy-models/llava_surgery.py +38 -0
  412. package/cpp/mtmd/legacy-models/llava_surgery_v2.py +180 -0
  413. package/cpp/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py +892 -0
  414. package/cpp/mtmd/legacy-models/minicpmv-surgery.py +47 -0
  415. package/cpp/mtmd/models/cogvlm.cpp +98 -0
  416. package/cpp/mtmd/models/conformer.cpp +216 -0
  417. package/cpp/mtmd/models/glm4v.cpp +122 -0
  418. package/cpp/mtmd/models/internvl.cpp +69 -0
  419. package/cpp/mtmd/models/kimik25.cpp +101 -0
  420. package/cpp/mtmd/models/kimivl.cpp +63 -0
  421. package/cpp/mtmd/models/llama4.cpp +96 -0
  422. package/cpp/mtmd/models/llava.cpp +374 -0
  423. package/cpp/mtmd/models/minicpmv.cpp +114 -0
  424. package/cpp/mtmd/models/mobilenetv5.cpp +451 -0
  425. package/cpp/mtmd/models/models.h +128 -0
  426. package/cpp/mtmd/models/nemotron-v2-vl.cpp +35 -0
  427. package/cpp/mtmd/models/paddleocr.cpp +52 -0
  428. package/cpp/mtmd/models/pixtral.cpp +86 -0
  429. package/cpp/mtmd/models/qwen2vl.cpp +183 -0
  430. package/cpp/mtmd/models/qwen3vl.cpp +193 -0
  431. package/cpp/mtmd/models/siglip.cpp +86 -0
  432. package/cpp/mtmd/models/whisper-enc.cpp +115 -0
  433. package/cpp/mtmd/models/youtuvl.cpp +179 -0
  434. package/cpp/mtmd/mtmd-audio.cpp +730 -0
  435. package/cpp/mtmd/mtmd-audio.h +113 -0
  436. package/cpp/mtmd/mtmd-cli.cpp +437 -0
  437. package/cpp/mtmd/mtmd-helper.cpp +521 -0
  438. package/cpp/mtmd/mtmd-helper.h +96 -0
  439. package/cpp/mtmd/mtmd.cpp +1156 -0
  440. package/cpp/mtmd/mtmd.h +319 -0
  441. package/cpp/mtmd/requirements.txt +5 -0
  442. package/cpp/mtmd/test-1.jpeg +0 -0
  443. package/cpp/mtmd/test-2.mp3 +0 -0
  444. package/cpp/mtmd/tests.sh +192 -0
  445. package/cpp/src/CMakeLists.txt +169 -0
  446. package/cpp/src/llama-adapter.cpp +488 -0
  447. package/cpp/src/llama-adapter.h +89 -0
  448. package/cpp/src/llama-arch.cpp +2855 -0
  449. package/cpp/src/llama-arch.h +619 -0
  450. package/cpp/src/llama-batch.cpp +917 -0
  451. package/cpp/src/llama-batch.h +173 -0
  452. package/cpp/src/llama-chat.cpp +896 -0
  453. package/cpp/src/llama-chat.h +71 -0
  454. package/cpp/src/llama-context.cpp +3512 -0
  455. package/cpp/src/llama-context.h +359 -0
  456. package/cpp/src/llama-cparams.cpp +5 -0
  457. package/cpp/src/llama-cparams.h +44 -0
  458. package/cpp/src/llama-grammar.cpp +1464 -0
  459. package/cpp/src/llama-grammar.h +194 -0
  460. package/cpp/src/llama-graph.cpp +2685 -0
  461. package/cpp/src/llama-graph.h +1026 -0
  462. package/cpp/src/llama-hparams.cpp +234 -0
  463. package/cpp/src/llama-hparams.h +339 -0
  464. package/cpp/src/llama-impl.cpp +171 -0
  465. package/cpp/src/llama-impl.h +73 -0
  466. package/cpp/src/llama-io.cpp +15 -0
  467. package/cpp/src/llama-io.h +35 -0
  468. package/cpp/src/llama-kv-cache-iswa.cpp +330 -0
  469. package/cpp/src/llama-kv-cache-iswa.h +137 -0
  470. package/cpp/src/llama-kv-cache.cpp +2271 -0
  471. package/cpp/src/llama-kv-cache.h +388 -0
  472. package/cpp/src/llama-kv-cells.h +533 -0
  473. package/cpp/src/llama-memory-hybrid-iswa.cpp +275 -0
  474. package/cpp/src/llama-memory-hybrid-iswa.h +140 -0
  475. package/cpp/src/llama-memory-hybrid.cpp +268 -0
  476. package/cpp/src/llama-memory-hybrid.h +139 -0
  477. package/cpp/src/llama-memory-recurrent.cpp +1165 -0
  478. package/cpp/src/llama-memory-recurrent.h +182 -0
  479. package/cpp/src/llama-memory.cpp +59 -0
  480. package/cpp/src/llama-memory.h +122 -0
  481. package/cpp/src/llama-mmap.cpp +785 -0
  482. package/cpp/src/llama-mmap.h +92 -0
  483. package/cpp/src/llama-model-loader.cpp +1414 -0
  484. package/cpp/src/llama-model-loader.h +203 -0
  485. package/cpp/src/llama-model-saver.cpp +286 -0
  486. package/cpp/src/llama-model-saver.h +37 -0
  487. package/cpp/src/llama-model.cpp +9253 -0
  488. package/cpp/src/llama-model.h +576 -0
  489. package/cpp/src/llama-quant.cpp +1119 -0
  490. package/cpp/src/llama-quant.h +1 -0
  491. package/cpp/src/llama-sampler.cpp +3885 -0
  492. package/cpp/src/llama-sampler.h +42 -0
  493. package/cpp/src/llama-vocab.cpp +3970 -0
  494. package/cpp/src/llama-vocab.h +187 -0
  495. package/cpp/src/llama.cpp +1313 -0
  496. package/cpp/src/models/afmoe.cpp +191 -0
  497. package/cpp/src/models/apertus.cpp +125 -0
  498. package/cpp/src/models/arcee.cpp +135 -0
  499. package/cpp/src/models/arctic.cpp +138 -0
  500. package/cpp/src/models/arwkv7.cpp +86 -0
  501. package/cpp/src/models/baichuan.cpp +122 -0
  502. package/cpp/src/models/bailingmoe.cpp +144 -0
  503. package/cpp/src/models/bailingmoe2.cpp +135 -0
  504. package/cpp/src/models/bert.cpp +178 -0
  505. package/cpp/src/models/bitnet.cpp +160 -0
  506. package/cpp/src/models/bloom.cpp +101 -0
  507. package/cpp/src/models/chameleon.cpp +178 -0
  508. package/cpp/src/models/chatglm.cpp +132 -0
  509. package/cpp/src/models/codeshell.cpp +111 -0
  510. package/cpp/src/models/cogvlm.cpp +102 -0
  511. package/cpp/src/models/cohere2-iswa.cpp +134 -0
  512. package/cpp/src/models/command-r.cpp +122 -0
  513. package/cpp/src/models/dbrx.cpp +123 -0
  514. package/cpp/src/models/deci.cpp +135 -0
  515. package/cpp/src/models/deepseek.cpp +144 -0
  516. package/cpp/src/models/deepseek2.cpp +262 -0
  517. package/cpp/src/models/delta-net-base.cpp +376 -0
  518. package/cpp/src/models/dots1.cpp +134 -0
  519. package/cpp/src/models/dream.cpp +105 -0
  520. package/cpp/src/models/ernie4-5-moe.cpp +150 -0
  521. package/cpp/src/models/ernie4-5.cpp +110 -0
  522. package/cpp/src/models/eurobert.cpp +97 -0
  523. package/cpp/src/models/exaone-moe.cpp +146 -0
  524. package/cpp/src/models/exaone.cpp +114 -0
  525. package/cpp/src/models/exaone4.cpp +123 -0
  526. package/cpp/src/models/falcon-h1.cpp +111 -0
  527. package/cpp/src/models/falcon.cpp +120 -0
  528. package/cpp/src/models/gemma-embedding.cpp +116 -0
  529. package/cpp/src/models/gemma.cpp +112 -0
  530. package/cpp/src/models/gemma2-iswa.cpp +128 -0
  531. package/cpp/src/models/gemma3.cpp +155 -0
  532. package/cpp/src/models/gemma3n-iswa.cpp +384 -0
  533. package/cpp/src/models/glm4-moe.cpp +170 -0
  534. package/cpp/src/models/glm4.cpp +157 -0
  535. package/cpp/src/models/gpt2.cpp +105 -0
  536. package/cpp/src/models/gptneox.cpp +144 -0
  537. package/cpp/src/models/granite-hybrid.cpp +196 -0
  538. package/cpp/src/models/granite.cpp +211 -0
  539. package/cpp/src/models/grok.cpp +159 -0
  540. package/cpp/src/models/grovemoe.cpp +141 -0
  541. package/cpp/src/models/hunyuan-dense.cpp +132 -0
  542. package/cpp/src/models/hunyuan-moe.cpp +154 -0
  543. package/cpp/src/models/internlm2.cpp +120 -0
  544. package/cpp/src/models/jais.cpp +86 -0
  545. package/cpp/src/models/jais2.cpp +123 -0
  546. package/cpp/src/models/jamba.cpp +106 -0
  547. package/cpp/src/models/kimi-linear.cpp +392 -0
  548. package/cpp/src/models/lfm2.cpp +190 -0
  549. package/cpp/src/models/llada-moe.cpp +122 -0
  550. package/cpp/src/models/llada.cpp +99 -0
  551. package/cpp/src/models/llama-iswa.cpp +178 -0
  552. package/cpp/src/models/llama.cpp +168 -0
  553. package/cpp/src/models/maincoder.cpp +117 -0
  554. package/cpp/src/models/mamba-base.cpp +285 -0
  555. package/cpp/src/models/mamba.cpp +54 -0
  556. package/cpp/src/models/mimo2-iswa.cpp +123 -0
  557. package/cpp/src/models/minicpm3.cpp +200 -0
  558. package/cpp/src/models/minimax-m2.cpp +124 -0
  559. package/cpp/src/models/mistral3.cpp +160 -0
  560. package/cpp/src/models/models.h +684 -0
  561. package/cpp/src/models/modern-bert.cpp +109 -0
  562. package/cpp/src/models/mpt.cpp +126 -0
  563. package/cpp/src/models/nemotron-h.cpp +148 -0
  564. package/cpp/src/models/nemotron.cpp +122 -0
  565. package/cpp/src/models/neo-bert.cpp +104 -0
  566. package/cpp/src/models/olmo.cpp +121 -0
  567. package/cpp/src/models/olmo2.cpp +150 -0
  568. package/cpp/src/models/olmoe.cpp +124 -0
  569. package/cpp/src/models/openai-moe-iswa.cpp +127 -0
  570. package/cpp/src/models/openelm.cpp +124 -0
  571. package/cpp/src/models/orion.cpp +123 -0
  572. package/cpp/src/models/paddleocr.cpp +122 -0
  573. package/cpp/src/models/pangu-embedded.cpp +121 -0
  574. package/cpp/src/models/phi2.cpp +121 -0
  575. package/cpp/src/models/phi3.cpp +152 -0
  576. package/cpp/src/models/plamo.cpp +110 -0
  577. package/cpp/src/models/plamo2.cpp +318 -0
  578. package/cpp/src/models/plamo3.cpp +128 -0
  579. package/cpp/src/models/plm.cpp +169 -0
  580. package/cpp/src/models/qwen.cpp +108 -0
  581. package/cpp/src/models/qwen2.cpp +126 -0
  582. package/cpp/src/models/qwen2moe.cpp +151 -0
  583. package/cpp/src/models/qwen2vl.cpp +117 -0
  584. package/cpp/src/models/qwen3.cpp +117 -0
  585. package/cpp/src/models/qwen35.cpp +386 -0
  586. package/cpp/src/models/qwen35moe.cpp +420 -0
  587. package/cpp/src/models/qwen3moe.cpp +124 -0
  588. package/cpp/src/models/qwen3next.cpp +525 -0
  589. package/cpp/src/models/qwen3vl-moe.cpp +140 -0
  590. package/cpp/src/models/qwen3vl.cpp +132 -0
  591. package/cpp/src/models/refact.cpp +94 -0
  592. package/cpp/src/models/rnd1.cpp +126 -0
  593. package/cpp/src/models/rwkv6-base.cpp +164 -0
  594. package/cpp/src/models/rwkv6.cpp +94 -0
  595. package/cpp/src/models/rwkv6qwen2.cpp +86 -0
  596. package/cpp/src/models/rwkv7-base.cpp +137 -0
  597. package/cpp/src/models/rwkv7.cpp +90 -0
  598. package/cpp/src/models/seed-oss.cpp +124 -0
  599. package/cpp/src/models/smallthinker.cpp +126 -0
  600. package/cpp/src/models/smollm3.cpp +128 -0
  601. package/cpp/src/models/stablelm.cpp +146 -0
  602. package/cpp/src/models/starcoder.cpp +100 -0
  603. package/cpp/src/models/starcoder2.cpp +121 -0
  604. package/cpp/src/models/step35-iswa.cpp +168 -0
  605. package/cpp/src/models/t5-dec.cpp +166 -0
  606. package/cpp/src/models/t5-enc.cpp +96 -0
  607. package/cpp/src/models/wavtokenizer-dec.cpp +149 -0
  608. package/cpp/src/models/xverse.cpp +108 -0
  609. package/cpp/src/unicode-data.cpp +7034 -0
  610. package/cpp/src/unicode-data.h +20 -0
  611. package/cpp/src/unicode.cpp +1103 -0
  612. package/cpp/src/unicode.h +111 -0
  613. package/cpp/vendor/nlohmann/json.hpp +25526 -0
  614. package/cpp/vendor/nlohmann/json_fwd.hpp +187 -0
  615. package/cpp/vendor/stb/stb_image.h +7988 -0
  616. package/ios/LocalLLM-Bridging-Header.h +2 -0
  617. package/ios/LocalLLM.h +5 -0
  618. package/ios/LocalLLM.mm +1267 -0
  619. package/local-llm-rn.podspec +60 -0
  620. package/package.json +35 -0
  621. package/src/NativeLocalLLM.ts +73 -0
  622. package/src/device.ts +50 -0
  623. package/src/download-adapter.ts +17 -0
  624. package/src/index.ts +21 -0
  625. package/src/native-bridge.ts +142 -0
  626. package/src/rn-downloader.ts +37 -0
@@ -0,0 +1,133 @@
1
+ #pragma once
2
+
3
+ #include "chat.h"
4
+ #include "chat-parser-xml-toolcall.h"
5
+ #include "json-partial.h"
6
+ #include "regex-partial.h"
7
+
8
+ #include <nlohmann/json_fwd.hpp>
9
+
10
+ #include <optional>
11
+ #include <string>
12
+ #include <vector>
13
+
14
+ class common_chat_msg_partial_exception : public std::runtime_error {
15
+ public:
16
+ common_chat_msg_partial_exception(const std::string & message) : std::runtime_error(message) {}
17
+ };
18
+
19
+ class common_chat_msg_parser {
20
+ std::string input_;
21
+ bool is_partial_;
22
+ common_chat_parser_params syntax_; // TODO: rename to params
23
+ std::string healing_marker_;
24
+
25
+ size_t pos_ = 0;
26
+ common_chat_msg result_;
27
+
28
+ public:
29
+ common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_parser_params & syntax);
30
+ const std::string & input() const { return input_; }
31
+ size_t pos() const { return pos_; }
32
+ const std::string & healing_marker() const { return healing_marker_; }
33
+ const bool & is_partial() const { return is_partial_; }
34
+ const common_chat_msg & result() const { return result_; }
35
+ const common_chat_parser_params & syntax() const { return syntax_; }
36
+
37
+ void move_to(size_t pos) {
38
+ if (pos > input_.size()) {
39
+ throw std::runtime_error("Invalid position!");
40
+ }
41
+ pos_ = pos;
42
+ }
43
+ void move_back(size_t n) {
44
+ if (pos_ < n) {
45
+ throw std::runtime_error("Can't move back that far!");
46
+ }
47
+ pos_ -= n;
48
+ }
49
+
50
+ // Get the substring of the input at the given range
51
+ std::string str(const common_string_range & rng) const;
52
+
53
+ // Appends to the result.content field
54
+ void add_content(const std::string & content);
55
+
56
+ // Appends to the result.reasoning_content field
57
+ void add_reasoning_content(const std::string & reasoning_content);
58
+
59
+ // Adds a tool call to the result. If the tool call is too incomplete (e.g. name empty), it won't add anything.
60
+ bool add_tool_call(const std::string & name, const std::string & id, const std::string & arguments);
61
+
62
+ // Adds a tool call using the "name", "id" and "arguments" fields of the json object
63
+ bool add_tool_call(const nlohmann::ordered_json & tool_call);
64
+
65
+ // Adds an array of tool calls using their "name", "id" and "arguments" fields.
66
+ bool add_tool_calls(const nlohmann::ordered_json & arr);
67
+
68
+ // Adds a tool call using the short form: { "tool_name": { "arg1": val, "arg2": val } }
69
+ bool add_tool_call_short_form(const nlohmann::ordered_json & tool_call);
70
+
71
+ void finish();
72
+
73
+ bool consume_spaces();
74
+
75
+ void consume_literal(const std::string & literal);
76
+
77
+ bool try_parse_reasoning(const std::string & start_think, const std::string & end_think);
78
+
79
+ std::string consume_rest();
80
+
81
+ struct find_regex_result {
82
+ std::string prelude;
83
+ std::vector<common_string_range> groups;
84
+ };
85
+
86
+ std::optional<find_regex_result> try_find_regex(const common_regex & regex, size_t from = std::string::npos, bool add_prelude_to_content = true);
87
+
88
+ bool try_consume_literal(const std::string & literal);
89
+
90
+ std::optional<find_regex_result> try_find_literal(const std::string & literal);
91
+
92
+ find_regex_result consume_regex(const common_regex & regex);
93
+
94
+ std::optional<find_regex_result> try_consume_regex(const common_regex & regex);
95
+
96
+ std::optional<common_json> try_consume_json();
97
+ common_json consume_json();
98
+
99
+ struct consume_json_result {
100
+ nlohmann::ordered_json value;
101
+ bool is_partial;
102
+ };
103
+
104
+ /*
105
+ Consume (possibly partial) json and converts specific subtrees to (possibly truncated) JSON strings.
106
+
107
+ By default, object keys can't be truncated, nor can string values (their corresponding key is removed,
108
+ e.g. `{"foo": "bar", "baz": "b` -> `{"foo": "bar"}`
109
+
110
+ But one can allow subpaths to be kept truncated, and possibly json-dumped to truncated json strings
111
+ - with `content_paths={{"foo"}}` -> `{"foo": "b` -> {"foo": "b"}`
112
+ - with `args_paths={{"foo"}}` -> `{"foo": {"b` -> `{"foo": "{b"}`
113
+ */
114
+ consume_json_result consume_json_with_dumped_args(
115
+ const std::vector<std::vector<std::string>> & args_paths = {},
116
+ const std::vector<std::vector<std::string>> & content_paths = {}
117
+ );
118
+ std::optional<consume_json_result> try_consume_json_with_dumped_args(
119
+ const std::vector<std::vector<std::string>> & args_paths = {},
120
+ const std::vector<std::vector<std::string>> & content_paths = {}
121
+ );
122
+
123
+ /**
124
+ * Parse XML-Style tool call for given xml_tool_call_format. Return false for invalid syntax and get the position untouched.
125
+ * form.scope_start, form.tool_sep and form.scope_end can be empty.
126
+ */
127
+ bool try_consume_xml_tool_calls(const struct xml_tool_call_format & form);
128
+
129
+ // Parse content uses reasoning and XML-Style tool call
130
+ void consume_reasoning_with_xml_tool_calls(const struct xml_tool_call_format & form, const std::string & start_think = "<think>", const std::string & end_think = "</think>");
131
+
132
+ void clear_tools();
133
+ };
@@ -0,0 +1,124 @@
1
+ #include "chat-peg-parser.h"
2
+
3
+ #include <nlohmann/json.hpp>
4
+
5
+ using json = nlohmann::json;
6
+
7
+ static std::string_view trim_trailing_space(std::string_view sv, int max = -1) {
8
+ int count = 0;
9
+ while (!sv.empty() && std::isspace(static_cast<unsigned char>(sv.back()))) {
10
+ if (max != -1 && count <= max) {
11
+ break;
12
+ }
13
+ sv.remove_suffix(1);
14
+ count++;
15
+ }
16
+ return sv;
17
+ }
18
+
19
+ void common_chat_peg_mapper::from_ast(const common_peg_ast_arena & arena, const common_peg_parse_result & result) {
20
+ arena.visit(result, [this](const common_peg_ast_node & node) {
21
+ map(node);
22
+ });
23
+ }
24
+
25
+ void common_chat_peg_mapper::map(const common_peg_ast_node & node) {
26
+ bool is_reasoning = node.tag == common_chat_peg_builder::REASONING;
27
+ bool is_content = node.tag == common_chat_peg_builder::CONTENT;
28
+
29
+ if (is_reasoning) {
30
+ result.reasoning_content = std::string(trim_trailing_space(node.text));
31
+ }
32
+
33
+ if (is_content) {
34
+ result.content = std::string(trim_trailing_space(node.text));
35
+ }
36
+ }
37
+
38
+ void common_chat_peg_native_mapper::map(const common_peg_ast_node & node) {
39
+ common_chat_peg_mapper::map(node);
40
+
41
+ bool is_tool_open = node.tag == common_chat_peg_native_builder::TOOL_OPEN;
42
+ bool is_tool_name = node.tag == common_chat_peg_native_builder::TOOL_NAME;
43
+ bool is_tool_id = node.tag == common_chat_peg_native_builder::TOOL_ID;
44
+ bool is_tool_args = node.tag == common_chat_peg_native_builder::TOOL_ARGS;
45
+
46
+ if (is_tool_open) {
47
+ result.tool_calls.emplace_back();
48
+ current_tool = &result.tool_calls.back();
49
+ }
50
+
51
+ if (is_tool_id && current_tool) {
52
+ current_tool->id = std::string(trim_trailing_space(node.text));
53
+ }
54
+
55
+ if (is_tool_name && current_tool) {
56
+ current_tool->name = std::string(trim_trailing_space(node.text));
57
+ }
58
+
59
+ if (is_tool_args && current_tool) {
60
+ current_tool->arguments = std::string(trim_trailing_space(node.text));
61
+ }
62
+ }
63
+
64
+ void common_chat_peg_constructed_mapper::map(const common_peg_ast_node & node) {
65
+ common_chat_peg_mapper::map(node);
66
+
67
+ bool is_tool_open = node.tag == common_chat_peg_constructed_builder::TOOL_OPEN;
68
+ bool is_tool_name = node.tag == common_chat_peg_constructed_builder::TOOL_NAME;
69
+ bool is_tool_close = node.tag == common_chat_peg_constructed_builder::TOOL_CLOSE;
70
+ bool is_arg_open = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_OPEN;
71
+ bool is_arg_close = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_CLOSE;
72
+ bool is_arg_name = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_NAME;
73
+ bool is_arg_string = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_STRING_VALUE;
74
+ bool is_arg_json = node.tag == common_chat_peg_constructed_builder::TOOL_ARG_JSON_VALUE;
75
+
76
+ if (is_tool_open) {
77
+ result.tool_calls.emplace_back();
78
+ current_tool = &result.tool_calls.back();
79
+ arg_count = 0;
80
+ }
81
+
82
+ if (is_tool_name) {
83
+ current_tool->name = std::string(node.text);
84
+ current_tool->arguments = "{";
85
+ }
86
+
87
+ if (is_arg_open) {
88
+ needs_closing_quote = false;
89
+ }
90
+
91
+ if (is_arg_name && current_tool) {
92
+ if (arg_count > 0) {
93
+ current_tool->arguments += ",";
94
+ }
95
+ current_tool->arguments += json(trim_trailing_space(node.text)).dump() + ":";
96
+ ++arg_count;
97
+ }
98
+
99
+ if (is_arg_string && current_tool) {
100
+ // Serialize to JSON, but exclude the end quote
101
+ std::string dumped = json(trim_trailing_space(node.text)).dump();
102
+ current_tool->arguments += dumped.substr(0, dumped.size() - 1);
103
+ needs_closing_quote = true;
104
+ }
105
+
106
+ if (is_arg_close && current_tool) {
107
+ if (needs_closing_quote) {
108
+ current_tool->arguments += "\"";
109
+ needs_closing_quote = false;
110
+ }
111
+ }
112
+
113
+ if (is_arg_json && current_tool) {
114
+ current_tool->arguments += std::string(trim_trailing_space(node.text));
115
+ }
116
+
117
+ if (is_tool_close && current_tool) {
118
+ if (needs_closing_quote) {
119
+ current_tool->arguments += "\"";
120
+ needs_closing_quote = false;
121
+ }
122
+ current_tool->arguments += "}";
123
+ }
124
+ }
@@ -0,0 +1,105 @@
1
+ #pragma once
2
+
3
+ #include "chat.h"
4
+ #include "peg-parser.h"
5
+
6
+ class common_chat_peg_builder : public common_peg_parser_builder {
7
+ public:
8
+ static constexpr const char * REASONING_BLOCK = "reasoning-block";
9
+ static constexpr const char * REASONING = "reasoning";
10
+ static constexpr const char * CONTENT = "content";
11
+
12
+ common_peg_parser reasoning_block(const common_peg_parser & p) { return tag(REASONING_BLOCK, p); }
13
+ common_peg_parser reasoning(const common_peg_parser & p) { return tag(REASONING, p); }
14
+ common_peg_parser content(const common_peg_parser & p) { return tag(CONTENT, p); }
15
+ };
16
+
17
+ inline common_peg_arena build_chat_peg_parser(const std::function<common_peg_parser(common_chat_peg_builder & builder)> & fn) {
18
+ common_chat_peg_builder builder;
19
+ builder.set_root(fn(builder));
20
+ return builder.build();
21
+ }
22
+
23
+ class common_chat_peg_mapper {
24
+ public:
25
+ common_chat_msg & result;
26
+
27
+ common_chat_peg_mapper(common_chat_msg & msg) : result(msg) {}
28
+
29
+ virtual void from_ast(const common_peg_ast_arena & arena, const common_peg_parse_result & result);
30
+ virtual void map(const common_peg_ast_node & node);
31
+ };
32
+
33
+ class common_chat_peg_native_builder : public common_chat_peg_builder {
34
+ public:
35
+ static constexpr const char * TOOL = "tool";
36
+ static constexpr const char * TOOL_OPEN = "tool-open";
37
+ static constexpr const char * TOOL_CLOSE = "tool-close";
38
+ static constexpr const char * TOOL_ID = "tool-id";
39
+ static constexpr const char * TOOL_NAME = "tool-name";
40
+ static constexpr const char * TOOL_ARGS = "tool-args";
41
+
42
+ common_peg_parser tool(const common_peg_parser & p) { return tag(TOOL, p); }
43
+ common_peg_parser tool_open(const common_peg_parser & p) { return atomic(tag(TOOL_OPEN, p)); }
44
+ common_peg_parser tool_close(const common_peg_parser & p) { return atomic(tag(TOOL_CLOSE, p)); }
45
+ common_peg_parser tool_id(const common_peg_parser & p) { return atomic(tag(TOOL_ID, p)); }
46
+ common_peg_parser tool_name(const common_peg_parser & p) { return atomic(tag(TOOL_NAME, p)); }
47
+ common_peg_parser tool_args(const common_peg_parser & p) { return tag(TOOL_ARGS, p); }
48
+ };
49
+
50
+ class common_chat_peg_native_mapper : public common_chat_peg_mapper {
51
+ common_chat_tool_call * current_tool;
52
+
53
+ public:
54
+ common_chat_peg_native_mapper(common_chat_msg & msg) : common_chat_peg_mapper(msg) {}
55
+
56
+ void map(const common_peg_ast_node & node) override;
57
+ };
58
+
59
+ inline common_peg_arena build_chat_peg_native_parser(const std::function<common_peg_parser(common_chat_peg_native_builder & builder)> & fn) {
60
+ common_chat_peg_native_builder builder;
61
+ builder.set_root(fn(builder));
62
+ return builder.build();
63
+ }
64
+
65
+ class common_chat_peg_constructed_builder : public common_chat_peg_builder {
66
+ public:
67
+ static constexpr const char * TOOL = "tool";
68
+ static constexpr const char * TOOL_OPEN = "tool-open";
69
+ static constexpr const char * TOOL_CLOSE = "tool-close";
70
+ static constexpr const char * TOOL_NAME = "tool-name";
71
+ static constexpr const char * TOOL_ARG = "tool-arg";
72
+ static constexpr const char * TOOL_ARG_OPEN = "tool-arg-open";
73
+ static constexpr const char * TOOL_ARG_CLOSE = "tool-arg-close";
74
+ static constexpr const char * TOOL_ARG_NAME = "tool-arg-name";
75
+ static constexpr const char * TOOL_ARG_STRING_VALUE = "tool-arg-string-value";
76
+ static constexpr const char * TOOL_ARG_JSON_VALUE = "tool-arg-json-value";
77
+
78
+ common_peg_parser tool(const common_peg_parser & p) { return tag(TOOL, p); }
79
+ common_peg_parser tool_open(const common_peg_parser & p) { return atomic(tag(TOOL_OPEN, p)); }
80
+ common_peg_parser tool_close(const common_peg_parser & p) { return atomic(tag(TOOL_CLOSE, p)); }
81
+ common_peg_parser tool_name(const common_peg_parser & p) { return atomic(tag(TOOL_NAME, p)); }
82
+ common_peg_parser tool_arg(const common_peg_parser & p) { return tag(TOOL_ARG, p); }
83
+ common_peg_parser tool_arg_open(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_OPEN, p)); }
84
+ common_peg_parser tool_arg_close(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_CLOSE, p)); }
85
+ common_peg_parser tool_arg_name(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_NAME, p)); }
86
+ common_peg_parser tool_arg_string_value(const common_peg_parser & p) { return tag(TOOL_ARG_STRING_VALUE, p); }
87
+ common_peg_parser tool_arg_json_value(const common_peg_parser & p) { return tag(TOOL_ARG_JSON_VALUE, p); }
88
+ };
89
+
90
+ class common_chat_peg_constructed_mapper : public common_chat_peg_mapper {
91
+ common_chat_tool_call * current_tool;
92
+ int arg_count = 0;
93
+ bool needs_closing_quote = false;
94
+
95
+ public:
96
+ common_chat_peg_constructed_mapper(common_chat_msg & msg) : common_chat_peg_mapper(msg) {}
97
+
98
+ void map(const common_peg_ast_node & node) override;
99
+ };
100
+
101
+ inline common_peg_arena build_chat_peg_constructed_parser(const std::function<common_peg_parser(common_chat_peg_constructed_builder & builder)> & fn) {
102
+ common_chat_peg_constructed_builder builder;
103
+ builder.set_root(fn(builder));
104
+ return builder.build();
105
+ }