whispercpp 1.3.2 → 1.3.4

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 (664) hide show
  1. checksums.yaml +4 -4
  2. data/.gitignore +6 -3
  3. data/README.md +71 -14
  4. data/Rakefile +20 -7
  5. data/ext/.gitignore +4 -6
  6. data/ext/dependencies.rb +36 -24
  7. data/ext/extconf.rb +1 -1
  8. data/ext/options.rb +48 -184
  9. data/ext/ruby_whisper.c +18 -0
  10. data/ext/ruby_whisper_context.c +43 -12
  11. data/ext/ruby_whisper_model.c +1 -1
  12. data/ext/ruby_whisper_params.c +59 -27
  13. data/ext/ruby_whisper_segment.c +81 -4
  14. data/ext/ruby_whisper_transcribe.cpp +13 -7
  15. data/ext/ruby_whisper_vad_params.c +1 -1
  16. data/ext/sources/CMakeLists.txt +5 -1
  17. data/ext/sources/bindings/javascript/package.json +1 -1
  18. data/ext/sources/build-xcframework.sh +24 -0
  19. data/ext/sources/examples/CMakeLists.txt +1 -0
  20. data/ext/sources/examples/addon.node/__test__/whisper.spec.js +120 -24
  21. data/ext/sources/examples/addon.node/addon.cpp +154 -35
  22. data/ext/sources/examples/addon.node/index.js +10 -5
  23. data/ext/sources/examples/addon.node/vad-example.js +132 -0
  24. data/ext/sources/examples/bench/bench.cpp +29 -18
  25. data/ext/sources/examples/bench.wasm/index-tmpl.html +10 -9
  26. data/ext/sources/examples/cli/cli.cpp +7 -4
  27. data/ext/sources/examples/command/command.cpp +58 -32
  28. data/ext/sources/examples/command.wasm/index-tmpl.html +5 -4
  29. data/ext/sources/examples/common-ggml.cpp +2 -0
  30. data/ext/sources/examples/common-whisper.cpp +14 -7
  31. data/ext/sources/examples/lsp/lsp.cpp +21 -17
  32. data/ext/sources/examples/quantize/quantize.cpp +3 -0
  33. data/ext/sources/examples/server/CMakeLists.txt +3 -0
  34. data/ext/sources/examples/server/server.cpp +193 -35
  35. data/ext/sources/examples/server.py +6 -1
  36. data/ext/sources/examples/stream/stream.cpp +10 -2
  37. data/ext/sources/examples/stream.wasm/emscripten.cpp +6 -6
  38. data/ext/sources/examples/stream.wasm/index-tmpl.html +82 -5
  39. data/ext/sources/examples/talk-llama/CMakeLists.txt +3 -0
  40. data/ext/sources/examples/talk-llama/llama-adapter.cpp +101 -4
  41. data/ext/sources/examples/talk-llama/llama-adapter.h +6 -0
  42. data/ext/sources/examples/talk-llama/llama-arch.cpp +756 -15
  43. data/ext/sources/examples/talk-llama/llama-arch.h +85 -1
  44. data/ext/sources/examples/talk-llama/llama-batch.cpp +773 -272
  45. data/ext/sources/examples/talk-llama/llama-batch.h +126 -55
  46. data/ext/sources/examples/talk-llama/llama-chat.cpp +150 -13
  47. data/ext/sources/examples/talk-llama/llama-chat.h +8 -0
  48. data/ext/sources/examples/talk-llama/llama-context.cpp +814 -542
  49. data/ext/sources/examples/talk-llama/llama-context.h +68 -32
  50. data/ext/sources/examples/talk-llama/llama-cparams.cpp +1 -1
  51. data/ext/sources/examples/talk-llama/llama-cparams.h +4 -4
  52. data/ext/sources/examples/talk-llama/llama-graph.cpp +787 -440
  53. data/ext/sources/examples/talk-llama/llama-graph.h +333 -153
  54. data/ext/sources/examples/talk-llama/llama-hparams.cpp +128 -6
  55. data/ext/sources/examples/talk-llama/llama-hparams.h +80 -17
  56. data/ext/sources/examples/talk-llama/llama-impl.h +2 -0
  57. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +326 -0
  58. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.h +137 -0
  59. data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +1248 -1967
  60. data/ext/sources/examples/talk-llama/llama-kv-cache.h +218 -345
  61. data/ext/sources/examples/talk-llama/llama-kv-cells.h +164 -52
  62. data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +266 -0
  63. data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +139 -0
  64. data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +1154 -0
  65. data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +182 -0
  66. data/ext/sources/examples/talk-llama/llama-memory.cpp +58 -0
  67. data/ext/sources/examples/talk-llama/llama-memory.h +94 -4
  68. data/ext/sources/examples/talk-llama/llama-mmap.cpp +1 -1
  69. data/ext/sources/examples/talk-llama/llama-model-loader.cpp +44 -17
  70. data/ext/sources/examples/talk-llama/llama-model-loader.h +3 -2
  71. data/ext/sources/examples/talk-llama/llama-model-saver.cpp +1 -0
  72. data/ext/sources/examples/talk-llama/llama-model.cpp +11377 -5248
  73. data/ext/sources/examples/talk-llama/llama-model.h +87 -9
  74. data/ext/sources/examples/talk-llama/llama-quant.cpp +137 -16
  75. data/ext/sources/examples/talk-llama/llama-sampling.cpp +226 -126
  76. data/ext/sources/examples/talk-llama/llama-vocab.cpp +502 -38
  77. data/ext/sources/examples/talk-llama/llama-vocab.h +46 -0
  78. data/ext/sources/examples/talk-llama/llama.cpp +76 -17
  79. data/ext/sources/examples/talk-llama/llama.h +176 -151
  80. data/ext/sources/examples/talk-llama/talk-llama.cpp +11 -6
  81. data/ext/sources/examples/talk-llama/unicode.cpp +212 -0
  82. data/ext/sources/examples/talk-llama/unicode.h +45 -0
  83. data/ext/sources/examples/vad-speech-segments/speech.cpp +6 -0
  84. data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +6 -2
  85. data/ext/sources/examples/whisper.wasm/index-tmpl.html +17 -16
  86. data/ext/sources/ggml/CMakeLists.txt +106 -33
  87. data/ext/sources/ggml/cmake/common.cmake +24 -0
  88. data/ext/sources/ggml/cmake/ggml-config.cmake.in +132 -93
  89. data/ext/sources/ggml/include/ggml-backend.h +18 -2
  90. data/ext/sources/ggml/include/ggml-cpu.h +2 -0
  91. data/ext/sources/ggml/include/ggml-metal.h +1 -6
  92. data/ext/sources/ggml/include/ggml-opt.h +25 -6
  93. data/ext/sources/ggml/include/ggml-webgpu.h +19 -0
  94. data/ext/sources/ggml/include/ggml-zdnn.h +17 -0
  95. data/ext/sources/ggml/include/ggml.h +365 -21
  96. data/ext/sources/ggml/src/CMakeLists.txt +98 -25
  97. data/ext/sources/ggml/src/ggml-alloc.c +265 -141
  98. data/ext/sources/ggml/src/ggml-backend-impl.h +4 -1
  99. data/ext/sources/ggml/src/ggml-backend-reg.cpp +35 -13
  100. data/ext/sources/ggml/src/ggml-backend.cpp +266 -60
  101. data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +4 -4
  102. data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +5 -4
  103. data/ext/sources/ggml/src/ggml-cann/CMakeLists.txt +15 -0
  104. data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +3 -1
  105. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +903 -717
  106. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +143 -25
  107. data/ext/sources/ggml/src/ggml-cann/common.h +149 -2
  108. data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +521 -78
  109. data/ext/sources/ggml/src/ggml-common.h +21 -0
  110. data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +165 -50
  111. data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +5 -3
  112. data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +11 -10
  113. data/ext/sources/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp +94 -0
  114. data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +3650 -0
  115. data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +1891 -0
  116. data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +2160 -0
  117. data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp +82 -0
  118. data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +2305 -0
  119. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +1897 -0
  120. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +342 -0
  121. data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +1468 -0
  122. data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +1221 -0
  123. data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +3820 -0
  124. data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +6307 -0
  125. data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +214 -0
  126. data/ext/sources/ggml/src/ggml-cpu/common.h +18 -3
  127. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +23 -7
  128. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +179 -110
  129. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +44 -33
  130. data/ext/sources/ggml/src/ggml-cpu/{ggml-cpu-hbm.cpp → hbm.cpp} +1 -1
  131. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +152 -18
  132. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.h +7 -1
  133. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +228 -98
  134. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +532 -1124
  135. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.h +5 -0
  136. data/ext/sources/ggml/src/ggml-cpu/ops.cpp +3374 -2081
  137. data/ext/sources/ggml/src/ggml-cpu/ops.h +13 -8
  138. data/ext/sources/ggml/src/ggml-cpu/quants.c +1193 -0
  139. data/ext/sources/ggml/src/ggml-cpu/{ggml-cpu-quants.h → quants.h} +34 -0
  140. data/ext/sources/ggml/src/ggml-cpu/repack.cpp +1982 -0
  141. data/ext/sources/ggml/src/ggml-cpu/repack.h +120 -0
  142. data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +367 -46
  143. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +1024 -0
  144. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.h +13 -0
  145. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +3196 -0
  146. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_kernels.h +26 -0
  147. data/ext/sources/ggml/src/ggml-cpu/{ggml-cpu-traits.cpp → traits.cpp} +3 -3
  148. data/ext/sources/ggml/src/ggml-cpu/{ggml-cpu-traits.h → traits.h} +1 -1
  149. data/ext/sources/ggml/src/ggml-cpu/vec.cpp +272 -35
  150. data/ext/sources/ggml/src/ggml-cpu/vec.h +794 -142
  151. data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +20 -16
  152. data/ext/sources/ggml/src/ggml-cuda/add-id.cu +58 -0
  153. data/ext/sources/ggml/src/ggml-cuda/add-id.cuh +3 -0
  154. data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +330 -191
  155. data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +2 -0
  156. data/ext/sources/ggml/src/ggml-cuda/common.cuh +291 -81
  157. data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cu +1 -4
  158. data/ext/sources/ggml/src/ggml-cuda/conv2d-dw.cu +161 -0
  159. data/ext/sources/ggml/src/ggml-cuda/conv2d-dw.cuh +5 -0
  160. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cu +91 -0
  161. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cuh +4 -0
  162. data/ext/sources/ggml/src/ggml-cuda/conv2d.cu +166 -0
  163. data/ext/sources/ggml/src/ggml-cuda/conv2d.cuh +5 -0
  164. data/ext/sources/ggml/src/ggml-cuda/convert.cu +117 -22
  165. data/ext/sources/ggml/src/ggml-cuda/convert.cuh +20 -0
  166. data/ext/sources/ggml/src/ggml-cuda/cpy-utils.cuh +217 -0
  167. data/ext/sources/ggml/src/ggml-cuda/cpy.cu +64 -307
  168. data/ext/sources/ggml/src/ggml-cuda/cross-entropy-loss.cu +2 -14
  169. data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +14 -40
  170. data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +499 -368
  171. data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +142 -93
  172. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +755 -0
  173. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +3 -0
  174. data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +593 -0
  175. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +90 -50
  176. data/ext/sources/ggml/src/ggml-cuda/fattn.cu +185 -198
  177. data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +2 -0
  178. data/ext/sources/ggml/src/ggml-cuda/getrows.cu +50 -39
  179. data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +636 -222
  180. data/ext/sources/ggml/src/ggml-cuda/im2col.cu +196 -35
  181. data/ext/sources/ggml/src/ggml-cuda/im2col.cuh +1 -0
  182. data/ext/sources/ggml/src/ggml-cuda/mean.cu +73 -0
  183. data/ext/sources/ggml/src/ggml-cuda/mean.cuh +3 -0
  184. data/ext/sources/ggml/src/ggml-cuda/mma.cuh +198 -45
  185. data/ext/sources/ggml/src/ggml-cuda/mmf.cu +123 -0
  186. data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +496 -0
  187. data/ext/sources/ggml/src/ggml-cuda/mmq.cu +206 -57
  188. data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +1262 -721
  189. data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +506 -0
  190. data/ext/sources/ggml/src/ggml-cuda/{mmv.cuh → mmvf.cuh} +4 -5
  191. data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +64 -73
  192. data/ext/sources/ggml/src/ggml-cuda/norm.cu +284 -12
  193. data/ext/sources/ggml/src/ggml-cuda/norm.cuh +7 -0
  194. data/ext/sources/ggml/src/ggml-cuda/opt-step-sgd.cu +49 -0
  195. data/ext/sources/ggml/src/ggml-cuda/opt-step-sgd.cuh +5 -0
  196. data/ext/sources/ggml/src/ggml-cuda/pad.cu +46 -23
  197. data/ext/sources/ggml/src/ggml-cuda/pad_reflect_1d.cu +91 -0
  198. data/ext/sources/ggml/src/ggml-cuda/pad_reflect_1d.cuh +5 -0
  199. data/ext/sources/ggml/src/ggml-cuda/quantize.cu +12 -10
  200. data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +53 -0
  201. data/ext/sources/ggml/src/ggml-cuda/roll.cu +67 -0
  202. data/ext/sources/ggml/src/ggml-cuda/roll.cuh +5 -0
  203. data/ext/sources/ggml/src/ggml-cuda/rope.cu +21 -27
  204. data/ext/sources/ggml/src/ggml-cuda/scale.cu +14 -11
  205. data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +276 -0
  206. data/ext/sources/ggml/src/ggml-cuda/set-rows.cuh +7 -0
  207. data/ext/sources/ggml/src/ggml-cuda/softcap.cu +34 -0
  208. data/ext/sources/ggml/src/ggml-cuda/softcap.cuh +5 -0
  209. data/ext/sources/ggml/src/ggml-cuda/softmax.cu +126 -59
  210. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +10 -2
  211. data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +322 -98
  212. data/ext/sources/ggml/src/ggml-cuda/sum.cu +6 -10
  213. data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +23 -19
  214. data/ext/sources/ggml/src/ggml-cuda/sumrows.cuh +0 -1
  215. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-f16.cu +7 -0
  216. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_0.cu +7 -0
  217. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_1.cu +7 -0
  218. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_0.cu +7 -0
  219. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_1.cu +7 -0
  220. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q8_0.cu +7 -0
  221. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-f16.cu +7 -0
  222. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_0.cu +7 -0
  223. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_1.cu +7 -0
  224. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_0.cu +7 -0
  225. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_1.cu +7 -0
  226. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q8_0.cu +7 -0
  227. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-f16.cu +7 -0
  228. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_0.cu +7 -0
  229. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_1.cu +7 -0
  230. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_0.cu +7 -0
  231. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_1.cu +7 -0
  232. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q8_0.cu +7 -0
  233. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-f16.cu +7 -0
  234. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_0.cu +7 -0
  235. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_1.cu +7 -0
  236. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_0.cu +7 -0
  237. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_1.cu +7 -0
  238. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q8_0.cu +7 -0
  239. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-f16.cu +7 -0
  240. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_0.cu +7 -0
  241. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_1.cu +7 -0
  242. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_0.cu +7 -0
  243. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_1.cu +7 -0
  244. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q8_0.cu +7 -0
  245. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-f16.cu +7 -0
  246. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_0.cu +7 -0
  247. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_1.cu +7 -0
  248. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_0.cu +7 -0
  249. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_1.cu +7 -0
  250. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q8_0.cu +7 -0
  251. data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +21 -18
  252. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_1.cu +5 -0
  253. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_10.cu +5 -0
  254. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_11.cu +5 -0
  255. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_12.cu +5 -0
  256. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_13.cu +5 -0
  257. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_14.cu +5 -0
  258. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_15.cu +5 -0
  259. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_16.cu +5 -0
  260. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_2.cu +5 -0
  261. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_3.cu +5 -0
  262. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_4.cu +5 -0
  263. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_5.cu +5 -0
  264. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_6.cu +5 -0
  265. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_7.cu +5 -0
  266. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_8.cu +5 -0
  267. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_9.cu +5 -0
  268. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu +5 -0
  269. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +259 -0
  270. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cuh +14 -0
  271. data/ext/sources/ggml/src/ggml-cuda/tsembd.cu +3 -3
  272. data/ext/sources/ggml/src/ggml-cuda/unary.cu +179 -0
  273. data/ext/sources/ggml/src/ggml-cuda/unary.cuh +15 -0
  274. data/ext/sources/ggml/src/ggml-cuda/upscale.cu +92 -6
  275. data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +110 -22
  276. data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +4 -0
  277. data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +58 -36
  278. data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +4 -3
  279. data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +14 -2
  280. data/ext/sources/ggml/src/ggml-impl.h +229 -175
  281. data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +21 -17
  282. data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp +446 -0
  283. data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.h +52 -0
  284. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.h +33 -0
  285. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.m +600 -0
  286. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +1376 -0
  287. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +226 -0
  288. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +1308 -0
  289. data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +163 -63
  290. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +3158 -0
  291. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +82 -0
  292. data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +718 -0
  293. data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +3208 -1575
  294. data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +18 -8
  295. data/ext/sources/ggml/src/ggml-musa/mudnn.cuh +2 -2
  296. data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +32 -0
  297. data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +4430 -792
  298. data/ext/sources/ggml/src/ggml-opencl/kernels/add.cl +107 -0
  299. data/ext/sources/ggml/src/ggml-opencl/kernels/add_id.cl +42 -0
  300. data/ext/sources/ggml/src/ggml-opencl/kernels/argsort.cl +86 -0
  301. data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +109 -0
  302. data/ext/sources/ggml/src/ggml-opencl/kernels/conv2d.cl +185 -0
  303. data/ext/sources/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl +176 -0
  304. data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +84 -0
  305. data/ext/sources/ggml/src/ggml-opencl/kernels/div.cl +138 -0
  306. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl +370 -0
  307. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +370 -0
  308. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl +373 -0
  309. data/ext/sources/ggml/src/ggml-opencl/kernels/gelu.cl +27 -0
  310. data/ext/sources/ggml/src/ggml-opencl/kernels/glu.cl +378 -0
  311. data/ext/sources/ggml/src/ggml-opencl/kernels/group_norm.cl +121 -0
  312. data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f16.cl +1 -1
  313. data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f32.cl +1 -1
  314. data/ext/sources/ggml/src/ggml-opencl/kernels/mul.cl +73 -0
  315. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl +130 -0
  316. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl +132 -0
  317. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl +133 -0
  318. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl +189 -0
  319. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl +176 -0
  320. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl +283 -0
  321. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl +140 -0
  322. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl +222 -0
  323. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl +144 -0
  324. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl +167 -0
  325. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl +125 -0
  326. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl +202 -0
  327. data/ext/sources/ggml/src/ggml-opencl/kernels/norm.cl +80 -0
  328. data/ext/sources/ggml/src/ggml-opencl/kernels/pad.cl +30 -0
  329. data/ext/sources/ggml/src/ggml-opencl/kernels/repeat.cl +39 -0
  330. data/ext/sources/ggml/src/ggml-opencl/kernels/rms_norm.cl +79 -0
  331. data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +3 -2
  332. data/ext/sources/ggml/src/ggml-opencl/kernels/set_rows.cl +189 -0
  333. data/ext/sources/ggml/src/ggml-opencl/kernels/sigmoid.cl +29 -0
  334. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +34 -13
  335. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +34 -13
  336. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f16.cl +34 -13
  337. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f32.cl +34 -13
  338. data/ext/sources/ggml/src/ggml-opencl/kernels/sub.cl +138 -0
  339. data/ext/sources/ggml/src/ggml-opencl/kernels/sum_rows.cl +39 -0
  340. data/ext/sources/ggml/src/ggml-opencl/kernels/tanh.cl +63 -0
  341. data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +20 -0
  342. data/ext/sources/ggml/src/ggml-opencl/kernels/tsembd.cl +48 -0
  343. data/ext/sources/ggml/src/ggml-opencl/kernels/upscale.cl +120 -0
  344. data/ext/sources/ggml/src/ggml-opt.cpp +97 -41
  345. data/ext/sources/ggml/src/ggml-quants.c +117 -24
  346. data/ext/sources/ggml/src/ggml-quants.h +6 -0
  347. data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +85 -62
  348. data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +3 -3
  349. data/ext/sources/ggml/src/ggml-sycl/backend.hpp +2 -0
  350. data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +9 -0
  351. data/ext/sources/ggml/src/ggml-sycl/binbcast.hpp +6 -0
  352. data/ext/sources/ggml/src/ggml-sycl/common.hpp +20 -48
  353. data/ext/sources/ggml/src/ggml-sycl/concat.cpp +13 -17
  354. data/ext/sources/ggml/src/ggml-sycl/convert.cpp +21 -2
  355. data/ext/sources/ggml/src/ggml-sycl/cpy.cpp +116 -211
  356. data/ext/sources/ggml/src/ggml-sycl/cpy.hpp +213 -1
  357. data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +32 -0
  358. data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +700 -1041
  359. data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +20 -9
  360. data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +17 -26
  361. data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +2 -96
  362. data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +393 -250
  363. data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +1 -1
  364. data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +32 -8
  365. data/ext/sources/ggml/src/ggml-sycl/quantize.hpp +133 -0
  366. data/ext/sources/ggml/src/ggml-sycl/quants.hpp +38 -11
  367. data/ext/sources/ggml/src/ggml-sycl/rope.cpp +125 -21
  368. data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +234 -0
  369. data/ext/sources/ggml/src/ggml-sycl/set_rows.hpp +8 -0
  370. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +3 -1
  371. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +3 -0
  372. data/ext/sources/ggml/src/ggml-sycl/tsembd.cpp +4 -3
  373. data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +105 -17
  374. data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +36 -32
  375. data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +4198 -1145
  376. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +4 -12
  377. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add.comp +41 -1
  378. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp +42 -0
  379. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp +13 -4
  380. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp +39 -29
  381. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +349 -0
  382. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp +98 -0
  383. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +2 -2
  384. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +66 -12
  385. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +154 -0
  386. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp +21 -0
  387. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp +1 -1
  388. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp +2 -1
  389. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp +6 -5
  390. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp +4 -2
  391. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp +32 -0
  392. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp +1 -1
  393. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp +1 -1
  394. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp +1 -1
  395. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp +1 -1
  396. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp +1 -1
  397. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +21 -0
  398. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +69 -24
  399. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp +60 -20
  400. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +98 -42
  401. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +64 -27
  402. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +74 -13
  403. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/geglu.comp +13 -0
  404. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/geglu_erf.comp +27 -0
  405. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/geglu_quick.comp +11 -0
  406. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_erf.comp +39 -0
  407. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp +4 -17
  408. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +19 -10
  409. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +25 -15
  410. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp +19 -0
  411. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.comp +29 -0
  412. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardsigmoid.comp +22 -0
  413. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardswish.comp +22 -0
  414. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +18 -14
  415. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +126 -0
  416. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp +65 -1
  417. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +11 -7
  418. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +140 -0
  419. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +144 -531
  420. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +206 -38
  421. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.comp +556 -0
  422. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +12 -5
  423. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp +15 -9
  424. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +111 -0
  425. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp +22 -0
  426. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/pad.comp +24 -3
  427. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp +53 -3
  428. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/reglu.comp +9 -0
  429. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +64 -11
  430. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_partials.comp +65 -0
  431. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/roll.comp +46 -0
  432. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp +1 -4
  433. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +7 -9
  434. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +7 -9
  435. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +7 -9
  436. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.comp +5 -0
  437. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/scale.comp +1 -1
  438. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +29 -7
  439. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp +4 -0
  440. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp +17 -0
  441. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp +38 -5
  442. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/swiglu.comp +9 -0
  443. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp +14 -0
  444. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp +4 -3
  445. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +101 -9
  446. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +69 -5
  447. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp +25 -0
  448. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +338 -71
  449. data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +54 -0
  450. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +1558 -0
  451. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add.tmpl.wgsl +44 -0
  452. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_in_place.tmpl.wgsl +41 -0
  453. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary_head.tmpl +45 -0
  454. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +930 -0
  455. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl +60 -0
  456. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +124 -0
  457. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.tmpl.wgsl +874 -0
  458. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl +40 -0
  459. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul.tmpl.wgsl +44 -0
  460. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_in_place.tmpl.wgsl +41 -0
  461. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl +907 -0
  462. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +57 -0
  463. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_in_place.wgsl +48 -0
  464. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +81 -0
  465. data/ext/sources/ggml/src/ggml-zdnn/CMakeLists.txt +36 -0
  466. data/ext/sources/ggml/src/ggml-zdnn/common.hpp +59 -0
  467. data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +628 -0
  468. data/ext/sources/ggml/src/ggml-zdnn/mmf.cpp +80 -0
  469. data/ext/sources/ggml/src/ggml-zdnn/mmf.hpp +12 -0
  470. data/ext/sources/ggml/src/ggml-zdnn/utils.cpp +79 -0
  471. data/ext/sources/ggml/src/ggml-zdnn/utils.hpp +19 -0
  472. data/ext/sources/ggml/src/ggml.c +802 -142
  473. data/ext/sources/ggml/src/ggml.cpp +26 -0
  474. data/ext/sources/ggml/src/gguf.cpp +32 -4
  475. data/ext/sources/include/whisper.h +2 -0
  476. data/ext/sources/src/CMakeLists.txt +2 -0
  477. data/ext/sources/src/coreml/whisper-compat.h +10 -0
  478. data/ext/sources/src/coreml/whisper-compat.m +35 -0
  479. data/ext/sources/src/coreml/whisper-decoder-impl.m +1 -0
  480. data/ext/sources/src/coreml/whisper-encoder-impl.m +1 -0
  481. data/ext/sources/src/whisper.cpp +241 -215
  482. data/ext/sources/tests/CMakeLists.txt +8 -1
  483. data/ext/sources/tests/test-vad-full.cpp +3 -3
  484. data/ext/sources/tests/test-vad.cpp +2 -2
  485. data/extsources.rb +15 -9
  486. data/lib/whisper/context.rb +15 -0
  487. data/lib/whisper/model/uri.rb +57 -2
  488. data/lib/whisper/segment.rb +58 -0
  489. data/sig/whisper.rbs +75 -38
  490. data/{tests → test}/helper.rb +1 -12
  491. data/{tests → test}/test_model.rb +9 -0
  492. data/test/test_package.rb +51 -0
  493. data/{tests → test}/test_params.rb +8 -0
  494. data/test/test_segment.rb +146 -0
  495. data/{tests → test}/test_whisper.rb +70 -0
  496. data/whispercpp.gemspec +2 -3
  497. metadata +246 -191
  498. data/ext/sources/.dockerignore +0 -3
  499. data/ext/sources/.github/workflows/bindings-ruby.yml +0 -21
  500. data/ext/sources/ci/run.sh +0 -336
  501. data/ext/sources/close-issue.yml +0 -28
  502. data/ext/sources/ggml/include/ggml-kompute.h +0 -50
  503. data/ext/sources/ggml/src/ggml-amx/CMakeLists.txt +0 -107
  504. data/ext/sources/ggml/src/ggml-amx/common.h +0 -94
  505. data/ext/sources/ggml/src/ggml-amx/ggml-amx.cpp +0 -446
  506. data/ext/sources/ggml/src/ggml-amx/mmq.cpp +0 -2510
  507. data/ext/sources/ggml/src/ggml-amx/mmq.h +0 -17
  508. data/ext/sources/ggml/src/ggml-cann/kernels/CMakeLists.txt +0 -30
  509. data/ext/sources/ggml/src/ggml-cann/kernels/ascendc_kernels.h +0 -19
  510. data/ext/sources/ggml/src/ggml-cann/kernels/dup.cpp +0 -234
  511. data/ext/sources/ggml/src/ggml-cann/kernels/get_row_f16.cpp +0 -197
  512. data/ext/sources/ggml/src/ggml-cann/kernels/get_row_f32.cpp +0 -190
  513. data/ext/sources/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp +0 -204
  514. data/ext/sources/ggml/src/ggml-cann/kernels/get_row_q8_0.cpp +0 -191
  515. data/ext/sources/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp +0 -218
  516. data/ext/sources/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp +0 -216
  517. data/ext/sources/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp +0 -295
  518. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +0 -6431
  519. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-aarch64.h +0 -8
  520. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-quants.c +0 -13747
  521. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cu +0 -357
  522. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cuh +0 -3
  523. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cu +0 -365
  524. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cuh +0 -3
  525. data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f16.cuh +0 -482
  526. data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f32.cuh +0 -472
  527. data/ext/sources/ggml/src/ggml-cuda/mmv.cu +0 -336
  528. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu +0 -5
  529. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu +0 -5
  530. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu +0 -5
  531. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu +0 -5
  532. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu +0 -5
  533. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu +0 -5
  534. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu +0 -5
  535. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu +0 -5
  536. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu +0 -5
  537. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu +0 -5
  538. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu +0 -5
  539. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu +0 -5
  540. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu +0 -5
  541. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu +0 -5
  542. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu +0 -5
  543. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu +0 -5
  544. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu +0 -5
  545. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu +0 -5
  546. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu +0 -5
  547. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu +0 -5
  548. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu +0 -5
  549. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu +0 -5
  550. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu +0 -5
  551. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu +0 -5
  552. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu +0 -5
  553. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu +0 -5
  554. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu +0 -5
  555. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu +0 -5
  556. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu +0 -5
  557. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu +0 -5
  558. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu +0 -5
  559. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu +0 -5
  560. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu +0 -5
  561. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu +0 -5
  562. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu +0 -5
  563. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu +0 -5
  564. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu +0 -5
  565. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu +0 -5
  566. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu +0 -5
  567. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu +0 -5
  568. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu +0 -5
  569. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu +0 -5
  570. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu +0 -5
  571. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu +0 -5
  572. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu +0 -5
  573. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu +0 -5
  574. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu +0 -5
  575. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu +0 -5
  576. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu +0 -5
  577. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu +0 -5
  578. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu +0 -5
  579. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu +0 -5
  580. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu +0 -5
  581. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu +0 -5
  582. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu +0 -5
  583. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu +0 -5
  584. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu +0 -5
  585. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu +0 -5
  586. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu +0 -5
  587. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu +0 -5
  588. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu +0 -5
  589. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu +0 -5
  590. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu +0 -5
  591. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu +0 -5
  592. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu +0 -5
  593. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu +0 -5
  594. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu +0 -5
  595. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu +0 -5
  596. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu +0 -5
  597. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu +0 -5
  598. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu +0 -5
  599. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu +0 -5
  600. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu +0 -5
  601. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu +0 -5
  602. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu +0 -5
  603. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu +0 -5
  604. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu +0 -5
  605. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu +0 -5
  606. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu +0 -5
  607. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu +0 -5
  608. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu +0 -5
  609. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu +0 -5
  610. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu +0 -5
  611. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu +0 -5
  612. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu +0 -5
  613. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu +0 -5
  614. data/ext/sources/ggml/src/ggml-kompute/CMakeLists.txt +0 -166
  615. data/ext/sources/ggml/src/ggml-kompute/ggml-kompute.cpp +0 -2251
  616. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/common.comp +0 -112
  617. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_add.comp +0 -58
  618. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp +0 -25
  619. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp +0 -52
  620. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp +0 -52
  621. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp +0 -52
  622. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp +0 -52
  623. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp +0 -30
  624. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp +0 -22
  625. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp +0 -17
  626. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp +0 -31
  627. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp +0 -31
  628. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp +0 -38
  629. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp +0 -39
  630. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp +0 -44
  631. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp +0 -52
  632. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp +0 -69
  633. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp +0 -51
  634. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp +0 -33
  635. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp +0 -35
  636. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp +0 -140
  637. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp +0 -106
  638. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp +0 -73
  639. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp +0 -52
  640. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp +0 -28
  641. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp +0 -84
  642. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp +0 -21
  643. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp +0 -53
  644. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp +0 -52
  645. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp +0 -52
  646. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp +0 -52
  647. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp +0 -52
  648. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp +0 -19
  649. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp +0 -23
  650. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp +0 -22
  651. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp +0 -72
  652. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp +0 -71
  653. data/ext/sources/ggml/src/ggml-metal/ggml-metal.m +0 -5998
  654. data/tests/test_package.rb +0 -46
  655. data/tests/test_segment.rb +0 -74
  656. /data/ext/sources/ggml/src/ggml-cpu/{cpu-feats-x86.cpp → arch/x86/cpu-feats.cpp} +0 -0
  657. /data/ext/sources/ggml/src/ggml-cpu/{ggml-cpu-hbm.h → hbm.h} +0 -0
  658. /data/{tests → test}/jfk_reader/.gitignore +0 -0
  659. /data/{tests → test}/jfk_reader/extconf.rb +0 -0
  660. /data/{tests → test}/jfk_reader/jfk_reader.c +0 -0
  661. /data/{tests → test}/test_callback.rb +0 -0
  662. /data/{tests → test}/test_error.rb +0 -0
  663. /data/{tests → test}/test_vad.rb +0 -0
  664. /data/{tests → test}/test_vad_params.rb +0 -0
data/ext/sources/examples/talk-llama/llama-arch.cpp CHANGED
@@ -20,7 +20,9 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
20
20
  { LLM_ARCH_BERT, "bert" },
21
21
  { LLM_ARCH_NOMIC_BERT, "nomic-bert" },
22
22
  { LLM_ARCH_NOMIC_BERT_MOE, "nomic-bert-moe" },
23
+ { LLM_ARCH_NEO_BERT, "neo-bert" },
23
24
  { LLM_ARCH_JINA_BERT_V2, "jina-bert-v2" },
25
+ { LLM_ARCH_JINA_BERT_V3, "jina-bert-v3" },
24
26
  { LLM_ARCH_BLOOM, "bloom" },
25
27
  { LLM_ARCH_STABLELM, "stablelm" },
26
28
  { LLM_ARCH_QWEN, "qwen" },
@@ -33,6 +35,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
33
35
  { LLM_ARCH_PHI3, "phi3" },
34
36
  { LLM_ARCH_PHIMOE, "phimoe" },
35
37
  { LLM_ARCH_PLAMO, "plamo" },
38
+ { LLM_ARCH_PLAMO2, "plamo2" },
36
39
  { LLM_ARCH_CODESHELL, "codeshell" },
37
40
  { LLM_ARCH_ORION, "orion" },
38
41
  { LLM_ARCH_INTERNLM2, "internlm2" },
@@ -41,8 +44,13 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
41
44
  { LLM_ARCH_GEMMA, "gemma" },
42
45
  { LLM_ARCH_GEMMA2, "gemma2" },
43
46
  { LLM_ARCH_GEMMA3, "gemma3" },
47
+ { LLM_ARCH_GEMMA3N, "gemma3n" },
48
+ { LLM_ARCH_GEMMA_EMBEDDING, "gemma-embedding" },
44
49
  { LLM_ARCH_STARCODER2, "starcoder2" },
45
50
  { LLM_ARCH_MAMBA, "mamba" },
51
+ { LLM_ARCH_MAMBA2, "mamba2" },
52
+ { LLM_ARCH_JAMBA, "jamba" },
53
+ { LLM_ARCH_FALCON_H1, "falcon-h1" },
46
54
  { LLM_ARCH_XVERSE, "xverse" },
47
55
  { LLM_ARCH_COMMAND_R, "command-r" },
48
56
  { LLM_ARCH_COHERE2, "cohere2" },
@@ -56,22 +64,41 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
56
64
  { LLM_ARCH_DEEPSEEK2, "deepseek2" },
57
65
  { LLM_ARCH_CHATGLM, "chatglm" },
58
66
  { LLM_ARCH_GLM4, "glm4" },
67
+ { LLM_ARCH_GLM4_MOE, "glm4moe" },
59
68
  { LLM_ARCH_BITNET, "bitnet" },
60
69
  { LLM_ARCH_T5, "t5" },
61
70
  { LLM_ARCH_T5ENCODER, "t5encoder" },
62
71
  { LLM_ARCH_JAIS, "jais" },
63
72
  { LLM_ARCH_NEMOTRON, "nemotron" },
73
+ { LLM_ARCH_NEMOTRON_H, "nemotron_h" },
64
74
  { LLM_ARCH_EXAONE, "exaone" },
75
+ { LLM_ARCH_EXAONE4, "exaone4" },
65
76
  { LLM_ARCH_RWKV6, "rwkv6" },
66
77
  { LLM_ARCH_RWKV6QWEN2, "rwkv6qwen2" },
67
78
  { LLM_ARCH_RWKV7, "rwkv7" },
68
79
  { LLM_ARCH_ARWKV7, "arwkv7" },
69
80
  { LLM_ARCH_GRANITE, "granite" },
70
81
  { LLM_ARCH_GRANITE_MOE, "granitemoe" },
82
+ { LLM_ARCH_GRANITE_HYBRID, "granitehybrid" },
71
83
  { LLM_ARCH_CHAMELEON, "chameleon" },
72
84
  { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
73
85
  { LLM_ARCH_PLM, "plm" },
74
86
  { LLM_ARCH_BAILINGMOE, "bailingmoe" },
87
+ { LLM_ARCH_DOTS1, "dots1" },
88
+ { LLM_ARCH_ARCEE, "arcee" },
89
+ { LLM_ARCH_ERNIE4_5, "ernie4_5" },
90
+ { LLM_ARCH_ERNIE4_5_MOE, "ernie4_5-moe" },
91
+ { LLM_ARCH_HUNYUAN_MOE, "hunyuan-moe" },
92
+ { LLM_ARCH_HUNYUAN_DENSE, "hunyuan-dense" },
93
+ { LLM_ARCH_SMOLLM3, "smollm3" },
94
+ { LLM_ARCH_OPENAI_MOE, "gpt-oss" },
95
+ { LLM_ARCH_LFM2, "lfm2" },
96
+ { LLM_ARCH_DREAM, "dream" },
97
+ { LLM_ARCH_SMALLTHINKER, "smallthinker" },
98
+ { LLM_ARCH_LLADA, "llada" },
99
+ { LLM_ARCH_LLADA_MOE, "llada-moe" },
100
+ { LLM_ARCH_SEED_OSS, "seed_oss" },
101
+ { LLM_ARCH_GROVEMOE, "grovemoe" },
75
102
  { LLM_ARCH_UNKNOWN, "(unknown)" },
76
103
  };
77
104
 
@@ -99,6 +126,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
99
126
  { LLM_KV_FEED_FORWARD_LENGTH, "%s.feed_forward_length" },
100
127
  { LLM_KV_EXPERT_FEED_FORWARD_LENGTH, "%s.expert_feed_forward_length" },
101
128
  { LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, "%s.expert_shared_feed_forward_length" },
129
+ { LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH, "%s.expert_chunk_feed_forward_length" },
102
130
  { LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" },
103
131
  { LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" },
104
132
  { LLM_KV_EXPERT_COUNT, "%s.expert_count" },
@@ -107,11 +135,16 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
107
135
  { LLM_KV_EXPERT_WEIGHTS_SCALE, "%s.expert_weights_scale" },
108
136
  { LLM_KV_EXPERT_WEIGHTS_NORM, "%s.expert_weights_norm" },
109
137
  { LLM_KV_EXPERT_GATING_FUNC, "%s.expert_gating_func" },
138
+ { LLM_KV_EXPERT_GROUP_SCALE, "%s.expert_group_scale" },
139
+ { LLM_KV_EXPERTS_PER_GROUP, "%s.experts_per_group" },
110
140
  { LLM_KV_MOE_EVERY_N_LAYERS, "%s.moe_every_n_layers" },
141
+ { LLM_KV_NEXTN_PREDICT_LAYERS, "%s.nextn_predict_layers" },
111
142
  { LLM_KV_POOLING_TYPE, "%s.pooling_type" },
112
143
  { LLM_KV_LOGIT_SCALE, "%s.logit_scale" },
113
144
  { LLM_KV_DECODER_START_TOKEN_ID, "%s.decoder_start_token_id" },
145
+ { LLM_KV_DECODER_BLOCK_COUNT, "%s.decoder_block_count" },
114
146
  { LLM_KV_ATTN_LOGIT_SOFTCAPPING, "%s.attn_logit_softcapping" },
147
+ { LLM_KV_ROUTER_LOGIT_SOFTCAPPING, "%s.router_logit_softcapping" },
115
148
  { LLM_KV_FINAL_LOGIT_SOFTCAPPING, "%s.final_logit_softcapping" },
116
149
  { LLM_KV_SWIN_NORM, "%s.swin_norm" },
117
150
  { LLM_KV_RESCALE_EVERY_N_LAYERS, "%s.rescale_every_n_layers" },
@@ -142,19 +175,25 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
142
175
  { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
143
176
  { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" },
144
177
  { LLM_KV_ATTENTION_SCALE, "%s.attention.scale" },
178
+ { LLM_KV_ATTENTION_OUTPUT_SCALE, "%s.attention.output_scale" },
179
+ { LLM_KV_ATTENTION_TEMPERATURE_LENGTH, "%s.attention.temperature_length" },
145
180
  { LLM_KV_ATTENTION_KEY_LENGTH_MLA, "%s.attention.key_length_mla" },
146
181
  { LLM_KV_ATTENTION_VALUE_LENGTH_MLA, "%s.attention.value_length_mla" },
147
182
 
148
- { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
149
- { LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" },
150
- { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" },
151
- { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" },
152
- { LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" },
153
- { LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" },
154
- { LLM_KV_ROPE_SCALING_ATTN_FACTOR, "%s.rope.scaling.attn_factor" },
155
- { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" },
156
- { LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" },
157
- { LLM_KV_ROPE_SCALING_YARN_LOG_MUL, "%s.rope.scaling.yarn_log_multiplier" },
183
+ { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
184
+ { LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" },
185
+ { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" },
186
+ { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" },
187
+ { LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" },
188
+ { LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" },
189
+ { LLM_KV_ROPE_SCALING_ATTN_FACTOR, "%s.rope.scaling.attn_factor" },
190
+ { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" },
191
+ { LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" },
192
+ { LLM_KV_ROPE_SCALING_YARN_LOG_MUL, "%s.rope.scaling.yarn_log_multiplier" },
193
+ { LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR, "%s.rope.scaling.yarn_ext_factor" },
194
+ { LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR, "%s.rope.scaling.yarn_attn_factor" },
195
+ { LLM_KV_ROPE_SCALING_YARN_BETA_FAST, "%s.rope.scaling.yarn_beta_fast" },
196
+ { LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, "%s.rope.scaling.yarn_beta_slow" },
158
197
 
159
198
  { LLM_KV_SPLIT_NO, "split.no" },
160
199
  { LLM_KV_SPLIT_COUNT, "split.count" },
@@ -164,6 +203,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
164
203
  { LLM_KV_SSM_INNER_SIZE, "%s.ssm.inner_size" },
165
204
  { LLM_KV_SSM_STATE_SIZE, "%s.ssm.state_size" },
166
205
  { LLM_KV_SSM_TIME_STEP_RANK, "%s.ssm.time_step_rank" },
206
+ { LLM_KV_SSM_GROUP_COUNT, "%s.ssm.group_count" },
167
207
  { LLM_KV_SSM_DT_B_C_RMS, "%s.ssm.dt_b_c_rms" },
168
208
 
169
209
  { LLM_KV_WKV_HEAD_SIZE, "%s.wkv.head_size" },
@@ -174,6 +214,10 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
174
214
  { LLM_KV_CONVNEXT_EMBEDDING_LENGTH, "%s.convnext.embedding_length" },
175
215
  { LLM_KV_CONVNEXT_BLOCK_COUNT, "%s.convnext.block_count" },
176
216
 
217
+ { LLM_KV_CLASSIFIER_OUTPUT_LABELS, "%s.classifier.output_labels" },
218
+
219
+ { LLM_KV_SHORTCONV_L_CACHE, "%s.shortconv.l_cache" },
220
+
177
221
  { LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" },
178
222
  { LLM_KV_TOKENIZER_PRE, "tokenizer.ggml.pre" },
179
223
  { LLM_KV_TOKENIZER_LIST, "tokenizer.ggml.tokens" },
@@ -192,13 +236,13 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
192
236
  { LLM_KV_TOKENIZER_MASK_ID, "tokenizer.ggml.mask_token_id" },
193
237
  { LLM_KV_TOKENIZER_ADD_BOS, "tokenizer.ggml.add_bos_token" },
194
238
  { LLM_KV_TOKENIZER_ADD_EOS, "tokenizer.ggml.add_eos_token" },
239
+ { LLM_KV_TOKENIZER_ADD_SEP, "tokenizer.ggml.add_sep_token" },
195
240
  { LLM_KV_TOKENIZER_ADD_PREFIX, "tokenizer.ggml.add_space_prefix" },
196
241
  { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS, "tokenizer.ggml.remove_extra_whitespaces" },
197
242
  { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap" },
198
243
  { LLM_KV_TOKENIZER_HF_JSON, "tokenizer.huggingface.json" },
199
244
  { LLM_KV_TOKENIZER_RWKV, "tokenizer.rwkv.world" },
200
245
  { LLM_KV_TOKENIZER_CHAT_TEMPLATE, "tokenizer.chat_template" },
201
- { LLM_KV_TOKENIZER_CHAT_TEMPLATE_N, "tokenizer.chat_template.%s" },
202
246
  { LLM_KV_TOKENIZER_FIM_PRE_ID, "tokenizer.ggml.fim_pre_token_id" },
203
247
  { LLM_KV_TOKENIZER_FIM_SUF_ID, "tokenizer.ggml.fim_suf_token_id" },
204
248
  { LLM_KV_TOKENIZER_FIM_MID_ID, "tokenizer.ggml.fim_mid_token_id" },
@@ -206,8 +250,11 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
206
250
  { LLM_KV_TOKENIZER_FIM_REP_ID, "tokenizer.ggml.fim_rep_token_id" },
207
251
  { LLM_KV_TOKENIZER_FIM_SEP_ID, "tokenizer.ggml.fim_sep_token_id" },
208
252
 
209
- { LLM_KV_ADAPTER_TYPE, "adapter.type" },
210
- { LLM_KV_ADAPTER_LORA_ALPHA, "adapter.lora.alpha" },
253
+ { LLM_KV_ADAPTER_TYPE, "adapter.type" },
254
+ { LLM_KV_ADAPTER_LORA_ALPHA, "adapter.lora.alpha" },
255
+ { LLM_KV_ADAPTER_LORA_TASK_NAME, "adapter.lora.task_name" },
256
+ { LLM_KV_ADAPTER_LORA_PROMPT_PREFIX, "adapter.lora.prompt_prefix" },
257
+ { LLM_KV_ADAPTER_ALORA_INVOCATION_TOKENS, "adapter.alora.invocation_tokens" },
211
258
 
212
259
  // deprecated
213
260
  { LLM_KV_TOKENIZER_PREFIX_ID, "tokenizer.ggml.prefix_token_id" },
@@ -242,6 +289,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
242
289
  { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
243
290
  },
244
291
  },
292
+ {
293
+ LLM_ARCH_ARCEE,
294
+ {
295
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
296
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
297
+ { LLM_TENSOR_OUTPUT, "output" },
298
+ { LLM_TENSOR_ROPE_FREQS, "rope_freqs" },
299
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
300
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
301
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
302
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
303
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
304
+ { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" },
305
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
306
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
307
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
308
+ },
309
+ },
245
310
  {
246
311
  LLM_ARCH_LLAMA4,
247
312
  {
@@ -345,12 +410,16 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
345
410
  { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" },
346
411
  { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
347
412
  { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
413
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
414
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
415
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
348
416
  { LLM_TENSOR_FFN_GATE_EXP, "blk.%d.ffn_gate.%d" },
349
417
  { LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" },
350
418
  { LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" },
351
419
  { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
352
420
  { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
353
421
  { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
422
+ { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" },
354
423
  { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" },
355
424
  { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" },
356
425
  },
@@ -448,6 +517,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
448
517
  { LLM_TENSOR_TOKEN_TYPES, "token_types" },
449
518
  { LLM_TENSOR_POS_EMBD, "position_embd" },
450
519
  { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" },
520
+ { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
451
521
  { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
452
522
  { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
453
523
  { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
@@ -492,6 +562,21 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
492
562
  { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
493
563
  },
494
564
  },
565
+ {
566
+ LLM_ARCH_NEO_BERT,
567
+ {
568
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
569
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
570
+ { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
571
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
572
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
573
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
574
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
575
+ { LLM_TENSOR_ENC_OUTPUT_NORM, "enc.output_norm" },
576
+ { LLM_TENSOR_CLS, "cls" },
577
+ { LLM_TENSOR_CLS_OUT, "cls.output" },
578
+ },
579
+ },
495
580
  {
496
581
  LLM_ARCH_JINA_BERT_V2,
497
582
  {
@@ -513,6 +598,20 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
513
598
  { LLM_TENSOR_CLS, "cls" },
514
599
  },
515
600
  },
601
+ {
602
+ LLM_ARCH_JINA_BERT_V3,
603
+ {
604
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
605
+ { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
606
+ { LLM_TENSOR_TOKEN_TYPES, "token_types" },
607
+ { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" },
608
+ { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
609
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
610
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
611
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
612
+ { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" },
613
+ },
614
+ },
516
615
  {
517
616
  LLM_ARCH_BLOOM,
518
617
  {
@@ -626,6 +725,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
626
725
  { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
627
726
  { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
628
727
  { LLM_TENSOR_OUTPUT, "output" },
728
+ { LLM_TENSOR_CLS_OUT, "cls.output" },
629
729
  { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
630
730
  { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
631
731
  { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
@@ -733,6 +833,36 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
733
833
  { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
734
834
  },
735
835
  },
836
+ {
837
+ LLM_ARCH_PLAMO2,
838
+ {
839
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
840
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
841
+ { LLM_TENSOR_OUTPUT, "output" },
842
+ { LLM_TENSOR_ROPE_FREQS, "rope_freqs" },
843
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
844
+ { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
845
+ { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
846
+ { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
847
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
848
+ { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" },
849
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
850
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
851
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
852
+ { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" },
853
+ { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" },
854
+ { LLM_TENSOR_SSM_X, "blk.%d.ssm_x" },
855
+ { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" },
856
+ { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" },
857
+ { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" },
858
+ { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" },
859
+ { LLM_TENSOR_SSM_DT_NORM, "blk.%d.ssm_dt_norm" },
860
+ { LLM_TENSOR_SSM_B_NORM, "blk.%d.ssm_b_norm" },
861
+ { LLM_TENSOR_SSM_C_NORM, "blk.%d.ssm_c_norm" },
862
+ { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" },
863
+ { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" },
864
+ },
865
+ },
736
866
  {
737
867
  LLM_ARCH_CODESHELL,
738
868
  {
@@ -892,6 +1022,63 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
892
1022
  { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" },
893
1023
  },
894
1024
  },
1025
+ {
1026
+ LLM_ARCH_GEMMA3N,
1027
+ {
1028
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
1029
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
1030
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
1031
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
1032
+ { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
1033
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
1034
+ { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
1035
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
1036
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
1037
+ { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" },
1038
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
1039
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
1040
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
1041
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1042
+ { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" },
1043
+ { LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "per_layer_token_embd" },
1044
+ { LLM_TENSOR_PER_LAYER_MODEL_PROJ, "per_layer_model_proj" },
1045
+ { LLM_TENSOR_PER_LAYER_PROJ_NORM, "per_layer_proj_norm" },
1046
+ { LLM_TENSOR_ALTUP_UNEMBD_PROJ, "altup_unembd_proj" },
1047
+ { LLM_TENSOR_ALTUP_PROJ, "altup_proj" },
1048
+ { LLM_TENSOR_PER_LAYER_INP_GATE, "blk.%d.inp_gate" },
1049
+ { LLM_TENSOR_PER_LAYER_PROJ, "blk.%d.proj" },
1050
+ { LLM_TENSOR_PER_LAYER_POST_NORM, "blk.%d.post_norm" },
1051
+ { LLM_TENSOR_ALTUP_CORRECT_COEF, "blk.%d.altup_correct_coef" },
1052
+ { LLM_TENSOR_ALTUP_CORRECT_SCALE, "blk.%d.altup_correct_scale" },
1053
+ { LLM_TENSOR_ALTUP_PREDICT_COEF, "blk.%d.altup_predict_coef" },
1054
+ { LLM_TENSOR_ALTUP_ROUTER, "blk.%d.altup_router" },
1055
+ { LLM_TENSOR_ALTUP_ROUTER_NORM, "blk.%d.altup_router_norm" },
1056
+ { LLM_TENSOR_LAUREL_L, "blk.%d.laurel_l" },
1057
+ { LLM_TENSOR_LAUREL_R, "blk.%d.laurel_r" },
1058
+ { LLM_TENSOR_LAUREL_POST_NORM, "blk.%d.laurel_post_norm" },
1059
+ },
1060
+ },
1061
+ {
1062
+ LLM_ARCH_GEMMA_EMBEDDING,
1063
+ {
1064
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
1065
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
1066
+ { LLM_TENSOR_OUTPUT, "output" },
1067
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
1068
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
1069
+ { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
1070
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
1071
+ { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
1072
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
1073
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
1074
+ { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" },
1075
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
1076
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
1077
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
1078
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1079
+ { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" },
1080
+ },
1081
+ },
895
1082
  {
896
1083
  LLM_ARCH_STARCODER2,
897
1084
  {
@@ -926,6 +1113,77 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
926
1113
  { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" },
927
1114
  },
928
1115
  },
1116
+ {
1117
+ LLM_ARCH_MAMBA2,
1118
+ {
1119
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
1120
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
1121
+ { LLM_TENSOR_OUTPUT, "output" },
1122
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
1123
+ { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" },
1124
+ { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" },
1125
+ { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" },
1126
+ { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" },
1127
+ { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" },
1128
+ { LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" },
1129
+ { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" },
1130
+ },
1131
+ },
1132
+ {
1133
+ LLM_ARCH_JAMBA,
1134
+ {
1135
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
1136
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
1137
+ { LLM_TENSOR_OUTPUT, "output" },
1138
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
1139
+ { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" },
1140
+ { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" },
1141
+ { LLM_TENSOR_SSM_X, "blk.%d.ssm_x" },
1142
+ { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" },
1143
+ { LLM_TENSOR_SSM_DT_NORM, "blk.%d.ssm_dt_norm" },
1144
+ { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" },
1145
+ { LLM_TENSOR_SSM_B_NORM, "blk.%d.ssm_b_norm" },
1146
+ { LLM_TENSOR_SSM_C_NORM, "blk.%d.ssm_c_norm" },
1147
+ { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" },
1148
+ { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" },
1149
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
1150
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
1151
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
1152
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
1153
+ { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
1154
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
1155
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
1156
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
1157
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1158
+ { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
1159
+ { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
1160
+ { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
1161
+ },
1162
+ },
1163
+ {
1164
+ LLM_ARCH_FALCON_H1,
1165
+ {
1166
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
1167
+ { LLM_TENSOR_OUTPUT, "output" },
1168
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
1169
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
1170
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
1171
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
1172
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
1173
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
1174
+ { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" },
1175
+ { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" },
1176
+ { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" },
1177
+ { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" },
1178
+ { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" },
1179
+ { LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" },
1180
+ { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" },
1181
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
1182
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
1183
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
1184
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1185
+ },
1186
+ },
929
1187
  {
930
1188
  LLM_ARCH_XVERSE,
931
1189
  {
@@ -1196,6 +1454,40 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
1196
1454
  { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" },
1197
1455
  },
1198
1456
  },
1457
+ {
1458
+ LLM_ARCH_GLM4_MOE,
1459
+ {
1460
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
1461
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
1462
+ { LLM_TENSOR_OUTPUT, "output" },
1463
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
1464
+ { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" },
1465
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
1466
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
1467
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
1468
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
1469
+ { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
1470
+ { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
1471
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
1472
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
1473
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1474
+ { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
1475
+ { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
1476
+ { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
1477
+ { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
1478
+ { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" },
1479
+ { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
1480
+ { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
1481
+ { LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
1482
+ // NextN/MTP tensors - preserved but unused (in final layer, dynamic layer number)
1483
+ { LLM_TENSOR_NEXTN_EH_PROJ, "blk.%d.nextn.eh_proj" },
1484
+ { LLM_TENSOR_NEXTN_EMBED_TOKENS, "blk.%d.nextn.embed_tokens" },
1485
+ { LLM_TENSOR_NEXTN_ENORM, "blk.%d.nextn.enorm" },
1486
+ { LLM_TENSOR_NEXTN_HNORM, "blk.%d.nextn.hnorm" },
1487
+ { LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "blk.%d.nextn.shared_head_head" },
1488
+ { LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "blk.%d.nextn.shared_head_norm" },
1489
+ },
1490
+ },
1199
1491
  {
1200
1492
  LLM_ARCH_BITNET,
1201
1493
  {
@@ -1300,6 +1592,31 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
1300
1592
  { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1301
1593
  },
1302
1594
  },
1595
+ {
1596
+ LLM_ARCH_NEMOTRON_H,
1597
+ {
1598
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
1599
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
1600
+ { LLM_TENSOR_OUTPUT, "output" },
1601
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
1602
+ // mamba(2) ssm layers
1603
+ { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" },
1604
+ { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" },
1605
+ { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" },
1606
+ { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" },
1607
+ { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" },
1608
+ { LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" },
1609
+ { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" },
1610
+ // attention layers
1611
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
1612
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
1613
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
1614
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
1615
+ // dense FFN
1616
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
1617
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1618
+ },
1619
+ },
1303
1620
  {
1304
1621
  LLM_ARCH_EXAONE,
1305
1622
  {
@@ -1319,6 +1636,26 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
1319
1636
  { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1320
1637
  },
1321
1638
  },
1639
+ {
1640
+ LLM_ARCH_EXAONE4,
1641
+ {
1642
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
1643
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
1644
+ { LLM_TENSOR_OUTPUT, "output" },
1645
+ { LLM_TENSOR_ROPE_FREQS, "rope_freqs" },
1646
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
1647
+ { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
1648
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
1649
+ { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
1650
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
1651
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
1652
+ { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" },
1653
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
1654
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
1655
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1656
+ { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" },
1657
+ }
1658
+ },
1322
1659
  {
1323
1660
  LLM_ARCH_RWKV6,
1324
1661
  {
@@ -1486,6 +1823,43 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
1486
1823
  { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
1487
1824
  },
1488
1825
  },
1826
+ {
1827
+ LLM_ARCH_GRANITE_HYBRID,
1828
+ {
1829
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
1830
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
1831
+ { LLM_TENSOR_OUTPUT, "output" },
1832
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
1833
+ // mamba(2) ssm layers
1834
+ { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" },
1835
+ { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" },
1836
+ { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" },
1837
+ { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" },
1838
+ { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" },
1839
+ { LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" },
1840
+ { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" },
1841
+ // attention layers
1842
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
1843
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
1844
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
1845
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
1846
+ // dense FFN
1847
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
1848
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
1849
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
1850
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1851
+ // moe FFN
1852
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
1853
+ { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
1854
+ { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
1855
+ { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
1856
+ { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
1857
+ // shared expert
1858
+ { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" },
1859
+ { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
1860
+ { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
1861
+ },
1862
+ },
1489
1863
  {
1490
1864
  LLM_ARCH_CHAMELEON,
1491
1865
  {
@@ -1553,6 +1927,292 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
1553
1927
  { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
1554
1928
  },
1555
1929
  },
1930
+ {
1931
+ LLM_ARCH_DOTS1,
1932
+ {
1933
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
1934
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
1935
+ { LLM_TENSOR_OUTPUT, "output" },
1936
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
1937
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
1938
+ { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
1939
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
1940
+ { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
1941
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
1942
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
1943
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
1944
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
1945
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1946
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
1947
+ { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
1948
+ { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
1949
+ { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
1950
+ { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
1951
+ { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
1952
+ { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" },
1953
+ { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
1954
+ { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
1955
+ { LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
1956
+ }
1957
+ },
1958
+ {
1959
+ LLM_ARCH_ERNIE4_5,
1960
+ {
1961
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
1962
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
1963
+ { LLM_TENSOR_OUTPUT, "output" },
1964
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
1965
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
1966
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
1967
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
1968
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
1969
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
1970
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
1971
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
1972
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1973
+ },
1974
+ },
1975
+ {
1976
+ LLM_ARCH_ERNIE4_5_MOE,
1977
+ {
1978
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
1979
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
1980
+ { LLM_TENSOR_OUTPUT, "output" },
1981
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
1982
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
1983
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
1984
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
1985
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
1986
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
1987
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
1988
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
1989
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1990
+ { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
1991
+ { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" },
1992
+ { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
1993
+ { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
1994
+ { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
1995
+ { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
1996
+ { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
1997
+ { LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
1998
+ },
1999
+ },
2000
+ {
2001
+ LLM_ARCH_HUNYUAN_MOE,
2002
+ {
2003
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
2004
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
2005
+ { LLM_TENSOR_OUTPUT, "output" },
2006
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
2007
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
2008
+ { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
2009
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
2010
+ { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
2011
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
2012
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
2013
+ { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
2014
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
2015
+ { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" },
2016
+ { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
2017
+ { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
2018
+ { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
2019
+ { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
2020
+ { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
2021
+ },
2022
+ },
2023
+ {
2024
+ LLM_ARCH_HUNYUAN_DENSE,
2025
+ {
2026
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
2027
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
2028
+ { LLM_TENSOR_OUTPUT, "output" },
2029
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
2030
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
2031
+ { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
2032
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
2033
+ { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
2034
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
2035
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
2036
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
2037
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
2038
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
2039
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
2040
+
2041
+ },
2042
+ },
2043
+ {
2044
+ LLM_ARCH_SMOLLM3,
2045
+ {
2046
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
2047
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
2048
+ { LLM_TENSOR_OUTPUT, "output" },
2049
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
2050
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
2051
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
2052
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
2053
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
2054
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
2055
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
2056
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
2057
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
2058
+ },
2059
+ },
2060
+ {
2061
+ LLM_ARCH_OPENAI_MOE,
2062
+ {
2063
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
2064
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
2065
+ { LLM_TENSOR_OUTPUT, "output" },
2066
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
2067
+ { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" },
2068
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
2069
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
2070
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
2071
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
2072
+ { LLM_TENSOR_ATTN_SINKS, "blk.%d.attn_sinks" },
2073
+ { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
2074
+ { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
2075
+ { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
2076
+ { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
2077
+ },
2078
+ },
2079
+ {
2080
+ LLM_ARCH_LFM2,
2081
+ {
2082
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
2083
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
2084
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
2085
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
2086
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
2087
+ { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
2088
+ { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
2089
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
2090
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
2091
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
2092
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
2093
+ { LLM_TENSOR_SHORTCONV_CONV, "blk.%d.shortconv.conv" },
2094
+ { LLM_TENSOR_SHORTCONV_INPROJ, "blk.%d.shortconv.in_proj" },
2095
+ { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
2096
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
2097
+ { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
2098
+ { LLM_TENSOR_OUTPUT, "output" },
2099
+ }
2100
+ },
2101
+ {
2102
+ LLM_ARCH_SMALLTHINKER,
2103
+ {
2104
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
2105
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
2106
+ { LLM_TENSOR_OUTPUT, "output" },
2107
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
2108
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
2109
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
2110
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
2111
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
2112
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
2113
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
2114
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
2115
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
2116
+ { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
2117
+ { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
2118
+ { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
2119
+ { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }
2120
+ },
2121
+ },
2122
+ {
2123
+ LLM_ARCH_DREAM,
2124
+ {
2125
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
2126
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
2127
+ { LLM_TENSOR_OUTPUT, "output" },
2128
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
2129
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
2130
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
2131
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
2132
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
2133
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
2134
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
2135
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
2136
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
2137
+ },
2138
+ },
2139
+ {
2140
+ LLM_ARCH_LLADA,
2141
+ {
2142
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
2143
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
2144
+ { LLM_TENSOR_OUTPUT, "output" },
2145
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
2146
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
2147
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
2148
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
2149
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
2150
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
2151
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
2152
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
2153
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
2154
+ },
2155
+ },
2156
+ {
2157
+ LLM_ARCH_LLADA_MOE,
2158
+ {
2159
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
2160
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
2161
+ { LLM_TENSOR_OUTPUT, "output" },
2162
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
2163
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
2164
+ { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
2165
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
2166
+ { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
2167
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
2168
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
2169
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
2170
+ { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
2171
+ { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
2172
+ { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
2173
+ { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
2174
+ },
2175
+ },
2176
+ {
2177
+ LLM_ARCH_SEED_OSS,
2178
+ {
2179
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
2180
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
2181
+ { LLM_TENSOR_OUTPUT, "output" },
2182
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
2183
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
2184
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
2185
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
2186
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
2187
+ { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" },
2188
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
2189
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
2190
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
2191
+ },
2192
+ },
2193
+ {
2194
+ LLM_ARCH_GROVEMOE,
2195
+ {
2196
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
2197
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
2198
+ { LLM_TENSOR_OUTPUT, "output" },
2199
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
2200
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
2201
+ { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
2202
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
2203
+ { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
2204
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
2205
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
2206
+ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
2207
+ { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
2208
+ { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
2209
+ { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
2210
+ { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
2211
+ { LLM_TENSOR_FFN_GATE_CHEXPS, "blk.%d.ffn_gate_chexps" },
2212
+ { LLM_TENSOR_FFN_DOWN_CHEXPS, "blk.%d.ffn_down_chexps" },
2213
+ { LLM_TENSOR_FFN_UP_CHEXPS, "blk.%d.ffn_up_chexps" },
2214
+ },
2215
+ },
1556
2216
  {
1557
2217
  LLM_ARCH_UNKNOWN,
1558
2218
  {
@@ -1592,6 +2252,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
1592
2252
  {LLM_TENSOR_ATTN_KV_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
1593
2253
  {LLM_TENSOR_ATTN_K_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
1594
2254
  {LLM_TENSOR_ATTN_V_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2255
+ {LLM_TENSOR_ATTN_SINKS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SCALE}},
1595
2256
  {LLM_TENSOR_DEC_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
1596
2257
  {LLM_TENSOR_DEC_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
1597
2258
  {LLM_TENSOR_DEC_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -1637,7 +2298,11 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
1637
2298
  {LLM_TENSOR_FFN_ACT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_DIV}},
1638
2299
  {LLM_TENSOR_SSM_CONV1D, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
1639
2300
  {LLM_TENSOR_SSM_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_SCAN}},
2301
+ {LLM_TENSOR_SSM_DT_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2302
+ {LLM_TENSOR_SSM_B_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2303
+ {LLM_TENSOR_SSM_C_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
1640
2304
  {LLM_TENSOR_SSM_D, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2305
+ {LLM_TENSOR_SSM_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
1641
2306
  {LLM_TENSOR_TIME_MIX_LERP_X, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
1642
2307
  {LLM_TENSOR_TIME_MIX_LN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
1643
2308
  {LLM_TENSOR_CHANNEL_MIX_LERP_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
@@ -1680,7 +2345,27 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
1680
2345
  {LLM_TENSOR_FFN_DOWN_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
1681
2346
  {LLM_TENSOR_FFN_GATE_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
1682
2347
  {LLM_TENSOR_FFN_UP_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
2348
+ {LLM_TENSOR_FFN_DOWN_CHEXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
2349
+ {LLM_TENSOR_FFN_GATE_CHEXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
2350
+ {LLM_TENSOR_FFN_UP_CHEXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
1683
2351
  {LLM_TENSOR_FFN_EXP_PROBS_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
2352
+ // altup / laurel (gemma 3n)
2353
+ {LLM_TENSOR_PER_LAYER_TOKEN_EMBD, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
2354
+ {LLM_TENSOR_PER_LAYER_MODEL_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
2355
+ {LLM_TENSOR_PER_LAYER_PROJ_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
2356
+ {LLM_TENSOR_ALTUP_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
2357
+ {LLM_TENSOR_ALTUP_UNEMBD_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
2358
+ {LLM_TENSOR_PER_LAYER_INP_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2359
+ {LLM_TENSOR_PER_LAYER_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2360
+ {LLM_TENSOR_PER_LAYER_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2361
+ {LLM_TENSOR_ALTUP_CORRECT_COEF, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2362
+ {LLM_TENSOR_ALTUP_CORRECT_SCALE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2363
+ {LLM_TENSOR_ALTUP_PREDICT_COEF, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2364
+ {LLM_TENSOR_ALTUP_ROUTER, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2365
+ {LLM_TENSOR_ALTUP_ROUTER_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2366
+ {LLM_TENSOR_LAUREL_L, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2367
+ {LLM_TENSOR_LAUREL_R, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2368
+ {LLM_TENSOR_LAUREL_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
1684
2369
  // this tensor is loaded for T5, but never used
1685
2370
  {LLM_TENSOR_DEC_CROSS_ATTN_REL_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
1686
2371
  {LLM_TENSOR_CONV1D, {LLM_TENSOR_LAYER_INPUT, GGML_OP_IM2COL}},
@@ -1699,13 +2384,30 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
1699
2384
  {LLM_TENSOR_CONVNEXT_PW1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
1700
2385
  {LLM_TENSOR_CONVNEXT_PW2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
1701
2386
  {LLM_TENSOR_CONVNEXT_GAMMA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2387
+ {LLM_TENSOR_SHORTCONV_CONV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
2388
+ {LLM_TENSOR_SHORTCONV_INPROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2389
+ {LLM_TENSOR_SHORTCONV_OUTPROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2390
+ // NextN/MTP tensors are currently ignored (reserved for future MTP support)
2391
+ // These tensors only exist in the last layer(s) and are treated as output tensors
2392
+ {LLM_TENSOR_NEXTN_EH_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
2393
+ {LLM_TENSOR_NEXTN_EMBED_TOKENS, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
2394
+ {LLM_TENSOR_NEXTN_ENORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
2395
+ {LLM_TENSOR_NEXTN_HNORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
2396
+ {LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
2397
+ {LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
1702
2398
  };
1703
2399
 
1704
2400
  LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}
1705
2401
 
1706
2402
  std::string LLM_KV::operator()(llm_kv kv) const {
1707
- return suffix ? ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch), suffix)
1708
- : ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
2403
+ std::string name = ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
2404
+
2405
+ if (suffix != nullptr) {
2406
+ name += ".";
2407
+ name += suffix;
2408
+ }
2409
+
2410
+ return name;
1709
2411
  }
1710
2412
 
1711
2413
  std::string LLM_TN_IMPL::str() const {
@@ -1744,3 +2446,42 @@ llm_arch llm_arch_from_string(const std::string & name) {
1744
2446
  const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor) {
1745
2447
  return LLM_TENSOR_INFOS.at(tensor);
1746
2448
  }
2449
+
2450
+ bool llm_arch_is_recurrent(const llm_arch & arch) {
2451
+ switch (arch) {
2452
+ case LLM_ARCH_MAMBA:
2453
+ case LLM_ARCH_MAMBA2:
2454
+ case LLM_ARCH_RWKV6:
2455
+ case LLM_ARCH_RWKV6QWEN2:
2456
+ case LLM_ARCH_RWKV7:
2457
+ case LLM_ARCH_ARWKV7:
2458
+ return true;
2459
+ default:
2460
+ return false;
2461
+ }
2462
+ }
2463
+
2464
+ bool llm_arch_is_hybrid(const llm_arch & arch) {
2465
+ switch (arch) {
2466
+ case LLM_ARCH_JAMBA:
2467
+ case LLM_ARCH_FALCON_H1:
2468
+ case LLM_ARCH_PLAMO2:
2469
+ case LLM_ARCH_GRANITE_HYBRID:
2470
+ case LLM_ARCH_LFM2:
2471
+ case LLM_ARCH_NEMOTRON_H:
2472
+ return true;
2473
+ default:
2474
+ return false;
2475
+ }
2476
+ }
2477
+
2478
+ bool llm_arch_is_diffusion(const llm_arch & arch) {
2479
+ switch (arch) {
2480
+ case LLM_ARCH_DREAM:
2481
+ case LLM_ARCH_LLADA:
2482
+ case LLM_ARCH_LLADA_MOE:
2483
+ return true;
2484
+ default:
2485
+ return false;
2486
+ }
2487
+ }