whispercpp 1.3.3 → 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 (586) hide show
  1. checksums.yaml +4 -4
  2. data/ext/ruby_whisper_params.c +55 -25
  3. data/ext/sources/CMakeLists.txt +1 -1
  4. data/ext/sources/bindings/javascript/package.json +1 -1
  5. data/ext/sources/build-xcframework.sh +24 -0
  6. data/ext/sources/examples/CMakeLists.txt +1 -0
  7. data/ext/sources/examples/addon.node/addon.cpp +19 -19
  8. data/ext/sources/examples/addon.node/index.js +7 -5
  9. data/ext/sources/examples/bench/bench.cpp +26 -16
  10. data/ext/sources/examples/bench.wasm/index-tmpl.html +10 -9
  11. data/ext/sources/examples/cli/cli.cpp +4 -2
  12. data/ext/sources/examples/command/command.cpp +26 -24
  13. data/ext/sources/examples/command.wasm/index-tmpl.html +5 -4
  14. data/ext/sources/examples/common-ggml.cpp +2 -0
  15. data/ext/sources/examples/lsp/lsp.cpp +19 -17
  16. data/ext/sources/examples/server/server.cpp +24 -13
  17. data/ext/sources/examples/server.py +6 -1
  18. data/ext/sources/examples/stream/stream.cpp +4 -2
  19. data/ext/sources/examples/stream.wasm/emscripten.cpp +6 -6
  20. data/ext/sources/examples/stream.wasm/index-tmpl.html +82 -5
  21. data/ext/sources/examples/talk-llama/CMakeLists.txt +2 -2
  22. data/ext/sources/examples/talk-llama/llama-adapter.cpp +101 -4
  23. data/ext/sources/examples/talk-llama/llama-adapter.h +6 -0
  24. data/ext/sources/examples/talk-llama/llama-arch.cpp +588 -15
  25. data/ext/sources/examples/talk-llama/llama-arch.h +58 -1
  26. data/ext/sources/examples/talk-llama/llama-batch.cpp +103 -71
  27. data/ext/sources/examples/talk-llama/llama-batch.h +31 -18
  28. data/ext/sources/examples/talk-llama/llama-chat.cpp +120 -5
  29. data/ext/sources/examples/talk-llama/llama-chat.h +7 -0
  30. data/ext/sources/examples/talk-llama/llama-context.cpp +460 -357
  31. data/ext/sources/examples/talk-llama/llama-context.h +44 -29
  32. data/ext/sources/examples/talk-llama/llama-cparams.h +4 -4
  33. data/ext/sources/examples/talk-llama/llama-graph.cpp +543 -271
  34. data/ext/sources/examples/talk-llama/llama-graph.h +278 -168
  35. data/ext/sources/examples/talk-llama/llama-hparams.cpp +118 -4
  36. data/ext/sources/examples/talk-llama/llama-hparams.h +61 -15
  37. data/ext/sources/examples/talk-llama/llama-impl.h +2 -0
  38. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +326 -0
  39. data/ext/sources/examples/talk-llama/{llama-kv-cache-unified-iswa.h → llama-kv-cache-iswa.h} +38 -29
  40. data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +2020 -0
  41. data/ext/sources/examples/talk-llama/llama-kv-cache.h +358 -27
  42. data/ext/sources/examples/talk-llama/llama-kv-cells.h +80 -28
  43. data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +56 -36
  44. data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +30 -29
  45. data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +48 -19
  46. data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +13 -14
  47. data/ext/sources/examples/talk-llama/llama-memory.h +16 -10
  48. data/ext/sources/examples/talk-llama/llama-model-loader.cpp +2 -0
  49. data/ext/sources/examples/talk-llama/llama-model-loader.h +3 -2
  50. data/ext/sources/examples/talk-llama/llama-model.cpp +7165 -2336
  51. data/ext/sources/examples/talk-llama/llama-model.h +60 -9
  52. data/ext/sources/examples/talk-llama/llama-quant.cpp +48 -10
  53. data/ext/sources/examples/talk-llama/llama-sampling.cpp +226 -126
  54. data/ext/sources/examples/talk-llama/llama-vocab.cpp +440 -13
  55. data/ext/sources/examples/talk-llama/llama-vocab.h +45 -0
  56. data/ext/sources/examples/talk-llama/llama.cpp +65 -10
  57. data/ext/sources/examples/talk-llama/llama.h +95 -177
  58. data/ext/sources/examples/talk-llama/talk-llama.cpp +9 -6
  59. data/ext/sources/examples/talk-llama/unicode.cpp +207 -0
  60. data/ext/sources/examples/talk-llama/unicode.h +45 -0
  61. data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +4 -2
  62. data/ext/sources/examples/whisper.wasm/index-tmpl.html +17 -16
  63. data/ext/sources/ggml/CMakeLists.txt +59 -31
  64. data/ext/sources/ggml/cmake/ggml-config.cmake.in +132 -93
  65. data/ext/sources/ggml/include/ggml-backend.h +17 -1
  66. data/ext/sources/ggml/include/ggml-cpu.h +1 -1
  67. data/ext/sources/ggml/include/ggml-metal.h +1 -6
  68. data/ext/sources/ggml/include/ggml-opt.h +25 -6
  69. data/ext/sources/ggml/include/ggml-webgpu.h +19 -0
  70. data/ext/sources/ggml/include/ggml-zdnn.h +17 -0
  71. data/ext/sources/ggml/include/ggml.h +221 -16
  72. data/ext/sources/ggml/src/CMakeLists.txt +17 -2
  73. data/ext/sources/ggml/src/ggml-alloc.c +265 -141
  74. data/ext/sources/ggml/src/ggml-backend-impl.h +4 -1
  75. data/ext/sources/ggml/src/ggml-backend-reg.cpp +30 -13
  76. data/ext/sources/ggml/src/ggml-backend.cpp +221 -38
  77. data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +1 -1
  78. data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +5 -4
  79. data/ext/sources/ggml/src/ggml-cann/CMakeLists.txt +14 -0
  80. data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +3 -1
  81. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +903 -717
  82. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +143 -25
  83. data/ext/sources/ggml/src/ggml-cann/common.h +143 -1
  84. data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +488 -69
  85. data/ext/sources/ggml/src/ggml-common.h +17 -0
  86. data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +40 -18
  87. data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +4 -2
  88. data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +132 -596
  89. data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +14 -286
  90. data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +103 -582
  91. data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +162 -589
  92. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +265 -437
  93. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +3 -58
  94. data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +521 -353
  95. data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +54 -314
  96. data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +184 -675
  97. data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +4679 -1657
  98. data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +32 -2
  99. data/ext/sources/ggml/src/ggml-cpu/common.h +14 -0
  100. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +13 -6
  101. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +70 -42
  102. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +35 -28
  103. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +152 -18
  104. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.h +7 -1
  105. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +227 -97
  106. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +474 -1116
  107. data/ext/sources/ggml/src/ggml-cpu/ops.cpp +1587 -1177
  108. data/ext/sources/ggml/src/ggml-cpu/ops.h +5 -8
  109. data/ext/sources/ggml/src/ggml-cpu/quants.c +35 -0
  110. data/ext/sources/ggml/src/ggml-cpu/quants.h +8 -0
  111. data/ext/sources/ggml/src/ggml-cpu/repack.cpp +458 -47
  112. data/ext/sources/ggml/src/ggml-cpu/repack.h +22 -0
  113. data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +89 -60
  114. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +1024 -0
  115. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.h +13 -0
  116. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +3196 -0
  117. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_kernels.h +26 -0
  118. data/ext/sources/ggml/src/ggml-cpu/traits.cpp +2 -2
  119. data/ext/sources/ggml/src/ggml-cpu/traits.h +1 -1
  120. data/ext/sources/ggml/src/ggml-cpu/vec.cpp +170 -26
  121. data/ext/sources/ggml/src/ggml-cpu/vec.h +506 -63
  122. data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +20 -16
  123. data/ext/sources/ggml/src/ggml-cuda/add-id.cu +58 -0
  124. data/ext/sources/ggml/src/ggml-cuda/add-id.cuh +3 -0
  125. data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +330 -191
  126. data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +2 -0
  127. data/ext/sources/ggml/src/ggml-cuda/common.cuh +250 -63
  128. data/ext/sources/ggml/src/ggml-cuda/conv-transpose-1d.cu +1 -4
  129. data/ext/sources/ggml/src/ggml-cuda/conv2d.cu +166 -0
  130. data/ext/sources/ggml/src/ggml-cuda/conv2d.cuh +5 -0
  131. data/ext/sources/ggml/src/ggml-cuda/convert.cu +95 -22
  132. data/ext/sources/ggml/src/ggml-cuda/convert.cuh +15 -0
  133. data/ext/sources/ggml/src/ggml-cuda/cpy-utils.cuh +217 -0
  134. data/ext/sources/ggml/src/ggml-cuda/cpy.cu +64 -307
  135. data/ext/sources/ggml/src/ggml-cuda/cross-entropy-loss.cu +2 -14
  136. data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +14 -40
  137. data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +498 -367
  138. data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +137 -91
  139. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +755 -0
  140. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +3 -0
  141. data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +593 -0
  142. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +86 -50
  143. data/ext/sources/ggml/src/ggml-cuda/fattn.cu +185 -198
  144. data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +2 -0
  145. data/ext/sources/ggml/src/ggml-cuda/getrows.cu +50 -39
  146. data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +379 -107
  147. data/ext/sources/ggml/src/ggml-cuda/im2col.cu +196 -35
  148. data/ext/sources/ggml/src/ggml-cuda/im2col.cuh +1 -0
  149. data/ext/sources/ggml/src/ggml-cuda/mean.cu +56 -2
  150. data/ext/sources/ggml/src/ggml-cuda/mma.cuh +198 -45
  151. data/ext/sources/ggml/src/ggml-cuda/mmf.cu +123 -0
  152. data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +496 -0
  153. data/ext/sources/ggml/src/ggml-cuda/mmq.cu +206 -57
  154. data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +1262 -721
  155. data/ext/sources/ggml/src/ggml-cuda/{mmv.cu → mmvf.cu} +53 -53
  156. data/ext/sources/ggml/src/ggml-cuda/{mmv.cuh → mmvf.cuh} +3 -3
  157. data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +64 -73
  158. data/ext/sources/ggml/src/ggml-cuda/norm.cu +284 -12
  159. data/ext/sources/ggml/src/ggml-cuda/norm.cuh +7 -0
  160. data/ext/sources/ggml/src/ggml-cuda/opt-step-sgd.cu +49 -0
  161. data/ext/sources/ggml/src/ggml-cuda/opt-step-sgd.cuh +5 -0
  162. data/ext/sources/ggml/src/ggml-cuda/pad.cu +46 -23
  163. data/ext/sources/ggml/src/ggml-cuda/pad_reflect_1d.cu +91 -0
  164. data/ext/sources/ggml/src/ggml-cuda/pad_reflect_1d.cuh +5 -0
  165. data/ext/sources/ggml/src/ggml-cuda/quantize.cu +12 -10
  166. data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +53 -0
  167. data/ext/sources/ggml/src/ggml-cuda/roll.cu +67 -0
  168. data/ext/sources/ggml/src/ggml-cuda/roll.cuh +5 -0
  169. data/ext/sources/ggml/src/ggml-cuda/rope.cu +21 -27
  170. data/ext/sources/ggml/src/ggml-cuda/scale.cu +14 -11
  171. data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +276 -0
  172. data/ext/sources/ggml/src/ggml-cuda/set-rows.cuh +7 -0
  173. data/ext/sources/ggml/src/ggml-cuda/softcap.cu +34 -0
  174. data/ext/sources/ggml/src/ggml-cuda/softcap.cuh +5 -0
  175. data/ext/sources/ggml/src/ggml-cuda/softmax.cu +126 -59
  176. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +10 -2
  177. data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +322 -100
  178. data/ext/sources/ggml/src/ggml-cuda/sum.cu +6 -10
  179. data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +21 -4
  180. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-f16.cu +7 -0
  181. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_0.cu +7 -0
  182. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q4_1.cu +7 -0
  183. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_0.cu +7 -0
  184. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q5_1.cu +7 -0
  185. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-q8_0.cu +7 -0
  186. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-f16.cu +7 -0
  187. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_0.cu +7 -0
  188. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q4_1.cu +7 -0
  189. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_0.cu +7 -0
  190. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q5_1.cu +7 -0
  191. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-q8_0.cu +7 -0
  192. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-f16.cu +7 -0
  193. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_0.cu +7 -0
  194. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q4_1.cu +7 -0
  195. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_0.cu +7 -0
  196. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q5_1.cu +7 -0
  197. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-q8_0.cu +7 -0
  198. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-f16.cu +7 -0
  199. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_0.cu +7 -0
  200. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q4_1.cu +7 -0
  201. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_0.cu +7 -0
  202. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q5_1.cu +7 -0
  203. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-q8_0.cu +7 -0
  204. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-f16.cu +7 -0
  205. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_0.cu +7 -0
  206. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q4_1.cu +7 -0
  207. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_0.cu +7 -0
  208. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q5_1.cu +7 -0
  209. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-q8_0.cu +7 -0
  210. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-f16.cu +7 -0
  211. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_0.cu +7 -0
  212. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q4_1.cu +7 -0
  213. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_0.cu +7 -0
  214. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q5_1.cu +7 -0
  215. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-q8_0.cu +7 -0
  216. data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +21 -18
  217. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_1.cu +5 -0
  218. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_10.cu +5 -0
  219. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_11.cu +5 -0
  220. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_12.cu +5 -0
  221. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_13.cu +5 -0
  222. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_14.cu +5 -0
  223. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_15.cu +5 -0
  224. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_16.cu +5 -0
  225. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_2.cu +5 -0
  226. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_3.cu +5 -0
  227. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_4.cu +5 -0
  228. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_5.cu +5 -0
  229. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_6.cu +5 -0
  230. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_7.cu +5 -0
  231. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_8.cu +5 -0
  232. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmf-instance-ncols_9.cu +5 -0
  233. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-mxfp4.cu +5 -0
  234. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +259 -0
  235. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cuh +14 -0
  236. data/ext/sources/ggml/src/ggml-cuda/tsembd.cu +3 -3
  237. data/ext/sources/ggml/src/ggml-cuda/unary.cu +90 -0
  238. data/ext/sources/ggml/src/ggml-cuda/unary.cuh +8 -0
  239. data/ext/sources/ggml/src/ggml-cuda/upscale.cu +92 -6
  240. data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +110 -22
  241. data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +4 -0
  242. data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +58 -36
  243. data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +4 -3
  244. data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +10 -2
  245. data/ext/sources/ggml/src/ggml-impl.h +119 -9
  246. data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +10 -7
  247. data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp +446 -0
  248. data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.h +52 -0
  249. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.h +33 -0
  250. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.m +600 -0
  251. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +1376 -0
  252. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +226 -0
  253. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +1308 -0
  254. data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +136 -63
  255. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +3158 -0
  256. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +82 -0
  257. data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +718 -0
  258. data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +2854 -1503
  259. data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +18 -8
  260. data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +18 -0
  261. data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +2510 -242
  262. data/ext/sources/ggml/src/ggml-opencl/kernels/add.cl +107 -0
  263. data/ext/sources/ggml/src/ggml-opencl/kernels/add_id.cl +42 -0
  264. data/ext/sources/ggml/src/ggml-opencl/kernels/conv2d.cl +185 -0
  265. data/ext/sources/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl +176 -0
  266. data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +84 -0
  267. data/ext/sources/ggml/src/ggml-opencl/kernels/div.cl +66 -0
  268. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl +370 -0
  269. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl +370 -0
  270. data/ext/sources/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl +373 -0
  271. data/ext/sources/ggml/src/ggml-opencl/kernels/gelu.cl +27 -0
  272. data/ext/sources/ggml/src/ggml-opencl/kernels/glu.cl +177 -0
  273. data/ext/sources/ggml/src/ggml-opencl/kernels/group_norm.cl +49 -0
  274. data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f16.cl +1 -1
  275. data/ext/sources/ggml/src/ggml-opencl/kernels/im2col_f32.cl +1 -1
  276. data/ext/sources/ggml/src/ggml-opencl/kernels/mul.cl +73 -0
  277. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl +130 -0
  278. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl +132 -0
  279. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl +133 -0
  280. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl +189 -0
  281. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl +176 -0
  282. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl +140 -0
  283. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl +222 -0
  284. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl +144 -0
  285. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl +167 -0
  286. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl +125 -0
  287. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl +202 -0
  288. data/ext/sources/ggml/src/ggml-opencl/kernels/norm.cl +80 -0
  289. data/ext/sources/ggml/src/ggml-opencl/kernels/rms_norm.cl +79 -0
  290. data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +3 -2
  291. data/ext/sources/ggml/src/ggml-opencl/kernels/set_rows.cl +189 -0
  292. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl +34 -13
  293. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl +34 -13
  294. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f16.cl +34 -13
  295. data/ext/sources/ggml/src/ggml-opencl/kernels/softmax_f32.cl +34 -13
  296. data/ext/sources/ggml/src/ggml-opencl/kernels/sub.cl +66 -0
  297. data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +20 -0
  298. data/ext/sources/ggml/src/ggml-opencl/kernels/tsembd.cl +2 -2
  299. data/ext/sources/ggml/src/ggml-opencl/kernels/upscale.cl +2 -3
  300. data/ext/sources/ggml/src/ggml-opt.cpp +97 -41
  301. data/ext/sources/ggml/src/ggml-quants.c +111 -16
  302. data/ext/sources/ggml/src/ggml-quants.h +6 -0
  303. data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +67 -47
  304. data/ext/sources/ggml/src/ggml-sycl/backend.hpp +2 -0
  305. data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +15 -5
  306. data/ext/sources/ggml/src/ggml-sycl/binbcast.hpp +6 -0
  307. data/ext/sources/ggml/src/ggml-sycl/concat.cpp +25 -16
  308. data/ext/sources/ggml/src/ggml-sycl/conv.cpp +10 -4
  309. data/ext/sources/ggml/src/ggml-sycl/convert.cpp +166 -99
  310. data/ext/sources/ggml/src/ggml-sycl/cpy.cpp +72 -306
  311. data/ext/sources/ggml/src/ggml-sycl/cpy.hpp +213 -1
  312. data/ext/sources/ggml/src/ggml-sycl/dmmv.cpp +67 -49
  313. data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +1 -31
  314. data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +79 -29
  315. data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -0
  316. data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +14 -26
  317. data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +9 -6
  318. data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +328 -323
  319. data/ext/sources/ggml/src/ggml-sycl/gla.cpp +2 -2
  320. data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +2 -2
  321. data/ext/sources/ggml/src/ggml-sycl/mmq.cpp +80 -60
  322. data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +201 -132
  323. data/ext/sources/ggml/src/ggml-sycl/norm.cpp +74 -55
  324. data/ext/sources/ggml/src/ggml-sycl/quantize.hpp +133 -0
  325. data/ext/sources/ggml/src/ggml-sycl/quants.hpp +8 -9
  326. data/ext/sources/ggml/src/ggml-sycl/rope.cpp +35 -42
  327. data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +234 -0
  328. data/ext/sources/ggml/src/ggml-sycl/set_rows.hpp +8 -0
  329. data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +3 -3
  330. data/ext/sources/ggml/src/ggml-sycl/tsembd.cpp +12 -6
  331. data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +2 -6
  332. data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +16 -12
  333. data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +3492 -883
  334. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add.comp +41 -1
  335. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/add_id.comp +42 -0
  336. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp +13 -4
  337. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp +39 -29
  338. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +349 -0
  339. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +2 -2
  340. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +66 -12
  341. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +154 -0
  342. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp +21 -0
  343. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp +1 -1
  344. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp +2 -1
  345. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp +6 -5
  346. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp +4 -2
  347. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp +32 -0
  348. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp +1 -1
  349. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp +1 -1
  350. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp +1 -1
  351. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp +1 -1
  352. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp +1 -1
  353. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +21 -0
  354. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +69 -24
  355. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp +60 -20
  356. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +98 -42
  357. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +64 -27
  358. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +74 -13
  359. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/geglu_erf.comp +27 -0
  360. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/geglu_quick.comp +11 -0
  361. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gelu_erf.comp +39 -0
  362. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp +4 -17
  363. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +19 -10
  364. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +25 -15
  365. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp +4 -0
  366. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardsigmoid.comp +22 -0
  367. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/hardswish.comp +22 -0
  368. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +18 -14
  369. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +126 -0
  370. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp +65 -1
  371. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +11 -7
  372. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +140 -0
  373. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +144 -531
  374. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +206 -38
  375. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.comp +556 -0
  376. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +12 -5
  377. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp +15 -9
  378. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +111 -0
  379. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp +22 -0
  380. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/pad.comp +24 -3
  381. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp +53 -3
  382. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +55 -11
  383. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_partials.comp +65 -0
  384. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/roll.comp +46 -0
  385. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp +1 -4
  386. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +7 -9
  387. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +7 -9
  388. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +7 -9
  389. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.comp +5 -0
  390. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/scale.comp +1 -1
  391. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp +29 -7
  392. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp +4 -0
  393. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sqrt.comp +17 -0
  394. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp +38 -5
  395. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/swiglu_oai.comp +14 -0
  396. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp +4 -3
  397. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +101 -9
  398. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +69 -5
  399. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/utils.comp +25 -0
  400. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +335 -77
  401. data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +54 -0
  402. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +1558 -0
  403. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add.tmpl.wgsl +44 -0
  404. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_in_place.tmpl.wgsl +41 -0
  405. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary_head.tmpl +45 -0
  406. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +930 -0
  407. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl +60 -0
  408. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +124 -0
  409. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/get_rows.tmpl.wgsl +874 -0
  410. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl +40 -0
  411. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul.tmpl.wgsl +44 -0
  412. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_in_place.tmpl.wgsl +41 -0
  413. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl +907 -0
  414. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +57 -0
  415. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_in_place.wgsl +48 -0
  416. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +81 -0
  417. data/ext/sources/ggml/src/ggml-zdnn/CMakeLists.txt +36 -0
  418. data/ext/sources/ggml/src/ggml-zdnn/common.hpp +59 -0
  419. data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +628 -0
  420. data/ext/sources/ggml/src/ggml-zdnn/mmf.cpp +80 -0
  421. data/ext/sources/ggml/src/ggml-zdnn/mmf.hpp +12 -0
  422. data/ext/sources/ggml/src/ggml-zdnn/utils.cpp +79 -0
  423. data/ext/sources/ggml/src/ggml-zdnn/utils.hpp +19 -0
  424. data/ext/sources/ggml/src/ggml.c +478 -98
  425. data/ext/sources/ggml/src/gguf.cpp +8 -1
  426. data/ext/sources/src/whisper.cpp +23 -46
  427. data/ext/sources/tests/CMakeLists.txt +8 -1
  428. data/ext/sources/tests/test-vad-full.cpp +3 -3
  429. data/ext/sources/tests/test-vad.cpp +2 -2
  430. data/lib/whisper/model/uri.rb +1 -1
  431. data/sig/whisper.rbs +7 -0
  432. data/test/test_params.rb +8 -0
  433. data/test/test_whisper.rb +1 -1
  434. data/whispercpp.gemspec +1 -1
  435. metadata +164 -157
  436. data/ext/sources/examples/talk-llama/llama-kv-cache-unified-iswa.cpp +0 -279
  437. data/ext/sources/examples/talk-llama/llama-kv-cache-unified.cpp +0 -1841
  438. data/ext/sources/examples/talk-llama/llama-kv-cache-unified.h +0 -303
  439. data/ext/sources/ggml/include/ggml-kompute.h +0 -50
  440. data/ext/sources/ggml/src/ggml-amx/CMakeLists.txt +0 -107
  441. data/ext/sources/ggml/src/ggml-amx/common.h +0 -94
  442. data/ext/sources/ggml/src/ggml-amx/ggml-amx.cpp +0 -446
  443. data/ext/sources/ggml/src/ggml-amx/mmq.cpp +0 -2510
  444. data/ext/sources/ggml/src/ggml-amx/mmq.h +0 -17
  445. data/ext/sources/ggml/src/ggml-cann/kernels/CMakeLists.txt +0 -30
  446. data/ext/sources/ggml/src/ggml-cann/kernels/ascendc_kernels.h +0 -19
  447. data/ext/sources/ggml/src/ggml-cann/kernels/dup.cpp +0 -234
  448. data/ext/sources/ggml/src/ggml-cann/kernels/get_row_f16.cpp +0 -197
  449. data/ext/sources/ggml/src/ggml-cann/kernels/get_row_f32.cpp +0 -190
  450. data/ext/sources/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp +0 -204
  451. data/ext/sources/ggml/src/ggml-cann/kernels/get_row_q8_0.cpp +0 -191
  452. data/ext/sources/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp +0 -218
  453. data/ext/sources/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp +0 -216
  454. data/ext/sources/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp +0 -295
  455. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cu +0 -357
  456. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f16.cuh +0 -3
  457. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cu +0 -365
  458. data/ext/sources/ggml/src/ggml-cuda/fattn-tile-f32.cuh +0 -3
  459. data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f16.cuh +0 -482
  460. data/ext/sources/ggml/src/ggml-cuda/fattn-vec-f32.cuh +0 -472
  461. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu +0 -5
  462. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu +0 -5
  463. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu +0 -5
  464. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu +0 -5
  465. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu +0 -5
  466. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu +0 -5
  467. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu +0 -5
  468. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu +0 -5
  469. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu +0 -5
  470. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu +0 -5
  471. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu +0 -5
  472. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu +0 -5
  473. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu +0 -5
  474. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu +0 -5
  475. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu +0 -5
  476. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu +0 -5
  477. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu +0 -5
  478. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu +0 -5
  479. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu +0 -5
  480. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu +0 -5
  481. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu +0 -5
  482. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu +0 -5
  483. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu +0 -5
  484. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu +0 -5
  485. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu +0 -5
  486. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu +0 -5
  487. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu +0 -5
  488. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu +0 -5
  489. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu +0 -5
  490. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu +0 -5
  491. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu +0 -5
  492. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu +0 -5
  493. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu +0 -5
  494. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu +0 -5
  495. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu +0 -5
  496. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu +0 -5
  497. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu +0 -5
  498. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu +0 -5
  499. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu +0 -5
  500. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu +0 -5
  501. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu +0 -5
  502. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu +0 -5
  503. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu +0 -5
  504. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu +0 -5
  505. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu +0 -5
  506. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu +0 -5
  507. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu +0 -5
  508. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu +0 -5
  509. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu +0 -5
  510. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu +0 -5
  511. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu +0 -5
  512. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu +0 -5
  513. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu +0 -5
  514. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu +0 -5
  515. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu +0 -5
  516. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu +0 -5
  517. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu +0 -5
  518. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu +0 -5
  519. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu +0 -5
  520. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu +0 -5
  521. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu +0 -5
  522. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu +0 -5
  523. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu +0 -5
  524. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu +0 -5
  525. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu +0 -5
  526. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu +0 -5
  527. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu +0 -5
  528. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu +0 -5
  529. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu +0 -5
  530. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu +0 -5
  531. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu +0 -5
  532. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu +0 -5
  533. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu +0 -5
  534. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu +0 -5
  535. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu +0 -5
  536. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu +0 -5
  537. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu +0 -5
  538. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu +0 -5
  539. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu +0 -5
  540. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu +0 -5
  541. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu +0 -5
  542. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu +0 -5
  543. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu +0 -5
  544. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu +0 -5
  545. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu +0 -5
  546. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu +0 -5
  547. data/ext/sources/ggml/src/ggml-kompute/CMakeLists.txt +0 -166
  548. data/ext/sources/ggml/src/ggml-kompute/ggml-kompute.cpp +0 -2251
  549. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/common.comp +0 -112
  550. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_add.comp +0 -58
  551. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp +0 -25
  552. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp +0 -52
  553. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp +0 -52
  554. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp +0 -52
  555. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp +0 -52
  556. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp +0 -30
  557. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp +0 -22
  558. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp +0 -17
  559. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp +0 -31
  560. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp +0 -31
  561. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp +0 -38
  562. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp +0 -39
  563. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp +0 -44
  564. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp +0 -52
  565. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp +0 -69
  566. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp +0 -51
  567. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp +0 -33
  568. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp +0 -35
  569. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp +0 -140
  570. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp +0 -106
  571. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp +0 -73
  572. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp +0 -52
  573. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp +0 -28
  574. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp +0 -84
  575. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp +0 -21
  576. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp +0 -53
  577. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp +0 -52
  578. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp +0 -52
  579. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp +0 -52
  580. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp +0 -52
  581. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp +0 -19
  582. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp +0 -23
  583. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp +0 -22
  584. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp +0 -72
  585. data/ext/sources/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp +0 -71
  586. data/ext/sources/ggml/src/ggml-metal/ggml-metal.m +0 -6280
@@ -22,6 +22,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
22
22
  { LLM_ARCH_NOMIC_BERT_MOE, "nomic-bert-moe" },
23
23
  { LLM_ARCH_NEO_BERT, "neo-bert" },
24
24
  { LLM_ARCH_JINA_BERT_V2, "jina-bert-v2" },
25
+ { LLM_ARCH_JINA_BERT_V3, "jina-bert-v3" },
25
26
  { LLM_ARCH_BLOOM, "bloom" },
26
27
  { LLM_ARCH_STABLELM, "stablelm" },
27
28
  { LLM_ARCH_QWEN, "qwen" },
@@ -34,6 +35,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
34
35
  { LLM_ARCH_PHI3, "phi3" },
35
36
  { LLM_ARCH_PHIMOE, "phimoe" },
36
37
  { LLM_ARCH_PLAMO, "plamo" },
38
+ { LLM_ARCH_PLAMO2, "plamo2" },
37
39
  { LLM_ARCH_CODESHELL, "codeshell" },
38
40
  { LLM_ARCH_ORION, "orion" },
39
41
  { LLM_ARCH_INTERNLM2, "internlm2" },
@@ -43,8 +45,12 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
43
45
  { LLM_ARCH_GEMMA2, "gemma2" },
44
46
  { LLM_ARCH_GEMMA3, "gemma3" },
45
47
  { LLM_ARCH_GEMMA3N, "gemma3n" },
48
+ { LLM_ARCH_GEMMA_EMBEDDING, "gemma-embedding" },
46
49
  { LLM_ARCH_STARCODER2, "starcoder2" },
47
50
  { LLM_ARCH_MAMBA, "mamba" },
51
+ { LLM_ARCH_MAMBA2, "mamba2" },
52
+ { LLM_ARCH_JAMBA, "jamba" },
53
+ { LLM_ARCH_FALCON_H1, "falcon-h1" },
48
54
  { LLM_ARCH_XVERSE, "xverse" },
49
55
  { LLM_ARCH_COMMAND_R, "command-r" },
50
56
  { LLM_ARCH_COHERE2, "cohere2" },
@@ -58,18 +64,22 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
58
64
  { LLM_ARCH_DEEPSEEK2, "deepseek2" },
59
65
  { LLM_ARCH_CHATGLM, "chatglm" },
60
66
  { LLM_ARCH_GLM4, "glm4" },
67
+ { LLM_ARCH_GLM4_MOE, "glm4moe" },
61
68
  { LLM_ARCH_BITNET, "bitnet" },
62
69
  { LLM_ARCH_T5, "t5" },
63
70
  { LLM_ARCH_T5ENCODER, "t5encoder" },
64
71
  { LLM_ARCH_JAIS, "jais" },
65
72
  { LLM_ARCH_NEMOTRON, "nemotron" },
73
+ { LLM_ARCH_NEMOTRON_H, "nemotron_h" },
66
74
  { LLM_ARCH_EXAONE, "exaone" },
75
+ { LLM_ARCH_EXAONE4, "exaone4" },
67
76
  { LLM_ARCH_RWKV6, "rwkv6" },
68
77
  { LLM_ARCH_RWKV6QWEN2, "rwkv6qwen2" },
69
78
  { LLM_ARCH_RWKV7, "rwkv7" },
70
79
  { LLM_ARCH_ARWKV7, "arwkv7" },
71
80
  { LLM_ARCH_GRANITE, "granite" },
72
81
  { LLM_ARCH_GRANITE_MOE, "granitemoe" },
82
+ { LLM_ARCH_GRANITE_HYBRID, "granitehybrid" },
73
83
  { LLM_ARCH_CHAMELEON, "chameleon" },
74
84
  { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
75
85
  { LLM_ARCH_PLM, "plm" },
@@ -77,6 +87,18 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
77
87
  { LLM_ARCH_DOTS1, "dots1" },
78
88
  { LLM_ARCH_ARCEE, "arcee" },
79
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" },
80
102
  { LLM_ARCH_UNKNOWN, "(unknown)" },
81
103
  };
82
104
 
@@ -104,6 +126,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
104
126
  { LLM_KV_FEED_FORWARD_LENGTH, "%s.feed_forward_length" },
105
127
  { LLM_KV_EXPERT_FEED_FORWARD_LENGTH, "%s.expert_feed_forward_length" },
106
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" },
107
130
  { LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" },
108
131
  { LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" },
109
132
  { LLM_KV_EXPERT_COUNT, "%s.expert_count" },
@@ -112,11 +135,16 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
112
135
  { LLM_KV_EXPERT_WEIGHTS_SCALE, "%s.expert_weights_scale" },
113
136
  { LLM_KV_EXPERT_WEIGHTS_NORM, "%s.expert_weights_norm" },
114
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" },
115
140
  { LLM_KV_MOE_EVERY_N_LAYERS, "%s.moe_every_n_layers" },
141
+ { LLM_KV_NEXTN_PREDICT_LAYERS, "%s.nextn_predict_layers" },
116
142
  { LLM_KV_POOLING_TYPE, "%s.pooling_type" },
117
143
  { LLM_KV_LOGIT_SCALE, "%s.logit_scale" },
118
144
  { LLM_KV_DECODER_START_TOKEN_ID, "%s.decoder_start_token_id" },
145
+ { LLM_KV_DECODER_BLOCK_COUNT, "%s.decoder_block_count" },
119
146
  { LLM_KV_ATTN_LOGIT_SOFTCAPPING, "%s.attn_logit_softcapping" },
147
+ { LLM_KV_ROUTER_LOGIT_SOFTCAPPING, "%s.router_logit_softcapping" },
120
148
  { LLM_KV_FINAL_LOGIT_SOFTCAPPING, "%s.final_logit_softcapping" },
121
149
  { LLM_KV_SWIN_NORM, "%s.swin_norm" },
122
150
  { LLM_KV_RESCALE_EVERY_N_LAYERS, "%s.rescale_every_n_layers" },
@@ -147,20 +175,25 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
147
175
  { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
148
176
  { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" },
149
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" },
150
180
  { LLM_KV_ATTENTION_KEY_LENGTH_MLA, "%s.attention.key_length_mla" },
151
181
  { LLM_KV_ATTENTION_VALUE_LENGTH_MLA, "%s.attention.value_length_mla" },
152
- { LLM_KV_ATTENTION_LAYER_INDICES, "%s.attention.layer_indices" },
153
182
 
154
- { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
155
- { LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" },
156
- { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" },
157
- { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" },
158
- { LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" },
159
- { LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" },
160
- { LLM_KV_ROPE_SCALING_ATTN_FACTOR, "%s.rope.scaling.attn_factor" },
161
- { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" },
162
- { LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" },
163
- { 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" },
164
197
 
165
198
  { LLM_KV_SPLIT_NO, "split.no" },
166
199
  { LLM_KV_SPLIT_COUNT, "split.count" },
@@ -170,6 +203,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
170
203
  { LLM_KV_SSM_INNER_SIZE, "%s.ssm.inner_size" },
171
204
  { LLM_KV_SSM_STATE_SIZE, "%s.ssm.state_size" },
172
205
  { LLM_KV_SSM_TIME_STEP_RANK, "%s.ssm.time_step_rank" },
206
+ { LLM_KV_SSM_GROUP_COUNT, "%s.ssm.group_count" },
173
207
  { LLM_KV_SSM_DT_B_C_RMS, "%s.ssm.dt_b_c_rms" },
174
208
 
175
209
  { LLM_KV_WKV_HEAD_SIZE, "%s.wkv.head_size" },
@@ -182,6 +216,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
182
216
 
183
217
  { LLM_KV_CLASSIFIER_OUTPUT_LABELS, "%s.classifier.output_labels" },
184
218
 
219
+ { LLM_KV_SHORTCONV_L_CACHE, "%s.shortconv.l_cache" },
220
+
185
221
  { LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" },
186
222
  { LLM_KV_TOKENIZER_PRE, "tokenizer.ggml.pre" },
187
223
  { LLM_KV_TOKENIZER_LIST, "tokenizer.ggml.tokens" },
@@ -214,8 +250,11 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
214
250
  { LLM_KV_TOKENIZER_FIM_REP_ID, "tokenizer.ggml.fim_rep_token_id" },
215
251
  { LLM_KV_TOKENIZER_FIM_SEP_ID, "tokenizer.ggml.fim_sep_token_id" },
216
252
 
217
- { LLM_KV_ADAPTER_TYPE, "adapter.type" },
218
- { 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" },
219
258
 
220
259
  // deprecated
221
260
  { LLM_KV_TOKENIZER_PREFIX_ID, "tokenizer.ggml.prefix_token_id" },
@@ -371,12 +410,16 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
371
410
  { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" },
372
411
  { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
373
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" },
374
416
  { LLM_TENSOR_FFN_GATE_EXP, "blk.%d.ffn_gate.%d" },
375
417
  { LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" },
376
418
  { LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" },
377
419
  { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
378
420
  { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
379
421
  { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
422
+ { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" },
380
423
  { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" },
381
424
  { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" },
382
425
  },
@@ -555,6 +598,20 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
555
598
  { LLM_TENSOR_CLS, "cls" },
556
599
  },
557
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
+ },
558
615
  {
559
616
  LLM_ARCH_BLOOM,
560
617
  {
@@ -668,6 +725,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
668
725
  { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
669
726
  { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
670
727
  { LLM_TENSOR_OUTPUT, "output" },
728
+ { LLM_TENSOR_CLS_OUT, "cls.output" },
671
729
  { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
672
730
  { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
673
731
  { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
@@ -775,6 +833,36 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
775
833
  { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
776
834
  },
777
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
+ },
778
866
  {
779
867
  LLM_ARCH_CODESHELL,
780
868
  {
@@ -970,6 +1058,27 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
970
1058
  { LLM_TENSOR_LAUREL_POST_NORM, "blk.%d.laurel_post_norm" },
971
1059
  },
972
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
+ },
973
1082
  {
974
1083
  LLM_ARCH_STARCODER2,
975
1084
  {
@@ -1004,6 +1113,77 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
1004
1113
  { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" },
1005
1114
  },
1006
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
+ },
1007
1187
  {
1008
1188
  LLM_ARCH_XVERSE,
1009
1189
  {
@@ -1274,6 +1454,40 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
1274
1454
  { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" },
1275
1455
  },
1276
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
+ },
1277
1491
  {
1278
1492
  LLM_ARCH_BITNET,
1279
1493
  {
@@ -1378,6 +1592,31 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
1378
1592
  { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1379
1593
  },
1380
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
+ },
1381
1620
  {
1382
1621
  LLM_ARCH_EXAONE,
1383
1622
  {
@@ -1397,6 +1636,26 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
1397
1636
  { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1398
1637
  },
1399
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
+ },
1400
1659
  {
1401
1660
  LLM_ARCH_RWKV6,
1402
1661
  {
@@ -1564,6 +1823,43 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
1564
1823
  { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
1565
1824
  },
1566
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
+ },
1567
1863
  {
1568
1864
  LLM_ARCH_CHAMELEON,
1569
1865
  {
@@ -1676,6 +1972,247 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
1676
1972
  { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
1677
1973
  },
1678
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
+ },
1679
2216
  {
1680
2217
  LLM_ARCH_UNKNOWN,
1681
2218
  {
@@ -1715,6 +2252,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
1715
2252
  {LLM_TENSOR_ATTN_KV_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
1716
2253
  {LLM_TENSOR_ATTN_K_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
1717
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}},
1718
2256
  {LLM_TENSOR_DEC_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
1719
2257
  {LLM_TENSOR_DEC_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
1720
2258
  {LLM_TENSOR_DEC_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -1760,7 +2298,11 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
1760
2298
  {LLM_TENSOR_FFN_ACT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_DIV}},
1761
2299
  {LLM_TENSOR_SSM_CONV1D, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
1762
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}},
1763
2304
  {LLM_TENSOR_SSM_D, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2305
+ {LLM_TENSOR_SSM_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
1764
2306
  {LLM_TENSOR_TIME_MIX_LERP_X, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
1765
2307
  {LLM_TENSOR_TIME_MIX_LN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
1766
2308
  {LLM_TENSOR_CHANNEL_MIX_LERP_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
@@ -1803,6 +2345,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
1803
2345
  {LLM_TENSOR_FFN_DOWN_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
1804
2346
  {LLM_TENSOR_FFN_GATE_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
1805
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}},
1806
2351
  {LLM_TENSOR_FFN_EXP_PROBS_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
1807
2352
  // altup / laurel (gemma 3n)
1808
2353
  {LLM_TENSOR_PER_LAYER_TOKEN_EMBD, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
@@ -1839,6 +2384,17 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
1839
2384
  {LLM_TENSOR_CONVNEXT_PW1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
1840
2385
  {LLM_TENSOR_CONVNEXT_PW2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
1841
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}},
1842
2398
  };
1843
2399
 
1844
2400
  LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}
@@ -1894,6 +2450,7 @@ const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor) {
1894
2450
  bool llm_arch_is_recurrent(const llm_arch & arch) {
1895
2451
  switch (arch) {
1896
2452
  case LLM_ARCH_MAMBA:
2453
+ case LLM_ARCH_MAMBA2:
1897
2454
  case LLM_ARCH_RWKV6:
1898
2455
  case LLM_ARCH_RWKV6QWEN2:
1899
2456
  case LLM_ARCH_RWKV7:
@@ -1905,9 +2462,25 @@ bool llm_arch_is_recurrent(const llm_arch & arch) {
1905
2462
  }
1906
2463
 
1907
2464
  bool llm_arch_is_hybrid(const llm_arch & arch) {
1908
- // TODO: There are currently no hybrid models! Once there are, this will be
1909
- // the place to identify them
1910
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;
1911
2484
  default:
1912
2485
  return false;
1913
2486
  }