whispercpp 1.3.5 → 1.3.7

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 (1017) hide show
  1. checksums.yaml +4 -4
  2. data/.document +3 -0
  3. data/.rdoc_options +2 -0
  4. data/LICENSE +1 -1
  5. data/README.md +133 -3
  6. data/Rakefile +18 -3
  7. data/ext/dependencies.rb +10 -4
  8. data/ext/dependencies_for_windows.rb +17 -0
  9. data/ext/extconf.rb +20 -7
  10. data/ext/options.rb +54 -14
  11. data/ext/options_for_windows.rb +51 -0
  12. data/ext/ruby_whisper.c +56 -46
  13. data/ext/ruby_whisper.h +165 -2
  14. data/ext/ruby_whisper_context.c +297 -126
  15. data/ext/ruby_whisper_context_params.c +163 -0
  16. data/ext/ruby_whisper_log_queue.c +180 -0
  17. data/ext/ruby_whisper_log_settable.h +47 -0
  18. data/ext/ruby_whisper_model.c +0 -1
  19. data/ext/ruby_whisper_parakeet.c +49 -0
  20. data/ext/ruby_whisper_parakeet_context.c +304 -0
  21. data/ext/ruby_whisper_parakeet_context_params.c +117 -0
  22. data/ext/ruby_whisper_parakeet_model.c +84 -0
  23. data/ext/ruby_whisper_parakeet_params.c +548 -0
  24. data/ext/ruby_whisper_parakeet_segment.c +157 -0
  25. data/ext/ruby_whisper_parakeet_token.c +188 -0
  26. data/ext/ruby_whisper_parakeet_transcribe.cpp +58 -0
  27. data/ext/ruby_whisper_params.c +256 -66
  28. data/ext/ruby_whisper_segment.c +6 -7
  29. data/ext/ruby_whisper_token.c +29 -9
  30. data/ext/ruby_whisper_transcribe.cpp +46 -16
  31. data/ext/ruby_whisper_vad_context.c +48 -1
  32. data/ext/ruby_whisper_vad_context_detect.cpp +6 -5
  33. data/ext/ruby_whisper_vad_params.c +0 -1
  34. data/ext/ruby_whisper_vad_segment.c +0 -1
  35. data/ext/ruby_whisper_vad_segments.c +0 -1
  36. data/ext/sources/CMakeLists.txt +41 -3
  37. data/ext/sources/CMakePresets.json +95 -0
  38. data/ext/sources/cmake/parakeet-config.cmake.in +30 -0
  39. data/ext/sources/cmake/parakeet.pc.in +10 -0
  40. data/ext/sources/cmake/whisper-config.cmake.in +5 -40
  41. data/ext/sources/cmake/whisper.pc.in +1 -1
  42. data/ext/sources/examples/CMakeLists.txt +4 -2
  43. data/ext/sources/examples/bench/bench.cpp +24 -19
  44. data/ext/sources/examples/cli/cli.cpp +51 -9
  45. data/ext/sources/examples/common-ggml.cpp +4 -0
  46. data/ext/sources/examples/common-whisper.cpp +139 -67
  47. data/ext/sources/examples/common-whisper.h +11 -0
  48. data/ext/sources/examples/ffmpeg-transcode.cpp +211 -341
  49. data/ext/sources/examples/miniaudio.h +4507 -2131
  50. data/ext/sources/examples/parakeet-cli/CMakeLists.txt +8 -0
  51. data/ext/sources/examples/parakeet-cli/parakeet-cli.cpp +243 -0
  52. data/ext/sources/examples/parakeet-quantize/CMakeLists.txt +7 -0
  53. data/ext/sources/examples/parakeet-quantize/parakeet-quantize.cpp +230 -0
  54. data/ext/sources/examples/server/server.cpp +213 -163
  55. data/ext/sources/ggml/CMakeLists.txt +29 -15
  56. data/ext/sources/ggml/cmake/FindNCCL.cmake +36 -0
  57. data/ext/sources/ggml/cmake/ggml-config.cmake.in +12 -2
  58. data/ext/sources/ggml/include/ggml-alloc.h +1 -0
  59. data/ext/sources/ggml/include/ggml-backend.h +73 -11
  60. data/ext/sources/ggml/include/ggml-cann.h +1 -1
  61. data/ext/sources/ggml/include/ggml-cpu.h +5 -0
  62. data/ext/sources/ggml/include/ggml-cuda.h +3 -0
  63. data/ext/sources/ggml/include/ggml-openvino.h +37 -0
  64. data/ext/sources/ggml/include/ggml-opt.h +1 -1
  65. data/ext/sources/ggml/include/ggml-rpc.h +8 -3
  66. data/ext/sources/ggml/include/ggml-virtgpu.h +14 -0
  67. data/ext/sources/ggml/include/ggml.h +155 -16
  68. data/ext/sources/ggml/include/gguf.h +10 -2
  69. data/ext/sources/ggml/src/CMakeLists.txt +25 -5
  70. data/ext/sources/ggml/src/ggml-alloc.c +9 -10
  71. data/ext/sources/ggml/src/ggml-backend-dl.cpp +48 -0
  72. data/ext/sources/ggml/src/ggml-backend-dl.h +45 -0
  73. data/ext/sources/ggml/src/ggml-backend-impl.h +22 -2
  74. data/ext/sources/ggml/src/ggml-backend-meta.cpp +2263 -0
  75. data/ext/sources/ggml/src/ggml-backend-reg.cpp +40 -86
  76. data/ext/sources/ggml/src/ggml-backend.cpp +114 -10
  77. data/ext/sources/ggml/src/ggml-blas/CMakeLists.txt +1 -1
  78. data/ext/sources/ggml/src/ggml-blas/ggml-blas.cpp +10 -2
  79. data/ext/sources/ggml/src/ggml-cann/acl_tensor.cpp +1 -1
  80. data/ext/sources/ggml/src/ggml-cann/acl_tensor.h +1 -1
  81. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.cpp +1016 -442
  82. data/ext/sources/ggml/src/ggml-cann/aclnn_ops.h +111 -85
  83. data/ext/sources/ggml/src/ggml-cann/common.h +23 -14
  84. data/ext/sources/ggml/src/ggml-cann/ggml-cann.cpp +255 -92
  85. data/ext/sources/ggml/src/ggml-common.h +22 -0
  86. data/ext/sources/ggml/src/ggml-cpu/CMakeLists.txt +68 -34
  87. data/ext/sources/ggml/src/ggml-cpu/amx/amx.cpp +44 -19
  88. data/ext/sources/ggml/src/ggml-cpu/amx/common.h +34 -10
  89. data/ext/sources/ggml/src/ggml-cpu/amx/mmq.cpp +101 -101
  90. data/ext/sources/ggml/src/ggml-cpu/arch/arm/quants.c +194 -1
  91. data/ext/sources/ggml/src/ggml-cpu/arch/arm/repack.cpp +2874 -613
  92. data/ext/sources/ggml/src/ggml-cpu/arch/loongarch/quants.c +151 -1
  93. data/ext/sources/ggml/src/ggml-cpu/arch/powerpc/quants.c +0 -1
  94. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/quants.c +5480 -840
  95. data/ext/sources/ggml/src/ggml-cpu/arch/riscv/repack.cpp +1361 -0
  96. data/ext/sources/ggml/src/ggml-cpu/arch/s390/quants.c +8 -11
  97. data/ext/sources/ggml/src/ggml-cpu/arch/wasm/quants.c +72 -1
  98. data/ext/sources/ggml/src/ggml-cpu/arch/x86/quants.c +186 -36
  99. data/ext/sources/ggml/src/ggml-cpu/arch/x86/repack.cpp +119 -19
  100. data/ext/sources/ggml/src/ggml-cpu/arch-fallback.h +112 -26
  101. data/ext/sources/ggml/src/ggml-cpu/binary-ops.cpp +2 -6
  102. data/ext/sources/ggml/src/ggml-cpu/cmake/FindSMTIME.cmake +32 -0
  103. data/ext/sources/ggml/src/ggml-cpu/common.h +8 -0
  104. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu-impl.h +13 -0
  105. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.c +153 -16
  106. data/ext/sources/ggml/src/ggml-cpu/ggml-cpu.cpp +17 -0
  107. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kernels.cpp +21 -20
  108. data/ext/sources/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +976 -251
  109. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm.cpp +671 -266
  110. data/ext/sources/ggml/src/ggml-cpu/ops.cpp +1277 -263
  111. data/ext/sources/ggml/src/ggml-cpu/ops.h +4 -0
  112. data/ext/sources/ggml/src/ggml-cpu/quants.c +95 -0
  113. data/ext/sources/ggml/src/ggml-cpu/quants.h +6 -0
  114. data/ext/sources/ggml/src/ggml-cpu/repack.cpp +2893 -679
  115. data/ext/sources/ggml/src/ggml-cpu/repack.h +119 -8
  116. data/ext/sources/ggml/src/ggml-cpu/simd-gemm.h +226 -0
  117. data/ext/sources/ggml/src/ggml-cpu/simd-mappings.h +114 -19
  118. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.cpp +1402 -687
  119. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime.h +8 -0
  120. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp +597 -2766
  121. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime2_kernels.cpp +5768 -0
  122. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.cpp +320 -0
  123. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_env.h +55 -0
  124. data/ext/sources/ggml/src/ggml-cpu/spacemit/ime_kernels.h +182 -19
  125. data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.cpp +1795 -0
  126. data/ext/sources/ggml/src/ggml-cpu/spacemit/repack.h +14 -0
  127. data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.cpp +3178 -0
  128. data/ext/sources/ggml/src/ggml-cpu/spacemit/rvv_kernels.h +95 -0
  129. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_barrier.h +34 -0
  130. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.cpp +760 -0
  131. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_mem_pool.h +32 -0
  132. data/ext/sources/ggml/src/ggml-cpu/spacemit/spine_tcm.h +409 -0
  133. data/ext/sources/ggml/src/ggml-cpu/unary-ops.cpp +1 -1
  134. data/ext/sources/ggml/src/ggml-cpu/vec.cpp +54 -53
  135. data/ext/sources/ggml/src/ggml-cpu/vec.h +225 -240
  136. data/ext/sources/ggml/src/ggml-cuda/CMakeLists.txt +18 -8
  137. data/ext/sources/ggml/src/ggml-cuda/allreduce.cu +971 -0
  138. data/ext/sources/ggml/src/ggml-cuda/allreduce.cuh +29 -0
  139. data/ext/sources/ggml/src/ggml-cuda/argsort.cu +73 -28
  140. data/ext/sources/ggml/src/ggml-cuda/binbcast.cu +69 -41
  141. data/ext/sources/ggml/src/ggml-cuda/binbcast.cuh +1 -0
  142. data/ext/sources/ggml/src/ggml-cuda/common.cuh +359 -29
  143. data/ext/sources/ggml/src/ggml-cuda/concat.cu +120 -114
  144. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cu +45 -21
  145. data/ext/sources/ggml/src/ggml-cuda/conv2d-transpose.cuh +1 -0
  146. data/ext/sources/ggml/src/ggml-cuda/convert.cu +94 -27
  147. data/ext/sources/ggml/src/ggml-cuda/convert.cuh +10 -0
  148. data/ext/sources/ggml/src/ggml-cuda/cpy.cu +20 -9
  149. data/ext/sources/ggml/src/ggml-cuda/dequantize.cuh +22 -0
  150. data/ext/sources/ggml/src/ggml-cuda/fattn-common.cuh +333 -85
  151. data/ext/sources/ggml/src/ggml-cuda/fattn-mma-f16.cuh +632 -190
  152. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cu +12 -0
  153. data/ext/sources/ggml/src/ggml-cuda/fattn-tile.cuh +162 -49
  154. data/ext/sources/ggml/src/ggml-cuda/fattn-vec.cuh +43 -18
  155. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cu +44 -14
  156. data/ext/sources/ggml/src/ggml-cuda/fattn-wmma-f16.cuh +1 -1
  157. data/ext/sources/ggml/src/ggml-cuda/fattn.cu +241 -23
  158. data/ext/sources/ggml/src/ggml-cuda/fattn.cuh +2 -0
  159. data/ext/sources/ggml/src/ggml-cuda/fwht.cu +101 -0
  160. data/ext/sources/ggml/src/ggml-cuda/fwht.cuh +4 -0
  161. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cu +312 -0
  162. data/ext/sources/ggml/src/ggml-cuda/gated_delta_net.cuh +4 -0
  163. data/ext/sources/ggml/src/ggml-cuda/getrows.cu +34 -12
  164. data/ext/sources/ggml/src/ggml-cuda/ggml-cuda.cu +1454 -599
  165. data/ext/sources/ggml/src/ggml-cuda/im2col.cu +32 -29
  166. data/ext/sources/ggml/src/ggml-cuda/mean.cu +13 -10
  167. data/ext/sources/ggml/src/ggml-cuda/mma.cuh +397 -183
  168. data/ext/sources/ggml/src/ggml-cuda/mmf.cu +30 -10
  169. data/ext/sources/ggml/src/ggml-cuda/mmf.cuh +161 -88
  170. data/ext/sources/ggml/src/ggml-cuda/mmq.cu +18 -12
  171. data/ext/sources/ggml/src/ggml-cuda/mmq.cuh +522 -431
  172. data/ext/sources/ggml/src/ggml-cuda/mmvf.cu +139 -72
  173. data/ext/sources/ggml/src/ggml-cuda/mmvf.cuh +2 -0
  174. data/ext/sources/ggml/src/ggml-cuda/mmvq.cu +608 -88
  175. data/ext/sources/ggml/src/ggml-cuda/mmvq.cuh +6 -0
  176. data/ext/sources/ggml/src/ggml-cuda/norm.cu +47 -79
  177. data/ext/sources/ggml/src/ggml-cuda/out-prod.cu +23 -7
  178. data/ext/sources/ggml/src/ggml-cuda/pad.cu +13 -10
  179. data/ext/sources/ggml/src/ggml-cuda/quantize.cu +134 -27
  180. data/ext/sources/ggml/src/ggml-cuda/quantize.cuh +1 -1
  181. data/ext/sources/ggml/src/ggml-cuda/reduce_rows.cuh +7 -17
  182. data/ext/sources/ggml/src/ggml-cuda/rope.cu +244 -137
  183. data/ext/sources/ggml/src/ggml-cuda/scale.cu +4 -1
  184. data/ext/sources/ggml/src/ggml-cuda/set-rows.cu +14 -6
  185. data/ext/sources/ggml/src/ggml-cuda/snake.cu +72 -0
  186. data/ext/sources/ggml/src/ggml-cuda/snake.cuh +8 -0
  187. data/ext/sources/ggml/src/ggml-cuda/softcap.cu +4 -1
  188. data/ext/sources/ggml/src/ggml-cuda/softmax.cu +8 -83
  189. data/ext/sources/ggml/src/ggml-cuda/solve_tri.cu +1 -1
  190. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cu +96 -40
  191. data/ext/sources/ggml/src/ggml-cuda/ssm-conv.cuh +1 -1
  192. data/ext/sources/ggml/src/ggml-cuda/ssm-scan.cu +40 -18
  193. data/ext/sources/ggml/src/ggml-cuda/sumrows.cu +8 -4
  194. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu +1 -0
  195. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu +6 -0
  196. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +2 -0
  197. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu +2 -0
  198. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu +1 -0
  199. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu +6 -0
  200. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu +2 -0
  201. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +2 -0
  202. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu +1 -0
  203. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu +2 -0
  204. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +2 -0
  205. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu +2 -0
  206. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +2 -0
  207. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq192-dv128.cu +5 -0
  208. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu +5 -0
  209. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu +5 -0
  210. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-bf16.cu +7 -0
  211. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-f16.cu +7 -0
  212. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_0.cu +7 -0
  213. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q4_1.cu +7 -0
  214. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_0.cu +7 -0
  215. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q5_1.cu +7 -0
  216. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-bf16-q8_0.cu +7 -0
  217. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-f16-bf16.cu +7 -0
  218. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_0-bf16.cu +7 -0
  219. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q4_1-bf16.cu +7 -0
  220. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_0-bf16.cu +7 -0
  221. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q5_1-bf16.cu +7 -0
  222. data/ext/sources/ggml/src/ggml-cuda/template-instances/fattn-vec-instance-q8_0-bf16.cu +7 -0
  223. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu +5 -0
  224. data/ext/sources/ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu +5 -0
  225. data/ext/sources/ggml/src/ggml-cuda/top-k.cu +5 -5
  226. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cu +202 -135
  227. data/ext/sources/ggml/src/ggml-cuda/topk-moe.cuh +20 -14
  228. data/ext/sources/ggml/src/ggml-cuda/unary.cu +86 -2
  229. data/ext/sources/ggml/src/ggml-cuda/unary.cuh +4 -0
  230. data/ext/sources/ggml/src/ggml-cuda/vecdotq.cuh +111 -17
  231. data/ext/sources/ggml/src/ggml-cuda/vendors/cuda.h +7 -2
  232. data/ext/sources/ggml/src/ggml-cuda/vendors/hip.h +30 -2
  233. data/ext/sources/ggml/src/ggml-cuda/vendors/musa.h +3 -0
  234. data/ext/sources/ggml/src/ggml-hexagon/CMakeLists.txt +84 -46
  235. data/ext/sources/ggml/src/ggml-hexagon/ggml-hexagon.cpp +1612 -753
  236. data/ext/sources/ggml/src/ggml-hexagon/htp/CMakeLists.txt +51 -11
  237. data/ext/sources/ggml/src/ggml-hexagon/htp/act-ops.c +361 -261
  238. data/ext/sources/ggml/src/ggml-hexagon/htp/argsort-ops.c +294 -0
  239. data/ext/sources/ggml/src/ggml-hexagon/htp/binary-ops.c +753 -241
  240. data/ext/sources/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake +5 -5
  241. data/ext/sources/ggml/src/ggml-hexagon/htp/concat-ops.c +277 -0
  242. data/ext/sources/ggml/src/ggml-hexagon/htp/cpy-ops.c +295 -0
  243. data/ext/sources/ggml/src/ggml-hexagon/htp/cumsum-ops.c +270 -0
  244. data/ext/sources/ggml/src/ggml-hexagon/htp/diag-ops.c +216 -0
  245. data/ext/sources/ggml/src/ggml-hexagon/htp/fill-ops.c +123 -0
  246. data/ext/sources/ggml/src/ggml-hexagon/htp/flash-attn-ops.c +471 -296
  247. data/ext/sources/ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c +1148 -0
  248. data/ext/sources/ggml/src/ggml-hexagon/htp/get-rows-ops.c +159 -53
  249. data/ext/sources/ggml/src/ggml-hexagon/htp/{htp-dma.c → hex-dma.c} +3 -3
  250. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dma.h +372 -0
  251. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-dump.h +86 -0
  252. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-fastdiv.h +37 -0
  253. data/ext/sources/ggml/src/ggml-hexagon/htp/hex-utils.h +137 -0
  254. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-flash-attn-ops.c +1878 -0
  255. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-matmul-ops.c +2066 -0
  256. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-ops.c +6 -0
  257. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-ops.h +88 -0
  258. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-profile.h +34 -0
  259. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.c +158 -0
  260. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-queue.h +134 -0
  261. data/ext/sources/ggml/src/ggml-hexagon/htp/hmx-utils.h +200 -0
  262. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ctx.h +97 -14
  263. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-ops.h +163 -67
  264. data/ext/sources/ggml/src/ggml-hexagon/htp/htp_iface.idl +9 -3
  265. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-arith.h +443 -0
  266. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-base.h +308 -0
  267. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-copy.h +262 -0
  268. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-div.h +291 -0
  269. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-dump.h +129 -0
  270. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.h +216 -0
  271. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-flash-attn.h +47 -0
  272. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-floor.h +100 -0
  273. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.h +210 -0
  274. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-log.h +65 -0
  275. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-pow.h +42 -0
  276. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-reduce.h +296 -0
  277. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-repl.h +74 -0
  278. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-scale.h +133 -0
  279. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.h +142 -0
  280. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sin-cos.h +90 -0
  281. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sqrt.h +126 -0
  282. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-types.h +36 -0
  283. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.h +18 -1348
  284. data/ext/sources/ggml/src/ggml-hexagon/htp/main.c +547 -635
  285. data/ext/sources/ggml/src/ggml-hexagon/htp/matmul-ops.c +3556 -1101
  286. data/ext/sources/ggml/src/ggml-hexagon/htp/pad-ops.c +547 -0
  287. data/ext/sources/ggml/src/ggml-hexagon/htp/repeat-ops.c +148 -0
  288. data/ext/sources/ggml/src/ggml-hexagon/htp/rope-ops.c +475 -269
  289. data/ext/sources/ggml/src/ggml-hexagon/htp/set-rows-ops.c +94 -72
  290. data/ext/sources/ggml/src/ggml-hexagon/htp/softmax-ops.c +222 -217
  291. data/ext/sources/ggml/src/ggml-hexagon/htp/solve-tri-ops.c +267 -0
  292. data/ext/sources/ggml/src/ggml-hexagon/htp/ssm-conv.c +432 -0
  293. data/ext/sources/ggml/src/ggml-hexagon/htp/sum-rows-ops.c +128 -0
  294. data/ext/sources/ggml/src/ggml-hexagon/htp/unary-ops.c +886 -117
  295. data/ext/sources/ggml/src/ggml-hexagon/htp/vtcm-utils.h +16 -0
  296. data/ext/sources/ggml/src/ggml-hexagon/htp/worker-pool.c +1 -5
  297. data/ext/sources/ggml/src/ggml-hexagon/htp-drv.cpp +418 -0
  298. data/ext/sources/ggml/src/ggml-hexagon/htp-drv.h +121 -0
  299. data/ext/sources/ggml/src/ggml-hexagon/htp-opnode.h +272 -0
  300. data/ext/sources/ggml/src/ggml-hexagon/libdl.h +79 -0
  301. data/ext/sources/ggml/src/ggml-hexagon/libggml-htp.inf +40 -0
  302. data/ext/sources/ggml/src/ggml-hip/CMakeLists.txt +28 -9
  303. data/ext/sources/ggml/src/ggml-impl.h +68 -1
  304. data/ext/sources/ggml/src/ggml-metal/CMakeLists.txt +10 -10
  305. data/ext/sources/ggml/src/ggml-metal/ggml-metal-common.cpp +13 -2
  306. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.h +8 -0
  307. data/ext/sources/ggml/src/ggml-metal/ggml-metal-context.m +147 -17
  308. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.cpp +409 -83
  309. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.h +54 -5
  310. data/ext/sources/ggml/src/ggml-metal/ggml-metal-device.m +254 -52
  311. data/ext/sources/ggml/src/ggml-metal/ggml-metal-impl.h +254 -23
  312. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.cpp +756 -285
  313. data/ext/sources/ggml/src/ggml-metal/ggml-metal-ops.h +7 -4
  314. data/ext/sources/ggml/src/ggml-metal/ggml-metal.cpp +359 -133
  315. data/ext/sources/ggml/src/ggml-metal/ggml-metal.metal +1867 -1123
  316. data/ext/sources/ggml/src/ggml-musa/CMakeLists.txt +5 -6
  317. data/ext/sources/ggml/src/ggml-opencl/CMakeLists.txt +71 -4
  318. data/ext/sources/ggml/src/ggml-opencl/ggml-opencl.cpp +14127 -5314
  319. data/ext/sources/ggml/src/ggml-opencl/kernels/concat.cl +97 -88
  320. data/ext/sources/ggml/src/ggml-opencl/kernels/cpy.cl +104 -0
  321. data/ext/sources/ggml/src/ggml-opencl/kernels/cumsum.cl +139 -0
  322. data/ext/sources/ggml/src/ggml-opencl/kernels/cvt.cl +1978 -67
  323. data/ext/sources/ggml/src/ggml-opencl/kernels/diag.cl +27 -0
  324. data/ext/sources/ggml/src/ggml-opencl/kernels/exp.cl +125 -0
  325. data/ext/sources/ggml/src/ggml-opencl/kernels/expm1.cl +87 -56
  326. data/ext/sources/ggml/src/ggml-opencl/kernels/gated_delta_net.cl +249 -0
  327. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32_ns.cl +306 -0
  328. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_0_f32_ns.cl +256 -0
  329. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_1_f32_ns.cl +258 -0
  330. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q4_k_f32_ns.cl +283 -0
  331. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_0_f32_ns.cl +260 -0
  332. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_1_f32_ns.cl +262 -0
  333. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q5_k_f32_ns.cl +288 -0
  334. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_moe_q6_k_f32_ns.cl +267 -0
  335. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq4_nl_f32.cl +150 -0
  336. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mat_Ab_Bi_8x4.cl → gemm_noshuffle_q4_0_f32.cl} +1 -1
  337. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_1_f32.cl +132 -0
  338. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q4_k_f32.cl +172 -0
  339. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_0_f32.cl +131 -0
  340. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_1_f32.cl +134 -0
  341. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl +176 -0
  342. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q6_k_f32.cl +140 -0
  343. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_noshuffle_q8_0_f32.cl +129 -0
  344. data/ext/sources/ggml/src/ggml-opencl/kernels/gemm_xmem_f16_f32_os8.cl +233 -0
  345. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32_ns.cl +165 -0
  346. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_0_f32_ns.cl +120 -0
  347. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_1_f32_ns.cl +123 -0
  348. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q4_k_f32_ns.cl +155 -0
  349. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_0_f32_ns.cl +123 -0
  350. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_1_f32_ns.cl +125 -0
  351. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q5_k_f32_ns.cl +160 -0
  352. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_moe_q6_k_f32_ns.cl +141 -0
  353. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq4_nl_f32.cl +302 -0
  354. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle_general.cl → gemv_noshuffle_q4_0_f32.cl} +5 -5
  355. data/ext/sources/ggml/src/ggml-opencl/kernels/{gemv_noshuffle.cl → gemv_noshuffle_q4_0_f32_spec.cl} +5 -5
  356. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_1_f32.cl +283 -0
  357. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q4_k_f32.cl +318 -0
  358. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_0_f32.cl +291 -0
  359. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_1_f32.cl +294 -0
  360. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl +326 -0
  361. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q6_k_f32.cl +293 -0
  362. data/ext/sources/ggml/src/ggml-opencl/kernels/gemv_noshuffle_q8_0_f32.cl +195 -0
  363. data/ext/sources/ggml/src/ggml-opencl/kernels/get_rows.cl +15 -9
  364. data/ext/sources/ggml/src/ggml-opencl/kernels/l2_norm.cl +71 -0
  365. data/ext/sources/ggml/src/ggml-opencl/kernels/mean.cl +114 -13
  366. data/ext/sources/ggml/src/ggml-opencl/kernels/moe_reorder_b.cl +30 -0
  367. data/ext/sources/ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl +82 -0
  368. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_iq4_nl_f32_l4_lm.cl +171 -0
  369. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_0_f32_l4_lm.cl +163 -0
  370. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_1_f32_l4_lm.cl +165 -0
  371. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q4_k_f32_l4_lm.cl +179 -0
  372. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_0_f32_l4_lm.cl +173 -0
  373. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_1_f32_l4_lm.cl +175 -0
  374. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q5_k_f32_l4_lm.cl +192 -0
  375. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl +158 -0
  376. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32.cl +164 -0
  377. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_iq4_nl_f32_flat.cl +202 -0
  378. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32.cl +219 -0
  379. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_1_f32_flat.cl +229 -0
  380. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl +180 -0
  381. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32_flat.cl +196 -0
  382. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32.cl +241 -0
  383. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_0_f32_flat.cl +243 -0
  384. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32.cl +243 -0
  385. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_1_f32_flat.cl +247 -0
  386. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32.cl +187 -0
  387. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q5_k_f32_flat.cl +203 -0
  388. data/ext/sources/ggml/src/ggml-opencl/kernels/{mul_mv_q6_k.cl → mul_mv_q6_k_f32.cl} +4 -0
  389. data/ext/sources/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl +178 -0
  390. data/ext/sources/ggml/src/ggml-opencl/kernels/neg.cl +125 -0
  391. data/ext/sources/ggml/src/ggml-opencl/kernels/repeat.cl +31 -32
  392. data/ext/sources/ggml/src/ggml-opencl/kernels/scale.cl +14 -4
  393. data/ext/sources/ggml/src/ggml-opencl/kernels/softplus.cl +88 -60
  394. data/ext/sources/ggml/src/ggml-opencl/kernels/solve_tri.cl +51 -0
  395. data/ext/sources/ggml/src/ggml-opencl/kernels/sum_rows.cl +114 -13
  396. data/ext/sources/ggml/src/ggml-opencl/kernels/tanh.cl +94 -48
  397. data/ext/sources/ggml/src/ggml-opencl/kernels/transpose.cl +26 -0
  398. data/ext/sources/ggml/src/ggml-opencl/kernels/tri.cl +32 -0
  399. data/ext/sources/ggml/src/ggml-openvino/.clang-format +154 -0
  400. data/ext/sources/ggml/src/ggml-openvino/CMakeLists.txt +22 -0
  401. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.cpp +985 -0
  402. data/ext/sources/ggml/src/ggml-openvino/ggml-decoder.h +294 -0
  403. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.cpp +380 -0
  404. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino-extra.h +182 -0
  405. data/ext/sources/ggml/src/ggml-openvino/ggml-openvino.cpp +1132 -0
  406. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.cpp +956 -0
  407. data/ext/sources/ggml/src/ggml-openvino/ggml-quants.h +153 -0
  408. data/ext/sources/ggml/src/ggml-openvino/openvino/decoder.h +74 -0
  409. data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.cpp +27 -0
  410. data/ext/sources/ggml/src/ggml-openvino/openvino/frontend.h +23 -0
  411. data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.cpp +17 -0
  412. data/ext/sources/ggml/src/ggml-openvino/openvino/input_model.h +29 -0
  413. data/ext/sources/ggml/src/ggml-openvino/openvino/node_context.h +112 -0
  414. data/ext/sources/ggml/src/ggml-openvino/openvino/op/cont.cpp +48 -0
  415. data/ext/sources/ggml/src/ggml-openvino/openvino/op/cpy.cpp +21 -0
  416. data/ext/sources/ggml/src/ggml-openvino/openvino/op/flash_attn_ext.cpp +90 -0
  417. data/ext/sources/ggml/src/ggml-openvino/openvino/op/get_rows.cpp +69 -0
  418. data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +61 -0
  419. data/ext/sources/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +62 -0
  420. data/ext/sources/ggml/src/ggml-openvino/openvino/op/mulmat.cpp +90 -0
  421. data/ext/sources/ggml/src/ggml-openvino/openvino/op/permute.cpp +102 -0
  422. data/ext/sources/ggml/src/ggml-openvino/openvino/op/reshape.cpp +83 -0
  423. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rms_norm.cpp +46 -0
  424. data/ext/sources/ggml/src/ggml-openvino/openvino/op/rope.cpp +149 -0
  425. data/ext/sources/ggml/src/ggml-openvino/openvino/op/scale.cpp +41 -0
  426. data/ext/sources/ggml/src/ggml-openvino/openvino/op/set_rows.cpp +76 -0
  427. data/ext/sources/ggml/src/ggml-openvino/openvino/op/softmax.cpp +89 -0
  428. data/ext/sources/ggml/src/ggml-openvino/openvino/op/transpose.cpp +23 -0
  429. data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_gelu.cpp +25 -0
  430. data/ext/sources/ggml/src/ggml-openvino/openvino/op/unary_silu.cpp +27 -0
  431. data/ext/sources/ggml/src/ggml-openvino/openvino/op/view.cpp +53 -0
  432. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.cpp +47 -0
  433. data/ext/sources/ggml/src/ggml-openvino/openvino/op_table.h +40 -0
  434. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.cpp +60 -0
  435. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/fuse_to_sdpa.h +17 -0
  436. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.h +29 -0
  437. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp +58 -0
  438. data/ext/sources/ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.h +17 -0
  439. data/ext/sources/ggml/src/ggml-openvino/openvino/rt_info/weightless_caching_attributes.hpp +41 -0
  440. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.cpp +317 -0
  441. data/ext/sources/ggml/src/ggml-openvino/openvino/translate_session.h +28 -0
  442. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.cpp +257 -0
  443. data/ext/sources/ggml/src/ggml-openvino/openvino/utils.h +86 -0
  444. data/ext/sources/ggml/src/ggml-openvino/utils.cpp +880 -0
  445. data/ext/sources/ggml/src/ggml-openvino/utils.h +143 -0
  446. data/ext/sources/ggml/src/ggml-opt.cpp +1 -0
  447. data/ext/sources/ggml/src/ggml-quants.c +385 -119
  448. data/ext/sources/ggml/src/ggml-quants.h +6 -0
  449. data/ext/sources/ggml/src/ggml-rpc/CMakeLists.txt +24 -0
  450. data/ext/sources/ggml/src/ggml-rpc/ggml-rpc.cpp +167 -311
  451. data/ext/sources/ggml/src/ggml-rpc/transport.cpp +683 -0
  452. data/ext/sources/ggml/src/ggml-rpc/transport.h +34 -0
  453. data/ext/sources/ggml/src/ggml-sycl/CMakeLists.txt +64 -91
  454. data/ext/sources/ggml/src/ggml-sycl/add-id.cpp +5 -1
  455. data/ext/sources/ggml/src/ggml-sycl/backend.hpp +4 -1
  456. data/ext/sources/ggml/src/ggml-sycl/binbcast.cpp +21 -20
  457. data/ext/sources/ggml/src/ggml-sycl/common.cpp +74 -2
  458. data/ext/sources/ggml/src/ggml-sycl/common.hpp +356 -11
  459. data/ext/sources/ggml/src/ggml-sycl/convert.cpp +184 -14
  460. data/ext/sources/ggml/src/ggml-sycl/convert.hpp +31 -1
  461. data/ext/sources/ggml/src/ggml-sycl/count-equal.cpp +1 -1
  462. data/ext/sources/ggml/src/ggml-sycl/cumsum.cpp +148 -0
  463. data/ext/sources/ggml/src/ggml-sycl/cumsum.hpp +5 -0
  464. data/ext/sources/ggml/src/ggml-sycl/dequantize.hpp +663 -0
  465. data/ext/sources/ggml/src/ggml-sycl/diag.cpp +67 -0
  466. data/ext/sources/ggml/src/ggml-sycl/diag.hpp +5 -0
  467. data/ext/sources/ggml/src/ggml-sycl/dmmv.cpp +586 -6
  468. data/ext/sources/ggml/src/ggml-sycl/dpct/helper.hpp +791 -47
  469. data/ext/sources/ggml/src/ggml-sycl/element_wise.cpp +77 -156
  470. data/ext/sources/ggml/src/ggml-sycl/element_wise.hpp +2 -2
  471. data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.cpp +56 -0
  472. data/ext/sources/ggml/src/ggml-sycl/fattn-buffers.hpp +63 -0
  473. data/ext/sources/ggml/src/ggml-sycl/fattn-common.hpp +1181 -0
  474. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.cpp +59 -0
  475. data/ext/sources/ggml/src/ggml-sycl/fattn-tile.hpp +1246 -0
  476. data/ext/sources/ggml/src/ggml-sycl/fattn-vec.hpp +674 -0
  477. data/ext/sources/ggml/src/ggml-sycl/fattn.cpp +227 -0
  478. data/ext/sources/ggml/src/ggml-sycl/fattn.hpp +22 -0
  479. data/ext/sources/ggml/src/ggml-sycl/fill.cpp +55 -0
  480. data/ext/sources/ggml/src/ggml-sycl/fill.hpp +5 -0
  481. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.cpp +347 -0
  482. data/ext/sources/ggml/src/ggml-sycl/gated_delta_net.hpp +9 -0
  483. data/ext/sources/ggml/src/ggml-sycl/gemm.hpp +3 -0
  484. data/ext/sources/ggml/src/ggml-sycl/getrows.cpp +79 -3
  485. data/ext/sources/ggml/src/ggml-sycl/ggml-sycl.cpp +1134 -236
  486. data/ext/sources/ggml/src/ggml-sycl/im2col.cpp +353 -89
  487. data/ext/sources/ggml/src/ggml-sycl/im2col.hpp +5 -3
  488. data/ext/sources/ggml/src/ggml-sycl/mmvq.cpp +1344 -26
  489. data/ext/sources/ggml/src/ggml-sycl/mmvq.hpp +16 -0
  490. data/ext/sources/ggml/src/ggml-sycl/norm.cpp +65 -66
  491. data/ext/sources/ggml/src/ggml-sycl/outprod.cpp +3 -3
  492. data/ext/sources/ggml/src/ggml-sycl/pad.cpp +27 -27
  493. data/ext/sources/ggml/src/ggml-sycl/presets.hpp +3 -0
  494. data/ext/sources/ggml/src/ggml-sycl/quants.hpp +72 -1
  495. data/ext/sources/ggml/src/ggml-sycl/rope.cpp +450 -287
  496. data/ext/sources/ggml/src/ggml-sycl/rope.hpp +6 -0
  497. data/ext/sources/ggml/src/ggml-sycl/set_rows.cpp +7 -1
  498. data/ext/sources/ggml/src/ggml-sycl/softmax.cpp +6 -6
  499. data/ext/sources/ggml/src/ggml-sycl/solve_tri.cpp +172 -0
  500. data/ext/sources/ggml/src/ggml-sycl/solve_tri.hpp +8 -0
  501. data/ext/sources/ggml/src/ggml-sycl/ssm_conv.cpp +6 -1
  502. data/ext/sources/ggml/src/ggml-sycl/ssm_scan.cpp +156 -0
  503. data/ext/sources/ggml/src/ggml-sycl/ssm_scan.hpp +5 -0
  504. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.cpp +62 -10
  505. data/ext/sources/ggml/src/ggml-sycl/sycl_hw.hpp +18 -6
  506. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq112-dv112.cpp +5 -0
  507. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq128-dv128.cpp +5 -0
  508. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq256-dv256.cpp +5 -0
  509. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq40-dv40.cpp +5 -0
  510. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq512-dv512.cpp +6 -0
  511. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq576-dv512.cpp +5 -0
  512. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq64-dv64.cpp +5 -0
  513. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq72-dv72.cpp +5 -0
  514. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq80-dv80.cpp +5 -0
  515. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-tile-instance-dkq96-dv96.cpp +5 -0
  516. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-f16.cpp +8 -0
  517. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_0.cpp +8 -0
  518. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q4_1.cpp +8 -0
  519. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_0.cpp +8 -0
  520. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q5_1.cpp +8 -0
  521. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-f16-q8_0.cpp +8 -0
  522. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-f16.cpp +8 -0
  523. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_0.cpp +8 -0
  524. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q4_1.cpp +8 -0
  525. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_0.cpp +8 -0
  526. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q5_1.cpp +8 -0
  527. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_0-q8_0.cpp +8 -0
  528. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-f16.cpp +8 -0
  529. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_0.cpp +8 -0
  530. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q4_1.cpp +8 -0
  531. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_0.cpp +8 -0
  532. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q5_1.cpp +8 -0
  533. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q4_1-q8_0.cpp +8 -0
  534. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-f16.cpp +8 -0
  535. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_0.cpp +8 -0
  536. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q4_1.cpp +8 -0
  537. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_0.cpp +8 -0
  538. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q5_1.cpp +8 -0
  539. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_0-q8_0.cpp +8 -0
  540. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-f16.cpp +8 -0
  541. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_0.cpp +8 -0
  542. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q4_1.cpp +8 -0
  543. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_0.cpp +8 -0
  544. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q5_1.cpp +8 -0
  545. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q5_1-q8_0.cpp +8 -0
  546. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-f16.cpp +8 -0
  547. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_0.cpp +8 -0
  548. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q4_1.cpp +8 -0
  549. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_0.cpp +8 -0
  550. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q5_1.cpp +8 -0
  551. data/ext/sources/ggml/src/ggml-sycl/template-instances/fattn-vec-instance-q8_0-q8_0.cpp +8 -0
  552. data/ext/sources/ggml/src/ggml-sycl/type.hpp +112 -0
  553. data/ext/sources/ggml/src/ggml-sycl/upscale.cpp +410 -0
  554. data/ext/sources/ggml/src/ggml-sycl/upscale.hpp +9 -0
  555. data/ext/sources/ggml/src/ggml-sycl/vecdotq.hpp +228 -53
  556. data/ext/sources/ggml/src/ggml-sycl/wkv.cpp +1 -1
  557. data/ext/sources/ggml/src/ggml-virtgpu/CMakeLists.txt +70 -0
  558. data/ext/sources/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp +87 -0
  559. data/ext/sources/ggml/src/ggml-virtgpu/backend/CMakeLists.txt +21 -0
  560. data/ext/sources/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp +115 -0
  561. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-convert.h +13 -0
  562. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp +102 -0
  563. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp +105 -0
  564. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp +179 -0
  565. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp +148 -0
  566. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp +51 -0
  567. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h +73 -0
  568. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-dispatched.h +27 -0
  569. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h +32 -0
  570. data/ext/sources/ggml/src/ggml-virtgpu/backend/backend.cpp +144 -0
  571. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h +95 -0
  572. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h +94 -0
  573. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h +50 -0
  574. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h +378 -0
  575. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h +232 -0
  576. data/ext/sources/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h +58 -0
  577. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp +81 -0
  578. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp +123 -0
  579. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-device.cpp +160 -0
  580. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp +213 -0
  581. data/ext/sources/ggml/src/ggml-virtgpu/ggml-backend.cpp +71 -0
  582. data/ext/sources/ggml/src/ggml-virtgpu/ggml-remoting.h +71 -0
  583. data/ext/sources/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml +166 -0
  584. data/ext/sources/ggml/src/ggml-virtgpu/include/apir_hw.h +9 -0
  585. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-apir.h +15 -0
  586. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp +58 -0
  587. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp +110 -0
  588. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp +173 -0
  589. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp +192 -0
  590. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h +36 -0
  591. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h +53 -0
  592. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.cpp +99 -0
  593. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-shm.h +23 -0
  594. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.cpp +179 -0
  595. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu-utils.h +86 -0
  596. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.cpp +545 -0
  597. data/ext/sources/ggml/src/ggml-virtgpu/virtgpu.h +115 -0
  598. data/ext/sources/ggml/src/ggml-vulkan/CMakeLists.txt +12 -1
  599. data/ext/sources/ggml/src/ggml-vulkan/ggml-vulkan.cpp +3250 -940
  600. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +4 -0
  601. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp +16 -8
  602. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +6 -2
  603. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp +146 -13
  604. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +3 -1
  605. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +1 -1
  606. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +25 -1
  607. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +88 -0
  608. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl +643 -1
  609. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_nvfp4.comp +32 -0
  610. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q1_0.comp +29 -0
  611. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/diag.comp +0 -1
  612. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/dot_product_funcs.glsl +27 -0
  613. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/elu.comp +27 -0
  614. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/exp.comp +0 -1
  615. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/feature-tests/coopmat2_decode_vector.comp +7 -0
  616. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp +533 -180
  617. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl +113 -68
  618. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp +412 -222
  619. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +222 -83
  620. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_dequant.glsl +131 -0
  621. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mask_opt.comp +162 -0
  622. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_mmq_funcs.glsl +203 -0
  623. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp +9 -8
  624. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/fwht.comp +115 -0
  625. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/gated_delta_net.comp +189 -0
  626. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl +0 -1
  627. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.glsl +10 -1
  628. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.glsl +16 -6
  629. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +76 -54
  630. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/im2col_3d.comp +0 -1
  631. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp +12 -9
  632. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/log.comp +0 -1
  633. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +122 -27
  634. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl +20 -17
  635. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iface.glsl +6 -6
  636. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +1 -1
  637. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp +1 -1
  638. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp +1 -1
  639. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +1 -0
  640. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +88 -55
  641. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +22 -20
  642. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +51 -14
  643. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +159 -125
  644. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +3 -1
  645. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +5 -3
  646. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.glsl +8 -8
  647. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_shmem_types.glsl +24 -9
  648. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp +0 -1
  649. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp +2 -3
  650. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +39 -63
  651. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.glsl +0 -1
  652. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +7 -4
  653. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +7 -4
  654. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +7 -4
  655. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +13 -7
  656. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +7 -4
  657. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/sgn.comp +21 -0
  658. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/snake.comp +49 -0
  659. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/ssm_conv.comp +27 -11
  660. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/tri.comp +0 -1
  661. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +79 -2
  662. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +193 -149
  663. data/ext/sources/ggml/src/ggml-webgpu/CMakeLists.txt +5 -2
  664. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp +3221 -97
  665. data/ext/sources/ggml/src/ggml-webgpu/ggml-webgpu.cpp +3493 -1997
  666. data/ext/sources/ggml/src/ggml-webgpu/pre_wgsl.hpp +37 -7
  667. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/add_id.wgsl +64 -0
  668. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argmax.wgsl +72 -0
  669. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl +106 -0
  670. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/argsort_merge.wgsl +134 -0
  671. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary.wgsl +142 -0
  672. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/common_decls.tmpl +115 -141
  673. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/concat.wgsl +93 -0
  674. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/conv2d.wgsl +165 -0
  675. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{cpy.tmpl.wgsl → cpy.wgsl} +25 -44
  676. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/cumsum.wgsl +66 -0
  677. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl +198 -230
  678. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_quant_staging.tmpl +124 -0
  679. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl +397 -0
  680. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_blk.wgsl +101 -0
  681. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_reduce.wgsl +84 -0
  682. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl +619 -0
  683. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/gated_delta_net.wgsl +149 -0
  684. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{get_rows.tmpl.wgsl → get_rows.wgsl} +234 -335
  685. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.wgsl +155 -0
  686. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/im2col.wgsl +101 -0
  687. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl +871 -42
  688. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id.wgsl +195 -0
  689. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_gather.wgsl +52 -0
  690. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_id_vec.wgsl +154 -0
  691. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.wgsl +149 -0
  692. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{mul_mat_subgroup_matrix.tmpl.wgsl → mul_mat_subgroup_matrix.wgsl} +36 -138
  693. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.wgsl +151 -0
  694. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_acc.tmpl +1432 -0
  695. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec_q_acc.tmpl +303 -0
  696. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/pad.wgsl +86 -0
  697. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quant_inner_loops.tmpl +21 -0
  698. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/quantize_q8.wgsl +173 -0
  699. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl +67 -0
  700. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm_mul.wgsl +152 -0
  701. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{rope.tmpl.wgsl → rope.wgsl} +71 -142
  702. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/row_norm.wgsl +153 -0
  703. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{scale.tmpl.wgsl → scale.wgsl} +15 -40
  704. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set.wgsl +109 -0
  705. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.wgsl +39 -12
  706. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows_quant.wgsl +224 -0
  707. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/{soft_max.tmpl.wgsl → soft_max.wgsl} +106 -206
  708. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/solve_tri.wgsl +121 -0
  709. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_conv.wgsl +65 -0
  710. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/ssm_scan.wgsl +193 -0
  711. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/sum_rows.wgsl +55 -0
  712. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl +213 -0
  713. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/upscale.wgsl +240 -0
  714. data/ext/sources/ggml/src/ggml-zdnn/ggml-zdnn.cpp +24 -15
  715. data/ext/sources/ggml/src/ggml-zendnn/CMakeLists.txt +31 -32
  716. data/ext/sources/ggml/src/ggml-zendnn/ggml-zendnn.cpp +253 -16
  717. data/ext/sources/ggml/src/ggml.c +268 -52
  718. data/ext/sources/ggml/src/gguf.cpp +377 -47
  719. data/ext/sources/include/parakeet.h +342 -0
  720. data/ext/sources/include/whisper.h +10 -0
  721. data/ext/sources/media/matmul.png +0 -0
  722. data/ext/sources/src/CMakeLists.txt +23 -0
  723. data/ext/sources/src/parakeet-arch.h +188 -0
  724. data/ext/sources/src/parakeet.cpp +3838 -0
  725. data/ext/sources/src/whisper.cpp +62 -40
  726. data/extsources.rb +26 -10
  727. data/lib/whisper/log_settable.rb +36 -0
  728. data/lib/whisper/model/uri.rb +13 -1
  729. data/lib/whisper/output.rb +74 -0
  730. data/sig/whisper.rbs +445 -55
  731. data/test/helper.rb +2 -0
  732. data/test/jfk_reader/jfk_reader.c +50 -7
  733. data/test/test_callback.rb +1 -0
  734. data/test/test_context_params.rb +82 -0
  735. data/test/test_package.rb +6 -5
  736. data/test/test_parakeet.rb +28 -0
  737. data/test/test_parakeet_callback.rb +107 -0
  738. data/test/test_parakeet_context.rb +116 -0
  739. data/test/test_parakeet_context_params.rb +24 -0
  740. data/test/test_parakeet_model.rb +21 -0
  741. data/test/test_parakeet_params.rb +78 -0
  742. data/test/test_parakeet_segment.rb +42 -0
  743. data/test/test_parakeet_token.rb +73 -0
  744. data/test/test_params.rb +2 -0
  745. data/test/test_token.rb +11 -0
  746. data/test/test_vad_context.rb +58 -8
  747. data/test/test_vad_segment.rb +1 -1
  748. data/test/test_whisper.rb +44 -6
  749. data/whispercpp.gemspec +2 -2
  750. metadata +426 -280
  751. data/ext/sources/bindings/javascript/CMakeLists.txt +0 -41
  752. data/ext/sources/bindings/javascript/emscripten.cpp +0 -93
  753. data/ext/sources/bindings/javascript/libwhisper.worker.js +0 -1
  754. data/ext/sources/bindings/javascript/package.json +0 -26
  755. data/ext/sources/bindings/javascript/whisper.js +0 -19
  756. data/ext/sources/examples/addon.node/CMakeLists.txt +0 -31
  757. data/ext/sources/examples/addon.node/__test__/whisper.spec.js +0 -133
  758. data/ext/sources/examples/addon.node/addon.cpp +0 -557
  759. data/ext/sources/examples/addon.node/index.js +0 -59
  760. data/ext/sources/examples/addon.node/package.json +0 -16
  761. data/ext/sources/examples/addon.node/vad-example.js +0 -132
  762. data/ext/sources/examples/bench.wasm/CMakeLists.txt +0 -49
  763. data/ext/sources/examples/bench.wasm/emscripten.cpp +0 -87
  764. data/ext/sources/examples/bench.wasm/index-tmpl.html +0 -285
  765. data/ext/sources/examples/coi-serviceworker.js +0 -146
  766. data/ext/sources/examples/command/CMakeLists.txt +0 -10
  767. data/ext/sources/examples/command/command.cpp +0 -802
  768. data/ext/sources/examples/command/commands.txt +0 -9
  769. data/ext/sources/examples/command.wasm/CMakeLists.txt +0 -50
  770. data/ext/sources/examples/command.wasm/emscripten.cpp +0 -327
  771. data/ext/sources/examples/command.wasm/index-tmpl.html +0 -415
  772. data/ext/sources/examples/generate-karaoke.sh +0 -57
  773. data/ext/sources/examples/helpers.js +0 -191
  774. data/ext/sources/examples/livestream.sh +0 -112
  775. data/ext/sources/examples/lsp/CMakeLists.txt +0 -10
  776. data/ext/sources/examples/lsp/lsp.cpp +0 -471
  777. data/ext/sources/examples/lsp/whisper.vim +0 -362
  778. data/ext/sources/examples/python/test_whisper_processor.py +0 -7
  779. data/ext/sources/examples/python/whisper_processor.py +0 -54
  780. data/ext/sources/examples/server/bench.js +0 -29
  781. data/ext/sources/examples/server.py +0 -120
  782. data/ext/sources/examples/stream/CMakeLists.txt +0 -10
  783. data/ext/sources/examples/stream/stream.cpp +0 -437
  784. data/ext/sources/examples/stream.wasm/CMakeLists.txt +0 -49
  785. data/ext/sources/examples/stream.wasm/emscripten.cpp +0 -216
  786. data/ext/sources/examples/stream.wasm/index-tmpl.html +0 -491
  787. data/ext/sources/examples/sycl/CMakeLists.txt +0 -9
  788. data/ext/sources/examples/sycl/build.sh +0 -22
  789. data/ext/sources/examples/sycl/ls-sycl-device.cpp +0 -11
  790. data/ext/sources/examples/sycl/run-whisper.sh +0 -17
  791. data/ext/sources/examples/talk-llama/CMakeLists.txt +0 -47
  792. data/ext/sources/examples/talk-llama/eleven-labs.py +0 -80
  793. data/ext/sources/examples/talk-llama/llama-adapter.cpp +0 -494
  794. data/ext/sources/examples/talk-llama/llama-adapter.h +0 -88
  795. data/ext/sources/examples/talk-llama/llama-arch.cpp +0 -2559
  796. data/ext/sources/examples/talk-llama/llama-arch.h +0 -586
  797. data/ext/sources/examples/talk-llama/llama-batch.cpp +0 -917
  798. data/ext/sources/examples/talk-llama/llama-batch.h +0 -173
  799. data/ext/sources/examples/talk-llama/llama-chat.cpp +0 -876
  800. data/ext/sources/examples/talk-llama/llama-chat.h +0 -70
  801. data/ext/sources/examples/talk-llama/llama-context.cpp +0 -3645
  802. data/ext/sources/examples/talk-llama/llama-context.h +0 -360
  803. data/ext/sources/examples/talk-llama/llama-cparams.cpp +0 -5
  804. data/ext/sources/examples/talk-llama/llama-cparams.h +0 -42
  805. data/ext/sources/examples/talk-llama/llama-grammar.cpp +0 -1464
  806. data/ext/sources/examples/talk-llama/llama-grammar.h +0 -194
  807. data/ext/sources/examples/talk-llama/llama-graph.cpp +0 -2282
  808. data/ext/sources/examples/talk-llama/llama-graph.h +0 -910
  809. data/ext/sources/examples/talk-llama/llama-hparams.cpp +0 -241
  810. data/ext/sources/examples/talk-llama/llama-hparams.h +0 -284
  811. data/ext/sources/examples/talk-llama/llama-impl.cpp +0 -171
  812. data/ext/sources/examples/talk-llama/llama-impl.h +0 -63
  813. data/ext/sources/examples/talk-llama/llama-io.cpp +0 -15
  814. data/ext/sources/examples/talk-llama/llama-io.h +0 -35
  815. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.cpp +0 -328
  816. data/ext/sources/examples/talk-llama/llama-kv-cache-iswa.h +0 -137
  817. data/ext/sources/examples/talk-llama/llama-kv-cache.cpp +0 -2100
  818. data/ext/sources/examples/talk-llama/llama-kv-cache.h +0 -390
  819. data/ext/sources/examples/talk-llama/llama-kv-cells.h +0 -533
  820. data/ext/sources/examples/talk-llama/llama-memory-hybrid.cpp +0 -268
  821. data/ext/sources/examples/talk-llama/llama-memory-hybrid.h +0 -139
  822. data/ext/sources/examples/talk-llama/llama-memory-recurrent.cpp +0 -1167
  823. data/ext/sources/examples/talk-llama/llama-memory-recurrent.h +0 -182
  824. data/ext/sources/examples/talk-llama/llama-memory.cpp +0 -59
  825. data/ext/sources/examples/talk-llama/llama-memory.h +0 -122
  826. data/ext/sources/examples/talk-llama/llama-mmap.cpp +0 -735
  827. data/ext/sources/examples/talk-llama/llama-mmap.h +0 -73
  828. data/ext/sources/examples/talk-llama/llama-model-loader.cpp +0 -1247
  829. data/ext/sources/examples/talk-llama/llama-model-loader.h +0 -176
  830. data/ext/sources/examples/talk-llama/llama-model-saver.cpp +0 -285
  831. data/ext/sources/examples/talk-llama/llama-model-saver.h +0 -37
  832. data/ext/sources/examples/talk-llama/llama-model.cpp +0 -8338
  833. data/ext/sources/examples/talk-llama/llama-model.h +0 -544
  834. data/ext/sources/examples/talk-llama/llama-quant.cpp +0 -1072
  835. data/ext/sources/examples/talk-llama/llama-quant.h +0 -1
  836. data/ext/sources/examples/talk-llama/llama-sampling.cpp +0 -3771
  837. data/ext/sources/examples/talk-llama/llama-sampling.h +0 -44
  838. data/ext/sources/examples/talk-llama/llama-vocab.cpp +0 -3900
  839. data/ext/sources/examples/talk-llama/llama-vocab.h +0 -182
  840. data/ext/sources/examples/talk-llama/llama.cpp +0 -1140
  841. data/ext/sources/examples/talk-llama/llama.h +0 -1540
  842. data/ext/sources/examples/talk-llama/models/afmoe.cpp +0 -191
  843. data/ext/sources/examples/talk-llama/models/apertus.cpp +0 -125
  844. data/ext/sources/examples/talk-llama/models/arcee.cpp +0 -135
  845. data/ext/sources/examples/talk-llama/models/arctic.cpp +0 -138
  846. data/ext/sources/examples/talk-llama/models/arwkv7.cpp +0 -86
  847. data/ext/sources/examples/talk-llama/models/baichuan.cpp +0 -122
  848. data/ext/sources/examples/talk-llama/models/bailingmoe.cpp +0 -144
  849. data/ext/sources/examples/talk-llama/models/bailingmoe2.cpp +0 -135
  850. data/ext/sources/examples/talk-llama/models/bert.cpp +0 -178
  851. data/ext/sources/examples/talk-llama/models/bitnet.cpp +0 -160
  852. data/ext/sources/examples/talk-llama/models/bloom.cpp +0 -101
  853. data/ext/sources/examples/talk-llama/models/chameleon.cpp +0 -178
  854. data/ext/sources/examples/talk-llama/models/chatglm.cpp +0 -132
  855. data/ext/sources/examples/talk-llama/models/codeshell.cpp +0 -111
  856. data/ext/sources/examples/talk-llama/models/cogvlm.cpp +0 -102
  857. data/ext/sources/examples/talk-llama/models/cohere2-iswa.cpp +0 -134
  858. data/ext/sources/examples/talk-llama/models/command-r.cpp +0 -122
  859. data/ext/sources/examples/talk-llama/models/dbrx.cpp +0 -123
  860. data/ext/sources/examples/talk-llama/models/deci.cpp +0 -135
  861. data/ext/sources/examples/talk-llama/models/deepseek.cpp +0 -144
  862. data/ext/sources/examples/talk-llama/models/deepseek2.cpp +0 -259
  863. data/ext/sources/examples/talk-llama/models/dots1.cpp +0 -134
  864. data/ext/sources/examples/talk-llama/models/dream.cpp +0 -105
  865. data/ext/sources/examples/talk-llama/models/ernie4-5-moe.cpp +0 -150
  866. data/ext/sources/examples/talk-llama/models/ernie4-5.cpp +0 -110
  867. data/ext/sources/examples/talk-llama/models/exaone.cpp +0 -114
  868. data/ext/sources/examples/talk-llama/models/exaone4.cpp +0 -123
  869. data/ext/sources/examples/talk-llama/models/falcon-h1.cpp +0 -113
  870. data/ext/sources/examples/talk-llama/models/falcon.cpp +0 -120
  871. data/ext/sources/examples/talk-llama/models/gemma-embedding.cpp +0 -116
  872. data/ext/sources/examples/talk-llama/models/gemma.cpp +0 -112
  873. data/ext/sources/examples/talk-llama/models/gemma2-iswa.cpp +0 -128
  874. data/ext/sources/examples/talk-llama/models/gemma3.cpp +0 -155
  875. data/ext/sources/examples/talk-llama/models/gemma3n-iswa.cpp +0 -384
  876. data/ext/sources/examples/talk-llama/models/glm4-moe.cpp +0 -170
  877. data/ext/sources/examples/talk-llama/models/glm4.cpp +0 -150
  878. data/ext/sources/examples/talk-llama/models/gpt2.cpp +0 -105
  879. data/ext/sources/examples/talk-llama/models/gptneox.cpp +0 -144
  880. data/ext/sources/examples/talk-llama/models/granite-hybrid.cpp +0 -196
  881. data/ext/sources/examples/talk-llama/models/granite.cpp +0 -211
  882. data/ext/sources/examples/talk-llama/models/graph-context-mamba.cpp +0 -283
  883. data/ext/sources/examples/talk-llama/models/grok.cpp +0 -159
  884. data/ext/sources/examples/talk-llama/models/grovemoe.cpp +0 -141
  885. data/ext/sources/examples/talk-llama/models/hunyuan-dense.cpp +0 -132
  886. data/ext/sources/examples/talk-llama/models/hunyuan-moe.cpp +0 -154
  887. data/ext/sources/examples/talk-llama/models/internlm2.cpp +0 -120
  888. data/ext/sources/examples/talk-llama/models/jais.cpp +0 -86
  889. data/ext/sources/examples/talk-llama/models/jamba.cpp +0 -106
  890. data/ext/sources/examples/talk-llama/models/lfm2.cpp +0 -175
  891. data/ext/sources/examples/talk-llama/models/llada-moe.cpp +0 -122
  892. data/ext/sources/examples/talk-llama/models/llada.cpp +0 -99
  893. data/ext/sources/examples/talk-llama/models/llama-iswa.cpp +0 -178
  894. data/ext/sources/examples/talk-llama/models/llama.cpp +0 -168
  895. data/ext/sources/examples/talk-llama/models/maincoder.cpp +0 -117
  896. data/ext/sources/examples/talk-llama/models/mamba.cpp +0 -55
  897. data/ext/sources/examples/talk-llama/models/mimo2-iswa.cpp +0 -123
  898. data/ext/sources/examples/talk-llama/models/minicpm3.cpp +0 -199
  899. data/ext/sources/examples/talk-llama/models/minimax-m2.cpp +0 -124
  900. data/ext/sources/examples/talk-llama/models/mistral3.cpp +0 -160
  901. data/ext/sources/examples/talk-llama/models/models.h +0 -569
  902. data/ext/sources/examples/talk-llama/models/modern-bert.cpp +0 -116
  903. data/ext/sources/examples/talk-llama/models/mpt.cpp +0 -126
  904. data/ext/sources/examples/talk-llama/models/nemotron-h.cpp +0 -150
  905. data/ext/sources/examples/talk-llama/models/nemotron.cpp +0 -122
  906. data/ext/sources/examples/talk-llama/models/neo-bert.cpp +0 -104
  907. data/ext/sources/examples/talk-llama/models/olmo.cpp +0 -121
  908. data/ext/sources/examples/talk-llama/models/olmo2.cpp +0 -150
  909. data/ext/sources/examples/talk-llama/models/olmoe.cpp +0 -124
  910. data/ext/sources/examples/talk-llama/models/openai-moe-iswa.cpp +0 -127
  911. data/ext/sources/examples/talk-llama/models/openelm.cpp +0 -124
  912. data/ext/sources/examples/talk-llama/models/orion.cpp +0 -123
  913. data/ext/sources/examples/talk-llama/models/pangu-embedded.cpp +0 -121
  914. data/ext/sources/examples/talk-llama/models/phi2.cpp +0 -121
  915. data/ext/sources/examples/talk-llama/models/phi3.cpp +0 -152
  916. data/ext/sources/examples/talk-llama/models/plamo.cpp +0 -110
  917. data/ext/sources/examples/talk-llama/models/plamo2.cpp +0 -316
  918. data/ext/sources/examples/talk-llama/models/plamo3.cpp +0 -128
  919. data/ext/sources/examples/talk-llama/models/plm.cpp +0 -168
  920. data/ext/sources/examples/talk-llama/models/qwen.cpp +0 -108
  921. data/ext/sources/examples/talk-llama/models/qwen2.cpp +0 -126
  922. data/ext/sources/examples/talk-llama/models/qwen2moe.cpp +0 -151
  923. data/ext/sources/examples/talk-llama/models/qwen2vl.cpp +0 -117
  924. data/ext/sources/examples/talk-llama/models/qwen3.cpp +0 -117
  925. data/ext/sources/examples/talk-llama/models/qwen3moe.cpp +0 -124
  926. data/ext/sources/examples/talk-llama/models/qwen3next.cpp +0 -873
  927. data/ext/sources/examples/talk-llama/models/qwen3vl-moe.cpp +0 -149
  928. data/ext/sources/examples/talk-llama/models/qwen3vl.cpp +0 -141
  929. data/ext/sources/examples/talk-llama/models/refact.cpp +0 -94
  930. data/ext/sources/examples/talk-llama/models/rnd1.cpp +0 -126
  931. data/ext/sources/examples/talk-llama/models/rwkv6-base.cpp +0 -162
  932. data/ext/sources/examples/talk-llama/models/rwkv6.cpp +0 -94
  933. data/ext/sources/examples/talk-llama/models/rwkv6qwen2.cpp +0 -86
  934. data/ext/sources/examples/talk-llama/models/rwkv7-base.cpp +0 -135
  935. data/ext/sources/examples/talk-llama/models/rwkv7.cpp +0 -90
  936. data/ext/sources/examples/talk-llama/models/seed-oss.cpp +0 -124
  937. data/ext/sources/examples/talk-llama/models/smallthinker.cpp +0 -126
  938. data/ext/sources/examples/talk-llama/models/smollm3.cpp +0 -128
  939. data/ext/sources/examples/talk-llama/models/stablelm.cpp +0 -146
  940. data/ext/sources/examples/talk-llama/models/starcoder.cpp +0 -100
  941. data/ext/sources/examples/talk-llama/models/starcoder2.cpp +0 -121
  942. data/ext/sources/examples/talk-llama/models/t5-dec.cpp +0 -166
  943. data/ext/sources/examples/talk-llama/models/t5-enc.cpp +0 -96
  944. data/ext/sources/examples/talk-llama/models/wavtokenizer-dec.cpp +0 -149
  945. data/ext/sources/examples/talk-llama/models/xverse.cpp +0 -108
  946. data/ext/sources/examples/talk-llama/prompts/talk-alpaca.txt +0 -23
  947. data/ext/sources/examples/talk-llama/speak +0 -40
  948. data/ext/sources/examples/talk-llama/speak.bat +0 -1
  949. data/ext/sources/examples/talk-llama/speak.ps1 +0 -14
  950. data/ext/sources/examples/talk-llama/talk-llama.cpp +0 -813
  951. data/ext/sources/examples/talk-llama/unicode-data.cpp +0 -7034
  952. data/ext/sources/examples/talk-llama/unicode-data.h +0 -20
  953. data/ext/sources/examples/talk-llama/unicode.cpp +0 -1147
  954. data/ext/sources/examples/talk-llama/unicode.h +0 -111
  955. data/ext/sources/examples/wchess/CMakeLists.txt +0 -10
  956. data/ext/sources/examples/wchess/libwchess/CMakeLists.txt +0 -19
  957. data/ext/sources/examples/wchess/libwchess/Chessboard.cpp +0 -803
  958. data/ext/sources/examples/wchess/libwchess/Chessboard.h +0 -33
  959. data/ext/sources/examples/wchess/libwchess/WChess.cpp +0 -193
  960. data/ext/sources/examples/wchess/libwchess/WChess.h +0 -63
  961. data/ext/sources/examples/wchess/libwchess/test-chessboard.cpp +0 -117
  962. data/ext/sources/examples/wchess/wchess.cmd/CMakeLists.txt +0 -8
  963. data/ext/sources/examples/wchess/wchess.cmd/wchess.cmd.cpp +0 -253
  964. data/ext/sources/examples/whisper.wasm/CMakeLists.txt +0 -50
  965. data/ext/sources/examples/whisper.wasm/emscripten.cpp +0 -118
  966. data/ext/sources/examples/whisper.wasm/index-tmpl.html +0 -659
  967. data/ext/sources/ggml/cmake/BuildTypes.cmake +0 -54
  968. data/ext/sources/ggml/src/ggml-cpu/llamafile/sgemm-ppc.h +0 -333
  969. data/ext/sources/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +0 -99
  970. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-dma.h +0 -157
  971. data/ext/sources/ggml/src/ggml-hexagon/htp/htp-msg.h +0 -165
  972. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-exp.c +0 -94
  973. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-inverse.c +0 -72
  974. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-sigmoid.c +0 -49
  975. data/ext/sources/ggml/src/ggml-hexagon/htp/hvx-utils.c +0 -1020
  976. data/ext/sources/ggml/src/ggml-hexagon/htp/ops-utils.h +0 -149
  977. data/ext/sources/ggml/src/ggml-hexagon/htp-utils.c +0 -454
  978. data/ext/sources/ggml/src/ggml-hexagon/htp-utils.h +0 -221
  979. data/ext/sources/ggml/src/ggml-hexagon/op-desc.h +0 -153
  980. data/ext/sources/ggml/src/ggml-opencl/kernels/embed_kernel.py +0 -26
  981. data/ext/sources/ggml/src/ggml-vulkan/vulkan-shaders/rte.glsl +0 -5
  982. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/bin_op.tmpl.wgsl +0 -188
  983. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/binary_head.tmpl +0 -45
  984. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py +0 -147
  985. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/glu.tmpl.wgsl +0 -323
  986. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl +0 -907
  987. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.tmpl.wgsl +0 -247
  988. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.tmpl.wgsl +0 -267
  989. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/rms_norm.wgsl +0 -123
  990. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/set_rows.tmpl.wgsl +0 -112
  991. data/ext/sources/ggml/src/ggml-webgpu/wgsl-shaders/unary_op.wgsl +0 -483
  992. data/ext/sources/tests/CMakeLists.txt +0 -112
  993. data/ext/sources/tests/earnings21/eval.mk +0 -58
  994. data/ext/sources/tests/earnings21/eval.py +0 -68
  995. data/ext/sources/tests/earnings21/normalizers/__init__.py +0 -2
  996. data/ext/sources/tests/earnings21/normalizers/basic.py +0 -80
  997. data/ext/sources/tests/earnings21/normalizers/english.json +0 -1741
  998. data/ext/sources/tests/earnings21/normalizers/english.py +0 -550
  999. data/ext/sources/tests/earnings21/requirements.txt +0 -6
  1000. data/ext/sources/tests/en-0-ref.txt +0 -1
  1001. data/ext/sources/tests/en-1-ref.txt +0 -1
  1002. data/ext/sources/tests/en-2-ref.txt +0 -1
  1003. data/ext/sources/tests/es-0-ref.txt +0 -1
  1004. data/ext/sources/tests/librispeech/eval.mk +0 -39
  1005. data/ext/sources/tests/librispeech/eval.py +0 -47
  1006. data/ext/sources/tests/librispeech/normalizers/__init__.py +0 -2
  1007. data/ext/sources/tests/librispeech/normalizers/basic.py +0 -80
  1008. data/ext/sources/tests/librispeech/normalizers/english.json +0 -1741
  1009. data/ext/sources/tests/librispeech/normalizers/english.py +0 -550
  1010. data/ext/sources/tests/librispeech/requirements.txt +0 -6
  1011. data/ext/sources/tests/run-tests.sh +0 -130
  1012. data/ext/sources/tests/test-c.c +0 -3
  1013. data/ext/sources/tests/test-vad-full.cpp +0 -56
  1014. data/ext/sources/tests/test-vad.cpp +0 -83
  1015. data/ext/sources/tests/test-whisper.js +0 -58
  1016. data/lib/whisper/context.rb +0 -15
  1017. data/lib/whisper/segment.rb +0 -58
@@ -1,2559 +0,0 @@
1
- #include "llama-arch.h"
2
-
3
- #include "llama-impl.h"
4
-
5
- #include <map>
6
- #include <set>
7
-
8
- static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
9
- { LLM_ARCH_CLIP, "clip" }, // dummy, only used by llama-quantize
10
- { LLM_ARCH_LLAMA, "llama" },
11
- { LLM_ARCH_LLAMA4, "llama4" },
12
- { LLM_ARCH_DECI, "deci" },
13
- { LLM_ARCH_FALCON, "falcon" },
14
- { LLM_ARCH_GROK, "grok" },
15
- { LLM_ARCH_GPT2, "gpt2" },
16
- { LLM_ARCH_GPTJ, "gptj" },
17
- { LLM_ARCH_GPTNEOX, "gptneox" },
18
- { LLM_ARCH_MPT, "mpt" },
19
- { LLM_ARCH_BAICHUAN, "baichuan" },
20
- { LLM_ARCH_STARCODER, "starcoder" },
21
- { LLM_ARCH_REFACT, "refact" },
22
- { LLM_ARCH_BERT, "bert" },
23
- { LLM_ARCH_MODERN_BERT, "modern-bert" },
24
- { LLM_ARCH_NOMIC_BERT, "nomic-bert" },
25
- { LLM_ARCH_NOMIC_BERT_MOE, "nomic-bert-moe" },
26
- { LLM_ARCH_NEO_BERT, "neo-bert" },
27
- { LLM_ARCH_JINA_BERT_V2, "jina-bert-v2" },
28
- { LLM_ARCH_JINA_BERT_V3, "jina-bert-v3" },
29
- { LLM_ARCH_BLOOM, "bloom" },
30
- { LLM_ARCH_STABLELM, "stablelm" },
31
- { LLM_ARCH_QWEN, "qwen" },
32
- { LLM_ARCH_QWEN2, "qwen2" },
33
- { LLM_ARCH_QWEN2MOE, "qwen2moe" },
34
- { LLM_ARCH_QWEN2VL, "qwen2vl" },
35
- { LLM_ARCH_QWEN3, "qwen3" },
36
- { LLM_ARCH_QWEN3MOE, "qwen3moe" },
37
- { LLM_ARCH_QWEN3NEXT, "qwen3next" },
38
- { LLM_ARCH_QWEN3VL, "qwen3vl" },
39
- { LLM_ARCH_QWEN3VLMOE, "qwen3vlmoe" },
40
- { LLM_ARCH_PHI2, "phi2" },
41
- { LLM_ARCH_PHI3, "phi3" },
42
- { LLM_ARCH_PHIMOE, "phimoe" },
43
- { LLM_ARCH_PLAMO, "plamo" },
44
- { LLM_ARCH_PLAMO2, "plamo2" },
45
- { LLM_ARCH_PLAMO3, "plamo3" },
46
- { LLM_ARCH_CODESHELL, "codeshell" },
47
- { LLM_ARCH_ORION, "orion" },
48
- { LLM_ARCH_INTERNLM2, "internlm2" },
49
- { LLM_ARCH_MINICPM, "minicpm" },
50
- { LLM_ARCH_MINICPM3, "minicpm3" },
51
- { LLM_ARCH_GEMMA, "gemma" },
52
- { LLM_ARCH_GEMMA2, "gemma2" },
53
- { LLM_ARCH_GEMMA3, "gemma3" },
54
- { LLM_ARCH_GEMMA3N, "gemma3n" },
55
- { LLM_ARCH_GEMMA_EMBEDDING, "gemma-embedding" },
56
- { LLM_ARCH_STARCODER2, "starcoder2" },
57
- { LLM_ARCH_MAMBA, "mamba" },
58
- { LLM_ARCH_MAMBA2, "mamba2" },
59
- { LLM_ARCH_JAMBA, "jamba" },
60
- { LLM_ARCH_FALCON_H1, "falcon-h1" },
61
- { LLM_ARCH_XVERSE, "xverse" },
62
- { LLM_ARCH_COMMAND_R, "command-r" },
63
- { LLM_ARCH_COHERE2, "cohere2" },
64
- { LLM_ARCH_DBRX, "dbrx" },
65
- { LLM_ARCH_OLMO, "olmo" },
66
- { LLM_ARCH_OLMO2, "olmo2" },
67
- { LLM_ARCH_OLMOE, "olmoe" },
68
- { LLM_ARCH_OPENELM, "openelm" },
69
- { LLM_ARCH_ARCTIC, "arctic" },
70
- { LLM_ARCH_DEEPSEEK, "deepseek" },
71
- { LLM_ARCH_DEEPSEEK2, "deepseek2" },
72
- { LLM_ARCH_CHATGLM, "chatglm" },
73
- { LLM_ARCH_GLM4, "glm4" },
74
- { LLM_ARCH_GLM4_MOE, "glm4moe" },
75
- { LLM_ARCH_BITNET, "bitnet" },
76
- { LLM_ARCH_T5, "t5" },
77
- { LLM_ARCH_T5ENCODER, "t5encoder" },
78
- { LLM_ARCH_JAIS, "jais" },
79
- { LLM_ARCH_NEMOTRON, "nemotron" },
80
- { LLM_ARCH_NEMOTRON_H, "nemotron_h" },
81
- { LLM_ARCH_NEMOTRON_H_MOE, "nemotron_h_moe" },
82
- { LLM_ARCH_EXAONE, "exaone" },
83
- { LLM_ARCH_EXAONE4, "exaone4" },
84
- { LLM_ARCH_RWKV6, "rwkv6" },
85
- { LLM_ARCH_RWKV6QWEN2, "rwkv6qwen2" },
86
- { LLM_ARCH_RWKV7, "rwkv7" },
87
- { LLM_ARCH_ARWKV7, "arwkv7" },
88
- { LLM_ARCH_GRANITE, "granite" },
89
- { LLM_ARCH_GRANITE_MOE, "granitemoe" },
90
- { LLM_ARCH_GRANITE_HYBRID, "granitehybrid" },
91
- { LLM_ARCH_CHAMELEON, "chameleon" },
92
- { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
93
- { LLM_ARCH_PLM, "plm" },
94
- { LLM_ARCH_BAILINGMOE, "bailingmoe" },
95
- { LLM_ARCH_BAILINGMOE2, "bailingmoe2" },
96
- { LLM_ARCH_DOTS1, "dots1" },
97
- { LLM_ARCH_ARCEE, "arcee" },
98
- { LLM_ARCH_AFMOE, "afmoe" },
99
- { LLM_ARCH_ERNIE4_5, "ernie4_5" },
100
- { LLM_ARCH_ERNIE4_5_MOE, "ernie4_5-moe" },
101
- { LLM_ARCH_HUNYUAN_MOE, "hunyuan-moe" },
102
- { LLM_ARCH_HUNYUAN_DENSE, "hunyuan-dense" },
103
- { LLM_ARCH_SMOLLM3, "smollm3" },
104
- { LLM_ARCH_OPENAI_MOE, "gpt-oss" },
105
- { LLM_ARCH_LFM2, "lfm2" },
106
- { LLM_ARCH_LFM2MOE, "lfm2moe" },
107
- { LLM_ARCH_DREAM, "dream" },
108
- { LLM_ARCH_SMALLTHINKER, "smallthinker" },
109
- { LLM_ARCH_LLADA, "llada" },
110
- { LLM_ARCH_LLADA_MOE, "llada-moe" },
111
- { LLM_ARCH_SEED_OSS, "seed_oss" },
112
- { LLM_ARCH_GROVEMOE, "grovemoe" },
113
- { LLM_ARCH_APERTUS, "apertus" },
114
- { LLM_ARCH_MINIMAX_M2, "minimax-m2" },
115
- { LLM_ARCH_COGVLM, "cogvlm" },
116
- { LLM_ARCH_RND1, "rnd1" },
117
- { LLM_ARCH_PANGU_EMBED, "pangu-embedded" },
118
- { LLM_ARCH_MISTRAL3, "mistral3" },
119
- { LLM_ARCH_MIMO2, "mimo2" },
120
- { LLM_ARCH_LLAMA_EMBED, "llama-embed" },
121
- { LLM_ARCH_MAINCODER, "maincoder" },
122
- { LLM_ARCH_UNKNOWN, "(unknown)" },
123
- };
124
-
125
- static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
126
- { LLM_KV_GENERAL_TYPE, "general.type" },
127
- { LLM_KV_GENERAL_ARCHITECTURE, "general.architecture" },
128
- { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version" },
129
- { LLM_KV_GENERAL_ALIGNMENT, "general.alignment" },
130
- { LLM_KV_GENERAL_FILE_TYPE, "general.file_type" },
131
- { LLM_KV_GENERAL_SAMPLING_SEQUENCE, "general.sampling.sequence" },
132
- { LLM_KV_GENERAL_SAMPLING_TOP_K, "general.sampling.top_k" },
133
- { LLM_KV_GENERAL_SAMPLING_TOP_P, "general.sampling.top_p" },
134
- { LLM_KV_GENERAL_SAMPLING_MIN_P, "general.sampling.min_p" },
135
- { LLM_KV_GENERAL_SAMPLING_XTC_PROBABILITY, "general.sampling.xtc_probability" },
136
- { LLM_KV_GENERAL_SAMPLING_XTC_THRESHOLD, "general.sampling.xtc_threshold" },
137
- { LLM_KV_GENERAL_SAMPLING_TEMP, "general.sampling.temp" },
138
- { LLM_KV_GENERAL_SAMPLING_PENALTY_LAST_N, "general.sampling.penalty_last_n" },
139
- { LLM_KV_GENERAL_SAMPLING_PENALTY_REPEAT, "general.sampling.penalty_repeat" },
140
- { LLM_KV_GENERAL_SAMPLING_MIROSTAT, "general.sampling.mirostat" },
141
- { LLM_KV_GENERAL_SAMPLING_MIROSTAT_TAU, "general.sampling.mirostat_tau" },
142
- { LLM_KV_GENERAL_SAMPLING_MIROSTAT_ETA, "general.sampling.mirostat_eta" },
143
- { LLM_KV_GENERAL_NAME, "general.name" },
144
- { LLM_KV_GENERAL_AUTHOR, "general.author" },
145
- { LLM_KV_GENERAL_VERSION, "general.version" },
146
- { LLM_KV_GENERAL_URL, "general.url" },
147
- { LLM_KV_GENERAL_DESCRIPTION, "general.description" },
148
- { LLM_KV_GENERAL_LICENSE, "general.license" },
149
- { LLM_KV_GENERAL_SOURCE_URL, "general.source.url" },
150
- { LLM_KV_GENERAL_SOURCE_HF_REPO, "general.source.huggingface.repository" },
151
-
152
- { LLM_KV_VOCAB_SIZE, "%s.vocab_size" },
153
- { LLM_KV_CONTEXT_LENGTH, "%s.context_length" },
154
- { LLM_KV_EMBEDDING_LENGTH, "%s.embedding_length" },
155
- { LLM_KV_EMBEDDING_LENGTH_OUT, "%s.embedding_length_out" },
156
- { LLM_KV_FEATURES_LENGTH, "%s.features_length" },
157
- { LLM_KV_BLOCK_COUNT, "%s.block_count" },
158
- { LLM_KV_LEADING_DENSE_BLOCK_COUNT, "%s.leading_dense_block_count" },
159
- { LLM_KV_FEED_FORWARD_LENGTH, "%s.feed_forward_length" },
160
- { LLM_KV_EXPERT_FEED_FORWARD_LENGTH, "%s.expert_feed_forward_length" },
161
- { LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, "%s.expert_shared_feed_forward_length" },
162
- { LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH, "%s.expert_chunk_feed_forward_length" },
163
- { LLM_KV_USE_PARALLEL_RESIDUAL, "%s.use_parallel_residual" },
164
- { LLM_KV_TENSOR_DATA_LAYOUT, "%s.tensor_data_layout" },
165
- { LLM_KV_EXPERT_COUNT, "%s.expert_count" },
166
- { LLM_KV_EXPERT_USED_COUNT, "%s.expert_used_count" },
167
- { LLM_KV_EXPERT_SHARED_COUNT, "%s.expert_shared_count" },
168
- { LLM_KV_EXPERT_GROUP_COUNT, "%s.expert_group_count" },
169
- { LLM_KV_EXPERT_GROUP_USED_COUNT, "%s.expert_group_used_count" },
170
- { LLM_KV_EXPERT_WEIGHTS_SCALE, "%s.expert_weights_scale" },
171
- { LLM_KV_EXPERT_WEIGHTS_NORM, "%s.expert_weights_norm" },
172
- { LLM_KV_EXPERT_GATING_FUNC, "%s.expert_gating_func" },
173
- { LLM_KV_EXPERT_GROUP_SCALE, "%s.expert_group_scale" },
174
- { LLM_KV_EXPERTS_PER_GROUP, "%s.experts_per_group" },
175
- { LLM_KV_MOE_EVERY_N_LAYERS, "%s.moe_every_n_layers" },
176
- { LLM_KV_NEXTN_PREDICT_LAYERS, "%s.nextn_predict_layers" },
177
- { LLM_KV_NUM_DEEPSTACK_LAYERS, "%s.n_deepstack_layers" },
178
- { LLM_KV_POOLING_TYPE, "%s.pooling_type" },
179
- { LLM_KV_LOGIT_SCALE, "%s.logit_scale" },
180
- { LLM_KV_DECODER_START_TOKEN_ID, "%s.decoder_start_token_id" },
181
- { LLM_KV_DECODER_BLOCK_COUNT, "%s.decoder_block_count" },
182
- { LLM_KV_ATTN_LOGIT_SOFTCAPPING, "%s.attn_logit_softcapping" },
183
- { LLM_KV_ROUTER_LOGIT_SOFTCAPPING, "%s.router_logit_softcapping" },
184
- { LLM_KV_FINAL_LOGIT_SOFTCAPPING, "%s.final_logit_softcapping" },
185
- { LLM_KV_SWIN_NORM, "%s.swin_norm" },
186
- { LLM_KV_RESCALE_EVERY_N_LAYERS, "%s.rescale_every_n_layers" },
187
- { LLM_KV_TIME_MIX_EXTRA_DIM, "%s.time_mix_extra_dim" },
188
- { LLM_KV_TIME_DECAY_EXTRA_DIM, "%s.time_decay_extra_dim" },
189
- { LLM_KV_RESIDUAL_SCALE, "%s.residual_scale" },
190
- { LLM_KV_EMBEDDING_SCALE, "%s.embedding_scale" },
191
- { LLM_KV_TOKEN_SHIFT_COUNT, "%s.token_shift_count" },
192
- { LLM_KV_INTERLEAVE_MOE_LAYER_STEP, "%s.interleave_moe_layer_step" },
193
-
194
- { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
195
- { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
196
- { LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" },
197
- { LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" },
198
- { LLM_KV_ATTENTION_KEY_LENGTH, "%s.attention.key_length" },
199
- { LLM_KV_ATTENTION_VALUE_LENGTH, "%s.attention.value_length" },
200
- { LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" },
201
- { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" },
202
- { LLM_KV_ATTENTION_GROUPNORM_EPS, "%s.attention.group_norm_epsilon" },
203
- { LLM_KV_ATTENTION_GROUPNORM_GROUPS, "%s.attention.group_norm_groups" },
204
- { LLM_KV_ATTENTION_CAUSAL, "%s.attention.causal" },
205
- { LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" },
206
- { LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" },
207
- { LLM_KV_ATTENTION_DECAY_LORA_RANK, "%s.attention.decay_lora_rank" },
208
- { LLM_KV_ATTENTION_ICLR_LORA_RANK, "%s.attention.iclr_lora_rank" },
209
- { LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, "%s.attention.value_residual_mix_lora_rank" },
210
- { LLM_KV_ATTENTION_GATE_LORA_RANK, "%s.attention.gate_lora_rank" },
211
- { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
212
- { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" },
213
- { LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, "%s.attention.sliding_window_pattern" },
214
- { LLM_KV_ATTENTION_SCALE, "%s.attention.scale" },
215
- { LLM_KV_ATTENTION_OUTPUT_SCALE, "%s.attention.output_scale" },
216
- { LLM_KV_ATTENTION_TEMPERATURE_LENGTH, "%s.attention.temperature_length" },
217
- { LLM_KV_ATTENTION_TEMPERATURE_SCALE, "%s.attention.temperature_scale" },
218
- { LLM_KV_ATTENTION_KEY_LENGTH_MLA, "%s.attention.key_length_mla" },
219
- { LLM_KV_ATTENTION_VALUE_LENGTH_MLA, "%s.attention.value_length_mla" },
220
-
221
- { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
222
- { LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" },
223
- { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" },
224
- { LLM_KV_ROPE_FREQ_BASE_SWA, "%s.rope.freq_base_swa" },
225
- { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" },
226
- { LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" },
227
- { LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" },
228
- { LLM_KV_ROPE_SCALING_ATTN_FACTOR, "%s.rope.scaling.attn_factor" },
229
- { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" },
230
- { LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" },
231
- { LLM_KV_ROPE_SCALING_YARN_LOG_MUL, "%s.rope.scaling.yarn_log_multiplier" },
232
- { LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR, "%s.rope.scaling.yarn_ext_factor" },
233
- { LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR, "%s.rope.scaling.yarn_attn_factor" },
234
- { LLM_KV_ROPE_SCALING_YARN_BETA_FAST, "%s.rope.scaling.yarn_beta_fast" },
235
- { LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, "%s.rope.scaling.yarn_beta_slow" },
236
-
237
- { LLM_KV_SPLIT_NO, "split.no" },
238
- { LLM_KV_SPLIT_COUNT, "split.count" },
239
- { LLM_KV_SPLIT_TENSORS_COUNT, "split.tensors.count" },
240
-
241
- { LLM_KV_SSM_CONV_KERNEL, "%s.ssm.conv_kernel" },
242
- { LLM_KV_SSM_INNER_SIZE, "%s.ssm.inner_size" },
243
- { LLM_KV_SSM_STATE_SIZE, "%s.ssm.state_size" },
244
- { LLM_KV_SSM_TIME_STEP_RANK, "%s.ssm.time_step_rank" },
245
- { LLM_KV_SSM_GROUP_COUNT, "%s.ssm.group_count" },
246
- { LLM_KV_SSM_DT_B_C_RMS, "%s.ssm.dt_b_c_rms" },
247
-
248
- { LLM_KV_WKV_HEAD_SIZE, "%s.wkv.head_size" },
249
-
250
- { LLM_KV_POSNET_EMBEDDING_LENGTH, "%s.posnet.embedding_length" },
251
- { LLM_KV_POSNET_BLOCK_COUNT, "%s.posnet.block_count" },
252
-
253
- { LLM_KV_CONVNEXT_EMBEDDING_LENGTH, "%s.convnext.embedding_length" },
254
- { LLM_KV_CONVNEXT_BLOCK_COUNT, "%s.convnext.block_count" },
255
-
256
- { LLM_KV_CLASSIFIER_OUTPUT_LABELS, "%s.classifier.output_labels" },
257
-
258
- { LLM_KV_SHORTCONV_L_CACHE, "%s.shortconv.l_cache" },
259
- // sentence-transformers dense modules feature dims
260
- { LLM_KV_DENSE_2_FEAT_IN, "%s.dense_2_feat_in" },
261
- { LLM_KV_DENSE_2_FEAT_OUT, "%s.dense_2_feat_out" },
262
- { LLM_KV_DENSE_3_FEAT_IN, "%s.dense_3_feat_in" },
263
- { LLM_KV_DENSE_3_FEAT_OUT, "%s.dense_3_feat_out" },
264
-
265
- { LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" },
266
- { LLM_KV_TOKENIZER_PRE, "tokenizer.ggml.pre" },
267
- { LLM_KV_TOKENIZER_LIST, "tokenizer.ggml.tokens" },
268
- { LLM_KV_TOKENIZER_TOKEN_TYPE, "tokenizer.ggml.token_type" },
269
- { LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, "tokenizer.ggml.token_type_count" },
270
- { LLM_KV_TOKENIZER_SCORES, "tokenizer.ggml.scores" },
271
- { LLM_KV_TOKENIZER_MERGES, "tokenizer.ggml.merges" },
272
- { LLM_KV_TOKENIZER_BOS_ID, "tokenizer.ggml.bos_token_id" },
273
- { LLM_KV_TOKENIZER_EOS_ID, "tokenizer.ggml.eos_token_id" },
274
- { LLM_KV_TOKENIZER_EOT_ID, "tokenizer.ggml.eot_token_id" },
275
- { LLM_KV_TOKENIZER_EOM_ID, "tokenizer.ggml.eom_token_id" },
276
- { LLM_KV_TOKENIZER_UNK_ID, "tokenizer.ggml.unknown_token_id" },
277
- { LLM_KV_TOKENIZER_SEP_ID, "tokenizer.ggml.seperator_token_id" },
278
- { LLM_KV_TOKENIZER_PAD_ID, "tokenizer.ggml.padding_token_id" },
279
- { LLM_KV_TOKENIZER_CLS_ID, "tokenizer.ggml.cls_token_id" },
280
- { LLM_KV_TOKENIZER_MASK_ID, "tokenizer.ggml.mask_token_id" },
281
- { LLM_KV_TOKENIZER_ADD_BOS, "tokenizer.ggml.add_bos_token" },
282
- { LLM_KV_TOKENIZER_ADD_EOS, "tokenizer.ggml.add_eos_token" },
283
- { LLM_KV_TOKENIZER_ADD_SEP, "tokenizer.ggml.add_sep_token" },
284
- { LLM_KV_TOKENIZER_ADD_PREFIX, "tokenizer.ggml.add_space_prefix" },
285
- { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS, "tokenizer.ggml.remove_extra_whitespaces" },
286
- { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap" },
287
- { LLM_KV_TOKENIZER_HF_JSON, "tokenizer.huggingface.json" },
288
- { LLM_KV_TOKENIZER_RWKV, "tokenizer.rwkv.world" },
289
- { LLM_KV_TOKENIZER_CHAT_TEMPLATE, "tokenizer.chat_template" },
290
- { LLM_KV_TOKENIZER_FIM_PRE_ID, "tokenizer.ggml.fim_pre_token_id" },
291
- { LLM_KV_TOKENIZER_FIM_SUF_ID, "tokenizer.ggml.fim_suf_token_id" },
292
- { LLM_KV_TOKENIZER_FIM_MID_ID, "tokenizer.ggml.fim_mid_token_id" },
293
- { LLM_KV_TOKENIZER_FIM_PAD_ID, "tokenizer.ggml.fim_pad_token_id" },
294
- { LLM_KV_TOKENIZER_FIM_REP_ID, "tokenizer.ggml.fim_rep_token_id" },
295
- { LLM_KV_TOKENIZER_FIM_SEP_ID, "tokenizer.ggml.fim_sep_token_id" },
296
-
297
- { LLM_KV_ADAPTER_TYPE, "adapter.type" },
298
- { LLM_KV_ADAPTER_LORA_ALPHA, "adapter.lora.alpha" },
299
- { LLM_KV_ADAPTER_LORA_TASK_NAME, "adapter.lora.task_name" },
300
- { LLM_KV_ADAPTER_LORA_PROMPT_PREFIX, "adapter.lora.prompt_prefix" },
301
- { LLM_KV_ADAPTER_ALORA_INVOCATION_TOKENS, "adapter.alora.invocation_tokens" },
302
-
303
- { LLM_KV_XIELU_ALPHA_N, "xielu.alpha_n" },
304
- { LLM_KV_XIELU_ALPHA_P, "xielu.alpha_p" },
305
- { LLM_KV_XIELU_BETA, "xielu.beta" },
306
- { LLM_KV_XIELU_EPS, "xielu.eps" },
307
-
308
- // deprecated
309
- { LLM_KV_TOKENIZER_PREFIX_ID, "tokenizer.ggml.prefix_token_id" },
310
- { LLM_KV_TOKENIZER_SUFFIX_ID, "tokenizer.ggml.suffix_token_id" },
311
- { LLM_KV_TOKENIZER_MIDDLE_ID, "tokenizer.ggml.middle_token_id" },
312
- };
313
-
314
- static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
315
- { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
316
- { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
317
- { LLM_TENSOR_OUTPUT_NORM_LFM2, "token_embd_norm" }, // fix for wrong tensor name
318
- { LLM_TENSOR_OUTPUT, "output" },
319
- { LLM_TENSOR_ROPE_FREQS, "rope_freqs" },
320
- { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
321
- { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
322
- { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
323
- { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
324
- { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
325
- { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" },
326
- { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
327
- { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
328
- { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
329
- { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
330
- { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
331
- { LLM_TENSOR_FFN_GATE_EXP, "blk.%d.ffn_gate.%d" },
332
- { LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" },
333
- { LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" },
334
- { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
335
- { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
336
- { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
337
- { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" },
338
- { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
339
- { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
340
- { LLM_TENSOR_ATTN_GATE, "blk.%d.attn_gate" },
341
- { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" },
342
- { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" },
343
- { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
344
- { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
345
- { LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" },
346
- { LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" },
347
- { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
348
- { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" },
349
- { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" },
350
- { LLM_TENSOR_POS_EMBD, "position_embd" },
351
- { LLM_TENSOR_FFN_ACT, "blk.%d.ffn.act" },
352
- { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
353
- { LLM_TENSOR_TOKEN_TYPES, "token_types" },
354
- { LLM_TENSOR_CLS, "cls" },
355
- { LLM_TENSOR_CLS_OUT, "cls.output" },
356
- { LLM_TENSOR_ENC_OUTPUT_NORM, "enc.output_norm" },
357
- { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
358
- { LLM_TENSOR_SSM_A_NOSCAN, "blk.%d.ssm_a" },
359
- { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" },
360
- { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" },
361
- { LLM_TENSOR_SSM_BETA_ALPHA, "blk.%d.ssm_ba" },
362
- { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" },
363
- { LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" },
364
- { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" },
365
- { LLM_TENSOR_ROPE_FACTORS_LONG, "rope_factors_long" },
366
- { LLM_TENSOR_ROPE_FACTORS_SHORT, "rope_factors_short" },
367
- { LLM_TENSOR_SSM_X, "blk.%d.ssm_x" },
368
- { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" },
369
- { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" },
370
- { LLM_TENSOR_SSM_DT_NORM, "blk.%d.ssm_dt_norm" },
371
- { LLM_TENSOR_SSM_B_NORM, "blk.%d.ssm_b_norm" },
372
- { LLM_TENSOR_SSM_C_NORM, "blk.%d.ssm_c_norm" },
373
- { LLM_TENSOR_ATTN_Q_A_NORM, "blk.%d.attn_q_a_norm" },
374
- { LLM_TENSOR_ATTN_KV_A_NORM, "blk.%d.attn_kv_a_norm" },
375
- { LLM_TENSOR_ATTN_Q_A, "blk.%d.attn_q_a" },
376
- { LLM_TENSOR_ATTN_Q_B, "blk.%d.attn_q_b" },
377
- { LLM_TENSOR_ATTN_KV_A_MQA, "blk.%d.attn_kv_a_mqa" },
378
- { LLM_TENSOR_ATTN_KV_B, "blk.%d.attn_kv_b" },
379
- { LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "per_layer_token_embd" },
380
- { LLM_TENSOR_PER_LAYER_MODEL_PROJ, "per_layer_model_proj" },
381
- { LLM_TENSOR_PER_LAYER_PROJ_NORM, "per_layer_proj_norm" },
382
- { LLM_TENSOR_ALTUP_UNEMBD_PROJ, "altup_unembd_proj" },
383
- { LLM_TENSOR_ALTUP_PROJ, "altup_proj" },
384
- { LLM_TENSOR_PER_LAYER_INP_GATE, "blk.%d.inp_gate" },
385
- { LLM_TENSOR_PER_LAYER_PROJ, "blk.%d.proj" },
386
- { LLM_TENSOR_PER_LAYER_POST_NORM, "blk.%d.post_norm" },
387
- { LLM_TENSOR_ALTUP_CORRECT_COEF, "blk.%d.altup_correct_coef" },
388
- { LLM_TENSOR_ALTUP_CORRECT_SCALE, "blk.%d.altup_correct_scale" },
389
- { LLM_TENSOR_ALTUP_PREDICT_COEF, "blk.%d.altup_predict_coef" },
390
- { LLM_TENSOR_ALTUP_ROUTER, "blk.%d.altup_router" },
391
- { LLM_TENSOR_ALTUP_ROUTER_NORM, "blk.%d.altup_router_norm" },
392
- { LLM_TENSOR_LAUREL_L, "blk.%d.laurel_l" },
393
- { LLM_TENSOR_LAUREL_R, "blk.%d.laurel_r" },
394
- { LLM_TENSOR_LAUREL_POST_NORM, "blk.%d.laurel_post_norm" },
395
- { LLM_TENSOR_DENSE_2_OUT, "dense_2" },
396
- { LLM_TENSOR_DENSE_3_OUT, "dense_3" },
397
- { LLM_TENSOR_FFN_NORM_EXPS, "blk.%d.ffn_norm_exps" },
398
- { LLM_TENSOR_ATTN_K_B, "blk.%d.attn_k_b" },
399
- { LLM_TENSOR_ATTN_V_B, "blk.%d.attn_v_b" },
400
- { LLM_TENSOR_NEXTN_EH_PROJ, "blk.%d.nextn.eh_proj" },
401
- { LLM_TENSOR_NEXTN_EMBED_TOKENS, "blk.%d.nextn.embed_tokens" },
402
- { LLM_TENSOR_NEXTN_ENORM, "blk.%d.nextn.enorm" },
403
- { LLM_TENSOR_NEXTN_HNORM, "blk.%d.nextn.hnorm" },
404
- { LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "blk.%d.nextn.shared_head_head" },
405
- { LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "blk.%d.nextn.shared_head_norm" },
406
- { LLM_TENSOR_ATTN_SUB_NORM, "blk.%d.attn_sub_norm" },
407
- { LLM_TENSOR_FFN_SUB_NORM, "blk.%d.ffn_sub_norm" },
408
- { LLM_TENSOR_DEC_OUTPUT_NORM, "dec.output_norm" },
409
- { LLM_TENSOR_DEC_ATTN_NORM, "dec.blk.%d.attn_norm" },
410
- { LLM_TENSOR_DEC_ATTN_Q, "dec.blk.%d.attn_q" },
411
- { LLM_TENSOR_DEC_ATTN_K, "dec.blk.%d.attn_k" },
412
- { LLM_TENSOR_DEC_ATTN_V, "dec.blk.%d.attn_v" },
413
- { LLM_TENSOR_DEC_ATTN_OUT, "dec.blk.%d.attn_o" },
414
- { LLM_TENSOR_DEC_ATTN_REL_B, "dec.blk.%d.attn_rel_b" },
415
- { LLM_TENSOR_DEC_CROSS_ATTN_NORM, "dec.blk.%d.cross_attn_norm" },
416
- { LLM_TENSOR_DEC_CROSS_ATTN_Q, "dec.blk.%d.cross_attn_q" },
417
- { LLM_TENSOR_DEC_CROSS_ATTN_K, "dec.blk.%d.cross_attn_k" },
418
- { LLM_TENSOR_DEC_CROSS_ATTN_V, "dec.blk.%d.cross_attn_v" },
419
- { LLM_TENSOR_DEC_CROSS_ATTN_OUT, "dec.blk.%d.cross_attn_o" },
420
- { LLM_TENSOR_DEC_CROSS_ATTN_REL_B, "dec.blk.%d.cross_attn_rel_b" },
421
- { LLM_TENSOR_DEC_FFN_NORM, "dec.blk.%d.ffn_norm" },
422
- { LLM_TENSOR_DEC_FFN_GATE, "dec.blk.%d.ffn_gate" },
423
- { LLM_TENSOR_DEC_FFN_DOWN, "dec.blk.%d.ffn_down" },
424
- { LLM_TENSOR_DEC_FFN_UP, "dec.blk.%d.ffn_up" },
425
- { LLM_TENSOR_ENC_ATTN_NORM, "enc.blk.%d.attn_norm" },
426
- { LLM_TENSOR_ENC_ATTN_Q, "enc.blk.%d.attn_q" },
427
- { LLM_TENSOR_ENC_ATTN_K, "enc.blk.%d.attn_k" },
428
- { LLM_TENSOR_ENC_ATTN_V, "enc.blk.%d.attn_v" },
429
- { LLM_TENSOR_ENC_ATTN_OUT, "enc.blk.%d.attn_o" },
430
- { LLM_TENSOR_ENC_ATTN_REL_B, "enc.blk.%d.attn_rel_b" },
431
- { LLM_TENSOR_ENC_FFN_NORM, "enc.blk.%d.ffn_norm" },
432
- { LLM_TENSOR_ENC_FFN_GATE, "enc.blk.%d.ffn_gate" },
433
- { LLM_TENSOR_ENC_FFN_DOWN, "enc.blk.%d.ffn_down" },
434
- { LLM_TENSOR_ENC_FFN_UP, "enc.blk.%d.ffn_up" },
435
- { LLM_TENSOR_TIME_MIX_W1, "blk.%d.time_mix_w1" },
436
- { LLM_TENSOR_TIME_MIX_W2, "blk.%d.time_mix_w2" },
437
- { LLM_TENSOR_TIME_MIX_LERP_X, "blk.%d.time_mix_lerp_x" },
438
- { LLM_TENSOR_TIME_MIX_LERP_W, "blk.%d.time_mix_lerp_w" },
439
- { LLM_TENSOR_TIME_MIX_LERP_K, "blk.%d.time_mix_lerp_k" },
440
- { LLM_TENSOR_TIME_MIX_LERP_V, "blk.%d.time_mix_lerp_v" },
441
- { LLM_TENSOR_TIME_MIX_LERP_R, "blk.%d.time_mix_lerp_r" },
442
- { LLM_TENSOR_TIME_MIX_LERP_G, "blk.%d.time_mix_lerp_g" },
443
- { LLM_TENSOR_TIME_MIX_LERP_FUSED, "blk.%d.time_mix_lerp_fused" },
444
- { LLM_TENSOR_TIME_MIX_FIRST, "blk.%d.time_mix_first" },
445
- { LLM_TENSOR_TIME_MIX_DECAY, "blk.%d.time_mix_decay" },
446
- { LLM_TENSOR_TIME_MIX_DECAY_W1, "blk.%d.time_mix_decay_w1" },
447
- { LLM_TENSOR_TIME_MIX_DECAY_W2, "blk.%d.time_mix_decay_w2" },
448
- { LLM_TENSOR_TIME_MIX_KEY, "blk.%d.time_mix_key" },
449
- { LLM_TENSOR_TIME_MIX_VALUE, "blk.%d.time_mix_value" },
450
- { LLM_TENSOR_TIME_MIX_RECEPTANCE, "blk.%d.time_mix_receptance" },
451
- { LLM_TENSOR_TIME_MIX_GATE, "blk.%d.time_mix_gate" },
452
- { LLM_TENSOR_TIME_MIX_LN, "blk.%d.time_mix_ln" },
453
- { LLM_TENSOR_TIME_MIX_OUTPUT, "blk.%d.time_mix_output" },
454
- { LLM_TENSOR_CHANNEL_MIX_LERP_K, "blk.%d.channel_mix_lerp_k" },
455
- { LLM_TENSOR_CHANNEL_MIX_LERP_R, "blk.%d.channel_mix_lerp_r" },
456
- { LLM_TENSOR_CHANNEL_MIX_KEY, "blk.%d.channel_mix_key" },
457
- { LLM_TENSOR_CHANNEL_MIX_VALUE, "blk.%d.channel_mix_value" },
458
- { LLM_TENSOR_CHANNEL_MIX_RECEPTANCE, "blk.%d.channel_mix_receptance" },
459
- { LLM_TENSOR_TIME_MIX_W0, "blk.%d.time_mix_w0" },
460
- { LLM_TENSOR_TIME_MIX_A0, "blk.%d.time_mix_a0" },
461
- { LLM_TENSOR_TIME_MIX_A1, "blk.%d.time_mix_a1" },
462
- { LLM_TENSOR_TIME_MIX_A2, "blk.%d.time_mix_a2" },
463
- { LLM_TENSOR_TIME_MIX_V0, "blk.%d.time_mix_v0" },
464
- { LLM_TENSOR_TIME_MIX_V1, "blk.%d.time_mix_v1" },
465
- { LLM_TENSOR_TIME_MIX_V2, "blk.%d.time_mix_v2" },
466
- { LLM_TENSOR_TIME_MIX_G1, "blk.%d.time_mix_g1" },
467
- { LLM_TENSOR_TIME_MIX_G2, "blk.%d.time_mix_g2" },
468
- { LLM_TENSOR_TIME_MIX_K_K, "blk.%d.time_mix_k_k" },
469
- { LLM_TENSOR_TIME_MIX_K_A, "blk.%d.time_mix_k_a" },
470
- { LLM_TENSOR_TIME_MIX_R_K, "blk.%d.time_mix_r_k" },
471
- { LLM_TENSOR_CONV1D, "conv1d" },
472
- { LLM_TENSOR_CONVNEXT_DW, "convnext.%d.dw" },
473
- { LLM_TENSOR_CONVNEXT_NORM, "convnext.%d.norm" },
474
- { LLM_TENSOR_CONVNEXT_PW1, "convnext.%d.pw1" },
475
- { LLM_TENSOR_CONVNEXT_PW2, "convnext.%d.pw2" },
476
- { LLM_TENSOR_CONVNEXT_GAMMA, "convnext.%d.gamma" },
477
- { LLM_TENSOR_POS_NET_CONV1, "posnet.%d.conv1" },
478
- { LLM_TENSOR_POS_NET_CONV2, "posnet.%d.conv2" },
479
- { LLM_TENSOR_POS_NET_NORM, "posnet.%d.norm" },
480
- { LLM_TENSOR_POS_NET_NORM1, "posnet.%d.norm1" },
481
- { LLM_TENSOR_POS_NET_NORM2, "posnet.%d.norm2" },
482
- { LLM_TENSOR_POS_NET_ATTN_NORM, "posnet.%d.attn_norm" },
483
- { LLM_TENSOR_POS_NET_ATTN_Q, "posnet.%d.attn_q" },
484
- { LLM_TENSOR_POS_NET_ATTN_K, "posnet.%d.attn_k" },
485
- { LLM_TENSOR_POS_NET_ATTN_V, "posnet.%d.attn_v" },
486
- { LLM_TENSOR_POS_NET_ATTN_OUT, "posnet.%d.attn_output" },
487
- { LLM_TENSOR_ATTN_SINKS, "blk.%d.attn_sinks" },
488
- { LLM_TENSOR_SHORTCONV_CONV, "blk.%d.shortconv.conv" },
489
- { LLM_TENSOR_SHORTCONV_INPROJ, "blk.%d.shortconv.in_proj" },
490
- { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
491
- { LLM_TENSOR_FFN_GATE_CHEXPS, "blk.%d.ffn_gate_chexps" },
492
- { LLM_TENSOR_FFN_DOWN_CHEXPS, "blk.%d.ffn_down_chexps" },
493
- { LLM_TENSOR_FFN_UP_CHEXPS, "blk.%d.ffn_up_chexps" },
494
- { LLM_TENSOR_VISEXP_ATTN_QKV, "blk.%d.vis_attn_qkv" },
495
- { LLM_TENSOR_VISEXP_ATTN_OUT, "blk.%d.vis_attn_output" },
496
- { LLM_TENSOR_VISEXP_FFN_GATE, "blk.%d.vis_gate" },
497
- { LLM_TENSOR_VISEXP_FFN_DOWN, "blk.%d.vis_down" },
498
- { LLM_TENSOR_VISEXP_FFN_UP, "blk.%d.vis_up" },
499
- };
500
-
501
- static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
502
- switch (arch) {
503
- case LLM_ARCH_CLIP:
504
- return {};
505
- case LLM_ARCH_LLAMA:
506
- case LLM_ARCH_DECI:
507
- case LLM_ARCH_MISTRAL3:
508
- case LLM_ARCH_LLAMA_EMBED:
509
- return {
510
- LLM_TENSOR_TOKEN_EMBD,
511
- LLM_TENSOR_OUTPUT_NORM,
512
- LLM_TENSOR_OUTPUT,
513
- LLM_TENSOR_ROPE_FREQS,
514
- LLM_TENSOR_ATTN_NORM,
515
- LLM_TENSOR_ATTN_Q,
516
- LLM_TENSOR_ATTN_K,
517
- LLM_TENSOR_ATTN_V,
518
- LLM_TENSOR_ATTN_OUT,
519
- LLM_TENSOR_ATTN_ROT_EMBD,
520
- LLM_TENSOR_FFN_GATE_INP,
521
- LLM_TENSOR_FFN_NORM,
522
- LLM_TENSOR_FFN_GATE,
523
- LLM_TENSOR_FFN_DOWN,
524
- LLM_TENSOR_FFN_UP,
525
- LLM_TENSOR_FFN_GATE_EXP,
526
- LLM_TENSOR_FFN_DOWN_EXP,
527
- LLM_TENSOR_FFN_UP_EXP,
528
- LLM_TENSOR_FFN_GATE_EXPS,
529
- LLM_TENSOR_FFN_DOWN_EXPS,
530
- LLM_TENSOR_FFN_UP_EXPS,
531
- };
532
- case LLM_ARCH_ARCEE:
533
- case LLM_ARCH_STARCODER2:
534
- case LLM_ARCH_NEMOTRON:
535
- return {
536
- LLM_TENSOR_TOKEN_EMBD,
537
- LLM_TENSOR_OUTPUT_NORM,
538
- LLM_TENSOR_OUTPUT,
539
- LLM_TENSOR_ROPE_FREQS,
540
- LLM_TENSOR_ATTN_NORM,
541
- LLM_TENSOR_ATTN_Q,
542
- LLM_TENSOR_ATTN_K,
543
- LLM_TENSOR_ATTN_V,
544
- LLM_TENSOR_ATTN_OUT,
545
- LLM_TENSOR_ATTN_ROT_EMBD,
546
- LLM_TENSOR_FFN_NORM,
547
- LLM_TENSOR_FFN_DOWN,
548
- LLM_TENSOR_FFN_UP,
549
- };
550
- case LLM_ARCH_AFMOE:
551
- return {
552
- LLM_TENSOR_TOKEN_EMBD,
553
- LLM_TENSOR_OUTPUT_NORM,
554
- LLM_TENSOR_OUTPUT,
555
- LLM_TENSOR_ATTN_NORM,
556
- LLM_TENSOR_ATTN_POST_NORM,
557
- LLM_TENSOR_ATTN_Q,
558
- LLM_TENSOR_ATTN_K,
559
- LLM_TENSOR_ATTN_V,
560
- LLM_TENSOR_ATTN_OUT,
561
- LLM_TENSOR_ATTN_Q_NORM,
562
- LLM_TENSOR_ATTN_K_NORM,
563
- LLM_TENSOR_ATTN_GATE,
564
- LLM_TENSOR_FFN_NORM,
565
- LLM_TENSOR_FFN_POST_NORM,
566
- LLM_TENSOR_FFN_GATE_INP,
567
- LLM_TENSOR_FFN_GATE,
568
- LLM_TENSOR_FFN_DOWN,
569
- LLM_TENSOR_FFN_UP,
570
- LLM_TENSOR_FFN_GATE_EXPS,
571
- LLM_TENSOR_FFN_DOWN_EXPS,
572
- LLM_TENSOR_FFN_UP_EXPS,
573
- LLM_TENSOR_FFN_GATE_SHEXP,
574
- LLM_TENSOR_FFN_UP_SHEXP,
575
- LLM_TENSOR_FFN_DOWN_SHEXP,
576
- LLM_TENSOR_FFN_EXP_PROBS_B,
577
- };
578
- case LLM_ARCH_LLAMA4:
579
- return {
580
- LLM_TENSOR_TOKEN_EMBD,
581
- LLM_TENSOR_OUTPUT_NORM,
582
- LLM_TENSOR_OUTPUT,
583
- LLM_TENSOR_ROPE_FREQS,
584
- LLM_TENSOR_ATTN_NORM,
585
- LLM_TENSOR_ATTN_Q,
586
- LLM_TENSOR_ATTN_K,
587
- LLM_TENSOR_ATTN_V,
588
- LLM_TENSOR_ATTN_OUT,
589
- LLM_TENSOR_ATTN_ROT_EMBD,
590
- LLM_TENSOR_FFN_GATE_INP,
591
- LLM_TENSOR_FFN_NORM,
592
- LLM_TENSOR_FFN_GATE,
593
- LLM_TENSOR_FFN_DOWN,
594
- LLM_TENSOR_FFN_UP,
595
- LLM_TENSOR_FFN_GATE_EXP,
596
- LLM_TENSOR_FFN_DOWN_EXP,
597
- LLM_TENSOR_FFN_UP_EXP,
598
- LLM_TENSOR_FFN_GATE_EXPS,
599
- LLM_TENSOR_FFN_DOWN_EXPS,
600
- LLM_TENSOR_FFN_UP_EXPS,
601
- LLM_TENSOR_FFN_GATE_SHEXP,
602
- LLM_TENSOR_FFN_DOWN_SHEXP,
603
- LLM_TENSOR_FFN_UP_SHEXP,
604
- };
605
- case LLM_ARCH_BAICHUAN:
606
- case LLM_ARCH_ORION:
607
- case LLM_ARCH_XVERSE:
608
- case LLM_ARCH_EXAONE:
609
- return {
610
- LLM_TENSOR_TOKEN_EMBD,
611
- LLM_TENSOR_OUTPUT_NORM,
612
- LLM_TENSOR_OUTPUT,
613
- LLM_TENSOR_ROPE_FREQS,
614
- LLM_TENSOR_ATTN_NORM,
615
- LLM_TENSOR_ATTN_Q,
616
- LLM_TENSOR_ATTN_K,
617
- LLM_TENSOR_ATTN_V,
618
- LLM_TENSOR_ATTN_OUT,
619
- LLM_TENSOR_ATTN_ROT_EMBD,
620
- LLM_TENSOR_FFN_NORM,
621
- LLM_TENSOR_FFN_GATE,
622
- LLM_TENSOR_FFN_DOWN,
623
- LLM_TENSOR_FFN_UP,
624
- };
625
- case LLM_ARCH_FALCON:
626
- return {
627
- LLM_TENSOR_TOKEN_EMBD,
628
- LLM_TENSOR_OUTPUT_NORM,
629
- LLM_TENSOR_OUTPUT,
630
- LLM_TENSOR_ATTN_NORM,
631
- LLM_TENSOR_ATTN_NORM_2,
632
- LLM_TENSOR_ATTN_QKV,
633
- LLM_TENSOR_ATTN_OUT,
634
- LLM_TENSOR_FFN_DOWN,
635
- LLM_TENSOR_FFN_UP,
636
- };
637
- case LLM_ARCH_GROK:
638
- return {
639
- LLM_TENSOR_TOKEN_EMBD,
640
- LLM_TENSOR_OUTPUT_NORM,
641
- LLM_TENSOR_OUTPUT,
642
- LLM_TENSOR_ROPE_FREQS,
643
- LLM_TENSOR_ATTN_NORM,
644
- LLM_TENSOR_ATTN_Q,
645
- LLM_TENSOR_ATTN_K,
646
- LLM_TENSOR_ATTN_V,
647
- LLM_TENSOR_ATTN_OUT,
648
- LLM_TENSOR_ATTN_ROT_EMBD,
649
- LLM_TENSOR_FFN_GATE_INP,
650
- LLM_TENSOR_FFN_NORM,
651
- LLM_TENSOR_FFN_GATE,
652
- LLM_TENSOR_FFN_DOWN,
653
- LLM_TENSOR_FFN_UP,
654
- LLM_TENSOR_FFN_GATE_EXP,
655
- LLM_TENSOR_FFN_DOWN_EXP,
656
- LLM_TENSOR_FFN_UP_EXP,
657
- LLM_TENSOR_FFN_GATE_EXPS,
658
- LLM_TENSOR_FFN_DOWN_EXPS,
659
- LLM_TENSOR_FFN_UP_EXPS,
660
- LLM_TENSOR_FFN_POST_NORM,
661
- LLM_TENSOR_LAYER_OUT_NORM,
662
- LLM_TENSOR_ATTN_OUT_NORM,
663
- };
664
- case LLM_ARCH_GPT2:
665
- case LLM_ARCH_STARCODER:
666
- return {
667
- LLM_TENSOR_TOKEN_EMBD,
668
- LLM_TENSOR_POS_EMBD,
669
- LLM_TENSOR_OUTPUT_NORM,
670
- LLM_TENSOR_OUTPUT,
671
- LLM_TENSOR_ATTN_NORM,
672
- LLM_TENSOR_ATTN_QKV,
673
- LLM_TENSOR_ATTN_OUT,
674
- LLM_TENSOR_FFN_NORM,
675
- LLM_TENSOR_FFN_UP,
676
- LLM_TENSOR_FFN_DOWN,
677
- };
678
- case LLM_ARCH_GPTNEOX:
679
- return {
680
- LLM_TENSOR_TOKEN_EMBD,
681
- LLM_TENSOR_OUTPUT_NORM,
682
- LLM_TENSOR_OUTPUT,
683
- LLM_TENSOR_ATTN_NORM,
684
- LLM_TENSOR_ATTN_QKV,
685
- LLM_TENSOR_ATTN_OUT,
686
- LLM_TENSOR_FFN_NORM,
687
- LLM_TENSOR_FFN_DOWN,
688
- LLM_TENSOR_FFN_UP,
689
- };
690
- case LLM_ARCH_MPT:
691
- return {
692
- LLM_TENSOR_TOKEN_EMBD,
693
- LLM_TENSOR_OUTPUT_NORM,
694
- LLM_TENSOR_OUTPUT,
695
- LLM_TENSOR_ATTN_NORM,
696
- LLM_TENSOR_FFN_NORM,
697
- LLM_TENSOR_ATTN_QKV,
698
- LLM_TENSOR_ATTN_OUT,
699
- LLM_TENSOR_FFN_DOWN,
700
- LLM_TENSOR_FFN_UP,
701
- LLM_TENSOR_FFN_ACT,
702
- LLM_TENSOR_POS_EMBD,
703
- LLM_TENSOR_ATTN_Q_NORM,
704
- LLM_TENSOR_ATTN_K_NORM,
705
- };
706
- case LLM_ARCH_REFACT:
707
- case LLM_ARCH_QWEN2:
708
- case LLM_ARCH_QWEN2VL:
709
- case LLM_ARCH_INTERNLM2:
710
- case LLM_ARCH_GRANITE:
711
- case LLM_ARCH_ERNIE4_5:
712
- case LLM_ARCH_SMOLLM3:
713
- case LLM_ARCH_DREAM:
714
- case LLM_ARCH_LLADA:
715
- case LLM_ARCH_PANGU_EMBED:
716
- return {
717
- LLM_TENSOR_TOKEN_EMBD,
718
- LLM_TENSOR_OUTPUT_NORM,
719
- LLM_TENSOR_OUTPUT,
720
- LLM_TENSOR_ATTN_NORM,
721
- LLM_TENSOR_ATTN_Q,
722
- LLM_TENSOR_ATTN_K,
723
- LLM_TENSOR_ATTN_V,
724
- LLM_TENSOR_ATTN_OUT,
725
- LLM_TENSOR_FFN_NORM,
726
- LLM_TENSOR_FFN_GATE,
727
- LLM_TENSOR_FFN_DOWN,
728
- LLM_TENSOR_FFN_UP,
729
- };
730
- case LLM_ARCH_BERT:
731
- return {
732
- LLM_TENSOR_TOKEN_EMBD,
733
- LLM_TENSOR_TOKEN_EMBD_NORM,
734
- LLM_TENSOR_TOKEN_TYPES,
735
- LLM_TENSOR_POS_EMBD,
736
- LLM_TENSOR_ATTN_OUT_NORM,
737
- LLM_TENSOR_ATTN_QKV,
738
- LLM_TENSOR_ATTN_Q,
739
- LLM_TENSOR_ATTN_K,
740
- LLM_TENSOR_ATTN_V,
741
- LLM_TENSOR_ATTN_OUT,
742
- LLM_TENSOR_LAYER_OUT_NORM,
743
- LLM_TENSOR_FFN_DOWN,
744
- LLM_TENSOR_FFN_UP,
745
- LLM_TENSOR_CLS,
746
- LLM_TENSOR_CLS_OUT,
747
- };
748
- case LLM_ARCH_NOMIC_BERT:
749
- return {
750
- LLM_TENSOR_TOKEN_EMBD,
751
- LLM_TENSOR_TOKEN_EMBD_NORM,
752
- LLM_TENSOR_TOKEN_TYPES,
753
- LLM_TENSOR_ATTN_OUT_NORM,
754
- LLM_TENSOR_ATTN_QKV,
755
- LLM_TENSOR_ATTN_OUT,
756
- LLM_TENSOR_LAYER_OUT_NORM,
757
- LLM_TENSOR_FFN_GATE,
758
- LLM_TENSOR_FFN_DOWN,
759
- LLM_TENSOR_FFN_UP,
760
- };
761
- case LLM_ARCH_NOMIC_BERT_MOE:
762
- return {
763
- LLM_TENSOR_TOKEN_EMBD,
764
- LLM_TENSOR_TOKEN_EMBD_NORM,
765
- LLM_TENSOR_TOKEN_TYPES,
766
- LLM_TENSOR_ATTN_OUT_NORM,
767
- LLM_TENSOR_ATTN_QKV,
768
- LLM_TENSOR_ATTN_OUT,
769
- LLM_TENSOR_LAYER_OUT_NORM,
770
- LLM_TENSOR_FFN_GATE,
771
- LLM_TENSOR_FFN_DOWN,
772
- LLM_TENSOR_FFN_UP,
773
- LLM_TENSOR_FFN_GATE_INP,
774
- LLM_TENSOR_FFN_DOWN_EXPS,
775
- LLM_TENSOR_FFN_UP_EXPS,
776
- };
777
- case LLM_ARCH_NEO_BERT:
778
- return {
779
- LLM_TENSOR_TOKEN_EMBD,
780
- LLM_TENSOR_ATTN_NORM,
781
- LLM_TENSOR_ATTN_QKV,
782
- LLM_TENSOR_ATTN_OUT,
783
- LLM_TENSOR_FFN_NORM,
784
- LLM_TENSOR_FFN_DOWN,
785
- LLM_TENSOR_FFN_UP,
786
- LLM_TENSOR_ENC_OUTPUT_NORM,
787
- LLM_TENSOR_CLS,
788
- LLM_TENSOR_CLS_OUT,
789
- };
790
- case LLM_ARCH_MODERN_BERT:
791
- return {
792
- LLM_TENSOR_TOKEN_EMBD,
793
- LLM_TENSOR_TOKEN_EMBD_NORM,
794
- LLM_TENSOR_OUTPUT_NORM,
795
- LLM_TENSOR_ATTN_NORM,
796
- LLM_TENSOR_ATTN_OUT,
797
- LLM_TENSOR_ATTN_QKV,
798
- LLM_TENSOR_FFN_DOWN,
799
- LLM_TENSOR_FFN_UP,
800
- LLM_TENSOR_FFN_NORM,
801
- LLM_TENSOR_CLS,
802
- LLM_TENSOR_CLS_OUT,
803
- };
804
- case LLM_ARCH_JINA_BERT_V2:
805
- return {
806
- LLM_TENSOR_TOKEN_EMBD,
807
- LLM_TENSOR_TOKEN_EMBD_NORM,
808
- LLM_TENSOR_TOKEN_TYPES,
809
- LLM_TENSOR_ATTN_NORM_2,
810
- LLM_TENSOR_ATTN_OUT_NORM,
811
- LLM_TENSOR_ATTN_Q,
812
- LLM_TENSOR_ATTN_Q_NORM,
813
- LLM_TENSOR_ATTN_K,
814
- LLM_TENSOR_ATTN_K_NORM,
815
- LLM_TENSOR_ATTN_V,
816
- LLM_TENSOR_ATTN_OUT,
817
- LLM_TENSOR_LAYER_OUT_NORM,
818
- LLM_TENSOR_FFN_DOWN,
819
- LLM_TENSOR_FFN_GATE,
820
- LLM_TENSOR_FFN_UP,
821
- LLM_TENSOR_CLS,
822
- };
823
- case LLM_ARCH_JINA_BERT_V3:
824
- return {
825
- LLM_TENSOR_TOKEN_EMBD,
826
- LLM_TENSOR_TOKEN_EMBD_NORM,
827
- LLM_TENSOR_TOKEN_TYPES,
828
- LLM_TENSOR_ATTN_OUT_NORM,
829
- LLM_TENSOR_ATTN_QKV,
830
- LLM_TENSOR_ATTN_OUT,
831
- LLM_TENSOR_FFN_DOWN,
832
- LLM_TENSOR_FFN_UP,
833
- LLM_TENSOR_LAYER_OUT_NORM,
834
- };
835
- case LLM_ARCH_BLOOM:
836
- return {
837
- LLM_TENSOR_TOKEN_EMBD,
838
- LLM_TENSOR_TOKEN_EMBD_NORM,
839
- LLM_TENSOR_OUTPUT_NORM,
840
- LLM_TENSOR_OUTPUT,
841
- LLM_TENSOR_ATTN_NORM,
842
- LLM_TENSOR_ATTN_QKV,
843
- LLM_TENSOR_ATTN_OUT,
844
- LLM_TENSOR_FFN_NORM,
845
- LLM_TENSOR_FFN_UP,
846
- LLM_TENSOR_FFN_DOWN,
847
- };
848
- case LLM_ARCH_STABLELM:
849
- return {
850
- LLM_TENSOR_TOKEN_EMBD,
851
- LLM_TENSOR_OUTPUT_NORM,
852
- LLM_TENSOR_OUTPUT,
853
- LLM_TENSOR_ROPE_FREQS,
854
- LLM_TENSOR_ATTN_NORM,
855
- LLM_TENSOR_ATTN_Q,
856
- LLM_TENSOR_ATTN_K,
857
- LLM_TENSOR_ATTN_V,
858
- LLM_TENSOR_ATTN_OUT,
859
- LLM_TENSOR_FFN_NORM,
860
- LLM_TENSOR_FFN_GATE,
861
- LLM_TENSOR_FFN_DOWN,
862
- LLM_TENSOR_FFN_UP,
863
- LLM_TENSOR_ATTN_Q_NORM,
864
- LLM_TENSOR_ATTN_K_NORM,
865
- };
866
- case LLM_ARCH_QWEN:
867
- return {
868
- LLM_TENSOR_TOKEN_EMBD,
869
- LLM_TENSOR_OUTPUT_NORM,
870
- LLM_TENSOR_OUTPUT,
871
- LLM_TENSOR_ROPE_FREQS,
872
- LLM_TENSOR_ATTN_NORM,
873
- LLM_TENSOR_ATTN_QKV,
874
- LLM_TENSOR_ATTN_OUT,
875
- LLM_TENSOR_FFN_NORM,
876
- LLM_TENSOR_FFN_GATE,
877
- LLM_TENSOR_FFN_DOWN,
878
- LLM_TENSOR_FFN_UP,
879
- };
880
- case LLM_ARCH_QWEN2MOE:
881
- return {
882
- LLM_TENSOR_TOKEN_EMBD,
883
- LLM_TENSOR_OUTPUT_NORM,
884
- LLM_TENSOR_OUTPUT,
885
- LLM_TENSOR_ATTN_NORM,
886
- LLM_TENSOR_ATTN_Q,
887
- LLM_TENSOR_ATTN_K,
888
- LLM_TENSOR_ATTN_V,
889
- LLM_TENSOR_ATTN_OUT,
890
- LLM_TENSOR_FFN_NORM,
891
- LLM_TENSOR_FFN_GATE_INP,
892
- LLM_TENSOR_FFN_GATE_EXPS,
893
- LLM_TENSOR_FFN_DOWN_EXPS,
894
- LLM_TENSOR_FFN_UP_EXPS,
895
- LLM_TENSOR_FFN_GATE_INP_SHEXP,
896
- LLM_TENSOR_FFN_GATE_SHEXP,
897
- LLM_TENSOR_FFN_DOWN_SHEXP,
898
- LLM_TENSOR_FFN_UP_SHEXP,
899
- };
900
- case LLM_ARCH_QWEN3:
901
- return {
902
- LLM_TENSOR_TOKEN_EMBD,
903
- LLM_TENSOR_OUTPUT_NORM,
904
- LLM_TENSOR_OUTPUT,
905
- LLM_TENSOR_CLS_OUT,
906
- LLM_TENSOR_ATTN_NORM,
907
- LLM_TENSOR_ATTN_Q,
908
- LLM_TENSOR_ATTN_Q_NORM,
909
- LLM_TENSOR_ATTN_K,
910
- LLM_TENSOR_ATTN_K_NORM,
911
- LLM_TENSOR_ATTN_V,
912
- LLM_TENSOR_ATTN_OUT,
913
- LLM_TENSOR_FFN_NORM,
914
- LLM_TENSOR_FFN_GATE,
915
- LLM_TENSOR_FFN_DOWN,
916
- LLM_TENSOR_FFN_UP,
917
- };
918
- case LLM_ARCH_QWEN3MOE:
919
- case LLM_ARCH_QWEN3VLMOE:
920
- case LLM_ARCH_OLMOE:
921
- case LLM_ARCH_LLADA_MOE:
922
- case LLM_ARCH_RND1:
923
- return {
924
- LLM_TENSOR_TOKEN_EMBD,
925
- LLM_TENSOR_OUTPUT_NORM,
926
- LLM_TENSOR_OUTPUT,
927
- LLM_TENSOR_ATTN_NORM,
928
- LLM_TENSOR_ATTN_Q,
929
- LLM_TENSOR_ATTN_Q_NORM,
930
- LLM_TENSOR_ATTN_K,
931
- LLM_TENSOR_ATTN_K_NORM,
932
- LLM_TENSOR_ATTN_V,
933
- LLM_TENSOR_ATTN_OUT,
934
- LLM_TENSOR_FFN_NORM,
935
- LLM_TENSOR_FFN_GATE_INP,
936
- LLM_TENSOR_FFN_GATE_EXPS,
937
- LLM_TENSOR_FFN_DOWN_EXPS,
938
- LLM_TENSOR_FFN_UP_EXPS,
939
- };
940
- case LLM_ARCH_QWEN3NEXT:
941
- return {
942
- LLM_TENSOR_TOKEN_EMBD,
943
- LLM_TENSOR_OUTPUT_NORM,
944
- LLM_TENSOR_OUTPUT,
945
- LLM_TENSOR_ATTN_NORM,
946
- LLM_TENSOR_ATTN_POST_NORM,
947
- LLM_TENSOR_ATTN_Q,
948
- LLM_TENSOR_ATTN_Q_NORM,
949
- LLM_TENSOR_ATTN_K,
950
- LLM_TENSOR_ATTN_K_NORM,
951
- LLM_TENSOR_ATTN_V,
952
- LLM_TENSOR_ATTN_OUT,
953
- LLM_TENSOR_ATTN_QKV,
954
- LLM_TENSOR_ATTN_GATE,
955
- LLM_TENSOR_FFN_NORM,
956
- LLM_TENSOR_FFN_GATE_INP,
957
- LLM_TENSOR_FFN_GATE_EXPS,
958
- LLM_TENSOR_FFN_DOWN_EXPS,
959
- LLM_TENSOR_FFN_UP_EXPS,
960
- LLM_TENSOR_FFN_GATE_INP_SHEXP,
961
- LLM_TENSOR_FFN_GATE_SHEXP,
962
- LLM_TENSOR_FFN_DOWN_SHEXP,
963
- LLM_TENSOR_FFN_UP_SHEXP,
964
- LLM_TENSOR_SSM_A_NOSCAN,
965
- LLM_TENSOR_SSM_CONV1D,
966
- LLM_TENSOR_SSM_DT,
967
- LLM_TENSOR_SSM_BETA_ALPHA,
968
- LLM_TENSOR_SSM_IN,
969
- LLM_TENSOR_SSM_NORM,
970
- LLM_TENSOR_SSM_OUT,
971
- };
972
- case LLM_ARCH_QWEN3VL:
973
- case LLM_ARCH_CHAMELEON:
974
- case LLM_ARCH_HUNYUAN_DENSE:
975
- return {
976
- LLM_TENSOR_TOKEN_EMBD,
977
- LLM_TENSOR_OUTPUT_NORM,
978
- LLM_TENSOR_OUTPUT,
979
- LLM_TENSOR_ATTN_NORM,
980
- LLM_TENSOR_ATTN_Q,
981
- LLM_TENSOR_ATTN_Q_NORM,
982
- LLM_TENSOR_ATTN_K,
983
- LLM_TENSOR_ATTN_K_NORM,
984
- LLM_TENSOR_ATTN_V,
985
- LLM_TENSOR_ATTN_OUT,
986
- LLM_TENSOR_FFN_NORM,
987
- LLM_TENSOR_FFN_GATE,
988
- LLM_TENSOR_FFN_DOWN,
989
- LLM_TENSOR_FFN_UP,
990
- };
991
- case LLM_ARCH_PHI2:
992
- return {
993
- LLM_TENSOR_TOKEN_EMBD,
994
- LLM_TENSOR_OUTPUT_NORM,
995
- LLM_TENSOR_OUTPUT,
996
- LLM_TENSOR_ATTN_NORM,
997
- LLM_TENSOR_ATTN_QKV,
998
- LLM_TENSOR_ATTN_Q,
999
- LLM_TENSOR_ATTN_K,
1000
- LLM_TENSOR_ATTN_V,
1001
- LLM_TENSOR_ATTN_OUT,
1002
- LLM_TENSOR_FFN_DOWN,
1003
- LLM_TENSOR_FFN_UP,
1004
- };
1005
- case LLM_ARCH_PHI3:
1006
- return {
1007
- LLM_TENSOR_TOKEN_EMBD,
1008
- LLM_TENSOR_OUTPUT_NORM,
1009
- LLM_TENSOR_OUTPUT,
1010
- LLM_TENSOR_ROPE_FACTORS_LONG,
1011
- LLM_TENSOR_ROPE_FACTORS_SHORT,
1012
- LLM_TENSOR_ATTN_NORM,
1013
- LLM_TENSOR_ATTN_QKV,
1014
- LLM_TENSOR_ATTN_Q,
1015
- LLM_TENSOR_ATTN_K,
1016
- LLM_TENSOR_ATTN_V,
1017
- LLM_TENSOR_ATTN_OUT,
1018
- LLM_TENSOR_FFN_NORM,
1019
- LLM_TENSOR_FFN_DOWN,
1020
- LLM_TENSOR_FFN_UP,
1021
- };
1022
- case LLM_ARCH_PHIMOE:
1023
- return {
1024
- LLM_TENSOR_TOKEN_EMBD,
1025
- LLM_TENSOR_OUTPUT_NORM,
1026
- LLM_TENSOR_OUTPUT,
1027
- LLM_TENSOR_ROPE_FACTORS_LONG,
1028
- LLM_TENSOR_ROPE_FACTORS_SHORT,
1029
- LLM_TENSOR_ATTN_NORM,
1030
- LLM_TENSOR_ATTN_QKV,
1031
- LLM_TENSOR_ATTN_Q,
1032
- LLM_TENSOR_ATTN_K,
1033
- LLM_TENSOR_ATTN_V,
1034
- LLM_TENSOR_ATTN_OUT,
1035
- LLM_TENSOR_FFN_NORM,
1036
- LLM_TENSOR_FFN_GATE_INP,
1037
- LLM_TENSOR_FFN_GATE_EXPS,
1038
- LLM_TENSOR_FFN_DOWN_EXPS,
1039
- LLM_TENSOR_FFN_UP_EXPS,
1040
- };
1041
- case LLM_ARCH_PLAMO:
1042
- return {
1043
- LLM_TENSOR_TOKEN_EMBD,
1044
- LLM_TENSOR_OUTPUT_NORM,
1045
- LLM_TENSOR_OUTPUT,
1046
- LLM_TENSOR_ROPE_FREQS,
1047
- LLM_TENSOR_ATTN_NORM,
1048
- LLM_TENSOR_ATTN_Q,
1049
- LLM_TENSOR_ATTN_K,
1050
- LLM_TENSOR_ATTN_V,
1051
- LLM_TENSOR_ATTN_OUT,
1052
- LLM_TENSOR_ATTN_ROT_EMBD,
1053
- LLM_TENSOR_FFN_GATE,
1054
- LLM_TENSOR_FFN_DOWN,
1055
- LLM_TENSOR_FFN_UP,
1056
- };
1057
- case LLM_ARCH_PLAMO2:
1058
- return {
1059
- LLM_TENSOR_TOKEN_EMBD,
1060
- LLM_TENSOR_OUTPUT_NORM,
1061
- LLM_TENSOR_OUTPUT,
1062
- LLM_TENSOR_ROPE_FREQS,
1063
- LLM_TENSOR_ATTN_NORM,
1064
- LLM_TENSOR_ATTN_QKV,
1065
- LLM_TENSOR_ATTN_Q_NORM,
1066
- LLM_TENSOR_ATTN_K_NORM,
1067
- LLM_TENSOR_ATTN_OUT,
1068
- LLM_TENSOR_ATTN_ROT_EMBD,
1069
- LLM_TENSOR_FFN_NORM,
1070
- LLM_TENSOR_FFN_DOWN,
1071
- LLM_TENSOR_FFN_UP,
1072
- LLM_TENSOR_SSM_IN,
1073
- LLM_TENSOR_SSM_CONV1D,
1074
- LLM_TENSOR_SSM_X,
1075
- LLM_TENSOR_SSM_DT,
1076
- LLM_TENSOR_SSM_A,
1077
- LLM_TENSOR_SSM_D,
1078
- LLM_TENSOR_SSM_OUT,
1079
- LLM_TENSOR_SSM_DT_NORM,
1080
- LLM_TENSOR_SSM_B_NORM,
1081
- LLM_TENSOR_SSM_C_NORM,
1082
- LLM_TENSOR_ATTN_POST_NORM,
1083
- LLM_TENSOR_FFN_POST_NORM,
1084
- };
1085
- case LLM_ARCH_PLAMO3:
1086
- return {
1087
- LLM_TENSOR_TOKEN_EMBD,
1088
- LLM_TENSOR_OUTPUT_NORM,
1089
- LLM_TENSOR_OUTPUT,
1090
- LLM_TENSOR_ATTN_NORM,
1091
- LLM_TENSOR_ATTN_QKV,
1092
- LLM_TENSOR_ATTN_Q_NORM,
1093
- LLM_TENSOR_ATTN_K_NORM,
1094
- LLM_TENSOR_ATTN_OUT,
1095
- LLM_TENSOR_ATTN_POST_NORM,
1096
- LLM_TENSOR_FFN_NORM,
1097
- LLM_TENSOR_FFN_POST_NORM,
1098
- LLM_TENSOR_FFN_DOWN,
1099
- LLM_TENSOR_FFN_UP,
1100
- };
1101
- case LLM_ARCH_CODESHELL:
1102
- return {
1103
- LLM_TENSOR_TOKEN_EMBD,
1104
- LLM_TENSOR_OUTPUT_NORM,
1105
- LLM_TENSOR_OUTPUT,
1106
- LLM_TENSOR_ROPE_FREQS,
1107
- LLM_TENSOR_ATTN_NORM,
1108
- LLM_TENSOR_ATTN_Q,
1109
- LLM_TENSOR_ATTN_K,
1110
- LLM_TENSOR_ATTN_V,
1111
- LLM_TENSOR_ATTN_QKV,
1112
- LLM_TENSOR_ATTN_OUT,
1113
- LLM_TENSOR_ATTN_ROT_EMBD,
1114
- LLM_TENSOR_FFN_NORM,
1115
- LLM_TENSOR_FFN_GATE,
1116
- LLM_TENSOR_FFN_DOWN,
1117
- LLM_TENSOR_FFN_UP,
1118
- };
1119
- case LLM_ARCH_MINICPM:
1120
- return {
1121
- LLM_TENSOR_TOKEN_EMBD,
1122
- LLM_TENSOR_OUTPUT_NORM,
1123
- LLM_TENSOR_OUTPUT,
1124
- LLM_TENSOR_ROPE_FREQS,
1125
- LLM_TENSOR_ROPE_FACTORS_LONG,
1126
- LLM_TENSOR_ROPE_FACTORS_SHORT,
1127
- LLM_TENSOR_ATTN_NORM,
1128
- LLM_TENSOR_ATTN_Q,
1129
- LLM_TENSOR_ATTN_K,
1130
- LLM_TENSOR_ATTN_V,
1131
- LLM_TENSOR_ATTN_OUT,
1132
- LLM_TENSOR_ATTN_ROT_EMBD,
1133
- LLM_TENSOR_FFN_GATE_INP,
1134
- LLM_TENSOR_FFN_NORM,
1135
- LLM_TENSOR_FFN_GATE,
1136
- LLM_TENSOR_FFN_DOWN,
1137
- LLM_TENSOR_FFN_UP,
1138
- LLM_TENSOR_FFN_GATE_EXP,
1139
- LLM_TENSOR_FFN_DOWN_EXP,
1140
- LLM_TENSOR_FFN_UP_EXP,
1141
- };
1142
- case LLM_ARCH_MINICPM3:
1143
- return {
1144
- LLM_TENSOR_TOKEN_EMBD,
1145
- LLM_TENSOR_OUTPUT_NORM,
1146
- LLM_TENSOR_OUTPUT,
1147
- LLM_TENSOR_ROPE_FACTORS_LONG,
1148
- LLM_TENSOR_ROPE_FACTORS_SHORT,
1149
- LLM_TENSOR_ATTN_NORM,
1150
- LLM_TENSOR_ATTN_Q_A_NORM,
1151
- LLM_TENSOR_ATTN_KV_A_NORM,
1152
- LLM_TENSOR_ATTN_Q,
1153
- LLM_TENSOR_ATTN_Q_A,
1154
- LLM_TENSOR_ATTN_Q_B,
1155
- LLM_TENSOR_ATTN_KV_A_MQA,
1156
- LLM_TENSOR_ATTN_KV_B,
1157
- LLM_TENSOR_ATTN_OUT,
1158
- LLM_TENSOR_FFN_NORM,
1159
- LLM_TENSOR_FFN_GATE,
1160
- LLM_TENSOR_FFN_UP,
1161
- LLM_TENSOR_FFN_DOWN,
1162
- };
1163
- case LLM_ARCH_GEMMA:
1164
- return {
1165
- LLM_TENSOR_TOKEN_EMBD,
1166
- LLM_TENSOR_OUTPUT_NORM,
1167
- LLM_TENSOR_ATTN_NORM,
1168
- LLM_TENSOR_ATTN_Q,
1169
- LLM_TENSOR_ATTN_K,
1170
- LLM_TENSOR_ATTN_V,
1171
- LLM_TENSOR_ATTN_OUT,
1172
- LLM_TENSOR_FFN_NORM,
1173
- LLM_TENSOR_FFN_GATE,
1174
- LLM_TENSOR_FFN_DOWN,
1175
- LLM_TENSOR_FFN_UP,
1176
- };
1177
- case LLM_ARCH_GEMMA2:
1178
- return {
1179
- LLM_TENSOR_TOKEN_EMBD,
1180
- LLM_TENSOR_OUTPUT_NORM,
1181
- LLM_TENSOR_ATTN_NORM,
1182
- LLM_TENSOR_ATTN_Q,
1183
- LLM_TENSOR_ATTN_K,
1184
- LLM_TENSOR_ATTN_V,
1185
- LLM_TENSOR_ATTN_OUT,
1186
- LLM_TENSOR_ATTN_POST_NORM,
1187
- LLM_TENSOR_FFN_NORM,
1188
- LLM_TENSOR_FFN_GATE,
1189
- LLM_TENSOR_FFN_DOWN,
1190
- LLM_TENSOR_FFN_UP,
1191
- LLM_TENSOR_FFN_POST_NORM,
1192
- };
1193
- case LLM_ARCH_GEMMA3:
1194
- return {
1195
- LLM_TENSOR_TOKEN_EMBD,
1196
- LLM_TENSOR_OUTPUT_NORM,
1197
- LLM_TENSOR_OUTPUT,
1198
- LLM_TENSOR_ATTN_NORM,
1199
- LLM_TENSOR_ATTN_Q,
1200
- LLM_TENSOR_ATTN_Q_NORM,
1201
- LLM_TENSOR_ATTN_K,
1202
- LLM_TENSOR_ATTN_K_NORM,
1203
- LLM_TENSOR_ATTN_V,
1204
- LLM_TENSOR_ATTN_OUT,
1205
- LLM_TENSOR_ATTN_POST_NORM,
1206
- LLM_TENSOR_FFN_NORM,
1207
- LLM_TENSOR_FFN_GATE,
1208
- LLM_TENSOR_FFN_DOWN,
1209
- LLM_TENSOR_FFN_UP,
1210
- LLM_TENSOR_FFN_POST_NORM,
1211
- };
1212
- case LLM_ARCH_GEMMA3N:
1213
- return {
1214
- LLM_TENSOR_TOKEN_EMBD,
1215
- LLM_TENSOR_OUTPUT_NORM,
1216
- LLM_TENSOR_ATTN_NORM,
1217
- LLM_TENSOR_ATTN_Q,
1218
- LLM_TENSOR_ATTN_Q_NORM,
1219
- LLM_TENSOR_ATTN_K,
1220
- LLM_TENSOR_ATTN_K_NORM,
1221
- LLM_TENSOR_ATTN_V,
1222
- LLM_TENSOR_ATTN_OUT,
1223
- LLM_TENSOR_ATTN_POST_NORM,
1224
- LLM_TENSOR_FFN_NORM,
1225
- LLM_TENSOR_FFN_GATE,
1226
- LLM_TENSOR_FFN_DOWN,
1227
- LLM_TENSOR_FFN_UP,
1228
- LLM_TENSOR_FFN_POST_NORM,
1229
- LLM_TENSOR_PER_LAYER_TOKEN_EMBD,
1230
- LLM_TENSOR_PER_LAYER_MODEL_PROJ,
1231
- LLM_TENSOR_PER_LAYER_PROJ_NORM,
1232
- LLM_TENSOR_ALTUP_UNEMBD_PROJ,
1233
- LLM_TENSOR_ALTUP_PROJ,
1234
- LLM_TENSOR_PER_LAYER_INP_GATE,
1235
- LLM_TENSOR_PER_LAYER_PROJ,
1236
- LLM_TENSOR_PER_LAYER_POST_NORM,
1237
- LLM_TENSOR_ALTUP_CORRECT_COEF,
1238
- LLM_TENSOR_ALTUP_CORRECT_SCALE,
1239
- LLM_TENSOR_ALTUP_PREDICT_COEF,
1240
- LLM_TENSOR_ALTUP_ROUTER,
1241
- LLM_TENSOR_ALTUP_ROUTER_NORM,
1242
- LLM_TENSOR_LAUREL_L,
1243
- LLM_TENSOR_LAUREL_R,
1244
- LLM_TENSOR_LAUREL_POST_NORM,
1245
- };
1246
- case LLM_ARCH_GEMMA_EMBEDDING:
1247
- return {
1248
- LLM_TENSOR_TOKEN_EMBD,
1249
- LLM_TENSOR_OUTPUT_NORM,
1250
- LLM_TENSOR_OUTPUT,
1251
- LLM_TENSOR_DENSE_2_OUT,
1252
- LLM_TENSOR_DENSE_3_OUT,
1253
- LLM_TENSOR_ATTN_NORM,
1254
- LLM_TENSOR_ATTN_Q,
1255
- LLM_TENSOR_ATTN_Q_NORM,
1256
- LLM_TENSOR_ATTN_K,
1257
- LLM_TENSOR_ATTN_K_NORM,
1258
- LLM_TENSOR_ATTN_V,
1259
- LLM_TENSOR_ATTN_OUT,
1260
- LLM_TENSOR_ATTN_POST_NORM,
1261
- LLM_TENSOR_FFN_NORM,
1262
- LLM_TENSOR_FFN_GATE,
1263
- LLM_TENSOR_FFN_DOWN,
1264
- LLM_TENSOR_FFN_UP,
1265
- LLM_TENSOR_FFN_POST_NORM,
1266
- };
1267
- case LLM_ARCH_MAMBA:
1268
- return {
1269
- LLM_TENSOR_TOKEN_EMBD,
1270
- LLM_TENSOR_OUTPUT_NORM,
1271
- LLM_TENSOR_OUTPUT,
1272
- LLM_TENSOR_ATTN_NORM,
1273
- LLM_TENSOR_SSM_IN,
1274
- LLM_TENSOR_SSM_CONV1D,
1275
- LLM_TENSOR_SSM_X,
1276
- LLM_TENSOR_SSM_DT,
1277
- LLM_TENSOR_SSM_A,
1278
- LLM_TENSOR_SSM_D,
1279
- LLM_TENSOR_SSM_OUT,
1280
- };
1281
- case LLM_ARCH_MAMBA2:
1282
- return {
1283
- LLM_TENSOR_TOKEN_EMBD,
1284
- LLM_TENSOR_OUTPUT_NORM,
1285
- LLM_TENSOR_OUTPUT,
1286
- LLM_TENSOR_ATTN_NORM,
1287
- LLM_TENSOR_SSM_IN,
1288
- LLM_TENSOR_SSM_CONV1D,
1289
- LLM_TENSOR_SSM_DT,
1290
- LLM_TENSOR_SSM_A,
1291
- LLM_TENSOR_SSM_D,
1292
- LLM_TENSOR_SSM_NORM,
1293
- LLM_TENSOR_SSM_OUT,
1294
- };
1295
- case LLM_ARCH_JAMBA:
1296
- return {
1297
- LLM_TENSOR_TOKEN_EMBD,
1298
- LLM_TENSOR_OUTPUT_NORM,
1299
- LLM_TENSOR_OUTPUT,
1300
- LLM_TENSOR_ATTN_NORM,
1301
- LLM_TENSOR_SSM_IN,
1302
- LLM_TENSOR_SSM_CONV1D,
1303
- LLM_TENSOR_SSM_X,
1304
- LLM_TENSOR_SSM_DT,
1305
- LLM_TENSOR_SSM_DT_NORM,
1306
- LLM_TENSOR_SSM_A,
1307
- LLM_TENSOR_SSM_B_NORM,
1308
- LLM_TENSOR_SSM_C_NORM,
1309
- LLM_TENSOR_SSM_D,
1310
- LLM_TENSOR_SSM_OUT,
1311
- LLM_TENSOR_ATTN_Q,
1312
- LLM_TENSOR_ATTN_K,
1313
- LLM_TENSOR_ATTN_V,
1314
- LLM_TENSOR_ATTN_OUT,
1315
- LLM_TENSOR_FFN_GATE_INP,
1316
- LLM_TENSOR_FFN_NORM,
1317
- LLM_TENSOR_FFN_GATE,
1318
- LLM_TENSOR_FFN_DOWN,
1319
- LLM_TENSOR_FFN_UP,
1320
- LLM_TENSOR_FFN_GATE_EXPS,
1321
- LLM_TENSOR_FFN_DOWN_EXPS,
1322
- LLM_TENSOR_FFN_UP_EXPS,
1323
- };
1324
- case LLM_ARCH_FALCON_H1:
1325
- return {
1326
- LLM_TENSOR_TOKEN_EMBD,
1327
- LLM_TENSOR_OUTPUT,
1328
- LLM_TENSOR_OUTPUT_NORM,
1329
- LLM_TENSOR_ATTN_NORM,
1330
- LLM_TENSOR_ATTN_Q,
1331
- LLM_TENSOR_ATTN_K,
1332
- LLM_TENSOR_ATTN_V,
1333
- LLM_TENSOR_ATTN_OUT,
1334
- LLM_TENSOR_SSM_IN,
1335
- LLM_TENSOR_SSM_CONV1D,
1336
- LLM_TENSOR_SSM_DT,
1337
- LLM_TENSOR_SSM_A,
1338
- LLM_TENSOR_SSM_D,
1339
- LLM_TENSOR_SSM_NORM,
1340
- LLM_TENSOR_SSM_OUT,
1341
- LLM_TENSOR_FFN_NORM,
1342
- LLM_TENSOR_FFN_GATE,
1343
- LLM_TENSOR_FFN_DOWN,
1344
- LLM_TENSOR_FFN_UP,
1345
- };
1346
- case LLM_ARCH_COMMAND_R:
1347
- return {
1348
- LLM_TENSOR_TOKEN_EMBD,
1349
- LLM_TENSOR_OUTPUT_NORM,
1350
- LLM_TENSOR_ATTN_NORM,
1351
- LLM_TENSOR_ATTN_Q,
1352
- LLM_TENSOR_ATTN_K,
1353
- LLM_TENSOR_ATTN_V,
1354
- LLM_TENSOR_ATTN_OUT,
1355
- LLM_TENSOR_FFN_GATE,
1356
- LLM_TENSOR_FFN_DOWN,
1357
- LLM_TENSOR_FFN_UP,
1358
- LLM_TENSOR_ATTN_Q_NORM,
1359
- LLM_TENSOR_ATTN_K_NORM,
1360
- };
1361
- case LLM_ARCH_COHERE2:
1362
- return {
1363
- LLM_TENSOR_TOKEN_EMBD,
1364
- LLM_TENSOR_OUTPUT_NORM,
1365
- LLM_TENSOR_ATTN_NORM,
1366
- LLM_TENSOR_ATTN_Q,
1367
- LLM_TENSOR_ATTN_K,
1368
- LLM_TENSOR_ATTN_V,
1369
- LLM_TENSOR_ATTN_OUT,
1370
- LLM_TENSOR_FFN_GATE,
1371
- LLM_TENSOR_FFN_DOWN,
1372
- LLM_TENSOR_FFN_UP,
1373
- };
1374
- case LLM_ARCH_DBRX:
1375
- return {
1376
- LLM_TENSOR_TOKEN_EMBD,
1377
- LLM_TENSOR_OUTPUT_NORM,
1378
- LLM_TENSOR_OUTPUT,
1379
- LLM_TENSOR_ATTN_QKV,
1380
- LLM_TENSOR_ATTN_NORM,
1381
- LLM_TENSOR_ATTN_OUT,
1382
- LLM_TENSOR_ATTN_OUT_NORM,
1383
- LLM_TENSOR_FFN_GATE_INP,
1384
- LLM_TENSOR_FFN_GATE_EXPS,
1385
- LLM_TENSOR_FFN_DOWN_EXPS,
1386
- LLM_TENSOR_FFN_UP_EXPS,
1387
- };
1388
- case LLM_ARCH_OLMO:
1389
- return {
1390
- LLM_TENSOR_TOKEN_EMBD,
1391
- LLM_TENSOR_OUTPUT,
1392
- LLM_TENSOR_ATTN_Q,
1393
- LLM_TENSOR_ATTN_K,
1394
- LLM_TENSOR_ATTN_V,
1395
- LLM_TENSOR_ATTN_OUT,
1396
- LLM_TENSOR_FFN_GATE,
1397
- LLM_TENSOR_FFN_DOWN,
1398
- LLM_TENSOR_FFN_UP,
1399
- };
1400
- case LLM_ARCH_OLMO2:
1401
- return {
1402
- LLM_TENSOR_TOKEN_EMBD,
1403
- LLM_TENSOR_OUTPUT_NORM,
1404
- LLM_TENSOR_OUTPUT,
1405
- LLM_TENSOR_ATTN_Q,
1406
- LLM_TENSOR_ATTN_K,
1407
- LLM_TENSOR_ATTN_V,
1408
- LLM_TENSOR_ATTN_OUT,
1409
- LLM_TENSOR_ATTN_POST_NORM,
1410
- LLM_TENSOR_ATTN_Q_NORM,
1411
- LLM_TENSOR_ATTN_K_NORM,
1412
- LLM_TENSOR_FFN_POST_NORM,
1413
- LLM_TENSOR_FFN_GATE,
1414
- LLM_TENSOR_FFN_DOWN,
1415
- LLM_TENSOR_FFN_UP,
1416
- };
1417
- case LLM_ARCH_OPENELM:
1418
- return {
1419
- LLM_TENSOR_TOKEN_EMBD,
1420
- LLM_TENSOR_OUTPUT_NORM,
1421
- LLM_TENSOR_ATTN_NORM,
1422
- LLM_TENSOR_ATTN_QKV,
1423
- LLM_TENSOR_ATTN_Q_NORM,
1424
- LLM_TENSOR_ATTN_K_NORM,
1425
- LLM_TENSOR_ATTN_OUT,
1426
- LLM_TENSOR_FFN_NORM,
1427
- LLM_TENSOR_FFN_GATE,
1428
- LLM_TENSOR_FFN_DOWN,
1429
- LLM_TENSOR_FFN_UP,
1430
- };
1431
- case LLM_ARCH_ARCTIC:
1432
- return {
1433
- LLM_TENSOR_TOKEN_EMBD,
1434
- LLM_TENSOR_OUTPUT_NORM,
1435
- LLM_TENSOR_OUTPUT,
1436
- LLM_TENSOR_ATTN_NORM,
1437
- LLM_TENSOR_ATTN_Q,
1438
- LLM_TENSOR_ATTN_K,
1439
- LLM_TENSOR_ATTN_V,
1440
- LLM_TENSOR_ATTN_OUT,
1441
- LLM_TENSOR_FFN_GATE_INP,
1442
- LLM_TENSOR_FFN_NORM,
1443
- LLM_TENSOR_FFN_GATE,
1444
- LLM_TENSOR_FFN_DOWN,
1445
- LLM_TENSOR_FFN_UP,
1446
- LLM_TENSOR_FFN_NORM_EXPS,
1447
- LLM_TENSOR_FFN_GATE_EXPS,
1448
- LLM_TENSOR_FFN_DOWN_EXPS,
1449
- LLM_TENSOR_FFN_UP_EXPS,
1450
- };
1451
- case LLM_ARCH_DEEPSEEK:
1452
- return {
1453
- LLM_TENSOR_TOKEN_EMBD,
1454
- LLM_TENSOR_OUTPUT_NORM,
1455
- LLM_TENSOR_OUTPUT,
1456
- LLM_TENSOR_ROPE_FREQS,
1457
- LLM_TENSOR_ATTN_NORM,
1458
- LLM_TENSOR_ATTN_Q,
1459
- LLM_TENSOR_ATTN_K,
1460
- LLM_TENSOR_ATTN_V,
1461
- LLM_TENSOR_ATTN_OUT,
1462
- LLM_TENSOR_ATTN_ROT_EMBD,
1463
- LLM_TENSOR_FFN_GATE_INP,
1464
- LLM_TENSOR_FFN_NORM,
1465
- LLM_TENSOR_FFN_GATE,
1466
- LLM_TENSOR_FFN_DOWN,
1467
- LLM_TENSOR_FFN_UP,
1468
- LLM_TENSOR_FFN_GATE_EXPS,
1469
- LLM_TENSOR_FFN_DOWN_EXPS,
1470
- LLM_TENSOR_FFN_UP_EXPS,
1471
- LLM_TENSOR_FFN_GATE_INP_SHEXP,
1472
- LLM_TENSOR_FFN_GATE_SHEXP,
1473
- LLM_TENSOR_FFN_DOWN_SHEXP,
1474
- LLM_TENSOR_FFN_UP_SHEXP,
1475
- };
1476
- case LLM_ARCH_DEEPSEEK2:
1477
- return {
1478
- LLM_TENSOR_TOKEN_EMBD,
1479
- LLM_TENSOR_OUTPUT_NORM,
1480
- LLM_TENSOR_OUTPUT,
1481
- LLM_TENSOR_ATTN_NORM,
1482
- LLM_TENSOR_ATTN_Q_A_NORM,
1483
- LLM_TENSOR_ATTN_KV_A_NORM,
1484
- LLM_TENSOR_ATTN_Q,
1485
- LLM_TENSOR_ATTN_Q_A,
1486
- LLM_TENSOR_ATTN_Q_B,
1487
- LLM_TENSOR_ATTN_KV_A_MQA,
1488
- LLM_TENSOR_ATTN_KV_B,
1489
- LLM_TENSOR_ATTN_K_B,
1490
- LLM_TENSOR_ATTN_V_B,
1491
- LLM_TENSOR_ATTN_OUT,
1492
- LLM_TENSOR_FFN_NORM,
1493
- LLM_TENSOR_FFN_GATE,
1494
- LLM_TENSOR_FFN_UP,
1495
- LLM_TENSOR_FFN_DOWN,
1496
- LLM_TENSOR_FFN_GATE_INP,
1497
- LLM_TENSOR_FFN_GATE_EXPS,
1498
- LLM_TENSOR_FFN_DOWN_EXPS,
1499
- LLM_TENSOR_FFN_UP_EXPS,
1500
- LLM_TENSOR_FFN_GATE_INP_SHEXP,
1501
- LLM_TENSOR_FFN_GATE_SHEXP,
1502
- LLM_TENSOR_FFN_DOWN_SHEXP,
1503
- LLM_TENSOR_FFN_UP_SHEXP,
1504
- LLM_TENSOR_FFN_EXP_PROBS_B,
1505
- };
1506
- case LLM_ARCH_PLM:
1507
- return {
1508
- LLM_TENSOR_TOKEN_EMBD,
1509
- LLM_TENSOR_OUTPUT_NORM,
1510
- LLM_TENSOR_ATTN_NORM,
1511
- LLM_TENSOR_ATTN_Q,
1512
- LLM_TENSOR_ATTN_KV_A_MQA,
1513
- LLM_TENSOR_ATTN_KV_A_NORM,
1514
- LLM_TENSOR_ATTN_KV_B,
1515
- LLM_TENSOR_ATTN_OUT,
1516
- LLM_TENSOR_FFN_NORM,
1517
- LLM_TENSOR_FFN_DOWN,
1518
- LLM_TENSOR_FFN_UP,
1519
- };
1520
- case LLM_ARCH_CHATGLM:
1521
- return {
1522
- LLM_TENSOR_TOKEN_EMBD,
1523
- LLM_TENSOR_ROPE_FREQS,
1524
- LLM_TENSOR_OUTPUT_NORM,
1525
- LLM_TENSOR_OUTPUT,
1526
- LLM_TENSOR_ATTN_NORM,
1527
- LLM_TENSOR_ATTN_QKV,
1528
- LLM_TENSOR_ATTN_Q,
1529
- LLM_TENSOR_ATTN_K,
1530
- LLM_TENSOR_ATTN_V,
1531
- LLM_TENSOR_ATTN_OUT,
1532
- LLM_TENSOR_FFN_NORM,
1533
- LLM_TENSOR_FFN_UP,
1534
- LLM_TENSOR_FFN_DOWN,
1535
- };
1536
- case LLM_ARCH_GLM4:
1537
- return {
1538
- LLM_TENSOR_TOKEN_EMBD,
1539
- LLM_TENSOR_ROPE_FREQS,
1540
- LLM_TENSOR_OUTPUT_NORM,
1541
- LLM_TENSOR_OUTPUT,
1542
- LLM_TENSOR_ATTN_NORM,
1543
- LLM_TENSOR_ATTN_Q,
1544
- LLM_TENSOR_ATTN_K,
1545
- LLM_TENSOR_ATTN_V,
1546
- LLM_TENSOR_ATTN_OUT,
1547
- LLM_TENSOR_FFN_NORM,
1548
- LLM_TENSOR_FFN_UP,
1549
- LLM_TENSOR_FFN_DOWN,
1550
- LLM_TENSOR_ATTN_POST_NORM,
1551
- LLM_TENSOR_FFN_POST_NORM,
1552
- };
1553
- case LLM_ARCH_GLM4_MOE:
1554
- return {
1555
- LLM_TENSOR_TOKEN_EMBD,
1556
- LLM_TENSOR_OUTPUT_NORM,
1557
- LLM_TENSOR_OUTPUT,
1558
- LLM_TENSOR_ATTN_NORM,
1559
- LLM_TENSOR_ATTN_POST_NORM,
1560
- LLM_TENSOR_ATTN_Q,
1561
- LLM_TENSOR_ATTN_K,
1562
- LLM_TENSOR_ATTN_V,
1563
- LLM_TENSOR_ATTN_OUT,
1564
- LLM_TENSOR_ATTN_Q_NORM,
1565
- LLM_TENSOR_ATTN_K_NORM,
1566
- LLM_TENSOR_FFN_GATE,
1567
- LLM_TENSOR_FFN_DOWN,
1568
- LLM_TENSOR_FFN_UP,
1569
- LLM_TENSOR_FFN_GATE_INP,
1570
- LLM_TENSOR_FFN_GATE_EXPS,
1571
- LLM_TENSOR_FFN_DOWN_EXPS,
1572
- LLM_TENSOR_FFN_UP_EXPS,
1573
- LLM_TENSOR_FFN_GATE_SHEXP,
1574
- LLM_TENSOR_FFN_DOWN_SHEXP,
1575
- LLM_TENSOR_FFN_UP_SHEXP,
1576
- LLM_TENSOR_FFN_EXP_PROBS_B,
1577
- LLM_TENSOR_NEXTN_EH_PROJ,
1578
- LLM_TENSOR_NEXTN_EMBED_TOKENS,
1579
- LLM_TENSOR_NEXTN_ENORM,
1580
- LLM_TENSOR_NEXTN_HNORM,
1581
- LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
1582
- LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
1583
- };
1584
- case LLM_ARCH_BITNET:
1585
- return {
1586
- LLM_TENSOR_TOKEN_EMBD,
1587
- LLM_TENSOR_OUTPUT_NORM,
1588
- LLM_TENSOR_ATTN_Q,
1589
- LLM_TENSOR_ATTN_K,
1590
- LLM_TENSOR_ATTN_V,
1591
- LLM_TENSOR_ATTN_OUT,
1592
- LLM_TENSOR_ATTN_NORM,
1593
- LLM_TENSOR_ATTN_SUB_NORM,
1594
- LLM_TENSOR_FFN_GATE,
1595
- LLM_TENSOR_FFN_DOWN,
1596
- LLM_TENSOR_FFN_UP,
1597
- LLM_TENSOR_FFN_NORM,
1598
- LLM_TENSOR_FFN_SUB_NORM,
1599
- };
1600
- case LLM_ARCH_T5:
1601
- return {
1602
- LLM_TENSOR_TOKEN_EMBD,
1603
- LLM_TENSOR_OUTPUT,
1604
- LLM_TENSOR_DEC_OUTPUT_NORM,
1605
- LLM_TENSOR_DEC_ATTN_NORM,
1606
- LLM_TENSOR_DEC_ATTN_Q,
1607
- LLM_TENSOR_DEC_ATTN_K,
1608
- LLM_TENSOR_DEC_ATTN_V,
1609
- LLM_TENSOR_DEC_ATTN_OUT,
1610
- LLM_TENSOR_DEC_ATTN_REL_B,
1611
- LLM_TENSOR_DEC_CROSS_ATTN_NORM,
1612
- LLM_TENSOR_DEC_CROSS_ATTN_Q,
1613
- LLM_TENSOR_DEC_CROSS_ATTN_K,
1614
- LLM_TENSOR_DEC_CROSS_ATTN_V,
1615
- LLM_TENSOR_DEC_CROSS_ATTN_OUT,
1616
- LLM_TENSOR_DEC_CROSS_ATTN_REL_B,
1617
- LLM_TENSOR_DEC_FFN_NORM,
1618
- LLM_TENSOR_DEC_FFN_GATE,
1619
- LLM_TENSOR_DEC_FFN_DOWN,
1620
- LLM_TENSOR_DEC_FFN_UP,
1621
- LLM_TENSOR_ENC_OUTPUT_NORM,
1622
- LLM_TENSOR_ENC_ATTN_NORM,
1623
- LLM_TENSOR_ENC_ATTN_Q,
1624
- LLM_TENSOR_ENC_ATTN_K,
1625
- LLM_TENSOR_ENC_ATTN_V,
1626
- LLM_TENSOR_ENC_ATTN_OUT,
1627
- LLM_TENSOR_ENC_ATTN_REL_B,
1628
- LLM_TENSOR_ENC_FFN_NORM,
1629
- LLM_TENSOR_ENC_FFN_GATE,
1630
- LLM_TENSOR_ENC_FFN_DOWN,
1631
- LLM_TENSOR_ENC_FFN_UP,
1632
- };
1633
- case LLM_ARCH_T5ENCODER:
1634
- return {
1635
- LLM_TENSOR_TOKEN_EMBD,
1636
- LLM_TENSOR_OUTPUT,
1637
- LLM_TENSOR_ENC_OUTPUT_NORM,
1638
- LLM_TENSOR_ENC_ATTN_NORM,
1639
- LLM_TENSOR_ENC_ATTN_Q,
1640
- LLM_TENSOR_ENC_ATTN_K,
1641
- LLM_TENSOR_ENC_ATTN_V,
1642
- LLM_TENSOR_ENC_ATTN_OUT,
1643
- LLM_TENSOR_ENC_ATTN_REL_B,
1644
- LLM_TENSOR_ENC_FFN_NORM,
1645
- LLM_TENSOR_ENC_FFN_GATE,
1646
- LLM_TENSOR_ENC_FFN_DOWN,
1647
- LLM_TENSOR_ENC_FFN_UP,
1648
- };
1649
- case LLM_ARCH_JAIS:
1650
- return {
1651
- LLM_TENSOR_TOKEN_EMBD,
1652
- LLM_TENSOR_OUTPUT_NORM,
1653
- LLM_TENSOR_OUTPUT,
1654
- LLM_TENSOR_ATTN_NORM,
1655
- LLM_TENSOR_ATTN_QKV,
1656
- LLM_TENSOR_ATTN_OUT,
1657
- LLM_TENSOR_FFN_NORM,
1658
- LLM_TENSOR_FFN_UP,
1659
- LLM_TENSOR_FFN_GATE,
1660
- LLM_TENSOR_FFN_DOWN,
1661
- };
1662
- case LLM_ARCH_NEMOTRON_H:
1663
- return {
1664
- LLM_TENSOR_TOKEN_EMBD,
1665
- LLM_TENSOR_OUTPUT_NORM,
1666
- LLM_TENSOR_OUTPUT,
1667
- LLM_TENSOR_ATTN_NORM,
1668
- LLM_TENSOR_SSM_IN,
1669
- LLM_TENSOR_SSM_CONV1D,
1670
- LLM_TENSOR_SSM_DT,
1671
- LLM_TENSOR_SSM_A,
1672
- LLM_TENSOR_SSM_D,
1673
- LLM_TENSOR_SSM_NORM,
1674
- LLM_TENSOR_SSM_OUT,
1675
- LLM_TENSOR_ATTN_Q,
1676
- LLM_TENSOR_ATTN_K,
1677
- LLM_TENSOR_ATTN_V,
1678
- LLM_TENSOR_ATTN_OUT,
1679
- LLM_TENSOR_FFN_DOWN,
1680
- LLM_TENSOR_FFN_UP,
1681
- };
1682
- case LLM_ARCH_NEMOTRON_H_MOE:
1683
- return {
1684
- LLM_TENSOR_TOKEN_EMBD,
1685
- LLM_TENSOR_OUTPUT_NORM,
1686
- LLM_TENSOR_OUTPUT,
1687
- LLM_TENSOR_ATTN_NORM,
1688
- // mamba(2) ssm layers
1689
- LLM_TENSOR_SSM_IN,
1690
- LLM_TENSOR_SSM_CONV1D,
1691
- LLM_TENSOR_SSM_DT,
1692
- LLM_TENSOR_SSM_A,
1693
- LLM_TENSOR_SSM_D,
1694
- LLM_TENSOR_SSM_NORM,
1695
- LLM_TENSOR_SSM_OUT,
1696
- // attention layers
1697
- LLM_TENSOR_ATTN_Q,
1698
- LLM_TENSOR_ATTN_K,
1699
- LLM_TENSOR_ATTN_V,
1700
- LLM_TENSOR_ATTN_OUT,
1701
- // dense FFN
1702
- LLM_TENSOR_FFN_DOWN,
1703
- LLM_TENSOR_FFN_UP,
1704
- // MoE FFN (for MoE layers)
1705
- LLM_TENSOR_FFN_GATE_INP,
1706
- LLM_TENSOR_FFN_UP_EXPS,
1707
- LLM_TENSOR_FFN_DOWN_EXPS,
1708
- LLM_TENSOR_FFN_EXP_PROBS_B,
1709
- // MoE shared expert layer
1710
- LLM_TENSOR_FFN_DOWN_SHEXP,
1711
- LLM_TENSOR_FFN_UP_SHEXP,
1712
- };
1713
- case LLM_ARCH_EXAONE4:
1714
- return {
1715
- LLM_TENSOR_TOKEN_EMBD,
1716
- LLM_TENSOR_OUTPUT_NORM,
1717
- LLM_TENSOR_OUTPUT,
1718
- LLM_TENSOR_ROPE_FREQS,
1719
- LLM_TENSOR_ATTN_Q,
1720
- LLM_TENSOR_ATTN_Q_NORM,
1721
- LLM_TENSOR_ATTN_K,
1722
- LLM_TENSOR_ATTN_K_NORM,
1723
- LLM_TENSOR_ATTN_V,
1724
- LLM_TENSOR_ATTN_OUT,
1725
- LLM_TENSOR_ATTN_POST_NORM,
1726
- LLM_TENSOR_FFN_GATE,
1727
- LLM_TENSOR_FFN_DOWN,
1728
- LLM_TENSOR_FFN_UP,
1729
- LLM_TENSOR_FFN_POST_NORM,
1730
- };
1731
- case LLM_ARCH_RWKV6:
1732
- return {
1733
- LLM_TENSOR_TOKEN_EMBD,
1734
- LLM_TENSOR_TOKEN_EMBD_NORM,
1735
- LLM_TENSOR_OUTPUT_NORM,
1736
- LLM_TENSOR_OUTPUT,
1737
- LLM_TENSOR_ATTN_NORM,
1738
- LLM_TENSOR_ATTN_NORM_2,
1739
- LLM_TENSOR_TIME_MIX_W1,
1740
- LLM_TENSOR_TIME_MIX_W2,
1741
- LLM_TENSOR_TIME_MIX_LERP_X,
1742
- LLM_TENSOR_TIME_MIX_LERP_W,
1743
- LLM_TENSOR_TIME_MIX_LERP_K,
1744
- LLM_TENSOR_TIME_MIX_LERP_V,
1745
- LLM_TENSOR_TIME_MIX_LERP_R,
1746
- LLM_TENSOR_TIME_MIX_LERP_G,
1747
- LLM_TENSOR_TIME_MIX_LERP_FUSED,
1748
- LLM_TENSOR_TIME_MIX_FIRST,
1749
- LLM_TENSOR_TIME_MIX_DECAY,
1750
- LLM_TENSOR_TIME_MIX_DECAY_W1,
1751
- LLM_TENSOR_TIME_MIX_DECAY_W2,
1752
- LLM_TENSOR_TIME_MIX_KEY,
1753
- LLM_TENSOR_TIME_MIX_VALUE,
1754
- LLM_TENSOR_TIME_MIX_RECEPTANCE,
1755
- LLM_TENSOR_TIME_MIX_GATE,
1756
- LLM_TENSOR_TIME_MIX_LN,
1757
- LLM_TENSOR_TIME_MIX_OUTPUT,
1758
- LLM_TENSOR_CHANNEL_MIX_LERP_K,
1759
- LLM_TENSOR_CHANNEL_MIX_LERP_R,
1760
- LLM_TENSOR_CHANNEL_MIX_KEY,
1761
- LLM_TENSOR_CHANNEL_MIX_VALUE,
1762
- LLM_TENSOR_CHANNEL_MIX_RECEPTANCE,
1763
- };
1764
- case LLM_ARCH_RWKV6QWEN2:
1765
- return {
1766
- LLM_TENSOR_TOKEN_EMBD,
1767
- LLM_TENSOR_OUTPUT_NORM,
1768
- LLM_TENSOR_OUTPUT,
1769
- LLM_TENSOR_ATTN_NORM,
1770
- LLM_TENSOR_TIME_MIX_W1,
1771
- LLM_TENSOR_TIME_MIX_W2,
1772
- LLM_TENSOR_TIME_MIX_LERP_X,
1773
- LLM_TENSOR_TIME_MIX_LERP_FUSED,
1774
- LLM_TENSOR_TIME_MIX_FIRST,
1775
- LLM_TENSOR_TIME_MIX_DECAY,
1776
- LLM_TENSOR_TIME_MIX_DECAY_W1,
1777
- LLM_TENSOR_TIME_MIX_DECAY_W2,
1778
- LLM_TENSOR_TIME_MIX_KEY,
1779
- LLM_TENSOR_TIME_MIX_VALUE,
1780
- LLM_TENSOR_TIME_MIX_RECEPTANCE,
1781
- LLM_TENSOR_TIME_MIX_GATE,
1782
- LLM_TENSOR_TIME_MIX_OUTPUT,
1783
- LLM_TENSOR_FFN_NORM,
1784
- LLM_TENSOR_FFN_GATE,
1785
- LLM_TENSOR_FFN_DOWN,
1786
- LLM_TENSOR_FFN_UP,
1787
- };
1788
- case LLM_ARCH_RWKV7:
1789
- return {
1790
- LLM_TENSOR_TOKEN_EMBD,
1791
- LLM_TENSOR_TOKEN_EMBD_NORM,
1792
- LLM_TENSOR_OUTPUT_NORM,
1793
- LLM_TENSOR_OUTPUT,
1794
- LLM_TENSOR_ATTN_NORM,
1795
- LLM_TENSOR_ATTN_NORM_2,
1796
- LLM_TENSOR_TIME_MIX_W0,
1797
- LLM_TENSOR_TIME_MIX_W1,
1798
- LLM_TENSOR_TIME_MIX_W2,
1799
- LLM_TENSOR_TIME_MIX_A0,
1800
- LLM_TENSOR_TIME_MIX_A1,
1801
- LLM_TENSOR_TIME_MIX_A2,
1802
- LLM_TENSOR_TIME_MIX_V0,
1803
- LLM_TENSOR_TIME_MIX_V1,
1804
- LLM_TENSOR_TIME_MIX_V2,
1805
- LLM_TENSOR_TIME_MIX_G1,
1806
- LLM_TENSOR_TIME_MIX_G2,
1807
- LLM_TENSOR_TIME_MIX_K_K,
1808
- LLM_TENSOR_TIME_MIX_K_A,
1809
- LLM_TENSOR_TIME_MIX_R_K,
1810
- LLM_TENSOR_TIME_MIX_LERP_FUSED,
1811
- LLM_TENSOR_TIME_MIX_KEY,
1812
- LLM_TENSOR_TIME_MIX_VALUE,
1813
- LLM_TENSOR_TIME_MIX_RECEPTANCE,
1814
- LLM_TENSOR_TIME_MIX_LN,
1815
- LLM_TENSOR_TIME_MIX_OUTPUT,
1816
- LLM_TENSOR_CHANNEL_MIX_LERP_K,
1817
- LLM_TENSOR_CHANNEL_MIX_KEY,
1818
- LLM_TENSOR_CHANNEL_MIX_VALUE,
1819
- };
1820
- case LLM_ARCH_ARWKV7:
1821
- return {
1822
- LLM_TENSOR_TOKEN_EMBD,
1823
- LLM_TENSOR_TOKEN_EMBD_NORM,
1824
- LLM_TENSOR_OUTPUT_NORM,
1825
- LLM_TENSOR_OUTPUT,
1826
- LLM_TENSOR_ATTN_NORM,
1827
- LLM_TENSOR_TIME_MIX_W0,
1828
- LLM_TENSOR_TIME_MIX_W1,
1829
- LLM_TENSOR_TIME_MIX_W2,
1830
- LLM_TENSOR_TIME_MIX_A0,
1831
- LLM_TENSOR_TIME_MIX_A1,
1832
- LLM_TENSOR_TIME_MIX_A2,
1833
- LLM_TENSOR_TIME_MIX_V0,
1834
- LLM_TENSOR_TIME_MIX_V1,
1835
- LLM_TENSOR_TIME_MIX_V2,
1836
- LLM_TENSOR_TIME_MIX_G1,
1837
- LLM_TENSOR_TIME_MIX_G2,
1838
- LLM_TENSOR_TIME_MIX_K_K,
1839
- LLM_TENSOR_TIME_MIX_K_A,
1840
- LLM_TENSOR_TIME_MIX_R_K,
1841
- LLM_TENSOR_TIME_MIX_LERP_FUSED,
1842
- LLM_TENSOR_TIME_MIX_KEY,
1843
- LLM_TENSOR_TIME_MIX_VALUE,
1844
- LLM_TENSOR_TIME_MIX_RECEPTANCE,
1845
- LLM_TENSOR_TIME_MIX_LN,
1846
- LLM_TENSOR_TIME_MIX_OUTPUT,
1847
- LLM_TENSOR_FFN_NORM,
1848
- LLM_TENSOR_FFN_GATE,
1849
- LLM_TENSOR_FFN_DOWN,
1850
- LLM_TENSOR_FFN_UP,
1851
- };
1852
- case LLM_ARCH_GRANITE_MOE:
1853
- return {
1854
- LLM_TENSOR_TOKEN_EMBD,
1855
- LLM_TENSOR_OUTPUT_NORM,
1856
- LLM_TENSOR_OUTPUT,
1857
- LLM_TENSOR_ATTN_NORM,
1858
- LLM_TENSOR_ATTN_Q,
1859
- LLM_TENSOR_ATTN_K,
1860
- LLM_TENSOR_ATTN_V,
1861
- LLM_TENSOR_ATTN_OUT,
1862
- LLM_TENSOR_FFN_NORM,
1863
- LLM_TENSOR_FFN_GATE_INP,
1864
- LLM_TENSOR_FFN_GATE_EXPS,
1865
- LLM_TENSOR_FFN_DOWN_EXPS,
1866
- LLM_TENSOR_FFN_UP_EXPS,
1867
- LLM_TENSOR_FFN_GATE_SHEXP,
1868
- LLM_TENSOR_FFN_DOWN_SHEXP,
1869
- LLM_TENSOR_FFN_UP_SHEXP,
1870
- };
1871
- case LLM_ARCH_GRANITE_HYBRID:
1872
- return {
1873
- LLM_TENSOR_TOKEN_EMBD,
1874
- LLM_TENSOR_OUTPUT_NORM,
1875
- LLM_TENSOR_OUTPUT,
1876
- LLM_TENSOR_ATTN_NORM,
1877
- LLM_TENSOR_SSM_IN,
1878
- LLM_TENSOR_SSM_CONV1D,
1879
- LLM_TENSOR_SSM_DT,
1880
- LLM_TENSOR_SSM_A,
1881
- LLM_TENSOR_SSM_D,
1882
- LLM_TENSOR_SSM_NORM,
1883
- LLM_TENSOR_SSM_OUT,
1884
- LLM_TENSOR_ATTN_Q,
1885
- LLM_TENSOR_ATTN_K,
1886
- LLM_TENSOR_ATTN_V,
1887
- LLM_TENSOR_ATTN_OUT,
1888
- LLM_TENSOR_FFN_NORM,
1889
- LLM_TENSOR_FFN_GATE,
1890
- LLM_TENSOR_FFN_DOWN,
1891
- LLM_TENSOR_FFN_UP,
1892
- LLM_TENSOR_FFN_NORM,
1893
- LLM_TENSOR_FFN_GATE_INP,
1894
- LLM_TENSOR_FFN_GATE_EXPS,
1895
- LLM_TENSOR_FFN_DOWN_EXPS,
1896
- LLM_TENSOR_FFN_UP_EXPS,
1897
- LLM_TENSOR_FFN_GATE_SHEXP,
1898
- LLM_TENSOR_FFN_DOWN_SHEXP,
1899
- LLM_TENSOR_FFN_UP_SHEXP,
1900
- };
1901
- case LLM_ARCH_WAVTOKENIZER_DEC:
1902
- return {
1903
- LLM_TENSOR_TOKEN_EMBD,
1904
- LLM_TENSOR_TOKEN_EMBD_NORM,
1905
- LLM_TENSOR_CONV1D,
1906
- LLM_TENSOR_CONVNEXT_DW,
1907
- LLM_TENSOR_CONVNEXT_NORM,
1908
- LLM_TENSOR_CONVNEXT_PW1,
1909
- LLM_TENSOR_CONVNEXT_PW2,
1910
- LLM_TENSOR_CONVNEXT_GAMMA,
1911
- LLM_TENSOR_OUTPUT_NORM,
1912
- LLM_TENSOR_OUTPUT,
1913
- LLM_TENSOR_POS_NET_CONV1,
1914
- LLM_TENSOR_POS_NET_CONV2,
1915
- LLM_TENSOR_POS_NET_NORM,
1916
- LLM_TENSOR_POS_NET_NORM1,
1917
- LLM_TENSOR_POS_NET_NORM2,
1918
- LLM_TENSOR_POS_NET_ATTN_NORM,
1919
- LLM_TENSOR_POS_NET_ATTN_Q,
1920
- LLM_TENSOR_POS_NET_ATTN_K,
1921
- LLM_TENSOR_POS_NET_ATTN_V,
1922
- LLM_TENSOR_POS_NET_ATTN_OUT,
1923
- };
1924
- case LLM_ARCH_BAILINGMOE:
1925
- return {
1926
- LLM_TENSOR_TOKEN_EMBD,
1927
- LLM_TENSOR_OUTPUT_NORM,
1928
- LLM_TENSOR_OUTPUT,
1929
- LLM_TENSOR_ROPE_FREQS,
1930
- LLM_TENSOR_ATTN_NORM,
1931
- LLM_TENSOR_ATTN_Q,
1932
- LLM_TENSOR_ATTN_K,
1933
- LLM_TENSOR_ATTN_V,
1934
- LLM_TENSOR_ATTN_OUT,
1935
- LLM_TENSOR_FFN_GATE_INP,
1936
- LLM_TENSOR_FFN_NORM,
1937
- LLM_TENSOR_FFN_GATE_EXPS,
1938
- LLM_TENSOR_FFN_DOWN_EXPS,
1939
- LLM_TENSOR_FFN_UP_EXPS,
1940
- LLM_TENSOR_FFN_GATE_INP_SHEXP,
1941
- LLM_TENSOR_FFN_GATE_SHEXP,
1942
- LLM_TENSOR_FFN_DOWN_SHEXP,
1943
- LLM_TENSOR_FFN_UP_SHEXP,
1944
- };
1945
- case LLM_ARCH_BAILINGMOE2:
1946
- return {
1947
- LLM_TENSOR_TOKEN_EMBD,
1948
- LLM_TENSOR_OUTPUT_NORM,
1949
- LLM_TENSOR_OUTPUT,
1950
- LLM_TENSOR_ATTN_NORM,
1951
- LLM_TENSOR_ATTN_Q_NORM,
1952
- LLM_TENSOR_ATTN_K_NORM,
1953
- LLM_TENSOR_ATTN_QKV,
1954
- LLM_TENSOR_ATTN_OUT,
1955
- LLM_TENSOR_FFN_GATE_INP,
1956
- LLM_TENSOR_FFN_EXP_PROBS_B,
1957
- LLM_TENSOR_FFN_NORM,
1958
- LLM_TENSOR_FFN_GATE,
1959
- LLM_TENSOR_FFN_DOWN,
1960
- LLM_TENSOR_FFN_UP,
1961
- LLM_TENSOR_FFN_GATE_EXPS,
1962
- LLM_TENSOR_FFN_DOWN_EXPS,
1963
- LLM_TENSOR_FFN_UP_EXPS,
1964
- LLM_TENSOR_FFN_GATE_SHEXP,
1965
- LLM_TENSOR_FFN_DOWN_SHEXP,
1966
- LLM_TENSOR_FFN_UP_SHEXP,
1967
- LLM_TENSOR_NEXTN_EH_PROJ,
1968
- LLM_TENSOR_NEXTN_EMBED_TOKENS,
1969
- LLM_TENSOR_NEXTN_ENORM,
1970
- LLM_TENSOR_NEXTN_HNORM,
1971
- LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
1972
- LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
1973
- LLM_TENSOR_LAYER_OUT_NORM,
1974
- };
1975
- case LLM_ARCH_DOTS1:
1976
- return {
1977
- LLM_TENSOR_TOKEN_EMBD,
1978
- LLM_TENSOR_OUTPUT_NORM,
1979
- LLM_TENSOR_OUTPUT,
1980
- LLM_TENSOR_ATTN_NORM,
1981
- LLM_TENSOR_ATTN_Q,
1982
- LLM_TENSOR_ATTN_Q_NORM,
1983
- LLM_TENSOR_ATTN_K,
1984
- LLM_TENSOR_ATTN_K_NORM,
1985
- LLM_TENSOR_ATTN_V,
1986
- LLM_TENSOR_ATTN_OUT,
1987
- LLM_TENSOR_FFN_NORM,
1988
- LLM_TENSOR_FFN_GATE,
1989
- LLM_TENSOR_FFN_UP,
1990
- LLM_TENSOR_FFN_DOWN,
1991
- LLM_TENSOR_FFN_GATE_INP,
1992
- LLM_TENSOR_FFN_GATE_EXPS,
1993
- LLM_TENSOR_FFN_DOWN_EXPS,
1994
- LLM_TENSOR_FFN_UP_EXPS,
1995
- LLM_TENSOR_FFN_GATE_INP_SHEXP,
1996
- LLM_TENSOR_FFN_GATE_SHEXP,
1997
- LLM_TENSOR_FFN_DOWN_SHEXP,
1998
- LLM_TENSOR_FFN_UP_SHEXP,
1999
- LLM_TENSOR_FFN_EXP_PROBS_B,
2000
- };
2001
- case LLM_ARCH_ERNIE4_5_MOE:
2002
- return {
2003
- LLM_TENSOR_TOKEN_EMBD,
2004
- LLM_TENSOR_OUTPUT_NORM,
2005
- LLM_TENSOR_OUTPUT,
2006
- LLM_TENSOR_ATTN_NORM,
2007
- LLM_TENSOR_ATTN_Q,
2008
- LLM_TENSOR_ATTN_K,
2009
- LLM_TENSOR_ATTN_V,
2010
- LLM_TENSOR_ATTN_OUT,
2011
- LLM_TENSOR_FFN_NORM,
2012
- LLM_TENSOR_FFN_GATE,
2013
- LLM_TENSOR_FFN_DOWN,
2014
- LLM_TENSOR_FFN_UP,
2015
- LLM_TENSOR_FFN_GATE_INP,
2016
- LLM_TENSOR_FFN_GATE_SHEXP,
2017
- LLM_TENSOR_FFN_DOWN_SHEXP,
2018
- LLM_TENSOR_FFN_UP_SHEXP,
2019
- LLM_TENSOR_FFN_GATE_EXPS,
2020
- LLM_TENSOR_FFN_DOWN_EXPS,
2021
- LLM_TENSOR_FFN_UP_EXPS,
2022
- LLM_TENSOR_FFN_EXP_PROBS_B,
2023
- };
2024
- case LLM_ARCH_HUNYUAN_MOE:
2025
- return {
2026
- LLM_TENSOR_TOKEN_EMBD,
2027
- LLM_TENSOR_OUTPUT_NORM,
2028
- LLM_TENSOR_OUTPUT,
2029
- LLM_TENSOR_ATTN_NORM,
2030
- LLM_TENSOR_ATTN_Q,
2031
- LLM_TENSOR_ATTN_Q_NORM,
2032
- LLM_TENSOR_ATTN_K,
2033
- LLM_TENSOR_ATTN_K_NORM,
2034
- LLM_TENSOR_ATTN_V,
2035
- LLM_TENSOR_ATTN_OUT,
2036
- LLM_TENSOR_FFN_GATE_INP,
2037
- LLM_TENSOR_FFN_NORM,
2038
- LLM_TENSOR_FFN_GATE_SHEXP,
2039
- LLM_TENSOR_FFN_DOWN_SHEXP,
2040
- LLM_TENSOR_FFN_UP_SHEXP,
2041
- LLM_TENSOR_FFN_GATE_EXPS,
2042
- LLM_TENSOR_FFN_DOWN_EXPS,
2043
- LLM_TENSOR_FFN_UP_EXPS,
2044
- };
2045
- case LLM_ARCH_OPENAI_MOE:
2046
- return {
2047
- LLM_TENSOR_TOKEN_EMBD,
2048
- LLM_TENSOR_OUTPUT_NORM,
2049
- LLM_TENSOR_OUTPUT,
2050
- LLM_TENSOR_ATTN_NORM,
2051
- LLM_TENSOR_ATTN_POST_NORM,
2052
- LLM_TENSOR_ATTN_Q,
2053
- LLM_TENSOR_ATTN_K,
2054
- LLM_TENSOR_ATTN_V,
2055
- LLM_TENSOR_ATTN_OUT,
2056
- LLM_TENSOR_ATTN_SINKS,
2057
- LLM_TENSOR_FFN_GATE_INP,
2058
- LLM_TENSOR_FFN_GATE_EXPS,
2059
- LLM_TENSOR_FFN_DOWN_EXPS,
2060
- LLM_TENSOR_FFN_UP_EXPS,
2061
- };
2062
- case LLM_ARCH_LFM2:
2063
- return {
2064
- LLM_TENSOR_ATTN_NORM,
2065
- LLM_TENSOR_ATTN_Q,
2066
- LLM_TENSOR_ATTN_K,
2067
- LLM_TENSOR_ATTN_V,
2068
- LLM_TENSOR_ATTN_OUT,
2069
- LLM_TENSOR_ATTN_K_NORM,
2070
- LLM_TENSOR_ATTN_Q_NORM,
2071
- LLM_TENSOR_FFN_DOWN,
2072
- LLM_TENSOR_FFN_GATE,
2073
- LLM_TENSOR_FFN_NORM,
2074
- LLM_TENSOR_FFN_UP,
2075
- LLM_TENSOR_SHORTCONV_CONV,
2076
- LLM_TENSOR_SHORTCONV_INPROJ,
2077
- LLM_TENSOR_SHORTCONV_OUTPROJ,
2078
- LLM_TENSOR_TOKEN_EMBD,
2079
- LLM_TENSOR_OUTPUT_NORM_LFM2,
2080
- LLM_TENSOR_OUTPUT,
2081
- LLM_TENSOR_DENSE_2_OUT,
2082
- };
2083
- case LLM_ARCH_LFM2MOE:
2084
- return {
2085
- LLM_TENSOR_ATTN_NORM,
2086
- LLM_TENSOR_ATTN_Q,
2087
- LLM_TENSOR_ATTN_K,
2088
- LLM_TENSOR_ATTN_V,
2089
- LLM_TENSOR_ATTN_OUT,
2090
- LLM_TENSOR_ATTN_K_NORM,
2091
- LLM_TENSOR_ATTN_Q_NORM,
2092
- LLM_TENSOR_FFN_DOWN,
2093
- LLM_TENSOR_FFN_GATE,
2094
- LLM_TENSOR_FFN_NORM,
2095
- LLM_TENSOR_FFN_UP,
2096
- LLM_TENSOR_SHORTCONV_CONV,
2097
- LLM_TENSOR_SHORTCONV_INPROJ,
2098
- LLM_TENSOR_SHORTCONV_OUTPROJ,
2099
- LLM_TENSOR_TOKEN_EMBD,
2100
- LLM_TENSOR_OUTPUT_NORM_LFM2,
2101
- LLM_TENSOR_FFN_GATE_INP,
2102
- LLM_TENSOR_FFN_GATE_EXPS,
2103
- LLM_TENSOR_FFN_DOWN_EXPS,
2104
- LLM_TENSOR_FFN_UP_EXPS,
2105
- LLM_TENSOR_FFN_EXP_PROBS_B,
2106
- };
2107
- case LLM_ARCH_SMALLTHINKER:
2108
- return {
2109
- LLM_TENSOR_TOKEN_EMBD,
2110
- LLM_TENSOR_OUTPUT_NORM,
2111
- LLM_TENSOR_OUTPUT,
2112
- LLM_TENSOR_ATTN_NORM,
2113
- LLM_TENSOR_ATTN_Q,
2114
- LLM_TENSOR_ATTN_K,
2115
- LLM_TENSOR_ATTN_V,
2116
- LLM_TENSOR_ATTN_OUT,
2117
- LLM_TENSOR_FFN_NORM,
2118
- LLM_TENSOR_FFN_GATE,
2119
- LLM_TENSOR_FFN_DOWN,
2120
- LLM_TENSOR_FFN_UP,
2121
- LLM_TENSOR_FFN_GATE_INP,
2122
- LLM_TENSOR_FFN_GATE_EXPS,
2123
- LLM_TENSOR_FFN_DOWN_EXPS,
2124
- LLM_TENSOR_FFN_UP_EXPS,
2125
- };
2126
- case LLM_ARCH_APERTUS:
2127
- return {
2128
- LLM_TENSOR_TOKEN_EMBD,
2129
- LLM_TENSOR_OUTPUT_NORM,
2130
- LLM_TENSOR_OUTPUT,
2131
- LLM_TENSOR_ROPE_FREQS,
2132
- LLM_TENSOR_ATTN_NORM,
2133
- LLM_TENSOR_ATTN_Q,
2134
- LLM_TENSOR_ATTN_K,
2135
- LLM_TENSOR_ATTN_V,
2136
- LLM_TENSOR_ATTN_OUT,
2137
- LLM_TENSOR_ATTN_Q_NORM,
2138
- LLM_TENSOR_ATTN_K_NORM,
2139
- LLM_TENSOR_FFN_NORM,
2140
- LLM_TENSOR_FFN_DOWN,
2141
- LLM_TENSOR_FFN_UP,
2142
- };
2143
- case LLM_ARCH_SEED_OSS:
2144
- return {
2145
- LLM_TENSOR_TOKEN_EMBD,
2146
- LLM_TENSOR_OUTPUT_NORM,
2147
- LLM_TENSOR_OUTPUT,
2148
- LLM_TENSOR_ATTN_NORM,
2149
- LLM_TENSOR_ATTN_Q,
2150
- LLM_TENSOR_ATTN_K,
2151
- LLM_TENSOR_ATTN_V,
2152
- LLM_TENSOR_ATTN_OUT,
2153
- LLM_TENSOR_ATTN_POST_NORM,
2154
- LLM_TENSOR_FFN_GATE,
2155
- LLM_TENSOR_FFN_DOWN,
2156
- LLM_TENSOR_FFN_UP,
2157
- };
2158
- case LLM_ARCH_GROVEMOE:
2159
- return {
2160
- LLM_TENSOR_TOKEN_EMBD,
2161
- LLM_TENSOR_OUTPUT_NORM,
2162
- LLM_TENSOR_OUTPUT,
2163
- LLM_TENSOR_ATTN_NORM,
2164
- LLM_TENSOR_ATTN_Q,
2165
- LLM_TENSOR_ATTN_Q_NORM,
2166
- LLM_TENSOR_ATTN_K,
2167
- LLM_TENSOR_ATTN_K_NORM,
2168
- LLM_TENSOR_ATTN_V,
2169
- LLM_TENSOR_ATTN_OUT,
2170
- LLM_TENSOR_FFN_NORM,
2171
- LLM_TENSOR_FFN_GATE_INP,
2172
- LLM_TENSOR_FFN_GATE_EXPS,
2173
- LLM_TENSOR_FFN_DOWN_EXPS,
2174
- LLM_TENSOR_FFN_UP_EXPS,
2175
- LLM_TENSOR_FFN_GATE_CHEXPS,
2176
- LLM_TENSOR_FFN_DOWN_CHEXPS,
2177
- LLM_TENSOR_FFN_UP_CHEXPS,
2178
- };
2179
- case LLM_ARCH_MINIMAX_M2:
2180
- return {
2181
- LLM_TENSOR_TOKEN_EMBD,
2182
- LLM_TENSOR_OUTPUT_NORM,
2183
- LLM_TENSOR_OUTPUT,
2184
- LLM_TENSOR_ATTN_NORM,
2185
- LLM_TENSOR_ATTN_Q,
2186
- LLM_TENSOR_ATTN_K,
2187
- LLM_TENSOR_ATTN_V,
2188
- LLM_TENSOR_ATTN_OUT,
2189
- LLM_TENSOR_ATTN_Q_NORM,
2190
- LLM_TENSOR_ATTN_K_NORM,
2191
- LLM_TENSOR_FFN_NORM,
2192
- LLM_TENSOR_FFN_GATE_INP,
2193
- LLM_TENSOR_FFN_GATE_EXPS,
2194
- LLM_TENSOR_FFN_DOWN_EXPS,
2195
- LLM_TENSOR_FFN_UP_EXPS,
2196
- LLM_TENSOR_FFN_EXP_PROBS_B,
2197
- };
2198
- case LLM_ARCH_COGVLM:
2199
- return {
2200
- LLM_TENSOR_TOKEN_EMBD,
2201
- LLM_TENSOR_OUTPUT_NORM,
2202
- LLM_TENSOR_OUTPUT,
2203
- LLM_TENSOR_ATTN_NORM,
2204
- LLM_TENSOR_ATTN_QKV,
2205
- LLM_TENSOR_ATTN_OUT,
2206
- LLM_TENSOR_FFN_NORM,
2207
- LLM_TENSOR_FFN_GATE,
2208
- LLM_TENSOR_FFN_DOWN,
2209
- LLM_TENSOR_FFN_UP,
2210
- LLM_TENSOR_VISEXP_ATTN_QKV,
2211
- LLM_TENSOR_VISEXP_ATTN_OUT,
2212
- LLM_TENSOR_VISEXP_FFN_GATE,
2213
- LLM_TENSOR_VISEXP_FFN_DOWN,
2214
- LLM_TENSOR_VISEXP_FFN_UP,
2215
- };
2216
- case LLM_ARCH_MIMO2:
2217
- return {
2218
- LLM_TENSOR_TOKEN_EMBD,
2219
- LLM_TENSOR_OUTPUT_NORM,
2220
- LLM_TENSOR_OUTPUT,
2221
- LLM_TENSOR_ATTN_NORM,
2222
- LLM_TENSOR_ATTN_Q,
2223
- LLM_TENSOR_ATTN_K,
2224
- LLM_TENSOR_ATTN_V,
2225
- LLM_TENSOR_ATTN_SINKS,
2226
- LLM_TENSOR_ATTN_OUT,
2227
- LLM_TENSOR_FFN_NORM,
2228
- LLM_TENSOR_FFN_GATE,
2229
- LLM_TENSOR_FFN_DOWN,
2230
- LLM_TENSOR_FFN_UP,
2231
- LLM_TENSOR_FFN_GATE_INP,
2232
- LLM_TENSOR_FFN_GATE_EXPS,
2233
- LLM_TENSOR_FFN_DOWN_EXPS,
2234
- LLM_TENSOR_FFN_UP_EXPS,
2235
- LLM_TENSOR_FFN_EXP_PROBS_B,
2236
- };
2237
- case LLM_ARCH_GPTJ:
2238
- case LLM_ARCH_UNKNOWN:
2239
- return {
2240
- LLM_TENSOR_TOKEN_EMBD,
2241
- };
2242
- case LLM_ARCH_MAINCODER:
2243
- return {
2244
- LLM_TENSOR_TOKEN_EMBD,
2245
- LLM_TENSOR_OUTPUT_NORM,
2246
- LLM_TENSOR_OUTPUT,
2247
- LLM_TENSOR_ATTN_NORM,
2248
- LLM_TENSOR_ATTN_Q,
2249
- LLM_TENSOR_ATTN_Q_NORM,
2250
- LLM_TENSOR_ATTN_K,
2251
- LLM_TENSOR_ATTN_K_NORM,
2252
- LLM_TENSOR_ATTN_V,
2253
- LLM_TENSOR_ATTN_OUT,
2254
- LLM_TENSOR_FFN_NORM,
2255
- LLM_TENSOR_FFN_GATE,
2256
- LLM_TENSOR_FFN_DOWN,
2257
- LLM_TENSOR_FFN_UP,
2258
- };
2259
- default:
2260
- GGML_ABORT("unknown architecture for tensor mapping");
2261
- }
2262
- }
2263
-
2264
- // declare information about the model weight tensors:
2265
- // - the layer in which the tensor is going to be used. this is needed in order to assign the correct buffer type for the weight
2266
- // - the operator which is going to use the weight. this is needed to determine if the respective backend supports the operator
2267
- //
2268
- // for example, input layers are usually assigned to CPU/host buffer types
2269
- //
2270
- // a mismatch between the declared information and the actual layer/op in which the tensor is used can lead to sub-optimal
2271
- // assignment of the buffer types and extra overhead during computation
2272
- // example: https://github.com/ggml-org/llama.cpp/pull/17548
2273
- //
2274
- static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
2275
- {LLM_TENSOR_TOKEN_EMBD, {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
2276
- {LLM_TENSOR_POS_EMBD, {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
2277
- {LLM_TENSOR_TOKEN_TYPES, {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
2278
- {LLM_TENSOR_TOKEN_EMBD_NORM, {LLM_TENSOR_LAYER_INPUT, GGML_OP_MUL}},
2279
- {LLM_TENSOR_OUTPUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
2280
- {LLM_TENSOR_CLS, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
2281
- {LLM_TENSOR_CLS_OUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
2282
- {LLM_TENSOR_DENSE_2_OUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, // Dense layer output
2283
- {LLM_TENSOR_DENSE_3_OUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, // Dense layer output
2284
- {LLM_TENSOR_OUTPUT_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
2285
- {LLM_TENSOR_OUTPUT_NORM_LFM2, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
2286
- {LLM_TENSOR_DEC_OUTPUT_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
2287
- {LLM_TENSOR_ENC_OUTPUT_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
2288
- {LLM_TENSOR_ROPE_FREQS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ROPE}},
2289
- {LLM_TENSOR_ROPE_FACTORS_LONG, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ROPE}},
2290
- {LLM_TENSOR_ROPE_FACTORS_SHORT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ROPE}},
2291
- {LLM_TENSOR_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2292
- {LLM_TENSOR_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2293
- {LLM_TENSOR_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2294
- {LLM_TENSOR_ATTN_QKV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2295
- {LLM_TENSOR_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2296
- {LLM_TENSOR_ATTN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2297
- {LLM_TENSOR_FFN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2298
- {LLM_TENSOR_FFN_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2299
- {LLM_TENSOR_FFN_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2300
- {LLM_TENSOR_FFN_DOWN_SHEXP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2301
- {LLM_TENSOR_FFN_GATE_SHEXP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2302
- {LLM_TENSOR_FFN_UP_SHEXP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2303
- {LLM_TENSOR_ATTN_Q_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2304
- {LLM_TENSOR_ATTN_Q_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2305
- {LLM_TENSOR_ATTN_KV_A_MQA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2306
- {LLM_TENSOR_ATTN_KV_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2307
- {LLM_TENSOR_ATTN_K_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2308
- {LLM_TENSOR_ATTN_V_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2309
- {LLM_TENSOR_ATTN_SINKS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SCALE}},
2310
- {LLM_TENSOR_DEC_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2311
- {LLM_TENSOR_DEC_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2312
- {LLM_TENSOR_DEC_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2313
- {LLM_TENSOR_DEC_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2314
- {LLM_TENSOR_DEC_CROSS_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2315
- {LLM_TENSOR_DEC_CROSS_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2316
- {LLM_TENSOR_DEC_CROSS_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2317
- {LLM_TENSOR_DEC_CROSS_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2318
- {LLM_TENSOR_DEC_FFN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2319
- {LLM_TENSOR_DEC_FFN_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2320
- {LLM_TENSOR_DEC_FFN_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2321
- {LLM_TENSOR_ENC_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2322
- {LLM_TENSOR_ENC_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2323
- {LLM_TENSOR_ENC_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2324
- {LLM_TENSOR_ENC_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2325
- {LLM_TENSOR_ENC_FFN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2326
- {LLM_TENSOR_ENC_FFN_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2327
- {LLM_TENSOR_ENC_FFN_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2328
- {LLM_TENSOR_FFN_GATE_INP_SHEXP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2329
- {LLM_TENSOR_FFN_GATE_INP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2330
- {LLM_TENSOR_SSM_IN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2331
- {LLM_TENSOR_SSM_X, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2332
- {LLM_TENSOR_SSM_DT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2333
- {LLM_TENSOR_SSM_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2334
- {LLM_TENSOR_SSM_BETA_ALPHA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2335
- {LLM_TENSOR_TIME_MIX_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2336
- {LLM_TENSOR_TIME_MIX_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2337
- {LLM_TENSOR_TIME_MIX_A1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2338
- {LLM_TENSOR_TIME_MIX_A2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2339
- {LLM_TENSOR_TIME_MIX_V1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2340
- {LLM_TENSOR_TIME_MIX_V2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2341
- {LLM_TENSOR_TIME_MIX_G1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2342
- {LLM_TENSOR_TIME_MIX_G2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2343
- {LLM_TENSOR_TIME_MIX_DECAY_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2344
- {LLM_TENSOR_TIME_MIX_DECAY_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2345
- {LLM_TENSOR_TIME_MIX_KEY, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2346
- {LLM_TENSOR_TIME_MIX_VALUE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2347
- {LLM_TENSOR_TIME_MIX_RECEPTANCE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2348
- {LLM_TENSOR_TIME_MIX_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2349
- {LLM_TENSOR_TIME_MIX_OUTPUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2350
- {LLM_TENSOR_CHANNEL_MIX_KEY, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2351
- {LLM_TENSOR_CHANNEL_MIX_RECEPTANCE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2352
- {LLM_TENSOR_CHANNEL_MIX_VALUE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2353
- {LLM_TENSOR_FFN_ACT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_DIV}},
2354
- {LLM_TENSOR_SSM_CONV1D, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
2355
- {LLM_TENSOR_SSM_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_SCAN}},
2356
- {LLM_TENSOR_SSM_A_NOSCAN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, // a version of SSM_A used for MUL instead of SSM_SCAN
2357
- {LLM_TENSOR_SSM_DT_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2358
- {LLM_TENSOR_SSM_B_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2359
- {LLM_TENSOR_SSM_C_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2360
- {LLM_TENSOR_SSM_D, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2361
- {LLM_TENSOR_SSM_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2362
- {LLM_TENSOR_TIME_MIX_LERP_X, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2363
- {LLM_TENSOR_TIME_MIX_LN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2364
- {LLM_TENSOR_CHANNEL_MIX_LERP_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2365
- {LLM_TENSOR_CHANNEL_MIX_LERP_R, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2366
- {LLM_TENSOR_TIME_MIX_K_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2367
- {LLM_TENSOR_TIME_MIX_K_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2368
- {LLM_TENSOR_TIME_MIX_R_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2369
- {LLM_TENSOR_TIME_MIX_LERP_W, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
2370
- {LLM_TENSOR_TIME_MIX_LERP_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
2371
- {LLM_TENSOR_TIME_MIX_LERP_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
2372
- {LLM_TENSOR_TIME_MIX_LERP_R, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
2373
- {LLM_TENSOR_TIME_MIX_LERP_G, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
2374
- {LLM_TENSOR_TIME_MIX_LERP_FUSED, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
2375
- {LLM_TENSOR_TIME_MIX_DECAY, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
2376
- {LLM_TENSOR_TIME_MIX_W0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
2377
- {LLM_TENSOR_TIME_MIX_A0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
2378
- {LLM_TENSOR_TIME_MIX_V0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
2379
- {LLM_TENSOR_TIME_MIX_FIRST, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_RWKV_WKV6}},
2380
- {LLM_TENSOR_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2381
- {LLM_TENSOR_ATTN_NORM_2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2382
- {LLM_TENSOR_ATTN_OUT_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2383
- {LLM_TENSOR_ATTN_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2384
- {LLM_TENSOR_FFN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2385
- {LLM_TENSOR_FFN_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2386
- {LLM_TENSOR_FFN_NORM_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2387
- {LLM_TENSOR_ATTN_Q_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2388
- {LLM_TENSOR_ATTN_K_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2389
- {LLM_TENSOR_LAYER_OUT_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2390
- {LLM_TENSOR_ATTN_Q_A_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2391
- {LLM_TENSOR_ATTN_KV_A_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2392
- {LLM_TENSOR_ATTN_SUB_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2393
- {LLM_TENSOR_FFN_SUB_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2394
- {LLM_TENSOR_DEC_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2395
- {LLM_TENSOR_DEC_CROSS_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2396
- {LLM_TENSOR_DEC_FFN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2397
- {LLM_TENSOR_ENC_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2398
- {LLM_TENSOR_ENC_FFN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2399
- {LLM_TENSOR_DEC_ATTN_REL_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},
2400
- {LLM_TENSOR_ENC_ATTN_REL_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},
2401
- {LLM_TENSOR_FFN_DOWN_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
2402
- {LLM_TENSOR_FFN_GATE_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
2403
- {LLM_TENSOR_FFN_UP_EXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
2404
- {LLM_TENSOR_FFN_DOWN_CHEXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
2405
- {LLM_TENSOR_FFN_GATE_CHEXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
2406
- {LLM_TENSOR_FFN_UP_CHEXPS, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
2407
- {LLM_TENSOR_FFN_EXP_PROBS_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
2408
- // altup / laurel (gemma 3n)
2409
- {LLM_TENSOR_PER_LAYER_TOKEN_EMBD, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
2410
- {LLM_TENSOR_PER_LAYER_MODEL_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
2411
- {LLM_TENSOR_PER_LAYER_PROJ_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
2412
- {LLM_TENSOR_ALTUP_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
2413
- {LLM_TENSOR_ALTUP_UNEMBD_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
2414
- {LLM_TENSOR_PER_LAYER_INP_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2415
- {LLM_TENSOR_PER_LAYER_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2416
- {LLM_TENSOR_PER_LAYER_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2417
- {LLM_TENSOR_ALTUP_CORRECT_COEF, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2418
- {LLM_TENSOR_ALTUP_CORRECT_SCALE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2419
- {LLM_TENSOR_ALTUP_PREDICT_COEF, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2420
- {LLM_TENSOR_ALTUP_ROUTER, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2421
- {LLM_TENSOR_ALTUP_ROUTER_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2422
- {LLM_TENSOR_LAUREL_L, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2423
- {LLM_TENSOR_LAUREL_R, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2424
- {LLM_TENSOR_LAUREL_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2425
- // this tensor is loaded for T5, but never used
2426
- {LLM_TENSOR_DEC_CROSS_ATTN_REL_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
2427
- {LLM_TENSOR_CONV1D, {LLM_TENSOR_LAYER_INPUT, GGML_OP_IM2COL}},
2428
- {LLM_TENSOR_POS_NET_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2429
- {LLM_TENSOR_POS_NET_NORM1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2430
- {LLM_TENSOR_POS_NET_NORM2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2431
- {LLM_TENSOR_POS_NET_CONV1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_IM2COL}},
2432
- {LLM_TENSOR_POS_NET_CONV2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_IM2COL}},
2433
- {LLM_TENSOR_POS_NET_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2434
- {LLM_TENSOR_POS_NET_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2435
- {LLM_TENSOR_POS_NET_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2436
- {LLM_TENSOR_POS_NET_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2437
- {LLM_TENSOR_POS_NET_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2438
- {LLM_TENSOR_CONVNEXT_DW, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_IM2COL}},
2439
- {LLM_TENSOR_CONVNEXT_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2440
- {LLM_TENSOR_CONVNEXT_PW1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2441
- {LLM_TENSOR_CONVNEXT_PW2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2442
- {LLM_TENSOR_CONVNEXT_GAMMA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
2443
- {LLM_TENSOR_SHORTCONV_CONV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
2444
- {LLM_TENSOR_SHORTCONV_INPROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2445
- {LLM_TENSOR_SHORTCONV_OUTPROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2446
- {LLM_TENSOR_VISEXP_ATTN_QKV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2447
- {LLM_TENSOR_VISEXP_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2448
- {LLM_TENSOR_VISEXP_FFN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2449
- {LLM_TENSOR_VISEXP_FFN_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2450
- {LLM_TENSOR_VISEXP_FFN_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
2451
- // NextN/MTP tensors are currently ignored (reserved for future MTP support)
2452
- // These tensors only exist in the last layer(s) and are treated as output tensors
2453
- {LLM_TENSOR_NEXTN_EH_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
2454
- {LLM_TENSOR_NEXTN_EMBED_TOKENS, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
2455
- {LLM_TENSOR_NEXTN_ENORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_GET_ROWS}},
2456
- {LLM_TENSOR_NEXTN_HNORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
2457
- {LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
2458
- {LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
2459
- };
2460
-
2461
- LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}
2462
-
2463
- std::string LLM_KV::operator()(llm_kv kv) const {
2464
- std::string name = ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
2465
-
2466
- if (suffix != nullptr) {
2467
- name += ".";
2468
- name += suffix;
2469
- }
2470
-
2471
- return name;
2472
- }
2473
-
2474
- LLM_TN_IMPL::LLM_TN_IMPL(llm_arch arch, llm_tensor tensor, const char * suffix, int bid, int xid)
2475
- : arch(arch), tensor(tensor), suffix(suffix), bid(bid), xid(xid),
2476
- model_tensors(llm_get_tensor_names(arch)) {}
2477
-
2478
- std::string LLM_TN_IMPL::str() const {
2479
- if (LLM_TENSOR_NAMES.find(tensor) == LLM_TENSOR_NAMES.end()) {
2480
- GGML_ABORT("unknown tensor name for tensor id %d", static_cast<int>(tensor));
2481
- }
2482
-
2483
- if (model_tensors.find(tensor) == model_tensors.end()) {
2484
- return LLM_TENSOR_NAMES.at(tensor);
2485
- }
2486
-
2487
- std::string name = ::format(LLM_TENSOR_NAMES.at(tensor), bid, xid);
2488
- if (suffix != nullptr) {
2489
- name += ".";
2490
- name += suffix;
2491
- }
2492
-
2493
- return name;
2494
- }
2495
-
2496
- const char * llm_arch_name(llm_arch arch) {
2497
- auto it = LLM_ARCH_NAMES.find(arch);
2498
- if (it == LLM_ARCH_NAMES.end()) {
2499
- return "unknown";
2500
- }
2501
- return it->second;
2502
- }
2503
-
2504
- llm_arch llm_arch_from_string(const std::string & name) {
2505
- for (const auto & kv : LLM_ARCH_NAMES) { // NOLINT
2506
- if (kv.second == name) {
2507
- return kv.first;
2508
- }
2509
- }
2510
-
2511
- return LLM_ARCH_UNKNOWN;
2512
- }
2513
-
2514
- const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor) {
2515
- return LLM_TENSOR_INFOS.at(tensor);
2516
- }
2517
-
2518
- bool llm_arch_is_recurrent(const llm_arch & arch) {
2519
- switch (arch) {
2520
- case LLM_ARCH_MAMBA:
2521
- case LLM_ARCH_MAMBA2:
2522
- case LLM_ARCH_RWKV6:
2523
- case LLM_ARCH_RWKV6QWEN2:
2524
- case LLM_ARCH_RWKV7:
2525
- case LLM_ARCH_ARWKV7:
2526
- return true;
2527
- default:
2528
- return false;
2529
- }
2530
- }
2531
-
2532
- bool llm_arch_is_hybrid(const llm_arch & arch) {
2533
- switch (arch) {
2534
- case LLM_ARCH_JAMBA:
2535
- case LLM_ARCH_FALCON_H1:
2536
- case LLM_ARCH_PLAMO2:
2537
- case LLM_ARCH_GRANITE_HYBRID:
2538
- case LLM_ARCH_LFM2:
2539
- case LLM_ARCH_LFM2MOE:
2540
- case LLM_ARCH_NEMOTRON_H:
2541
- case LLM_ARCH_NEMOTRON_H_MOE:
2542
- case LLM_ARCH_QWEN3NEXT:
2543
- return true;
2544
- default:
2545
- return false;
2546
- }
2547
- }
2548
-
2549
- bool llm_arch_is_diffusion(const llm_arch & arch) {
2550
- switch (arch) {
2551
- case LLM_ARCH_DREAM:
2552
- case LLM_ARCH_LLADA:
2553
- case LLM_ARCH_LLADA_MOE:
2554
- case LLM_ARCH_RND1:
2555
- return true;
2556
- default:
2557
- return false;
2558
- }
2559
- }