llama-cpp-pydist 0.19.0__py3-none-any.whl → 0.21.0__py3-none-any.whl

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 (242) hide show
  1. llama_cpp/binaries/{llama-b7488-bin-win-cpu-x64.zip → llama-b7631-bin-win-cpu-x64.zip} +0 -0
  2. llama_cpp_pydist-0.21.0.dist-info/METADATA +4684 -0
  3. {llama_cpp_pydist-0.19.0.dist-info → llama_cpp_pydist-0.21.0.dist-info}/RECORD +240 -222
  4. {llama_cpp_pydist-0.19.0.dist-info → llama_cpp_pydist-0.21.0.dist-info}/WHEEL +1 -1
  5. vendor_llama_cpp_pydist/llama.cpp/.devops/cuda-new.Dockerfile +95 -0
  6. vendor_llama_cpp_pydist/llama.cpp/.gemini/settings.json +1 -0
  7. vendor_llama_cpp_pydist/llama.cpp/.github/ISSUE_TEMPLATE/010-bug-compilation.yml +2 -1
  8. vendor_llama_cpp_pydist/llama.cpp/.github/ISSUE_TEMPLATE/011-bug-results.yml +13 -2
  9. vendor_llama_cpp_pydist/llama.cpp/.github/ISSUE_TEMPLATE/019-bug-misc.yml +13 -2
  10. vendor_llama_cpp_pydist/llama.cpp/.github/workflows/build.yml +18 -6
  11. vendor_llama_cpp_pydist/llama.cpp/.github/workflows/docker.yml +25 -13
  12. vendor_llama_cpp_pydist/llama.cpp/.github/workflows/release.yml +9 -5
  13. vendor_llama_cpp_pydist/llama.cpp/.github/workflows/server.yml +18 -0
  14. vendor_llama_cpp_pydist/llama.cpp/AGENTS.md +81 -0
  15. vendor_llama_cpp_pydist/llama.cpp/CLAUDE.md +1 -0
  16. vendor_llama_cpp_pydist/llama.cpp/CONTRIBUTING.md +34 -5
  17. vendor_llama_cpp_pydist/llama.cpp/ci/run.sh +2 -1
  18. vendor_llama_cpp_pydist/llama.cpp/common/CMakeLists.txt +4 -3
  19. vendor_llama_cpp_pydist/llama.cpp/common/arg.cpp +46 -14
  20. vendor_llama_cpp_pydist/llama.cpp/common/arg.h +1 -0
  21. vendor_llama_cpp_pydist/llama.cpp/common/chat-parser.cpp +11 -0
  22. vendor_llama_cpp_pydist/llama.cpp/common/chat.cpp +36 -7
  23. vendor_llama_cpp_pydist/llama.cpp/common/chat.h +1 -0
  24. vendor_llama_cpp_pydist/llama.cpp/common/common.cpp +42 -23
  25. vendor_llama_cpp_pydist/llama.cpp/common/common.h +7 -2
  26. vendor_llama_cpp_pydist/llama.cpp/common/llguidance.cpp +10 -6
  27. vendor_llama_cpp_pydist/llama.cpp/common/regex-partial.cpp +13 -13
  28. vendor_llama_cpp_pydist/llama.cpp/common/sampling.cpp +58 -14
  29. vendor_llama_cpp_pydist/llama.cpp/common/sampling.h +3 -1
  30. vendor_llama_cpp_pydist/llama.cpp/convert_hf_to_gguf.py +424 -103
  31. vendor_llama_cpp_pydist/llama.cpp/convert_hf_to_gguf_update.py +5 -0
  32. vendor_llama_cpp_pydist/llama.cpp/docs/backend/CANN.md +4 -0
  33. vendor_llama_cpp_pydist/llama.cpp/docs/backend/OPENCL.md +51 -1
  34. vendor_llama_cpp_pydist/llama.cpp/docs/backend/SYCL.md +1 -1
  35. vendor_llama_cpp_pydist/llama.cpp/docs/backend/hexagon/README.md +5 -5
  36. vendor_llama_cpp_pydist/llama.cpp/docs/backend/hexagon/developer.md +1 -1
  37. vendor_llama_cpp_pydist/llama.cpp/docs/build.md +21 -2
  38. vendor_llama_cpp_pydist/llama.cpp/docs/development/parsing.md +2 -2
  39. vendor_llama_cpp_pydist/llama.cpp/docs/ops/Metal.csv +360 -322
  40. vendor_llama_cpp_pydist/llama.cpp/docs/ops.md +1 -1
  41. vendor_llama_cpp_pydist/llama.cpp/ggml/CMakeLists.txt +13 -1
  42. vendor_llama_cpp_pydist/llama.cpp/ggml/include/ggml-backend.h +1 -1
  43. vendor_llama_cpp_pydist/llama.cpp/ggml/src/CMakeLists.txt +23 -9
  44. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-backend.cpp +11 -11
  45. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cann/aclnn_ops.cpp +303 -19
  46. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cann/aclnn_ops.h +17 -0
  47. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cann/common.h +153 -9
  48. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp +51 -158
  49. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt +12 -2
  50. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h +1 -1
  51. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cpu/kleidiai/kernels.cpp +86 -25
  52. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +15 -8
  53. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cpu/llamafile/sgemm.cpp +768 -0
  54. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cpu/simd-mappings.h +0 -4
  55. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/CMakeLists.txt +66 -1
  56. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/argsort.cu +48 -27
  57. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/argsort.cuh +16 -0
  58. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/common.cuh +45 -9
  59. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/cpy.cu +117 -103
  60. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/cumsum.cu +105 -35
  61. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/fattn-common.cuh +3 -1
  62. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/fattn-mma-f16.cuh +2 -2
  63. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu +83 -33
  64. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/mean.cu +3 -0
  65. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/mma.cuh +21 -0
  66. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/mmq.cu +34 -8
  67. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/mmq.cuh +168 -13
  68. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/quantize.cu +151 -0
  69. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/quantize.cuh +14 -0
  70. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/softmax.cu +203 -6
  71. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/top-k.cu +96 -0
  72. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/top-k.cuh +3 -0
  73. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/topk-moe.cu +17 -2
  74. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/topk-moe.cuh +6 -1
  75. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/vendors/cuda.h +4 -0
  76. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/vendors/hip.h +3 -0
  77. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-cuda/vendors/musa.h +1 -0
  78. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-hexagon/ggml-hexagon.cpp +224 -758
  79. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-hexagon/htp/act-ops.c +316 -164
  80. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-hexagon/htp/htp-dma.c +5 -11
  81. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-hexagon/htp/htp-dma.h +46 -15
  82. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-hexagon/htp/hvx-utils.h +9 -3
  83. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-hexagon/htp/main.c +2 -1
  84. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-hexagon/htp/matmul-ops.c +20 -20
  85. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-hexagon/htp-utils.h +1 -0
  86. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-hexagon/op-desc.h +153 -0
  87. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-impl.h +0 -4
  88. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-metal/ggml-metal-device.cpp +57 -0
  89. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-metal/ggml-metal-device.h +2 -0
  90. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-metal/ggml-metal-device.m +5 -0
  91. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-metal/ggml-metal-impl.h +20 -0
  92. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-metal/ggml-metal-ops.cpp +71 -2
  93. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-metal/ggml-metal-ops.h +1 -0
  94. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-metal/ggml-metal.metal +73 -6
  95. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp +134 -13
  96. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-opencl/kernels/cvt.cl +21 -0
  97. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-opencl/kernels/transpose.cl +13 -0
  98. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-rpc/ggml-rpc.cpp +14 -7
  99. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-sycl/CMakeLists.txt +42 -1
  100. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp +742 -315
  101. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/count_experts.comp +51 -0
  102. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/cumsum.comp +28 -14
  103. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/cumsum_multipass1.comp +60 -0
  104. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/cumsum_multipass2.comp +66 -0
  105. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl +1 -7
  106. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/generic_head.glsl +2 -0
  107. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +17 -4
  108. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp +42 -24
  109. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq.comp +11 -0
  110. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vecq_funcs.glsl +115 -0
  111. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +10 -4
  112. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +29 -18
  113. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +19 -16
  114. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_id_funcs.glsl +3 -1
  115. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +10 -4
  116. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp +8 -8
  117. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl +11 -4
  118. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp +4 -1
  119. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +4 -1
  120. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +4 -1
  121. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl +1 -0
  122. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp +4 -1
  123. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/topk_moe.comp +57 -22
  124. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/types.glsl +312 -6
  125. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp +54 -0
  126. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +10 -2
  127. vendor_llama_cpp_pydist/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/xielu.comp +35 -0
  128. vendor_llama_cpp_pydist/llama.cpp/gguf-py/gguf/constants.py +99 -0
  129. vendor_llama_cpp_pydist/llama.cpp/gguf-py/gguf/gguf_writer.py +38 -2
  130. vendor_llama_cpp_pydist/llama.cpp/gguf-py/gguf/tensor_mapping.py +26 -0
  131. vendor_llama_cpp_pydist/llama.cpp/gguf-py/gguf/utility.py +0 -8
  132. vendor_llama_cpp_pydist/llama.cpp/grammars/README.md +3 -0
  133. vendor_llama_cpp_pydist/llama.cpp/include/llama.h +99 -12
  134. vendor_llama_cpp_pydist/llama.cpp/scripts/snapdragon/adb/run-cli.sh +9 -9
  135. vendor_llama_cpp_pydist/llama.cpp/scripts/snapdragon/adb/run-completion.sh +53 -0
  136. vendor_llama_cpp_pydist/llama.cpp/scripts/sync-ggml.last +1 -1
  137. vendor_llama_cpp_pydist/llama.cpp/src/CMakeLists.txt +4 -0
  138. vendor_llama_cpp_pydist/llama.cpp/src/llama-adapter.cpp +12 -3
  139. vendor_llama_cpp_pydist/llama.cpp/src/llama-adapter.h +7 -1
  140. vendor_llama_cpp_pydist/llama.cpp/src/llama-arch.cpp +76 -0
  141. vendor_llama_cpp_pydist/llama.cpp/src/llama-arch.h +7 -0
  142. vendor_llama_cpp_pydist/llama.cpp/src/llama-chat.cpp +11 -0
  143. vendor_llama_cpp_pydist/llama.cpp/src/llama-chat.h +1 -0
  144. vendor_llama_cpp_pydist/llama.cpp/src/llama-context.cpp +625 -40
  145. vendor_llama_cpp_pydist/llama.cpp/src/llama-context.h +43 -1
  146. vendor_llama_cpp_pydist/llama.cpp/src/llama-grammar.cpp +40 -13
  147. vendor_llama_cpp_pydist/llama.cpp/src/llama-grammar.h +2 -0
  148. vendor_llama_cpp_pydist/llama.cpp/src/llama-graph.cpp +166 -2
  149. vendor_llama_cpp_pydist/llama.cpp/src/llama-graph.h +71 -6
  150. vendor_llama_cpp_pydist/llama.cpp/src/llama-hparams.h +6 -5
  151. vendor_llama_cpp_pydist/llama.cpp/src/llama-kv-cache.h +1 -1
  152. vendor_llama_cpp_pydist/llama.cpp/src/llama-mmap.cpp +11 -4
  153. vendor_llama_cpp_pydist/llama.cpp/src/llama-model-loader.cpp +23 -0
  154. vendor_llama_cpp_pydist/llama.cpp/src/llama-model-loader.h +2 -0
  155. vendor_llama_cpp_pydist/llama.cpp/src/llama-model.cpp +329 -26
  156. vendor_llama_cpp_pydist/llama.cpp/src/llama-model.h +13 -2
  157. vendor_llama_cpp_pydist/llama.cpp/src/llama-sampling.cpp +1259 -186
  158. vendor_llama_cpp_pydist/llama.cpp/src/llama-sampling.h +19 -7
  159. vendor_llama_cpp_pydist/llama.cpp/src/llama-vocab.cpp +101 -33
  160. vendor_llama_cpp_pydist/llama.cpp/src/llama-vocab.h +2 -0
  161. vendor_llama_cpp_pydist/llama.cpp/src/llama.cpp +53 -38
  162. vendor_llama_cpp_pydist/llama.cpp/src/models/afmoe.cpp +9 -5
  163. vendor_llama_cpp_pydist/llama.cpp/src/models/bert.cpp +4 -2
  164. vendor_llama_cpp_pydist/llama.cpp/src/models/cogvlm.cpp +5 -3
  165. vendor_llama_cpp_pydist/llama.cpp/src/models/cohere2-iswa.cpp +3 -0
  166. vendor_llama_cpp_pydist/llama.cpp/src/models/deepseek2.cpp +1 -1
  167. vendor_llama_cpp_pydist/llama.cpp/src/models/gemma-embedding.cpp +2 -6
  168. vendor_llama_cpp_pydist/llama.cpp/src/models/gemma2-iswa.cpp +5 -2
  169. vendor_llama_cpp_pydist/llama.cpp/src/models/gemma3.cpp +3 -4
  170. vendor_llama_cpp_pydist/llama.cpp/src/models/gemma3n-iswa.cpp +4 -7
  171. vendor_llama_cpp_pydist/llama.cpp/src/models/llama-iswa.cpp +6 -2
  172. vendor_llama_cpp_pydist/llama.cpp/src/models/llama.cpp +19 -6
  173. vendor_llama_cpp_pydist/llama.cpp/src/models/maincoder.cpp +117 -0
  174. vendor_llama_cpp_pydist/llama.cpp/src/models/mimo2-iswa.cpp +123 -0
  175. vendor_llama_cpp_pydist/llama.cpp/src/models/models.h +18 -0
  176. vendor_llama_cpp_pydist/llama.cpp/src/models/modern-bert.cpp +116 -0
  177. vendor_llama_cpp_pydist/llama.cpp/src/models/openai-moe-iswa.cpp +5 -2
  178. vendor_llama_cpp_pydist/llama.cpp/src/models/plamo3.cpp +128 -0
  179. vendor_llama_cpp_pydist/llama.cpp/src/models/smallthinker.cpp +11 -5
  180. vendor_llama_cpp_pydist/llama.cpp/src/unicode.cpp +23 -14
  181. vendor_llama_cpp_pydist/llama.cpp/tests/CMakeLists.txt +12 -2
  182. vendor_llama_cpp_pydist/llama.cpp/tests/test-backend-ops.cpp +286 -65
  183. vendor_llama_cpp_pydist/llama.cpp/tests/test-backend-sampler.cpp +1237 -0
  184. vendor_llama_cpp_pydist/llama.cpp/tests/test-chat.cpp +29 -3
  185. vendor_llama_cpp_pydist/llama.cpp/tests/test-grammar-llguidance.cpp +3 -0
  186. vendor_llama_cpp_pydist/llama.cpp/tests/test-regex-partial.cpp +14 -14
  187. vendor_llama_cpp_pydist/llama.cpp/tests/test-tokenizer-0.cpp +1 -1
  188. vendor_llama_cpp_pydist/llama.cpp/tests/test-tokenizer-1-bpe.cpp +1 -1
  189. vendor_llama_cpp_pydist/llama.cpp/tests/test-tokenizer-1-spm.cpp +1 -1
  190. vendor_llama_cpp_pydist/llama.cpp/tools/batched-bench/batched-bench.cpp +11 -0
  191. vendor_llama_cpp_pydist/llama.cpp/tools/cli/README.md +187 -1
  192. vendor_llama_cpp_pydist/llama.cpp/tools/cli/cli.cpp +1 -3
  193. vendor_llama_cpp_pydist/llama.cpp/tools/completion/README.md +179 -7
  194. vendor_llama_cpp_pydist/llama.cpp/tools/completion/completion.cpp +4 -1
  195. vendor_llama_cpp_pydist/llama.cpp/tools/fit-params/fit-params.cpp +3 -3
  196. vendor_llama_cpp_pydist/llama.cpp/tools/llama-bench/llama-bench.cpp +18 -1
  197. vendor_llama_cpp_pydist/llama.cpp/tools/mtmd/CMakeLists.txt +1 -0
  198. vendor_llama_cpp_pydist/llama.cpp/tools/mtmd/clip-impl.h +12 -7
  199. vendor_llama_cpp_pydist/llama.cpp/tools/mtmd/clip-model.h +3 -1
  200. vendor_llama_cpp_pydist/llama.cpp/tools/mtmd/clip.cpp +118 -4
  201. vendor_llama_cpp_pydist/llama.cpp/tools/mtmd/models/models.h +10 -0
  202. vendor_llama_cpp_pydist/llama.cpp/tools/mtmd/models/siglip.cpp +9 -4
  203. vendor_llama_cpp_pydist/llama.cpp/tools/mtmd/models/whisper-enc.cpp +9 -0
  204. vendor_llama_cpp_pydist/llama.cpp/tools/mtmd/models/youtuvl.cpp +179 -0
  205. vendor_llama_cpp_pydist/llama.cpp/tools/mtmd/mtmd.cpp +5 -1
  206. vendor_llama_cpp_pydist/llama.cpp/tools/mtmd/mtmd.h +3 -0
  207. vendor_llama_cpp_pydist/llama.cpp/tools/quantize/quantize.cpp +6 -0
  208. vendor_llama_cpp_pydist/llama.cpp/tools/server/CMakeLists.txt +0 -8
  209. vendor_llama_cpp_pydist/llama.cpp/tools/server/README-dev.md +2 -0
  210. vendor_llama_cpp_pydist/llama.cpp/tools/server/README.md +27 -14
  211. vendor_llama_cpp_pydist/llama.cpp/tools/server/public/index.html.gz +0 -0
  212. vendor_llama_cpp_pydist/llama.cpp/tools/server/server-common.cpp +22 -24
  213. vendor_llama_cpp_pydist/llama.cpp/tools/server/server-common.h +2 -3
  214. vendor_llama_cpp_pydist/llama.cpp/tools/server/server-context.cpp +453 -267
  215. vendor_llama_cpp_pydist/llama.cpp/tools/server/server-context.h +52 -15
  216. vendor_llama_cpp_pydist/llama.cpp/tools/server/server-http.cpp +16 -10
  217. vendor_llama_cpp_pydist/llama.cpp/tools/server/server-models.cpp +174 -62
  218. vendor_llama_cpp_pydist/llama.cpp/tools/server/server-models.h +14 -5
  219. vendor_llama_cpp_pydist/llama.cpp/tools/server/server-queue.cpp +78 -21
  220. vendor_llama_cpp_pydist/llama.cpp/tools/server/server-queue.h +48 -10
  221. vendor_llama_cpp_pydist/llama.cpp/tools/server/server-task.cpp +36 -11
  222. vendor_llama_cpp_pydist/llama.cpp/tools/server/server-task.h +28 -35
  223. vendor_llama_cpp_pydist/llama.cpp/tools/server/server.cpp +9 -5
  224. vendor_llama_cpp_pydist/llama.cpp/tools/server/tests/unit/test_chat_completion.py +11 -2
  225. vendor_llama_cpp_pydist/llama.cpp/tools/server/tests/unit/test_sleep.py +39 -0
  226. vendor_llama_cpp_pydist/llama.cpp/tools/server/tests/utils.py +3 -0
  227. vendor_llama_cpp_pydist/llama.cpp/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageAssistant.svelte +25 -1
  228. vendor_llama_cpp_pydist/llama.cpp/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageStatistics.svelte +66 -13
  229. vendor_llama_cpp_pydist/llama.cpp/tools/server/webui/src/lib/components/app/chat/ChatSettings/ChatSettings.svelte +5 -0
  230. vendor_llama_cpp_pydist/llama.cpp/tools/server/webui/src/lib/constants/settings-config.ts +3 -0
  231. vendor_llama_cpp_pydist/llama.cpp/tools/server/webui/src/lib/hooks/use-processing-state.svelte.ts +125 -11
  232. vendor_llama_cpp_pydist/llama.cpp/tools/server/webui/src/lib/services/chat.ts +15 -8
  233. vendor_llama_cpp_pydist/llama.cpp/tools/server/webui/src/lib/stores/chat.svelte.ts +12 -3
  234. vendor_llama_cpp_pydist/llama.cpp/tools/server/webui/src/lib/stores/settings.svelte.ts +4 -5
  235. vendor_llama_cpp_pydist/llama.cpp/tools/server/webui/src/lib/types/api.d.ts +5 -0
  236. vendor_llama_cpp_pydist/llama.cpp/tools/server/webui/src/lib/types/settings.d.ts +2 -1
  237. vendor_llama_cpp_pydist/llama.cpp/tools/server/webui/src/lib/utils/clipboard.ts +1 -4
  238. vendor_llama_cpp_pydist/llama.cpp/tools/server/webui/src/routes/+layout.svelte +1 -1
  239. llama_cpp_pydist-0.19.0.dist-info/METADATA +0 -2506
  240. vendor_llama_cpp_pydist/llama.cpp/.github/copilot-instructions.md +0 -262
  241. {llama_cpp_pydist-0.19.0.dist-info/licenses → llama_cpp_pydist-0.21.0.dist-info}/LICENSE +0 -0
  242. {llama_cpp_pydist-0.19.0.dist-info → llama_cpp_pydist-0.21.0.dist-info}/top_level.txt +0 -0
vendor_llama_cpp_pydist/llama.cpp/src/models/plamo3.cpp ADDED
@@ -0,0 +1,128 @@
1
+ #include "models.h"
2
+
3
+ template <bool iswa>
4
+ llm_build_plamo3<iswa>::llm_build_plamo3(const llama_model & model, const llm_graph_params & params) :
5
+ llm_graph_context(params) {
6
+ const int64_t head_dim_q = hparams.n_embd_head_k;
7
+ const int64_t head_dim_v = hparams.n_embd_head_v;
8
+
9
+ ggml_tensor * cur;
10
+ ggml_tensor * inpL = build_inp_embd(model.tok_embd);
11
+ ggml_tensor * inp_pos = build_inp_pos();
12
+
13
+ using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
14
+ inp_attn_type * inp_attn = nullptr;
15
+
16
+ if constexpr (iswa) {
17
+ inp_attn = build_attn_inp_kv_iswa();
18
+ } else {
19
+ inp_attn = build_attn_inp_kv();
20
+ }
21
+
22
+ ggml_tensor * inp_out_ids = build_inp_out_ids();
23
+
24
+ for (int il = 0; il < n_layer; ++il) {
25
+ ggml_tensor * residual = inpL;
26
+
27
+ float freq_base_l = 0.0f;
28
+ float freq_scale_l = 0.0f;
29
+ if constexpr (iswa) {
30
+ freq_base_l = model.get_rope_freq_base (cparams, il);
31
+ freq_scale_l = model.get_rope_freq_scale(cparams, il);
32
+ } else {
33
+ freq_base_l = freq_base;
34
+ freq_scale_l = freq_scale;
35
+ }
36
+
37
+ cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
38
+ cb(cur, "attn_norm", il);
39
+
40
+ ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);
41
+ cb(cur, "wqkv", il);
42
+
43
+ const int32_t n_head = hparams.n_head(il);
44
+ const int32_t n_head_kv = hparams.n_head_kv(il);
45
+
46
+ const int64_t q_offset = 0;
47
+ const int64_t k_offset = head_dim_q * n_head;
48
+ const int64_t v_offset = k_offset + head_dim_q * n_head_kv;
49
+
50
+ ggml_tensor * Qcur = ggml_view_3d(ctx0, qkv, head_dim_q, n_head, n_tokens,
51
+ head_dim_q * sizeof(float), qkv->nb[1], q_offset * ggml_element_size(qkv));
52
+ ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, head_dim_q, n_head_kv, n_tokens,
53
+ head_dim_q * sizeof(float), qkv->nb[1], k_offset * ggml_element_size(qkv));
54
+ ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, head_dim_v, n_head_kv, n_tokens,
55
+ head_dim_v * sizeof(float), qkv->nb[1], v_offset * ggml_element_size(qkv));
56
+
57
+ cb(Qcur, "Qcur", il);
58
+ cb(Kcur, "Kcur", il);
59
+ cb(Vcur, "Vcur", il);
60
+
61
+ Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
62
+ cb(Qcur, "attn_q_norm", il);
63
+ Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
64
+ cb(Kcur, "attn_k_norm", il);
65
+
66
+ Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr,
67
+ n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
68
+ ext_factor, attn_factor, beta_fast, beta_slow);
69
+ Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr,
70
+ n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
71
+ ext_factor, attn_factor, beta_fast, beta_slow);
72
+
73
+ const float attn_scale = 1.0f / sqrtf(float(head_dim_q));
74
+
75
+ cur = build_attn(inp_attn,
76
+ model.layers[il].wo, NULL,
77
+ Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, attn_scale, il);
78
+ cb(cur, "attn_out", il);
79
+
80
+ if (il == n_layer - 1 && inp_out_ids) {
81
+ cur = ggml_get_rows(ctx0, cur, inp_out_ids);
82
+ residual = ggml_get_rows(ctx0, residual, inp_out_ids);
83
+ }
84
+
85
+ cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
86
+ cb(cur, "attn_post_norm", il);
87
+
88
+ cur = ggml_add(ctx0, cur, residual);
89
+ cb(cur, "attn_residual", il);
90
+
91
+ residual = cur;
92
+
93
+ cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
94
+ cb(cur, "ffn_norm", il);
95
+
96
+ cur = build_ffn(cur,
97
+ model.layers[il].ffn_up, NULL, NULL,
98
+ NULL, NULL, NULL,
99
+ model.layers[il].ffn_down, NULL, NULL,
100
+ NULL,
101
+ LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
102
+ cb(cur, "ffn_out", il);
103
+
104
+ cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il);
105
+ cb(cur, "ffn_post_norm", il);
106
+
107
+ cur = ggml_add(ctx0, cur, residual);
108
+ cb(cur, "ffn_residual", il);
109
+
110
+ cur = build_cvec(cur, il);
111
+ cb(cur, "l_out", il);
112
+ inpL = cur;
113
+ }
114
+
115
+ cur = inpL;
116
+
117
+ cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
118
+ res->t_embd = cur;
119
+
120
+ cur = build_lora_mm(model.output, cur);
121
+ res->t_logits = cur;
122
+
123
+ ggml_build_forward_expand(gf, cur);
124
+ }
125
+
126
+ // Explicit template instantiations
127
+ template struct llm_build_plamo3<false>;
128
+ template struct llm_build_plamo3<true>;
vendor_llama_cpp_pydist/llama.cpp/src/models/smallthinker.cpp CHANGED
@@ -26,10 +26,16 @@ llm_build_smallthinker<iswa>::llm_build_smallthinker(const llama_model & model,
26
26
  ggml_tensor * inp_out_ids = build_inp_out_ids();
27
27
 
28
28
  for (int il = 0; il < n_layer; ++il) {
29
+ const float freq_base_l = model.get_rope_freq_base (cparams, il);
30
+ const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
31
+
29
32
  ggml_tensor * inpSA = inpL;
30
- ggml_tensor * probs = nullptr;
31
33
 
32
- probs = build_lora_mm(model.layers[il].ffn_gate_inp, inpL); // [n_expert, n_tokens]
34
+ // This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous
35
+ const bool use_rope = hparams.n_no_rope_layer_step == n_layer ||
36
+ il % hparams.n_no_rope_layer_step != 0;
37
+
38
+ ggml_tensor * probs = build_lora_mm(model.layers[il].ffn_gate_inp, inpL); // [n_expert, n_tokens]
33
39
  cb(probs, "ffn_moe_logits", il);
34
40
 
35
41
  // norm
@@ -52,11 +58,11 @@ llm_build_smallthinker<iswa>::llm_build_smallthinker(const llama_model & model,
52
58
  Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
53
59
  Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
54
60
 
55
- if (hparams.n_no_rope_layer_step == n_layer || il % hparams.n_no_rope_layer_step != 0) {
56
- Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
61
+ if (use_rope) {
62
+ Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
57
63
  ext_factor, attn_factor, beta_fast, beta_slow);
58
64
 
59
- Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
65
+ Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
60
66
  ext_factor, attn_factor, beta_fast, beta_slow);
61
67
  }
62
68
  cb(Qcur, "Qcur", il);
vendor_llama_cpp_pydist/llama.cpp/src/unicode.cpp CHANGED
@@ -964,6 +964,11 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
964
964
  { "\\p{P}", unicode_cpt_flags::PUNCTUATION },
965
965
  { "\\p{M}", unicode_cpt_flags::ACCENT_MARK },
966
966
  { "\\p{S}", unicode_cpt_flags::SYMBOL },
967
+ { "\\p{Lu}", unicode_cpt_flags::LETTER }, // Uppercase letter
968
+ { "\\p{Ll}", unicode_cpt_flags::LETTER }, // Lowercase letter
969
+ { "\\p{Lt}", unicode_cpt_flags::LETTER }, // Titlecase letter
970
+ { "\\p{Lm}", unicode_cpt_flags::LETTER }, // Modifier letter
971
+ { "\\p{Lo}", unicode_cpt_flags::LETTER }, // Other letter
967
972
  };
968
973
 
969
974
  static const std::map<int, int> k_ucat_cpt = {
@@ -1074,22 +1079,26 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
1074
1079
  continue;
1075
1080
  }
1076
1081
 
1077
- if (regex_expr[i + 0] == '\\' && i + 4 < regex_expr.size() &&
1082
+ // Match \p{...} Unicode properties of varying lengths
1083
+ if (regex_expr[i + 0] == '\\' && i + 3 < regex_expr.size() &&
1078
1084
  regex_expr[i + 1] == 'p' &&
1079
- regex_expr[i + 2] == '{' &&
1080
- regex_expr[i + 4] == '}') {
1081
- const std::string pat = regex_expr.substr(i, 5);
1082
- if (k_ucat_enum.find(pat) != k_ucat_enum.end()) {
1083
- if (!inside) {
1084
- regex_expr_collapsed += '[';
1085
+ regex_expr[i + 2] == '{') {
1086
+ // Find the closing brace
1087
+ size_t closing_brace = regex_expr.find('}', i + 3);
1088
+ if (closing_brace != std::string::npos && closing_brace <= i + 10) { // reasonable limit
1089
+ const std::string pat = regex_expr.substr(i, closing_brace - i + 1);
1090
+ if (k_ucat_enum.find(pat) != k_ucat_enum.end()) {
1091
+ if (!inside) {
1092
+ regex_expr_collapsed += '[';
1093
+ }
1094
+ regex_expr_collapsed += k_ucat_cpt.at(k_ucat_enum.at(pat));
1095
+ regex_expr_collapsed += k_ucat_map.at(k_ucat_enum.at(pat));
1096
+ if (!inside) {
1097
+ regex_expr_collapsed += ']';
1098
+ }
1099
+ i = closing_brace;
1100
+ continue;
1085
1101
  }
1086
- regex_expr_collapsed += k_ucat_cpt.at(k_ucat_enum.at(pat));
1087
- regex_expr_collapsed += k_ucat_map.at(k_ucat_enum.at(pat));
1088
- if (!inside) {
1089
- regex_expr_collapsed += ']';
1090
- }
1091
- i += 4;
1092
- continue;
1093
1102
  }
1094
1103
  }
1095
1104
 
vendor_llama_cpp_pydist/llama.cpp/tests/CMakeLists.txt CHANGED
@@ -219,8 +219,18 @@ endif()
219
219
  llama_build_and_test(test-gguf.cpp)
220
220
  llama_build_and_test(test-backend-ops.cpp)
221
221
 
222
- llama_build_and_test(test-model-load-cancel.cpp LABEL "model")
223
- llama_build_and_test(test-autorelease.cpp LABEL "model")
222
+ llama_build_and_test(test-model-load-cancel.cpp LABEL "model")
223
+ llama_build_and_test(test-autorelease.cpp LABEL "model")
224
+ llama_build_and_test(test-backend-sampler.cpp LABEL "model")
225
+
226
+ llama_test(test-backend-sampler NAME test-backend-sampler-greedy ARGS --test greedy)
227
+ llama_test(test-backend-sampler NAME test-backend-sampler-temp ARGS --test temp)
228
+ llama_test(test-backend-sampler NAME test-backend-sampler-top_k ARGS --test top_k)
229
+ llama_test(test-backend-sampler NAME test-backend-sampler-dist ARGS --test dist)
230
+ llama_test(test-backend-sampler NAME test-backend-sampler-dist-and-cpu ARGS --test dist_and_cpu)
231
+ llama_test(test-backend-sampler NAME test-backend-sampler-logit-bias ARGS --test logit_bias)
232
+ llama_test(test-backend-sampler NAME test-backend-sampler-mul_seq ARGS --test multi_sequence)
233
+ llama_test(test-backend-sampler NAME test-backend-sampler-set-sampler ARGS --test set_sampler)
224
234
 
225
235
  # Test for state restore with fragmented KV cache
226
236
  # Requires a model, uses same args pattern as test-thread-safety