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/tools/server/README.md CHANGED
@@ -23,9 +23,11 @@ For the ful list of features, please refer to [server's changelog](https://githu
23
23
 
24
24
  ## Usage
25
25
 
26
- <!-- Note for contributors: The list below is generated by llama-gen-docs -->
26
+ <!-- HELP_START -->
27
27
 
28
- **Common params**
28
+ <!-- IMPORTANT: The list below is auto-generated by llama-gen-docs; do NOT modify it manually -->
29
+
30
+ ### Common params
29
31
 
30
32
  | Argument | Explanation |
31
33
  | -------- | ----------- |
@@ -38,13 +40,13 @@ For the ful list of features, please refer to [server's changelog](https://githu
38
40
  | `-tb, --threads-batch N` | number of threads to use during batch and prompt processing (default: same as --threads) |
39
41
  | `-C, --cpu-mask M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: "") |
40
42
  | `-Cr, --cpu-range lo-hi` | range of CPUs for affinity. Complements --cpu-mask |
41
- | `--cpu-strict <0\|1>` | use strict CPU placement (default: 0)<br/> |
42
- | `--prio N` | set process/thread priority : low(-1), normal(0), medium(1), high(2), realtime(3) (default: 0)<br/> |
43
- | `--poll <0...100>` | use polling level to wait for work (0 - no polling, default: 50)<br/> |
43
+ | `--cpu-strict <0\|1>` | use strict CPU placement (default: 0) |
44
+ | `--prio N` | set process/thread priority : low(-1), normal(0), medium(1), high(2), realtime(3) (default: 0) |
45
+ | `--poll <0...100>` | use polling level to wait for work (0 - no polling, default: 50) |
44
46
  | `-Cb, --cpu-mask-batch M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask) |
45
47
  | `-Crb, --cpu-range-batch lo-hi` | ranges of CPUs for affinity. Complements --cpu-mask-batch |
46
48
  | `--cpu-strict-batch <0\|1>` | use strict CPU placement (default: same as --cpu-strict) |
47
- | `--prio-batch N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0)<br/> |
49
+ | `--prio-batch N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0) |
48
50
  | `--poll-batch <0\|1>` | use polling to wait for work (default: same as --poll) |
49
51
  | `-c, --ctx-size N` | size of the prompt context (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE) |
50
52
  | `-n, --predict, --n-predict N` | number of tokens to predict (default: -1, -1 = infinity)<br/>(env: LLAMA_ARG_N_PREDICT) |
@@ -114,7 +116,7 @@ For the ful list of features, please refer to [server's changelog](https://githu
114
116
  | `-ctvd, --cache-type-v-draft TYPE` | KV cache data type for V for the draft model<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V_DRAFT) |
115
117
 
116
118
 
117
- **Sampling params**
119
+ ### Sampling params
118
120
 
119
121
  | Argument | Explanation |
120
122
  | -------- | ----------- |
@@ -138,7 +140,7 @@ For the ful list of features, please refer to [server's changelog](https://githu
138
140
  | `--dry-base N` | set DRY sampling base value (default: 1.75) |
139
141
  | `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) |
140
142
  | `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
141
- | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers<br/> |
143
+ | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers |
142
144
  | `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
143
145
  | `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
144
146
  | `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
@@ -151,7 +153,7 @@ For the ful list of features, please refer to [server's changelog](https://githu
151
153
  | `-jf, --json-schema-file FILE` | File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
152
154
 
153
155
 
154
- **Server-specific params**
156
+ ### Server-specific params
155
157
 
156
158
  | Argument | Explanation |
157
159
  | -------- | ----------- |
@@ -159,7 +161,7 @@ For the ful list of features, please refer to [server's changelog](https://githu
159
161
  | `-cram, --cache-ram N` | set the maximum cache size in MiB (default: 8192, -1 - no limit, 0 - disable)[(more info)](https://github.com/ggml-org/llama.cpp/pull/16391)<br/>(env: LLAMA_ARG_CACHE_RAM) |
160
162
  | `-kvu, --kv-unified` | use single unified KV buffer shared across all sequences (default: enabled if number of slots is auto)<br/>(env: LLAMA_ARG_KV_UNIFIED) |
161
163
  | `--context-shift, --no-context-shift` | whether to use context shift on infinite text generation (default: disabled)<br/>(env: LLAMA_ARG_CONTEXT_SHIFT) |
162
- | `-r, --reverse-prompt PROMPT` | halt generation at PROMPT, return control in interactive mode<br/> |
164
+ | `-r, --reverse-prompt PROMPT` | halt generation at PROMPT, return control in interactive mode |
163
165
  | `-sp, --special` | special tokens output enabled (default: false) |
164
166
  | `--warmup, --no-warmup` | whether to perform warmup with an empty run (default: enabled) |
165
167
  | `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) |
@@ -208,8 +210,9 @@ For the ful list of features, please refer to [server's changelog](https://githu
208
210
  | `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
209
211
  | `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
210
212
  | `--prefill-assistant, --no-prefill-assistant` | whether to prefill the assistant's response if the last message is an assistant message (default: prefill enabled)<br/>when this flag is set, if the last message is an assistant message then it will be treated as a full message and not prefilled<br/><br/>(env: LLAMA_ARG_PREFILL_ASSISTANT) |
211
- | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.10, 0.0 = disabled)<br/> |
213
+ | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.10, 0.0 = disabled) |
212
214
  | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
215
+ | `--sleep-idle-seconds SECONDS` | number of seconds of idleness after which the server will sleep (default: -1; -1 = disabled) |
213
216
  | `-td, --threads-draft N` | number of threads to use during generation (default: same as --threads) |
214
217
  | `-tbd, --threads-batch-draft N` | number of threads to use during batch and prompt processing (default: same as --threads-draft) |
215
218
  | `--draft, --draft-n, --draft-max N` | number of tokens to draft for speculative decoding (default: 16)<br/>(env: LLAMA_ARG_DRAFT_MAX) |
@@ -234,6 +237,7 @@ For the ful list of features, please refer to [server's changelog](https://githu
234
237
  | `--vision-gemma-4b-default` | use Gemma 3 4B QAT (note: can download weights from the internet) |
235
238
  | `--vision-gemma-12b-default` | use Gemma 3 12B QAT (note: can download weights from the internet) |
236
239
 
240
+ <!-- HELP_END -->
237
241
 
238
242
  Note: If both command line argument and environment variable are both set for the same param, the argument will take precedence over env var.
239
243
 
@@ -1482,6 +1486,7 @@ The precedence rule for preset options is as follows:
1482
1486
 
1483
1487
  We also offer additional options that are exclusive to presets (these aren't treated as command-line arguments):
1484
1488
  - `load-on-startup` (boolean): Controls whether the model loads automatically when the server starts
1489
+ - `stop-timeout` (int, seconds): After requested unload, wait for this many seconds before forcing termination (default: 10)
1485
1490
 
1486
1491
  ### Routing requests
1487
1492
 
@@ -1567,12 +1572,10 @@ Load a model
1567
1572
 
1568
1573
  Payload:
1569
1574
  - `model`: name of the model to be loaded.
1570
- - `extra_args`: (optional) an array of additional arguments to be passed to the model instance. Note: you must start the server with `--models-allow-extra-args` to enable this feature.
1571
1575
 
1572
1576
  ```json
1573
1577
  {
1574
- "model": "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M",
1575
- "extra_args": ["-n", "128", "--top-k", "4"]
1578
+ "model": "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
1576
1579
  }
1577
1580
  ```
1578
1581
 
@@ -1621,6 +1624,16 @@ Example of an error:
1621
1624
  }
1622
1625
  ```
1623
1626
 
1627
+ ## Sleeping on Idle
1628
+
1629
+ The server supports an automatic sleep mode that activates after a specified period of inactivity (no incoming tasks). This feature, introduced in [PR #18228](https://github.com/ggml-org/llama.cpp/pull/18228), can be enabled using the `--sleep-idle-seconds` command-line argument. It works seamlessly in both single-model and multi-model configurations.
1630
+
1631
+ When the server enters sleep mode, the model and its associated memory (including the KV cache) are unloaded from RAM to conserve resources. Any new incoming task will automatically trigger the model to reload.
1632
+
1633
+ Note that the following endpoints are exempt from being considered as incoming tasks. They do not trigger model reloading and do not reset the idle timer:
1634
+ - `GET /health`
1635
+ - `GET /props`
1636
+
1624
1637
  ## More examples
1625
1638
 
1626
1639
  ### Interactive mode
vendor_llama_cpp_pydist/llama.cpp/tools/server/server-common.cpp CHANGED
@@ -115,26 +115,14 @@ bool lora_should_clear_cache(
115
115
  !lora_all_alora(next));
116
116
  }
117
117
 
118
- std::vector<common_adapter_lora_info> parse_lora_request(
119
- const std::vector<common_adapter_lora_info> & lora_base,
120
- const json & data) {
121
- std::vector<common_adapter_lora_info> lora(lora_base);
122
- int max_idx = lora.size();
123
-
124
- // clear existing value
125
- for (auto & entry : lora) {
126
- entry.scale = 0.0f;
127
- }
118
+ std::map<int, float> parse_lora_request(const json & data) {
119
+ std::map<int, float> lora;
128
120
 
129
121
  // set value
130
122
  for (const auto & entry : data) {
131
123
  int id = json_value(entry, "id", -1);
132
124
  float scale = json_value(entry, "scale", 0.0f);
133
- if (0 <= id && id < max_idx) {
134
- lora[id].scale = scale;
135
- } else {
136
- throw std::runtime_error("invalid adapter id");
137
- }
125
+ lora[id] = scale;
138
126
  }
139
127
 
140
128
  return lora;
@@ -1397,16 +1385,21 @@ json format_response_rerank(
1397
1385
 
1398
1386
  std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int idx) {
1399
1387
  std::vector<llama_token_data> cur;
1400
- const auto * logits = llama_get_logits_ith(ctx, idx);
1401
1388
 
1402
- const llama_model * model = llama_get_model(ctx);
1403
- const llama_vocab * vocab = llama_model_get_vocab(model);
1389
+ const auto * logits = llama_get_logits_ith(ctx, idx);
1390
+ const llama_token * sampled_ids = llama_get_sampled_candidates_ith(ctx, idx);
1404
1391
 
1405
- const int n_vocab = llama_vocab_n_tokens(vocab);
1392
+ const int n_logits = llama_get_sampled_logits_count_ith(ctx, idx);
1406
1393
 
1407
- cur.resize(n_vocab);
1408
- for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
1409
- cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
1394
+ cur.resize(n_logits);
1395
+ if (sampled_ids) {
1396
+ for (int i = 0; i < n_logits; i++) {
1397
+ cur[i] = llama_token_data{sampled_ids[i], logits[i], 0.0f};
1398
+ }
1399
+ } else {
1400
+ for (llama_token token_id = 0; token_id < n_logits; token_id++) {
1401
+ cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
1402
+ }
1410
1403
  }
1411
1404
 
1412
1405
  // sort tokens by logits
@@ -1435,7 +1428,7 @@ std::string safe_json_to_str(const json & data) {
1435
1428
 
1436
1429
  // TODO: reuse llama_detokenize
1437
1430
  template <class Iter>
1438
- static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
1431
+ static std::string tokens_to_str(const llama_vocab * ctx, Iter begin, Iter end) {
1439
1432
  std::string ret;
1440
1433
  for (; begin != end; ++begin) {
1441
1434
  ret += common_token_to_piece(ctx, *begin);
@@ -1445,7 +1438,12 @@ static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
1445
1438
  }
1446
1439
 
1447
1440
  std::string tokens_to_str(llama_context * ctx, const llama_tokens & tokens) {
1448
- return tokens_to_str(ctx, tokens.begin(), tokens.end());
1441
+ auto model = llama_get_model(ctx);
1442
+ return tokens_to_str(llama_model_get_vocab(model), tokens.begin(), tokens.end());
1443
+ }
1444
+
1445
+ std::string tokens_to_str(const llama_vocab * vocab, const llama_tokens & tokens) {
1446
+ return tokens_to_str(vocab, tokens.begin(), tokens.end());
1449
1447
  }
1450
1448
 
1451
1449
  // format incomplete utf-8 multibyte character for output
vendor_llama_cpp_pydist/llama.cpp/tools/server/server-common.h CHANGED
@@ -107,9 +107,7 @@ bool lora_should_clear_cache(
107
107
  const std::vector<common_adapter_lora_info> & current,
108
108
  const std::vector<common_adapter_lora_info> & next);
109
109
 
110
- std::vector<common_adapter_lora_info> parse_lora_request(
111
- const std::vector<common_adapter_lora_info> & lora_base,
112
- const json & data);
110
+ std::map<int, float> parse_lora_request(const json & data);
113
111
 
114
112
  bool are_lora_equal(
115
113
  const std::vector<common_adapter_lora_info> & l1,
@@ -325,6 +323,7 @@ std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int i
325
323
  std::string safe_json_to_str(const json & data);
326
324
 
327
325
  std::string tokens_to_str(llama_context * ctx, const llama_tokens & tokens);
326
+ std::string tokens_to_str(const llama_vocab * vocab, const llama_tokens & tokens);
328
327
 
329
328
  // format incomplete utf-8 multibyte character for output
330
329
  std::string tokens_to_output_formatted_string(const llama_context * ctx, const llama_token token);