bitsandbytes-cpu-fork 0.50.2.dev9__tar.gz → 0.50.2.dev11__tar.gz

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 (166) hide show
  1. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/LICENSE +21 -21
  2. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/MANIFEST.in +54 -35
  3. {bitsandbytes_cpu_fork-0.50.2.dev9/bitsandbytes_cpu_fork.egg-info → bitsandbytes_cpu_fork-0.50.2.dev11}/PKG-INFO +73 -11
  4. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/README.md +72 -10
  5. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/__init__.py +84 -84
  6. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/__main__.py +4 -4
  7. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/_cpu_caps.py +110 -110
  8. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/_ops.py +526 -526
  9. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/autograd/_functions.py +491 -491
  10. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/cpu/ops.py +819 -819
  11. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/cuda/ops.py +1212 -1212
  12. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/default/ops.py +632 -632
  13. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/hpu/ops.py +53 -53
  14. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/mps/ops.py +277 -277
  15. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/triton/kernels_4bit.py +583 -583
  16. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/triton/kernels_8bit_quant.py +195 -195
  17. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/triton/kernels_optim.py +1177 -1177
  18. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/triton/ops.py +304 -304
  19. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/utils.py +94 -94
  20. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/xpu/ops.py +305 -305
  21. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/cextension.py +447 -447
  22. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/cli.py +50 -50
  23. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/consts.py +12 -12
  24. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/cuda_specs.py +111 -111
  25. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/detect_cpu.py +189 -189
  26. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/diagnostics/cuda.py +193 -193
  27. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/diagnostics/main.py +134 -134
  28. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/diagnostics/utils.py +12 -12
  29. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/disk_balancer.py +1687 -1687
  30. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/efst.py +782 -782
  31. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/functional.py +1836 -1836
  32. bitsandbytes_cpu_fork-0.50.2.dev11/bitsandbytes/fused/__init__.py +21 -0
  33. bitsandbytes_cpu_fork-0.50.2.dev11/bitsandbytes/fused/_native.py +66 -0
  34. bitsandbytes_cpu_fork-0.50.2.dev11/bitsandbytes/fused/_ref.py +62 -0
  35. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/gdn_cpu.py +616 -616
  36. bitsandbytes_cpu_fork-0.50.2.dev11/bitsandbytes/libbitsandbytes_cpu.dll +0 -0
  37. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/nn/__init__.py +19 -19
  38. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/nn/modules.py +1220 -1220
  39. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/nn/parametrize.py +206 -206
  40. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/optim/__init__.py +26 -26
  41. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/optim/adagrad.py +187 -187
  42. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/optim/adam.py +346 -346
  43. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/optim/adamw.py +337 -337
  44. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/optim/adamw4bit.py +239 -239
  45. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/optim/ademamix.py +410 -410
  46. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/optim/lamb.py +190 -190
  47. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/optim/lars.py +259 -259
  48. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/optim/lion.py +266 -266
  49. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/optim/optimizer.py +756 -756
  50. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/optim/rmsprop.py +170 -170
  51. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/optim/sgd.py +152 -152
  52. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/q3bit.py +487 -481
  53. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/torch_cpu_kit.py +794 -794
  54. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/utils.py +208 -208
  55. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes_cpu_cli.py +1062 -1062
  56. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11/bitsandbytes_cpu_fork.egg-info}/PKG-INFO +73 -11
  57. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes_cpu_fork.egg-info/SOURCES.txt +10 -0
  58. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/build_manual/build_manual.bat +234 -234
  59. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/build_manual/export.def +44 -44
  60. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/build_manual/export_generic.def +39 -39
  61. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/build_manual/selftest_win.bat +37 -37
  62. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/avx2_gemv_4bit.h +25 -3
  63. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/common.cuh +105 -105
  64. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/compat.cuh +206 -206
  65. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/compat_device.cuh +51 -51
  66. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/cpu_cache.h +160 -15
  67. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/cpu_ops.cpp +114 -32
  68. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/gemm_4bit.cu +168 -168
  69. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/gemm_4bit_common.cuh +105 -105
  70. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/gemm_4bit_mma.cuh +125 -125
  71. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/gemm_4bit_simt.cu +547 -547
  72. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/gemm_4bit_simt.cuh +16 -16
  73. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/gemm_4bit_sm75.cu +416 -416
  74. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/gemm_4bit_sm75.cuh +16 -16
  75. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/gemm_4bit_sm80.cu +699 -699
  76. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/gemm_4bit_sm80.cuh +16 -16
  77. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/kernels.cu +1863 -1863
  78. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/kernels.cuh +85 -85
  79. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/ops.cu +602 -602
  80. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/ops.cuh +151 -151
  81. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/test_avx2_gemv_4bit.cpp +51 -4
  82. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/test_cpu_cache.cpp +1 -0
  83. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/docs_cpu/API_REFERENCE.md +17 -2
  84. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/docs_cpu/API_REFERENCE_EN.md +2 -2
  85. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/docs_cpu/ISA_MATRIX.md +2 -2
  86. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/docs_cpu/MOTION_BOUNDARY.md +3 -3
  87. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/docs_cpu/MOTION_BOUNDARY_EN.md +3 -3
  88. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/docs_cpu/QUICKSTART.md +111 -20
  89. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/docs_cpu/QUICKSTART_EN.md +126 -20
  90. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/docs_cpu/TECHNICAL_GUIDE.md +17 -15
  91. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/docs_cpu/TECHNICAL_GUIDE_EN.md +17 -15
  92. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/docs_cpu/TECH_REPORT.md +99 -8
  93. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/docs_cpu/TECH_REPORT_EN.md +117 -8
  94. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/docs_cpu/check_docs.py +180 -157
  95. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/examples/compile_inference.py +32 -32
  96. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/examples/cpu/cpu_training.py +370 -370
  97. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/examples/int8_inference_huggingface.py +19 -19
  98. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/examples/xpu/benchmark_paged_memory.py +253 -253
  99. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/examples/xpu/paged_xpu_training.py +369 -369
  100. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/pyproject.toml +124 -124
  101. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/setup.py +175 -175
  102. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tests/test_autograd.py +232 -232
  103. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tests/test_cpu_e2e.py +201 -201
  104. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tests/test_cuda_setup_evaluator.py +135 -135
  105. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tests/test_functional.py +1049 -1049
  106. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tests/test_generation.py +122 -122
  107. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tests/test_linear4bit.py +614 -614
  108. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tests/test_linear8bitlt.py +366 -366
  109. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tests/test_modules.py +494 -494
  110. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tests/test_ops.py +474 -474
  111. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tests/test_optim.py +743 -743
  112. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tests/test_parametrize.py +499 -499
  113. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tests/test_rmsnorm_fwd_cpu.py +198 -198
  114. bitsandbytes_cpu_fork-0.50.2.dev11/tools/carve_selftest.py +225 -0
  115. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tools/create_github_release.ps1 +169 -169
  116. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tools/dump_signatures.py +64 -64
  117. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tools/gen_api_reference.py +43 -43
  118. bitsandbytes_cpu_fork-0.50.2.dev11/tools/mft_recover.c +148 -0
  119. bitsandbytes_cpu_fork-0.50.2.dev11/tools/mft_recover_gui.c +643 -0
  120. bitsandbytes_cpu_fork-0.50.2.dev11/tools/mft_recover_gui.manifest +37 -0
  121. bitsandbytes_cpu_fork-0.50.2.dev11/tools/mft_selftest.py +515 -0
  122. bitsandbytes_cpu_fork-0.50.2.dev11/tools/mftcore.h +690 -0
  123. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tools/patch_dev9.py +75 -75
  124. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tools/publish.ps1 +99 -99
  125. bitsandbytes_cpu_fork-0.50.2.dev11/tools/recovery_pipeline_selftest.py +347 -0
  126. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tools/sector_carve.c +233 -15
  127. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tools/sector_carve_gui.c +222 -16
  128. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tools/sector_mirror.c +97 -2
  129. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tools/sector_mirror_gui.c +318 -20
  130. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tools/train_llm_sft.py +136 -111
  131. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tools/verify_help_examples.py +213 -213
  132. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tools/verify_train_release.py +90 -90
  133. bitsandbytes_cpu_fork-0.50.2.dev9/bitsandbytes/libbitsandbytes_cpu.dll +0 -0
  134. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/autograd/__init__.py +0 -0
  135. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/__init__.py +0 -0
  136. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/cpu/__init__.py +0 -0
  137. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/cuda/__init__.py +0 -0
  138. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/default/__init__.py +0 -0
  139. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/hpu/__init__.py +0 -0
  140. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/mps/__init__.py +0 -0
  141. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/triton/__init__.py +0 -0
  142. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/backends/xpu/__init__.py +0 -0
  143. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/diagnostics/__init__.py +0 -0
  144. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/py.typed +0 -0
  145. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes/vcomp140.dll +0 -0
  146. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes_cpu_fork.egg-info/dependency_links.txt +0 -0
  147. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes_cpu_fork.egg-info/entry_points.txt +0 -0
  148. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes_cpu_fork.egg-info/requires.txt +0 -0
  149. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/bitsandbytes_cpu_fork.egg-info/top_level.txt +0 -0
  150. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/build_linux.sh +0 -0
  151. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/build_manual/export_avx512.def +0 -0
  152. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/build_termux.sh +0 -0
  153. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/common.h +0 -0
  154. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/cpu_features.h +0 -0
  155. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/cpu_gdn.cpp +0 -0
  156. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/cpu_ops.h +0 -0
  157. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/pythonInterface.cpp +0 -0
  158. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/xpu_kernels.cpp +0 -0
  159. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/xpu_kernels.h +0 -0
  160. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/xpu_ops.cpp +0 -0
  161. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/csrc/xpu_ops.h +0 -0
  162. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/docs_cpu/EPYC_VALIDATION.md +0 -0
  163. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/docs_cpu/WHEEL_PLATFORM_NOTES.md +0 -0
  164. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/selftest_cpu.c +0 -0
  165. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/setup.cfg +0 -0
  166. {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev11}/tools/verify_linux_cli.sh +0 -0
@@ -1,21 +1,21 @@
1
- MIT License
2
-
3
- Copyright (c) Facebook, Inc. and its affiliates.
4
-
5
- Permission is hereby granted, free of charge, to any person obtaining a copy
6
- of this software and associated documentation files (the "Software"), to deal
7
- in the Software without restriction, including without limitation the rights
8
- to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
- copies of the Software, and to permit persons to whom the Software is
10
- furnished to do so, subject to the following conditions:
11
-
12
- The above copyright notice and this permission notice shall be included in all
13
- copies or substantial portions of the Software.
14
-
15
- THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
- IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
- FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
- AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
- LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
- OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
- SOFTWARE.
1
+ MIT License
2
+
3
+ Copyright (c) Facebook, Inc. and its affiliates.
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
@@ -1,35 +1,54 @@
1
- # Files for the source distribution (sdist).
2
- #
3
- # The inherited manifest referenced the upstream CUDA layout:
4
- # include CMakeLists.txt
5
- # graft csrc
6
- # graft include
7
- # `include/` does not exist in this fork -- it held upstream's CUDA headers -- so
8
- # every sdist build printed "no directories found matching 'include'", and
9
- # CMakeLists.txt is unused now that the wheel ships a prebuilt CPU DLL instead of
10
- # compiling anything.
11
- #
12
- # csrc is kept deliberately: it is what build_manual.bat / build_linux.sh /
13
- # build_termux.sh compile when someone wants to rebuild the kernel themselves,
14
- # which is the difference between an sdist that is merely installable and one that
15
- # is actually useful.
16
- graft csrc
17
- graft build_manual
18
- graft docs_cpu
19
- graft examples
20
- # tools/ holds gen_api_reference.py and verify_help_examples.py, which regenerate and
21
- # check docs_cpu/API_REFERENCE*.md from the tables in bitsandbytes_cpu_cli.py. The
22
- # generated documents ARE shipped, so leaving the generators out would hand users the
23
- # output of scripts they cannot run.
24
- graft tools
25
- # ...but not the Windows binaries inside it. `graft tools` takes the whole directory,
26
- # and the four sector_*.exe disaster-recovery tools were 0.65 MB of a 1.36 MB sdist --
27
- # nearly half its weight, on every platform, for binaries that cannot run on two of the
28
- # three. Their .c sources stay in the sdist and in the repository.
29
- global-exclude *.exe
30
- global-exclude *.exe.manifest
31
- include build_linux.sh
32
- include build_termux.sh
33
- include selftest_cpu.c
34
- include README.md
35
- include LICENSE
1
+ # Files for the source distribution (sdist).
2
+ #
3
+ # The inherited manifest referenced the upstream CUDA layout:
4
+ # include CMakeLists.txt
5
+ # graft csrc
6
+ # graft include
7
+ # `include/` does not exist in this fork -- it held upstream's CUDA headers -- so
8
+ # every sdist build printed "no directories found matching 'include'", and
9
+ # CMakeLists.txt is unused now that the wheel ships a prebuilt CPU DLL instead of
10
+ # compiling anything.
11
+ #
12
+ # csrc is kept deliberately: it is what build_manual.bat / build_linux.sh /
13
+ # build_termux.sh compile when someone wants to rebuild the kernel themselves,
14
+ # which is the difference between an sdist that is merely installable and one that
15
+ # is actually useful.
16
+ graft csrc
17
+ graft build_manual
18
+ graft docs_cpu
19
+ graft examples
20
+ # tools/ holds gen_api_reference.py and verify_help_examples.py, which regenerate and
21
+ # check docs_cpu/API_REFERENCE*.md from the tables in bitsandbytes_cpu_cli.py. The
22
+ # generated documents ARE shipped, so leaving the generators out would hand users the
23
+ # output of scripts they cannot run.
24
+ graft tools
25
+ # ...but not the Windows binaries inside it. `graft tools` takes the whole directory,
26
+ # and the four sector_*.exe disaster-recovery tools were 0.65 MB of a 1.36 MB sdist --
27
+ # nearly half its weight, on every platform, for binaries that cannot run on two of the
28
+ # three. Their .c sources stay in the sdist and in the repository.
29
+ global-exclude *.exe
30
+ global-exclude *.exe.manifest
31
+ # Not shipping byte-code: a stale __pycache__ inside csrc/ or tools/ can otherwise end up in the
32
+ # sdist, and a .pyc built by a different interpreter is worse than useless to the recipient
33
+ # (it is silently ignored, or it shadows nothing). The source is always there.
34
+ global-exclude *.pyc
35
+ global-exclude *.pyo
36
+ prune tools/__pycache__
37
+ # Build outputs are not sources. The tools/ directory holds compiled .obj files next to the .c
38
+ # they came from, and a local Linux build leaves libbitsandbytes_*.so in bitsandbytes/. A sdist
39
+ # built from such a tree carried 5.5 MB of them (a 4.1 MB AVX-512 .so among them) on every
40
+ # platform, because an egg-info SOURCES.txt written by that earlier build kept them listed.
41
+ # The Windows DLL stays: it is the prebuilt kernel this fork ships.
42
+ global-exclude *.obj
43
+ global-exclude *.pdb
44
+ global-exclude *.o
45
+ global-exclude *.a
46
+ global-exclude *.lib
47
+ global-exclude *.exp
48
+ global-exclude *.so
49
+ global-exclude *.dylib
50
+ include build_linux.sh
51
+ include build_termux.sh
52
+ include selftest_cpu.c
53
+ include README.md
54
+ include LICENSE
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: bitsandbytes-cpu-fork
3
- Version: 0.50.2.dev9
3
+ Version: 0.50.2.dev11
4
4
  Summary: bitsandbytes with a pure-CPU training backend (AVX2 / NEON): 8-bit optimizers, fused GDN, gemm_8bit, disk balancer. No GPU required.
5
5
  License-Expression: MIT
6
6
  Project-URL: repository, https://github.com/Ganyvnanoaa09-0906/bitsandbytes-CPU
@@ -36,12 +36,12 @@ Dynamic: license-file
36
36
  > toolkit and no driver** — just an AVX2 x86-64 CPU (or ARM64 NEON), the CPU build of PyTorch,
37
37
  > and 12–16 GB of RAM.
38
38
 
39
- **Version: `0.50.2.dev9`** · import name is unchanged (`import bitsandbytes as bnb`) · MIT.
39
+ **Version: `0.50.2.dev11`** · import name is unchanged (`import bitsandbytes as bnb`) · MIT.
40
40
 
41
41
  What the CPU backend provides: 8-bit blockwise optimizers (≈2.3 bytes of state per parameter
42
42
  instead of 8; the 4-bit optimizer ≈1.03), 4/8-bit layers that run on CPU tensors, the fused
43
43
  Gated DeltaNet kernel for Qwen3-Next / Qwen3.5 linear attention, fused `gemm_8bit`, fused
44
- RMSNorm/SwiGLU, an SSD cold-memory balancer, and a CPU training/concurrency toolbox.
44
+ RMSNorm/SwiGLU (with NT-store fast paths for large outputs), an SSD cold-memory balancer, and a CPU training/concurrency toolbox.
45
45
 
46
46
  ---
47
47
 
@@ -215,6 +215,48 @@ The library never loads a CUDA/ROCm/XPU library: `bnb.cuda`/`bnb.optim` CPU path
215
215
 
216
216
  ---
217
217
 
218
+ ## Full fine-tuning vs LoRA
219
+
220
+ `--mode full` (and `bnb.optim.*` directly) fine-tunes **every parameter**; measured on
221
+ Qwen3.5-0.8B (752 M, 100% trainable, fp32 + AdamW8bit, seq 64, batch 1): **peak RSS
222
+ 10.4 GB, 8.8 s/step**, versus LoRA (5.4 M trainable) at **5.9 GB / 2.65 s/step**. On a
223
+ 15.4 GB machine ~0.8 B full fine-tuning is comfortable; ~1 B is tight (estimator:
224
+ 10.7 GB weights+grads+8-bit state before activations), and above that use LoRA / EFST /
225
+ q3bit, or add activation checkpointing (not yet wired into the generic trainer).
226
+
227
+ ## iGPU image inference (experimental)
228
+
229
+ The OpenCL kit is GEMM-oriented (`nn.Linear`/`Conv1D`); diffusion UNets are conv2d-heavy and
230
+ need an im2col+dGEMM conv engine before the OpenCL path can run them. As a proof of concept,
231
+ `gt1030/igpu_image_ort.py` runs an ONNX-exported SD1.5 through ONNX Runtime's
232
+ **DmlExecutionProvider** on the integrated GPU (local `_sd15_combined`, 256x256, DDIM):
233
+
234
+ | path | 8 steps | per step |
235
+ |---|---|---|
236
+ | CPU diffusers (Torch 2.13, 6 threads) | 25.29 s | 3.16 s |
237
+ | **iGPU ONNX Runtime DirectML** | **13.05 s** | **1.63 s** |
238
+
239
+ ≈**1.9x** faster at 256x256; 512x512 currently fails inside the UNet on the DML EP (activation
240
+ memory / operator coverage) while the same ONNX runs on `CPUExecutionProvider`. Export once with
241
+ `optimum-cli export onnx --task stable-diffusion --opset 17`. A diffusers + `torch_directml`
242
+ variant (`igpu_image_infer.py`) is kept as a comparison; it needs legacy attention and still hits
243
+ DML `chunk`/reshape issues. Long-term the OpenCL backend gets an **fp16 device-resident conv engine**: measured probes
244
+ show per-op OpenCL conv is 2-3x slower than CPU oneDNN and fp32 resident only reaches parity,
245
+ but fp16 resident im2col+Hgemm is **1.3-2x faster than the CPU conv** at SD1.5 shapes
246
+ (538 vs 342 GFLOPS; 1280-channel 16x16 layer 17.3 vs 34.8 ms). The 256-resolution demo
247
+ therefore uses DirectML, while the OpenCL engine is the planned production backend.
248
+
249
+ The fp16 device-resident building blocks are in `gt1030/opencl_fp16_ops.py` +
250
+ `gt1030/kernels_attention.cl` (im2col + CLBlast Hgemm conv, GroupNorm+SiLU, LayerNorm,
251
+ GEGLU, multi-head attention, nearest-2x upsample, add). Clean measurements on the
252
+ Vega 6 (no CPU contention), parity 0.02-1.1% relL2:
253
+ ResNetBlock2D 0.61-0.77x, VAE up-block 0.40x, FFN 0.16-0.23x, attention 0.42-0.99x,
254
+ full Transformer block 0.51-0.71x of CPU time -- i.e. 1.3-6x faster for most blocks.
255
+ The **VAE decoder is now end-to-end on the iGPU** (`gt1030/opencl_vae.py`): fp16 parity
256
+ 2.7-3.7% vs CPU, 1.81 s vs 2.07 s at 256 (0.87x), 7.93 s vs 8.34 s at 512 (0.95x, warm).
257
+ The remaining step is assembling the full 256px UNet graph (fp16 weights ~1.7 GB,
258
+ under the 5,090 MB single-allocation limit; 512px needs activation tiling).
259
+
218
260
  ## 3-bit frozen base + LoRA (`bitsandbytes.q3bit`, experimental)
219
261
 
220
262
  A frozen weight can be stored as blockwise **3-bit** codes (≈11% of fp32 memory) and
@@ -257,6 +299,7 @@ R5-4500U (6C/6T, AVX2, 15.4 GB) / i5-10400 (6C/12T, AVX2, 12 GB), no NVIDIA GPU.
257
299
  | Real model (Qwen3.5-0.8B LoRA, seq 64) | 4 threads beat 6 by **7–10%** — measure with `tune_threads()` |
258
300
  | Fused lm_head 3-bit GEMM (research) | 14–22% faster than fp32 matmul, 89% less weight memory |
259
301
  | 3-bit frozen base + LoRA (Qwen3-1.7B, CPU) | 80 steps: eval loss 4.34 → **4.03**; RSS **2.9 GB vs 8.7 GB** fp32; step 7.7 s vs 5.5 s (native Windows) |
302
+ | Full-parameter fine-tune (Qwen3.5-0.8B, 752 M, fp32 + AdamW8bit, CPU) | **10.4 GB peak RSS**, 8.8 s/step; LoRA on the same model: 5.9 GB, 2.65 s/step |
260
303
  | AdamW4bit step (1M params) | 28.6 ms vs 3.3 ms for the AVX2 8-bit kernel; state 1.031 vs 2.062 B/param |
261
304
 
262
305
  Not every workload is faster than fp32: 4-bit/8-bit kernels trade compute for memory traffic,
@@ -275,8 +318,8 @@ the full trade-off tables, including negative results.
275
318
  - **manylinux_2_17 wheel**: the AVX-512 library inside it is single-threaded (the parallel
276
319
  build is only for the native `manylinux_2_39` wheel). On AVX-512 machines set
277
320
  `BNB_CPU_DISABLE_AVX512=1` to run the AVX2 OpenMP path.
278
- - **aarch64**: there is no PyPI wheel yet; build from the sdist (`build_linux.sh`) or use the
279
- Termux instructions in `docs_cpu/TERMUX_GUIDE.md`.
321
+ - **aarch64**: there is no PyPI wheel yet; build from the sdist (`build_linux.sh`) or follow the
322
+ Termux guide [`TERMUX_GUIDE.md`](https://github.com/Ganyvnanoaa09-0906/bitsandbytes-CPU/blob/HEAD/bitsandbytes/TERMUX_GUIDE.md).
280
323
  - **macOS**: no wheel is published and the CPU backend is not tested there; the sdist build may
281
324
  work with clang + OpenMP, but treat it as unsupported.
282
325
  - 3-bit (and 2-bit dequant-only) is **experimental**: the API and kernels are covered
@@ -295,6 +338,23 @@ bash build_linux.sh --selftest # also build + run the torch-free C se
295
338
  build_manual\build_manual.bat amd # run inside "x64 Native Tools Command Prompt"
296
339
  ```
297
340
 
341
+ ### Termux (Android / ARM64)
342
+
343
+ On-device builds work; there is no aarch64 PyPI wheel, so build from the sdist:
344
+
345
+ ```bash
346
+ pkg update && pkg install -y clang libomp make python
347
+ bash build_termux.sh # -> bitsandbytes/libbitsandbytes_cpu.so
348
+ bash build_termux.sh --selftest # also build + run the torch-free C self-test
349
+ ```
350
+
351
+ Three Termux-specific traps the script already handles: the compiler is `clang++` (there is no
352
+ `g++`), OpenMP needs `pkg install libomp`, and `-march=native` is unsafe under Android's clang.
353
+ The script is also what the automated flow uses to build on a phone **driven from a PC over ADB**
354
+ — see [`TERMUX_GUIDE.md`](https://github.com/Ganyvnanoaa09-0906/bitsandbytes-CPU/blob/HEAD/bitsandbytes/TERMUX_GUIDE.md)
355
+ for the phone-side preparation, the ADB/wireless-debugging path, the ARM64 kernel notes and the
356
+ verification matrix.
357
+
298
358
  Two special wheel builds used by this project are documented in
299
359
  [`docs_cpu/WHEEL_PLATFORM_NOTES.md`](https://github.com/Ganyvnanoaa09-0906/bitsandbytes-CPU/blob/HEAD/bitsandbytes/docs_cpu/WHEEL_PLATFORM_NOTES.md):
300
360
  the zig manylinux_2_17 portable build and the devtoolset-8 manylinux_2_17 **OpenMP** build
@@ -305,12 +365,12 @@ the zig manylinux_2_17 portable build and the devtoolset-8 manylinux_2_17 **Open
305
365
  ```bash
306
366
  bash build_linux.sh --selftest # C kernels: 9 checks (quantize/gemm_8bit/gemv_4bit/
307
367
  # AdamW8bit/vector-vs-scalar/2-3bit dequant/fused GEMM)
308
- python test_efst_regressions.py # EFST: nested blocks, 3-D tensor experts, routing usage
309
- python test_q3bit_kernels.py # 2/3-bit research prototype (old DLL-based tests)
310
- python test_q3bit_api.py # public q3bit API: pack/dequant/gemm, QuantLinear3bit, fallback
311
- python test_adamw4bit.py # restored 4-bit optimizer: state bytes, native vs torch, fp32 tracking
312
- python test_concurrency_plan.py # planner / memory estimate / tune_threads / launcher
313
- python test_disk_balancer_safety.py # data-safety rules for the disk balancer
368
+ python scripts/test/test_efst_regressions.py # EFST: nested blocks, 3-D tensor experts, routing usage
369
+ python scripts/test/test_q3bit_kernels.py # 2/3-bit research prototype (old DLL-based tests)
370
+ python scripts/test/test_q3bit_api.py # public q3bit API: pack/dequant/gemm, QuantLinear3bit, fallback
371
+ python scripts/test/test_adamw4bit.py # restored 4-bit optimizer: state bytes, native vs torch, fp32 tracking
372
+ python scripts/test/test_concurrency_plan.py # planner / memory estimate / tune_threads / launcher
373
+ python scripts/test/test_disk_balancer_safety.py # data-safety rules for the disk balancer
314
374
  ```
315
375
 
316
376
  ## Documentation
@@ -337,6 +397,8 @@ The long-form docs live in the repository (`bitsandbytes/docs_cpu/`, Chinese and
337
397
  | `0.50.2.dev7` | documentation release: README rewritten for the CPU fork (install/wheels/platform matrix, feature table, measured performance, limitations, tests, full dev1→dev7 changelog, EFST/disk-balancer/toolbox examples) |
338
398
  | `0.50.2.dev8` | **restored `AdamW4bit`** (the kernel was dropped when dev4 vendored upstream; now scalar C kernel + torch fallback, 1.031 B/param); **new `bitsandbytes.q3bit` API** — `quantize_3bit`/`dequantize_3bit`/`gemm_3bit`/`gemm_3bit_transposed`, `QuantLinear3bit` (frozen 3-bit base + LoRA), `convert_linears_3bit`, native kernels on Linux with identical pure-torch fallback; CLI reference now includes 4bitopt/q3bit and all 43 examples execute in CI-style verification |
339
399
  | `0.50.2.dev9` | **Windows native kernels**: CI now compiles the MSVC DLL from the current csrc and the export list includes the 4-bit optimizer and 2/3-bit symbols (they were silently missing before); **q3bit real-model validation** on Qwen3-1.7B (frozen 3-bit base + LoRA, eval 4.34→4.03 over 80 steps, RSS −67%, native step 7.7 s vs fp32 5.5 s); fixed the causal-LM **double-shift** label bug in the iGPU training kit (Windows package v0.1.2) |
400
+ | `0.50.2.dev10` | **NTFS recovery toolkit** (`sector_carve`, `mft_recover`, CLI + GUI, selftests) and the **video E2E pipeline** (generation/training, LCM and motion diagnostics); AVX2 4-bit GEMV fixes (wrong FP4 codebook, lazy-init race, non-divisible blocksize); gt1030 probe scripts and docs |
401
+ | `0.50.2.dev11` | **NT-store fast paths**: 4-bit dequant writes bypass the cache above 16 MiB (1.7–2.0× at 32–64 MB) and RMSNorm chooses NT by total footprint (x+res+out vs per-CCX L3); CPUID L3-sharing mask fix and the NT threshold wired to one macro; **byte-reproducible sdist** with an identity gate that audits the uploaded bytes (`--sdist`, 2-DLL policy, 20 adversarial cases) plus a docs gate requiring all six ZH/EN docs; repo regrouping into `scripts/*`; UTF-8 BOM strip; 5 reproducible CPU audio benches; 6 video perf fixes and 3 AR fixes |
340
402
 
341
403
  ## License and attribution
342
404
 
@@ -5,12 +5,12 @@
5
5
  > toolkit and no driver** — just an AVX2 x86-64 CPU (or ARM64 NEON), the CPU build of PyTorch,
6
6
  > and 12–16 GB of RAM.
7
7
 
8
- **Version: `0.50.2.dev9`** · import name is unchanged (`import bitsandbytes as bnb`) · MIT.
8
+ **Version: `0.50.2.dev11`** · import name is unchanged (`import bitsandbytes as bnb`) · MIT.
9
9
 
10
10
  What the CPU backend provides: 8-bit blockwise optimizers (≈2.3 bytes of state per parameter
11
11
  instead of 8; the 4-bit optimizer ≈1.03), 4/8-bit layers that run on CPU tensors, the fused
12
12
  Gated DeltaNet kernel for Qwen3-Next / Qwen3.5 linear attention, fused `gemm_8bit`, fused
13
- RMSNorm/SwiGLU, an SSD cold-memory balancer, and a CPU training/concurrency toolbox.
13
+ RMSNorm/SwiGLU (with NT-store fast paths for large outputs), an SSD cold-memory balancer, and a CPU training/concurrency toolbox.
14
14
 
15
15
  ---
16
16
 
@@ -184,6 +184,48 @@ The library never loads a CUDA/ROCm/XPU library: `bnb.cuda`/`bnb.optim` CPU path
184
184
 
185
185
  ---
186
186
 
187
+ ## Full fine-tuning vs LoRA
188
+
189
+ `--mode full` (and `bnb.optim.*` directly) fine-tunes **every parameter**; measured on
190
+ Qwen3.5-0.8B (752 M, 100% trainable, fp32 + AdamW8bit, seq 64, batch 1): **peak RSS
191
+ 10.4 GB, 8.8 s/step**, versus LoRA (5.4 M trainable) at **5.9 GB / 2.65 s/step**. On a
192
+ 15.4 GB machine ~0.8 B full fine-tuning is comfortable; ~1 B is tight (estimator:
193
+ 10.7 GB weights+grads+8-bit state before activations), and above that use LoRA / EFST /
194
+ q3bit, or add activation checkpointing (not yet wired into the generic trainer).
195
+
196
+ ## iGPU image inference (experimental)
197
+
198
+ The OpenCL kit is GEMM-oriented (`nn.Linear`/`Conv1D`); diffusion UNets are conv2d-heavy and
199
+ need an im2col+dGEMM conv engine before the OpenCL path can run them. As a proof of concept,
200
+ `gt1030/igpu_image_ort.py` runs an ONNX-exported SD1.5 through ONNX Runtime's
201
+ **DmlExecutionProvider** on the integrated GPU (local `_sd15_combined`, 256x256, DDIM):
202
+
203
+ | path | 8 steps | per step |
204
+ |---|---|---|
205
+ | CPU diffusers (Torch 2.13, 6 threads) | 25.29 s | 3.16 s |
206
+ | **iGPU ONNX Runtime DirectML** | **13.05 s** | **1.63 s** |
207
+
208
+ ≈**1.9x** faster at 256x256; 512x512 currently fails inside the UNet on the DML EP (activation
209
+ memory / operator coverage) while the same ONNX runs on `CPUExecutionProvider`. Export once with
210
+ `optimum-cli export onnx --task stable-diffusion --opset 17`. A diffusers + `torch_directml`
211
+ variant (`igpu_image_infer.py`) is kept as a comparison; it needs legacy attention and still hits
212
+ DML `chunk`/reshape issues. Long-term the OpenCL backend gets an **fp16 device-resident conv engine**: measured probes
213
+ show per-op OpenCL conv is 2-3x slower than CPU oneDNN and fp32 resident only reaches parity,
214
+ but fp16 resident im2col+Hgemm is **1.3-2x faster than the CPU conv** at SD1.5 shapes
215
+ (538 vs 342 GFLOPS; 1280-channel 16x16 layer 17.3 vs 34.8 ms). The 256-resolution demo
216
+ therefore uses DirectML, while the OpenCL engine is the planned production backend.
217
+
218
+ The fp16 device-resident building blocks are in `gt1030/opencl_fp16_ops.py` +
219
+ `gt1030/kernels_attention.cl` (im2col + CLBlast Hgemm conv, GroupNorm+SiLU, LayerNorm,
220
+ GEGLU, multi-head attention, nearest-2x upsample, add). Clean measurements on the
221
+ Vega 6 (no CPU contention), parity 0.02-1.1% relL2:
222
+ ResNetBlock2D 0.61-0.77x, VAE up-block 0.40x, FFN 0.16-0.23x, attention 0.42-0.99x,
223
+ full Transformer block 0.51-0.71x of CPU time -- i.e. 1.3-6x faster for most blocks.
224
+ The **VAE decoder is now end-to-end on the iGPU** (`gt1030/opencl_vae.py`): fp16 parity
225
+ 2.7-3.7% vs CPU, 1.81 s vs 2.07 s at 256 (0.87x), 7.93 s vs 8.34 s at 512 (0.95x, warm).
226
+ The remaining step is assembling the full 256px UNet graph (fp16 weights ~1.7 GB,
227
+ under the 5,090 MB single-allocation limit; 512px needs activation tiling).
228
+
187
229
  ## 3-bit frozen base + LoRA (`bitsandbytes.q3bit`, experimental)
188
230
 
189
231
  A frozen weight can be stored as blockwise **3-bit** codes (≈11% of fp32 memory) and
@@ -226,6 +268,7 @@ R5-4500U (6C/6T, AVX2, 15.4 GB) / i5-10400 (6C/12T, AVX2, 12 GB), no NVIDIA GPU.
226
268
  | Real model (Qwen3.5-0.8B LoRA, seq 64) | 4 threads beat 6 by **7–10%** — measure with `tune_threads()` |
227
269
  | Fused lm_head 3-bit GEMM (research) | 14–22% faster than fp32 matmul, 89% less weight memory |
228
270
  | 3-bit frozen base + LoRA (Qwen3-1.7B, CPU) | 80 steps: eval loss 4.34 → **4.03**; RSS **2.9 GB vs 8.7 GB** fp32; step 7.7 s vs 5.5 s (native Windows) |
271
+ | Full-parameter fine-tune (Qwen3.5-0.8B, 752 M, fp32 + AdamW8bit, CPU) | **10.4 GB peak RSS**, 8.8 s/step; LoRA on the same model: 5.9 GB, 2.65 s/step |
229
272
  | AdamW4bit step (1M params) | 28.6 ms vs 3.3 ms for the AVX2 8-bit kernel; state 1.031 vs 2.062 B/param |
230
273
 
231
274
  Not every workload is faster than fp32: 4-bit/8-bit kernels trade compute for memory traffic,
@@ -244,8 +287,8 @@ the full trade-off tables, including negative results.
244
287
  - **manylinux_2_17 wheel**: the AVX-512 library inside it is single-threaded (the parallel
245
288
  build is only for the native `manylinux_2_39` wheel). On AVX-512 machines set
246
289
  `BNB_CPU_DISABLE_AVX512=1` to run the AVX2 OpenMP path.
247
- - **aarch64**: there is no PyPI wheel yet; build from the sdist (`build_linux.sh`) or use the
248
- Termux instructions in `docs_cpu/TERMUX_GUIDE.md`.
290
+ - **aarch64**: there is no PyPI wheel yet; build from the sdist (`build_linux.sh`) or follow the
291
+ Termux guide [`TERMUX_GUIDE.md`](https://github.com/Ganyvnanoaa09-0906/bitsandbytes-CPU/blob/HEAD/bitsandbytes/TERMUX_GUIDE.md).
249
292
  - **macOS**: no wheel is published and the CPU backend is not tested there; the sdist build may
250
293
  work with clang + OpenMP, but treat it as unsupported.
251
294
  - 3-bit (and 2-bit dequant-only) is **experimental**: the API and kernels are covered
@@ -264,6 +307,23 @@ bash build_linux.sh --selftest # also build + run the torch-free C se
264
307
  build_manual\build_manual.bat amd # run inside "x64 Native Tools Command Prompt"
265
308
  ```
266
309
 
310
+ ### Termux (Android / ARM64)
311
+
312
+ On-device builds work; there is no aarch64 PyPI wheel, so build from the sdist:
313
+
314
+ ```bash
315
+ pkg update && pkg install -y clang libomp make python
316
+ bash build_termux.sh # -> bitsandbytes/libbitsandbytes_cpu.so
317
+ bash build_termux.sh --selftest # also build + run the torch-free C self-test
318
+ ```
319
+
320
+ Three Termux-specific traps the script already handles: the compiler is `clang++` (there is no
321
+ `g++`), OpenMP needs `pkg install libomp`, and `-march=native` is unsafe under Android's clang.
322
+ The script is also what the automated flow uses to build on a phone **driven from a PC over ADB**
323
+ — see [`TERMUX_GUIDE.md`](https://github.com/Ganyvnanoaa09-0906/bitsandbytes-CPU/blob/HEAD/bitsandbytes/TERMUX_GUIDE.md)
324
+ for the phone-side preparation, the ADB/wireless-debugging path, the ARM64 kernel notes and the
325
+ verification matrix.
326
+
267
327
  Two special wheel builds used by this project are documented in
268
328
  [`docs_cpu/WHEEL_PLATFORM_NOTES.md`](https://github.com/Ganyvnanoaa09-0906/bitsandbytes-CPU/blob/HEAD/bitsandbytes/docs_cpu/WHEEL_PLATFORM_NOTES.md):
269
329
  the zig manylinux_2_17 portable build and the devtoolset-8 manylinux_2_17 **OpenMP** build
@@ -274,12 +334,12 @@ the zig manylinux_2_17 portable build and the devtoolset-8 manylinux_2_17 **Open
274
334
  ```bash
275
335
  bash build_linux.sh --selftest # C kernels: 9 checks (quantize/gemm_8bit/gemv_4bit/
276
336
  # AdamW8bit/vector-vs-scalar/2-3bit dequant/fused GEMM)
277
- python test_efst_regressions.py # EFST: nested blocks, 3-D tensor experts, routing usage
278
- python test_q3bit_kernels.py # 2/3-bit research prototype (old DLL-based tests)
279
- python test_q3bit_api.py # public q3bit API: pack/dequant/gemm, QuantLinear3bit, fallback
280
- python test_adamw4bit.py # restored 4-bit optimizer: state bytes, native vs torch, fp32 tracking
281
- python test_concurrency_plan.py # planner / memory estimate / tune_threads / launcher
282
- python test_disk_balancer_safety.py # data-safety rules for the disk balancer
337
+ python scripts/test/test_efst_regressions.py # EFST: nested blocks, 3-D tensor experts, routing usage
338
+ python scripts/test/test_q3bit_kernels.py # 2/3-bit research prototype (old DLL-based tests)
339
+ python scripts/test/test_q3bit_api.py # public q3bit API: pack/dequant/gemm, QuantLinear3bit, fallback
340
+ python scripts/test/test_adamw4bit.py # restored 4-bit optimizer: state bytes, native vs torch, fp32 tracking
341
+ python scripts/test/test_concurrency_plan.py # planner / memory estimate / tune_threads / launcher
342
+ python scripts/test/test_disk_balancer_safety.py # data-safety rules for the disk balancer
283
343
  ```
284
344
 
285
345
  ## Documentation
@@ -306,6 +366,8 @@ The long-form docs live in the repository (`bitsandbytes/docs_cpu/`, Chinese and
306
366
  | `0.50.2.dev7` | documentation release: README rewritten for the CPU fork (install/wheels/platform matrix, feature table, measured performance, limitations, tests, full dev1→dev7 changelog, EFST/disk-balancer/toolbox examples) |
307
367
  | `0.50.2.dev8` | **restored `AdamW4bit`** (the kernel was dropped when dev4 vendored upstream; now scalar C kernel + torch fallback, 1.031 B/param); **new `bitsandbytes.q3bit` API** — `quantize_3bit`/`dequantize_3bit`/`gemm_3bit`/`gemm_3bit_transposed`, `QuantLinear3bit` (frozen 3-bit base + LoRA), `convert_linears_3bit`, native kernels on Linux with identical pure-torch fallback; CLI reference now includes 4bitopt/q3bit and all 43 examples execute in CI-style verification |
308
368
  | `0.50.2.dev9` | **Windows native kernels**: CI now compiles the MSVC DLL from the current csrc and the export list includes the 4-bit optimizer and 2/3-bit symbols (they were silently missing before); **q3bit real-model validation** on Qwen3-1.7B (frozen 3-bit base + LoRA, eval 4.34→4.03 over 80 steps, RSS −67%, native step 7.7 s vs fp32 5.5 s); fixed the causal-LM **double-shift** label bug in the iGPU training kit (Windows package v0.1.2) |
369
+ | `0.50.2.dev10` | **NTFS recovery toolkit** (`sector_carve`, `mft_recover`, CLI + GUI, selftests) and the **video E2E pipeline** (generation/training, LCM and motion diagnostics); AVX2 4-bit GEMV fixes (wrong FP4 codebook, lazy-init race, non-divisible blocksize); gt1030 probe scripts and docs |
370
+ | `0.50.2.dev11` | **NT-store fast paths**: 4-bit dequant writes bypass the cache above 16 MiB (1.7–2.0× at 32–64 MB) and RMSNorm chooses NT by total footprint (x+res+out vs per-CCX L3); CPUID L3-sharing mask fix and the NT threshold wired to one macro; **byte-reproducible sdist** with an identity gate that audits the uploaded bytes (`--sdist`, 2-DLL policy, 20 adversarial cases) plus a docs gate requiring all six ZH/EN docs; repo regrouping into `scripts/*`; UTF-8 BOM strip; 5 reproducible CPU audio benches; 6 video perf fixes and 3 AR fixes |
309
371
 
310
372
  ## License and attribution
311
373
 
@@ -1,84 +1,84 @@
1
- # Copyright (c) Facebook, Inc. and its affiliates.
2
- #
3
- # This source code is licensed under the MIT license found in the
4
- # LICENSE file in the root directory of this source tree.
5
-
6
-
7
- import importlib
8
- import sys
9
-
10
- import torch
11
-
12
- from . import _ops, utils
13
- from .autograd._functions import (
14
- MatmulLtState,
15
- matmul,
16
- matmul_4bit,
17
- )
18
- from .backends.cpu import ops as cpu_ops
19
- from .backends.default import ops as default_ops
20
- from .nn import modules
21
- from .optim import adam
22
-
23
- # This is a signal for integrations with transformers/diffusers.
24
- # Eventually we may remove this but it is currently required for compatibility.
25
- features = {"multi_backend"}
26
- supported_torch_devices = {
27
- "cpu",
28
- "cuda", # NVIDIA/AMD GPU
29
- "xpu", # Intel GPU
30
- "hpu", # Intel Gaudi
31
- "npu", # Ascend NPU
32
- "mps", # Apple Silicon
33
- }
34
-
35
- if torch.cuda.is_available():
36
- from .backends.cuda import ops as cuda_ops
37
-
38
- if hasattr(torch, "xpu") and torch.xpu.is_available():
39
- from .backends.xpu import ops as xpu_ops
40
-
41
- if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
42
- from .backends.mps import ops as mps_ops
43
-
44
- if importlib.util.find_spec("habana_frameworks") and importlib.util.find_spec("habana_frameworks.torch"):
45
- # In case not automatically imported
46
- import habana_frameworks.torch
47
-
48
- if hasattr(torch, "hpu") and torch.hpu.is_available():
49
- from .backends.hpu import ops as hpu_ops
50
-
51
-
52
- def _import_backends():
53
- """
54
- Discover and autoload all available backends installed as separate packages.
55
- Packages with an entrypoint for "bitsandbytes.backends" will be loaded.
56
- Inspired by PyTorch implementation: https://pytorch.org/tutorials/prototype/python_extension_autoload.html
57
- """
58
- from importlib.metadata import entry_points
59
-
60
- extensions = entry_points(group="bitsandbytes.backends")
61
-
62
- for ext in extensions:
63
- try:
64
- entry = ext.load()
65
- entry()
66
- except Exception as e:
67
- raise RuntimeError(f"bitsandbytes: failed to load backend {ext.name}: {e}") from e
68
-
69
-
70
- _import_backends()
71
-
72
- __pdoc__ = {
73
- "libbitsandbytes": False,
74
- "optim.optimizer.Optimizer8bit": False,
75
- "optim.optimizer.MockArgs": False,
76
- }
77
-
78
- __version__ = "0.50.2.dev9"
79
-
80
- # 挂载 gdn_cpu(GDN 融合内核 / patch_transformers 入口),让 bnb.gdn_cpu 可直接访问
81
- from . import gdn_cpu # noqa: E402,F401
82
-
83
- # 挂载 q3bit(实验性 2/3-bit 量化 / 冻结 3-bit base + LoRA),让 bnb.q3bit 可直接访问
84
- from . import q3bit # noqa: E402,F401
1
+ # Copyright (c) Facebook, Inc. and its affiliates.
2
+ #
3
+ # This source code is licensed under the MIT license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+
7
+ import importlib
8
+ import sys
9
+
10
+ import torch
11
+
12
+ from . import _ops, utils
13
+ from .autograd._functions import (
14
+ MatmulLtState,
15
+ matmul,
16
+ matmul_4bit,
17
+ )
18
+ from .backends.cpu import ops as cpu_ops
19
+ from .backends.default import ops as default_ops
20
+ from .nn import modules
21
+ from .optim import adam
22
+
23
+ # This is a signal for integrations with transformers/diffusers.
24
+ # Eventually we may remove this but it is currently required for compatibility.
25
+ features = {"multi_backend"}
26
+ supported_torch_devices = {
27
+ "cpu",
28
+ "cuda", # NVIDIA/AMD GPU
29
+ "xpu", # Intel GPU
30
+ "hpu", # Intel Gaudi
31
+ "npu", # Ascend NPU
32
+ "mps", # Apple Silicon
33
+ }
34
+
35
+ if torch.cuda.is_available():
36
+ from .backends.cuda import ops as cuda_ops
37
+
38
+ if hasattr(torch, "xpu") and torch.xpu.is_available():
39
+ from .backends.xpu import ops as xpu_ops
40
+
41
+ if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
42
+ from .backends.mps import ops as mps_ops
43
+
44
+ if importlib.util.find_spec("habana_frameworks") and importlib.util.find_spec("habana_frameworks.torch"):
45
+ # In case not automatically imported
46
+ import habana_frameworks.torch
47
+
48
+ if hasattr(torch, "hpu") and torch.hpu.is_available():
49
+ from .backends.hpu import ops as hpu_ops
50
+
51
+
52
+ def _import_backends():
53
+ """
54
+ Discover and autoload all available backends installed as separate packages.
55
+ Packages with an entrypoint for "bitsandbytes.backends" will be loaded.
56
+ Inspired by PyTorch implementation: https://pytorch.org/tutorials/prototype/python_extension_autoload.html
57
+ """
58
+ from importlib.metadata import entry_points
59
+
60
+ extensions = entry_points(group="bitsandbytes.backends")
61
+
62
+ for ext in extensions:
63
+ try:
64
+ entry = ext.load()
65
+ entry()
66
+ except Exception as e:
67
+ raise RuntimeError(f"bitsandbytes: failed to load backend {ext.name}: {e}") from e
68
+
69
+
70
+ _import_backends()
71
+
72
+ __pdoc__ = {
73
+ "libbitsandbytes": False,
74
+ "optim.optimizer.Optimizer8bit": False,
75
+ "optim.optimizer.MockArgs": False,
76
+ }
77
+
78
+ __version__ = "0.50.2.dev11"
79
+
80
+ # 挂载 gdn_cpu(GDN 融合内核 / patch_transformers 入口),让 bnb.gdn_cpu 可直接访问
81
+ from . import gdn_cpu # noqa: E402,F401
82
+
83
+ # 挂载 q3bit(实验性 2/3-bit 量化 / 冻结 3-bit base + LoRA),让 bnb.q3bit 可直接访问
84
+ from . import q3bit # noqa: E402,F401
@@ -1,4 +1,4 @@
1
- if __name__ == "__main__":
2
- from bitsandbytes.diagnostics.main import main
3
-
4
- main()
1
+ if __name__ == "__main__":
2
+ from bitsandbytes.diagnostics.main import main
3
+
4
+ main()