bitsandbytes-cpu-fork 0.50.2.dev8__tar.gz → 0.50.2.dev9__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 (156) hide show
  1. {bitsandbytes_cpu_fork-0.50.2.dev8/bitsandbytes_cpu_fork.egg-info → bitsandbytes_cpu_fork-0.50.2.dev9}/PKG-INFO +9 -6
  2. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/README.md +8 -5
  3. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/__init__.py +1 -1
  4. bitsandbytes_cpu_fork-0.50.2.dev9/bitsandbytes/libbitsandbytes_cpu.dll +0 -0
  5. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes_cpu_cli.py +5 -5
  6. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9/bitsandbytes_cpu_fork.egg-info}/PKG-INFO +9 -6
  7. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes_cpu_fork.egg-info/SOURCES.txt +1 -0
  8. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/build_manual/export.def +44 -39
  9. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/docs_cpu/API_REFERENCE.md +4 -4
  10. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/docs_cpu/API_REFERENCE_EN.md +4 -4
  11. bitsandbytes_cpu_fork-0.50.2.dev9/tools/patch_dev9.py +75 -0
  12. bitsandbytes_cpu_fork-0.50.2.dev8/bitsandbytes/libbitsandbytes_cpu.dll +0 -0
  13. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/LICENSE +0 -0
  14. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/MANIFEST.in +0 -0
  15. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/__main__.py +0 -0
  16. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/_cpu_caps.py +0 -0
  17. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/_ops.py +0 -0
  18. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/autograd/__init__.py +0 -0
  19. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/autograd/_functions.py +0 -0
  20. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/__init__.py +0 -0
  21. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/cpu/__init__.py +0 -0
  22. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/cpu/ops.py +0 -0
  23. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/cuda/__init__.py +0 -0
  24. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/cuda/ops.py +0 -0
  25. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/default/__init__.py +0 -0
  26. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/default/ops.py +0 -0
  27. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/hpu/__init__.py +0 -0
  28. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/hpu/ops.py +0 -0
  29. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/mps/__init__.py +0 -0
  30. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/mps/ops.py +0 -0
  31. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/triton/__init__.py +0 -0
  32. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/triton/kernels_4bit.py +0 -0
  33. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/triton/kernels_8bit_quant.py +0 -0
  34. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/triton/kernels_optim.py +0 -0
  35. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/triton/ops.py +0 -0
  36. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/utils.py +0 -0
  37. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/xpu/__init__.py +0 -0
  38. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/backends/xpu/ops.py +0 -0
  39. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/cextension.py +0 -0
  40. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/cli.py +0 -0
  41. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/consts.py +0 -0
  42. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/cuda_specs.py +0 -0
  43. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/detect_cpu.py +0 -0
  44. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/diagnostics/__init__.py +0 -0
  45. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/diagnostics/cuda.py +0 -0
  46. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/diagnostics/main.py +0 -0
  47. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/diagnostics/utils.py +0 -0
  48. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/disk_balancer.py +0 -0
  49. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/efst.py +0 -0
  50. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/functional.py +0 -0
  51. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/gdn_cpu.py +0 -0
  52. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/nn/__init__.py +0 -0
  53. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/nn/modules.py +0 -0
  54. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/nn/parametrize.py +0 -0
  55. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/optim/__init__.py +0 -0
  56. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/optim/adagrad.py +0 -0
  57. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/optim/adam.py +0 -0
  58. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/optim/adamw.py +0 -0
  59. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/optim/adamw4bit.py +0 -0
  60. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/optim/ademamix.py +0 -0
  61. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/optim/lamb.py +0 -0
  62. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/optim/lars.py +0 -0
  63. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/optim/lion.py +0 -0
  64. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/optim/optimizer.py +0 -0
  65. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/optim/rmsprop.py +0 -0
  66. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/optim/sgd.py +0 -0
  67. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/py.typed +0 -0
  68. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/q3bit.py +0 -0
  69. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/torch_cpu_kit.py +0 -0
  70. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/utils.py +0 -0
  71. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes/vcomp140.dll +0 -0
  72. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes_cpu_fork.egg-info/dependency_links.txt +0 -0
  73. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes_cpu_fork.egg-info/entry_points.txt +0 -0
  74. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes_cpu_fork.egg-info/requires.txt +0 -0
  75. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/bitsandbytes_cpu_fork.egg-info/top_level.txt +0 -0
  76. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/build_linux.sh +0 -0
  77. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/build_manual/build_manual.bat +0 -0
  78. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/build_manual/export_avx512.def +0 -0
  79. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/build_manual/export_generic.def +0 -0
  80. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/build_manual/selftest_win.bat +0 -0
  81. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/build_termux.sh +0 -0
  82. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/avx2_gemv_4bit.h +0 -0
  83. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/common.cuh +0 -0
  84. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/common.h +0 -0
  85. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/compat.cuh +0 -0
  86. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/compat_device.cuh +0 -0
  87. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/cpu_cache.h +0 -0
  88. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/cpu_features.h +0 -0
  89. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/cpu_gdn.cpp +0 -0
  90. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/cpu_ops.cpp +0 -0
  91. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/cpu_ops.h +0 -0
  92. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/gemm_4bit.cu +0 -0
  93. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/gemm_4bit_common.cuh +0 -0
  94. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/gemm_4bit_mma.cuh +0 -0
  95. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/gemm_4bit_simt.cu +0 -0
  96. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/gemm_4bit_simt.cuh +0 -0
  97. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/gemm_4bit_sm75.cu +0 -0
  98. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/gemm_4bit_sm75.cuh +0 -0
  99. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/gemm_4bit_sm80.cu +0 -0
  100. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/gemm_4bit_sm80.cuh +0 -0
  101. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/kernels.cu +0 -0
  102. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/kernels.cuh +0 -0
  103. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/ops.cu +0 -0
  104. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/ops.cuh +0 -0
  105. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/pythonInterface.cpp +0 -0
  106. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/test_avx2_gemv_4bit.cpp +0 -0
  107. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/test_cpu_cache.cpp +0 -0
  108. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/xpu_kernels.cpp +0 -0
  109. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/xpu_kernels.h +0 -0
  110. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/xpu_ops.cpp +0 -0
  111. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/csrc/xpu_ops.h +0 -0
  112. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/docs_cpu/EPYC_VALIDATION.md +0 -0
  113. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/docs_cpu/ISA_MATRIX.md +0 -0
  114. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/docs_cpu/MOTION_BOUNDARY.md +0 -0
  115. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/docs_cpu/MOTION_BOUNDARY_EN.md +0 -0
  116. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/docs_cpu/QUICKSTART.md +0 -0
  117. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/docs_cpu/QUICKSTART_EN.md +0 -0
  118. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/docs_cpu/TECHNICAL_GUIDE.md +0 -0
  119. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/docs_cpu/TECHNICAL_GUIDE_EN.md +0 -0
  120. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/docs_cpu/TECH_REPORT.md +0 -0
  121. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/docs_cpu/TECH_REPORT_EN.md +0 -0
  122. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/docs_cpu/WHEEL_PLATFORM_NOTES.md +0 -0
  123. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/docs_cpu/check_docs.py +0 -0
  124. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/examples/compile_inference.py +0 -0
  125. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/examples/cpu/cpu_training.py +0 -0
  126. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/examples/int8_inference_huggingface.py +0 -0
  127. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/examples/xpu/benchmark_paged_memory.py +0 -0
  128. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/examples/xpu/paged_xpu_training.py +0 -0
  129. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/pyproject.toml +0 -0
  130. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/selftest_cpu.c +0 -0
  131. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/setup.cfg +0 -0
  132. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/setup.py +0 -0
  133. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tests/test_autograd.py +0 -0
  134. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tests/test_cpu_e2e.py +0 -0
  135. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tests/test_cuda_setup_evaluator.py +0 -0
  136. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tests/test_functional.py +0 -0
  137. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tests/test_generation.py +0 -0
  138. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tests/test_linear4bit.py +0 -0
  139. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tests/test_linear8bitlt.py +0 -0
  140. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tests/test_modules.py +0 -0
  141. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tests/test_ops.py +0 -0
  142. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tests/test_optim.py +0 -0
  143. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tests/test_parametrize.py +0 -0
  144. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tests/test_rmsnorm_fwd_cpu.py +0 -0
  145. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tools/create_github_release.ps1 +0 -0
  146. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tools/dump_signatures.py +0 -0
  147. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tools/gen_api_reference.py +0 -0
  148. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tools/publish.ps1 +0 -0
  149. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tools/sector_carve.c +0 -0
  150. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tools/sector_carve_gui.c +0 -0
  151. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tools/sector_mirror.c +0 -0
  152. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tools/sector_mirror_gui.c +0 -0
  153. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tools/train_llm_sft.py +0 -0
  154. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tools/verify_help_examples.py +0 -0
  155. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tools/verify_linux_cli.sh +0 -0
  156. {bitsandbytes_cpu_fork-0.50.2.dev8 → bitsandbytes_cpu_fork-0.50.2.dev9}/tools/verify_train_release.py +0 -0
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: bitsandbytes-cpu-fork
3
- Version: 0.50.2.dev8
3
+ Version: 0.50.2.dev9
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
@@ -36,7 +36,7 @@ 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.dev8`** · import name is unchanged (`import bitsandbytes as bnb`) · MIT.
39
+ **Version: `0.50.2.dev9`** · 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
@@ -202,7 +202,7 @@ Also: `estimate_training_memory_gb()`, `spawn_training_jobs(cmds, plan)` (logs p
202
202
  | Area | What it does |
203
203
  |---|---|
204
204
  | 8-bit optimizers | AdamW/Adam/AdEMAMix/LAMB/LARS/Lion + paged variants, blockwise 8-bit state with single-pass dequant→update→requant kernels (AVX2/NEON) |
205
- | 4-bit optimizer | `AdamW4bit`: ≈1.03 bytes of state per parameter (restored in dev8; scalar kernel + torch fallback) |
205
+ | 4-bit optimizer | `AdamW4bit`: ≈1.03 bytes of state per parameter (scalar kernel; native on Linux and, since dev9, Windows) |
206
206
  | Quantisation | blockwise 8-bit, NF4/FP4 4-bit quantize/dequantize; 4-bit GEMV inference kernel |
207
207
  | `gemm_8bit` | `out = A @ dequant8(B)` with weights kept uint8 (¼ DRAM traffic), no fp32 temporary (`bnb.functional.fused_dequant_linear_8bit`) |
208
208
  | GDN | fused Gated DeltaNet forward + chunk-checkpointed backward (`gdn_fwd_cpu` / `gdn_bwd_cpu`), plus `patch_transformers()` / `patch_fla()` |
@@ -256,6 +256,7 @@ R5-4500U (6C/6T, AVX2, 15.4 GB) / i5-10400 (6C/12T, AVX2, 12 GB), no NVIDIA GPU.
256
256
  | Concurrency packing (synthetic jobs) | 6 jobs × 1 thread **5619 tok/s** vs 1 job × 6 threads **3843 tok/s** (+46%) |
257
257
  | Real model (Qwen3.5-0.8B LoRA, seq 64) | 4 threads beat 6 by **7–10%** — measure with `tune_threads()` |
258
258
  | Fused lm_head 3-bit GEMM (research) | 14–22% faster than fp32 matmul, 89% less weight memory |
259
+ | 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) |
259
260
  | 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 |
260
261
 
261
262
  Not every workload is faster than fp32: 4-bit/8-bit kernels trade compute for memory traffic,
@@ -267,9 +268,10 @@ the full trade-off tables, including negative results.
267
268
  ## Platform limitations
268
269
 
269
270
  - **CPU only.** CUDA/ROCm/XPU kernels are not built; a GPU is never required or used.
270
- - **Windows wheel** ships a prebuilt DLL. That DLL predates the dev5/dev6 2/3-bit
271
- kernels and the restored 4-bit optimizer kernel, so those run the pure-torch
272
- fallback there (same numbers, lower speed).
271
+ - **Windows wheel**: from dev9 on it ships the MSVC DLL built by CI from the current
272
+ `csrc`, which exports the 2/3-bit dequant/GEMM and 4-bit optimizer kernels; they run
273
+ natively on Windows (verified: dequant maxdiff 1.2e-6, 4-bit native vs torch 1.2e-7).
274
+ Wheels older than dev9 fall back to pure torch for those symbols.
273
275
  - **manylinux_2_17 wheel**: the AVX-512 library inside it is single-threaded (the parallel
274
276
  build is only for the native `manylinux_2_39` wheel). On AVX-512 machines set
275
277
  `BNB_CPU_DISABLE_AVX512=1` to run the AVX2 OpenMP path.
@@ -334,6 +336,7 @@ The long-form docs live in the repository (`bitsandbytes/docs_cpu/`, Chinese and
334
336
  | `0.50.2.dev6` | manylinux_2_17 wheel is now **OpenMP** (devtoolset-8, bundled libgomp → 3–4× optimizer-kernel speedup on old glibc); csrc 2/3-bit dequant + fused GEMM kernels with 9-item C self-test; EFST fused tensor-expert routing fix; GCC 8 `<cpuid.h>` portability fix |
335
337
  | `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) |
336
338
  | `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
+ | `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) |
337
340
 
338
341
  ## License and attribution
339
342
 
@@ -5,7 +5,7 @@
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.dev8`** · import name is unchanged (`import bitsandbytes as bnb`) · MIT.
8
+ **Version: `0.50.2.dev9`** · 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
@@ -171,7 +171,7 @@ Also: `estimate_training_memory_gb()`, `spawn_training_jobs(cmds, plan)` (logs p
171
171
  | Area | What it does |
172
172
  |---|---|
173
173
  | 8-bit optimizers | AdamW/Adam/AdEMAMix/LAMB/LARS/Lion + paged variants, blockwise 8-bit state with single-pass dequant→update→requant kernels (AVX2/NEON) |
174
- | 4-bit optimizer | `AdamW4bit`: ≈1.03 bytes of state per parameter (restored in dev8; scalar kernel + torch fallback) |
174
+ | 4-bit optimizer | `AdamW4bit`: ≈1.03 bytes of state per parameter (scalar kernel; native on Linux and, since dev9, Windows) |
175
175
  | Quantisation | blockwise 8-bit, NF4/FP4 4-bit quantize/dequantize; 4-bit GEMV inference kernel |
176
176
  | `gemm_8bit` | `out = A @ dequant8(B)` with weights kept uint8 (¼ DRAM traffic), no fp32 temporary (`bnb.functional.fused_dequant_linear_8bit`) |
177
177
  | GDN | fused Gated DeltaNet forward + chunk-checkpointed backward (`gdn_fwd_cpu` / `gdn_bwd_cpu`), plus `patch_transformers()` / `patch_fla()` |
@@ -225,6 +225,7 @@ R5-4500U (6C/6T, AVX2, 15.4 GB) / i5-10400 (6C/12T, AVX2, 12 GB), no NVIDIA GPU.
225
225
  | Concurrency packing (synthetic jobs) | 6 jobs × 1 thread **5619 tok/s** vs 1 job × 6 threads **3843 tok/s** (+46%) |
226
226
  | Real model (Qwen3.5-0.8B LoRA, seq 64) | 4 threads beat 6 by **7–10%** — measure with `tune_threads()` |
227
227
  | Fused lm_head 3-bit GEMM (research) | 14–22% faster than fp32 matmul, 89% less weight memory |
228
+ | 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) |
228
229
  | 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 |
229
230
 
230
231
  Not every workload is faster than fp32: 4-bit/8-bit kernels trade compute for memory traffic,
@@ -236,9 +237,10 @@ the full trade-off tables, including negative results.
236
237
  ## Platform limitations
237
238
 
238
239
  - **CPU only.** CUDA/ROCm/XPU kernels are not built; a GPU is never required or used.
239
- - **Windows wheel** ships a prebuilt DLL. That DLL predates the dev5/dev6 2/3-bit
240
- kernels and the restored 4-bit optimizer kernel, so those run the pure-torch
241
- fallback there (same numbers, lower speed).
240
+ - **Windows wheel**: from dev9 on it ships the MSVC DLL built by CI from the current
241
+ `csrc`, which exports the 2/3-bit dequant/GEMM and 4-bit optimizer kernels; they run
242
+ natively on Windows (verified: dequant maxdiff 1.2e-6, 4-bit native vs torch 1.2e-7).
243
+ Wheels older than dev9 fall back to pure torch for those symbols.
242
244
  - **manylinux_2_17 wheel**: the AVX-512 library inside it is single-threaded (the parallel
243
245
  build is only for the native `manylinux_2_39` wheel). On AVX-512 machines set
244
246
  `BNB_CPU_DISABLE_AVX512=1` to run the AVX2 OpenMP path.
@@ -303,6 +305,7 @@ The long-form docs live in the repository (`bitsandbytes/docs_cpu/`, Chinese and
303
305
  | `0.50.2.dev6` | manylinux_2_17 wheel is now **OpenMP** (devtoolset-8, bundled libgomp → 3–4× optimizer-kernel speedup on old glibc); csrc 2/3-bit dequant + fused GEMM kernels with 9-item C self-test; EFST fused tensor-expert routing fix; GCC 8 `<cpuid.h>` portability fix |
304
306
  | `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) |
305
307
  | `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
+ | `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) |
306
309
 
307
310
  ## License and attribution
308
311
 
@@ -75,7 +75,7 @@ __pdoc__ = {
75
75
  "optim.optimizer.MockArgs": False,
76
76
  }
77
77
 
78
- __version__ = "0.50.2.dev8"
78
+ __version__ = "0.50.2.dev9"
79
79
 
80
80
  # 挂载 gdn_cpu(GDN 融合内核 / patch_transformers 入口),让 bnb.gdn_cpu 可直接访问
81
81
  from . import gdn_cpu # noqa: E402,F401
@@ -312,15 +312,15 @@ SECTIONS = [
312
312
  "# what runs today is the verified scalar path",
313
313
  "# the optimizer is a few percent of a real training step at 30M params,",
314
314
  "# so the wall-clock cost is small while the state saving is 7.8x",
315
- "# the pure-torch fallback (Windows wheel) is slower still",
315
+ "# older Windows wheels run the pure-torch fallback (dev9+ is native)",
316
316
  ], PY),
317
317
  ("traps",
318
318
  "the failure modes seen while building this, each one silent at first", [
319
319
  "# 1. torch.optim.AdamW is NOT replaced: this is a separate class, and",
320
320
  "# AdamW4bit requires fp32 params (the kernel's dtype 0).",
321
321
  "# 2. blocksize is compiled in at 256; passing another value raises.",
322
- "# 3. if the shared library does not export",
323
- "# coptimizer_update_4bit_blockwise_cpu (the Windows wheel does not),",
322
+ "# 3. older Windows wheels (< dev9) do not export",
323
+ "# coptimizer_update_4bit_blockwise_cpu (pre-dev9 Windows wheels),",
324
324
  "# this class transparently uses a pure-torch implementation with the",
325
325
  "# same numbers and the same 1.031 B/param state, just slower.",
326
326
  "# 4. quantised state is not a drop-in for exact reproducibility runs:",
@@ -565,8 +565,8 @@ SECTIONS = [
565
565
  ("convert_linears_3bit(model, lora_r=8)",
566
566
  "replaces every nn.Linear in place and returns the count; lm_head stays fp32", []),
567
567
  ("experimental / portability",
568
- "the Windows wheel predates the 2/3-bit C kernels, so it runs the pure-torch "
569
- "fallback; the Linux wheels use the native kernels", []),
568
+ "since dev9 the Windows wheel ships a CI-built MSVC DLL; older wheels use the pure-torch "
569
+ "fallback, while the dev9 Windows and Linux wheels use the native kernels", []),
570
570
  ]),
571
571
  ("layers", "Reference: bitsandbytes.nn", [
572
572
  ("Linear4bit(input_features, output_features, bias=True, compute_dtype=None,\n"
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: bitsandbytes-cpu-fork
3
- Version: 0.50.2.dev8
3
+ Version: 0.50.2.dev9
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
@@ -36,7 +36,7 @@ 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.dev8`** · import name is unchanged (`import bitsandbytes as bnb`) · MIT.
39
+ **Version: `0.50.2.dev9`** · 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
@@ -202,7 +202,7 @@ Also: `estimate_training_memory_gb()`, `spawn_training_jobs(cmds, plan)` (logs p
202
202
  | Area | What it does |
203
203
  |---|---|
204
204
  | 8-bit optimizers | AdamW/Adam/AdEMAMix/LAMB/LARS/Lion + paged variants, blockwise 8-bit state with single-pass dequant→update→requant kernels (AVX2/NEON) |
205
- | 4-bit optimizer | `AdamW4bit`: ≈1.03 bytes of state per parameter (restored in dev8; scalar kernel + torch fallback) |
205
+ | 4-bit optimizer | `AdamW4bit`: ≈1.03 bytes of state per parameter (scalar kernel; native on Linux and, since dev9, Windows) |
206
206
  | Quantisation | blockwise 8-bit, NF4/FP4 4-bit quantize/dequantize; 4-bit GEMV inference kernel |
207
207
  | `gemm_8bit` | `out = A @ dequant8(B)` with weights kept uint8 (¼ DRAM traffic), no fp32 temporary (`bnb.functional.fused_dequant_linear_8bit`) |
208
208
  | GDN | fused Gated DeltaNet forward + chunk-checkpointed backward (`gdn_fwd_cpu` / `gdn_bwd_cpu`), plus `patch_transformers()` / `patch_fla()` |
@@ -256,6 +256,7 @@ R5-4500U (6C/6T, AVX2, 15.4 GB) / i5-10400 (6C/12T, AVX2, 12 GB), no NVIDIA GPU.
256
256
  | Concurrency packing (synthetic jobs) | 6 jobs × 1 thread **5619 tok/s** vs 1 job × 6 threads **3843 tok/s** (+46%) |
257
257
  | Real model (Qwen3.5-0.8B LoRA, seq 64) | 4 threads beat 6 by **7–10%** — measure with `tune_threads()` |
258
258
  | Fused lm_head 3-bit GEMM (research) | 14–22% faster than fp32 matmul, 89% less weight memory |
259
+ | 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) |
259
260
  | 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 |
260
261
 
261
262
  Not every workload is faster than fp32: 4-bit/8-bit kernels trade compute for memory traffic,
@@ -267,9 +268,10 @@ the full trade-off tables, including negative results.
267
268
  ## Platform limitations
268
269
 
269
270
  - **CPU only.** CUDA/ROCm/XPU kernels are not built; a GPU is never required or used.
270
- - **Windows wheel** ships a prebuilt DLL. That DLL predates the dev5/dev6 2/3-bit
271
- kernels and the restored 4-bit optimizer kernel, so those run the pure-torch
272
- fallback there (same numbers, lower speed).
271
+ - **Windows wheel**: from dev9 on it ships the MSVC DLL built by CI from the current
272
+ `csrc`, which exports the 2/3-bit dequant/GEMM and 4-bit optimizer kernels; they run
273
+ natively on Windows (verified: dequant maxdiff 1.2e-6, 4-bit native vs torch 1.2e-7).
274
+ Wheels older than dev9 fall back to pure torch for those symbols.
273
275
  - **manylinux_2_17 wheel**: the AVX-512 library inside it is single-threaded (the parallel
274
276
  build is only for the native `manylinux_2_39` wheel). On AVX-512 machines set
275
277
  `BNB_CPU_DISABLE_AVX512=1` to run the AVX2 OpenMP path.
@@ -334,6 +336,7 @@ The long-form docs live in the repository (`bitsandbytes/docs_cpu/`, Chinese and
334
336
  | `0.50.2.dev6` | manylinux_2_17 wheel is now **OpenMP** (devtoolset-8, bundled libgomp → 3–4× optimizer-kernel speedup on old glibc); csrc 2/3-bit dequant + fused GEMM kernels with 9-item C self-test; EFST fused tensor-expert routing fix; GCC 8 `<cpuid.h>` portability fix |
335
337
  | `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) |
336
338
  | `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
+ | `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) |
337
340
 
338
341
  ## License and attribution
339
342
 
@@ -141,6 +141,7 @@ tests/test_rmsnorm_fwd_cpu.py
141
141
  tools/create_github_release.ps1
142
142
  tools/dump_signatures.py
143
143
  tools/gen_api_reference.py
144
+ tools/patch_dev9.py
144
145
  tools/publish.ps1
145
146
  tools/sector_carve.c
146
147
  tools/sector_carve_gui.c
@@ -1,39 +1,44 @@
1
- LIBRARY libbitsandbytes_cpu
2
- EXPORTS
3
- gdn_fwd_cpu
4
- gdn_bwd_cpu
5
- cquantize_blockwise_cpu_fp32
6
- cquantize_blockwise_cpu_bf16
7
- cquantize_blockwise_cpu_fp16
8
- cdequantize_blockwise_cpu_fp32
9
- cdequantize_blockwise_cpu_bf16
10
- cdequantize_blockwise_cpu_fp16
11
- cdequantize_blockwise_cpu_fp4_fp32
12
- cdequantize_blockwise_cpu_fp4_bf16
13
- cdequantize_blockwise_cpu_fp4_fp16
14
- cdequantize_blockwise_cpu_nf4_fp32
15
- cdequantize_blockwise_cpu_nf4_bf16
16
- cdequantize_blockwise_cpu_nf4_fp16
17
- cquantize_blockwise_cpu_4bit_fp32
18
- cquantize_blockwise_cpu_4bit_bf16
19
- cquantize_blockwise_cpu_4bit_fp16
20
- cint8_vector_quant_cpu_fp32
21
- cint8_vector_quant_cpu_bf16
22
- cint8_vector_quant_cpu_fp16
23
- cgemv_4bit_inference_cpu_fp32
24
- cgemv_4bit_inference_cpu_bf16
25
- cgemv_4bit_inference_cpu_fp16
26
- cgemv_fp16w_inference_cpu_fp32
27
- cgemv_fp16w_inference_cpu_bf16
28
- cgemv_fp16w_inference_cpu_fp16
29
- cgemm_8bit_inference_cpu_fp32
30
- coptimizer_update_8bit_blockwise_cpu
31
- cfused_rmsnorm_fwd_cpu
32
- cfused_rmsnorm_bwd_cpu
33
- cfused_swiglu_fwd_cpu
34
- cfused_swiglu_bwd_cpu
35
- cfused_add_scale_cpu
36
- cfused_swiglu_strided_fwd_cpu
37
- cfused_swiglu_strided_bwd_cpu
38
- cfused_add_scale_inplace_cpu
39
- cfused_add_scale_rmsnorm_cpu
1
+ LIBRARY libbitsandbytes_cpu
2
+ EXPORTS
3
+ gdn_fwd_cpu
4
+ gdn_bwd_cpu
5
+ cquantize_blockwise_cpu_fp32
6
+ cquantize_blockwise_cpu_bf16
7
+ cquantize_blockwise_cpu_fp16
8
+ cdequantize_blockwise_cpu_fp32
9
+ cdequantize_blockwise_cpu_bf16
10
+ cdequantize_blockwise_cpu_fp16
11
+ cdequantize_blockwise_cpu_fp4_fp32
12
+ cdequantize_blockwise_cpu_fp4_bf16
13
+ cdequantize_blockwise_cpu_fp4_fp16
14
+ cdequantize_blockwise_cpu_nf4_fp32
15
+ cdequantize_blockwise_cpu_nf4_bf16
16
+ cdequantize_blockwise_cpu_nf4_fp16
17
+ cquantize_blockwise_cpu_4bit_fp32
18
+ cquantize_blockwise_cpu_4bit_bf16
19
+ cquantize_blockwise_cpu_4bit_fp16
20
+ cint8_vector_quant_cpu_fp32
21
+ cint8_vector_quant_cpu_bf16
22
+ cint8_vector_quant_cpu_fp16
23
+ cgemv_4bit_inference_cpu_fp32
24
+ cgemv_4bit_inference_cpu_bf16
25
+ cgemv_4bit_inference_cpu_fp16
26
+ cgemv_fp16w_inference_cpu_fp32
27
+ cgemv_fp16w_inference_cpu_bf16
28
+ cgemv_fp16w_inference_cpu_fp16
29
+ cgemm_8bit_inference_cpu_fp32
30
+ coptimizer_update_8bit_blockwise_cpu
31
+ coptimizer_update_4bit_blockwise_cpu
32
+ cbdequantize_3bit_cpu
33
+ cbdequantize_2bit_cpu
34
+ cbq3bit_gemm_cpu
35
+ cbq3bit_gemm_t_cpu
36
+ cfused_rmsnorm_fwd_cpu
37
+ cfused_rmsnorm_bwd_cpu
38
+ cfused_swiglu_fwd_cpu
39
+ cfused_swiglu_bwd_cpu
40
+ cfused_add_scale_cpu
41
+ cfused_swiglu_strided_fwd_cpu
42
+ cfused_swiglu_strided_bwd_cpu
43
+ cfused_add_scale_inplace_cpu
44
+ cfused_add_scale_rmsnorm_cpu
@@ -277,7 +277,7 @@ print(opt.state_bytes(), 'bytes for', p.numel(), 'params =',
277
277
  # what runs today is the verified scalar path
278
278
  # the optimizer is a few percent of a real training step at 30M params,
279
279
  # so the wall-clock cost is small while the state saving is 7.8x
280
- # the pure-torch fallback (Windows wheel) is slower still
280
+ # older Windows wheels run the pure-torch fallback (dev9+ is native)
281
281
  ```
282
282
 
283
283
  **`traps`** -- the failure modes seen while building this, each one silent at first
@@ -286,8 +286,8 @@ print(opt.state_bytes(), 'bytes for', p.numel(), 'params =',
286
286
  # 1. torch.optim.AdamW is NOT replaced: this is a separate class, and
287
287
  # AdamW4bit requires fp32 params (the kernel's dtype 0).
288
288
  # 2. blocksize is compiled in at 256; passing another value raises.
289
- # 3. if the shared library does not export
290
- # coptimizer_update_4bit_blockwise_cpu (the Windows wheel does not),
289
+ # 3. older Windows wheels (< dev9) do not export
290
+ # coptimizer_update_4bit_blockwise_cpu (pre-dev9 Windows wheels),
291
291
  # this class transparently uses a pure-torch implementation with the
292
292
  # same numbers and the same 1.031 B/param state, just slower.
293
293
  # 4. quantised state is not a drop-in for exact reproducibility runs:
@@ -571,7 +571,7 @@ assert [n for n, p in ql.named_parameters() if p.requires_grad] == ['lora_A', 'l
571
571
 
572
572
  **`convert_linears_3bit(model, lora_r=8)`** -- replaces every nn.Linear in place and returns the count; lm_head stays fp32
573
573
 
574
- **`experimental / portability`** -- the Windows wheel predates the 2/3-bit C kernels, so it runs the pure-torch fallback; the Linux wheels use the native kernels
574
+ **`experimental / portability`** -- since dev9 the Windows wheel ships a CI-built MSVC DLL; older wheels use the pure-torch fallback, while the dev9 Windows and Linux wheels use the native kernels
575
575
 
576
576
  ## Reference: bitsandbytes.nn
577
577
 
@@ -276,7 +276,7 @@ print(opt.state_bytes(), 'bytes for', p.numel(), 'params =',
276
276
  # what runs today is the verified scalar path
277
277
  # the optimizer is a few percent of a real training step at 30M params,
278
278
  # so the wall-clock cost is small while the state saving is 7.8x
279
- # the pure-torch fallback (Windows wheel) is slower still
279
+ # older Windows wheels run the pure-torch fallback (dev9+ is native)
280
280
  ```
281
281
 
282
282
  **`traps`** -- the failure modes seen while building this, each one silent at first
@@ -285,8 +285,8 @@ print(opt.state_bytes(), 'bytes for', p.numel(), 'params =',
285
285
  # 1. torch.optim.AdamW is NOT replaced: this is a separate class, and
286
286
  # AdamW4bit requires fp32 params (the kernel's dtype 0).
287
287
  # 2. blocksize is compiled in at 256; passing another value raises.
288
- # 3. if the shared library does not export
289
- # coptimizer_update_4bit_blockwise_cpu (the Windows wheel does not),
288
+ # 3. older Windows wheels (< dev9) do not export
289
+ # coptimizer_update_4bit_blockwise_cpu (pre-dev9 Windows wheels),
290
290
  # this class transparently uses a pure-torch implementation with the
291
291
  # same numbers and the same 1.031 B/param state, just slower.
292
292
  # 4. quantised state is not a drop-in for exact reproducibility runs:
@@ -570,7 +570,7 @@ assert [n for n, p in ql.named_parameters() if p.requires_grad] == ['lora_A', 'l
570
570
 
571
571
  **`convert_linears_3bit(model, lora_r=8)`** -- replaces every nn.Linear in place and returns the count; lm_head stays fp32
572
572
 
573
- **`experimental / portability`** -- the Windows wheel predates the 2/3-bit C kernels, so it runs the pure-torch fallback; the Linux wheels use the native kernels
573
+ **`experimental / portability`** -- since dev9 the Windows wheel ships a CI-built MSVC DLL; older wheels use the pure-torch fallback, while the dev9 Windows and Linux wheels use the native kernels
574
574
 
575
575
  ## Reference: bitsandbytes.nn
576
576
 
@@ -0,0 +1,75 @@
1
+ """dev9: Windows native DLL, q3bit real-model validation, iGPU label fix."""
2
+ from __future__ import annotations
3
+
4
+ import io
5
+
6
+ ROOT = r'D:\work\bitsandbytes-CPU'
7
+ INNER = ROOT + r'\bitsandbytes\README.md'
8
+ EN = ROOT + r'\bitsandbytes\README_EN.md'
9
+ OUTER = ROOT + r'\README.md'
10
+ CLI = ROOT + r'\bitsandbytes\bitsandbytes_cpu_cli.py'
11
+ INIT = ROOT + r'\bitsandbytes\bitsandbytes\__init__.py'
12
+
13
+
14
+ def patch(path: str, pairs: list[tuple[str, str]]) -> None:
15
+ t = io.open(path, encoding='utf-8').read()
16
+ for old, new in pairs:
17
+ assert t.count(old) == 1, (path, old[:70], t.count(old))
18
+ t = t.replace(old, new, 1)
19
+ io.open(path, 'w', encoding='utf-8', newline='').write(t)
20
+ print('updated', path)
21
+
22
+
23
+ patch(INIT, [('__version__ = "0.50.2.dev8"', '__version__ = "0.50.2.dev9"')])
24
+
25
+ patch(INNER, [
26
+ ('**Version: `0.50.2.dev8`**', '**Version: `0.50.2.dev9`**'),
27
+ ('- **Windows wheel** ships a prebuilt DLL. That DLL predates the dev5/dev6 2/3-bit\n'
28
+ ' kernels and the restored 4-bit optimizer kernel, so those run the pure-torch\n'
29
+ ' fallback there (same numbers, lower speed).',
30
+ '- **Windows wheel**: from dev9 on it ships the MSVC DLL built by CI from the current\n'
31
+ ' `csrc`, which exports the 2/3-bit dequant/GEMM and 4-bit optimizer kernels; they run\n'
32
+ ' natively on Windows (verified: dequant maxdiff 1.2e-6, 4-bit native vs torch 1.2e-7).\n'
33
+ ' Wheels older than dev9 fall back to pure torch for those symbols.'),
34
+ ('| 4-bit optimizer | `AdamW4bit`: ≈1.03 bytes of state per parameter (restored in dev8; scalar kernel + torch fallback) |',
35
+ '| 4-bit optimizer | `AdamW4bit`: ≈1.03 bytes of state per parameter (scalar kernel; native on Linux and, since dev9, Windows) |'),
36
+ ('| Fused lm_head 3-bit GEMM (research) | 14–22% faster than fp32 matmul, 89% less weight memory |',
37
+ '| Fused lm_head 3-bit GEMM (research) | 14–22% faster than fp32 matmul, 89% less weight memory |\n'
38
+ '| 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) |'),
39
+ ('| `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 |',
40
+ '| `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 |\n'
41
+ '| `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) |'),
42
+ ])
43
+
44
+ patch(EN, [
45
+ ('(import name unchanged: `import bitsandbytes as bnb`). Current release: **`0.50.2.dev8`**.',
46
+ '(import name unchanged: `import bitsandbytes as bnb`). Current release: **`0.50.2.dev9`**.'),
47
+ ('dev8 restored `AdamW4bit` (1.031 B/param, scalar kernel + torch fallback) and added the\n'
48
+ '`bitsandbytes.q3bit` API (3-bit quant/dequant/fused GEMM, `QuantLinear3bit` frozen base +\n'
49
+ 'LoRA, `convert_linears_3bit`).',
50
+ 'dev8 restored `AdamW4bit` and added the `bitsandbytes.q3bit` API ·\n'
51
+ 'dev9 ships the CI-built MSVC DLL that now exports the 4-bit optimizer and 2/3-bit kernels\n'
52
+ '(native on Windows), validates the 3-bit frozen base + LoRA on Qwen3-1.7B, and fixes the\n'
53
+ 'causal-LM double-shift label bug in the iGPU training kit.'),
54
+ ])
55
+
56
+ patch(OUTER, [
57
+ ('current\nrelease `0.50.2.dev8`', 'current\nrelease `0.50.2.dev9`'),
58
+ ('| `0.50.2.dev8` | restored `AdamW4bit` (1.031 B of state/param, scalar kernel + torch fallback) + new `bitsandbytes.q3bit` API (blockwise 3-bit quant/dequant/fused GEMM, `QuantLinear3bit` frozen 3-bit base + LoRA, `convert_linears_3bit`) |',
59
+ '| `0.50.2.dev8` | restored `AdamW4bit` (1.031 B of state/param, scalar kernel + torch fallback) + new `bitsandbytes.q3bit` API (blockwise 3-bit quant/dequant/fused GEMM, `QuantLinear3bit` frozen 3-bit base + LoRA, `convert_linears_3bit`) |\n'
60
+ '| `0.50.2.dev9` | Windows native kernels (CI-built MSVC DLL now exports the 4-bit optimizer + 2/3-bit symbols), q3bit real-model validation on Qwen3-1.7B, iGPU causal-LM double-shift fix (package v0.1.2) |'),
61
+ ])
62
+
63
+ patch(CLI, [
64
+ ('the Windows wheel predates the 2/3-bit C kernels, so it runs the pure-torch ',
65
+ 'since dev9 the Windows wheel ships a CI-built MSVC DLL; older wheels use the pure-torch '),
66
+ ('fallback; the Linux wheels use the native kernels',
67
+ 'fallback, while the dev9 Windows and Linux wheels use the native kernels'),
68
+ ('# 3. if the shared library does not export',
69
+ '# 3. older Windows wheels (< dev9) do not export'),
70
+ ('# coptimizer_update_4bit_blockwise_cpu (the Windows wheel does not),',
71
+ '# coptimizer_update_4bit_blockwise_cpu (pre-dev9 Windows wheels),'),
72
+ ('# the pure-torch fallback (Windows wheel) is slower still',
73
+ '# older Windows wheels run the pure-torch fallback (dev9+ is native)'),
74
+ ])
75
+ print('dev9 docs/version updated')