bitsandbytes-cpu-fork 0.50.2.dev9__tar.gz → 0.50.2.dev10__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.
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/LICENSE +21 -21
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/MANIFEST.in +35 -35
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/PKG-INFO +63 -3
- bitsandbytes_cpu_fork-0.50.2.dev9/bitsandbytes_cpu_fork.egg-info/PKG-INFO → bitsandbytes_cpu_fork-0.50.2.dev10/README.md +375 -346
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/__init__.py +84 -84
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/__main__.py +4 -4
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/_cpu_caps.py +110 -110
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/_ops.py +526 -526
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/autograd/_functions.py +491 -491
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/cpu/ops.py +819 -819
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/cuda/ops.py +1212 -1212
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/default/ops.py +632 -632
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/hpu/ops.py +53 -53
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/mps/ops.py +277 -277
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/triton/kernels_4bit.py +583 -583
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/triton/kernels_8bit_quant.py +195 -195
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/triton/kernels_optim.py +1177 -1177
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/triton/ops.py +304 -304
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/utils.py +94 -94
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/xpu/ops.py +305 -305
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/cextension.py +447 -447
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/cli.py +50 -50
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/consts.py +12 -12
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/cuda_specs.py +111 -111
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/detect_cpu.py +189 -189
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/diagnostics/cuda.py +193 -193
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/diagnostics/main.py +134 -134
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/diagnostics/utils.py +12 -12
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/disk_balancer.py +1687 -1687
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/efst.py +782 -782
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/functional.py +1836 -1836
- bitsandbytes_cpu_fork-0.50.2.dev10/bitsandbytes/fused/__init__.py +21 -0
- bitsandbytes_cpu_fork-0.50.2.dev10/bitsandbytes/fused/_native.py +66 -0
- bitsandbytes_cpu_fork-0.50.2.dev10/bitsandbytes/fused/_ref.py +62 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/gdn_cpu.py +616 -616
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/libbitsandbytes_cpu.dll +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/nn/__init__.py +19 -19
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/nn/modules.py +1220 -1220
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/nn/parametrize.py +206 -206
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/optim/__init__.py +26 -26
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/optim/adagrad.py +187 -187
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/optim/adam.py +346 -346
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/optim/adamw.py +337 -337
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/optim/adamw4bit.py +239 -239
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/optim/ademamix.py +410 -410
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/optim/lamb.py +190 -190
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/optim/lars.py +259 -259
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/optim/lion.py +266 -266
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/optim/optimizer.py +756 -756
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/optim/rmsprop.py +170 -170
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/optim/sgd.py +152 -152
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/q3bit.py +487 -481
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/torch_cpu_kit.py +794 -794
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/utils.py +208 -208
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes_cpu_cli.py +1062 -1062
- bitsandbytes_cpu_fork-0.50.2.dev9/README.md → bitsandbytes_cpu_fork-0.50.2.dev10/bitsandbytes_cpu_fork.egg-info/PKG-INFO +406 -315
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes_cpu_fork.egg-info/SOURCES.txt +10 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/build_manual/build_manual.bat +234 -234
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/build_manual/export.def +44 -44
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/build_manual/export_generic.def +39 -39
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/build_manual/selftest_win.bat +37 -37
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/avx2_gemv_4bit.h +25 -3
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/common.cuh +105 -105
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/compat.cuh +206 -206
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/compat_device.cuh +51 -51
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/cpu_ops.cpp +45 -19
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/gemm_4bit.cu +168 -168
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/gemm_4bit_common.cuh +105 -105
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/gemm_4bit_mma.cuh +125 -125
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/gemm_4bit_simt.cu +547 -547
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/gemm_4bit_simt.cuh +16 -16
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/gemm_4bit_sm75.cu +416 -416
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/gemm_4bit_sm75.cuh +16 -16
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/gemm_4bit_sm80.cu +699 -699
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/gemm_4bit_sm80.cuh +16 -16
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/kernels.cu +1863 -1863
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/kernels.cuh +85 -85
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/ops.cu +602 -602
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/ops.cuh +151 -151
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/test_avx2_gemv_4bit.cpp +51 -4
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/docs_cpu/API_REFERENCE.md +16 -1
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/docs_cpu/API_REFERENCE_EN.md +1 -1
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/docs_cpu/ISA_MATRIX.md +2 -2
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/docs_cpu/QUICKSTART.md +77 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/docs_cpu/QUICKSTART_EN.md +88 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/docs_cpu/TECH_REPORT.md +71 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/docs_cpu/TECH_REPORT_EN.md +84 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/docs_cpu/check_docs.py +157 -157
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/examples/compile_inference.py +32 -32
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/examples/cpu/cpu_training.py +370 -370
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/examples/int8_inference_huggingface.py +19 -19
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/examples/xpu/benchmark_paged_memory.py +253 -253
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/examples/xpu/paged_xpu_training.py +369 -369
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/pyproject.toml +124 -124
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/setup.py +175 -175
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tests/test_autograd.py +232 -232
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tests/test_cpu_e2e.py +201 -201
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tests/test_cuda_setup_evaluator.py +135 -135
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tests/test_functional.py +1049 -1049
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tests/test_generation.py +122 -122
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tests/test_linear4bit.py +614 -614
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tests/test_linear8bitlt.py +366 -366
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tests/test_modules.py +494 -494
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tests/test_ops.py +474 -474
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tests/test_optim.py +743 -743
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tests/test_parametrize.py +499 -499
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tests/test_rmsnorm_fwd_cpu.py +198 -198
- bitsandbytes_cpu_fork-0.50.2.dev10/tools/carve_selftest.py +225 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tools/create_github_release.ps1 +169 -169
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tools/dump_signatures.py +64 -64
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tools/gen_api_reference.py +43 -43
- bitsandbytes_cpu_fork-0.50.2.dev10/tools/mft_recover.c +148 -0
- bitsandbytes_cpu_fork-0.50.2.dev10/tools/mft_recover_gui.c +643 -0
- bitsandbytes_cpu_fork-0.50.2.dev10/tools/mft_recover_gui.manifest +37 -0
- bitsandbytes_cpu_fork-0.50.2.dev10/tools/mft_selftest.py +515 -0
- bitsandbytes_cpu_fork-0.50.2.dev10/tools/mftcore.h +690 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tools/patch_dev9.py +75 -75
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tools/publish.ps1 +99 -99
- bitsandbytes_cpu_fork-0.50.2.dev10/tools/recovery_pipeline_selftest.py +347 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tools/sector_carve.c +233 -15
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tools/sector_carve_gui.c +222 -16
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tools/sector_mirror.c +97 -2
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tools/sector_mirror_gui.c +318 -20
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tools/train_llm_sft.py +136 -111
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tools/verify_help_examples.py +213 -213
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/tools/verify_train_release.py +90 -90
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/autograd/__init__.py +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/__init__.py +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/cpu/__init__.py +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/cuda/__init__.py +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/default/__init__.py +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/hpu/__init__.py +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/mps/__init__.py +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/triton/__init__.py +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/backends/xpu/__init__.py +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/diagnostics/__init__.py +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/py.typed +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes/vcomp140.dll +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes_cpu_fork.egg-info/dependency_links.txt +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes_cpu_fork.egg-info/entry_points.txt +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes_cpu_fork.egg-info/requires.txt +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/bitsandbytes_cpu_fork.egg-info/top_level.txt +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/build_linux.sh +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/build_manual/export_avx512.def +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/build_termux.sh +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/common.h +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/cpu_cache.h +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/cpu_features.h +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/cpu_gdn.cpp +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/cpu_ops.h +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/pythonInterface.cpp +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/test_cpu_cache.cpp +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/xpu_kernels.cpp +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/xpu_kernels.h +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/xpu_ops.cpp +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/csrc/xpu_ops.h +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/docs_cpu/EPYC_VALIDATION.md +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/docs_cpu/MOTION_BOUNDARY.md +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/docs_cpu/MOTION_BOUNDARY_EN.md +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/docs_cpu/TECHNICAL_GUIDE.md +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/docs_cpu/TECHNICAL_GUIDE_EN.md +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/docs_cpu/WHEEL_PLATFORM_NOTES.md +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/selftest_cpu.c +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/setup.cfg +0 -0
- {bitsandbytes_cpu_fork-0.50.2.dev9 → bitsandbytes_cpu_fork-0.50.2.dev10}/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 +1,35 @@
|
|
|
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
|
+
include build_linux.sh
|
|
32
|
+
include build_termux.sh
|
|
33
|
+
include selftest_cpu.c
|
|
34
|
+
include README.md
|
|
35
|
+
include LICENSE
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: bitsandbytes-cpu-fork
|
|
3
|
-
Version: 0.50.2.
|
|
3
|
+
Version: 0.50.2.dev10
|
|
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
|
|
@@ -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
|
|
279
|
-
Termux
|
|
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
|