MCPcopy Create free account

hub / github.com/bitsandbytes-foundation/bitsandbytes / functions

Functions720 in github.com/bitsandbytes-foundation/bitsandbytes

⨍ Functions 720◇ Types & classes 117↳ Endpoints 78
↓ 2 callersFunctionneon_f32_to_fp16x4
NEON-optimized float to FP16 conversion (4 values at a time)
csrc/cpu_ops.cpp:156
↓ 2 callersFunctionneon_fp16x4_to_f32
NEON-optimized FP16 to float conversion (4 values at a time)
csrc/cpu_ops.cpp:163
↓ 2 callersMethodoverride_config
Override initial optimizer config with specific hyperparameters. The key-values of the optimizer config for the input parameters are
bitsandbytes/optim/optimizer.py:60
↓ 2 callersFunctionprefetch_tensor
(A: torch.Tensor, to_cpu=False)
bitsandbytes/functional.py:102
↓ 2 callersFunctionquantize_4bit
Quantize tensor A in blocks of 4-bit values. Quantizes tensor A by dividing it into blocks which are independently quantized. Args:
bitsandbytes/functional.py:884
↓ 2 callersFunctionquantize_blockwise
Quantize a tensor in blocks of values. The input tensor is quantized by dividing it into blocks of `blocksize` values. The the absolute maxim
bitsandbytes/functional.py:613
↓ 2 callersMethodregister_module_override
(self, module, param_name, config)
bitsandbytes/optim/optimizer.py:113
↓ 2 callersMethodregister_parameters
(self, params)
bitsandbytes/optim/optimizer.py:50
↓ 2 callersFunctionreplace_parameter_4bit_prequantized
( module: nn.Module, param_name: str, qs_dict: dict[str, Any], device: torch.device )
bitsandbytes/nn/parametrize.py:42
↓ 2 callersMethodreset_grads
(self)
bitsandbytes/autograd/_functions.py:92
↓ 2 callersFunctionsync_gpu
(t: torch.Tensor)
bitsandbytes/utils.py:204
↓ 2 callersFunctiontorch_load_from_buffer
(buffer)
tests/helpers.py:60
↓ 2 callersFunctiontorch_save_to_buffer
(obj)
tests/helpers.py:53
↓ 2 callersFunctiontrain_loop
Run training and return list of (step, loss, time) tuples.
examples/xpu/paged_xpu_training.py:107
↓ 2 callersFunctiontrain_loop
(model, optimizer, dataloader, steps, log_interval)
examples/cpu/cpu_training.py:114
↓ 2 callersMethodvectorwise_mm_dequant
Reference implementation for the F.int8_mm_dequant function.
tests/test_functional.py:352
↓ 1 callersMethod__getattr__
(self, name)
bitsandbytes/cextension.py:91
↓ 1 callersMethod__getattr__
Return a dummy function that throws when called, rather than on attribute access
bitsandbytes/cextension.py:322
↓ 1 callersMethod__setstate__
(self, state)
bitsandbytes/optim/optimizer.py:189
↓ 1 callersFunction_compute_update_norm_and_scale
Compute trust-ratio scaling factor for LAMB/LARS and store update norm.
bitsandbytes/backends/cpu/ops.py:327
↓ 1 callersFunction_convert_weight_packed_for_cpu
qweight: (K * N / 2) uint8 return: packed_weight
bitsandbytes/functional.py:1676
↓ 1 callersFunction_current_accelerator_type
()
tests/fsdp_state_dict_save.py:23
↓ 1 callersFunction_decode
(subprocess_err_out_tuple)
bitsandbytes/utils.py:105
↓ 1 callersFunction_dequant_blockwise_fp32_direct
( A_uint8: torch.Tensor, absmax: torch.Tensor, code: torch.Tensor, blocksize: int )
bitsandbytes/backends/cpu/ops.py:455
↓ 1 callersFunction_dequant_linear_fallback
Unfused fallback shared by CUDA and ROCm: reconstruct the (optionally nested) absmax, dequantize the 4-bit weight via the backend dequant impls (r
bitsandbytes/backends/cuda/ops.py:904
↓ 1 callersFunction_dequantize_4bit_impl
( A: torch.Tensor, absmax: torch.Tensor, blocksize: int, quant_type: str, dtype: torch.dty
bitsandbytes/backends/xpu/ops.py:30
↓ 1 callersFunction_dequantize_blockwise_compute
( A_flat: torch.Tensor, absmax: torch.Tensor, code: torch.Tensor, blocksize: int, dtype: torch.dtype )
bitsandbytes/backends/default/ops.py:200
↓ 1 callersFunction_dequantize_blockwise_impl
( A: torch.Tensor, absmax: torch.Tensor, code: torch.Tensor, blocksize: int, dtype: torch.dtype, out: torc
bitsandbytes/backends/xpu/ops.py:66
↓ 1 callersFunction_dequantize_blockwise_pytorch
Pure PyTorch reference implementation for block-wise dequantization.
bitsandbytes/backends/triton/kernels_optim.py:509
↓ 1 callersFunction_disable_parametrization_cache
(module: nn.Module, inputs: tuple[Any, ...], output: Any)
bitsandbytes/nn/parametrize.py:129
↓ 1 callersMethod_fill_padding_idx_with_zero
(self)
bitsandbytes/nn/modules.py:112
↓ 1 callersMethod_fill_padding_idx_with_zero
(self)
bitsandbytes/nn/modules.py:194
↓ 1 callersMethod_format_dependency_error
Format error message for missing shared libraries
bitsandbytes/cextension.py:294
↓ 1 callersMethod_format_lib_error_message
Format detailed error message for library loading failures
bitsandbytes/cextension.py:211
↓ 1 callersMethod_forward_with_partial_dequantize
(self, input: Tensor)
bitsandbytes/nn/modules.py:930
↓ 1 callersFunction_gemm_4bit_default_impl
( A: torch.Tensor, B: torch.Tensor, shapeB: Sequence[int], absmax: torch.Tensor, blocksize
bitsandbytes/backends/default/ops.py:323
↓ 1 callersFunction_gemm_4bit_kernel_impl
Invoke the fused cgemm_4bit_* kernel (shared by the CUDA and ROCm dispatch; the C dispatch in gemm_4bit.cu picks SIMT vs MMA per arch/shape). A is
bitsandbytes/backends/cuda/ops.py:846
↓ 1 callersFunction_gemv_4bit_impl
( A: torch.Tensor, B: torch.Tensor, shapeB: Sequence[int], absmax: torch.Tensor, code: tor
bitsandbytes/backends/mps/ops.py:193
↓ 1 callersFunction_gemv_4bit_impl
( A: torch.Tensor, B: torch.Tensor, shapeB: Sequence[int], absmax: torch.Tensor, code: tor
bitsandbytes/backends/xpu/ops.py:88
↓ 1 callersFunction_gemv_4bit_impl
( A: torch.Tensor, B: torch.Tensor, shapeB: Sequence[int], absmax: torch.Tensor, code: tor
bitsandbytes/backends/cuda/ops.py:522
↓ 1 callersFunction_get_col_absmax
( A: torch.Tensor, threshold=0.0, )
bitsandbytes/backends/cuda/ops.py:277
↓ 1 callersFunction_get_device_and_backend
Auto-detect accelerator device and distributed backend.
tests/fsdp_state_dict_save.py:43
↓ 1 callersMethod_get_state_double_buffer
(self, p, dtype=torch.float32)
bitsandbytes/optim/ademamix.py:260
↓ 1 callersFunction_gpu_dispatch_props
(device_index)
bitsandbytes/backends/cuda/ops.py:73
↓ 1 callersFunction_import_backends
Discover and autoload all available backends installed as separate packages. Packages with an entrypoint for "bitsandbytes.backends" will be
bitsandbytes/__init__.py:52
↓ 1 callersFunction_int8_linear_matmul_impl
(A: torch.Tensor, B: torch.Tensor, out: Optional[torch.Tensor] = None)
bitsandbytes/backends/default/ops.py:134
↓ 1 callersFunction_int8_linear_matmul_impl
(A: torch.Tensor, B: torch.Tensor, out: torch.Tensor)
bitsandbytes/backends/cuda/ops.py:89
↓ 1 callersFunction_optimizer_precondition_32bit
Preprocessing optimizer, computing update norm
bitsandbytes/backends/default/ops.py:369
↓ 1 callersFunction_optimizer_update_32bit
Unified optimizer update kernel
bitsandbytes/backends/default/ops.py:430
↓ 1 callersFunction_print_cuda_diagnostics
(cuda_specs: CUDASpecs)
bitsandbytes/diagnostics/cuda.py:113
↓ 1 callersFunction_print_hip_diagnostics
(cuda_specs: CUDASpecs)
bitsandbytes/diagnostics/cuda.py:139
↓ 1 callersFunction_quant_blockwise_fp32_direct
( A_fp32: torch.Tensor, code: torch.Tensor, absmax_out: torch.Tensor, out_uint8: torch.Tensor, blocksize:
bitsandbytes/backends/cpu/ops.py:461
↓ 1 callersMethod_quantize
(self, device)
bitsandbytes/nn/modules.py:381
↓ 1 callersMethod_quantize
(self, device)
bitsandbytes/nn/modules.py:737
↓ 1 callersFunction_quantize_4bit_compute
( A_flat: torch.Tensor, blocksize: int, bounds: torch.Tensor, order: torch.Tensor, nf4: bo
bitsandbytes/backends/mps/ops.py:91
↓ 1 callersFunction_quantize_4bit_fallback
( A: torch.Tensor, blocksize: int, quant_type: str, quant_storage: torch.dtype )
bitsandbytes/backends/mps/ops.py:120
↓ 1 callersFunction_quantize_blockwise_compute
On torch <= 2.12, torch.bucketize does not perform well. Implements blockwise quantization using a binary search instead of using the default
bitsandbytes/backends/mps/ops.py:53
↓ 1 callersFunction_quantize_blockwise_pytorch
Pure PyTorch reference implementation for block-wise quantization.
bitsandbytes/backends/triton/kernels_optim.py:544
↓ 1 callersFunction_reverse_4bit_compress_format
(weight: torch.Tensor)
bitsandbytes/backends/hpu/ops.py:12
↓ 1 callersFunction_rocm_gfx_arch
gfx arch string (e.g. 'gfx1100') for a ROCm device, feature flags stripped.
bitsandbytes/backends/cuda/ops.py:840
↓ 1 callersMethod_save_to_state_dict
save weight and bias, then fill state_dict with components of quant_state
bitsandbytes/nn/modules.py:593
↓ 1 callersFunction_set_device_index
(index: int, device_type: str)
tests/fsdp_state_dict_save.py:33
↓ 1 callersFunction_set_seed
()
tests/conftest.py:9
↓ 1 callersFunction_setup_ctypes
(names, argtypes, restype=None)
bitsandbytes/backends/cuda/ops.py:16
↓ 1 callersFunctionadjust_num_threads
csrc/cpu_ops.h:71
↓ 1 callersFunctionassert_all_approx_close
(a, b, atol=1e-8, rtol=1e-5, count=10)
tests/test_modules.py:55
↓ 1 callersFunctionbatched_igemm
( A: Tensor, B: Tensor, out: Optional[torch.Tensor] = None, transposed_A=False, transposed
bitsandbytes/functional.py:1440
↓ 1 callersFunctionblock_size_m
block size for AMX gemm
csrc/cpu_ops.h:28
↓ 1 callersFunctionblock_size_n
csrc/cpu_ops.h:30
↓ 1 callersFunctionbuild_quantize_lut
csrc/cpu_ops.cpp:501
↓ 1 callersMethodcast
r"""Make a deep copy of value, casting all tensors to device of param.
bitsandbytes/optim/optimizer.py:232
↓ 1 callersMethodcheck_overrides
(self)
bitsandbytes/optim/optimizer.py:286
↓ 1 callersFunctioncount_params
(model)
examples/xpu/benchmark_paged_memory.py:48
↓ 1 callersMethodcpu
(self)
bitsandbytes/nn/modules.py:397
↓ 1 callersFunctioncvt_fp32_to_bf16
csrc/cpu_ops.cpp:239
↓ 1 callersFunctioncvt_fp32_to_fp16
csrc/cpu_ops.cpp:235
↓ 1 callersFunctiondDequantizeFP4
csrc/xpu_kernels.cpp:8
↓ 1 callersFunctiondDequantizeNF4
csrc/xpu_kernels.cpp:50
↓ 1 callersFunctiondecorator
(fn)
bitsandbytes/backends/default/ops.py:17
↓ 1 callersFunctiondequant_4bit_body_util
(a, offsets, quant_ptr, absmax_ptr, n_elems, QUANT_BLOCK: tl.constexpr)
bitsandbytes/backends/triton/kernels_4bit.py:189
↓ 1 callersFunctiondequant_8bit_blockwise
( a: torch.Tensor, absmax: torch.Tensor, quant_state_code: torch.Tensor, quant_blocksize: int
bitsandbytes/backends/triton/kernels_8bit_quant.py:45
↓ 1 callersFunctiondequant_fp4_body_util
(a, offsets, absmax_ptr, n_elems, QUANT_BLOCK: tl.constexpr)
bitsandbytes/backends/triton/kernels_4bit.py:235
↓ 1 callersFunctiondequant_nf4_body_util
(a, offsets, absmax_ptr, n_elems, QUANT_BLOCK: tl.constexpr)
bitsandbytes/backends/triton/kernels_4bit.py:291
↓ 1 callersFunctiondequantizeBlockwise_bf16
csrc/pythonInterface.cpp:205
↓ 1 callersFunctiondequantizeBlockwise_bf16_fp4
csrc/pythonInterface.cpp:211
↓ 1 callersFunctiondequantizeBlockwise_bf16_nf4
csrc/pythonInterface.cpp:217
↓ 1 callersFunctiondequantizeBlockwise_fp16
csrc/pythonInterface.cpp:169
↓ 1 callersFunctiondequantizeBlockwise_fp16_fp4
csrc/pythonInterface.cpp:175
↓ 1 callersFunctiondequantizeBlockwise_fp16_nf4
csrc/pythonInterface.cpp:181
↓ 1 callersFunctiondequantizeBlockwise_fp32
csrc/pythonInterface.cpp:187
↓ 1 callersFunctiondequantizeBlockwise_fp32_fp4
csrc/pythonInterface.cpp:193
↓ 1 callersFunctiondequantizeBlockwise_fp32_nf4
csrc/pythonInterface.cpp:199
↓ 1 callersFunctiondequantize_fp4_tree
(val, absmax)
bitsandbytes/backends/triton/kernels_4bit.py:211
↓ 1 callersFunctiondequantize_nf4_tree
(val)
bitsandbytes/backends/triton/kernels_4bit.py:251
↓ 1 callersFunctiondescribe_dtype
(dtype: torch.dtype)
tests/helpers.py:102
↓ 1 callersFunctionexecute_and_return_decoded_std_streams
(command_string)
bitsandbytes/utils.py:108
↓ 1 callersFunctionfetch_all_issues
Fetch issues with pagination and exponential backoff.
agents/fetch_issues.py:161
↓ 1 callersFunctionfind_cuda_libraries_in_path_list
(paths_list_candidate: str)
bitsandbytes/diagnostics/cuda.py:50
↓ 1 callersFunctionfind_cudart_libraries
Searches for a cuda installations, in the following order of priority: 1. active conda env 2. LD_LIBRARY_PATH 3. any othe
bitsandbytes/diagnostics/cuda.py:89
↓ 1 callersFunctionfind_outlier_dims
(weight, reduction_dim=0, zscore=4.0, topk=None, rdm=False)
bitsandbytes/utils.py:86
← previousnext →101–200 of 720, ranked by callers