From 8258b4364a21a4da2572cb644d0926080c3268da Mon Sep 17 00:00:00 2001 From: Max Ryabinin Date: Fri, 1 Jul 2022 17:16:10 +0300 Subject: Add a CPU-only build option --- bitsandbytes/__init__.py | 11 ++++++--- bitsandbytes/cextension.py | 13 +++++++++++ bitsandbytes/functional.py | 52 +++++++++++++++++++++--------------------- bitsandbytes/optim/__init__.py | 20 +++++++++------- bitsandbytes/optim/rmsprop.py | 2 +- 5 files changed, 60 insertions(+), 38 deletions(-) create mode 100644 bitsandbytes/cextension.py (limited to 'bitsandbytes') diff --git a/bitsandbytes/__init__.py b/bitsandbytes/__init__.py index 6e29322..22fb841 100644 --- a/bitsandbytes/__init__.py +++ b/bitsandbytes/__init__.py @@ -2,9 +2,14 @@ # # This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. -from .optim import adam + from .nn import modules -__pdoc__ = {'libBitsNBytes' : False, +from cextension import COMPILED_WITH_CUDA + +if COMPILED_WITH_CUDA: + from .optim import adam + +__pdoc__ = {'libBitsNBytes': False, 'optim.optimizer.Optimizer8bit': False, 'optim.optimizer.MockArgs': False - } + } diff --git a/bitsandbytes/cextension.py b/bitsandbytes/cextension.py new file mode 100644 index 0000000..63d627e --- /dev/null +++ b/bitsandbytes/cextension.py @@ -0,0 +1,13 @@ +import ctypes as ct +import os +from warnings import warn + +lib = ct.cdll.LoadLibrary(os.path.dirname(__file__) + '/libbitsandbytes.so') + +try: + lib.cadam32bit_g32 + COMPILED_WITH_CUDA = True +except AttributeError: + warn("The installed version of bitsandbytes was compiled without GPU support. " + "8-bit optimizers and GPU quantization are unavailable.") + COMPILED_WITH_CUDA = False diff --git a/bitsandbytes/functional.py b/bitsandbytes/functional.py index fbd7564..68b1d78 100644 --- a/bitsandbytes/functional.py +++ b/bitsandbytes/functional.py @@ -3,38 +3,38 @@ # This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. import ctypes as ct -import os import random from typing import Tuple import torch from torch import Tensor -lib = ct.cdll.LoadLibrary(os.path.dirname(__file__) + '/libbitsandbytes.so') +from cextension import lib, COMPILED_WITH_CUDA + name2qmap = {} -''' C FUNCTIONS FOR OPTIMIZERS ''' - -str2optimizer32bit = {} -str2optimizer32bit['adam'] = (lib.cadam32bit_g32, lib.cadam32bit_g16) -str2optimizer32bit['momentum'] = (lib.cmomentum32bit_g32, lib.cmomentum32bit_g16) -str2optimizer32bit['rmsprop'] = (lib.crmsprop32bit_g32, lib.crmsprop32bit_g16) -str2optimizer32bit['adagrad'] = (lib.cadagrad32bit_g32, lib.cadagrad32bit_g16) -str2optimizer32bit['lars'] = (lib.cmomentum32bit_g32, lib.cmomentum32bit_g16) -str2optimizer32bit['lamb'] = (lib.cadam32bit_g32, lib.cadam32bit_g16) - -str2optimizer8bit = {} -str2optimizer8bit['adam'] = (lib.cadam_static_8bit_g32, lib.cadam_static_8bit_g16) -str2optimizer8bit['momentum'] = (lib.cmomentum_static_8bit_g32, lib.cmomentum_static_8bit_g16) -str2optimizer8bit['rmsprop'] = (lib.crmsprop_static_8bit_g32, lib.crmsprop_static_8bit_g16) -str2optimizer8bit['lamb'] = (lib.cadam_static_8bit_g32, lib.cadam_static_8bit_g16) -str2optimizer8bit['lars'] = (lib.cmomentum_static_8bit_g32, lib.cmomentum_static_8bit_g16) - -str2optimizer8bit_blockwise = {} -str2optimizer8bit_blockwise['adam'] = (lib.cadam_8bit_blockwise_fp32, lib.cadam_8bit_blockwise_fp16) -str2optimizer8bit_blockwise['momentum'] = (lib.cmomentum_8bit_blockwise_fp32, lib.cmomentum_8bit_blockwise_fp16) -str2optimizer8bit_blockwise['rmsprop'] = (lib.crmsprop_8bit_blockwise_fp32, lib.crmsprop_8bit_blockwise_fp16) -str2optimizer8bit_blockwise['adagrad'] = (lib.cadagrad_8bit_blockwise_fp32, lib.cadagrad_8bit_blockwise_fp16) +if COMPILED_WITH_CUDA: + ''' C FUNCTIONS FOR OPTIMIZERS ''' + str2optimizer32bit = {} + str2optimizer32bit['adam'] = (lib.cadam32bit_g32, lib.cadam32bit_g16) + str2optimizer32bit['momentum'] = (lib.cmomentum32bit_g32, lib.cmomentum32bit_g16) + str2optimizer32bit['rmsprop'] = (lib.crmsprop32bit_g32, lib.crmsprop32bit_g16) + str2optimizer32bit['adagrad'] = (lib.cadagrad32bit_g32, lib.cadagrad32bit_g16) + str2optimizer32bit['lars'] = (lib.cmomentum32bit_g32, lib.cmomentum32bit_g16) + str2optimizer32bit['lamb'] = (lib.cadam32bit_g32, lib.cadam32bit_g16) + + str2optimizer8bit = {} + str2optimizer8bit['adam'] = (lib.cadam_static_8bit_g32, lib.cadam_static_8bit_g16) + str2optimizer8bit['momentum'] = (lib.cmomentum_static_8bit_g32, lib.cmomentum_static_8bit_g16) + str2optimizer8bit['rmsprop'] = (lib.crmsprop_static_8bit_g32, lib.crmsprop_static_8bit_g16) + str2optimizer8bit['lamb'] = (lib.cadam_static_8bit_g32, lib.cadam_static_8bit_g16) + str2optimizer8bit['lars'] = (lib.cmomentum_static_8bit_g32, lib.cmomentum_static_8bit_g16) + + str2optimizer8bit_blockwise = {} + str2optimizer8bit_blockwise['adam'] = (lib.cadam_8bit_blockwise_fp32, lib.cadam_8bit_blockwise_fp16) + str2optimizer8bit_blockwise['momentum'] = (lib.cmomentum_8bit_blockwise_fp32, lib.cmomentum_8bit_blockwise_fp16) + str2optimizer8bit_blockwise['rmsprop'] = (lib.crmsprop_8bit_blockwise_fp32, lib.crmsprop_8bit_blockwise_fp16) + str2optimizer8bit_blockwise['adagrad'] = (lib.cadagrad_8bit_blockwise_fp32, lib.cadagrad_8bit_blockwise_fp16) optimal_normal = [-0.9939730167388916, -0.8727636337280273, -0.8097418546676636, -0.7660024166107178, -0.7318882346153259, -0.6793879270553589, -0.657649040222168, -0.6385974884033203, -0.6211113333702087, -0.5901028513908386, -0.5762918591499329, -0.5630806684494019, -0.5509274005889893, -0.5394591689109802, -0.5283197164535522, -0.517780065536499, -0.5074946284294128, -0.4980469048023224, -0.48867011070251465, -0.48003149032592773, -0.47125306725502014, -0.4629971981048584, -0.4547359049320221, -0.446626216173172, -0.43902668356895447, -0.43158355355262756, -0.4244747757911682, -0.4173796474933624, -0.41038978099823, -0.4055633544921875, -0.4035947024822235, -0.39701032638549805, -0.39057496190071106, -0.38439232110977173, -0.3782760500907898, -0.3721940815448761, -0.3661896586418152, -0.3604033589363098, -0.354605108499527, -0.34892538189888, -0.34320303797721863, -0.3376772701740265, -0.3323028087615967, -0.3269782066345215, -0.32166096568107605, -0.316457599401474, -0.3112771809101105, -0.3061025142669678, -0.30106794834136963, -0.2961243987083435, -0.2912728488445282, -0.28644347190856934, -0.28165507316589355, -0.2769731283187866, -0.2722635865211487, -0.26779335737228394, -0.26314786076545715, -0.2586647868156433, -0.2541804611682892, -0.2496625930070877, -0.24527113139629364, -0.24097171425819397, -0.23659978806972504, -0.23218469321727753, -0.22799566388130188, -0.22380566596984863, -0.21965542435646057, -0.2154538631439209, -0.2113603949546814, -0.20735277235507965, -0.20334717631340027, -0.19932441413402557, -0.19530178606510162, -0.19136647880077362, -0.18736697733402252, -0.18337111175060272, -0.17951400578022003, -0.1757056713104248, -0.17182783782482147, -0.1680615097284317, -0.16431649029254913, -0.16053077578544617, -0.15685945749282837, -0.15298527479171753, -0.1493264138698578, -0.14566898345947266, -0.14188314974308014, -0.13819937407970428, -0.1344561129808426, -0.1306886374950409, -0.1271020770072937, -0.12346585839986801, -0.11981867253780365, -0.11614970862865448, -0.11256207525730133, -0.10889036953449249, -0.10525048524141312, -0.1016591489315033, -0.09824034571647644, -0.09469068050384521, -0.0911419615149498, -0.08773849159479141, -0.08416644483804703, -0.08071305602788925, -0.07720902562141418, -0.07371306419372559, -0.07019119709730148, -0.06673648208379745, -0.06329209357500076, -0.059800852090120316, -0.0564190037548542, -0.05296570807695389, -0.049522045999765396, -0.04609023034572601, -0.04262964054942131, -0.039246633648872375, -0.03577171266078949, -0.03236335143446922, -0.028855687007308006, -0.02542758360505104, -0.022069433704018593, -0.018754752352833748, -0.015386369079351425, -0.01194947212934494, -0.008439815603196621, -0.004995611496269703, -0.0016682245768606663, 0.0, 0.0015510577941313386, 0.005062474869191647, 0.008417150937020779, 0.011741090565919876, 0.015184164978563786, 0.018582714721560478, 0.02204744517803192, 0.025471193715929985, 0.02889077737927437, 0.0323684960603714, 0.03579240292310715, 0.039281025528907776, 0.0427563451230526, 0.04619763046503067, 0.04968220740556717, 0.05326594039797783, 0.05679265409708023, 0.060245808213949203, 0.06372645497322083, 0.06721872836351395, 0.0706876739859581, 0.0742349922657013, 0.07774098962545395, 0.08123527467250824, 0.08468879014253616, 0.08810535818338394, 0.09155989438295364, 0.09498448669910431, 0.0985206812620163, 0.10206405073404312, 0.10563778132200241, 0.10921968519687653, 0.11284469068050385, 0.11653254181146622, 0.12008969485759735, 0.12368203699588776, 0.1272617131471634, 0.13089501857757568, 0.134552001953125, 0.1382799744606018, 0.14194637537002563, 0.14563234150409698, 0.14930322766304016, 0.15303383767604828, 0.1567956507205963, 0.16050070524215698, 0.16431072354316711, 0.16813558340072632, 0.17204202711582184, 0.1758781224489212, 0.17973239719867706, 0.1836014688014984, 0.18753431737422943, 0.19138391315937042, 0.19535475969314575, 0.19931404292583466, 0.20333819091320038, 0.20738255977630615, 0.21152682602405548, 0.21568812429904938, 0.21978361904621124, 0.22393859922885895, 0.22814159095287323, 0.23241068422794342, 0.23675410449504852, 0.24123944342136383, 0.24569889903068542, 0.2500703036785126, 0.25904011726379395, 0.26349544525146484, 0.2682226300239563, 0.272907555103302, 0.2774306833744049, 0.28220856189727783, 0.2869136929512024, 0.2916390895843506, 0.29649388790130615, 0.30142995715141296, 0.3065022826194763, 0.3114383816719055, 0.31648796796798706, 0.3216581642627716, 0.32700115442276, 0.3322487473487854, 0.33778008818626404, 0.3431521952152252, 0.3487405776977539, 0.3543166518211365, 0.3601346015930176, 0.36605337262153625, 0.37217751145362854, 0.378179669380188, 0.3843980133533478, 0.3906566798686981, 0.39714935421943665, 0.40357843041419983, 0.4104187488555908, 0.4171563684940338, 0.42418959736824036, 0.43136918544769287, 0.4389212429523468, 0.44673123955726624, 0.45457619428634644, 0.4627031683921814, 0.47130417823791504, 0.4798591434955597, 0.48897242546081543, 0.4979848861694336, 0.5, 0.5076631307601929, 0.5177803635597229, 0.5282770991325378, 0.5392990112304688, 0.5506287813186646, 0.5632893443107605, 0.5764452815055847, 0.5903191566467285, 0.6051878333091736, 0.6209936141967773, 0.6382884979248047, 0.6573970913887024, 0.6795773506164551, 0.7037051916122437, 0.7327037453651428, 0.7677436470985413, 0.8111193776130676, 0.875165581703186, 1.0] @@ -138,7 +138,7 @@ def estimate_quantiles(A: Tensor, out: Tensor=None, offset: float=1/512) -> Tens elif A.dtype == torch.float16: lib.cestimate_quantiles_fp16(get_ptr(A), get_ptr(out), ct.c_float(offset), ct.c_int(A.numel())) else: - raise NotImplementError(f'Not supported data type {A.dtype}') + raise NotImplementedError(f'Not supported data type {A.dtype}') return out def quantize_blockwise(A: Tensor, code: Tensor=None, absmax: Tensor=None, rand=None, out: Tensor=None) -> Tensor: @@ -384,7 +384,7 @@ def optimizer_update_32bit(optimizer_name:str, g: Tensor, p: Tensor, state1: Ten param_norm = torch.norm(p.data.float()) if optimizer_name not in str2optimizer32bit: - raise NotImplementError(f'Optimizer not implemented: {optimizer_name}. Choices: {",".join(str2optimizer32bit.keys())}') + raise NotImplementedError(f'Optimizer not implemented: {optimizer_name}. Choices: {",".join(str2optimizer32bit.keys())}') if g.dtype == torch.float32 and state1.dtype == torch.float32: str2optimizer32bit[optimizer_name][0](get_ptr(g), get_ptr(p), get_ptr(state1), get_ptr(state2), get_ptr(unorm_vec), ct.c_float(max_unorm), diff --git a/bitsandbytes/optim/__init__.py b/bitsandbytes/optim/__init__.py index 5e73414..e833ecc 100644 --- a/bitsandbytes/optim/__init__.py +++ b/bitsandbytes/optim/__init__.py @@ -2,11 +2,15 @@ # # This source code is licensed under the MIT license found in the # LICENSE file in the root directory of this source tree. -from .adam import Adam, Adam8bit, Adam32bit -from .adamw import AdamW, AdamW8bit, AdamW32bit -from .sgd import SGD, SGD8bit, SGD32bit -from .lars import LARS, LARS8bit, LARS32bit, PytorchLARS -from .lamb import LAMB, LAMB8bit, LAMB32bit -from .rmsprop import RMSprop, RMSprop8bit, RMSprop32bit -from .adagrad import Adagrad, Adagrad8bit, Adagrad32bit -from .optimizer import GlobalOptimManager + +from bitsandbytes.cextension import COMPILED_WITH_CUDA + +if COMPILED_WITH_CUDA: + from .adam import Adam, Adam8bit, Adam32bit + from .adamw import AdamW, AdamW8bit, AdamW32bit + from .sgd import SGD, SGD8bit, SGD32bit + from .lars import LARS, LARS8bit, LARS32bit, PytorchLARS + from .lamb import LAMB, LAMB8bit, LAMB32bit + from .rmsprop import RMSprop, RMSprop8bit, RMSprop32bit + from .adagrad import Adagrad, Adagrad8bit, Adagrad32bit + from .optimizer import GlobalOptimManager diff --git a/bitsandbytes/optim/rmsprop.py b/bitsandbytes/optim/rmsprop.py index 7909d5d..0f1ffaa 100644 --- a/bitsandbytes/optim/rmsprop.py +++ b/bitsandbytes/optim/rmsprop.py @@ -31,6 +31,6 @@ class RMSprop32bit(Optimizer1State): if alpha == 0: raise NotImplementedError(f'RMSprop with alpha==0.0 is not supported!') if centered: - raise NotImplementError(f'Centered RMSprop is not supported!') + raise NotImplementedError(f'Centered RMSprop is not supported!') super(RMSprop32bit, self).__init__('rmsprop', params, lr, (alpha, momentum), eps, weight_decay, 32, args, min_8bit_size, percentile_clipping, block_wise) -- cgit v1.2.3