From 8258b4364a21a4da2572cb644d0926080c3268da Mon Sep 17 00:00:00 2001
From: Max Ryabinin <mryabinin0@gmail.com>
Date: Fri, 1 Jul 2022 17:16:10 +0300
Subject: [PATCH 1/7] Add a CPU-only build option

---
 Makefile                       |  19 +-
 bitsandbytes/__init__.py       |  11 +-
 bitsandbytes/cextension.py     |  13 +
 bitsandbytes/functional.py     |  48 ++--
 bitsandbytes/optim/__init__.py |  20 +-
 bitsandbytes/optim/rmsprop.py  |   2 +-
 csrc/common.cpp                |  39 +++
 csrc/common.h                  |  23 ++
 csrc/cpu_ops.cpp               |  57 +++++
 csrc/cpu_ops.h                 |   9 +
 csrc/ops.cu                    | 453 ++++++++++++++-------------------
 csrc/ops.cuh                   |  10 -
 csrc/pythonInterface.c         | 110 ++++----
 setup.py                       |  22 +-
 14 files changed, 454 insertions(+), 382 deletions(-)
 create mode 100644 bitsandbytes/cextension.py
 create mode 100644 csrc/common.cpp
 create mode 100644 csrc/common.h
 create mode 100644 csrc/cpu_ops.cpp
 create mode 100644 csrc/cpu_ops.h

diff --git a/Makefile b/Makefile
index f21fde3..b58e233 100644
--- a/Makefile
+++ b/Makefile
@@ -10,10 +10,10 @@ NVCC := $(CUDA_HOME)/bin/nvcc
 ###########################################
 
 CSRC := $(ROOT_DIR)/csrc
-BUILD_DIR:= $(ROOT_DIR)/cuda_build
+BUILD_DIR:= $(ROOT_DIR)/build
 
 FILES_CUDA := $(CSRC)/ops.cu $(CSRC)/kernels.cu
-FILES_CPP := $(CSRC)/pythonInterface.c
+FILES_CPP := $(CSRC)/common.cpp $(CSRC)/cpu_ops.cpp $(CSRC)/pythonInterface.c
 
 INCLUDE :=  -I $(CUDA_HOME)/include -I $(ROOT_DIR)/csrc -I $(CONDA_PREFIX)/include -I $(ROOT_DIR)/dependencies/cub -I $(ROOT_DIR)/include
 LIB := -L $(CUDA_HOME)/lib64 -lcudart -lcuda -lcublas -lcurand -lcusparse -L $(CONDA_PREFIX)/lib
@@ -46,27 +46,30 @@ CC_CUDA11x += -gencode arch=compute_86,code=sm_86
 all: $(ROOT_DIR)/dependencies/cub $(BUILD_DIR) env
 	$(NVCC) $(COMPUTE_CAPABILITY) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR)
 	$(NVCC) $(COMPUTE_CAPABILITY) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o 
-	$(GPP) -std=c++14 -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
+	$(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
 
 cuda92: $(ROOT_DIR)/dependencies/cub $(BUILD_DIR) env
 	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA92) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR)
 	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA92) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o 
-	$(GPP) -std=c++14 -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
+	$(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
 
 cuda10x: $(ROOT_DIR)/dependencies/cub $(BUILD_DIR) env
 	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA10x) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR)
 	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA10x) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o 
-	$(GPP) -std=c++14 -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
+	$(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
 
 cuda110: $(BUILD_DIR) env
 	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA110) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR)
 	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA110) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o 
-	$(GPP) -std=c++14 -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
+	$(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
 
 cuda11x: $(BUILD_DIR) env
 	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA11x) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR)
 	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA11x) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o 
-	$(GPP) -std=c++14 -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
+	$(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so $(LIB)
+
+cpuonly: $(BUILD_DIR) env
+	$(GPP) -std=c++14 -shared -fPIC -I $(ROOT_DIR)/csrc -I $(ROOT_DIR)/include $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes.so
 
 env:
 	@echo "ENVIRONMENT"
@@ -80,7 +83,7 @@ env:
 	@echo "============================"
 
 $(BUILD_DIR):
-	mkdir -p cuda_build
+	mkdir -p build
 	mkdir -p dependencies
 
 $(ROOT_DIR)/dependencies/cub:
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 '''
+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)
 
-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 = {}
-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)
+    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)
diff --git a/csrc/common.cpp b/csrc/common.cpp
new file mode 100644
index 0000000..972602b
--- /dev/null
+++ b/csrc/common.cpp
@@ -0,0 +1,39 @@
+#include <common.h>
+#include <float.h>
+
+void *quantize_block(void *arguments) {
+    // 1. find absmax in block
+    // 2. divide input value by absmax to normalize into [-1.0, 1.0]
+    // 3. do binary search to find the closest value
+    // 4. check minimal distance
+    // 5. store index
+
+    struct quantize_block_args *args = (quantize_block_args *) arguments;
+
+    // 1. find absmax in block
+    float absmax_block = -FLT_MAX;
+    for (int i = args->block_idx; i < args->block_end; i++)
+        absmax_block = fmax(absmax_block, fabs(args->A[i]));
+
+    args->absmax[args->block_idx / BLOCK_SIZE] = absmax_block;
+
+    for (int i = args->block_idx; i < args->block_end; i++) {
+        // 2. divide input value by absmax to normalize into [-1.0, 1.0]
+        // 3. do binary search to find the closest value
+        float normed_value = args->A[i] / absmax_block;
+        int idx = args->bin_searcher->scalar(normed_value);
+
+        // 4. check minimal distance
+        // The binary search returns always the value to the left, which might not be the closest value
+        if (idx < 255) {
+            float dist_left = fabs(normed_value - (args->code[idx]));
+            float dist_right = fabs(normed_value - (args->code[idx + 1]));
+            if (dist_right < dist_left) { idx += 1; }
+        }
+
+        // 5. store index
+        args->out[i] = (unsigned char) idx;
+    }
+
+    return NULL;
+}
diff --git a/csrc/common.h b/csrc/common.h
new file mode 100644
index 0000000..35f2463
--- /dev/null
+++ b/csrc/common.h
@@ -0,0 +1,23 @@
+#include <BinSearch.h>
+
+#ifndef common
+#define common
+
+using namespace BinSearch;
+
+struct quantize_block_args {
+    BinAlgo<Scalar, float, Direct2> *bin_searcher;
+    float *code;
+    float *A;
+    float *absmax;
+    unsigned char *out;
+    int block_end;
+    int block_idx;
+    int threadidx;
+};
+
+#define BLOCK_SIZE 4096
+
+void *quantize_block(void *arguments);
+
+#endif
\ No newline at end of file
diff --git a/csrc/cpu_ops.cpp b/csrc/cpu_ops.cpp
new file mode 100644
index 0000000..11a2615
--- /dev/null
+++ b/csrc/cpu_ops.cpp
@@ -0,0 +1,57 @@
+#include <BinSearch.h>
+#include <pthread.h>
+#include <common.h>
+
+using namespace BinSearch;
+
+void dequantize_cpu(float *code, unsigned char *A, float *absmax, float *out, int n) {
+    for (int block_idx = 0; block_idx < n; block_idx += BLOCK_SIZE) {
+        int valid_items = n - block_idx >= BLOCK_SIZE ? BLOCK_SIZE : n - block_idx;
+        int block_end = block_idx + valid_items;
+        for (int i = block_idx; i < block_end; i++)
+            out[i] = code[A[i]] * absmax[block_idx / BLOCK_SIZE];
+    }
+}
+
+void quantize_cpu(float *code, float *A, float *absmax, unsigned char *out, int n) {
+
+    // the default code is has range [-0.993, 1.0] which can cause an error in the binary search algorithm used below
+    code[0] = -1.0f;
+
+    int num_blocks = n / BLOCK_SIZE;
+    num_blocks += n % BLOCK_SIZE == 0 ? 0 : 1;
+
+    pthread_t *threads = (pthread_t *) malloc(sizeof(pthread_t) * num_blocks);
+    struct quantize_block_args **args = (quantize_block_args **) malloc(num_blocks * sizeof(quantize_block_args *));
+
+    for (int i = 0; i < num_blocks; i++)
+        args[i] = (quantize_block_args *) malloc(sizeof(quantize_block_args));
+
+    const uint32 elements_code = 256;
+    BinAlgo<Scalar, float, Direct2> bin_searcher(code, elements_code);
+
+    for (int block_idx = 0; block_idx < n; block_idx += BLOCK_SIZE) {
+        int valid_items = n - block_idx >= BLOCK_SIZE ? BLOCK_SIZE : n - block_idx;
+        int block_end = block_idx + valid_items;
+
+        struct quantize_block_args *arg = args[block_idx / BLOCK_SIZE];
+        arg->bin_searcher = &bin_searcher;
+        arg->code = code;
+        arg->A = A;
+        arg->absmax = absmax;
+        arg->out = out;
+        arg->block_end = block_end;
+        arg->block_idx = block_idx;
+        arg->threadidx = block_idx / BLOCK_SIZE;
+
+        pthread_create(&threads[block_idx / BLOCK_SIZE], NULL, &quantize_block, (void *) arg);
+    }
+
+    for (int i = 0; i < num_blocks; i++)
+        int err = pthread_join(threads[i], NULL);
+
+    free(threads);
+    for (int i = 0; i < num_blocks; i++)
+        free(args[i]);
+    free(args);
+}
\ No newline at end of file
diff --git a/csrc/cpu_ops.h b/csrc/cpu_ops.h
new file mode 100644
index 0000000..57145a9
--- /dev/null
+++ b/csrc/cpu_ops.h
@@ -0,0 +1,9 @@
+#ifndef BITSANDBYTES_CPU_OPS_H
+#define BITSANDBYTES_CPU_OPS_H
+
+
+void quantize_cpu(float *code, float *A, float *absmax, unsigned char *out, int n);
+
+void dequantize_cpu(float *code, unsigned char *A, float *absmax, float *out, int n);
+
+#endif
diff --git a/csrc/ops.cu b/csrc/ops.cu
index 9691241..464ea2e 100644
--- a/csrc/ops.cu
+++ b/csrc/ops.cu
@@ -8,251 +8,141 @@
 #include <cub/device/device_scan.cuh>
 #include <limits>
 #include <BinSearch.h>
+#include <common.h>
 
 
 using namespace BinSearch;
 using std::cout;
 using std::endl;
 
-#define BLOCK_SIZE 4096
+void histogramScatterAdd2D(float *histogram, int *index1, int *index2, float *src, int maxidx1, int n) {
+    int threads = 512;
+    int blocks = n / threads;
+    blocks = n % threads == 0 ? blocks : blocks + 1;
+    kHistogramScatterAdd2D<<<blocks, 512>>>(histogram, index1, index2, src, maxidx1, n);
+    CUDA_CHECK_RETURN(cudaPeekAtLastError());
+}
 
-struct quantize_block_args
-{
-  BinAlgo<Scalar, float, Direct2> *bin_searcher;
-  float *code;
-  float *A;
-  float *absmax;
-  unsigned char *out;
-  int block_end;
-  int block_idx;
-  int threadidx;
-};
+template<typename T>
+void estimateQuantiles(T *A, float *code, float offset, int n) {
+    int blocks = n / 4096;
+    blocks = n % 4096 == 0 ? blocks : blocks + 1;
+    CUDA_CHECK_RETURN(cudaMemset(code, 0, 256 * sizeof(float)));
+    kEstimateQuantiles < T ><<<blocks, 512>>>(A, code, offset, std::numeric_limits<T>::max(), n);
+    CUDA_CHECK_RETURN(cudaPeekAtLastError());
+}
 
-void *quantize_block(void *arguments)
-{
-  // 1. find absmax in block
-  // 2. divide input value by absmax to normalize into [-1.0, 1.0]
-  // 3. do binary search to find the closest value
-  // 4. check minimal distance
-  // 5. store index
+void quantize(float *code, float *A, unsigned char *out, int n) {
+    int blocks = n / 1024;
+    blocks = n % 1024 == 0 ? blocks : blocks + 1;
+    kQuantize<<<blocks, 1024>>>(code, A, out, n);
+    CUDA_CHECK_RETURN(cudaPeekAtLastError());
+}
 
-  struct quantize_block_args *args = (quantize_block_args*)arguments;
+void dequantize(float *code, unsigned char *A, float *out, int n) {
+    int blocks = n / 1024;
+    blocks = n % 1024 == 0 ? blocks : blocks + 1;
+    kDequantize<<<blocks, 1024>>>(code, A, out, n);
+    CUDA_CHECK_RETURN(cudaPeekAtLastError());
+}
 
-  // 1. find absmax in block
-  float absmax_block = -FLT_MAX;
-  for (int i = args->block_idx; i < args->block_end; i++)
-    absmax_block = fmax(absmax_block, fabs(args->A[i]));
+template<typename T, int STOCHASTIC>
+void quantizeBlockwise(float *code, T *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n) {
+    int blocks = n / 4096;
+    blocks = n % 4096 == 0 ? blocks : blocks + 1;
+    kQuantizeBlockwise < T, 4096, 4, STOCHASTIC ><<<blocks, 1024>>>(code, A, absmax, out, rand, rand_offset, n);
+    CUDA_CHECK_RETURN(cudaPeekAtLastError());
+}
 
-  args->absmax[args->block_idx/BLOCK_SIZE] = absmax_block;
+template<typename T>
+void dequantizeBlockwise(float *code, unsigned char *A, float *absmax, T *out, int blocksize, const int n) {
+    int blocks = n / blocksize;
+    blocks = n % blocksize == 0 ? blocks : blocks + 1;
+    if (blocksize == 4096)
+        kDequantizeBlockwise < T, 4096, 1024, 4 ><<<blocks, 4096 / 4>>>(code, A, absmax, out, n);
+    else if (blocksize == 2048)
+        kDequantizeBlockwise < T, 2048, 512, 4 ><<<blocks, 2048 / 4>>>(code, A, absmax, out, n);
+    CUDA_CHECK_RETURN(cudaPeekAtLastError());
+}
 
-  for (int i = args->block_idx; i < args->block_end; i++)
-  {
-    // 2. divide input value by absmax to normalize into [-1.0, 1.0]
-    // 3. do binary search to find the closest value
-    float normed_value = args->A[i]/absmax_block;
-    int idx = args->bin_searcher->scalar(normed_value);
+template<typename T, int OPTIMIZER>
+void optimizer32bit(T *g, T *p,
+                    float *state1, float *state2, float *unorm, float max_unorm, float param_norm,
+                    const float beta1, const float beta2, const float eps, const float weight_decay,
+                    const int step, const float lr, const float gnorm_scale, bool skip_zeros, const int n) {
+    int blocks = n / 4096;
+    blocks = n % 4096 == 0 ? blocks : blocks + 1;
+    switch (OPTIMIZER) {
+        case ADAM:
+            if (max_unorm > 0.0f) {
+                CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1 * sizeof(float)));
+                kPreconditionOptimizer32bit2State < T, OPTIMIZER, 4096,
+                        8 ><<<blocks, 512>>>(g, p, state1, state2, unorm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, n);
+                CUDA_CHECK_RETURN(cudaPeekAtLastError());
+            }
+            kOptimizer32bit2State < T,
+                    OPTIMIZER ><<<blocks, 1024>>>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
+            CUDA_CHECK_RETURN(cudaPeekAtLastError());
+            break;
+        case MOMENTUM:
+        case RMSPROP:
+        case ADAGRAD:
 
-    // 4. check minimal distance
-    // The binary search returns always the value to the left, which might not be the closest value
-    if(idx < 255)
-    {
-      float dist_left = fabs(normed_value-(args->code[idx]));
-      float dist_right = fabs(normed_value-(args->code[idx+1]));
-      if(dist_right < dist_left){ idx+=1; }
+            if (max_unorm > 0.0f) {
+                CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1 * sizeof(float)));
+                kPreconditionOptimizer32bit1State < T, OPTIMIZER, 4096,
+                        8 ><<<blocks, 512>>>(g, p, state1, unorm, beta1, eps, weight_decay, step, lr, gnorm_scale, n);
+                CUDA_CHECK_RETURN(cudaPeekAtLastError());
+            }
+
+            kOptimizer32bit1State < T,
+                    OPTIMIZER ><<<blocks, 1024>>>(g, p, state1, unorm, max_unorm, param_norm, beta1, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
+            CUDA_CHECK_RETURN(cudaPeekAtLastError());
+            break;
     }
-
-    // 5. store index
-    args->out[i] = (unsigned char)idx;
-  }
-
-  return NULL;
 }
 
-void quantize_cpu(float *code, float *A, float *absmax, unsigned char *out, int n)
-{
+template<typename T, int OPTIMIZER>
+void optimizerStatic8bit(T *p, T *g,
+                         unsigned char *state1, unsigned char *state2,
+                         float *unorm, float max_unorm, float param_norm,
+                         float beta1, float beta2,
+                         float eps, int step, float lr,
+                         float *quantiles1, float *quantiles2,
+                         float *max1, float *max2, float *new_max1, float *new_max2,
+                         float weight_decay,
+                         const float gnorm_scale, int n) {
+    int blocks = n / 4096;
+    blocks = n % 4096 == 0 ? blocks : blocks + 1;
 
-  // the default code is has range [-0.993, 1.0] which can cause an error in the binary search algorithm used below
-  code[0] = -1.0f; 
+    if (max_unorm > 0.0f) { CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1 * sizeof(float))); }
 
-  int num_blocks = n/BLOCK_SIZE;
-  num_blocks += n % BLOCK_SIZE == 0 ? 0 : 1;
-
-  pthread_t *threads = (pthread_t*)malloc(sizeof(pthread_t)*num_blocks);
-  struct quantize_block_args **args = (quantize_block_args**)malloc(num_blocks*sizeof(quantize_block_args*));
-
-  for(int i = 0; i < num_blocks; i++)
-    args[i] = (quantize_block_args*)malloc(sizeof(quantize_block_args));
-
-  const uint32 elements_code = 256;
-  BinAlgo<Scalar, float, Direct2> bin_searcher(code, elements_code);
-
-  for(int block_idx = 0; block_idx < n; block_idx+=BLOCK_SIZE)
-  {
-    int valid_items = n-block_idx >= BLOCK_SIZE ? BLOCK_SIZE : n - block_idx;
-    int block_end = block_idx + valid_items;
-
-    struct quantize_block_args *arg = args[block_idx/BLOCK_SIZE];
-    arg->bin_searcher = &bin_searcher;
-    arg->code = code;
-    arg->A = A;
-    arg->absmax = absmax;
-    arg->out = out;
-    arg->block_end = block_end;
-    arg->block_idx = block_idx;
-    arg->threadidx = block_idx/BLOCK_SIZE;
- 
-    pthread_create(&threads[block_idx/BLOCK_SIZE], NULL, &quantize_block, (void *)arg);
-  }
-
-  for(int i = 0; i < num_blocks; i++)
-    int err = pthread_join(threads[i], NULL);
-
-  free(threads);
-  for(int i = 0; i < num_blocks; i++)
-    free(args[i]);
-  free(args);
-}
-
-
-void dequantize_cpu(float *code, unsigned char *A, float *absmax, float *out, int n)
-{
-  for(int block_idx = 0; block_idx < n; block_idx+=BLOCK_SIZE)
-  {
-    int valid_items = n-block_idx >= BLOCK_SIZE ? BLOCK_SIZE : n - block_idx;
-    int block_end = block_idx + valid_items;
-    for (int i = block_idx; i < block_end; i++)
-      out[i] = code[A[i]]*absmax[block_idx/BLOCK_SIZE];
-  }
-}
-
-void histogramScatterAdd2D(float* histogram, int *index1, int *index2, float *src, int maxidx1, int n)
-{
-  int threads = 512;
-  int blocks = n/threads;
-  blocks = n % threads == 0 ? blocks : blocks + 1;
-  kHistogramScatterAdd2D<<<blocks, 512>>>(histogram, index1, index2, src, maxidx1, n);
-  CUDA_CHECK_RETURN(cudaPeekAtLastError());
-}
-
-template <typename T> void estimateQuantiles(T *A, float *code, float offset, int n)
-{
-  int blocks = n/4096;
-  blocks = n % 4096 == 0 ? blocks : blocks + 1;
-	CUDA_CHECK_RETURN(cudaMemset(code, 0, 256*sizeof(float)));
-  kEstimateQuantiles<T><<<blocks, 512>>>(A, code, offset, std::numeric_limits<T>::max(), n);
-  CUDA_CHECK_RETURN(cudaPeekAtLastError());
-}
-
-void quantize(float *code, float *A, unsigned char *out, int n)
-{
-  int blocks = n/1024;
-  blocks = n % 1024 == 0 ? blocks : blocks + 1;
-  kQuantize<<<blocks, 1024>>>(code, A, out, n);
-  CUDA_CHECK_RETURN(cudaPeekAtLastError());
-}
-
-void dequantize(float *code, unsigned char *A, float *out, int n)
-{
-  int blocks = n/1024;
-  blocks = n % 1024 == 0 ? blocks : blocks + 1;
-  kDequantize<<<blocks, 1024>>>(code, A, out, n);
-  CUDA_CHECK_RETURN(cudaPeekAtLastError());
-}
-
-template <typename T, int STOCHASTIC> void quantizeBlockwise(float * code, T *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n)
-{
-  int blocks = n/4096;
-  blocks = n % 4096 == 0 ? blocks : blocks + 1;
-  kQuantizeBlockwise<T, 4096, 4, STOCHASTIC><<<blocks, 1024>>>(code, A, absmax, out, rand, rand_offset, n);
-  CUDA_CHECK_RETURN(cudaPeekAtLastError());
-}
-
-template<typename T> void dequantizeBlockwise(float *code, unsigned char *A, float *absmax, T *out, int blocksize, const int n)
-{
-  int blocks = n/blocksize;
-  blocks = n % blocksize == 0 ? blocks : blocks + 1;
-  if(blocksize == 4096)
-    kDequantizeBlockwise<T, 4096, 1024, 4><<<blocks, 4096/4>>>(code, A, absmax, out, n);
-  else if(blocksize == 2048)
-    kDequantizeBlockwise<T, 2048, 512, 4><<<blocks, 2048/4>>>(code, A, absmax, out, n);
-  CUDA_CHECK_RETURN(cudaPeekAtLastError());
-}
-
-template<typename T, int OPTIMIZER> void optimizer32bit(T* g, T* p, 
-                float* state1, float* state2, float *unorm, float max_unorm, float param_norm,
-                const float beta1, const float beta2, const float eps, const float weight_decay,
-                const int step, const float lr, const float gnorm_scale, bool skip_zeros, const int n)
-{
-  int blocks = n/4096;
-  blocks = n % 4096 == 0 ? blocks : blocks + 1;
-	switch(OPTIMIZER)
-	{
-		case ADAM:
-      if(max_unorm > 0.0f)
-			{ 
-				CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1*sizeof(float)));
-        kPreconditionOptimizer32bit2State<T, OPTIMIZER, 4096, 8><<<blocks, 512>>>(g, p, state1, state2, unorm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, n);
-        CUDA_CHECK_RETURN(cudaPeekAtLastError());
-      }
-			kOptimizer32bit2State<T, OPTIMIZER><<<blocks, 1024>>>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
-      CUDA_CHECK_RETURN(cudaPeekAtLastError());
-			break;
-		case MOMENTUM:
-    case RMSPROP:
-    case ADAGRAD:
-
-      if(max_unorm > 0.0f)
-			{ 
-				CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1*sizeof(float)));
-				kPreconditionOptimizer32bit1State<T, OPTIMIZER, 4096, 8><<<blocks, 512>>>(g, p, state1, unorm, beta1, eps, weight_decay, step, lr, gnorm_scale, n);
-        CUDA_CHECK_RETURN(cudaPeekAtLastError());
-			}
-
-			kOptimizer32bit1State<T, OPTIMIZER><<<blocks, 1024>>>(g, p, state1, unorm, max_unorm, param_norm, beta1, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
-      CUDA_CHECK_RETURN(cudaPeekAtLastError());
-			break;
-	}
-}
-
-template<typename T, int OPTIMIZER> void optimizerStatic8bit(T* p, T* g,
-                unsigned char* state1, unsigned char* state2,
-                float *unorm, float max_unorm, float param_norm,
-                float beta1, float beta2,
-                float eps, int step, float lr, 
-                float* quantiles1, float* quantiles2,
-                float* max1, float* max2, float* new_max1, float* new_max2,
-                float weight_decay,
-                const float gnorm_scale, int n)
-{
-  int blocks = n/4096;
-  blocks = n % 4096 == 0 ? blocks : blocks + 1;
-
-  if(max_unorm > 0.0f){ CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1*sizeof(float))); }
-
-	switch(OPTIMIZER)
-	{
-		case ADAM:
-			CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1*sizeof(float)));
-			CUDA_CHECK_RETURN(cudaMemset(new_max2, 0, 1*sizeof(float)));
-			kPreconditionOptimizerStatic8bit2State<T, OPTIMIZER><<<blocks, 256>>>(p, g, state1, state2, unorm, beta1, beta2, eps, step, quantiles1, quantiles2, max1, max2, new_max1, new_max2, gnorm_scale, n);
-			CUDA_CHECK_RETURN(cudaPeekAtLastError());
-			kOptimizerStatic8bit2State<T, OPTIMIZER><<<blocks, 1024>>>(p, g, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr,
-																														quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n);
-			CUDA_CHECK_RETURN(cudaPeekAtLastError());
-		break;
-		case MOMENTUM:
-    case RMSPROP:
-    case ADAGRAD:
-			CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1*sizeof(float)));
-			kPreconditionOptimizerStatic8bit1State<T, OPTIMIZER><<<blocks, 256>>>(p, g, state1, unorm, beta1, eps, step, quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
-			CUDA_CHECK_RETURN(cudaPeekAtLastError());
-			kOptimizerStatic8bit1State<T, OPTIMIZER><<<blocks, 1024>>>(p, g, state1, unorm, max_unorm, param_norm, beta1, eps, step, lr,
-																														quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
-			CUDA_CHECK_RETURN(cudaPeekAtLastError());
-			break;
-		default:
-			break;
-	}
+    switch (OPTIMIZER) {
+        case ADAM:
+            CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1 * sizeof(float)));
+            CUDA_CHECK_RETURN(cudaMemset(new_max2, 0, 1 * sizeof(float)));
+            kPreconditionOptimizerStatic8bit2State < T,
+                    OPTIMIZER ><<<blocks, 256>>>(p, g, state1, state2, unorm, beta1, beta2, eps, step, quantiles1, quantiles2, max1, max2, new_max1, new_max2, gnorm_scale, n);
+            CUDA_CHECK_RETURN(cudaPeekAtLastError());
+            kOptimizerStatic8bit2State < T,
+                    OPTIMIZER ><<<blocks, 1024>>>(p, g, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr,
+            quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n);
+            CUDA_CHECK_RETURN(cudaPeekAtLastError());
+            break;
+        case MOMENTUM:
+        case RMSPROP:
+        case ADAGRAD:
+            CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1 * sizeof(float)));
+            kPreconditionOptimizerStatic8bit1State < T,
+                    OPTIMIZER ><<<blocks, 256>>>(p, g, state1, unorm, beta1, eps, step, quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
+            CUDA_CHECK_RETURN(cudaPeekAtLastError());
+            kOptimizerStatic8bit1State < T, OPTIMIZER ><<<blocks, 1024>>>(p, g, state1, unorm, max_unorm, param_norm, beta1, eps, step, lr,
+            quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
+            CUDA_CHECK_RETURN(cudaPeekAtLastError());
+            break;
+        default:
+            break;
+    }
 }
 
 #define BLOCKSIZE_2STATE 2048
@@ -260,42 +150,43 @@ template<typename T, int OPTIMIZER> void optimizerStatic8bit(T* p, T* g,
 #define BLOCKSIZE_1STATE 2048
 #define NUM_1STATE 8
 
-template<typename T, int OPTIMIZER> void optimizerStatic8bitBlockwise(T* p, T* g,
-                unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr, 
-                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n)
-{
+template<typename T, int OPTIMIZER>
+void optimizerStatic8bitBlockwise(T *p, T *g,
+                                  unsigned char *state1, unsigned char *state2, float beta1, float beta2, float eps, int step, float lr,
+                                  float *quantiles1, float *quantiles2, float *absmax1, float *absmax2, float weight_decay,
+                                  const float gnorm_scale, bool skip_zeros, int n) {
 
-	int blocks = 0;
-	switch(OPTIMIZER)
-	{
-		case ADAM:
-			blocks = n/BLOCKSIZE_2STATE;
-			blocks = n % BLOCKSIZE_2STATE == 0 ? blocks : blocks + 1;
-			kOptimizerStatic8bit2StateBlockwise<T, OPTIMIZER, BLOCKSIZE_2STATE, NUM_2STATE><<<blocks, BLOCKSIZE_2STATE/NUM_2STATE>>>(p, g, state1, state2, beta1, beta2, eps, step, lr,
-																														quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n);
-			CUDA_CHECK_RETURN(cudaPeekAtLastError());
-		break;
-		case MOMENTUM:
-		case RMSPROP:
-    case ADAGRAD:
-			blocks = n/BLOCKSIZE_1STATE;
-			blocks = n % BLOCKSIZE_1STATE == 0 ? blocks : blocks + 1;
-			kOptimizerStatic8bit1StateBlockwise<T, OPTIMIZER, BLOCKSIZE_1STATE, NUM_1STATE><<<blocks, BLOCKSIZE_1STATE/NUM_1STATE>>>(p, g, state1, beta1, beta2, eps, step, lr,
-																														quantiles1, absmax1, weight_decay, gnorm_scale, skip_zeros, n);
-			CUDA_CHECK_RETURN(cudaPeekAtLastError());
-		break;
-	}
+    int blocks = 0;
+    switch (OPTIMIZER) {
+        case ADAM:
+            blocks = n / BLOCKSIZE_2STATE;
+            blocks = n % BLOCKSIZE_2STATE == 0 ? blocks : blocks + 1;
+            kOptimizerStatic8bit2StateBlockwise < T, OPTIMIZER, BLOCKSIZE_2STATE, NUM_2STATE ><<<blocks, BLOCKSIZE_2STATE /
+                                                                                                         NUM_2STATE>>>(p, g, state1, state2, beta1, beta2, eps, step, lr,
+            quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n);
+            CUDA_CHECK_RETURN(cudaPeekAtLastError());
+            break;
+        case MOMENTUM:
+        case RMSPROP:
+        case ADAGRAD:
+            blocks = n / BLOCKSIZE_1STATE;
+            blocks = n % BLOCKSIZE_1STATE == 0 ? blocks : blocks + 1;
+            kOptimizerStatic8bit1StateBlockwise < T, OPTIMIZER, BLOCKSIZE_1STATE, NUM_1STATE ><<<blocks, BLOCKSIZE_1STATE /
+                                                                                                         NUM_1STATE>>>(p, g, state1, beta1, beta2, eps, step, lr,
+            quantiles1, absmax1, weight_decay, gnorm_scale, skip_zeros, n);
+            CUDA_CHECK_RETURN(cudaPeekAtLastError());
+            break;
+    }
 }
 
 
-
-template<typename T> void percentileClipping(T * g, float *gnorm_vec, int step, const int n)
-{
-  int blocks = n/2048;
-  blocks = n % 2048 == 0 ? blocks : blocks + 1;
-	CUDA_CHECK_RETURN(cudaMemset(&gnorm_vec[step % 100], 0, 1*sizeof(float)));
-  kPercentileClipping<T, 2048, 4><<<blocks, 512>>>(g, gnorm_vec, step, n);
-  CUDA_CHECK_RETURN(cudaPeekAtLastError());
+template<typename T>
+void percentileClipping(T *g, float *gnorm_vec, int step, const int n) {
+    int blocks = n / 2048;
+    blocks = n % 2048 == 0 ? blocks : blocks + 1;
+    CUDA_CHECK_RETURN(cudaMemset(&gnorm_vec[step % 100], 0, 1 * sizeof(float)));
+    kPercentileClipping < T, 2048, 4 ><<<blocks, 512>>>(g, gnorm_vec, step, n);
+    CUDA_CHECK_RETURN(cudaPeekAtLastError());
 }
 
 
@@ -304,13 +195,23 @@ template<typename T> void percentileClipping(T * g, float *gnorm_vec, int step,
 //==============================================================
 
 template void estimateQuantiles(half *A, float *code, float offset, int n);
+
 template void estimateQuantiles(float *A, float *code, float offset, int n);
 
-template void quantizeBlockwise<half, 0>(float * code, half *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n);
-template void quantizeBlockwise<float, 0>(float * code, float *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n);
-template void quantizeBlockwise<half, 1>(float * code, half *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n);
-template void quantizeBlockwise<float, 1>(float * code, float *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n);
+template void
+quantizeBlockwise<half, 0>(float *code, half *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n);
+
+template void
+quantizeBlockwise<float, 0>(float *code, float *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n);
+
+template void
+quantizeBlockwise<half, 1>(float *code, half *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n);
+
+template void
+quantizeBlockwise<float, 1>(float *code, float *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n);
+
 template void dequantizeBlockwise<half>(float *code, unsigned char *A, float *absmax, half *out, int blocksize, const int n);
+
 template void dequantizeBlockwise<float>(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n);
 
 #define MAKE_optimizer32bit(name, gtype) \
@@ -320,12 +221,19 @@ template void optimizer32bit<gtype, name>(gtype* g, gtype* p, \
                 const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n);
 
 MAKE_optimizer32bit(ADAM, half)
+
 MAKE_optimizer32bit(ADAM, float)
+
 MAKE_optimizer32bit(MOMENTUM, half)
+
 MAKE_optimizer32bit(MOMENTUM, float)
+
 MAKE_optimizer32bit(RMSPROP, half)
+
 MAKE_optimizer32bit(RMSPROP, float)
+
 MAKE_optimizer32bit(ADAGRAD, half)
+
 MAKE_optimizer32bit(ADAGRAD, float)
 
 #define MAKE_optimizerStatic8bit(name, gtype) \
@@ -338,11 +246,17 @@ template void optimizerStatic8bit<gtype, name>(gtype* p, gtype* g, unsigned char
                 float weight_decay, \
                 const float gnorm_scale, int n); \
 
+
 MAKE_optimizerStatic8bit(ADAM, half)
+
 MAKE_optimizerStatic8bit(ADAM, float)
+
 MAKE_optimizerStatic8bit(MOMENTUM, half)
+
 MAKE_optimizerStatic8bit(MOMENTUM, float)
+
 MAKE_optimizerStatic8bit(RMSPROP, half)
+
 MAKE_optimizerStatic8bit(RMSPROP, float)
 
 #define MAKE_optimizerStatic8bitBlockwise(gtype, optim_name) \
@@ -350,14 +264,23 @@ template void optimizerStatic8bitBlockwise<gtype, optim_name>(gtype* p, gtype* g
                 unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr,  \
                 float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n); \
 
+
 MAKE_optimizerStatic8bitBlockwise(half, ADAM);
+
 MAKE_optimizerStatic8bitBlockwise(float, ADAM);
+
 MAKE_optimizerStatic8bitBlockwise(half, MOMENTUM);
+
 MAKE_optimizerStatic8bitBlockwise(float, MOMENTUM);
+
 MAKE_optimizerStatic8bitBlockwise(half, RMSPROP);
+
 MAKE_optimizerStatic8bitBlockwise(float, RMSPROP);
+
 MAKE_optimizerStatic8bitBlockwise(half, ADAGRAD);
+
 MAKE_optimizerStatic8bitBlockwise(float, ADAGRAD);
 
-template void percentileClipping(float * g, float *gnorm_vec, int step, const int n);
-template void percentileClipping(half * g, float *gnorm_vec, int step, const int n);
+template void percentileClipping(float *g, float *gnorm_vec, int step, const int n);
+
+template void percentileClipping(half *g, float *gnorm_vec, int step, const int n);
diff --git a/csrc/ops.cuh b/csrc/ops.cuh
index 1bc13fb..8fb4cec 100644
--- a/csrc/ops.cuh
+++ b/csrc/ops.cuh
@@ -68,16 +68,6 @@ template<typename T, int OPTIMIZER> void optimizerStatic8bitBlockwise(T* p, T* g
 
 template<typename T> void percentileClipping(T * g, float *gnorm_vec, int step, const int n);
 
-void quantize_cpu(float *code, float *A, float *absmax, unsigned char *out, int n);
-void dequantize_cpu(float *code, unsigned char *A, float *absmax, float *out, int n);
-
 void histogramScatterAdd2D(float* histogram, int *index1, int *index2, float *src, int maxidx1, int n);
 
 #endif
-
-
-
-
-
-
-
diff --git a/csrc/pythonInterface.c b/csrc/pythonInterface.c
index e0b0d59..229b7ed 100644
--- a/csrc/pythonInterface.c
+++ b/csrc/pythonInterface.c
@@ -3,7 +3,10 @@
 // This source code is licensed under the MIT license found in the 
 // LICENSE file in the root directory of this source tree.
 
+#if BUILD_CUDA
 #include <ops.cuh>
+#endif
+#include <cpu_ops.h>
 
 // We cannot call templated code from C, so we wrap the template in a C compatible call here if necessary.
 // We use macro functions to expand all the different optimizers. Looks ugly, and is ugly, but its better than to 
@@ -12,6 +15,7 @@
 //                               UNMANGLED CALLS
 //===================================================================================
 
+#if BUILD_CUDA
 void estimateQuantiles_fp32(float *A, float *code, float offset, int n){ estimateQuantiles<float>(A, code, offset, n); }
 void estimateQuantiles_fp16(half *A, float *code, float offset, int n){ estimateQuantiles<half>(A, code, offset, n); }
 
@@ -34,15 +38,15 @@ MAKE_FUNC32(adagrad, ADAGRAD, half, 16)
 
 #define MAKE_FUNC8(fname, oname, gtype, gbits) \
 void fname##_static_8bit_g##gbits(gtype* p, gtype* g, unsigned char* state1, unsigned char* state2, \
-								float *unorm, float max_unorm, float param_norm, \
+                                float *unorm, float max_unorm, float param_norm, \
                 float beta1, float beta2, \
                 float eps, int step, float lr,  \
                 float* quantiles1, float* quantiles2, \
                 float* max1, float* max2, float* new_max1, float* new_max2, \
                 float weight_decay, float gnorm_scale, int n) \
 {  \
-	optimizerStatic8bit<gtype, oname>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \
-			                                  quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n); \
+    optimizerStatic8bit<gtype, oname>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \
+                                              quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n); \
 } \
 
 MAKE_FUNC8(adam, ADAM, float, 32)
@@ -78,39 +82,41 @@ void quantizeBlockwise_stochastic_fp32(float * code, float *A, float *absmax, un
 
 void dequantizeBlockwise_fp16(float *code, unsigned char *A, float *absmax, half *out, int blocksize, const int n){ dequantizeBlockwise<half>(code, A, absmax, out, blocksize, n); } \
 void dequantizeBlockwise_fp32(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n){ dequantizeBlockwise<float>(code, A, absmax, out, blocksize, n); }
+#endif
 
 extern "C"
 {
-	void cestimate_quantiles_fp32(float *A, float *code, float offset, int n){ estimateQuantiles_fp32(A, code, offset, n); }
-	void cestimate_quantiles_fp16(half *A, float *code, float offset, int n){ estimateQuantiles_fp16(A, code, offset, n); }
-	void cquantize(float *code, float *A, unsigned char *out, int n){ quantize(code, A, out, n); }
-	void cdequantize(float *code, unsigned char *A, float *out, int n){ dequantize(code, A, out, n); }
-  void cquantize_blockwise_fp16(float * code, half *A, float *absmax, unsigned char *out, const int n){ quantizeBlockwise_fp16(code, A, absmax, out, n); }
-  void cquantize_blockwise_fp32(float * code, float *A, float *absmax, unsigned char *out, const int n){ quantizeBlockwise_fp32(code, A, absmax, out, n); }
-  void cquantize_blockwise_stochastic_fp16(float * code, half *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n){ quantizeBlockwise_stochastic_fp16(code, A, absmax, out, rand, rand_offset, n); }
-  void cquantize_blockwise_stochastic_fp32(float * code, float *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n){ quantizeBlockwise_stochastic_fp32(code, A, absmax, out, rand, rand_offset, n); }
+#if BUILD_CUDA
+void cestimate_quantiles_fp32(float *A, float *code, float offset, int n){ estimateQuantiles_fp32(A, code, offset, n); }
+void cestimate_quantiles_fp16(half *A, float *code, float offset, int n){ estimateQuantiles_fp16(A, code, offset, n); }
+void cquantize(float *code, float *A, unsigned char *out, int n){ quantize(code, A, out, n); }
+void cdequantize(float *code, unsigned char *A, float *out, int n){ dequantize(code, A, out, n); }
+void cquantize_blockwise_fp16(float * code, half *A, float *absmax, unsigned char *out, const int n){ quantizeBlockwise_fp16(code, A, absmax, out, n); }
+void cquantize_blockwise_fp32(float * code, float *A, float *absmax, unsigned char *out, const int n){ quantizeBlockwise_fp32(code, A, absmax, out, n); }
+void cquantize_blockwise_stochastic_fp16(float * code, half *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n){ quantizeBlockwise_stochastic_fp16(code, A, absmax, out, rand, rand_offset, n); }
+void cquantize_blockwise_stochastic_fp32(float * code, float *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n){ quantizeBlockwise_stochastic_fp32(code, A, absmax, out, rand, rand_offset, n); }
 
-  void cdequantize_blockwise_fp16(float *code, unsigned char *A, float *absmax, half *out, int blocksize, const int n){ dequantizeBlockwise_fp16(code, A, absmax, out, blocksize, n); }
-  void cdequantize_blockwise_fp32(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n){ dequantizeBlockwise_fp32(code, A, absmax, out, blocksize, n); }
+void cdequantize_blockwise_fp16(float *code, unsigned char *A, float *absmax, half *out, int blocksize, const int n){ dequantizeBlockwise_fp16(code, A, absmax, out, blocksize, n); }
+void cdequantize_blockwise_fp32(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n){ dequantizeBlockwise_fp32(code, A, absmax, out, blocksize, n); }
 
-	#define MAKE_CFUNC32(name, gtype, gbits) \
-	void c##name##32bit_g##gbits(gtype *g, gtype *p, \
-								 float* state1, float* state2, float *unorm, float max_unorm, float param_norm, \
-								 const float beta1, const float beta2, const float eps, const float weight_decay, \
-								 const int step, const float lr, const float gnorm_scale, bool skip_zeros, const int n) \
-	{ name##32bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); } \
+#define MAKE_CFUNC32(name, gtype, gbits) \
+    void c##name##32bit_g##gbits(gtype *g, gtype *p, \
+                                 float* state1, float* state2, float *unorm, float max_unorm, float param_norm, \
+                                 const float beta1, const float beta2, const float eps, const float weight_decay, \
+                                 const int step, const float lr, const float gnorm_scale, bool skip_zeros, const int n) \
+    { name##32bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); } \
 
-	MAKE_CFUNC32(adam, float, 32)
-	MAKE_CFUNC32(adam, half, 16)
-	MAKE_CFUNC32(momentum, float, 32)
-	MAKE_CFUNC32(momentum, half, 16)
-	MAKE_CFUNC32(rmsprop, float, 32)
-	MAKE_CFUNC32(rmsprop, half, 16)
-	MAKE_CFUNC32(adagrad, float, 32)
-	MAKE_CFUNC32(adagrad, half, 16)
+MAKE_CFUNC32(adam, float, 32)
+MAKE_CFUNC32(adam, half, 16)
+MAKE_CFUNC32(momentum, float, 32)
+MAKE_CFUNC32(momentum, half, 16)
+MAKE_CFUNC32(rmsprop, float, 32)
+MAKE_CFUNC32(rmsprop, half, 16)
+MAKE_CFUNC32(adagrad, float, 32)
+MAKE_CFUNC32(adagrad, half, 16)
 
-	#define MAKE_CFUNC8(name, gtype, gbits) \
-	void c##name##_static_8bit_g##gbits(gtype* p, gtype* g, unsigned char* state1, unsigned char* state2, \
+#define MAKE_CFUNC8(name, gtype, gbits) \
+    void c##name##_static_8bit_g##gbits(gtype* p, gtype* g, unsigned char* state1, unsigned char* state2, \
                 float *unorm, float max_unorm, float param_norm, \
                 float beta1, float beta2, \
                 float eps, int step, float lr,  \
@@ -118,40 +124,40 @@ extern "C"
                 float* max1, float* max2, float* new_max1, float* new_max2, \
                 float weight_decay, float gnorm_scale, int n) \
   {  \
-	    name##_static_8bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \
-			                                 quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n); \
+        name##_static_8bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \
+                                             quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n); \
   } \
 
-	MAKE_CFUNC8(adam, float, 32)
-	MAKE_CFUNC8(adam, half, 16)
-	MAKE_CFUNC8(momentum, float, 32)
-	MAKE_CFUNC8(momentum, half, 16)
-	MAKE_CFUNC8(rmsprop, float, 32)
-	MAKE_CFUNC8(rmsprop, half, 16)
+MAKE_CFUNC8(adam, float, 32)
+MAKE_CFUNC8(adam, half, 16)
+MAKE_CFUNC8(momentum, float, 32)
+MAKE_CFUNC8(momentum, half, 16)
+MAKE_CFUNC8(rmsprop, float, 32)
+MAKE_CFUNC8(rmsprop, half, 16)
 
-  #define MAKE_CBLOCKWISE8(fname, optim_name, gtype, gbits) \
+#define MAKE_CBLOCKWISE8(fname, optim_name, gtype, gbits) \
   void c##fname##_8bit_blockwise_fp##gbits(gtype* p, gtype* g, \
                 unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr,  \
                 float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n) \
   {	fname##_8bit_blockwise_fp##gbits(p, g, state1, state2, beta1, beta2, eps, step, lr, quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n); } \
 
-	MAKE_CBLOCKWISE8(adam, ADAM, half, 16)
-	MAKE_CBLOCKWISE8(adam, ADAM, float, 32)
-	MAKE_CBLOCKWISE8(momentum, MOMENTUM, half, 16)
-	MAKE_CBLOCKWISE8(momentum, MOMENTUM, float, 32)
-	MAKE_CBLOCKWISE8(rmsprop, RMSPROP, half, 16)
-	MAKE_CBLOCKWISE8(rmsprop, RMSPROP, float, 32)
-	MAKE_CBLOCKWISE8(adagrad, ADAGRAD, half, 16)
-	MAKE_CBLOCKWISE8(adagrad, ADAGRAD, float, 32)
+MAKE_CBLOCKWISE8(adam, ADAM, half, 16)
+MAKE_CBLOCKWISE8(adam, ADAM, float, 32)
+MAKE_CBLOCKWISE8(momentum, MOMENTUM, half, 16)
+MAKE_CBLOCKWISE8(momentum, MOMENTUM, float, 32)
+MAKE_CBLOCKWISE8(rmsprop, RMSPROP, half, 16)
+MAKE_CBLOCKWISE8(rmsprop, RMSPROP, float, 32)
+MAKE_CBLOCKWISE8(adagrad, ADAGRAD, half, 16)
+MAKE_CBLOCKWISE8(adagrad, ADAGRAD, float, 32)
 
 
-	void cpercentile_clipping_g32(float * g, float *gnorm_vec, int step, const int n){ percentileClipping_g32(g, gnorm_vec, step, n); }
-	void cpercentile_clipping_g16(half * g, float *gnorm_vec, int step, const int n){ percentileClipping_g16(g, gnorm_vec, step, n); }
+void cpercentile_clipping_g32(float * g, float *gnorm_vec, int step, const int n){ percentileClipping_g32(g, gnorm_vec, step, n); }
+void cpercentile_clipping_g16(half * g, float *gnorm_vec, int step, const int n){ percentileClipping_g16(g, gnorm_vec, step, n); }
+void chistogram_scatter_add_2d(float* histogram, int *index1, int *index2, float *src, int maxidx1, int n){ histogramScatterAdd2D(histogram, index1, index2, src, maxidx1, n); }
+#endif
 
-	void cquantize_blockwise_cpu_fp32(float *code, float *A, float *absmax, unsigned char *out, const int n){ quantize_cpu(code, A, absmax, out, n); }
-	void cdequantize_blockwise_cpu_fp32(float *code, unsigned char *A, float *absmax, float *out, const int n){ dequantize_cpu(code, A, absmax, out, n); }
-
-	void chistogram_scatter_add_2d(float* histogram, int *index1, int *index2, float *src, int maxidx1, int n){ histogramScatterAdd2D(histogram, index1, index2, src, maxidx1, n); }
+void cquantize_blockwise_cpu_fp32(float *code, float *A, float *absmax, unsigned char *out, const int n){ quantize_cpu(code, A, absmax, out, n); }
+void cdequantize_blockwise_cpu_fp32(float *code, unsigned char *A, float *absmax, float *out, const int n){ dequantize_cpu(code, A, absmax, out, n); }
 }
 
 
diff --git a/setup.py b/setup.py
index 59cd78e..2402c02 100644
--- a/setup.py
+++ b/setup.py
@@ -6,27 +6,27 @@ import os
 from setuptools import setup, find_packages
 
 
-
 def read(fname):
     return open(os.path.join(os.path.dirname(__file__), fname)).read()
 
 
+version = os.getenv("CUDA_VERSION", "cpu")
+
 setup(
-    name = f"bitsandbytes-cuda{os.environ['CUDA_VERSION']}",
-    version = "0.26.0",
-    author = "Tim Dettmers",
-    author_email = "dettmers@cs.washington.edu",
-    description = ("8-bit optimizers and quantization routines."),
-    license = "MIT",
-    keywords = "gpu optimizers optimization 8-bit quantization compression",
-    url = "http://packages.python.org/bitsandbytes",
+    name="bitsandbytes",
+    version=f"0.26.0+{version}",
+    author="Tim Dettmers",
+    author_email="dettmers@cs.washington.edu",
+    description="8-bit optimizers and quantization routines.",
+    license="MIT",
+    keywords="gpu optimizers optimization 8-bit quantization compression",
+    url="http://packages.python.org/bitsandbytes",
     packages=find_packages(),
     package_data={'': ['libbitsandbytes.so']},
     long_description=read('README.md'),
-    long_description_content_type = 'text/markdown',
+    long_description_content_type='text/markdown',
     classifiers=[
         "Development Status :: 4 - Beta",
         'Topic :: Scientific/Engineering :: Artificial Intelligence'
     ],
 )
-

From e4cf33f2a38e54ca2a523bc5d5923237c68b0112 Mon Sep 17 00:00:00 2001
From: Max Ryabinin <mryabinin0@gmail.com>
Date: Fri, 1 Jul 2022 17:25:44 +0300
Subject: [PATCH 2/7] Fix imports

---
 bitsandbytes/__init__.py       | 2 +-
 bitsandbytes/functional.py     | 2 +-
 bitsandbytes/optim/__init__.py | 3 ++-
 3 files changed, 4 insertions(+), 3 deletions(-)

diff --git a/bitsandbytes/__init__.py b/bitsandbytes/__init__.py
index 22fb841..02ca804 100644
--- a/bitsandbytes/__init__.py
+++ b/bitsandbytes/__init__.py
@@ -4,7 +4,7 @@
 # LICENSE file in the root directory of this source tree.
 
 from .nn import modules
-from cextension import COMPILED_WITH_CUDA
+from .cextension import COMPILED_WITH_CUDA
 
 if COMPILED_WITH_CUDA:
     from .optim import adam
diff --git a/bitsandbytes/functional.py b/bitsandbytes/functional.py
index 68b1d78..ab4e565 100644
--- a/bitsandbytes/functional.py
+++ b/bitsandbytes/functional.py
@@ -9,7 +9,7 @@ from typing import Tuple
 import torch
 from torch import Tensor
 
-from cextension import lib, COMPILED_WITH_CUDA
+from .cextension import lib, COMPILED_WITH_CUDA
 
 name2qmap = {}
 
diff --git a/bitsandbytes/optim/__init__.py b/bitsandbytes/optim/__init__.py
index e833ecc..42b5bc0 100644
--- a/bitsandbytes/optim/__init__.py
+++ b/bitsandbytes/optim/__init__.py
@@ -13,4 +13,5 @@ if COMPILED_WITH_CUDA:
     from .lamb import LAMB, LAMB8bit, LAMB32bit
     from .rmsprop import RMSprop, RMSprop8bit, RMSprop32bit
     from .adagrad import Adagrad, Adagrad8bit, Adagrad32bit
-    from .optimizer import GlobalOptimManager
+
+from .optimizer import GlobalOptimManager

From 31ce1b3708751016bf5e14beff7ae0a99c975991 Mon Sep 17 00:00:00 2001
From: Max Ryabinin <mryabinin0@gmail.com>
Date: Fri, 1 Jul 2022 17:36:30 +0300
Subject: [PATCH 3/7] Reduce diff

---
 csrc/common.h    |   2 +-
 csrc/cpu_ops.cpp |   2 +-
 csrc/ops.cu      | 280 +++++++++++++++++++++--------------------------
 3 files changed, 124 insertions(+), 160 deletions(-)

diff --git a/csrc/common.h b/csrc/common.h
index 35f2463..2f25a58 100644
--- a/csrc/common.h
+++ b/csrc/common.h
@@ -20,4 +20,4 @@ struct quantize_block_args {
 
 void *quantize_block(void *arguments);
 
-#endif
\ No newline at end of file
+#endif
diff --git a/csrc/cpu_ops.cpp b/csrc/cpu_ops.cpp
index 11a2615..89de52d 100644
--- a/csrc/cpu_ops.cpp
+++ b/csrc/cpu_ops.cpp
@@ -54,4 +54,4 @@ void quantize_cpu(float *code, float *A, float *absmax, unsigned char *out, int
     for (int i = 0; i < num_blocks; i++)
         free(args[i]);
     free(args);
-}
\ No newline at end of file
+}
diff --git a/csrc/ops.cu b/csrc/ops.cu
index 464ea2e..b2a1105 100644
--- a/csrc/ops.cu
+++ b/csrc/ops.cu
@@ -46,103 +46,100 @@ void dequantize(float *code, unsigned char *A, float *out, int n) {
     CUDA_CHECK_RETURN(cudaPeekAtLastError());
 }
 
-template<typename T, int STOCHASTIC>
-void quantizeBlockwise(float *code, T *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n) {
-    int blocks = n / 4096;
-    blocks = n % 4096 == 0 ? blocks : blocks + 1;
-    kQuantizeBlockwise < T, 4096, 4, STOCHASTIC ><<<blocks, 1024>>>(code, A, absmax, out, rand, rand_offset, n);
-    CUDA_CHECK_RETURN(cudaPeekAtLastError());
+template <typename T, int STOCHASTIC> void quantizeBlockwise(float * code, T *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n)
+{
+  int blocks = n/4096;
+  blocks = n % 4096 == 0 ? blocks : blocks + 1;
+  kQuantizeBlockwise<T, 4096, 4, STOCHASTIC><<<blocks, 1024>>>(code, A, absmax, out, rand, rand_offset, n);
+  CUDA_CHECK_RETURN(cudaPeekAtLastError());
 }
 
-template<typename T>
-void dequantizeBlockwise(float *code, unsigned char *A, float *absmax, T *out, int blocksize, const int n) {
-    int blocks = n / blocksize;
-    blocks = n % blocksize == 0 ? blocks : blocks + 1;
-    if (blocksize == 4096)
-        kDequantizeBlockwise < T, 4096, 1024, 4 ><<<blocks, 4096 / 4>>>(code, A, absmax, out, n);
-    else if (blocksize == 2048)
-        kDequantizeBlockwise < T, 2048, 512, 4 ><<<blocks, 2048 / 4>>>(code, A, absmax, out, n);
-    CUDA_CHECK_RETURN(cudaPeekAtLastError());
+template<typename T> void dequantizeBlockwise(float *code, unsigned char *A, float *absmax, T *out, int blocksize, const int n)
+{
+  int blocks = n/blocksize;
+  blocks = n % blocksize == 0 ? blocks : blocks + 1;
+  if(blocksize == 4096)
+    kDequantizeBlockwise<T, 4096, 1024, 4><<<blocks, 4096/4>>>(code, A, absmax, out, n);
+  else if(blocksize == 2048)
+    kDequantizeBlockwise<T, 2048, 512, 4><<<blocks, 2048/4>>>(code, A, absmax, out, n);
+  CUDA_CHECK_RETURN(cudaPeekAtLastError());
 }
 
-template<typename T, int OPTIMIZER>
-void optimizer32bit(T *g, T *p,
-                    float *state1, float *state2, float *unorm, float max_unorm, float param_norm,
-                    const float beta1, const float beta2, const float eps, const float weight_decay,
-                    const int step, const float lr, const float gnorm_scale, bool skip_zeros, const int n) {
-    int blocks = n / 4096;
-    blocks = n % 4096 == 0 ? blocks : blocks + 1;
-    switch (OPTIMIZER) {
-        case ADAM:
-            if (max_unorm > 0.0f) {
-                CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1 * sizeof(float)));
-                kPreconditionOptimizer32bit2State < T, OPTIMIZER, 4096,
-                        8 ><<<blocks, 512>>>(g, p, state1, state2, unorm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, n);
-                CUDA_CHECK_RETURN(cudaPeekAtLastError());
-            }
-            kOptimizer32bit2State < T,
-                    OPTIMIZER ><<<blocks, 1024>>>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
-            CUDA_CHECK_RETURN(cudaPeekAtLastError());
-            break;
-        case MOMENTUM:
-        case RMSPROP:
-        case ADAGRAD:
+template<typename T, int OPTIMIZER> void optimizer32bit(T* g, T* p,
+                float* state1, float* state2, float *unorm, float max_unorm, float param_norm,
+                const float beta1, const float beta2, const float eps, const float weight_decay,
+                const int step, const float lr, const float gnorm_scale, bool skip_zeros, const int n)
+{
+  int blocks = n/4096;
+  blocks = n % 4096 == 0 ? blocks : blocks + 1;
+	switch(OPTIMIZER)
+	{
+		case ADAM:
+      if(max_unorm > 0.0f)
+			{
+				CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1*sizeof(float)));
+        kPreconditionOptimizer32bit2State<T, OPTIMIZER, 4096, 8><<<blocks, 512>>>(g, p, state1, state2, unorm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, n);
+        CUDA_CHECK_RETURN(cudaPeekAtLastError());
+      }
+			kOptimizer32bit2State<T, OPTIMIZER><<<blocks, 1024>>>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
+      CUDA_CHECK_RETURN(cudaPeekAtLastError());
+			break;
+		case MOMENTUM:
+    case RMSPROP:
+    case ADAGRAD:
 
-            if (max_unorm > 0.0f) {
-                CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1 * sizeof(float)));
-                kPreconditionOptimizer32bit1State < T, OPTIMIZER, 4096,
-                        8 ><<<blocks, 512>>>(g, p, state1, unorm, beta1, eps, weight_decay, step, lr, gnorm_scale, n);
-                CUDA_CHECK_RETURN(cudaPeekAtLastError());
-            }
+      if(max_unorm > 0.0f)
+			{
+				CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1*sizeof(float)));
+				kPreconditionOptimizer32bit1State<T, OPTIMIZER, 4096, 8><<<blocks, 512>>>(g, p, state1, unorm, beta1, eps, weight_decay, step, lr, gnorm_scale, n);
+        CUDA_CHECK_RETURN(cudaPeekAtLastError());
+			}
 
-            kOptimizer32bit1State < T,
-                    OPTIMIZER ><<<blocks, 1024>>>(g, p, state1, unorm, max_unorm, param_norm, beta1, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
-            CUDA_CHECK_RETURN(cudaPeekAtLastError());
-            break;
-    }
+			kOptimizer32bit1State<T, OPTIMIZER><<<blocks, 1024>>>(g, p, state1, unorm, max_unorm, param_norm, beta1, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n);
+      CUDA_CHECK_RETURN(cudaPeekAtLastError());
+			break;
+	}
 }
 
-template<typename T, int OPTIMIZER>
-void optimizerStatic8bit(T *p, T *g,
-                         unsigned char *state1, unsigned char *state2,
-                         float *unorm, float max_unorm, float param_norm,
-                         float beta1, float beta2,
-                         float eps, int step, float lr,
-                         float *quantiles1, float *quantiles2,
-                         float *max1, float *max2, float *new_max1, float *new_max2,
-                         float weight_decay,
-                         const float gnorm_scale, int n) {
-    int blocks = n / 4096;
-    blocks = n % 4096 == 0 ? blocks : blocks + 1;
+template<typename T, int OPTIMIZER> void optimizerStatic8bit(T* p, T* g,
+                unsigned char* state1, unsigned char* state2,
+                float *unorm, float max_unorm, float param_norm,
+                float beta1, float beta2,
+                float eps, int step, float lr,
+                float* quantiles1, float* quantiles2,
+                float* max1, float* max2, float* new_max1, float* new_max2,
+                float weight_decay,
+                const float gnorm_scale, int n)
+{
+  int blocks = n/4096;
+  blocks = n % 4096 == 0 ? blocks : blocks + 1;
 
-    if (max_unorm > 0.0f) { CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1 * sizeof(float))); }
+  if(max_unorm > 0.0f){ CUDA_CHECK_RETURN(cudaMemset(unorm, 0, 1*sizeof(float))); }
 
-    switch (OPTIMIZER) {
-        case ADAM:
-            CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1 * sizeof(float)));
-            CUDA_CHECK_RETURN(cudaMemset(new_max2, 0, 1 * sizeof(float)));
-            kPreconditionOptimizerStatic8bit2State < T,
-                    OPTIMIZER ><<<blocks, 256>>>(p, g, state1, state2, unorm, beta1, beta2, eps, step, quantiles1, quantiles2, max1, max2, new_max1, new_max2, gnorm_scale, n);
-            CUDA_CHECK_RETURN(cudaPeekAtLastError());
-            kOptimizerStatic8bit2State < T,
-                    OPTIMIZER ><<<blocks, 1024>>>(p, g, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr,
-            quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n);
-            CUDA_CHECK_RETURN(cudaPeekAtLastError());
-            break;
-        case MOMENTUM:
-        case RMSPROP:
-        case ADAGRAD:
-            CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1 * sizeof(float)));
-            kPreconditionOptimizerStatic8bit1State < T,
-                    OPTIMIZER ><<<blocks, 256>>>(p, g, state1, unorm, beta1, eps, step, quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
-            CUDA_CHECK_RETURN(cudaPeekAtLastError());
-            kOptimizerStatic8bit1State < T, OPTIMIZER ><<<blocks, 1024>>>(p, g, state1, unorm, max_unorm, param_norm, beta1, eps, step, lr,
-            quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
-            CUDA_CHECK_RETURN(cudaPeekAtLastError());
-            break;
-        default:
-            break;
-    }
+	switch(OPTIMIZER)
+	{
+		case ADAM:
+			CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1*sizeof(float)));
+			CUDA_CHECK_RETURN(cudaMemset(new_max2, 0, 1*sizeof(float)));
+			kPreconditionOptimizerStatic8bit2State<T, OPTIMIZER><<<blocks, 256>>>(p, g, state1, state2, unorm, beta1, beta2, eps, step, quantiles1, quantiles2, max1, max2, new_max1, new_max2, gnorm_scale, n);
+			CUDA_CHECK_RETURN(cudaPeekAtLastError());
+			kOptimizerStatic8bit2State<T, OPTIMIZER><<<blocks, 1024>>>(p, g, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr,
+																														quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n);
+			CUDA_CHECK_RETURN(cudaPeekAtLastError());
+		break;
+		case MOMENTUM:
+    case RMSPROP:
+    case ADAGRAD:
+			CUDA_CHECK_RETURN(cudaMemset(new_max1, 0, 1*sizeof(float)));
+			kPreconditionOptimizerStatic8bit1State<T, OPTIMIZER><<<blocks, 256>>>(p, g, state1, unorm, beta1, eps, step, quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
+			CUDA_CHECK_RETURN(cudaPeekAtLastError());
+			kOptimizerStatic8bit1State<T, OPTIMIZER><<<blocks, 1024>>>(p, g, state1, unorm, max_unorm, param_norm, beta1, eps, step, lr,
+																														quantiles1, max1, new_max1, weight_decay, gnorm_scale, n);
+			CUDA_CHECK_RETURN(cudaPeekAtLastError());
+			break;
+		default:
+			break;
+	}
 }
 
 #define BLOCKSIZE_2STATE 2048
@@ -150,43 +147,42 @@ void optimizerStatic8bit(T *p, T *g,
 #define BLOCKSIZE_1STATE 2048
 #define NUM_1STATE 8
 
-template<typename T, int OPTIMIZER>
-void optimizerStatic8bitBlockwise(T *p, T *g,
-                                  unsigned char *state1, unsigned char *state2, float beta1, float beta2, float eps, int step, float lr,
-                                  float *quantiles1, float *quantiles2, float *absmax1, float *absmax2, float weight_decay,
-                                  const float gnorm_scale, bool skip_zeros, int n) {
+template<typename T, int OPTIMIZER> void optimizerStatic8bitBlockwise(T* p, T* g,
+                unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr,
+                float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n)
+{
 
-    int blocks = 0;
-    switch (OPTIMIZER) {
-        case ADAM:
-            blocks = n / BLOCKSIZE_2STATE;
-            blocks = n % BLOCKSIZE_2STATE == 0 ? blocks : blocks + 1;
-            kOptimizerStatic8bit2StateBlockwise < T, OPTIMIZER, BLOCKSIZE_2STATE, NUM_2STATE ><<<blocks, BLOCKSIZE_2STATE /
-                                                                                                         NUM_2STATE>>>(p, g, state1, state2, beta1, beta2, eps, step, lr,
-            quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n);
-            CUDA_CHECK_RETURN(cudaPeekAtLastError());
-            break;
-        case MOMENTUM:
-        case RMSPROP:
-        case ADAGRAD:
-            blocks = n / BLOCKSIZE_1STATE;
-            blocks = n % BLOCKSIZE_1STATE == 0 ? blocks : blocks + 1;
-            kOptimizerStatic8bit1StateBlockwise < T, OPTIMIZER, BLOCKSIZE_1STATE, NUM_1STATE ><<<blocks, BLOCKSIZE_1STATE /
-                                                                                                         NUM_1STATE>>>(p, g, state1, beta1, beta2, eps, step, lr,
-            quantiles1, absmax1, weight_decay, gnorm_scale, skip_zeros, n);
-            CUDA_CHECK_RETURN(cudaPeekAtLastError());
-            break;
-    }
+	int blocks = 0;
+	switch(OPTIMIZER)
+	{
+		case ADAM:
+			blocks = n/BLOCKSIZE_2STATE;
+			blocks = n % BLOCKSIZE_2STATE == 0 ? blocks : blocks + 1;
+			kOptimizerStatic8bit2StateBlockwise<T, OPTIMIZER, BLOCKSIZE_2STATE, NUM_2STATE><<<blocks, BLOCKSIZE_2STATE/NUM_2STATE>>>(p, g, state1, state2, beta1, beta2, eps, step, lr,
+																														quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n);
+			CUDA_CHECK_RETURN(cudaPeekAtLastError());
+		break;
+		case MOMENTUM:
+		case RMSPROP:
+    case ADAGRAD:
+			blocks = n/BLOCKSIZE_1STATE;
+			blocks = n % BLOCKSIZE_1STATE == 0 ? blocks : blocks + 1;
+			kOptimizerStatic8bit1StateBlockwise<T, OPTIMIZER, BLOCKSIZE_1STATE, NUM_1STATE><<<blocks, BLOCKSIZE_1STATE/NUM_1STATE>>>(p, g, state1, beta1, beta2, eps, step, lr,
+																														quantiles1, absmax1, weight_decay, gnorm_scale, skip_zeros, n);
+			CUDA_CHECK_RETURN(cudaPeekAtLastError());
+		break;
+	}
 }
 
 
-template<typename T>
-void percentileClipping(T *g, float *gnorm_vec, int step, const int n) {
-    int blocks = n / 2048;
-    blocks = n % 2048 == 0 ? blocks : blocks + 1;
-    CUDA_CHECK_RETURN(cudaMemset(&gnorm_vec[step % 100], 0, 1 * sizeof(float)));
-    kPercentileClipping < T, 2048, 4 ><<<blocks, 512>>>(g, gnorm_vec, step, n);
-    CUDA_CHECK_RETURN(cudaPeekAtLastError());
+
+template<typename T> void percentileClipping(T * g, float *gnorm_vec, int step, const int n)
+{
+  int blocks = n/2048;
+  blocks = n % 2048 == 0 ? blocks : blocks + 1;
+	CUDA_CHECK_RETURN(cudaMemset(&gnorm_vec[step % 100], 0, 1*sizeof(float)));
+  kPercentileClipping<T, 2048, 4><<<blocks, 512>>>(g, gnorm_vec, step, n);
+  CUDA_CHECK_RETURN(cudaPeekAtLastError());
 }
 
 
@@ -195,23 +191,13 @@ void percentileClipping(T *g, float *gnorm_vec, int step, const int n) {
 //==============================================================
 
 template void estimateQuantiles(half *A, float *code, float offset, int n);
-
 template void estimateQuantiles(float *A, float *code, float offset, int n);
 
-template void
-quantizeBlockwise<half, 0>(float *code, half *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n);
-
-template void
-quantizeBlockwise<float, 0>(float *code, float *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n);
-
-template void
-quantizeBlockwise<half, 1>(float *code, half *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n);
-
-template void
-quantizeBlockwise<float, 1>(float *code, float *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n);
-
+template void quantizeBlockwise<half, 0>(float * code, half *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n);
+template void quantizeBlockwise<float, 0>(float * code, float *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n);
+template void quantizeBlockwise<half, 1>(float * code, half *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n);
+template void quantizeBlockwise<float, 1>(float * code, float *A, float *absmax, unsigned char *out, float* rand, int rand_offset, const int n);
 template void dequantizeBlockwise<half>(float *code, unsigned char *A, float *absmax, half *out, int blocksize, const int n);
-
 template void dequantizeBlockwise<float>(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n);
 
 #define MAKE_optimizer32bit(name, gtype) \
@@ -221,19 +207,12 @@ template void optimizer32bit<gtype, name>(gtype* g, gtype* p, \
                 const int step, const float lr, const float gnorm_scale, const bool skip_zeros, const int n);
 
 MAKE_optimizer32bit(ADAM, half)
-
 MAKE_optimizer32bit(ADAM, float)
-
 MAKE_optimizer32bit(MOMENTUM, half)
-
 MAKE_optimizer32bit(MOMENTUM, float)
-
 MAKE_optimizer32bit(RMSPROP, half)
-
 MAKE_optimizer32bit(RMSPROP, float)
-
 MAKE_optimizer32bit(ADAGRAD, half)
-
 MAKE_optimizer32bit(ADAGRAD, float)
 
 #define MAKE_optimizerStatic8bit(name, gtype) \
@@ -246,17 +225,11 @@ template void optimizerStatic8bit<gtype, name>(gtype* p, gtype* g, unsigned char
                 float weight_decay, \
                 const float gnorm_scale, int n); \
 
-
 MAKE_optimizerStatic8bit(ADAM, half)
-
 MAKE_optimizerStatic8bit(ADAM, float)
-
 MAKE_optimizerStatic8bit(MOMENTUM, half)
-
 MAKE_optimizerStatic8bit(MOMENTUM, float)
-
 MAKE_optimizerStatic8bit(RMSPROP, half)
-
 MAKE_optimizerStatic8bit(RMSPROP, float)
 
 #define MAKE_optimizerStatic8bitBlockwise(gtype, optim_name) \
@@ -264,23 +237,14 @@ template void optimizerStatic8bitBlockwise<gtype, optim_name>(gtype* p, gtype* g
                 unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr,  \
                 float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n); \
 
-
 MAKE_optimizerStatic8bitBlockwise(half, ADAM);
-
 MAKE_optimizerStatic8bitBlockwise(float, ADAM);
-
 MAKE_optimizerStatic8bitBlockwise(half, MOMENTUM);
-
 MAKE_optimizerStatic8bitBlockwise(float, MOMENTUM);
-
 MAKE_optimizerStatic8bitBlockwise(half, RMSPROP);
-
 MAKE_optimizerStatic8bitBlockwise(float, RMSPROP);
-
 MAKE_optimizerStatic8bitBlockwise(half, ADAGRAD);
-
 MAKE_optimizerStatic8bitBlockwise(float, ADAGRAD);
 
-template void percentileClipping(float *g, float *gnorm_vec, int step, const int n);
-
-template void percentileClipping(half *g, float *gnorm_vec, int step, const int n);
+template void percentileClipping(float * g, float *gnorm_vec, int step, const int n);
+template void percentileClipping(half * g, float *gnorm_vec, int step, const int n);

From 4d1d5b569f55dd613bea26714eb1ad931a10be35 Mon Sep 17 00:00:00 2001
From: Max Ryabinin <mryabinin0@gmail.com>
Date: Fri, 1 Jul 2022 17:40:02 +0300
Subject: [PATCH 4/7] Reduce diff

---
 csrc/pythonInterface.c | 107 ++++++++++++++++++++---------------------
 1 file changed, 53 insertions(+), 54 deletions(-)

diff --git a/csrc/pythonInterface.c b/csrc/pythonInterface.c
index 229b7ed..1f690c5 100644
--- a/csrc/pythonInterface.c
+++ b/csrc/pythonInterface.c
@@ -38,15 +38,15 @@ MAKE_FUNC32(adagrad, ADAGRAD, half, 16)
 
 #define MAKE_FUNC8(fname, oname, gtype, gbits) \
 void fname##_static_8bit_g##gbits(gtype* p, gtype* g, unsigned char* state1, unsigned char* state2, \
-                                float *unorm, float max_unorm, float param_norm, \
+								float *unorm, float max_unorm, float param_norm, \
                 float beta1, float beta2, \
                 float eps, int step, float lr,  \
                 float* quantiles1, float* quantiles2, \
                 float* max1, float* max2, float* new_max1, float* new_max2, \
                 float weight_decay, float gnorm_scale, int n) \
 {  \
-    optimizerStatic8bit<gtype, oname>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \
-                                              quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n); \
+	optimizerStatic8bit<gtype, oname>(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \
+			                                  quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n); \
 } \
 
 MAKE_FUNC8(adam, ADAM, float, 32)
@@ -86,37 +86,37 @@ void dequantizeBlockwise_fp32(float *code, unsigned char *A, float *absmax, floa
 
 extern "C"
 {
-#if BUILD_CUDA
-void cestimate_quantiles_fp32(float *A, float *code, float offset, int n){ estimateQuantiles_fp32(A, code, offset, n); }
-void cestimate_quantiles_fp16(half *A, float *code, float offset, int n){ estimateQuantiles_fp16(A, code, offset, n); }
-void cquantize(float *code, float *A, unsigned char *out, int n){ quantize(code, A, out, n); }
-void cdequantize(float *code, unsigned char *A, float *out, int n){ dequantize(code, A, out, n); }
-void cquantize_blockwise_fp16(float * code, half *A, float *absmax, unsigned char *out, const int n){ quantizeBlockwise_fp16(code, A, absmax, out, n); }
-void cquantize_blockwise_fp32(float * code, float *A, float *absmax, unsigned char *out, const int n){ quantizeBlockwise_fp32(code, A, absmax, out, n); }
-void cquantize_blockwise_stochastic_fp16(float * code, half *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n){ quantizeBlockwise_stochastic_fp16(code, A, absmax, out, rand, rand_offset, n); }
-void cquantize_blockwise_stochastic_fp32(float * code, float *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n){ quantizeBlockwise_stochastic_fp32(code, A, absmax, out, rand, rand_offset, n); }
+    if #BUILD_CUDA
+	void cestimate_quantiles_fp32(float *A, float *code, float offset, int n){ estimateQuantiles_fp32(A, code, offset, n); }
+	void cestimate_quantiles_fp16(half *A, float *code, float offset, int n){ estimateQuantiles_fp16(A, code, offset, n); }
+	void cquantize(float *code, float *A, unsigned char *out, int n){ quantize(code, A, out, n); }
+	void cdequantize(float *code, unsigned char *A, float *out, int n){ dequantize(code, A, out, n); }
+  void cquantize_blockwise_fp16(float * code, half *A, float *absmax, unsigned char *out, const int n){ quantizeBlockwise_fp16(code, A, absmax, out, n); }
+  void cquantize_blockwise_fp32(float * code, float *A, float *absmax, unsigned char *out, const int n){ quantizeBlockwise_fp32(code, A, absmax, out, n); }
+  void cquantize_blockwise_stochastic_fp16(float * code, half *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n){ quantizeBlockwise_stochastic_fp16(code, A, absmax, out, rand, rand_offset, n); }
+  void cquantize_blockwise_stochastic_fp32(float * code, float *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n){ quantizeBlockwise_stochastic_fp32(code, A, absmax, out, rand, rand_offset, n); }
 
-void cdequantize_blockwise_fp16(float *code, unsigned char *A, float *absmax, half *out, int blocksize, const int n){ dequantizeBlockwise_fp16(code, A, absmax, out, blocksize, n); }
-void cdequantize_blockwise_fp32(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n){ dequantizeBlockwise_fp32(code, A, absmax, out, blocksize, n); }
+  void cdequantize_blockwise_fp16(float *code, unsigned char *A, float *absmax, half *out, int blocksize, const int n){ dequantizeBlockwise_fp16(code, A, absmax, out, blocksize, n); }
+  void cdequantize_blockwise_fp32(float *code, unsigned char *A, float *absmax, float *out, int blocksize, const int n){ dequantizeBlockwise_fp32(code, A, absmax, out, blocksize, n); }
 
-#define MAKE_CFUNC32(name, gtype, gbits) \
-    void c##name##32bit_g##gbits(gtype *g, gtype *p, \
-                                 float* state1, float* state2, float *unorm, float max_unorm, float param_norm, \
-                                 const float beta1, const float beta2, const float eps, const float weight_decay, \
-                                 const int step, const float lr, const float gnorm_scale, bool skip_zeros, const int n) \
-    { name##32bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); } \
+	#define MAKE_CFUNC32(name, gtype, gbits) \
+	void c##name##32bit_g##gbits(gtype *g, gtype *p, \
+								 float* state1, float* state2, float *unorm, float max_unorm, float param_norm, \
+								 const float beta1, const float beta2, const float eps, const float weight_decay, \
+								 const int step, const float lr, const float gnorm_scale, bool skip_zeros, const int n) \
+	{ name##32bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, weight_decay, step, lr, gnorm_scale, skip_zeros, n); } \
 
-MAKE_CFUNC32(adam, float, 32)
-MAKE_CFUNC32(adam, half, 16)
-MAKE_CFUNC32(momentum, float, 32)
-MAKE_CFUNC32(momentum, half, 16)
-MAKE_CFUNC32(rmsprop, float, 32)
-MAKE_CFUNC32(rmsprop, half, 16)
-MAKE_CFUNC32(adagrad, float, 32)
-MAKE_CFUNC32(adagrad, half, 16)
+	MAKE_CFUNC32(adam, float, 32)
+	MAKE_CFUNC32(adam, half, 16)
+	MAKE_CFUNC32(momentum, float, 32)
+	MAKE_CFUNC32(momentum, half, 16)
+	MAKE_CFUNC32(rmsprop, float, 32)
+	MAKE_CFUNC32(rmsprop, half, 16)
+	MAKE_CFUNC32(adagrad, float, 32)
+	MAKE_CFUNC32(adagrad, half, 16)
 
-#define MAKE_CFUNC8(name, gtype, gbits) \
-    void c##name##_static_8bit_g##gbits(gtype* p, gtype* g, unsigned char* state1, unsigned char* state2, \
+	#define MAKE_CFUNC8(name, gtype, gbits) \
+	void c##name##_static_8bit_g##gbits(gtype* p, gtype* g, unsigned char* state1, unsigned char* state2, \
                 float *unorm, float max_unorm, float param_norm, \
                 float beta1, float beta2, \
                 float eps, int step, float lr,  \
@@ -124,40 +124,39 @@ MAKE_CFUNC32(adagrad, half, 16)
                 float* max1, float* max2, float* new_max1, float* new_max2, \
                 float weight_decay, float gnorm_scale, int n) \
   {  \
-        name##_static_8bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \
-                                             quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n); \
+	    name##_static_8bit_g##gbits(g, p, state1, state2, unorm, max_unorm, param_norm, beta1, beta2, eps, step, lr, \
+			                                 quantiles1, quantiles2, max1, max2, new_max1, new_max2, weight_decay, gnorm_scale, n); \
   } \
 
-MAKE_CFUNC8(adam, float, 32)
-MAKE_CFUNC8(adam, half, 16)
-MAKE_CFUNC8(momentum, float, 32)
-MAKE_CFUNC8(momentum, half, 16)
-MAKE_CFUNC8(rmsprop, float, 32)
-MAKE_CFUNC8(rmsprop, half, 16)
+	MAKE_CFUNC8(adam, float, 32)
+	MAKE_CFUNC8(adam, half, 16)
+	MAKE_CFUNC8(momentum, float, 32)
+	MAKE_CFUNC8(momentum, half, 16)
+	MAKE_CFUNC8(rmsprop, float, 32)
+	MAKE_CFUNC8(rmsprop, half, 16)
 
-#define MAKE_CBLOCKWISE8(fname, optim_name, gtype, gbits) \
+  #define MAKE_CBLOCKWISE8(fname, optim_name, gtype, gbits) \
   void c##fname##_8bit_blockwise_fp##gbits(gtype* p, gtype* g, \
                 unsigned char* state1, unsigned char* state2, float beta1, float beta2, float eps, int step, float lr,  \
                 float* quantiles1, float* quantiles2, float* absmax1, float* absmax2, float weight_decay, const float gnorm_scale, bool skip_zeros, int n) \
   {	fname##_8bit_blockwise_fp##gbits(p, g, state1, state2, beta1, beta2, eps, step, lr, quantiles1, quantiles2, absmax1, absmax2, weight_decay, gnorm_scale, skip_zeros, n); } \
 
-MAKE_CBLOCKWISE8(adam, ADAM, half, 16)
-MAKE_CBLOCKWISE8(adam, ADAM, float, 32)
-MAKE_CBLOCKWISE8(momentum, MOMENTUM, half, 16)
-MAKE_CBLOCKWISE8(momentum, MOMENTUM, float, 32)
-MAKE_CBLOCKWISE8(rmsprop, RMSPROP, half, 16)
-MAKE_CBLOCKWISE8(rmsprop, RMSPROP, float, 32)
-MAKE_CBLOCKWISE8(adagrad, ADAGRAD, half, 16)
-MAKE_CBLOCKWISE8(adagrad, ADAGRAD, float, 32)
+	MAKE_CBLOCKWISE8(adam, ADAM, half, 16)
+	MAKE_CBLOCKWISE8(adam, ADAM, float, 32)
+	MAKE_CBLOCKWISE8(momentum, MOMENTUM, half, 16)
+	MAKE_CBLOCKWISE8(momentum, MOMENTUM, float, 32)
+	MAKE_CBLOCKWISE8(rmsprop, RMSPROP, half, 16)
+	MAKE_CBLOCKWISE8(rmsprop, RMSPROP, float, 32)
+	MAKE_CBLOCKWISE8(adagrad, ADAGRAD, half, 16)
+	MAKE_CBLOCKWISE8(adagrad, ADAGRAD, float, 32)
 
 
-void cpercentile_clipping_g32(float * g, float *gnorm_vec, int step, const int n){ percentileClipping_g32(g, gnorm_vec, step, n); }
-void cpercentile_clipping_g16(half * g, float *gnorm_vec, int step, const int n){ percentileClipping_g16(g, gnorm_vec, step, n); }
-void chistogram_scatter_add_2d(float* histogram, int *index1, int *index2, float *src, int maxidx1, int n){ histogramScatterAdd2D(histogram, index1, index2, src, maxidx1, n); }
-#endif
+	void cpercentile_clipping_g32(float * g, float *gnorm_vec, int step, const int n){ percentileClipping_g32(g, gnorm_vec, step, n); }
+	void cpercentile_clipping_g16(half * g, float *gnorm_vec, int step, const int n){ percentileClipping_g16(g, gnorm_vec, step, n); }
+	void chistogram_scatter_add_2d(float* histogram, int *index1, int *index2, float *src, int maxidx1, int n){ histogramScatterAdd2D(histogram, index1, index2, src, maxidx1, n); }
 
-void cquantize_blockwise_cpu_fp32(float *code, float *A, float *absmax, unsigned char *out, const int n){ quantize_cpu(code, A, absmax, out, n); }
-void cdequantize_blockwise_cpu_fp32(float *code, unsigned char *A, float *absmax, float *out, const int n){ dequantize_cpu(code, A, absmax, out, n); }
+    #endif
+	void cquantize_blockwise_cpu_fp32(float *code, float *A, float *absmax, unsigned char *out, const int n){ quantize_cpu(code, A, absmax, out, n); }
+	void cdequantize_blockwise_cpu_fp32(float *code, unsigned char *A, float *absmax, float *out, const int n){ dequantize_cpu(code, A, absmax, out, n); }
 }
 
-

From 575aa698fa53df2f5c584413aed7bf7714f86039 Mon Sep 17 00:00:00 2001
From: Max Ryabinin <mryabinin0@gmail.com>
Date: Fri, 1 Jul 2022 17:41:48 +0300
Subject: [PATCH 5/7] Reduce diff

---
 csrc/ops.cu            | 45 +++++++++++++++++++-----------------------
 csrc/pythonInterface.c |  2 +-
 2 files changed, 21 insertions(+), 26 deletions(-)

diff --git a/csrc/ops.cu b/csrc/ops.cu
index b2a1105..dbb50be 100644
--- a/csrc/ops.cu
+++ b/csrc/ops.cu
@@ -15,35 +15,30 @@ using namespace BinSearch;
 using std::cout;
 using std::endl;
 
-void histogramScatterAdd2D(float *histogram, int *index1, int *index2, float *src, int maxidx1, int n) {
-    int threads = 512;
-    int blocks = n / threads;
-    blocks = n % threads == 0 ? blocks : blocks + 1;
-    kHistogramScatterAdd2D<<<blocks, 512>>>(histogram, index1, index2, src, maxidx1, n);
-    CUDA_CHECK_RETURN(cudaPeekAtLastError());
+void histogramScatterAdd2D(float* histogram, int *index1, int *index2, float *src, int maxidx1, int n)
+{
+  int threads = 512;
+  int blocks = n/threads;
+  blocks = n % threads == 0 ? blocks : blocks + 1;
+  kHistogramScatterAdd2D<<<blocks, 512>>>(histogram, index1, index2, src, maxidx1, n);
+  CUDA_CHECK_RETURN(cudaPeekAtLastError());
 }
 
-template<typename T>
-void estimateQuantiles(T *A, float *code, float offset, int n) {
-    int blocks = n / 4096;
-    blocks = n % 4096 == 0 ? blocks : blocks + 1;
-    CUDA_CHECK_RETURN(cudaMemset(code, 0, 256 * sizeof(float)));
-    kEstimateQuantiles < T ><<<blocks, 512>>>(A, code, offset, std::numeric_limits<T>::max(), n);
-    CUDA_CHECK_RETURN(cudaPeekAtLastError());
+template <typename T> void estimateQuantiles(T *A, float *code, float offset, int n)
+{
+  int blocks = n/4096;
+  blocks = n % 4096 == 0 ? blocks : blocks + 1;
+	CUDA_CHECK_RETURN(cudaMemset(code, 0, 256*sizeof(float)));
+  kEstimateQuantiles<T><<<blocks, 512>>>(A, code, offset, std::numeric_limits<T>::max(), n);
+  CUDA_CHECK_RETURN(cudaPeekAtLastError());
 }
 
-void quantize(float *code, float *A, unsigned char *out, int n) {
-    int blocks = n / 1024;
-    blocks = n % 1024 == 0 ? blocks : blocks + 1;
-    kQuantize<<<blocks, 1024>>>(code, A, out, n);
-    CUDA_CHECK_RETURN(cudaPeekAtLastError());
-}
-
-void dequantize(float *code, unsigned char *A, float *out, int n) {
-    int blocks = n / 1024;
-    blocks = n % 1024 == 0 ? blocks : blocks + 1;
-    kDequantize<<<blocks, 1024>>>(code, A, out, n);
-    CUDA_CHECK_RETURN(cudaPeekAtLastError());
+void quantize(float *code, float *A, unsigned char *out, int n)
+{
+  int blocks = n/1024;
+  blocks = n % 1024 == 0 ? blocks : blocks + 1;
+  kQuantize<<<blocks, 1024>>>(code, A, out, n);
+  CUDA_CHECK_RETURN(cudaPeekAtLastError());
 }
 
 template <typename T, int STOCHASTIC> void quantizeBlockwise(float * code, T *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n)
diff --git a/csrc/pythonInterface.c b/csrc/pythonInterface.c
index 1f690c5..c2fed6b 100644
--- a/csrc/pythonInterface.c
+++ b/csrc/pythonInterface.c
@@ -86,7 +86,7 @@ void dequantizeBlockwise_fp32(float *code, unsigned char *A, float *absmax, floa
 
 extern "C"
 {
-    if #BUILD_CUDA
+    #if BUILD_CUDA
 	void cestimate_quantiles_fp32(float *A, float *code, float offset, int n){ estimateQuantiles_fp32(A, code, offset, n); }
 	void cestimate_quantiles_fp16(half *A, float *code, float offset, int n){ estimateQuantiles_fp16(A, code, offset, n); }
 	void cquantize(float *code, float *A, unsigned char *out, int n){ quantize(code, A, out, n); }

From 025824d29b38f6b981bbcea8a61bc23e7f2b3e02 Mon Sep 17 00:00:00 2001
From: Max Ryabinin <mryabinin0@gmail.com>
Date: Fri, 1 Jul 2022 17:42:58 +0300
Subject: [PATCH 6/7] Reduce diff

---
 csrc/ops.cu | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/csrc/ops.cu b/csrc/ops.cu
index dbb50be..40c185c 100644
--- a/csrc/ops.cu
+++ b/csrc/ops.cu
@@ -41,6 +41,14 @@ void quantize(float *code, float *A, unsigned char *out, int n)
   CUDA_CHECK_RETURN(cudaPeekAtLastError());
 }
 
+void dequantize(float *code, unsigned char *A, float *out, int n)
+{
+  int blocks = n/1024;
+  blocks = n % 1024 == 0 ? blocks : blocks + 1;
+  kDequantize<<<blocks, 1024>>>(code, A, out, n);
+  CUDA_CHECK_RETURN(cudaPeekAtLastError());
+}
+
 template <typename T, int STOCHASTIC> void quantizeBlockwise(float * code, T *A, float *absmax, unsigned char *out, float *rand, int rand_offset, const int n)
 {
   int blocks = n/4096;

From fd750cd2370b3b12e216a9148b23aaae63a80989 Mon Sep 17 00:00:00 2001
From: Max Ryabinin <mryabinin0@gmail.com>
Date: Fri, 1 Jul 2022 17:46:29 +0300
Subject: [PATCH 7/7] Update README.md

---
 compile_from_source.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/compile_from_source.md b/compile_from_source.md
index c6a8b18..71b0c09 100644
--- a/compile_from_source.md
+++ b/compile_from_source.md
@@ -1,12 +1,12 @@
 # Compiling from source
 
 Basic steps.
-1. `make cudaXXX` where `cudaXXX` is among `cuda92, cuda10x, cuda110, cuda11x`
+1. `make [target]` where `[target]` is among `cuda92, cuda10x, cuda110, cuda11x, cpuonly`
 2. `CUDA_VERSION=XXX python setup.py install`
 
 To run these steps you will need to have the nvcc compiler installed that comes with a CUDA installation. If you use anaconda (recommended) then you can figure out which version of CUDA you are using with PyTorch via the command `conda list | grep cudatoolkit`. Then you can install the nvcc compiler by downloading and installing the same CUDA version from the [CUDA toolkit archive](https://developer.nvidia.com/cuda-toolkit-archive). 
 
-For your convenience, there is a install script int he root directory that installs CUDA 11.1 locally and configures it automatically. After installing you should add the `bin` sub-directory to the `$PATH` variable to make the compiler visible to your system. To do this you can add this to your `.bashrc` by executing these commands:
+For your convenience, there is an installation script in the root directory that installs CUDA 11.1 locally and configures it automatically. After installing you should add the `bin` sub-directory to the `$PATH` variable to make the compiler visible to your system. To do this you can add this to your `.bashrc` by executing these commands:
 ```bash
 echo "export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64/" >> ~/.bashrc
 echo "export PATH=$PATH:/usr/local/cuda/bin/" >> ~/.bashrc