Merge pull request #76 from tomaarsen/cleanup

Cleanup involving a handful of failures, some optimization and a lot of code quality improvements

Author: TimDettmers

CHANGELOG.md

@@ -49,7 +49,7 @@ Features:
 Bug fixes:
  - Fixed a bug where weight decay was incorrectly applied to 32-bit Adam. #13
  - Fixed an unsafe use of eval. #8
- - Fixed a bug where the StableEmbedding layer 32-bit optimizer override would not work without registering the whole model first (`bnb.optim.GlobalOptimManager.get_instance().register_parameters(model.parameters())`).  #13 #15 
+ - Fixed a bug where the StableEmbedding layer 32-bit optimizer override would not work without registering the whole model first (`bnb.optim.GlobalOptimManager.get_instance().register_parameters(model.parameters())`).  #13 #15
 
 Docs:
  - Added instructions how to solve "\_\_fatbinwrap_" errors.

CONTRIBUTING.md

@@ -28,4 +28,4 @@ outlined on that page and do not file a public issue.
 
 ## License
 By contributing to bitsandbytes, you agree that your contributions will be licensed
-under the LICENSE file in the root directory of this source tree.
+under the LICENSE file in the root directory of this source tree.

Makefile

@@ -26,14 +26,14 @@ INCLUDE :=  -I $(CUDA_HOME)/include -I $(ROOT_DIR)/csrc -I $(CONDA_PREFIX)/inclu
 LIB := -L $(CUDA_HOME)/lib64 -lcudart -lcublas -lcublasLt -lcurand -lcusparse -L $(CONDA_PREFIX)/lib
 
 # NVIDIA NVCC compilation flags
-COMPUTE_CAPABILITY := -gencode arch=compute_35,code=sm_35 # Kepler 
-COMPUTE_CAPABILITY += -gencode arch=compute_37,code=sm_37 # Kepler 
+COMPUTE_CAPABILITY := -gencode arch=compute_35,code=sm_35 # Kepler
+COMPUTE_CAPABILITY += -gencode arch=compute_37,code=sm_37 # Kepler
 COMPUTE_CAPABILITY += -gencode arch=compute_50,code=sm_50 # Maxwell
 COMPUTE_CAPABILITY += -gencode arch=compute_52,code=sm_52 # Maxwell
 COMPUTE_CAPABILITY += -gencode arch=compute_60,code=sm_60 # Pascal
 COMPUTE_CAPABILITY += -gencode arch=compute_61,code=sm_61 # Pascal
 COMPUTE_CAPABILITY += -gencode arch=compute_70,code=sm_70 # Volta
-COMPUTE_CAPABILITY += -gencode arch=compute_72,code=sm_72 # Volta 
+COMPUTE_CAPABILITY += -gencode arch=compute_72,code=sm_72 # Volta
 
 # CUDA 9.2 supports CC 3.0, but CUDA >= 11.0 does not
 CC_CUDA92 := -gencode arch=compute_30,code=sm_30
@@ -58,38 +58,38 @@ CC_cublasLt111 += -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 
+	$(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 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes_cuda$(CUDA_VERSION).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) -D NO_CUBLASLT
-	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA92) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o 
+	$(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 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes_cuda$(CUDA_VERSION)_nocublaslt.so $(LIB)
 
 cuda10x_nomatmul: $(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) -D NO_CUBLASLT
-	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA10x) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o 
+	$(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 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes_cuda$(CUDA_VERSION)_nocublaslt.so $(LIB)
 
 cuda110_nomatmul: $(BUILD_DIR) env
 	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA110) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) -D NO_CUBLASLT
-	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA110) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o 
+	$(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 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes_cuda$(CUDA_VERSION)_nocublaslt.so $(LIB)
 
 cuda11x_nomatmul: $(BUILD_DIR) env
 	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA11x) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) -D NO_CUBLASLT
-	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA11x) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o 
+	$(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 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes_cuda$(CUDA_VERSION)_nocublaslt.so $(LIB)
 
 cuda110: $(BUILD_DIR) env
 	$(NVCC) $(CC_cublasLt110) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR)
-	$(NVCC) $(CC_cublasLt110) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o 
+	$(NVCC) $(CC_cublasLt110) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o
 	$(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes_cuda$(CUDA_VERSION).so $(LIB)
 
 cuda11x: $(BUILD_DIR) env
 	$(NVCC) $(CC_cublasLt111) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR)
-	$(NVCC) $(CC_cublasLt111) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o 
+	$(NVCC) $(CC_cublasLt111) -Xcompiler '-fPIC' -dlink $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o -o $(BUILD_DIR)/link.o
 	$(GPP) -std=c++14 -DBUILD_CUDA -shared -fPIC $(INCLUDE) $(BUILD_DIR)/ops.o $(BUILD_DIR)/kernels.o $(BUILD_DIR)/link.o $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes_cuda$(CUDA_VERSION).so $(LIB)
 
 cpuonly: $(BUILD_DIR) env
@@ -117,7 +117,7 @@ $(ROOT_DIR)/dependencies/cub:
 	cd dependencies/cub; git checkout 1.11.0
 
 clean:
-	rm build/* 
+	rm build/*
 
 cleaneggs:
 	rm -rf *.egg*

README.md

@@ -1,6 +1,6 @@
 # bitsandbytes
 
-The bitsandbytes is a lightweight wrapper around CUDA custom functions, in particular 8-bit optimizers, matrix multiplication (LLM.int8()), and quantization functions. 
+The bitsandbytes is a lightweight wrapper around CUDA custom functions, in particular 8-bit optimizers, matrix multiplication (LLM.int8()), and quantization functions.
 
 
 
@@ -48,7 +48,7 @@ out = linear(x.to(torch.float16))
 
 Requirements: anaconda, cudatoolkit, pytorch
 
-Hardware requirements: 
+Hardware requirements:
  - LLM.int8(): NVIDIA Turing (RTX 20xx; T4) or Ampere GPU (RTX 30xx; A4-A100); (a GPU from 2018 or older).
  - 8-bit optimizers and quantization: NVIDIA Maxwell GPU or newer (>=GTX 9XX).
 
@@ -87,7 +87,7 @@ Note that by default all parameter tensors with less than 4096 elements are kept
 ```
 # parameter tensors with less than 16384 values are optimized in 32-bit
 # it is recommended to use multiplies of 4096
-adam = bnb.optim.Adam8bit(model.parameters(), min_8bit_size=16384) 
+adam = bnb.optim.Adam8bit(model.parameters(), min_8bit_size=16384)
 ```
 
 ### Change Bits and other Hyperparameters for Individual Parameters

bitsandbytes/__init__.py

@@ -3,6 +3,7 @@
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
 
+from . import cuda_setup, utils
 from .autograd._functions import (
     MatmulLtState,
     bmm_cublas,
@@ -12,7 +13,6 @@
 )
 from .cextension import COMPILED_WITH_CUDA
 from .nn import modules
-from . import cuda_setup, utils
 
 if COMPILED_WITH_CUDA:
     from .optim import adam

bitsandbytes/__main__.py

@@ -1,6 +1,3 @@
-# from bitsandbytes.debug_cli import cli
-
-# cli()
 import os
 import sys
 from warnings import warn
@@ -31,8 +28,8 @@ def print_debug_info() -> None:
 
 
 from . import COMPILED_WITH_CUDA, PACKAGE_GITHUB_URL
-from .cuda_setup.main import get_compute_capabilities, get_cuda_lib_handle
 from .cuda_setup.env_vars import to_be_ignored
+from .cuda_setup.main import get_compute_capabilities, get_cuda_lib_handle
 
 print_header("POTENTIALLY LIBRARY-PATH-LIKE ENV VARS")
 for k, v in os.environ.items():

bitsandbytes/autograd/_functions.py

@@ -1,11 +1,12 @@
 import operator
 import warnings
+from dataclasses import dataclass
+from functools import reduce  # Required in Python 3
 
 import torch
+
 import bitsandbytes.functional as F
 
-from dataclasses import dataclass
-from functools import reduce  # Required in Python 3
 
 # math.prod not compatible with python < 3.8
 def prod(iterable):
@@ -15,10 +16,10 @@ def prod(iterable):
 
 """
     This class pools outlier dimensions across layers.
-    This is particularly important for small models where outlier features 
+    This is particularly important for small models where outlier features
     are less systematic and occur with low frequency.
 """
-class GlobalOutlierPooler(object):
+class GlobalOutlierPooler:
     _instance = None
 
     def __init__(self):
@@ -49,8 +50,9 @@ def get_current_outlier_idx(self):
 
 class MatMul8bit(torch.autograd.Function):
     @staticmethod
-    def forward(ctx, A, B, out=None, quant_type="vector", precision=[8, 8, 8]):
-
+    def forward(ctx, A, B, out=None, quant_type="vector", precision=None):
+        if precision is None:
+            precision = [8, 8, 8]
         if precision[0] != 8:
             with torch.no_grad():
                 output = torch.matmul(A, B)

bitsandbytes/cextension.py

@@ -1,11 +1,11 @@
 import ctypes as ct
-import torch
-
 from pathlib import Path
 from warnings import warn
 
+import torch
+
 
-class CUDASetup(object):
+class CUDASetup:
     _instance = None
 
     def __init__(self):
@@ -122,7 +122,7 @@ def get_instance(cls):
         CUDASetup.get_instance().generate_instructions()
         CUDASetup.get_instance().print_log_stack()
         raise RuntimeError('''
-        CUDA Setup failed despite GPU being available. Inspect the CUDA SETUP outputs aboveto fix your environment!
+        CUDA Setup failed despite GPU being available. Inspect the CUDA SETUP outputs above to fix your environment!
         If you cannot find any issues and suspect a bug, please open an issue with detals about your environment:
         https://github.com/TimDettmers/bitsandbytes/issues''')
     lib.cadam32bit_g32

bitsandbytes/cuda_setup/__init__.py

@@ -1,2 +1,6 @@
-from .paths import CUDA_RUNTIME_LIB, extract_candidate_paths, determine_cuda_runtime_lib_path
 from .main import evaluate_cuda_setup
+from .paths import (
+    CUDA_RUNTIME_LIB,
+    determine_cuda_runtime_lib_path,
+    extract_candidate_paths,
+)

bitsandbytes/cuda_setup/main.py

@@ -17,11 +17,13 @@
 """
 
 import ctypes
+
 import torch
 
-from .paths import determine_cuda_runtime_lib_path
 from bitsandbytes.cextension import CUDASetup
 
+from .paths import determine_cuda_runtime_lib_path
+
 
 def check_cuda_result(cuda, result_val):
     # 3. Check for CUDA errors
@@ -48,7 +50,7 @@ def get_cuda_version(cuda, cudart_path):
     minor = (version-(major*1000))//10
 
     if major < 11:
-       CUDASetup.get_instance().add_log_entry('CUDA SETUP: CUDA version lower than 11 are currenlty not supported for LLM.int8(). You will be only to use 8-bit optimizers and quantization routines!!')
+        CUDASetup.get_instance().add_log_entry('CUDA SETUP: CUDA version lower than 11 are currently not supported for LLM.int8(). You will be only to use 8-bit optimizers and quantization routines!!')
 
     return f'{major}{minor}'
 
@@ -129,7 +131,7 @@ def evaluate_cuda_setup():
         failure = True
         cuda_setup.add_log_entry("WARNING: No libcudart.so found! Install CUDA or the cudatoolkit package (anaconda)!", is_warning=True)
     else:
-        cuda_setup.add_log_entry((f"CUDA SETUP: CUDA runtime path found: {cudart_path}"))
+        cuda_setup.add_log_entry(f"CUDA SETUP: CUDA runtime path found: {cudart_path}")
 
     if cc == '' or cc is None:
         failure = True