Added CUDA 12.0 support; removed CC 3.0 support.

CHANGELOG.md

@@ -139,7 +139,6 @@ Features:
 Bug fixes:
  - Fixed a problem where warning messages would be displayed even though everything worked correctly.
 
-
 ### 0.35.2
 
 Bug fixes:
@@ -155,3 +154,38 @@ Bug fixes:
 Bug fixes:
  - Fixed a bug in the CUDA Setup failed with the cuda runtime was found, but not the cuda library.
  - Fixed a bug where not finding the cuda runtime led to an incomprehensible error.
+
+
+### 0.36.0
+
+#### Improvements, Ada/Hopper support, fake k-bit quantization.
+
+Features:
+ - CUDA 11.8 and 12.0 support added
+ - support for Ada and Hopper GPUs added (compute capability 8.9 and 9.0)
+ - support for fake k-bit block-wise quantization for Int, Float, quantile quantization, and dynamic exponent data types added
+ - Added CUDA instruction generator to fix some installations.
+ - Added additional block sizes for quantization {64, 128, 256, 512, 1024}
+ - Added SRAM Quantile algorithm to quickly estimate less than 256 quantiles
+ - Added option to suppress the bitsandbytes welcome message (@Cyberes)
+
+Regression:
+ - Compute capability 3.0 removed: GTX 600s and 700s series is no longer supported (except GTX 780 and GTX 780 Ti)
+
+Bug fixes:
+ - fixed a bug where too long directory names would crash the CUDA SETUP #35 (@tomaarsen)
+ - fixed a bug where CPU installations on Colab would run into an error  #34 (@tomaarsen)
+ - fixed an issue where the default CUDA version with fast-DreamBooth was not supported #52
+ - fixed a bug where the CUDA setup failed due to a wrong function call.
+ - fixed a bug in the CUDA Setup which led to an incomprehensible error if no GPU was detected.
+ - fixed a bug in the CUDA Setup failed with the cuda runtime was found, but not the cuda library.
+ - fixed a bug where not finding the cuda runtime led to an incomprehensible error.
+ - fixed a bug where with missing CUDA the default was an error instead of the loading the CPU library
+ - fixed a bug where the CC version of the GPU was not detected appropriately (@BlackHC)
+ - fixed a bug in CPU quantization which lead to errors when the input buffer exceeded 2^31 elements
+
+Improvements:
+ - multiple improvements in formatting, removal of unused imports, and slight performance improvements (@tomaarsen)
+ - StableEmbedding layer now has device and dtype parameters to make it 1:1 replaceable with regular Embedding layers (@lostmsu)
+ - runtime performance of block-wise quantization slightly improved
+ - added error message for the case multiple libcudart.so are installed and bitsandbytes picks the wrong one

Makefile

@@ -22,12 +22,11 @@ BUILD_DIR:= $(ROOT_DIR)/build
 FILES_CUDA := $(CSRC)/ops.cu $(CSRC)/kernels.cu
 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
+INCLUDE :=  -I $(CUDA_HOME)/include -I $(ROOT_DIR)/csrc -I $(CONDA_PREFIX)/include -I $(ROOT_DIR)/include
+INCLUDE_10x :=  -I $(CUDA_HOME)/include -I $(ROOT_DIR)/csrc -I $(ROOT_DIR)/dependencies/cub -I $(ROOT_DIR)/include
 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_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
@@ -35,11 +34,10 @@ 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
 
-# CUDA 9.2 supports CC 3.0, but CUDA >= 11.0 does not
-CC_CUDA92 := -gencode arch=compute_30,code=sm_30
+CC_KEPLER := -gencode arch=compute_35,code=sm_35 # Kepler
+CC_KEPLER += -gencode arch=compute_37,code=sm_37 # Kepler
 
 # Later versions of CUDA support the new architectures
-CC_CUDA10x := -gencode arch=compute_30,code=sm_30
 CC_CUDA10x += -gencode arch=compute_75,code=sm_75
 
 CC_CUDA110 := -gencode arch=compute_75,code=sm_75
@@ -49,6 +47,7 @@ CC_CUDA11x := -gencode arch=compute_75,code=sm_75
 CC_CUDA11x += -gencode arch=compute_80,code=sm_80
 CC_CUDA11x += -gencode arch=compute_86,code=sm_86
 
+
 CC_cublasLt110 := -gencode arch=compute_75,code=sm_75
 CC_cublasLt110 += -gencode arch=compute_80,code=sm_80
 
@@ -56,30 +55,38 @@ CC_cublasLt111 := -gencode arch=compute_75,code=sm_75
 CC_cublasLt111 += -gencode arch=compute_80,code=sm_80
 CC_cublasLt111 += -gencode arch=compute_86,code=sm_86
 
+CC_ADA_HOPPER := -gencode arch=compute_89,code=sm_89
+CC_ADA_HOPPER += -gencode arch=compute_90,code=sm_90
+
 
 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) $(CC_KEPLER) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR)
+	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_KEPLER) -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) $(CC_KEPLER) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) -D NO_CUBLASLT
+	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA92) $(CC_KEPLER) -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) $(CC_KEPLER) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE_10x) $(LIB) --output-directory $(BUILD_DIR) -D NO_CUBLASLT
+	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA10x) $(CC_KEPLER) -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) $(CC_KEPLER) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) -D NO_CUBLASLT
+	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA110) $(CC_KEPLER) -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) $(CC_KEPLER) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) -D NO_CUBLASLT
+	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA11x) $(CC_KEPLER) -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)
+
+cuda12x_nomatmul: $(BUILD_DIR) env
+	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA11x) $(CC_ADA_HOPPER) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR) -D NO_CUBLASLT
+	$(NVCC) $(COMPUTE_CAPABILITY) $(CC_CUDA11x) $(CC_ADA_HOPPER) -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
@@ -92,6 +99,11 @@ cuda11x: $(BUILD_DIR) env
 	$(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)
 
+cuda12x: $(BUILD_DIR) env
+	$(NVCC) $(CC_cublasLt111) $(CC_ADA_HOPPER) -Xcompiler '-fPIC' --use_fast_math -Xptxas=-v -dc $(FILES_CUDA) $(INCLUDE) $(LIB) --output-directory $(BUILD_DIR)
+	$(NVCC) $(CC_cublasLt111) $(CC_ADA_HOPPER) -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
 	$(GPP) -std=c++14 -shared -fPIC -I $(ROOT_DIR)/csrc -I $(ROOT_DIR)/include $(FILES_CPP) -o ./bitsandbytes/libbitsandbytes_cpu.so
 

README.md

@@ -50,9 +50,9 @@ Requirements: anaconda, cudatoolkit, pytorch
 
 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).
+ - 8-bit optimizers and quantization: NVIDIA Kepler GPU or newer (>=GTX 78X).
 
-Supported CUDA versions: 10.2 - 11.7
+Supported CUDA versions: 10.2 - 12.0
 
 The bitsandbytes library is currently only supported on Linux distributions. Windows is not supported at the moment.
 

cuda_install.sh

@@ -11,6 +11,7 @@ URL115=https://developer.download.nvidia.com/compute/cuda/11.5.2/local_installer
 URL116=https://developer.download.nvidia.com/compute/cuda/11.6.2/local_installers/cuda_11.6.2_510.47.03_linux.run
 URL117=https://developer.download.nvidia.com/compute/cuda/11.7.0/local_installers/cuda_11.7.0_515.43.04_linux.run
 URL118=https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda_11.8.0_520.61.05_linux.run
+URL120=https://developer.download.nvidia.com/compute/cuda/12.0.0/local_installers/cuda_12.0.0_525.60.13_linux.run
 
 
 CUDA_VERSION=$1
@@ -56,6 +57,9 @@ if [[ -n "$CUDA_VERSION" ]]; then
   elif [[ "$CUDA_VERSION" -eq "118" ]]; then
     URL=$URL118
     FOLDER=cuda-11.8
+  elif [[ "$CUDA_VERSION" -eq "120" ]]; then
+    URL=$URL120
+    FOLDER=cuda-12.0
   else
     echo "argument error: No cuda version passed as input. Choose among: {111, 115}"
   fi

deploy.sh

@@ -110,7 +110,7 @@ fi
 
 make clean
 export CUDA_HOME=$BASE_PATH/cuda-11.8
-make cuda11x CUDA_VERSION=118
+make cuda12x CUDA_VERSION=118
 
 if [ ! -f "./bitsandbytes/libbitsandbytes_cuda118.so" ]; then
   # Control will enter here if $DIRECTORY doesn't exist.
@@ -118,6 +118,16 @@ if [ ! -f "./bitsandbytes/libbitsandbytes_cuda118.so" ]; then
   exit 64
 fi
 
+make clean
+export CUDA_HOME=$BASE_PATH/cuda-12.0
+make cuda12x CUDA_VERSION=120
+
+if [ ! -f "./bitsandbytes/libbitsandbytes_cuda120.so" ]; then
+  # Control will enter here if $DIRECTORY doesn't exist.
+  echo "Compilation unsuccessul!" 1>&2
+  exit 64
+fi
+
 
 make clean
 export CUDA_HOME=$BASE_PATH/cuda-10.2
@@ -213,7 +223,7 @@ fi
 
 make clean
 export CUDA_HOME=$BASE_PATH/cuda-11.8
-make cuda11x_nomatmul CUDA_VERSION=118
+make cuda12x_nomatmul CUDA_VERSION=118
 
 if [ ! -f "./bitsandbytes/libbitsandbytes_cuda118_nocublaslt.so" ]; then
   # Control will enter here if $DIRECTORY doesn't exist.
@@ -221,5 +231,15 @@ if [ ! -f "./bitsandbytes/libbitsandbytes_cuda118_nocublaslt.so" ]; then
   exit 64
 fi
 
+make clean
+export CUDA_HOME=$BASE_PATH/cuda-12.0
+make cuda12x_nomatmul CUDA_VERSION=120
+
+if [ ! -f "./bitsandbytes/libbitsandbytes_cuda120_nocublaslt.so" ]; then
+  # Control will enter here if $DIRECTORY doesn't exist.
+  echo "Compilation unsuccessul!" 1>&2
+  exit 64
+fi
+
 python -m build
 python -m twine upload dist/* --verbose

setup.py

@@ -18,7 +18,7 @@ def read(fname):
 
 setup(
     name=f"bitsandbytes",
-    version=f"0.35.4",
+    version=f"0.36.0",
     author="Tim Dettmers",
     author_email="dettmers@cs.washington.edu",
     description="8-bit optimizers and matrix multiplication routines.",