Added H100 support for CUDA 11.8 precompiled binaries.

2023-07-13 21:16:23 +07:00 · 2023-07-13 21:16:23 +07:00 · 7b6cfe1738
parent 817bdf6325
commit 7b6cfe1738
3 changed files with 85 additions and 69 deletions
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@ -258,3 +258,9 @@ Bug fixes:

 Deprecated:
 - Binaries for CUDA 11.2, 11.6 no longer ship with `pip install bitsandbytes` and need to be compiled from source.
+
+
+### 0.40.1
+
+Features:
+ - Added precompiled CUDA 11.8 binaries to support H100 GPUs without compilation #571
--- a/10
+++ b/10
@ -80,6 +80,11 @@ cuda11x_nomatmul: $(BUILD_DIR) env
 	$(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)

+cuda118_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)
+
 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
@ -95,6 +100,11 @@ 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)

+cuda118: $(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)
+
 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 $(BUILD_DIR)/link.o
--- a/deploy.sh
+++ b/deploy.sh
@ -10,77 +10,77 @@ if [[ ! -z "${LD_LIBRARY_PATH}" ]]; then
 fi


-#module unload cuda && echo "no module function available. Probably not on a slurm cluster."
-#module unload gcc && echo "no module function available. Probably not on a slurm cluster."
-#
-#rm -rf dist build
-#make cleaneggs
-#make cleanlibs
-#
-#make clean
-#export CUDA_HOME=
-#export CUDA_VERSION=
-#make cpuonly CUDA_VERSION="CPU"
-#
-#if [ ! -f "./bitsandbytes/libbitsandbytes_cpu.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-11.0
-#make cuda110 CUDA_VERSION=110
-#
-#if [ ! -f "./bitsandbytes/libbitsandbytes_cuda110.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-11.1
-#make cuda11x CUDA_VERSION=111
-#
-#if [ ! -f "./bitsandbytes/libbitsandbytes_cuda111.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-11.4
-#make cuda11x CUDA_VERSION=114
-#
-#if [ ! -f "./bitsandbytes/libbitsandbytes_cuda114.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-11.5
-#make cuda11x CUDA_VERSION=115
-#
-#if [ ! -f "./bitsandbytes/libbitsandbytes_cuda115.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-11.7
-#make cuda11x CUDA_VERSION=117
-#
-#if [ ! -f "./bitsandbytes/libbitsandbytes_cuda117.so" ]; then
-#  # Control will enter here if $DIRECTORY doesn't exist.
-#  echo "Compilation unsuccessul!" 1>&2
-#  exit 64
-#fi
+module unload cuda && echo "no module function available. Probably not on a slurm cluster."
+module unload gcc && echo "no module function available. Probably not on a slurm cluster."
+
+rm -rf dist build
+make cleaneggs
+make cleanlibs
+
+make clean
+export CUDA_HOME=
+export CUDA_VERSION=
+make cpuonly CUDA_VERSION="CPU"
+
+if [ ! -f "./bitsandbytes/libbitsandbytes_cpu.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-11.0
+make cuda110 CUDA_VERSION=110
+
+if [ ! -f "./bitsandbytes/libbitsandbytes_cuda110.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-11.1
+make cuda11x CUDA_VERSION=111
+
+if [ ! -f "./bitsandbytes/libbitsandbytes_cuda111.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-11.4
+make cuda11x CUDA_VERSION=114
+
+if [ ! -f "./bitsandbytes/libbitsandbytes_cuda114.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-11.5
+make cuda11x CUDA_VERSION=115
+
+if [ ! -f "./bitsandbytes/libbitsandbytes_cuda115.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-11.7
+make cuda11x CUDA_VERSION=117
+
+if [ ! -f "./bitsandbytes/libbitsandbytes_cuda117.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-11.8
-make cuda12x CUDA_VERSION=118
+make cuda118 CUDA_VERSION=118

 if [ ! -f "./bitsandbytes/libbitsandbytes_cuda118.so" ]; then
  # Control will enter here if $DIRECTORY doesn't exist.
@ -172,7 +172,7 @@ fi

 make clean
 export CUDA_HOME=$BASE_PATH/cuda-11.8
-make cuda12x_nomatmul CUDA_VERSION=118
+make cuda118_nomatmul CUDA_VERSION=118

 if [ ! -f "./bitsandbytes/libbitsandbytes_cuda118_nocublaslt.so" ]; then
  # Control will enter here if $DIRECTORY doesn't exist.