Release 0.39.1.

Author: TimDettmers

CHANGELOG.md

@@ -239,3 +239,11 @@ Features:
  - Doubled quantization routines for 4-bit quantization
  - Paged optimizers for Adam and Lion.
  - bfloat16 gradient / weight support for Adam and Lion with 8 or 32-bit states.
+
+Bug fixes:
+ - Fixed a bug where 8-bit models consumed twice the memory as expected after serialization
+
+Deprecated:
+ - Kepler binaries (GTX 700s and Tesla K40/K80) are not longer provided via pip and need to be compiled from source. Kepler support might be fully removed in the future.
+
+

Makefile

@@ -59,16 +59,27 @@ all: $(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)
 
-cuda110_nomatmul: $(BUILD_DIR) env
+cuda110_nomatmul_kepler: $(BUILD_DIR) env
 	$(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
+cuda11x_nomatmul_kepler: $(BUILD_DIR) env
 	$(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)
 
+
+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
+	$(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
+	$(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

compile_from_source.md

@@ -33,3 +33,8 @@ You can set `CUDA_HOME` to `/usr/local/cuda-11.7`. For example, you might be abl
 
 
 If you have problems compiling the library with these instructions from source, please open an issue.
+
+## Compilation with Kepler
+
+Since 0.39.1 bitsandbytes installed via pip no longer provides Kepler binaries and these need to be compiled from source. Follow the steps above and instead of `cuda11x_nomatmul` etc use `cuda11x_nomatmul_kepler`
+

setup.py

@@ -18,7 +18,7 @@ def read(fname):
 
 setup(
     name=f"bitsandbytes",
-    version=f"0.39.0",
+    version=f"0.39.1",
     author="Tim Dettmers",
     author_email="[email protected]",
     description="k-bit optimizers and matrix multiplication routines.",