Fixed CMAKE CUDA Version. Updated compile doc

CMakelists.txt

@@ -9,26 +9,31 @@ set(FILES_CPP csrc/common.cpp csrc/cpu_ops.cpp csrc/pythonInterface.c)
 
 option(MAKE_CUDA_BUILD "Build using CUDA" ON)
 option(NO_CUBLASLT "Don't use CUBLAST" OFF)
-
-set(COMPUTE_CAPABILITY 
-        50 52 60 61 70 72
-    )
-
-set(CC_KEPLER 35 37)
+set(CUDA_VERSION "11.7" CACHE STRING "CUDA Version DLL Name: 11.0, 11.7, 11.6")
+set(CUDA_VERSION_FIXED "")
+set(CUDA_VERSION_MAJOR "")
+string(REPLACE "." "" CUDA_VERSION_FIXED "${CUDA_VERSION}")
+string(REGEX MATCH "[^\.]+" CUDA_VERSION_MAJOR "${CUDA_VERSION}")
+    
 # Later versions of CUDA support the new architectures
 set(CC_CUDA10x 75)
-
 set(CC_CUDA110 75 80)
 set(CC_CUDA11x 75 80 86)
-set(CC_cublasLt110 75 80)
-
-set(CC_cublasLt111 75 80 86)
-set(CC_ADA_HOPPER 89 90)
+set(CC_CUDA12x 89 90)
 
 add_definitions(-DUSE_AVX2 -DUSE_AVX)
 if( MAKE_CUDA_BUILD ) 
     if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-        set(CMAKE_CUDA_ARCHITECTURES 75 80 86)
+        if(${CUDA_VERSION} EQUAL "11.0")
+            set(CMAKE_CUDA_ARCHITECTURES ${CC_CUDA110})
+        elseif(${CUDA_VERSION_MAJOR} EQUAL "11")
+            set(CMAKE_CUDA_ARCHITECTURES ${CC_CUDA11x})
+        elseif(${CUDA_VERSION_MAJOR} EQUAL "12")
+            set(CMAKE_CUDA_ARCHITECTURES ${CC_CUDA12x})
+        else()
+            set(CMAKE_CUDA_ARCHITECTURES ${CC_CUDA10x})
+            set(NO_CUBLASLT ON)
+        endif()
     endif()
 
     if(NOT DEFINED CMAKE_CUDA_STANDARD)
@@ -54,7 +59,7 @@ if( MAKE_CUDA_BUILD )
         POSITION_INDEPENDENT_CODE ON
         CUDA_SEPARABLE_COMPILATION ON
         PREFIX ""
-        OUTPUT_NAME "libbitsandbytes_cuda${CMAKE_CUDA_STANDARD}7"
+        OUTPUT_NAME "libbitsandbytes_cuda${CUDA_VERSION_FIXED}"
         LINKER_LANGUAGE C 
         WINDOWS_EXPORT_ALL_SYMBOLS TRUE)
 #    add_link_options(-Wl,--export-all-symbols)

compile_from_source.md

@@ -1,5 +1,59 @@
 # Compiling from source
 
+## Windows
+
+CPU NOT TESTED
+
+Ensure you have your environment you want to bring in bitsandbytes. (a bloom setup, textgen-ui, etc) - via conda.
+I'd suggest to install MAMBA  and use it, as it's way faster.
+
+IMPORTANT - ensure your environment.yml matches your installed CUDA version in Visual Studio. Keep One VERSION.
+
+1. environment.yml. Replace/tweak as necesarry (remove cuda, change version):
+```
+name: mycompileenv
+channels:
+  - pytorch
+  - nvidia
+  - huggingface
+  - conda-forge
+  - anaconda
+    
+dependencies:
+  - python=3.10.9
+  - pytorch-cuda=11.7 
+  - pytorch
+  - torchvision
+  - torchaudio 
+  - transformers
+  - cudatoolkit=11.7
+  - jupyter
+  - notebook
+  - pytest
+```
+
+2. Then open POWERSHELL
+```
+conda install -c conda-forge mamba
+mamba env create
+mamba env activate mycompileenv
+& "C:\Program Files\Microsoft Visual Studio\2022\Community\Common7\Tools\Launch-VsDevShell.ps1" -List -Arch amd64
+```
+At this point - select your visual studio installation - aka hit 1
+
+3. Go into your bitsandbytes folder and run `cmake-gui .` 
+4. Make sure you put the build folder correctly, append "build" to Where to build the binaries
+5. Hit Configure
+6. Set your CUDA_VERSION to whatever you have. If you deselect MAKE_CUDA_BUILD, leave as is
+7. Hit Configure again, then Generate
+8. Open Visual Studio and select Release as configuration. Build Solution
+9. copy everything from `build\Release\*.*` over in the `bitsandbytes` folder (the one with the python modules)
+10. run tests `python -m pytest`. You may need to use `mamba` to install other modules
+11. build wheel `mamba install build` and then `python -m build --wheel`
+12. install wheel  `pip install .\dist\*.whl`
+
+
+## Linux
 Basic steps.
 1. `make [target]` where `[target]` is among `cuda92, cuda10x, cuda110, cuda11x, cuda12x, cpuonly`
 2. `CUDA_VERSION=XXX python setup.py install`