Fix build issues on Windows

build_win.bat

@@ -0,0 +1,19 @@
+@echo off
+
+set DS_BUILD_AIO=0
+set DS_BUILD_SPARSE_ATTN=0
+
+echo Administrative permissions required. Detecting permissions...
+
+net session >nul 2>&1
+if %errorLevel% == 0 (
+    echo Success: Administrative permissions confirmed.
+) else (
+    echo Failure: Current permissions inadequate.
+    goto end
+)
+
+
+python setup.py bdist_wheel
+
+:end

csrc/transformer/inference/includes/inference_context.h

@@ -157,7 +157,7 @@ class Context {
     void* GetWorkSpace() { return _workspace; }
     void* GetAttentionUnfusedWorkspace()
     {
-        return _workspace + _attention_unfused_workspace_offset;
+        return (char*)_workspace + _attention_unfused_workspace_offset;
     }
 
     inline unsigned new_token(unsigned layer_id)

op_builder/builder.py

@@ -638,6 +638,7 @@ def nvcc_args(self):
         else:
             cuda_major, _ = installed_cuda_version()
             args += [
+                '-allow-unsupported-compiler' if sys.platform == "win32" else '',
                 '--use_fast_math',
                 '-std=c++17'
                 if sys.platform == "win32" and cuda_major > 10 else '-std=c++14',

op_builder/fused_adam.py

@@ -3,6 +3,8 @@
 """
 from .builder import CUDAOpBuilder
 
+import sys
+
 
 class FusedAdamBuilder(CUDAOpBuilder):
     BUILD_VAR = "DS_BUILD_FUSED_ADAM"
@@ -27,6 +29,9 @@ def cxx_args(self):
     def nvcc_args(self):
         nvcc_flags = ['-O3'] + self.version_dependent_macros()
         if not self.is_rocm_pytorch():
-            nvcc_flags.extend(['-lineinfo',
-                               '--use_fast_math'] + self.compute_capability_args())
+            nvcc_flags.extend([
+                '-allow-unsupported-compiler' if sys.platform == "win32" else '',
+                '-lineinfo',
+                '--use_fast_math'
+            ] + self.compute_capability_args())
         return nvcc_flags

op_builder/fused_lamb.py

@@ -3,6 +3,8 @@
 """
 from .builder import CUDAOpBuilder
 
+import sys
+
 
 class FusedLambBuilder(CUDAOpBuilder):
     BUILD_VAR = 'DS_BUILD_FUSED_LAMB'
@@ -33,6 +35,9 @@ def nvcc_args(self):
                 '-DROCM_VERSION_MINOR=%s' % ROCM_MINOR
             ]
         else:
-            nvcc_flags.extend(['-lineinfo',
-                               '--use_fast_math'] + self.compute_capability_args())
+            nvcc_flags.extend([
+                '-allow-unsupported-compiler' if sys.platform == "win32" else '',
+                '-lineinfo',
+                '--use_fast_math'
+            ] + self.compute_capability_args())
         return nvcc_flags

setup.py

@@ -4,12 +4,13 @@
 DeepSpeed library
 
 To build wheel on Windows:
-    1. Install pytorch, such as pytorch 1.8 + cuda 11.1
+    1. Install pytorch, such as pytorch 1.12 + cuda 11.6
     2. Install visual cpp build tool
-    3. Launch cmd console with Administrator privilege for creating required symlink folders
+    3. Include cuda toolkit
+    4. Launch cmd console with Administrator privilege for creating required symlink folders
 
 Create a new wheel via the following command:
-    python setup.py bdist_wheel
+    build_win.bat
 
 The wheel will be located at: dist/*.whl
 """
@@ -224,11 +225,12 @@ def create_dir_symlink(src, dest):
 hip_version = "0.0"
 if torch_available and torch.version.cuda is not None:
     cuda_version = ".".join(torch.version.cuda.split('.')[:2])
-    if isinstance(torch.cuda.nccl.version(), int):
-        # This will break if minor version > 9
-        nccl_version = ".".join(str(torch.cuda.nccl.version())[:2])
-    else:
-        nccl_version = ".".join(map(str, torch.cuda.nccl.version()[:2]))
+    if sys.platform != "win32":
+        if isinstance(torch.cuda.nccl.version(), int):
+            # This will break if minor version > 9
+            nccl_version = ".".join(str(torch.cuda.nccl.version())[:2])
+        else:
+            nccl_version = ".".join(map(str, torch.cuda.nccl.version()[:2]))
     if hasattr(torch.cuda, 'is_bf16_supported') and torch.cuda.is_available():
         bf16_support = torch.cuda.is_bf16_supported()
 if torch_available and hasattr(torch.version, 'hip') and torch.version.hip is not None: