Graceful handling of cpp extensions

README.md

@@ -39,12 +39,15 @@ python setup.py develop
 
 If you want to install from source run
 ```Shell
-python setup.py install
+python setup.py install 
 ```
 
 ** Note:
-Since we are building pytorch c++/cuda extensions by default, running `pip install .` will
-not work.
+If you are running into any issues while building `ao` cpp extensions you can instead build using
+
+```shell
+USE_CPP=0 python setup.py install
+```
 
 ### Quantization
 

setup.py

@@ -19,6 +19,8 @@ def read_requirements(file_path):
 # Determine the package name based on the presence of an environment variable
 package_name = "torchao-nightly" if os.environ.get("TORCHAO_NIGHTLY") else "torchao"
 version_suffix = os.getenv("VERSION_SUFFIX", "")
+use_cpp = os.getenv('USE_CPP')
+
 
 # Version is year.month.date if using nightlies
 version = current_date if package_name == "torchao-nightly" else "0.2.0"
@@ -92,7 +94,7 @@ def get_extensions():
     package_data={
         "torchao.kernel.configs": ["*.pkl"],
     },
-    ext_modules=get_extensions(),
+    ext_modules=get_extensions() if use_cpp != "0" else None,
     install_requires=read_requirements("requirements.txt"),
     extras_require={"dev": read_requirements("dev-requirements.txt")},
     description="Package for applying ao techniques to GPU models",

test/dtypes/test_float6_e3m2.py

@@ -5,6 +5,13 @@
     parametrize,
     run_tests,
 )
+
+try:
+    import torchao.ops
+except RuntimeError:
+    pytest.skip("torchao.ops not available")
+
+
 from torchao.dtypes.float6_e3m2 import to_float6_e3m2, from_float6_e3m2
 
 

test/test_ops.py

@@ -7,6 +7,11 @@
 from parameterized import parameterized
 import pytest
 
+try:
+    import torchao.ops
+except RuntimeError:
+    pytest.skip("torchao.ops not available")
+
 
 # torch.testing._internal.optests.generate_tests.OpCheckError: opcheck(op, ...):
 # test_faketensor failed with module 'torch' has no attribute '_custom_ops' (scroll up for stack trace)

torchao/__init__.py

@@ -1,12 +1,18 @@
 import torch
+import logging
+
 _IS_FBCODE = (
     hasattr(torch._utils_internal, "IS_FBSOURCE") and
     torch._utils_internal.IS_FBSOURCE
 )
 
 if not _IS_FBCODE:
-    from . import _C
-    from . import ops
+    try:
+        from . import _C
+        from . import ops
+    except:
+        _C = None
+        logging.info("Skipping import of cpp extensions")
 
 from torchao.quantization import (
     apply_weight_only_int8_quant,

torchao/dtypes/__init__.py

@@ -1,14 +1,18 @@
 from .nf4tensor import NF4Tensor, to_nf4
 from .uint4 import UInt4Tensor
 from .aqt import AffineQuantizedTensor, to_aq
-from .float6_e3m2 import to_float6_e3m2, from_float6_e3m2
 
 __all__ = [
     "NF4Tensor",
     "to_nf4",
     "UInt4Tensor"
     "AffineQuantizedTensor",
     "to_aq",
-    "to_float6_e3m2",
-    "from_float6_e3m2",
 ]
+
+# CPP extensions
+try:
+    from .float6_e3m2 import to_float6_e3m2, from_float6_e3m2
+    __all__.extend(["to_float6_e3m2", "from_float6_e3m2"])
+except RuntimeError:
+    pass