Reenable int4 integration tests (#91)

Author: cpuhrsch

.github/workflows/regression_test.yml

@@ -22,9 +22,9 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
+        pip install torch
         pip install -r requirements.txt
         pip install -r dev-requirements.txt
-        pip install torch
         
 
     - name: Install package
@@ -48,9 +48,9 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
+        pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu121
         pip install -r requirements.txt
         pip install -r dev-requirements.txt
-        pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu121
         
 
     - name: Install package
@@ -74,9 +74,9 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
+        pip install torch --index-url https://download.pytorch.org/whl/cpu
         pip install -r requirements.txt
         pip install -r dev-requirements.txt
-        pip install torch --index-url https://download.pytorch.org/whl/cpu
         
 
     - name: Install package
@@ -100,9 +100,9 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
+        pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cpu
         pip install -r requirements.txt
         pip install -r dev-requirements.txt
-        pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cpu
         
 
     - name: Install package

test/dtypes/test_nf4.py

@@ -195,10 +195,13 @@ def test_smoketest_linear(self):
     @unittest.skipIf(torch.__version__.split('+')[0] == '2.2.1', "Broken on stable.")
     @unittest.skipIf(not torch.cuda.is_available(), "Need CUDA available")
     def test_smoketest_linear_compile(self):
-        a = torch.randn(32, 32, dtype=torch.bfloat16, device='cuda')
-        a_nf4 = torchao.dtypes.to_nf4(a, 16, 2)
-        inp = torch.randn(2, 32, 32, dtype=a.dtype, device=a.device)
-        out3 = torch.compile(torch.nn.functional.linear, mode='max-autotune')(inp, a_nf4)
+        for dtype in [torch.bfloat16, torch.float16]:
+            if torch.cuda.is_available() and torch.cuda.get_device_capability() < (8, 0) and dtype == torch.bfloat16:
+                self.skipTest("test requires SM capability of at least (8, 0).")
+            a = torch.randn(32, 32, dtype=dtype, device='cuda')
+            a_nf4 = torchao.dtypes.to_nf4(a, 16, 2)
+            inp = torch.randn(2, 32, 32, dtype=a.dtype, device=a.device)
+            out3 = torch.compile(torch.nn.functional.linear, mode='max-autotune')(inp, a_nf4)