update readme

pytorch · Sep 23, 2024 · 053a97d · 053a97d
1 parent 18585d2
commit 053a97d
Show file tree

Hide file tree

Showing 2 changed files with 3 additions and 8 deletions.
diff --git a/test/integration/test_integration.py b/test/integration/test_integration.py
@@ -102,9 +102,7 @@
 def _int8wo_api(mod):
     if TORCH_VERSION_AT_LEAST_2_4:
         quantize_(mod, int8_weight_only(), set_inductor_config=False)
-        if (
-            not TORCH_VERSION_AT_LEAST_2_5
-        ) or torch._inductor.config.freezing:
+        if not TORCH_VERSION_AT_LEAST_2_5:
             unwrap_tensor_subclass(mod)
     else:
         change_linear_weights_to_int8_woqtensors(mod)
@@ -1041,10 +1039,7 @@ def test_save_load_dqtensors(self, device, dtype):
     @parameterized.expand(COMMON_DEVICE_DTYPE)
     @torch.no_grad()
     @unittest.skipIf(is_fbcode(), "broken in fbcode")
-    def test_save_load_int8woqtensors(self, device, dtype):      
-        if TORCH_VERSION_AT_LEAST_2_5 and device == "cpu":
-            self.skipTest(f"Regression introduced in PT nightlies")
-
+    def test_save_load_int8woqtensors(self, device, dtype):
         undo_recommended_configs()
         self._test_handle_save_load_meta_impl(_int8wo_api, device, test_dtype=dtype)
 

diff --git a/torchao/quantization/README.md b/torchao/quantization/README.md
@@ -246,7 +246,7 @@ but if you use 2.4 or before, you'll need to use `unwrap_tensor_subclass` as wel
 
 Note that the workaround will not be needed after https://github.com/pytorch/pytorch/issues/129682 is fixed.
 
-
+Note that the workaround is also needed for `torch.compile` with `freezing` (`torch._inductor.config.freezing=True`) for now and will not be needed after https://github.com/pytorch/pytorch/pull/136265 landed.
 
 ## Other Available Quantization Techniques