Deprecate config functions like int4_weight_only

**Summary:** These have been superseded by `AOBaseConfig` objects
for several releases already, but we never deprecated them.
We will keep them around for another release before breaking BC
and removing them.

**Test Plan:**

```
python test/quantization/test_quant_api.py -k test_config_deprecation
```

[ghstack-poisoned]

Author: andrewor14

test/quantization/test_quant_api.py

@@ -10,6 +10,7 @@
 import gc
 import tempfile
 import unittest
+import warnings
 from pathlib import Path
 
 import torch
@@ -752,6 +753,50 @@ def test_int4wo_cuda_serialization(self):
             # load state_dict in cuda
             model.load_state_dict(sd, assign=True)
 
+    def test_config_deprecation(self):
+        """
+        Test that old config functions like `int4_weight_only` trigger deprecation warnings.
+        """
+        from torchao.quantization import (
+            float8_weight_only,
+            fpx_weight_only,
+            gemlite_uintx_weight_only,
+            int4_weight_only,
+            int8_dynamic_activation_int4_weight,
+            int8_dynamic_activation_int8_weight,
+            int8_weight_only,
+            uintx_weight_only,
+        )
+
+        # Reset deprecation warning state, otherwise we won't log warnings here
+        warnings.resetwarnings()
+
+        # Map from deprecated API to the args needed to instantiate it
+        deprecated_apis_to_args = {
+            float8_weight_only: (),
+            fpx_weight_only: (3, 2),
+            gemlite_uintx_weight_only: (),
+            int4_weight_only: (),
+            int8_dynamic_activation_int4_weight: (),
+            int8_dynamic_activation_int8_weight: (),
+            int8_weight_only: (),
+            uintx_weight_only: (torch.uint4,),
+        }
+
+        with warnings.catch_warnings(record=True) as _warnings:
+            # Call each deprecated API twice
+            for cls, args in deprecated_apis_to_args.items():
+                cls(*args)
+                cls(*args)
+
+            # Each call should trigger the warning only once
+            self.assertEqual(len(_warnings), len(deprecated_apis_to_args))
+            for w in _warnings:
+                self.assertIn(
+                    "is deprecated and will be removed in a future release",
+                    str(w.message),
+                )
+
 
 common_utils.instantiate_parametrized_tests(TestQuantFlow)
 

torchao/quantization/quant_api.py

@@ -92,6 +92,7 @@
     to_weight_tensor_with_linear_activation_quantization_metadata,
 )
 from torchao.utils import (
+    _ConfigDeprecationWrapper,
     _is_fbgemm_genai_gpu_available,
     is_MI300,
     is_sm_at_least_89,
@@ -639,7 +640,9 @@ def __post_init__(self):
 
 
 # for BC
-int8_dynamic_activation_int4_weight = Int8DynamicActivationInt4WeightConfig
+int8_dynamic_activation_int4_weight = _ConfigDeprecationWrapper(
+    "int8_dynamic_activation_int4_weight", Int8DynamicActivationInt4WeightConfig
+)
 
 
 @register_quantize_module_handler(Int8DynamicActivationInt4WeightConfig)
@@ -1018,7 +1021,9 @@ def __post_init__(self):
 
 
 # for BC
-gemlite_uintx_weight_only = GemliteUIntXWeightOnlyConfig
+gemlite_uintx_weight_only = _ConfigDeprecationWrapper(
+    "gemlite_uintx_weight_only", GemliteUIntXWeightOnlyConfig
+)
 
 
 @register_quantize_module_handler(GemliteUIntXWeightOnlyConfig)
@@ -1100,7 +1105,7 @@ def __post_init__(self):
 
 # for BC
 # TODO maybe change other callsites
-int4_weight_only = Int4WeightOnlyConfig
+int4_weight_only = _ConfigDeprecationWrapper("int4_weight_only", Int4WeightOnlyConfig)
 
 
 def _int4_weight_only_quantize_tensor(weight, config):
@@ -1310,7 +1315,7 @@ def __post_init__(self):
 
 
 # for BC
-int8_weight_only = Int8WeightOnlyConfig
+int8_weight_only = _ConfigDeprecationWrapper("int8_weight_only", Int8WeightOnlyConfig)
 
 
 def _int8_weight_only_quantize_tensor(weight, config):
@@ -1471,7 +1476,9 @@ def __post_init__(self):
 
 
 # for BC
-int8_dynamic_activation_int8_weight = Int8DynamicActivationInt8WeightConfig
+int8_dynamic_activation_int8_weight = _ConfigDeprecationWrapper(
+    "int8_dynamic_activation_int8_weight", Int8DynamicActivationInt8WeightConfig
+)
 
 
 def _int8_dynamic_activation_int8_weight_quantize_tensor(weight, config):
@@ -1580,7 +1587,9 @@ def __post_init__(self):
 
 
 # for BC
-float8_weight_only = Float8WeightOnlyConfig
+float8_weight_only = _ConfigDeprecationWrapper(
+    "float8_weight_only", Float8WeightOnlyConfig
+)
 
 
 def _float8_weight_only_quant_tensor(weight, config):
@@ -1994,7 +2003,9 @@ def __post_init__(self):
 
 
 # for BC
-uintx_weight_only = UIntXWeightOnlyConfig
+uintx_weight_only = _ConfigDeprecationWrapper(
+    "uintx_weight_only", UIntXWeightOnlyConfig
+)
 
 
 @register_quantize_module_handler(UIntXWeightOnlyConfig)
@@ -2234,7 +2245,7 @@ def __post_init__(self):
 
 
 # for BC
-fpx_weight_only = FPXWeightOnlyConfig
+fpx_weight_only = _ConfigDeprecationWrapper("fpx_weight_only", FPXWeightOnlyConfig)
 
 
 @register_quantize_module_handler(FPXWeightOnlyConfig)

torchao/utils.py

@@ -12,7 +12,7 @@
 from functools import reduce
 from importlib.metadata import version
 from math import gcd
-from typing import Any, Callable, Optional
+from typing import Any, Callable, Optional, Type
 
 import torch
 import torch.nn.utils.parametrize as parametrize
@@ -432,6 +432,25 @@ def __eq__(self, other):
 TORCH_VERSION_AFTER_2_2 = _deprecated_torch_version_after("2.2.0.dev")
 
 
+class _ConfigDeprecationWrapper:
+    """
+    A deprecation wrapper that directs users from a deprecated "config function"
+    (e.g. `int4_weight_only`) to the replacement config class.
+    """
+
+    def __init__(self, deprecated_name: str, config_cls: Type):
+        self.deprecated_name = deprecated_name
+        self.config_cls = config_cls
+
+    def __call__(self, *args, **kwargs):
+        warnings.warn(
+            f"`{self.deprecated_name}` is deprecated and will be removed in a future release. "
+            f"Please use `{self.config_cls.__name__}` instead. Example usage:\n"
+            f"    quantize_(model, {self.config_cls.__name__}(...))"
+        )
+        return self.config_cls(*args, **kwargs)
+
+
 """
 Helper function for implementing aten op or torch function dispatch
 and dispatching to these implementations.