Add register_apply_tensor_subclass (pytorch#366)

Summary:
`register_apply_tensor_subclass` allows users to add a string shortcut for
a new apply_tensor_subclass function, they can use this to test their new dtype tensor subclass

see `test/quantization/test_quant_api.py -k test_register_apply_tensor_subclass` for detail

Test Plan:
python test/quantization/test_quant_api.py -k test_register_apply_tensor_subclass

Reviewers:

Subscribers:

Tasks:

Tags:

test/quantization/test_quant_api.py

@@ -31,11 +31,11 @@
     Int8WeightOnlyQuantizedLinearWeight,
     Int4WeightOnlyQuantizedLinearWeight,
 )
+from torchao import quantize
 from torchao.quantization.quant_api import (
     _replace_with_custom_fn_if_matches_filter,
     Quantizer,
     TwoStepQuantizer,
-    quantize,
     int8da_int4w,
     int4wo,
     int8wo,
@@ -51,6 +51,7 @@
 from torchao.utils import unwrap_tensor_subclass
 import copy
 import tempfile
+from torch.testing._internal.common_utils import TestCase
 
 
 def dynamic_quant(model, example_inputs):
@@ -147,7 +148,7 @@ def _ref_change_linear_weights_to_woqtensors(model, filter_fn=None, **kwargs):
 _ref_change_linear_weights_to_int8_woqtensors = _get_ref_change_linear_weights_to_woqtensors(Int8WeightOnlyQuantizedLinearWeight)
 _ref_change_linear_weights_to_int4_woqtensors = _get_ref_change_linear_weights_to_woqtensors(Int4WeightOnlyQuantizedLinearWeight)
 
-class TestQuantFlow(unittest.TestCase):
+class TestQuantFlow(TestCase):
     def test_dynamic_quant_gpu_singleline(self):
         m = ToyLinearModel().eval()
         example_inputs = m.example_inputs()
@@ -601,5 +602,20 @@ def test_quantized_tensor_subclass_int8_dyn_quant(self):
         # make sure it compiles
         torch._export.aot_compile(m_unwrapped, example_inputs)
 
+    def test_register_apply_tensor_subclass(self):
+        from torchao import register_apply_tensor_subclass
+        def apply_my_dtype(weight):
+            return weight * 2
+
+        m = ToyLinearModel().eval()
+        example_inputs = m.example_inputs()
+        with self.assertRaisesRegex(ValueError, "not supported"):
+            quantize(m, "my_dtype")
+
+        register_apply_tensor_subclass("my_dtype", apply_my_dtype)
+        # make sure it runs
+        quantize(m, "my_dtype")
+        m(*example_inputs)
+
 if __name__ == "__main__":
     unittest.main()

torchao/__init__.py

@@ -25,10 +25,14 @@
 
 from torchao.quantization import (
     autoquant,
+    quantize,
+    register_apply_tensor_subclass,
 )
 from . import dtypes
 
 __all__ = [
     "dtypes",
     "autoquant",
+    "quantize",
+    "register_apply_tensor_subclass",
 ]

torchao/quantization/__init__.py

@@ -14,12 +14,6 @@
 from .autoquant import *
 
 __all__ = [
-    "DynamicallyPerAxisQuantizedLinear",
-    "apply_weight_only_int8_quant",
-    "apply_dynamic_quant",
-    "change_linear_weights_to_int8_dqtensors",
-    "change_linear_weights_to_int8_woqtensors",
-    "change_linear_weights_to_int4_woqtensors",
     "swap_conv2d_1x1_to_linear"
     "safe_int_mm",
     "autoquant",
@@ -31,14 +25,12 @@
     "swap_linear_with_smooth_fq_linear",
     "smooth_fq_linear_to_inference",
     "set_smooth_fq_attribute",
-    "Int8DynamicallyQuantizedLinearWeight",
-    "Int8WeightOnlyQuantizedLinearWeight",
-    "Int4WeightOnlyQuantizedLinearWeight",
     "compute_error",
-    "WeightOnlyInt8QuantLinear",
     "Int4WeightOnlyGPTQQuantizer",
     "Int4WeightOnlyQuantizer",
     "quantize_affine",
     "dequantize_affine",
     "choose_qprams_affine",
+    "quantize",
+    "register_apply_tensor_subclass",
 ]

torchao/quantization/quant_api.py

@@ -41,6 +41,7 @@
     Int4WeightOnlyGPTQQuantizer,
     Int4WeightOnlyQuantizer,
 )
+import logging
 from .autoquant import autoquant, AutoQuantizableLinearWeight
 
 
@@ -50,13 +51,14 @@
     "TwoStepQuantizer",
     "Int4WeightOnlyGPTQQuantizer",
     "Int4WeightOnlyQuantizer",
-    "quantize",
     "autoquant",
     "_get_subclass_inserter",
+    "quantize",
     "int8da_int4w",
     "int8da_int8w",
     "int4wo",
     "int8wo",
+    "register_apply_tensor_subclass",
 ]
 
 from .GPTQ import (
@@ -292,7 +294,8 @@ def filter_fn(module, fqn):
         m = quantize(m, apply_weight_quant, filter_fn)
     """
     if isinstance(apply_tensor_subclass, str):
-        assert apply_tensor_subclass in _APPLY_TS_TABLE, f"{apply_tensor_subclass} not supported: {_APPLY_TS_TABLE.keys()}"
+        if apply_tensor_subclass not in _APPLY_TS_TABLE:
+            raise ValueError(f"{apply_tensor_subclass} not supported: {_APPLY_TS_TABLE.keys()}")
         apply_tensor_subclass = _APPLY_TS_TABLE[apply_tensor_subclass]
 
     assert not isinstance(apply_tensor_subclass, str)
@@ -438,3 +441,19 @@ def get_per_token_block_size(x):
     "int8_weight_only": int8wo(),
     "int8_dynamic": int8da_int8w(),
 }
+
+def register_apply_tensor_subclass(name: str, apply_tensor_subclass: Callable):
+    """Register a string shortcut for `apply_tensor_subclass` that takes a weight Tensor
+    as input and ouptuts a tensor with tensor subclass applied
+
+    Example:
+        def apply_my_dtype(weight):
+            return weight * 2
+
+        register_apply_tensor_subclass("my_dtype", apply_my_dtype)
+        # calls `apply_my_dtype` on weights
+        quantize(m, "my_dtype")
+    """
+    if name in _APPLY_TS_TABLE:
+        logging.warning(f"shortcut string {name} already exist, overwriting")
+    _APPLY_TS_TABLE[name] = apply_tensor_subclass