Move QAT out of prototype

Summary: Move QAT out of prototype so we can provide stronger
BC guarantees moving forward.

**BC-breaking notes**

Before:
```
from torchao.quantization.prototype.qat import (
    disable_4w_fake_quant,
    disable_8da4w_fake_quant,
    enable_4w_fake_quant,
    enable_8da4w_fake_quant,
    ComposableQATQuantizer,
    Int4WeightOnlyQATQuantizer,
    Int4WeightOnlyEmbeddingQATQuantizer
    Int8DynActInt4WeightQATQuantizer,
    Int8DynActInt4WeightQATLinear,
)
from torchao.quantization.prototype.qat.api import (
    FakeQuantizeConfig,
)
from torchao.quantization.prototype.qat.fake_quantizer import (
    FakeQuantizer,
)
```

After:
```
from torchao.quantization.qat import (
    ComposableQATQuantizer,
    Int4WeightOnlyQATQuantizer,
    Int4WeightOnlyEmbeddingQATQuantizer
    Int8DynActInt4WeightQATQuantizer,
)
from torchao.quantization.qat.linear import (
    disable_4w_fake_quant,
    disable_8da4w_fake_quant,
    enable_4w_fake_quant,
    enable_8da4w_fake_quant,
    Int8DynActInt4WeightQATLinear,
)
from torchao.quantization.qat.api import (
    FakeQuantizeConfig,
)
from torchao.quantization.qat.fake_quantizer import (
    FakeQuantizer,
)
```

Test Plan:
python test/quantization/test_qat.py

ghstack-source-id: 9ebc45d
Pull Request resolved: #1091

Author: andrewor14

README.md

@@ -59,7 +59,7 @@ In practice these features alongside int4 weight only quantization allow us to *
 Post-training quantization can result in a fast and compact model, but may also lead to accuracy degradation. We recommend exploring Quantization Aware Training (QAT) to overcome this limitation. In collaboration with Torchtune, we've developed a QAT recipe that demonstrates significant accuracy improvements over traditional PTQ, recovering **96% of the accuracy degradation on hellaswag and 68% of the perplexity degradation on wikitext** for Llama3 compared to post-training quantization (PTQ). And we've provided a full recipe [here](https://pytorch.org/blog/quantization-aware-training/)
 
 ```python
-from torchao.quantization.prototype.qat import Int8DynActInt4WeightQATQuantizer
+from torchao.quantization.qat import Int8DynActInt4WeightQATQuantizer
 
 qat_quantizer = Int8DynActInt4WeightQATQuantizer()
 

test/quantization/test_qat.py

@@ -22,20 +22,20 @@
     PerRow,
     PerToken,
 )
-from torchao.quantization.prototype.qat.api import (
+from torchao.quantization.qat.api import (
     ComposableQATQuantizer,
     FakeQuantizeConfig,
 )
-from torchao.quantization.prototype.qat.fake_quantizer import (
+from torchao.quantization.qat.fake_quantizer import (
     FakeQuantizer,
 )
-from torchao.quantization.prototype.qat.embedding import (
+from torchao.quantization.qat.embedding import (
     FakeQuantizedEmbedding,
 )
-from torchao.quantization.prototype.qat.linear import (
+from torchao.quantization.qat.linear import (
     FakeQuantizedLinear,
 )
-from torchao.quantization.prototype.qat.utils import (
+from torchao.quantization.qat.utils import (
     _choose_qparams_per_token_asymmetric,
     _fake_quantize_per_channel_group,
     _fake_quantize_per_token,
@@ -175,7 +175,7 @@ def _set_ptq_weight(
             Int8DynActInt4WeightLinear,
             WeightOnlyInt4Linear,
         )
-        from torchao.quantization.prototype.qat.linear import (
+        from torchao.quantization.qat.linear import (
             Int8DynActInt4WeightQATLinear,
             Int4WeightOnlyQATLinear,
         )
@@ -207,7 +207,7 @@ def _set_ptq_weight(
 
     @unittest.skipIf(not TORCH_VERSION_AT_LEAST_2_4, "skipping when torch version is 2.4 or lower")
     def test_qat_8da4w_linear(self):
-        from torchao.quantization.prototype.qat.linear import Int8DynActInt4WeightQATLinear
+        from torchao.quantization.qat.linear import Int8DynActInt4WeightQATLinear
         from torchao.quantization.GPTQ import Int8DynActInt4WeightLinear
 
         group_size = 128
@@ -232,7 +232,7 @@ def test_qat_8da4w_linear(self):
 
     @unittest.skipIf(not TORCH_VERSION_AT_LEAST_2_4, "skipping when torch version is 2.4 or lower")
     def test_qat_8da4w_quantizer(self):
-        from torchao.quantization.prototype.qat import Int8DynActInt4WeightQATQuantizer
+        from torchao.quantization.qat import Int8DynActInt4WeightQATQuantizer
         from torchao.quantization.GPTQ import Int8DynActInt4WeightQuantizer
 
         group_size = 16
@@ -266,7 +266,7 @@ def test_qat_8da4w_quantizer(self):
 
     @unittest.skipIf(not TORCH_VERSION_AT_LEAST_2_4, "skipping when torch version is 2.4 or lower")
     def test_qat_8da4w_quantizer_meta_weights(self):
-        from torchao.quantization.prototype.qat import Int8DynActInt4WeightQATQuantizer
+        from torchao.quantization.qat import Int8DynActInt4WeightQATQuantizer
 
         with torch.device("meta"):
             m = M()
@@ -281,7 +281,7 @@ def test_qat_8da4w_quantizer_disable_fake_quant(self):
         """
         Test that 8da4w QAT with disabled fake quant matches nn.Linear in forward.
         """
-        from torchao.quantization.prototype.qat import (
+        from torchao.quantization.qat.linear import (
             Int8DynActInt4WeightQATQuantizer,
             disable_8da4w_fake_quant,
             enable_8da4w_fake_quant,
@@ -340,7 +340,7 @@ def test_qat_8da4w_quantizer_disable_fake_quant_backward(self):
         """
         Test that 8da4w QAT with disabled fake quant matches nn.Linear in backward.
         """
-        from torchao.quantization.prototype.qat import (
+        from torchao.quantization.qat.linear import (
             Int8DynActInt4WeightQATQuantizer,
             disable_8da4w_fake_quant,
         )
@@ -422,7 +422,7 @@ def _test_qat_quantized_gradients(self, quantizer):
 
     @unittest.skipIf(not TORCH_VERSION_AT_LEAST_2_4, "skipping when torch version is 2.4 or lower")
     def test_qat_8da4w_quantizer_gradients(self):
-        from torchao.quantization.prototype.qat import Int8DynActInt4WeightQATQuantizer
+        from torchao.quantization.qat import Int8DynActInt4WeightQATQuantizer
         quantizer = Int8DynActInt4WeightQATQuantizer(groupsize=16)
         self._test_qat_quantized_gradients(quantizer)
 
@@ -512,7 +512,7 @@ def test_qat_4w_primitives(self):
     @unittest.skipIf(not TORCH_VERSION_AT_LEAST_2_4, "skipping when torch version is 2.4 or lower")
     @unittest.skipIf(not _CUDA_IS_AVAILABLE, "skipping when cuda is not available")
     def test_qat_4w_linear(self):
-        from torchao.quantization.prototype.qat.linear import Int4WeightOnlyQATLinear
+        from torchao.quantization.qat.linear import Int4WeightOnlyQATLinear
         from torchao.quantization.GPTQ import WeightOnlyInt4Linear
 
         group_size = 128
@@ -539,14 +539,14 @@ def test_qat_4w_linear(self):
 
     @unittest.skipIf(not TORCH_VERSION_AT_LEAST_2_4, "skipping when torch version is 2.4 or lower")
     def test_qat_4w_quantizer_gradients(self):
-        from torchao.quantization.prototype.qat import Int4WeightOnlyQATQuantizer
+        from torchao.quantization.qat import Int4WeightOnlyQATQuantizer
         quantizer = Int4WeightOnlyQATQuantizer(groupsize=32, inner_k_tiles=8)
         self._test_qat_quantized_gradients(quantizer)
 
     @unittest.skipIf(not TORCH_VERSION_AT_LEAST_2_4, "skipping when torch version is 2.4 or lower")
     @unittest.skipIf(not _CUDA_IS_AVAILABLE, "skipping when cuda is not available")
     def test_qat_4w_quantizer(self):
-        from torchao.quantization.prototype.qat import Int4WeightOnlyQATQuantizer
+        from torchao.quantization.qat import Int4WeightOnlyQATQuantizer
         from torchao.quantization.GPTQ import Int4WeightOnlyQuantizer
 
         group_size = 32
@@ -624,7 +624,7 @@ def test_composable_qat_quantizer(self):
 
     @unittest.skipIf(not TORCH_VERSION_AT_LEAST_2_4, "skipping when torch version is 2.4 or lower")
     def test_qat_4w_embedding(self):
-        from torchao.quantization.prototype.qat import Int4WeightOnlyEmbeddingQATQuantizer
+        from torchao.quantization.qat import Int4WeightOnlyEmbeddingQATQuantizer
         model = M2()
         x = model.example_inputs()
         out = model(*x)

torchao/quantization/prototype/qat/_module_swap_api.py

@@ -41,7 +41,7 @@ For example, on a single GPU:
 ```python
 import torch
 from torchtune.models.llama3 import llama3
-from torchao.quantization.prototype.qat import Int8DynActInt4WeightQATQuantizer
+from torchao.quantization.qat import Int8DynActInt4WeightQATQuantizer
 
 # Smaller version of llama3 to fit in a single GPU
 model = llama3(

…hao/quantization/prototype/qat/README.md torchao/quantization/qat/README.mdtorchao/quantization/prototype/qat/README.md renamed to torchao/quantization/qat/README.md torchao/quantization/prototype/qat/README.md renamed to torchao/quantization/qat/README.md

@@ -2,26 +2,16 @@
     ComposableQATQuantizer,
 )
 from .linear import (
-    disable_4w_fake_quant,
-    disable_8da4w_fake_quant,
-    enable_4w_fake_quant,
-    enable_8da4w_fake_quant,
     Int4WeightOnlyQATQuantizer,
-    Int8DynActInt4WeightQATLinear,
     Int8DynActInt4WeightQATQuantizer,
 )
 from .embedding import (
     Int4WeightOnlyEmbeddingQATQuantizer,
 )
 
 __all__ = [
-    "disable_4w_fake_quant",
-    "disable_8da4w_fake_quant",
-    "enable_4w_fake_quant",
-    "enable_8da4w_fake_quant",
     "ComposableQATQuantizer",
     "Int4WeightOnlyQATQuantizer",
     "Int4WeightOnlyEmbeddingQATQuantizer"
     "Int8DynActInt4WeightQATQuantizer",
-    "Int8DynActInt4WeightQATLinear",
 ]