integrate new float8 quantization primitives into AQT

ghstack-source-id: c1deeeb
ghstack-comment-id: 2608090492
Pull Request resolved: #1598

Author: danielvegamyhre

torchao/dtypes/affine_quantized_tensor.py

@@ -4,27 +4,22 @@
 
 import torch
 
-from torchao.dtypes.utils import (
-    AQTTensorImpl,
-    Layout,
-    PlainLayout,
-)
+from torchao.dtypes.utils import AQTTensorImpl, Layout, PlainLayout
 from torchao.quantization.quant_primitives import (
     FP8_TYPES,
     MappingType,
     ZeroPointDomain,
     choose_qparams_affine,
+    choose_qparams_affine_float8,
     choose_qparams_affine_floatx,
     choose_qparams_and_quantize_affine_hqq,
     dequantize_affine,
     dequantize_affine_floatx,
     quantize_affine,
+    quantize_affine_float8,
     quantize_affine_floatx,
 )
-from torchao.utils import (
-    TORCH_VERSION_AT_LEAST_2_5,
-    TorchAOBaseTensor,
-)
+from torchao.utils import TORCH_VERSION_AT_LEAST_2_5, TorchAOBaseTensor
 
 logger = logging.getLogger(__name__)
 aten = torch.ops.aten
@@ -422,6 +417,39 @@ def from_hp_to_fpx(
         tensor_impl = tensor_impl_ctr(floatx_packed, scale, None, _layout)
         return cls(tensor_impl, block_size, original_shape, dtype=input_float.dtype)
 
+    @classmethod
+    def from_hp_to_float8(
+        cls,
+        input_float: torch.Tensor,
+        target_dtype: torch.dtype,
+        block_size: Tuple[int, ...],
+        _layout: Layout = PlainLayout(),
+    ):
+        assert target_dtype in FP8_TYPES, f"Unsupported dtype {target_dtype} for float8"
+
+        # to avoid circular dependency
+        from torchao.dtypes.floatx import Float8AQTTensorImpl
+
+        original_shape = input_float.shape
+        scale = choose_qparams_affine_float8(
+            input_float,
+            target_dtype,
+            target_dtype,
+        )
+        fp8_data = quantize_affine_float8(
+            input_float,
+            scale,
+            target_dtype,
+        )
+        fp8_data = _layout.post_process(fp8_data)
+        tensor_impl = Float8AQTTensorImpl(fp8_data, scale, None, _layout)
+        return cls(
+            tensor_impl,
+            block_size,
+            original_shape,
+            dtype=input_float.dtype,
+        )
+
     @property
     def _layout(self) -> Layout:
         return self.tensor_impl._layout
@@ -477,6 +505,7 @@ def _apply_fn_to_data(self, fn):
 to_affine_quantized_intx_static = AffineQuantizedTensor.from_hp_to_intx_static
 to_affine_quantized_floatx = AffineQuantizedTensor.from_hp_to_floatx
 to_affine_quantized_floatx_static = AffineQuantizedTensor.from_hp_to_floatx_static
+to_affine_quantized_float8 = AffineQuantizedTensor.from_hp_to_float8
 # experimental will be merged in to floatx
 to_affine_quantized_fpx = AffineQuantizedTensor.from_hp_to_fpx
 

torchao/quantization/quant_primitives.py

@@ -5,13 +5,15 @@
 # LICENSE file in the root directory of this source tree.
 
 import math
-from enum import auto, Enum
+from enum import Enum, auto
 from typing import Callable, Dict, List, Optional, Tuple, Union
 
 import torch
 
 from torchao.float8.float8_utils import (
     ScalingGranularity,
+)
+from torchao.float8.float8_utils import (
     tensor_to_scale as tensor_to_float8_scale,
 )
 from torchao.prototype.custom_fp_utils import (
@@ -20,11 +22,11 @@
     _n_ones,
 )
 from torchao.utils import (
-    _is_float8_type,
-    _register_custom_op,
     TORCH_VERSION_AT_LEAST_2_3,
     TORCH_VERSION_AT_LEAST_2_5,
     TORCH_VERSION_AT_LEAST_2_6,
+    _is_float8_type,
+    _register_custom_op,
 )
 
 __all__ = [