add support for power of 2 scaling in float8 training

Author: danielvegamyhre

torchao/float8/config.py

@@ -146,6 +146,20 @@ class Float8GemmConfig:
     use_fast_accum: bool = False
 
 
+@dataclass(frozen=True)
+class Float8ScalingFactorConfig:
+    """
+    Configuration for scaling factor used for float8 quantization.
+    """
+
+    # If this option is enabled, the scaling factor used for float8 quantization
+    # will be rounded down to the nearest power of 2. This has been shown to help
+    # reduce quantization error by avoiding rounding errors when multiplying/dividing
+    # by the scaling factor, as well as ensuring large values are quantized to the
+    # same value in the forward pass as the backward pass.
+    power_of_2_scale: bool = False
+
+
 @dataclass(frozen=True)
 class Float8LinearConfig:
     """
@@ -234,6 +248,9 @@ class Float8LinearConfig:
     # tests so that the warning does not spam the CI stdout.
     force_recompute_fp8_weight_in_bwd: bool = False
 
+    # configuration used for calculating the scaling factor used in float8 quantization.
+    scaling_factor_config: Float8ScalingFactorConfig = None
+
     def __post_init__(self):
         # Populate the additional cast overrides, if the user did not specify them
         # Note: this hacks around the frozen-ness of this dataclass

torchao/float8/float8_scaling_utils.py

@@ -12,7 +12,7 @@
 
 import torch
 
-from torchao.float8.config import ScalingGranularity
+from torchao.float8.config import Float8ScalingFactorConfig, ScalingGranularity
 from torchao.float8.distributed_utils import tensor_already_casted_to_fp8
 from torchao.float8.float8_tensor import (
     Float8Tensor,
@@ -36,6 +36,7 @@ def hp_tensor_to_float8_dynamic(
     device_mesh=None,
     scaling_granularity: ScalingGranularity = ScalingGranularity.TENSORWISE,
     axiswise_dim: Optional[int] = None,
+    scaling_factor_config: Float8ScalingFactorConfig = None,
 ) -> Float8Tensor:
     """
     Given a high precision tensor `hp_tensor`,
@@ -51,6 +52,10 @@ def hp_tensor_to_float8_dynamic(
           the 3 fwd/bwd gemms of linear
         scaling_granularity: Defines the scaling granularity
         axiswise_dim: if axiswise granularity is used, defines the dim to scale across
+        scaling_factor_config: optional configurations used to calculate the scaling factor.
+          * for row-wise scaling granularity, power of 2 scaling factor will be used by default,
+            but can be disabled via this config.
+          * for all other scaling granularities, power of 2 scaling factors are not used by default.
     """
     scale = tensor_to_scale(
         hp_tensor,
@@ -60,6 +65,11 @@ def hp_tensor_to_float8_dynamic(
         scaling_granularity,
         axiswise_dim,
     )
+
+    if _use_power_of_2_scale(scaling_granularity, scaling_factor_config):
+        # this rounds the scaling factor down to the nearest power of 2.
+        scale = torch.exp2(torch.floor(torch.log2(scale)))
+
     return hp_tensor_and_scale_to_float8(
         hp_tensor,
         scale,
@@ -70,6 +80,36 @@ def hp_tensor_to_float8_dynamic(
     )
 
 
+def _use_power_of_2_scale(
+    scaling_granularity: ScalingGranularity,
+    scaling_factor_config: Float8ScalingFactorConfig = None,
+) -> bool:
+    """
+    Returns boolean indicating if scaling factor should be rounded down to
+    the nearest power of 2.
+
+    Returns true in these cases:
+    1. The caller has enabled it in the scaling factor config.
+    2. Default on for row-wise scaling unless user has explicitly disabled
+       it in the scaling factor config.
+
+    Otherwise, returns false.
+    """
+    power_of_2_scale_enabled = (
+        scaling_factor_config is not None
+        and scaling_factor_config.power_of_2_scale is True
+    )
+    power_of_2_scale_explicitly_disabled = (
+        scaling_factor_config is not None
+        and scaling_factor_config.power_of_2_scale is False
+    )
+    use_power_of_2_scale = power_of_2_scale_enabled or (
+        scaling_granularity == ScalingGranularity.AXISWISE
+        and not power_of_2_scale_explicitly_disabled
+    )
+    return use_power_of_2_scale
+
+
 def hp_tensor_to_float8_delayed(
     hp_tensor: torch.Tensor,
     s: torch.Tensor,