Support power of 2 scaling factors in float8 training and use e4m3 everywhere

test/float8/test_base.py

@@ -164,7 +164,8 @@ def test_transpose(self):
 
     @pytest.mark.parametrize("shape", [(8, 16), (4, 8, 16), (2, 4, 8, 16)])
     @pytest.mark.parametrize("axiswise_dim", [0, -1])
-    def test_axiswise_dynamic_cast(self, shape, axiswise_dim):
+    @pytest.mark.parametrize("power_of_2_scale", [True, False])
+    def test_axiswise_dynamic_cast(self, shape, axiswise_dim, power_of_2_scale):
         a = torch.randn(*shape, dtype=torch.bfloat16)
         linear_mm_config = LinearMMConfig()
         a_fp8 = hp_tensor_to_float8_dynamic(
@@ -173,6 +174,7 @@ def test_axiswise_dynamic_cast(self, shape, axiswise_dim):
             linear_mm_config,
             scaling_granularity=ScalingGranularity.AXISWISE,
             axiswise_dim=axiswise_dim,
+            power_of_2_scale=power_of_2_scale,
         )
         a_dq = a_fp8.to_original_precision()
         sqnr = compute_error(a, a_dq)

test/float8/test_compile.py

@@ -45,11 +45,7 @@
     hp_tensor_to_float8_delayed,
     hp_tensor_to_float8_dynamic,
 )
-from torchao.float8.float8_tensor import (
-    GemmInputRole,
-    LinearMMConfig,
-    ScaledMMConfig,
-)
+from torchao.float8.float8_tensor import GemmInputRole, LinearMMConfig, ScaledMMConfig
 from torchao.float8.float8_utils import config_has_stateful_scaling
 from torchao.float8.stateful_float8_linear import StatefulFloat8Linear
 from torchao.testing.float8.test_utils import get_test_float8_linear_config
@@ -420,13 +416,21 @@ def test_sync_amax_func_cuda_graph_success():
         torch.float16,
     ],
 )
-def test_dynamic_scale_numeric_parity(dtype: torch.dtype):
+@pytest.mark.parametrize(
+    "power_of_2_scale",
+    [
+        True,
+        False,
+    ],
+)
+def test_dynamic_scale_numeric_parity(dtype: torch.dtype, power_of_2_scale: bool):
     scaling_type_weight = ScalingType.DYNAMIC
     torch.manual_seed(42)
     hp_tensor1 = torch.randn(16, 16, device="cuda", dtype=dtype)
     hp_tensor2 = hp_tensor1.detach().clone()
     float8_config = Float8LinearConfig(
         cast_config_weight=CastConfig(scaling_type=scaling_type_weight),
+        power_of_2_scale=power_of_2_scale,
     )
     linear_mm_config = LinearMMConfig(
         # output
@@ -456,13 +460,15 @@ def test_dynamic_scale_numeric_parity(dtype: torch.dtype):
         e4m3_dtype,
         linear_mm_config,
         gemm_input_role=GemmInputRole.WEIGHT,
+        power_of_2_scale=float8_config.power_of_2_scale,
     )
     torch._dynamo.reset()
     float8_compile = torch.compile(hp_tensor_to_float8_dynamic)(
         hp_tensor2,
         e4m3_dtype,
         linear_mm_config,
         gemm_input_role=GemmInputRole.WEIGHT,
+        power_of_2_scale=float8_config.power_of_2_scale,
     )
     assert torch.equal(float8_eager._scale, float8_compile._scale)
     assert torch.equal(float8_eager._data, float8_compile._data)

torchao/float8/config.py

@@ -234,6 +234,13 @@ class Float8LinearConfig:
     # tests so that the warning does not spam the CI stdout.
     force_recompute_fp8_weight_in_bwd: bool = False
 
+    # 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 passes.
+    power_of_2_scale: bool = False
+
     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
@@ -336,6 +343,8 @@ def recipe_name_to_linear_config(
             cast_config_input=cc_i,
             cast_config_weight=cc_w,
             cast_config_grad_output=cc_go,
+            # enable power of 2 scaling factors by default for row-wise scaling
+            power_of_2_scale=True,
         )
 
     elif recipe_name is Float8LinearRecipeName.LW_AXISWISE_WITH_GW_HP:

torchao/float8/float8_linear.py

@@ -96,6 +96,7 @@ def forward(
                 axiswise_dim=get_maybe_axiswise_dim(
                     -1, c.cast_config_input.scaling_granularity
                 ),
+                power_of_2_scale=c.power_of_2_scale,
             )
 
         if tensor_already_casted_to_fp8(weight_hp_t):
@@ -112,6 +113,7 @@ def forward(
                 axiswise_dim=get_maybe_axiswise_dim(
                     0, c.cast_config_weight.scaling_granularity
                 ),
+                power_of_2_scale=c.power_of_2_scale,
             )
 
         # the reshapes are needed in order to make the shapes compatible with
@@ -151,6 +153,7 @@ def backward(ctx, grad_output):
                 axiswise_dim=get_maybe_axiswise_dim(
                     -1, c.cast_config_grad_output.scaling_granularity
                 ),
+                power_of_2_scale=c.power_of_2_scale,
             )
 
         if tensor_already_casted_to_fp8(weight_hp_t):
@@ -181,6 +184,7 @@ def backward(ctx, grad_output):
                 axiswise_dim=get_maybe_axiswise_dim(
                     -1, c.cast_config_weight_for_grad_input.scaling_granularity
                 ),
+                power_of_2_scale=c.power_of_2_scale,
             )
 
         grad_input = torch.mm(
@@ -216,6 +220,7 @@ def backward(ctx, grad_output):
                 axiswise_dim=get_maybe_axiswise_dim(
                     0, c.cast_config_grad_output_for_grad_weight.scaling_granularity
                 ),
+                power_of_2_scale=c.power_of_2_scale,
             )
 
         if tensor_already_casted_to_fp8(input_hp_reshaped):
@@ -233,6 +238,7 @@ def backward(ctx, grad_output):
                 axiswise_dim=get_maybe_axiswise_dim(
                     0, c.cast_config_input_for_grad_weight.scaling_granularity
                 ),
+                power_of_2_scale=c.power_of_2_scale,
             )
 
         grad_weight = torch.mm(

torchao/float8/float8_scaling_utils.py

@@ -36,6 +36,7 @@ def hp_tensor_to_float8_dynamic(
     device_mesh=None,
     scaling_granularity: ScalingGranularity = ScalingGranularity.TENSORWISE,
     axiswise_dim: Optional[int] = None,
+    power_of_2_scale: bool = False,
 ) -> Float8Tensor:
     """
     Given a high precision tensor `hp_tensor`,
@@ -51,6 +52,7 @@ 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
+        power_of_2_scale: if true, round scaling factor down to the nearest power of 2.
     """
     scale = tensor_to_scale(
         hp_tensor,
@@ -59,6 +61,7 @@ def hp_tensor_to_float8_dynamic(
         device_mesh,
         scaling_granularity,
         axiswise_dim,
+        power_of_2_scale,
     )
     return hp_tensor_and_scale_to_float8(
         hp_tensor,

torchao/float8/float8_utils.py

@@ -10,11 +10,7 @@
 import torch.distributed as dist
 from torch.distributed._functional_collectives import AsyncCollectiveTensor, all_reduce
 
-from torchao.float8.config import (
-    Float8LinearConfig,
-    ScalingGranularity,
-    ScalingType,
-)
+from torchao.float8.config import Float8LinearConfig, ScalingGranularity, ScalingType
 
 # Helpful visualizer for debugging (only supports fp32):
 # https://www.h-schmidt.net/FloatConverter/IEEE754.html
@@ -33,11 +29,14 @@
 
 
 @torch.no_grad()
-def amax_to_scale(amax: torch.Tensor, float8_dtype: torch.dtype):
+def amax_to_scale(
+    amax: torch.Tensor, float8_dtype: torch.dtype, power_of_2_scale: bool = False
+):
     """Converts the amax value of a tensor to the fp8 scale.
     Args:
         amax: The amax value of the tensor.
         float8_dtype: The float8 dtype.
+        power_of_2_scale: if true, round scaling factor down to the nearest power of 2.
     """
     # torch.compile and eager show different numerics for 1.0 / float32,
     # upcast to float64 to ensure same numeric between compile and eager
@@ -46,7 +45,9 @@ def amax_to_scale(amax: torch.Tensor, float8_dtype: torch.dtype):
         res = torch.finfo(float8_dtype).max / torch.clamp(amax, min=EPS)
     else:
         raise ValueError(f"Unsupported float8_dtype: {float8_dtype}")
-
+    if power_of_2_scale:
+        # rounds down to the nearest power of 2.
+        res = torch.exp2(torch.floor(torch.log2(res)))
     return res.to(torch.float32)
 
 
@@ -119,21 +120,33 @@ def tensor_to_amax(
 
 @torch.no_grad()
 def tensor_to_scale(
-    x: torch.Tensor,
+    hp_tensor: torch.Tensor,
     float8_dtype: torch.dtype,
     reduce_amax: bool = False,
     device_mesh=None,
     scaling_granularity: ScalingGranularity = ScalingGranularity.TENSORWISE,
     axiswise_dim: Optional[int] = None,
+    power_of_2_scale: bool = False,
 ) -> torch.Tensor:
+    """
+    Compute scaling factor for the given high precision tensor.
+
+    Args:
+        hp_tensor: high precision tensor
+        float8_dtype: the float8 dtype to use
+        reduce_amax: whether to reduce the max(abs(hp_tensor)) value across distributed ranks
+        scaling_granularity: Defines the scaling granularity
+        axiswise_dim: if axiswise granularity is used, defines the dim to scale across
+        power_of_2_scale: if true, round scaling factor down to the nearest power of 2.
+    """
     amax = tensor_to_amax(
-        x,
+        hp_tensor,
         reduce_amax,
         device_mesh,
         scaling_granularity,
         axiswise_dim,
     )
-    return amax_to_scale(amax, float8_dtype)
+    return amax_to_scale(amax, float8_dtype, power_of_2_scale=power_of_2_scale)
 
 
 def to_fp8_saturated(x: torch.Tensor, float8_dtype: torch.dtype):