Add rowwwise scaling to Float8Inference module

ghstack-source-id: 20cfa0a
Pull Request resolved: #305

Author: drisspg

float8_experimental/float8_python_api.py

@@ -55,7 +55,7 @@ def addmm_float8_unwrapped(
         a_data,
         b_data,
         scale_a=a_inverse_scale,
-        scale_b=b_inverse_scale,
+        scale_b=b_inverse_scale.T,
         bias=bias,
         scale_result=output_scale,
         out_dtype=output_dtype,

float8_experimental/float8_tensor.py

@@ -109,7 +109,7 @@ def to_fp8_no_autograd(
         mm_config: Defines the configuration for the scaled_mm
     """
 
-    x_scaled = x * x_scale
+    x_scaled = x * x_scale.to(dtype=x.dtype)
     bits_fp8 = to_fp8_saturated(x_scaled, float8_dtype)
 
     if isinstance(bits_fp8, DTensor):
@@ -195,7 +195,9 @@ class FromFloat8ConstrFunc(torch.autograd.Function):
 
     @staticmethod
     def forward(ctx, tensor):
-        return tensor._data.to(tensor._orig_dtype) / tensor._scale
+        return tensor._data.to(tensor._orig_dtype) / tensor._scale.to(
+            tensor._orig_dtype
+        )
 
     @staticmethod
     def backward(ctx, g):
@@ -253,11 +255,11 @@ def __init__(
         orig_dtype: torch.dtype,
         mm_config: Optional[ScaledMMConfig],
     ):
-        assert (
-            scale.numel() == 1
-        ), "Scale should contain a single value, but got: {} elements".format(
-            scale.numel()
-        )
+        # assert (
+        #     scale.numel() == 1
+        # ), "Scale should contain a single value, but got: {} elements".format(
+        #     scale.numel()
+        # )
 
         self._data = data
         self._scale = scale

float8_experimental/float8_utils.py

@@ -4,7 +4,7 @@
 # This source code is licensed under the BSD 3-Clause license found in the
 # LICENSE file in the root directory of this source tree.
 
-from typing import Iterable, Literal, Tuple, Union
+from typing import Iterable, Literal, Optional, Tuple, Union
 
 import float8_experimental.config as config
 import torch
@@ -32,6 +32,12 @@
 e5m2_dtype = torch.float8_e5m2 if not config.use_fnuz_dtype else torch.float8_e5m2fnuz
 
 
+def get_supported_granularity():
+    from float8_experimental.float8_tensor import ScalingGranularity
+
+    return [ScalingGranularity.TensorWise, ScalingGranularity.AxisWise]
+
+
 @torch.no_grad()
 def amax_to_scale(
     amax: torch.Tensor, float8_dtype: torch.dtype, orig_dtype: torch.dtype
@@ -103,20 +109,34 @@ def amax_history_to_scale_stack(
 def tensor_to_amax(
     x: torch.Tensor,
     scaling_granularity,
+    dim: Optional[int] = None,
     reduce_amax: bool = False,
 ) -> torch.Tensor:
     """Calculates the amax of a tensor.
     Args:
         x: The tensor to calculate the amax for.
         scaling_granularity: The granularity of with which to calcualte the tensor amax
+        dim: The dimension along which to calculate the amax. This is only used if scaling_granularity is AxisWise.
         reduce_amax: Whether to perform a distributed reduction on the amax.
     """
     from float8_experimental.float8_tensor import ScalingGranularity
 
-    assert (
-        scaling_granularity == ScalingGranularity.TensorWise
-    ), f"Currently only TensorWise is supported for but given scaling_granularity: {scaling_granularity}"
-    amax = torch.max(torch.abs(x))
+    supported_granularities = get_supported_granularity()
+
+    if scaling_granularity not in supported_granularities:
+        raise ValueError(
+            f"Currently only {supported_granularities} are supported. Given scaling_granularity: {scaling_granularity}"
+        )
+
+    if scaling_granularity == ScalingGranularity.TensorWise:
+        amax = torch.max(torch.abs(x))
+    elif scaling_granularity == ScalingGranularity.AxisWise:
+        if dim is None:
+            raise ValueError("For AxisWise scaling, a dim must be passed in!")
+        amax = torch.max(torch.abs(x), dim=dim, keepdim=True).values
+    else:
+        # This should never be reached due to the earlier check, but it's here for completeness
+        raise ValueError(f"Unsupported scaling_granularity: {scaling_granularity}")
 
     # If the user asked for distributed reduction, do it.
     # If the user did not ask for it, assume that it will
@@ -132,16 +152,20 @@ def tensor_to_scale(
     x: torch.Tensor,
     float8_dtype: torch.dtype,
     scaling_granularity,
+    dim: Optional[int] = None,
     reduce_amax: bool = False,
+    collapse_leading_dims: bool = False,
 ) -> torch.Tensor:
     """Calculates the scale that will be used for quantization to Float8Tensor
     Args:
         x: The tensor to calculate the scale for.
         float8_dtype: The Float8 dtype to use.
         scaling_granularity: The granularity of the scale. See ScalingGranularity for more details.
+        dim: The dimension along which to calculate the scale. This is only used if scaling_granularity is AxisWise.
         reduce_amax: Whether to perform a distributed reduction on the amax.
+        collapse_leading_dims: Whether to collapse leading dimensions of the tensor.
     """
-    amax = tensor_to_amax(x, scaling_granularity, reduce_amax=reduce_amax)
+    amax = tensor_to_amax(x, scaling_granularity, dim=dim, reduce_amax=reduce_amax)
     return amax_to_scale(amax, float8_dtype, x.dtype)
 
 

float8_experimental/inference.py

@@ -25,7 +25,13 @@
     tensor_already_casted_to_fp8,
     to_fp8_no_autograd,
 )
-from float8_experimental.float8_utils import e4m3_dtype, tensor_to_scale
+from float8_experimental.float8_utils import (
+    e4m3_dtype,
+    get_supported_granularity,
+    tensor_to_scale,
+)
+
+SUPPORTED_GRANULARITY = get_supported_granularity()
 
 
 class ActivationCasting(Enum):
@@ -75,7 +81,7 @@ def __init__(
         # FP8 specific arguments
         quant_config: QuantConfig,
         forward_config: ScaledMMConfig,
-        scaling_granularity: ScalingGranularity,
+        scaling_granularity: Optional[ScalingGranularity],
         # nn.Linear arguments
         in_features: int,
         out_features: int,
@@ -86,7 +92,26 @@ def __init__(
         # Construct the superclass this will create dummy weights and biases
         super().__init__(in_features, out_features, bias, device, dtype)
         self.forward_config = forward_config
-        self.scaling_granularity = scaling_granularity
+        if scaling_granularity is None:
+            self.scaling_granularity = (
+                ScalingGranularity.AxisWise
+                if dtype == torch.bfloat16
+                and quant_config.static_quantization_scale is None
+                else ScalingGranularity.TensorWise
+            )
+        else:
+            assert (
+                scaling_granularity in SUPPORTED_GRANULARITY
+            ), f"scaling_granularity must be in {SUPPORTED_GRANULARITY} but got {scaling_granularity}"
+            if (
+                scaling_granularity == ScalingGranularity.AxisWise
+                and dtype != torch.bfloat16
+            ):
+                raise ValueError(
+                    "AxisWise scaling granularity is only supported for bfloat16."
+                )
+            self.scaling_granularity = scaling_granularity
+
         self.activation_casting = quant_config.activation_casting
         if self.activation_casting == ActivationCasting.STATIC:
             self.register_buffer(
@@ -101,13 +126,19 @@ def forward(self, input: torch.Tensor) -> torch.Tensor:
                 input, self.weight.to_original_precision()
             )
 
+        # TODO we arent folding leading dims yet, but need it to calculate the proper scale.. this sucks
+        original_m = input.shape[:-1]
+        input = input.view(-1, input.shape[-1])
+
         x_fp8 = cast_to_float8_e4m3_inference(
             input,
             self.forward_config,
             static_quantization_scale=self.static_quantization_scale,
             scaling_granularity=self.scaling_granularity,
         )
-        return torch.nn.functional.linear(x_fp8, self.weight, self.bias)
+        return torch.nn.functional.linear(x_fp8, self.weight, self.bias).view(
+            *original_m, -1
+        )
 
     # Builder functions for Float8LinearInference
     def quantize_weight(self, dtype: torch.dtype = e4m3_dtype) -> None:
@@ -124,7 +155,12 @@ def quantize_weight(self, dtype: torch.dtype = e4m3_dtype) -> None:
         assert not isinstance(
             self.weight, Float8Tensor
         ), "Weight has already been quantized, cannot quantize again."
-        scale = tensor_to_scale(self.weight, dtype, self.scaling_granularity)
+
+        # For weight tensors + AxisWise we calculate scales along columns
+        dim = None
+        if self.scaling_granularity == ScalingGranularity.AxisWise:
+            dim = 1
+        scale = tensor_to_scale(self.weight, dtype, self.scaling_granularity, dim=dim)
         quantized_weight = to_fp8_no_autograd(
             self.weight, scale, dtype, self.forward_config
         )
@@ -143,19 +179,20 @@ def from_float(
         module: nn.Module,
         quant_config: QuantConfig,
         use_fast_accum: bool,
+        scaling_granularity: Optional[ScalingGranularity],
     ) -> "Float8InferenceLinear":
         """
         Create an nn.Linear with fp8 compute from another nn.Linear
 
         Args:
             mod (torch.nn.Linear): nn.Linear to convert
             quant_config (QuantConfig): Configuration for the weight and activation casting
+            use_fast_accum (bool): Whether to enable fast accumulation for the Float8InferenceLinear.
+            scaling_granularity: The granularity of the scale. See ScalingGranularity for more details.
         """
         forward_config = ScaledMMConfig(
             False, use_fast_accum, pad_inner_dim=config.pad_inner_dim
         )
-        # TODO: For now hardcode TensorWise scaling
-        scaling_granularity = ScalingGranularity.TensorWise
         linear = cls(
             quant_config,
             forward_config,
@@ -164,6 +201,7 @@ def from_float(
             module.out_features,
             False,
             device=torch.device("meta"),
+            dtype=module.weight.dtype,
         )
         linear.set_weight_and_bias(module.weight, module.bias)
         linear.quantize_weight()
@@ -194,18 +232,29 @@ def cast_to_float8_e4m3_inference(
     """
     if tensor_already_casted_to_fp8(inpt_tensor):
         return inpt_tensor
+
+    # For input tensors + AxisWise we calculate scales along rows
+    dim = None
+    if scaling_granularity == ScalingGranularity.AxisWise:
+        dim = 1
+
     scale = (
         static_quantization_scale
         if static_quantization_scale is not None
         else tensor_to_scale(
-            inpt_tensor, e4m3_dtype, scaling_granularity, reduce_amax=reduce_amax
+            inpt_tensor,
+            e4m3_dtype,
+            scaling_granularity,
+            dim=dim,
+            reduce_amax=reduce_amax,
         )
     )
     return Float8Tensor.to_float8(
         inpt_tensor,
         scale,
         e4m3_dtype,
         mm_config=mm_config,
+        scaling_granularity=scaling_granularity,
     )
 
 
@@ -215,6 +264,7 @@ def quantize_to_float8(
     *,
     skip_fqn_list: Optional[List[str]] = None,
     use_fast_accum: bool = True,
+    scaling_granularity: Optional[ScalingGranularity] = None,
 ) -> Optional[nn.Module]:
     """
     Converts torch.nn.Linear layers in the given module to Float8InferenceLinear.
@@ -228,6 +278,7 @@ def quantize_to_float8(
         quant_config (QuantConfig): Quantization configuration for Float8 conversion.
         skip_fqn_list (List[str], optional): List of module FQNs to skip during conversion.
         use_fast_accum : Whether to enable fast accumulation for the Float8InferenceLinear. Defaults to True.
+        scaling_granularity: The granularity of the scale. See ScalingGranularity for more details.
 
     Returns:
         nn.Module: The modified module with applicable Linear layers converted to Float8.
@@ -237,6 +288,8 @@ def quantize_to_float8(
     """
     return swap_linear_layers(
         module,
-        lambda m: Float8InferenceLinear.from_float(m, quant_config, use_fast_accum),
+        lambda m: Float8InferenceLinear.from_float(
+            m, quant_config, use_fast_accum, scaling_granularity
+        ),
         skip_fqn_list=skip_fqn_list,
     )