[wip] add scaling granularity

Summary:

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

ghstack-source-id: d66074dab84512d32b149f9df79987c5133fd293
Pull Request resolved: #338

float8_experimental/config.py

@@ -21,6 +21,18 @@ def short_str(self):
             return "dyn"
 
 
+class ScalingGranularity(enum.Enum):
+    """
+    Defines the granularity of scaling strategies for casting to float8
+    """
+
+    # A single scaling factor for the entire tensor
+    TENSORWISE = "tensorwise"
+    # Scaling factors computed along one axis of the tensor, reducing it to 
+    # size 1.
+    AXISWISE = "axiswise"
+
+
 @dataclass(frozen=True)
 class CastConfig:
     """

float8_experimental/float8_dynamic_utils.py

@@ -4,8 +4,11 @@
 # 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 Optional, Tuple, Union
+
 import torch
 
+from float8_experimental.config import ScalingGranularity
 from float8_experimental.float8_tensor import (
     Float8Tensor,
     GemmInputRole,
@@ -52,10 +55,12 @@ def cast_to_float8_e4m3_dynamic(
     linear_mm_config: LinearMMConfig,
     reduce_amax: bool = False,
     gemm_input_role: GemmInputRole = GemmInputRole.INPUT,
+    granularity: ScalingGranularity = ScalingGranularity.TENSORWISE,
+    dim: Optional[Union[int, Tuple[int]]] = None,
 ) -> Float8Tensor:
     if tensor_already_casted_to_fp8(inpt_tensor):
         return inpt_tensor
-    scale = tensor_to_scale(inpt_tensor, e4m3_dtype, reduce_amax)
+    scale = tensor_to_scale(inpt_tensor, e4m3_dtype, reduce_amax, granularity, dim)
     return Float8Tensor.to_float8(
         inpt_tensor,
         scale,

float8_experimental/float8_tensor.py

@@ -290,12 +290,6 @@ def __new__(
         linear_mm_config: Optional[LinearMMConfig],
         gemm_input_role: Optional[GemmInputRole] = GemmInputRole.INPUT,
     ):
-        assert (
-            scale.numel() == 1
-        ), "Scale should contain a single value, but got: {} elements".format(
-            scale.numel()
-        )
-
         self = torch.Tensor._make_wrapper_subclass(
             cls,
             data.size(),

float8_experimental/float8_utils.py

@@ -4,9 +4,10 @@
 # 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
+from float8_experimental.config import ScalingGranularity
 
 import torch
 import torch.distributed as dist
@@ -100,8 +101,18 @@ def amax_history_to_scale_stack(
 
 
 @torch.no_grad()
-def tensor_to_amax(x: torch.Tensor, reduce_amax: bool = False) -> torch.Tensor:
-    amax = torch.max(torch.abs(x))
+def tensor_to_amax(
+    x: torch.Tensor, 
+    reduce_amax: bool = False,
+    granularity: ScalingGranularity = ScalingGranularity.AXISWISE,
+    dim: Optional[Union[int, Tuple[int]]] = None,
+) -> torch.Tensor:
+    if granularity is ScalingGranularity.TENSORWISE:
+        amax = torch.max(torch.abs(x))
+    else:
+        assert granularity is ScalingGranularity.AXISWISE, "unsupported"
+        assert dim is not None, "unsupported"
+        amax = torch.amax(torch.abs(x), dim=dim, keepdim=True)
 
     # If the user asked for distributed reduction, do it.
     # If the user did not ask for it, assume that it will
@@ -114,9 +125,13 @@ def tensor_to_amax(x: torch.Tensor, reduce_amax: bool = False) -> torch.Tensor:
 
 @torch.no_grad()
 def tensor_to_scale(
-    x: torch.Tensor, float8_dtype: torch.dtype, reduce_amax: bool = False
+    x: torch.Tensor, 
+    float8_dtype: torch.dtype, 
+    reduce_amax: bool = False,
+    granularity: ScalingGranularity = ScalingGranularity.AXISWISE,
+    dim: Optional[Union[int, Tuple[int]]] = None,
 ) -> torch.Tensor:
-    amax = tensor_to_amax(x, reduce_amax=reduce_amax)
+    amax = tensor_to_amax(x, reduce_amax=reduce_amax, granularity=granularity, dim=dim)
     return amax_to_scale(amax, float8_dtype, x.dtype)
 
 

test/test_base.py

@@ -143,6 +143,17 @@ def test_weights_only_load(self):
         buffer.seek(0)
         _ = torch.load(buffer, weights_only=True)
 
+    def test_axiswise_dynamic_cast(self):
+        a = torch.randn(16, 32, dtype=torch.bfloat16)
+        linear_mm_config = LinearMMConfig()
+        a_fp8 = cast_to_float8_e4m3_dynamic(
+            a, 
+            linear_mm_config, 
+            granularity=ScalingGranularity.AXISWISE, 
+            dim=0,
+        )
+        print(a_fp8)
+
 
 class TestFloat8Linear:
     def _test_linear_impl(