Add a Float8LinearInference module to support static, dynamic, and wo quant (#287)

Summary:
# Perf script:
https://gist.github.com/drisspg/f7a553710d64cce013227a2249d582d2

## Performance

In eager this produces:

| Operation                         | Time (μs)  |
|-----------------------------------|------------|
| bf16                              | 2667.9172  |
| fp8_dynamic_activations           | 2494.7294  |
| fp8_static_activations            | 2449.1784  |
| fp8_weight_only_activations       | 4084.7190  |

With compile this produces:
| Operation                    | Time (μs)  |
|------------------------------|------------|
| bf16                         | 2547.1938  |
| fp8_dynamic_activations      | 1542.0729  |
| fp8_static_activations       | 1407.0310  |
| fp8_weight_only_activations  | 2750.6369  |

## UX

#### Dynamic activation quantization
``` Python

original_mlp = FeedForward().to("cuda", dtype=dtype)
original_mlp.reset_parameters()

dynamic_fp8_mlp = copy.deepcopy(original_mlp)

quant_config = QuantConfig(ActivationCasting.DYNAMIC)
quantize_to_float8(dynamic_fp8_mlp, quant_config)
```

#### Static activation quantization
```Python
original_mlp = FeedForward().to("cuda", dtype=dtype)
original_mlp.reset_parameters()

static_fp8_mlp = copy.deepcopy(original_mlp)
quant_config = QuantConfig(
    ActivationCasting.STATIC,
    static_quantization_scale=torch.tensor(
        [1.0], device="cuda", dtype=torch.float32
    ),
)
quantize_to_float8(static_fp8_mlp, quant_config)
```

#### Weight Only quantization
``` Python
  original_mlp = FeedForward().to("cuda", dtype=dtype)
  original_mlp.reset_parameters()

  wo_fp8_mlp = copy.deepcopy(original_mlp)
  quant_config = QuantConfig(ActivationCasting.WEIGHT_ONLY)
  quantize_to_float8(wo_fp8_mlp, quant_config)
```

All of these are using Per-Tensor scaling will add in a follow up PR row-wise scaling and likely make this the default.

Pull Request resolved: #287

Reviewed By: vkuzo

Differential Revision: D59179113

Pulled By: drisspg

fbshipit-source-id: 7938efbcbc51109d2ff7261275ca04d1b90732d3

Author: drisspg

.github/workflows/python-app.yml

@@ -25,6 +25,7 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
+        pip3 install -U --pre torch --index-url https://download.pytorch.org/whl/nightly/cu121
         pip install -e .
         pip install -e .'[dev]'
         pip install -e .'[test]'

benchmarks/profile_linear_float8.py

@@ -21,7 +21,6 @@
 from float8_experimental.float8_dynamic_linear import Float8DynamicLinear
 from float8_experimental.float8_linear import Float8Linear
 from float8_experimental.float8_linear_utils import (
-    get_float8_linear,
     linear_requires_sync,
     LinearType,
     swap_linear_with_float8_linear,

benchmarks/utils.py

@@ -5,7 +5,6 @@
 # LICENSE file in the root directory of this source tree.
 
 import collections
-import json
 import re
 
 

float8_experimental/__init__.py

@@ -5,6 +5,11 @@
 # LICENSE file in the root directory of this source tree.
 # Lets define a few top level things here
 from float8_experimental.float8_linear import Float8Linear
-from float8_experimental.float8_tensor import Float8Tensor
+from float8_experimental.float8_tensor import Float8Tensor, ScaledMMConfig
+
+# Needed to load Float8Tensor with weights_only = True
+from torch.serialization import add_safe_globals
+
+add_safe_globals([Float8Tensor, ScaledMMConfig])
 
 __all__ = ["Float8Tensor", "Float8Linear"]

float8_experimental/float8_dynamic_linear.py

@@ -62,8 +62,8 @@ class Float8DynamicLinear(torch.nn.Linear):
     def __init__(self, **super_kwargs):
         super().__init__(**super_kwargs)
 
-    def forward(self, x):
-        x_fp8 = cast_to_float8_e4m3fn(x, self.forward_config)
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        x_fp8 = cast_to_float8_e4m3fn(input, self.forward_config)
         if isinstance(self.weight, Float8Tensor):  # cast by FSDP
             w_fp8 = self.weight
         else:

float8_experimental/float8_linear.py

@@ -312,10 +312,10 @@ def float8_post_forward(self):
         self.is_amax_initialized = True
         self.amax_and_scale_synced = False
 
-    def forward(self, x):
-        self.float8_pre_forward(x)
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        self.float8_pre_forward(input)
 
-        x_fp8 = self.cast_x_to_float8(x, self.is_amax_initialized)
+        x_fp8 = self.cast_x_to_float8(input, self.is_amax_initialized)
         w_fp8 = self.cast_w_to_float8(self.weight, self.is_amax_initialized)
 
         y = torch.matmul(x_fp8, w_fp8.t())

float8_experimental/float8_linear_utils.py

@@ -6,7 +6,7 @@
 import copy
 import logging
 from enum import auto, Enum
-from typing import Callable, List, Optional, Type
+from typing import Callable, List, Optional, Type, Union
 
 import torch
 import torch.distributed as dist
@@ -97,45 +97,51 @@ def filter_out_small_unaligned_layers(size_limit: int) -> Callable[[nn.Linear],
     )
 
 
-def swap_linear_with_float8_linear(
+def swap_linear_layers(
     module: nn.Module,
-    module_cls: Type[nn.Module],
+    from_float_func: Callable[[nn.Linear], nn.Linear],
     *,
     skip_fqn_list: Optional[List[str]] = None,
-    emulate: bool = False,
     linear_layer_filter: Optional[Callable[[nn.Linear], bool]] = None,
-) -> nn.Module:
+) -> Optional[nn.Module]:
     """
-    Replaces all instances of ``torch.nn.Linear`` in ``module`` with instances
-    of ``module_cls`` (either ``Float8Linear`` or ``Float8DynamicLinear``).
+    Generic function to swap linear layers in a module with a new type of linear layer.
+
+    Note:
+        If applied to a root-level nn.Linear, the module will not be modified in place
+        and returned instead
 
     Args:
-        module (torch.nn.Module): Module to modify.
-        module_cls (Union[Type[Float8Linear], Type[Float8DynamicLinear]]): Float8 linear class for the swap.
-        skip_fqn_list (List[str], optional): If specified, a list of module FQNs to skip.
-            Linear submodules of these skipped modules will also be skipped.
-        emulate (bool): Whether to emulate the fp8 matmul logic in fp32.
-        linear_layer_filter (Optional[Callable[[nn.Linear], bool]]): If specified, only the linear layers
+        module: Module to modify.
+        from_float_func: Function that accepts a linear layer and returns a new type of linear layer.
+        skip_fqn_list: If specified, a list of module FQNs to skip.
+        linear_layer_filter: If specified, only the linear layers
             that pass the filter function will be swapped.
+        from_float_kwargs: Additional keyword arguments for from_float_func.
+
+    Returns:
+     nn.Module: The modified module with swapped linear layers.
     """
     module_names_to_skip = set(skip_fqn_list or [])
+
     if isinstance(module, nn.Linear) and (
         linear_layer_filter is None or linear_layer_filter(module)
     ):
         if len(list(module.children())) > 0:
             raise AssertionError(
                 f"Does not support a root nn.Linear with children: {module}"
             )
-        return module_cls.from_float(module, emulate=emulate)
+        return from_float_func(
+            module,
+        )
 
-    # Mark all modules to skip as visited
     root_module = module
     visited_modules = {root_module}
+
     for module_name, module in root_module.named_modules():
         if module_name in module_names_to_skip:
             visited_modules.add(module)
 
-    # Run a post-order traversal to swap linears
     def post_order_traversal(
         module: nn.Module, module_name: str, parent_module: Optional[nn.Module]
     ):
@@ -144,14 +150,15 @@ def post_order_traversal(
             if child_module not in visited_modules:
                 visited_modules.add(child_module)
                 post_order_traversal(child_module, child_module_name, module)
+
         if isinstance(module, nn.Linear) and (
             linear_layer_filter is None or linear_layer_filter(module)
         ):
             assert (
                 parent_module is not None
             ), f"Linear root module should return early: {module}"
-            float8linear_module = module_cls.from_float(module, emulate=emulate)
-            setattr(parent_module, module_name, float8linear_module)
+            new_linear_module = from_float_func(module)
+            setattr(parent_module, module_name, new_linear_module)
 
     post_order_traversal(root_module, "", None)
     # Without this explicit `del`, this set only gets deleted upon an explicit
@@ -160,6 +167,22 @@ def post_order_traversal(
     return root_module
 
 
+def swap_linear_with_float8_linear(
+    module: nn.Module,
+    module_cls: Union[Type[Float8Linear], Type[Float8DynamicLinear]],
+    *,
+    skip_fqn_list: Optional[List[str]] = None,
+    emulate: bool = False,
+    linear_layer_filter: Optional[Callable[[nn.Linear], bool]] = None,
+) -> Optional[nn.Module]:
+    return swap_linear_layers(
+        module,
+        lambda m: module_cls.from_float(m, emulate=emulate),
+        skip_fqn_list=skip_fqn_list,
+        linear_layer_filter=linear_layer_filter,
+    )
+
+
 def get_float8_layers(model: torch.nn.Module):
     """Iterates through the model and returns all the Float8Linear layers.
     Args:

float8_experimental/float8_tensor.py

@@ -266,6 +266,7 @@ def to_float8(
             scale: the scale to use to convert the tensor
             float8_dtype: the float8 dtype to use
             amax_buffer: a buffer to store the amax value in prior to conversion
+            mm_config: Defines the configuration for the scaled_mm
 
         Returns:
             Float8Tensor: a float8 tensor

float8_experimental/float8_utils.py

@@ -139,7 +139,7 @@ def to_fp8_saturated(x: torch.Tensor, float8_dtype: torch.dtype):
         raise ValueError(f"Unsupported float8_dtype: {float8_dtype}")
 
 
-def compute_error(x: torch.Tensor, y: torch.Tensor):
+def compute_error(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
     """Computes the error between two tensors in dB.
 
     For more details see: