support delayed scaling of weight in float8 all-gather (#312)

Summary:
Pull Request resolved: #312

Adds support for delayed scaling in FSDP2 float8 all-gather. In detail:
1. add `WeightWithDelayedFloat8CastTensor`, note that we don't reuse
   code with the dynamic version because I'd rather not deal with
   plumbing optional tensors through dynamo. We can try that in a
   separate PR later.
2. wire `Float8Linear` to use (1)
3. add weight amax syncing back, since we need it for float8 all-gather
4. add test coverage for eager mode numerics

Next up (in separate PRs) will be training run validation for numerics, and
taking a look at performance.

Reviewed By: awgu

Differential Revision: D59685258

fbshipit-source-id: 9ff18d7649cc6e0e3c9e2a64a30a5ff8bc4108be

float8_experimental/float8_linear.py

@@ -34,7 +34,10 @@
     tensor_to_amax,
 )
 
-from float8_experimental.fsdp_utils import WeightWithDynamicFloat8CastTensor
+from float8_experimental.fsdp_utils import (
+    WeightWithDelayedFloat8CastTensor,
+    WeightWithDynamicFloat8CastTensor,
+)
 
 
 def _maybe_initialize_amaxes_scales_for_float8_cast(
@@ -316,28 +319,30 @@ def cast_w_to_float8(
         self, w: torch.Tensor, is_amax_initialized: bool
     ) -> torch.Tensor:
         if self.scaling_type_w is TensorScalingType.DELAYED:
-            scale_fn_name = self.recipe.scale_fn_name
-            _maybe_initialize_amaxes_scales_for_float8_cast(
-                w,
-                self.fp8_amax_w,
-                self.fp8_amax_history_w,
-                self.fp8_scale_w,
-                scale_fn_name,
-                e4m3_dtype,
-                is_amax_initialized,
-                reduce_amax=False,
-            )
-
-            w_fp8 = Float8Tensor.to_float8(
-                w,
-                self.fp8_scale_w,
-                e4m3_dtype,
-                self.fp8_amax_w,
-                self.forward_config,
-            )
+            if isinstance(self.weight, Float8Tensor):  # cast by FSDP
+                w_fp8 = self.weight
+            else:
+                scale_fn_name = self.recipe.scale_fn_name
+                _maybe_initialize_amaxes_scales_for_float8_cast(
+                    w,
+                    self.fp8_amax_w,
+                    self.fp8_amax_history_w,
+                    self.fp8_scale_w,
+                    scale_fn_name,
+                    e4m3_dtype,
+                    is_amax_initialized,
+                    reduce_amax=False,
+                )
+
+                w_fp8 = Float8Tensor.to_float8(
+                    w,
+                    self.fp8_scale_w,
+                    e4m3_dtype,
+                    self.fp8_amax_w,
+                    self.forward_config,
+                )
         else:
             assert self.scaling_type_w is TensorScalingType.DYNAMIC
-            # TODO(future): also support FSDP integration in delayed scaling path
             if isinstance(self.weight, Float8Tensor):  # cast by FSDP
                 w_fp8 = self.weight
             else:
@@ -436,18 +441,36 @@ def from_float(
                 scaling_type_dL_dY=scaling_type_dL_dY,
                 emulate=emulate,
             )
-        if (
-            scaling_type_w == TensorScalingType.DYNAMIC
-            and config.enable_fsdp_fp8_all_gather
-        ):
-            new_mod.weight = torch.nn.Parameter(
-                WeightWithDynamicFloat8CastTensor(mod.weight, new_mod.forward_config)
-            )
-        else:
-            assert not config.enable_fsdp_fp8_all_gather, "unsupported"
-            new_mod.weight = mod.weight
+        new_mod.weight = mod.weight
         new_mod.bias = mod.bias
         # need to create buffers again when moving from meta device to
         # real device
         new_mod.create_buffers()
+
+        # If FSDP float8 all-gather is on, wrap the weight in a float8-aware
+        # tensor subclass. This must happen last because:
+        # 1. weight needs to be on the correct device to create the buffers
+        # 2. buffers need to be already created for the delayed scaling version
+        #    of the weight wrapper to be initialized
+        if config.enable_fsdp_fp8_all_gather:
+            if scaling_type_w is TensorScalingType.DYNAMIC:
+                new_mod.weight = torch.nn.Parameter(
+                    WeightWithDynamicFloat8CastTensor(
+                        new_mod.weight,
+                        new_mod.forward_config,
+                    )
+                )
+            else:
+                assert scaling_type_w is TensorScalingType.DELAYED
+                new_mod.weight = torch.nn.Parameter(
+                    WeightWithDelayedFloat8CastTensor(
+                        new_mod.weight,
+                        new_mod.fp8_amax_w,
+                        new_mod.fp8_amax_history_w,
+                        new_mod.fp8_scale_w,
+                        new_mod.forward_config,
+                        new_mod.is_amax_initialized,
+                    )
+                )
+
         return new_mod

float8_experimental/float8_linear_utils.py

@@ -289,11 +289,10 @@ def inner_func():
         ), "Mismatched lengths of amax tensors."
 
         if dist.is_initialized():
-            # Combine all the amax tensors into one tensor and reduce it
-            # Note: do not reduce the weight values, because FSDP already ensures
-            # the weight values on all ranks are the same after all-gather.
             all_amax_tensors = torch.cat(
-                fp8_amax_x_tensor_list + fp8_amax_dL_dY_tensor_list
+                fp8_amax_x_tensor_list
+                + fp8_amax_w_tensor_list
+                + fp8_amax_dL_dY_tensor_list
             )
             all_reduced_amax_tensor = all_reduce(
                 all_amax_tensors, "MAX", list(range(dist.get_world_size()))
@@ -302,12 +301,14 @@ def inner_func():
                 all_reduced_amax_tensor = all_reduced_amax_tensor.wait()
 
             (
-                reduced_fp8_amax_tensor,
+                reduced_fp8_amax_x_tensor,
+                reduced_fp8_amax_w_tensor,
                 reduced_fp8_amax_dL_dY_tensor,
             ) = torch.split(all_reduced_amax_tensor, len(fp8_amax_x_tensor_list))
 
             for idx, child in enumerate(fp8_layers):
-                child.fp8_amax_x.copy_(reduced_fp8_amax_tensor[idx])
+                child.fp8_amax_x.copy_(reduced_fp8_amax_x_tensor[idx])
+                child.fp8_amax_w.copy_(reduced_fp8_amax_w_tensor[idx])
                 child.fp8_amax_dL_dY.copy_(reduced_fp8_amax_dL_dY_tensor[idx])
 
         # We create two stacked tensor groups, one for the amax history and one for the current scales

float8_experimental/fsdp_utils.py

@@ -18,7 +18,7 @@
     ScaledMMConfig,
 )
 
-from float8_experimental.float8_utils import EPS
+from float8_experimental.float8_utils import e4m3_dtype, EPS
 from torch._prims_common import suggest_memory_format
 
 
@@ -189,3 +189,182 @@ def fsdp_post_all_gather(
             out._scale = scale
             return
         return Float8Tensor(data, scale, param_dtype, self._mm_config), (data,)
+
+
+class WeightWithDelayedFloat8CastTensor(torch.Tensor):
+    @staticmethod
+    def __new__(
+        cls,
+        tensor: torch.Tensor,
+        amax_buffer: torch.Tensor,
+        amax_history_buffer: torch.Tensor,
+        scale_buffer: torch.Tensor,
+        mm_config: ScaledMMConfig,
+        is_amax_initialized: bool,
+    ):
+        return torch.Tensor._make_wrapper_subclass(
+            cls,
+            tensor.size(),
+            strides=tensor.stride(),
+            storage_offset=tensor.storage_offset(),
+            memory_format=suggest_memory_format(tensor),
+            dtype=tensor.dtype,
+            layout=tensor.layout,
+            device=tensor.device,
+            pin_memory=tensor.is_pinned(),
+            requires_grad=tensor.requires_grad,
+        )
+
+    def __init__(
+        self,
+        tensor: torch.Tensor,
+        amax_buffer: torch.Tensor,
+        amax_history_buffer: torch.Tensor,
+        scale_buffer: torch.Tensor,
+        mm_config: ScaledMMConfig,
+        is_amax_initialized: bool,
+    ):
+        self._tensor = tensor
+        self._amax_buffer = amax_buffer
+        self._amax_history_buffer = amax_history_buffer
+        self._scale_buffer = scale_buffer
+        self._mm_config = mm_config
+
+        # Note: is_amax_initialized is not a buffer to avoid data dependent
+        # control flow visible to dynamo
+        # TODO(future PR): add serialization for this flag
+        self.is_amax_initialized = is_amax_initialized
+
+    @classmethod
+    def __torch_dispatch__(cls, func, types, args, kwargs=None):
+        if func == torch.ops.aten.detach.default:
+            return WeightWithDelayedFloat8CastTensor(
+                args[0]._tensor,
+                args[0]._amax_buffer,
+                args[0]._amax_history_buffer,
+                args[0]._scale_buffer,
+                args[0]._mm_config,
+                args[0].is_amax_initialized,
+            )
+        mm_config: Optional[ScaledMMConfig] = None
+        amax_buffer: Optional[torch.Tensor] = None
+        amax_history_buffer: Optional[torch.Tensor] = None
+        scale_buffer: Optional[torch.Tensor] = None
+        is_amax_initialized: Optional[bool] = None
+
+        def unwrap(t):
+            nonlocal mm_config
+            if mm_config is None:
+                mm_config = t._mm_config
+            else:
+                mm_config = merge_mm_configs(mm_config, t._mm_config)
+            nonlocal amax_buffer
+            if amax_buffer is None:
+                amax_buffer = t._amax_buffer
+            nonlocal amax_history_buffer
+            if amax_history_buffer is None:
+                amax_history_buffer = t._amax_history_buffer
+            nonlocal scale_buffer
+            if scale_buffer is None:
+                scale_buffer = t._scale_buffer
+            nonlocal is_amax_initialized
+            if is_amax_initialized is None:
+                is_amax_initialized = t.is_amax_initialized
+            return t._tensor
+
+        args, kwargs = pytree.tree_map_only(
+            WeightWithDelayedFloat8CastTensor, unwrap, (args, kwargs or {})
+        )
+        out = func(*args, **kwargs)
+        if func not in _ops_to_preserve_subclass:
+            return out
+        return pytree.tree_map_only(
+            torch.Tensor,
+            lambda x: WeightWithDelayedFloat8CastTensor(
+                x,
+                amax_buffer,
+                amax_history_buffer,
+                scale_buffer,
+                mm_config,
+                is_amax_initialized,
+            ),
+            out,
+        )
+
+    def __tensor_flatten__(self):
+        return (
+            [
+                "_tensor",
+                "_amax_buffer",
+                "_amax_history_buffer",
+                "_scale_buffer",
+            ],
+            {
+                "mm_config": self._mm_config,
+                "is_amax_initialized": is_amax_initialized,
+            },
+        )
+
+    @staticmethod
+    def __tensor_unflatten__(inner_tensors, metadata, outer_size, outer_stride):
+        return WeightWithDelayedFloat8CastTensor(
+            inner_tensors["_tensor"],
+            inner_tensors["_amax_buffer"],
+            inner_tensors["_amax_history_buffer"],
+            inner_tensors["_scale_buffer"],
+            metadata["mm_config"],
+            metadata["is_amax_initialized"],
+        )
+
+    def __repr__(self):
+        return f"WeightWithDelayedFloat8CastTensor(tensor={self._tensor}, amax_buffer={self._amax_buffer}, scale_buffer={self._scale_buffer}, mm_config={self._mm_config})"
+
+    def fsdp_pre_all_gather(self, mesh):
+        # initialize if needed
+        # TODO(before land): ensure settings are consistent between Float8Linear and here
+        if not self.is_amax_initialized:
+            from float8_experimental.float8_linear import (
+                _maybe_initialize_amaxes_scales_for_float8_cast,
+            )
+
+            _maybe_initialize_amaxes_scales_for_float8_cast(
+                self._tensor,
+                self._amax_buffer,
+                self._amax_history_buffer,
+                self._scale_buffer,
+                "max",  # TODO(before land): read this from parent
+                e4m3_dtype,
+                self.is_amax_initialized,
+                reduce_amax=True,
+            )
+            self.is_amax_initialized = True
+
+        # this will:
+        # 1. cast the tensor to float8 using `_scale_buffer`
+        # 2. populate `_amax_buffer` inplace
+        # TODO(future PR): clean up all the casting functions and clearly
+        # separate dynamic vs delayed, tech debt has accumulated
+        float8_tensor = Float8Tensor.to_float8(
+            self._tensor,
+            self._scale_buffer,
+            e4m3_dtype,
+            self._amax_buffer,
+            self._mm_config,
+        )
+        return (float8_tensor._data,), (float8_tensor._scale,)
+
+    def fsdp_post_all_gather(
+        self,
+        all_gather_outputs: Tuple[torch.Tensor, ...],
+        metadata: Any,
+        param_dtype: torch.dtype,
+        *,
+        out: Optional[torch.Tensor] = None,
+    ):
+        (data,) = all_gather_outputs
+        (scale,) = metadata
+        if out is not None:
+            assert isinstance(out, Float8Tensor), f"{type(out)}"
+            out._scale = scale
+            return
+        return Float8Tensor(data, scale, param_dtype, self._mm_config), (data,)

test/test_fsdp2/test_fsdp2_common.py

@@ -6,6 +6,11 @@
 import torch
 import torch.distributed as dist
 import torch.nn as nn
+from float8_experimental.float8_linear import Float8Linear, TensorScalingType
+from float8_experimental.float8_linear_utils import (
+    linear_requires_sync,
+    sync_float8_amax_and_scale_history,
+)
 from float8_experimental.fsdp_utils import precompute_float8_dynamic_scale_for_fsdp
 
 
@@ -17,6 +22,7 @@ def check_parity_no_mp(
     fsdp_optim: torch.optim.Optimizer,
     local_inp: torch.Tensor,
     precompute: bool = False,
+    scaling_type_w: TensorScalingType = TensorScalingType.DYNAMIC,
 ):
     for iter_idx in range(10):
         losses: List[torch.Tensor] = []
@@ -28,10 +34,18 @@ def check_parity_no_mp(
                 for param in model.parameters():
                     dist.all_reduce(param.grad)
                     param.grad.div_(dist.get_world_size())
-            # TODO(future): add amax syncing once delayed scaling is supported
+
+            if linear_requires_sync(scaling_type_w=scaling_type_w):
+                sync_float8_amax_and_scale_history(model)
+
             optim.step()
-            if model is fsdp_model and precompute:
+            if (
+                model is fsdp_model
+                and precompute
+                and scaling_type_w is TensorScalingType.DYNAMIC
+            ):
                 precompute_float8_dynamic_scale_for_fsdp(model)
+
         test_cls.assertEqual(losses[0], losses[1])