Add Parallel Cross Entropy (huggingface#2017)

optimum/fx/parallelization/decomp.py

@@ -197,7 +197,7 @@ def run(self, *args, **kwargs):
 def decompose_and_functionalize(
     graph_module: GraphModule,
     decomposition_table: Dict[torch._ops.OperatorBase, Callable] = core_aten_decompositions(),
-    leaf_function_targets: List[Callable] = [F.scaled_dot_product_attention],
+    leaf_function_targets: List[Callable] = [F.scaled_dot_product_attention, F.cross_entropy],
 ) -> Callable:
     """
     API to decompose and functionalize a high-level graph module.

optimum/fx/parallelization/op_registry/op_handlers.py

@@ -19,7 +19,7 @@
 from torch.fx import Node
 
 from ..core import Config
-from ..utils import is_activation, is_embedding, is_linear
+from ..utils import is_activation, is_cross_entropy, is_cross_entropy_parallel_compatible, is_embedding, is_linear
 
 
 class Registry:
@@ -334,7 +334,16 @@ def propagate(self) -> List[int]:
         ndim = arg.meta["val"].ndim
         slice_dim = (slice_dim + ndim) % ndim
         if slice_dim == axis:
-            # slice on the parallel axis is not allowed
+            # slice on the parallel axis is not allowed, except it's a nop
+            start, stop, step = 0, arg.meta["val"].shape[axis], 1
+            if len(self.node.args) > 2:
+                start = self.node.args[2]
+            elif len(self.node.args) > 3:
+                stop = self.node.args[3]
+            elif len(self.node.args) > 4:
+                step = self.node.args[4]
+            if start == 0 and stop >= arg.meta["val"].shape[axis] and step == 1:
+                return [axis]
             return []
         return [axis]
 
@@ -404,12 +413,12 @@ def propagate(self) -> List[int]:
         if self.node.op in ["placeholder", "get_attr"]:
             return [None]
         elif self.node.op == "output":
-            for node in self.node.all_input_nodes:
-                # TODO: allow parallelized nodes in output, and append comm ops in graph tp all-gather
-                # parallelized output if intructed
-                if self.extract_axis(node) is not None:
-                    return []
-            return [None]
+            # does not care about if output is being parallelized right now, because if the output is loss,
+            # then it must be not parallelized as long as it comes from sharded cross entropy.
+            # TODO: append all-gather comm ops before all parallelized output nodes if instructed.
+            input_arg = self.node.all_input_nodes[0]
+            axis = self.extract_axis(input_arg)
+            return [axis]
         elif is_linear(self.node):
             input_arg = self.node.all_input_nodes[0]
             axis = self.extract_axis(input_arg)
@@ -438,6 +447,16 @@ def propagate(self) -> List[int]:
                 return [1, None] if self.config.enable_sequence_parallel else [None]
             else:
                 return []
+        elif is_cross_entropy(self.node):
+            logits = self.node.all_input_nodes[0]
+            axis = self.extract_axis(logits)
+            if axis is None or (
+                is_cross_entropy_parallel_compatible(self.node) and axis == logits.meta["val"].ndim - 1
+            ):
+                # for cross entropy, the input logits parallel axis can only be the last axis or None
+                return [None]
+            else:
+                return []
         elif is_activation(self.node):
             return UnaryOpParallelAxisPropagateHandler(self.node, self.meta_key, self.config).propagate()
 

optimum/fx/parallelization/parallel_layers/__init__.py

@@ -14,3 +14,4 @@
 # limitations under the License.
 from .embedding import VocabParallelEmbedding
 from .linear import ColumnParallelLinear, RowParallelLinear
+from .loss import VocabParallelCrossEntropyLoss, sharded_cross_entropy_wrapper_fn

optimum/fx/parallelization/parallel_layers/loss.py

@@ -0,0 +1,163 @@
+# coding=utf-8
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from functools import wraps
+from typing import Optional
+
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+
+from ..core import ParallelExecutionCtx
+
+
+# Adapted from https://github.com/huggingface/nanotron/blob/main/src/nanotron/parallel/tensor_parallel/functional.py
+class _ShardedCrossEntropy(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        sharded_logits: torch.Tensor,  # (batch_size, length, sharded_hidden_size)
+        target: torch.Tensor,  # (batch_size, length)
+        group: dist.ProcessGroup,
+    ):
+        # Maximum value along last dimension across all GPUs.
+        logits_max = torch.max(sharded_logits, dim=-1)[0]
+        dist.all_reduce(logits_max, op=dist.ReduceOp.MAX, group=group)
+        # Subtract the maximum value.
+        sharded_logits = sharded_logits - logits_max.unsqueeze(dim=-1)
+
+        # Get the shard's indices
+        sharded_hidden_size = sharded_logits.shape[-1]
+        rank = dist.get_rank(group)
+        start_index = rank * sharded_hidden_size
+        end_index = start_index + sharded_hidden_size
+
+        # Create a mask of valid ids (1 means it needs to be masked).
+        target_mask = (target < start_index) | (target >= end_index)
+        masked_target = target.clone() - start_index
+        masked_target[target_mask] = 0
+
+        # Get predicted-logits = logits[target].
+        # For Simplicity, we convert logits to a 2-D tensor with size
+        # [*, shard-size] and target to a 1-D tensor of size [*].
+        logits_2d = sharded_logits.view(-1, sharded_hidden_size)
+        masked_target_1d = masked_target.view(-1)
+        arange_1d = torch.arange(start=0, end=logits_2d.shape[0], device=logits_2d.device)
+        predicted_logits_1d = logits_2d[arange_1d, masked_target_1d]
+        if predicted_logits_1d.is_contiguous():
+            predicted_logits_1d = predicted_logits_1d.clone()
+        else:
+            predicted_logits_1d = predicted_logits_1d.contiguous()
+        predicted_logits = predicted_logits_1d.view_as(target)
+        predicted_logits[target_mask] = 0.0
+        # All reduce is needed to get the chunks from other GPUs.
+        dist.all_reduce(predicted_logits, op=dist.ReduceOp.SUM, group=group)
+
+        # Sum of exponential of logits along vocab dimension across all GPUs.
+        exp_logits = sharded_logits
+        torch.exp(sharded_logits, out=exp_logits)
+        sum_exp_logits = exp_logits.sum(dim=-1)
+        dist.all_reduce(sum_exp_logits, op=dist.ReduceOp.SUM, group=group)
+
+        # Loss = log(sum(exp(logits))) - predicted-logit.
+        loss = torch.log(sum_exp_logits) - predicted_logits
+
+        # Normalize and optionally smooth logits
+        exp_logits.div_(sum_exp_logits.unsqueeze(dim=-1))
+
+        # Store softmax, target-mask and masked-target for backward pass.
+        ctx.save_for_backward(exp_logits, target_mask, masked_target_1d)
+
+        return loss
+
+    @staticmethod
+    def backward(ctx, grad_output: torch.Tensor):
+        # Retrieve tensors from the forward path.
+        softmax, target_mask, masked_target_1d = ctx.saved_tensors
+
+        # All the inputs have softmax as their gradient.
+        grad_input = softmax
+        # For simplicity, work with the 2D gradient.
+        sharded_hidden_size = softmax.size()[-1]
+        grad_2d = grad_input.view(-1, sharded_hidden_size)
+
+        # Add the gradient from matching classes.
+        arange_1d = torch.arange(start=0, end=grad_2d.size()[0], device=grad_2d.device)
+        grad_2d[arange_1d, masked_target_1d] -= 1.0 - target_mask.view(-1).float()
+
+        # Finally elementwise multiplication with the output gradients.
+        grad_input.mul_(grad_output.unsqueeze(dim=-1))
+
+        return grad_input, None, None
+
+
+def sharded_cross_entropy(sharded_logits: torch.Tensor, target: torch.Tensor, process_group: dist.ProcessGroup):
+    return _ShardedCrossEntropy.apply(sharded_logits, target, process_group)
+
+
+def sharded_cross_entropy_wrapper_fn(process_group: dist.ProcessGroup):
+    @wraps(sharded_cross_entropy)
+    def wrapper(
+        sharded_logits: torch.Tensor,
+        target: torch.Tensor,
+        weight: Optional[torch.Tensor] = None,
+        size_average: Optional[bool] = None,
+        ignore_index: int = -100,
+        reduce: Optional[bool] = None,
+        reduction: str = "mean",
+        label_smoothing: float = 0.0,
+    ):
+        if weight is not None or ignore_index != -100 or label_smoothing != 0.0:
+            raise ValueError(
+                "Does not support weighted mode, index ignoring and label smoothing in current parallel cross entropy implementation."
+            )
+        loss: torch.Tensor = sharded_cross_entropy(sharded_logits, target, process_group)
+
+        if size_average is not None or reduce is not None:
+            size_average = True if size_average is None else size_average
+            reduce = True if reduce is None else reduce
+
+            if size_average and reduce:
+                reduction = "mean"
+            elif reduce:
+                reduction = "sum"
+            else:
+                reduction = "none"
+
+        if reduction == "mean":
+            return loss.mean()
+        elif reduction == "sum":
+            return loss.sum()
+        return loss
+
+    return wrapper
+
+
+class VocabParallelCrossEntropyLoss(nn.Module):
+    """
+    Simple parallel cross entropy implementation which does not support weighted mode and label smoothing yet.
+    """
+
+    def __init__(self, ctx: ParallelExecutionCtx, reduction: str = "mean") -> None:
+        super(VocabParallelCrossEntropyLoss, self).__init__()
+        self.process_group = ctx.tp_group
+        self.reduction = reduction
+
+    def forward(self, sharded_logits: torch.Tensor, target: torch.Tensor):
+        loss: torch.Tensor = _ShardedCrossEntropy.apply(sharded_logits, target, self.process_group)
+        if self.reduction == "mean":
+            return loss.mean()
+        elif self.reduction == "sum":
+            return loss.sum()
+        return loss

optimum/fx/parallelization/passes.py

@@ -26,8 +26,15 @@
 from .decomp import decompose_and_functionalize
 from .distributed import scatter
 from .op_registry import REGISTRY, FallbackParallelAxisPropagateHandler
-from .parallel_layers import ColumnParallelLinear, RowParallelLinear, VocabParallelEmbedding
+from .parallel_layers import (
+    ColumnParallelLinear,
+    RowParallelLinear,
+    VocabParallelCrossEntropyLoss,
+    VocabParallelEmbedding,
+    sharded_cross_entropy_wrapper_fn,
+)
 from .utils import (
+    is_cross_entropy,
     is_embedding,
     is_linear,
     is_shape_consumer,
@@ -273,6 +280,11 @@ def run(self, graph_module: GraphModule, ctx: ParallelExecutionCtx, config: Conf
                     info["sequence_parallel"] = False
                 self.place_marker_per_node(node, info)
 
+            elif is_cross_entropy(node):
+                axis_before = ParallelAxisSolverPass.get_stored_field_info(node.args[0], "parallel_axis")
+                if axis_before is not None:
+                    self.place_marker_per_node(node, {"axis": "vocab"})
+
         return graph_module
 
 
@@ -343,6 +355,35 @@ def handle_embedding(node: Node, ctx: ParallelExecutionCtx) -> None:
             layer_cache[key] = new_mod
         setattr(parent_mod, field, new_mod)
 
+    @staticmethod
+    def handle_cross_entropy(node: Node, ctx: ParallelExecutionCtx) -> None:
+        axis = ParallelLayerAnnotatePass.get_stored_field_info(node, field="axis")
+        if axis is None:
+            return
+
+        assert axis in {"vocab"}, "Only support parallelization on vocab dim for now."
+        if node.op == "call_module":
+            graph_module = node.graph.owning_module
+            prefix_and_field = node.target.rsplit(".", maxsplit=1)
+            if len(prefix_and_field) == 2:
+                parent_mod = graph_module.get_submodule(prefix_and_field[0])
+                field = prefix_and_field[1]
+            else:
+                parent_mod = graph_module
+                field = node.target
+
+            mod: nn.CrossEntropyLoss = graph_module.get_submodule(node.target)
+            key, layer_cache = node.target, ctx.parallel_layer_cache
+            if key in layer_cache:
+                new_mod = layer_cache[key]
+            else:
+                assert ctx.compile_times == 0, "illegal path for recompilation"
+                new_mod = VocabParallelCrossEntropyLoss(ctx, reduction=mod.reduction)
+                layer_cache[key] = new_mod
+            setattr(parent_mod, field, new_mod)
+        else:
+            node.target = sharded_cross_entropy_wrapper_fn(process_group=ctx.tp_group)
+
     @staticmethod
     def handle_hard_coded_axis_param(node: Node, ctx: ParallelExecutionCtx) -> None:
         def extract_shape_from_node(node: Node) -> List[Any]:
@@ -384,6 +425,8 @@ def run(self, graph_module: GraphModule, ctx: ParallelExecutionCtx, config: Conf
                 self.handle_linear(node, ctx)
             elif is_embedding(node):
                 self.handle_embedding(node, ctx)
+            elif is_cross_entropy(node):
+                self.handle_cross_entropy(node, ctx)
             # correct the attention head num in parallel setting
             elif is_shape_consumer(node):
                 self.handle_hard_coded_axis_param(node, ctx)

optimum/fx/parallelization/utils.py

@@ -82,6 +82,40 @@ def is_shape_generator(node: Node) -> bool:
     return node.op == "call_method" and node.target == "size"
 
 
+def is_cross_entropy(node: Node) -> bool:
+    if node.op == "call_function":
+        return node.target is F.cross_entropy
+    elif node.op == "call_module":
+        mod = node.graph.owning_module
+        return isinstance(mod.get_submodule(node.target), nn.CrossEntropyLoss)
+    return False
+
+
+def is_cross_entropy_parallel_compatible(node: Node) -> bool:
+    """
+    For now `VocabParallelCrossEntropyLoss` does not support weighted mode, index ignoring and label smoothing.
+    """
+    if node.op == "call_function":
+        weight = node.kwargs.get("weight", None)
+        ignore_index = node.kwargs.get("ignore_index", -100)
+        label_smoothing = node.kwargs.get("label_smoothing", 0.0)
+        if len(node.args) > 2 and weight is None:
+            weight = node.args[2]
+        if len(node.args) > 4 and ignore_index == -100:
+            ignore_index = node.args[4]
+        if len(node.args) > 7 and label_smoothing == 0.0:
+            label_smoothing = node.args[7]
+
+        return weight is None and ignore_index == -100 and label_smoothing == 0.0
+
+    elif node.op == "call_module":
+        mod: nn.CrossEntropyLoss = node.graph.owning_module.get_submodule(node.target)
+        weight, label_smoothing, ignore_index = mod.weight, mod.label_smoothing, mod.ignore_index
+        return weight is None and ignore_index == -100 and label_smoothing == 0.0
+
+    return False
+
+
 def stable_topological_sort(graph: Graph):
     def _args(n: torch.fx.Node) -> List[torch.fx.node.Argument]:
         args: List[torch.fx.node.Argument] = []