Distributed layers (#1270)

python/mlx/distributed_run.py

@@ -761,6 +761,8 @@ def main():
         "--cwd", help="Set the working directory on each node to the provided one"
     )
     args, rest = parser.parse_known_args()
+    if rest[0] == "--":
+        rest.pop(0)
 
     if args.print_python:
         print(sys.executable)

python/mlx/nn/layers/__init__.py

@@ -60,6 +60,12 @@ from mlx.nn.layers.convolution_transpose import (
     ConvTranspose2d,
     ConvTranspose3d,
 )
+from mlx.nn.layers.distributed import (
+    AllToShardedLinear,
+    QuantizedAllToShardedLinear,
+    QuantizedShardedToAllLinear,
+    ShardedToAllLinear,
+)
 from mlx.nn.layers.dropout import Dropout, Dropout2d, Dropout3d
 from mlx.nn.layers.embedding import Embedding
 from mlx.nn.layers.linear import Bilinear, Identity, Linear

python/mlx/nn/layers/distributed.py

@@ -0,0 +1,599 @@
+# Copyright © 2024 Apple Inc.
+
+import math
+from functools import lru_cache
+from typing import Callable, Optional, Union
+
+import mlx.core as mx
+from mlx.nn.layers.base import Module
+from mlx.nn.layers.linear import Linear
+from mlx.nn.layers.quantized import QuantizedLinear
+from mlx.utils import tree_map_with_path
+
+
+@lru_cache
+def sum_gradients(group):
+    if group.size() == 1:
+        return lambda x: x
+
+    @mx.custom_function
+    def f(x):
+        return x
+
+    @f.vjp
+    def f(x, dx, _):
+        return mx.distributed.all_sum(dx, group=group)
+
+    return f
+
+
+def _split(weight, segments, axis):
+    """Equivalent to mx.split but allows for fractional segments."""
+    if isinstance(segments, int) or isinstance(segments[0], int):
+        return mx.split(weight, segments, axis=axis)
+
+    N = weight.shape[axis]
+    indices = [int(s * N) for s in segments]
+    return mx.split(weight, indices, axis=axis)
+
+
+def _shard(
+    parameters: dict,
+    sharding_predicate: Callable,
+    group: Optional[mx.distributed.Group] = None,
+):
+    """Returns a new parameter tree with the weights sharded according to the
+    sharding_predicate.
+
+    The sharding predicate should return the sharding axis and optionally also
+    the segments that comprise the weight.
+    """
+    group = group or mx.distributed.init()
+    N = group.size()
+    r = group.rank()
+
+    def _shard_fn(path, weight):
+        if not isinstance(weight, mx.array):
+            return weight
+
+        s = sharding_predicate(path, weight)
+        if s is None:
+            return weight
+
+        axis = None
+        segments = 1
+        if isinstance(s, int):
+            axis = s
+        elif isinstance(s, tuple):
+            axis, segments = s
+        else:
+            raise ValueError(
+                "The sharding function should return int or tuple[int, list]"
+            )
+
+        return mx.contiguous(
+            mx.concatenate(
+                [_split(part, N, axis)[r] for part in _split(weight, segments, axis)],
+                axis=axis,
+            )
+        )
+
+    return tree_map_with_path(_shard_fn, parameters)
+
+
+def _all_to_sharded(segments):
+    """Simple predicate to shard fully connected layers such that a common
+    representation becomes a sharded representation."""
+
+    def _shard_fn(path, weight):
+        return max(weight.ndim - 2, 0), segments
+
+    return _shard_fn
+
+
+def _sharded_to_all(segments):
+    """Simple predicate to shard fully connected layers such that a sharded
+    representation becomes a common representation."""
+
+    def _shard_fn(path, weight):
+        if path.endswith("bias"):
+            return None
+        return -1, segments
+
+    return _shard_fn
+
+
+def _check_sharding(sharding):
+    if sharding not in ("all-to-sharded", "sharded-to-all"):
+        raise ValueError(
+            (
+                f"Sharding type {sharding=} not supported, "
+                "choose one of 'all-to-sharded' or 'sharded-to-all'"
+            )
+        )
+
+
+def shard_inplace(
+    module: Module,
+    sharding: Union[str, Callable],
+    *,
+    segments: Union[int, list] = 1,
+    group: Optional[mx.distributed.Group] = None,
+):
+    """Shard a module in-place by updating its parameter dictionary with the
+    sharded parameter dictionary.
+
+    The ``sharding`` argument can be any callable that given the path and the
+    weight returns the sharding axis and optionally also the segments that
+    comprise the unsharded weight. For instance if the weight is a fused QKV
+    matrix the segments should be 3.
+
+    .. note::
+        The module doesn't change so in order for distributed communication to
+        happen the module needs to natively support it and for it to be enabled.
+
+    Args:
+        module (mlx.nn.Module): The parameters of this module will be sharded
+            in-place.
+        sharding (str or callable): One of "all-to-sharded" and
+            "sharded-to-all" or a callable that returns the sharding axis and
+            segments.
+        segments (int or list): The segments to use if ``sharding`` is a
+            string. Default: ``1``.
+        group (mlx.core.distributed.Group): The distributed group to shard
+            across. If not set, the global group will be used. Default: ``None``.
+    """
+    if isinstance(sharding, str):
+        _check_sharding(sharding)
+        sharding = (
+            _all_to_sharded(segments)
+            if sharding == "all-to-sharded"
+            else _sharded_to_all(segments)
+        )
+    module.update(_shard(module.parameters(), sharding, group))
+
+
+def shard_linear(
+    module: Module,
+    sharding: str,
+    *,
+    segments: Union[int, list] = 1,
+    group: Optional[mx.distributed.Group] = None,
+):
+    """Create a new linear layer that has its parameters sharded and also
+    performs distributed communication either in the forward or backward
+    pass.
+
+    .. note::
+        Contrary to ``shard_inplace``, the original layer is not changed but a
+        new layer is returned.
+
+    Args:
+        module (mlx.nn.Module): The linear layer to be sharded.
+        sharding (str): One of "all-to-sharded" and
+            "sharded-to-all" that defines the type of sharding to perform.
+        segments (int or list): The segments to use. Default: ``1``.
+        group (mlx.core.distributed.Group): The distributed group to shard
+            across. If not set, the global group will be used. Default: ``None``.
+    """
+    _check_sharding(sharding)
+    fns = {
+        ("all-to-sharded", True): AllToShardedLinear.from_linear,
+        ("all-to-sharded", False): QuantizedAllToShardedLinear.from_quantized_linear,
+        ("sharded-to-all", True): ShardedToAllLinear.from_linear,
+        ("sharded-to-all", False): QuantizedShardedToAllLinear.from_quantized_linear,
+    }
+    return fns[sharding, isinstance(module, Linear)](
+        module, segments=segments, group=group
+    )
+
+
+class AllToShardedLinear(Module):
+    """Each member of the group applies part of the affine transformation such
+    that the result is sharded across the group.
+
+    The gradients are automatically aggregated from each member of the group.
+
+    Args:
+        input_dims (int): The dimensionality of the input features
+        output_dims (int): The dimensionality of the output features
+        bias (bool, optional): If set to ``False`` the the layer will not use a
+            bias. Default is ``True``.
+        group (mx.distributed.Group, optional): The sharding will happen across
+            this group. If not set then the global group is used. Default is
+            ``None``.
+    """
+
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        bias: bool = True,
+        group: Optional[mx.distributed.Group] = None,
+    ):
+        super().__init__()
+
+        # Initialize the parameters
+        scale = math.sqrt(1.0 / input_dims)
+        self.group = group or mx.distributed.init()
+        N = self.group.size()
+
+        if (output_dims % N) != 0:
+            raise ValueError(
+                f"Cannot shard the output of size {output_dims} across {N} devices."
+            )
+
+        self.weight = mx.random.uniform(
+            low=-scale,
+            high=scale,
+            shape=(output_dims // N, input_dims),
+        )
+        if bias:
+            self.bias = mx.random.uniform(
+                low=-scale,
+                high=scale,
+                shape=(output_dims // N,),
+            )
+
+    def _extra_repr(self) -> str:
+        out_dims, in_dims = self.weight.shape
+        N = self.group.size()
+        out_dims *= N
+        return f"input_dims={in_dims}, output_dims={out_dims}, bias={'bias' in self}"
+
+    def __call__(self, x: mx.array) -> mx.array:
+        # Aggregate the gradients coming from each shard
+        x = sum_gradients(self.group)(x)
+
+        # Compute the affine projection
+        if "bias" in self:
+            x = mx.addmm(self["bias"], x, self["weight"].T)
+        else:
+            x = x @ self["weight"].T
+        return x
+
+    @classmethod
+    def from_linear(
+        cls,
+        linear_layer: Module,
+        *,
+        segments: Union[int, list] = 1,
+        group: Optional[mx.distributed.Group] = None,
+    ):
+        group = group or mx.distributed.init()
+        output_dims, input_dims = linear_layer.weight.shape
+
+        sl = cls(input_dims, output_dims, hasattr(linear_layer, "bias"), group)
+        sl.update(_shard(linear_layer.parameters(), _all_to_sharded(segments), group))
+
+        return sl
+
+
+class ShardedToAllLinear(Module):
+    """Each member of the group applies part of the affine transformation and
+    then aggregates the results.
+
+    All nodes will have the same exact result after this layer.
+
+    :class:`ShardedToAllLinear` provides a classmethod :meth:`from_linear` to
+    convert linear layers to sharded :obj:`ShardedToAllLinear` layers.
+
+    Args:
+        input_dims (int): The dimensionality of the input features
+        output_dims (int): The dimensionality of the output features
+        bias (bool, optional): If set to ``False`` the the layer will not use a
+            bias. Default is ``True``.
+        group (mx.distributed.Group, optional): The sharding will happen across
+            this group. If not set then the global group is used. Default is
+            ``None``.
+    """
+
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        bias: bool = True,
+        group: Optional[mx.distributed.Group] = None,
+    ):
+        super().__init__()
+
+        # Initialize the parameters
+        scale = math.sqrt(1.0 / input_dims)
+        self.group = group or mx.distributed.init()
+        N = self.group.size()
+
+        if (input_dims % N) != 0:
+            raise ValueError(
+                f"The input of size {input_dims} cannot be sharded across {N} devices."
+            )
+
+        self.weight = mx.random.uniform(
+            low=-scale,
+            high=scale,
+            shape=(output_dims, input_dims // N),
+        )
+        if bias:
+            self.bias = mx.random.uniform(
+                low=-scale,
+                high=scale,
+                shape=(output_dims,),
+            )
+
+    def _extra_repr(self) -> str:
+        N = self.group.size()
+        out_dims, in_dims = self.weight.shape
+        in_dims *= N
+        return f"input_dims={in_dims}, output_dims={out_dims}, bias={'bias' in self}"
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = x @ self["weight"].T
+
+        x = mx.distributed.all_sum(x, group=self.group)
+
+        if "bias" in self:
+            x = x + self["bias"]
+
+        return x
+
+    @classmethod
+    def from_linear(
+        cls,
+        linear_layer: Module,
+        *,
+        segments: Union[int, list] = 1,
+        group: Optional[mx.distributed.Group] = None,
+    ):
+        group = group or mx.distributed.init()
+        output_dims, input_dims = linear_layer.weight.shape
+
+        sl = cls(input_dims, output_dims, hasattr(linear_layer, "bias"), group)
+        sl.update(_shard(linear_layer.parameters(), _sharded_to_all(segments), group))
+
+        return sl
+
+
+class QuantizedAllToShardedLinear(Module):
+    """Each member of the group applies part of the affine transformation with
+    a quantized matrix such that the result is sharded across the group.
+
+    It is the quantized equivalent of :class:`mlx.nn.AllToShardedLinear`.
+    Similar to :class:`mlx.nn.QuantizedLinear` its parameters are frozen and
+    will not be included in any gradient computation.
+
+    Args:
+        input_dims (int): The dimensionality of the input features.
+        output_dims (int): The dimensionality of the output features.
+        bias (bool, optional): If set to ``False`` then the layer will not use
+            a bias. Default: ``True``.
+        group_size (int, optional): The group size to use for the quantized
+            weight. See :func:`~mlx.core.quantize`. Default: ``64``.
+        bits (int, optional): The bit width to use for the quantized weight.
+            See :func:`~mlx.core.quantize`. Default: ``4``.
+        group (mx.distributed.Group, optional): The sharding will happen across
+            this group. If not set then the global group is used. Default is
+            ``None``.
+    """
+
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        bias: bool = True,
+        group_size: int = 64,
+        bits: int = 4,
+        group: Optional[mx.distributed.Group] = None,
+    ):
+        super().__init__()
+
+        # Quantization config
+        self.group_size = group_size
+        self.bits = bits
+
+        # Initialize the quantized weight
+        scale = math.sqrt(1.0 / input_dims)
+        self.group = group or mx.distributed.init()
+        N = self.group.size()
+
+        if (output_dims % N) != 0:
+            raise ValueError(
+                f"Cannot shard the output of size {output_dims} across {N} devices."
+            )
+
+        weight = mx.random.uniform(
+            low=-scale,
+            high=scale,
+            shape=(output_dims // N, input_dims),
+        )
+        self.weight, self.scales, self.biases = mx.quantize(weight, group_size, bits)
+
+        # And bias if needed
+        if bias:
+            self.bias = mx.zeros((output_dims // N,))
+
+        # Freeze this model's parameters
+        self.freeze()
+
+    def unfreeze(self, *args, **kwargs):
+        """Wrap unfreeze so that we unfreeze any layers we might contain but
+        our parameters will remain frozen."""
+        super().unfreeze(*args, **kwargs)
+        self.freeze(recurse=False)
+
+    def _extra_repr(self) -> str:
+        out_dims, in_dims = self.weight.shape
+        in_dims *= 32 // self.bits
+        out_dims *= self.group.size()
+        return (
+            f"input_dims={in_dims}, output_dims={out_dims}, bias={'bias' in self}, "
+            f"group_size={self.group_size}, bits={self.bits}"
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        # Aggregate the gradients coming from each shard
+        x = sum_gradients(self.group)(x)
+
+        x = mx.quantized_matmul(
+            x,
+            self["weight"],
+            scales=self["scales"],
+            biases=self["biases"],
+            transpose=True,
+            group_size=self.group_size,
+            bits=self.bits,
+        )
+        if "bias" in self:
+            x = x + self["bias"]
+        return x
+
+    @classmethod
+    def from_quantized_linear(
+        cls,
+        quantized_linear_layer: Module,
+        *,
+        segments: Union[int, list] = 1,
+        group: Optional[mx.distributed.Group] = None,
+    ):
+        group = group or mx.distributed.init()
+        output_dims, input_dims = quantized_linear_layer.weight.shape
+        input_dims *= 32 // quantized_linear_layer.bits
+
+        sl = cls(
+            input_dims,
+            output_dims,
+            hasattr(quantized_linear_layer, "bias"),
+            group_size=quantized_linear_layer.group_size,
+            bits=quantized_linear_layer.bits,
+            group=group,
+        )
+        sl.update(
+            _shard(
+                quantized_linear_layer.parameters(),
+                _all_to_sharded(segments),
+                group,
+            )
+        )
+
+        return sl
+
+
+class QuantizedShardedToAllLinear(Module):
+    """Each member of the group applies part of the affine transformation using
+    the quantized matrix and then aggregates the results.
+
+    All nodes will have the same exact result after this layer.
+
+    It is the quantized equivalent of :class:`mlx.nn.ShardedToAllLinear`.
+    Similar to :class:`mlx.nn.QuantizedLinear` its parameters are frozen and
+    will not be included in any gradient computation.
+
+    Args:
+        input_dims (int): The dimensionality of the input features.
+        output_dims (int): The dimensionality of the output features.
+        bias (bool, optional): If set to ``False`` then the layer will not use
+            a bias. Default: ``True``.
+        group_size (int, optional): The group size to use for the quantized
+            weight. See :func:`~mlx.core.quantize`. Default: ``64``.
+        bits (int, optional): The bit width to use for the quantized weight.
+            See :func:`~mlx.core.quantize`. Default: ``4``.
+        group (mx.distributed.Group, optional): The sharding will happen across
+            this group. If not set then the global group is used. Default is
+            ``None``.
+    """
+
+    def __init__(
+        self,
+        input_dims: int,
+        output_dims: int,
+        bias: bool = True,
+        group_size: int = 64,
+        bits: int = 4,
+        group: Optional[mx.distributed.Group] = None,
+    ):
+        super().__init__()
+
+        # Quantization config
+        self.group_size = group_size
+        self.bits = bits
+
+        # Initialize the quantized weight
+        scale = math.sqrt(1.0 / input_dims)
+        self.group = group or mx.distributed.init()
+        N = self.group.size()
+
+        if (input_dims % N) != 0:
+            raise ValueError(
+                f"The input of size {input_dims} cannot be sharded across {N} devices."
+            )
+
+        weight = mx.random.uniform(
+            low=-scale,
+            high=scale,
+            shape=(output_dims, input_dims // N),
+        )
+        self.weight, self.scales, self.biases = mx.quantize(weight, group_size, bits)
+
+        # And bias if needed
+        if bias:
+            self.bias = mx.zeros((output_dims,))
+
+        # Freeze this model's parameters
+        self.freeze()
+
+    def unfreeze(self, *args, **kwargs):
+        """Wrap unfreeze so that we unfreeze any layers we might contain but
+        our parameters will remain frozen."""
+        super().unfreeze(*args, **kwargs)
+        self.freeze(recurse=False)
+
+    def _extra_repr(self) -> str:
+        out_dims, in_dims = self.weight.shape
+        in_dims *= (32 // self.bits) * self.group.size()
+        return (
+            f"input_dims={in_dims}, output_dims={out_dims}, bias={'bias' in self}, "
+            f"group_size={self.group_size}, bits={self.bits}"
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = mx.quantized_matmul(
+            x,
+            self["weight"],
+            scales=self["scales"],
+            biases=self["biases"],
+            transpose=True,
+            group_size=self.group_size,
+            bits=self.bits,
+        )
+        x = mx.distributed.all_sum(x, group=self.group)
+        if "bias" in self:
+            x = x + self["bias"]
+        return x
+
+    @classmethod
+    def from_quantized_linear(
+        cls,
+        quantized_linear_layer: Module,
+        *,
+        segments: Union[int, list] = 1,
+        group: Optional[mx.distributed.Group] = None,
+    ):
+        group = group or mx.distributed.init()
+        output_dims, input_dims = quantized_linear_layer.weight.shape
+        input_dims *= 32 // quantized_linear_layer.bits
+
+        sl = cls(
+            input_dims,
+            output_dims,
+            hasattr(quantized_linear_layer, "bias"),
+            group_size=quantized_linear_layer.group_size,
+            bits=quantized_linear_layer.bits,
+            group=group,
+        )
+        sl.update(
+            _shard(
+                quantized_linear_layer.parameters(),
+                _sharded_to_all(segments),
+                group,
+            )
+        )
+
+        return sl