Data parallel helper (#1407)

mlx/distributed/mpi/mpi.cpp

@@ -32,8 +32,29 @@ array ensure_row_contiguous(const array& arr) {
   }
 }
 
+template <typename T>
+void simple_sum(
+    void* input,
+    void* accumulator,
+    int* len,
+    MPI_Datatype* datatype) {
+  T* in = (T*)input;
+  T* acc = (T*)accumulator;
+  int N = *len;
+
+  while (N-- > 0) {
+    *acc += *in;
+    acc++;
+    in++;
+  }
+}
+template void simple_sum<float16_t>(void*, void*, int*, MPI_Datatype*);
+template void simple_sum<bfloat16_t>(void*, void*, int*, MPI_Datatype*);
+
 struct MPIWrapper {
   MPIWrapper() {
+    initialized_ = false;
+
     libmpi_handle_ = dlopen("libmpi.dylib", RTLD_NOW | RTLD_GLOBAL);
     if (libmpi_handle_ == nullptr) {
       return;
@@ -50,6 +71,9 @@ struct MPIWrapper {
     LOAD_SYMBOL(MPI_Allgather, all_gather);
     LOAD_SYMBOL(MPI_Send, send);
     LOAD_SYMBOL(MPI_Recv, recv);
+    LOAD_SYMBOL(MPI_Type_contiguous, mpi_type_contiguous);
+    LOAD_SYMBOL(MPI_Type_commit, mpi_type_commit);
+    LOAD_SYMBOL(MPI_Op_create, mpi_op_create);
 
     // Objects
     LOAD_SYMBOL(ompi_mpi_comm_world, comm_world_);
@@ -79,7 +103,24 @@ struct MPIWrapper {
     if (!is_available()) {
       return false;
     }
-    return init(nullptr, nullptr) == MPI_SUCCESS;
+    bool success = init(nullptr, nullptr) == MPI_SUCCESS;
+
+    // Initialize custom types and ops
+    if (success && !initialized_) {
+      // Custom float16 dtypes
+      mpi_type_contiguous(2, mpi_uint8_, &mpi_float16_);
+      mpi_type_commit(&mpi_float16_);
+      mpi_type_contiguous(2, mpi_uint8_, &mpi_bfloat16_);
+      mpi_type_commit(&mpi_bfloat16_);
+
+      // Custom sum ops
+      mpi_op_create(&simple_sum<float16_t>, 1, &op_sum_f16_);
+      mpi_op_create(&simple_sum<bfloat16_t>, 1, &op_sum_bf16_);
+
+      initialized_ = true;
+    }
+
+    return success;
   }
 
   void finalize_safe() {
@@ -117,13 +158,21 @@ struct MPIWrapper {
       case complex64:
         return mpi_complex_;
       case float16:
+        return mpi_float16_;
       case bfloat16:
-        throw std::runtime_error("MPI doesn't support 16-bit floats");
+        return mpi_bfloat16_;
     }
   }
 
-  MPI_Op op_sum() {
-    return op_sum_;
+  MPI_Op op_sum(const array& arr) {
+    switch (arr.dtype()) {
+      case float16:
+        return op_sum_f16_;
+      case bfloat16:
+        return op_sum_bf16_;
+      default:
+        return op_sum_;
+    }
   }
 
   void* libmpi_handle_;
@@ -152,6 +201,8 @@ struct MPIWrapper {
 
   // Ops
   MPI_Op op_sum_;
+  MPI_Op op_sum_f16_;
+  MPI_Op op_sum_bf16_;
 
   // Datatypes
   MPI_Datatype mpi_bool_;
@@ -165,6 +216,16 @@ struct MPIWrapper {
   MPI_Datatype mpi_uint64_;
   MPI_Datatype mpi_float_;
   MPI_Datatype mpi_complex_;
+  MPI_Datatype mpi_float16_;
+  MPI_Datatype mpi_bfloat16_;
+
+ private:
+  bool initialized_;
+
+  // Private API
+  int (*mpi_type_contiguous)(int, MPI_Datatype, MPI_Datatype*);
+  int (*mpi_type_commit)(MPI_Datatype*);
+  int (*mpi_op_create)(MPI_User_function*, int, MPI_Op*);
 };
 
 MPIWrapper& mpi() {
@@ -276,7 +337,7 @@ void all_sum(Group group, const array& input_, array& output) {
       output.data<void>(),
       input.size(),
       mpi().datatype(input),
-      mpi().op_sum(),
+      mpi().op_sum(input),
       to_comm(group));
 }
 

python/mlx/nn/utils.py

@@ -1,10 +1,11 @@
 # Copyright © 2023-2024 Apple Inc.
 
-from functools import wraps
-from typing import Callable, Optional
+from functools import reduce, wraps
+from typing import Any, Callable, Optional
 
 import mlx.core as mx
 
+from ..utils import tree_flatten, tree_map, tree_unflatten
 from .layers.base import Module
 
 
@@ -68,3 +69,93 @@ def wrapped_checkpointed_fn(*args, **kwargs):
         return checkpointed_fn(module.trainable_parameters(), *args, **kwargs)
 
     return wrapped_checkpointed_fn
+
+
+def average_gradients(
+    gradients: Any,
+    group: Optional[mx.distributed.Group] = None,
+    all_reduce_size: int = 32 * 1024**2,
+    communication_type: Optional[mx.Dtype] = None,
+):
+    """Average the gradients across the distributed processes in the passed group.
+
+    This helper enables concatenating several gradients of small arrays to one
+    big all reduce call for better networking performance.
+
+    Args:
+        gradients (Any): The Python tree containing the gradients (it should
+            have the same structure across processes)
+        group (Optional[mlx.core.distributed.Group]): The group of processes to
+            average the gradients. If set to ``None`` the global group is used.
+            Default: ``None``.
+        all_reduce_size (int): Group arrays until their size in bytes exceeds
+            this number. Perform one communication step per group of arrays. If
+            less or equal to 0 array grouping is disabled. Default: ``32MiB``.
+        communication_type (Optional[mlx.core.Dtype]): If provided cast to this
+            type before performing the communication. Typically cast to a
+            smaller float to reduce the communication size. Default: ``None``.
+    """
+    group = group or mx.distributed.init()
+    N = group.size()
+
+    if N == 1:
+        return gradients
+
+    def _average(x):
+        dt = x.dtype
+        x = x.astype(communication_type) if communication_type is not None else x
+        return mx.distributed.all_sum(x, stream=mx.cpu).astype(dt) / N
+
+    if all_reduce_size <= 0:
+        return tree_map(_average, gradients)
+
+    else:
+        flat_grads = tree_flatten(gradients)
+        if len(flat_grads) == 0:
+            return gradients
+
+        # Extract some info for the gradient
+        keys = [k for k, _ in flat_grads]
+        shapes = [v.shape for _, v in flat_grads]
+        sizes = [v.size for _, v in flat_grads]
+        dtypes = [v.dtype for _, v in flat_grads]
+
+        # We can't group them if they have mixed types
+        if not all(dt == dtypes[0] for dt in dtypes):
+            return average_gradients(gradients, group, 0, communication_type)
+        itemsize = (
+            communication_type.size
+            if communication_type is not None
+            else dtypes[0].size
+        )
+
+        # Gather the gradients in groups that are just above or equal to all_reduce_size
+        grad_groups = []
+        grad_group = []
+        grad_group_size = 0
+        for i in range(len(keys)):
+            grad_group.append(i)
+            grad_group_size += sizes[i] * itemsize
+            if grad_group_size >= all_reduce_size:
+                grad_groups.append(grad_group)
+                grad_group = []
+                grad_group_size = 0
+        if grad_group:
+            grad_groups.append(grad_group)
+            grad_group = []
+
+        # Concatenate-reduce-split
+        new_flat_grads = []
+        for grad_group in grad_groups:
+            indices = reduce(lambda x, y: x + [x[-1] + sizes[y]], grad_group, [0])
+            big_grad = mx.concatenate(
+                [flat_grads[i][1].reshape(-1) for i in grad_group]
+            )
+            big_grad = _average(big_grad)
+            big_grad = mx.split(big_grad, indices[1:-1])
+            new_flat_grads.extend(
+                (keys[j], big_grad[i].reshape(shapes[j]))
+                for i, j in enumerate(grad_group)
+            )
+
+        return tree_unflatten(new_flat_grads)

python/tests/mpi_test_distributed.py

@@ -4,6 +4,7 @@
 
 import mlx.core as mx
 import mlx_tests
+from mlx.nn.utils import average_gradients
 
 
 class TestDistributed(mlx_tests.MLXTestCase):
@@ -110,6 +111,59 @@ def test_send_recv(self):
 
         self.assertTrue(mx.all(x == (1024 if pairs.rank() == 0 else 512)))
 
+    def test_average_gradients(self):
+        original_all_sum = mx.distributed.all_sum
+        n_calls = 0
+        xtype = None
+
+        def new_all_sum(x, **kwargs):
+            nonlocal n_calls
+            nonlocal xtype
+
+            n_calls += 1
+            if xtype is not None:
+                self.assertEqual(xtype, x.dtype)
+
+            return original_all_sum(x, **kwargs)
+
+        mx.distributed.all_sum = new_all_sum
+
+        try:
+            grads = [mx.ones(10) for i in range(10)]
+            new_grads = average_gradients(grads)
+            mx.eval(new_grads)
+            self.assertEqual(len(new_grads), 10)
+            self.assertTrue(all(mx.all(g == 1) for g in new_grads))
+            self.assertEqual(n_calls, 1)
+
+            n_calls = 0
+            new_grads = average_gradients(grads, all_reduce_size=4 * 50)
+            mx.eval(new_grads)
+            self.assertEqual(len(new_grads), 10)
+            self.assertTrue(all(mx.all(g == 1) for g in new_grads))
+            self.assertEqual(n_calls, 2)
+
+            n_calls = 0
+            new_grads = average_gradients(grads, all_reduce_size=0)
+            mx.eval(new_grads)
+            self.assertEqual(len(new_grads), 10)
+            self.assertTrue(all(mx.all(g == 1) for g in new_grads))
+            self.assertEqual(n_calls, 10)
+
+            n_calls = 0
+            xtype = mx.float16
+            new_grads = average_gradients(
+                grads, all_reduce_size=2 * 50, communication_type=mx.float16
+            )
+            mx.eval(new_grads)
+            self.assertEqual(len(new_grads), 10)
+            self.assertTrue(all(g.dtype == mx.float32 for g in new_grads))
+            self.assertTrue(all(mx.all(g == 1) for g in new_grads))
+            self.assertEqual(n_calls, 2)
+
+        finally:
+            mx.distributed.all_sum = original_all_sum
+
 
 if __name__ == "__main__":
     unittest.main()