MCPcopy Create free account
hub / github.com/MotrixLab/AiOS / allreduce_grads

Function allreduce_grads

detrsmpl/utils/dist_utils.py:34–44  ·  view source on GitHub ↗
(params, coalesce=True, bucket_size_mb=-1)

Source from the content-addressed store, hash-verified

32
33
34def allreduce_grads(params, coalesce=True, bucket_size_mb=-1):
35 grads = [
36 param.grad.data for param in params
37 if param.requires_grad and param.grad is not None
38 ]
39 world_size = dist.get_world_size()
40 if coalesce:
41 _allreduce_coalesced(grads, world_size, bucket_size_mb)
42 else:
43 for tensor in grads:
44 dist.all_reduce(tensor.div_(world_size))
45
46
47class DistOptimizerHook(OptimizerHook):

Callers

nothing calls this directly

Calls 1

_allreduce_coalescedFunction · 0.85

Tested by

no test coverage detected