Fix distributedType -> distributed_type typo in downcast_bf16 guard - #4149

Summary

The downcast_bf16 validation in Accelerator.__init__ references self.state.distributedType, which does not exist — the attribute is distributed_type:

if (
    (mixed_precision != "bf16")
    and getattr(self.state, "downcast_bfloat", False)
    and (self.state.distributedType != DistributedType.XLA)   # <-- typo
):
    raise ValueError("Can only use `downcast_bf16` when using `mixed_precision='bf16'` and on a TPU")

self.state is an AcceleratorState, whose attribute is distributed_type (used one line above at ~520: self.state.distributed_type == DistributedType.FSDP, and throughout state.py). distributedType appears only on this line in the entire src/ tree.

Impact

Because of short-circuit evaluation, this line is reached when mixed_precision != "bf16" and downcast_bfloat is set. In that case — a user enabling downcast_bf16 without mixed_precision="bf16" — the code raises:

AttributeError: 'AcceleratorState' object has no attribute 'distributedType'

instead of the intended, actionable ValueError that explains the misconfiguration.

Fix

-            and (self.state.distributedType != DistributedType.XLA)
+            and (self.state.distributed_type != DistributedType.XLA)

One-token typo fix; the guard now raises the intended ValueError.

(Happy to add a regression test for the downcast_bf16-without-bf16 path if you'd like — let me know the preferred fixture for constructing that state.)