Commit b2f78686 authored by Derek Murray's avatar Derek Murray Committed by TensorFlower Gardener
Browse files

Always use the local worker name in CreateWorkerSession when not doing ClusterSpec propagation.

Previously, the master would send a job name and task index in an
otherwise-empty ServerDef, and the worker would unquestioningly use
those to build its worker name. However, this would lead to errors if
the worker had a local name like "/job:worker/replica:1/task:0",
because the ServerDef doesn't support non-zero replica IDs, and so the
local worker would end up an inconsistent view of what its worker name
should be. In particular `WorkerSession::worker_name` would disagree
with the device names added during graph partitioning by the master,
which would lead to runtime failures ("InvalidArgumentError: Invalid
rendezvous key").

PiperOrigin-RevId: 193733855
parent cadbb0b7
Loading
Loading
Loading
Loading
0% Loading or .
You are about to add 0 people to the discussion. Proceed with caution.
Please to comment