v0.14.0
What's new
Added 🎉
- Adds a function to modify a Hugging Face transformer with IA3 adaptors
- Added a
BeakerScheduler
registrable class, specified as the argumentscheduler
toBeakerExecutor
, which controls the resources assigned to steps ran on Beaker.
Users can implement their ownBeakerScheduler
subclasses to customize the resource assignment behavior.
Changed ⚠️
- In the
tango run
command,--no-server
is now the default. Use--server
to start the server.
Fixed ✅
- Made
BeakerExecutor
more robust to connection, timeout, SSL, and other recoverable HTTP errors. - Made the
BeakerStepLock
more robust, and as a resultBeakerWorkspace
is more
robust and should require less manual intervention for locks in a bad state. - Fixed a bug with the internal scheduling logic of the
BeakerExecutor
which
could delay submitting some steps in parallel. - Fixed a bug where creating a
StepInfo
object from params might result in unnecessary imports. - Fixed a bug where canceling the Beaker executor might not work properly.
- Fixed a bug where the trainer trains too much when
train_epochs
is set and you're using gradient accumulation. - Fixed how the results of uncacheable steps are displayed by
tango run
. - Beaker executor won't run duplicate cacheable steps at the same time.
Commits
0828adc BeakerExecutor
won't run duplicate cacheable steps (#414)
7382019 IA3 adaptors (#403)
d498cf7 Hot fix to final output
bff9ebf Add warning when steps can't be run yet, bug fixes (#408)
c72552e Don't start server by default (#409)
d34fe09 Added BeakerScheduler
class for handling resource assignment (#407)
5dcbb56 Gradient accumulation and train_epochs
(#402)
d27bbef Make BeakerStepLock
more robust (#401)
cd9b5fd Fix bug with StepInfo.from_params()
, canceling BeakerExecutor
, reserve "ref" name (#400)
6ff6b9e Fix bug with scheduling logic (#399)
15196f2 Deterministic hashing for tensors (#398)
230d78e Make BeakerExecutor
more robust to all recoverable errors types (connection, HTTP, SSL, timeout, etc) (#397)
5f63a27 Bump fairscale from 0.4.8 to 0.4.9 (#391)