You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
- [ ] Inject an env var into hyperparameter to tell the selector script which entrypoint to choose
Inject the entire command passed in from Flytekit as one env var
Change the image lookup logic to take runnable container into consideration
Change the Output Path prefix to RawOutputPrefix
Build a separate CustomTrainingJob plugin, in which we need to process each literal in the input literal map, and inject inputs properly: inject blob-like inputs to the input channels, and inject others as command-line arguments (We currently do not know how properly pass complicated data type such as Generic (structs or maps) to SageMaker. An easy way might be using JSON string and let pyflyte-execute-alt use JSON parser when the input is a Generic type.)
Let Flytekit take care of the output handling for now. Do not write anything to the output channel (i.e., do not write to the /opt/ml/model path in the container)
Testing end2end
Handling the flyte container's args and environment
The only way to pass non-blob-type inputs to our container via SageMaker is to use the hyperparameter field of the SageMaker CRD. With that, the inputs will be passed in as command-line arguments. That means we have to pass in the container's args and environment via command line.
* updated flytestdlib
Signed-off-by: Daniel Rammer <[email protected]>
* set flytestdlib to new release
Signed-off-by: Daniel Rammer <[email protected]>
…yteorg#451)
Setting different values in multiple parts of the configs applied resulted in the least specific spec being applied
Added extra tests verifying mainly interruptible override fallbacks including application configuration
Signed-off-by: Nick Müller <[email protected]>
* updated flytestdlib
Signed-off-by: Daniel Rammer <[email protected]>
* set flytestdlib to new release
Signed-off-by: Daniel Rammer <[email protected]>
…yteorg#451)
Setting different values in multiple parts of the configs applied resulted in the least specific spec being applied
Added extra tests verifying mainly interruptible override fallbacks including application configuration
Signed-off-by: Nick Müller <[email protected]>
Related PRs
flyteorg/flyteplugins#113
- [ ] Inject an env var into hyperparameter to tell the selector script which entrypoint to choose(We currently do not know how properly pass complicated data type such as Generic (structs or maps) to SageMaker. An easy way might be using JSON string and letpyflyte-execute-alt
use JSON parser when the input is a Generic type.)Handling the flyte container's
args
andenvironment
The only way to pass non-blob-type inputs to our container via SageMaker is to use the hyperparameter field of the SageMaker CRD. With that, the inputs will be passed in as command-line arguments. That means we have to pass in the container's
args
andenvironment
via command line.To pass in
env1=val1 env2=val2 service_venv pyflyte-execute --task-module blah --task-name bloh --output-prefix s3://fake-bucket --inputs s3://fake-bucket
Our flyteplugin need to rewrite it into and our runner script needs to parse it from the following format:
We added the prefix and suffix in order to prepare for the future hyperparameter jobs.
The text was updated successfully, but these errors were encountered: