Ensure proxy and server images are only provided with --use-pathways. #100
Conversation
RoshaniN
force-pushed
the
roshanin-check-use-pathways
branch
from
9a719a3 to
24f8d5b
Compare
RoshaniN
force-pushed
the
roshanin-check-use-pathways
branch
from
570f4ae to
86bdd91
Compare
| --proxy-server-image=<Pathways proxy server image> \ | ||
| --docker-name='user-workload' \ | ||
| --docker-image=<maxtext docker image> \ | ||
| --command='bash /usr/pathways/ifrt/maxtext_entrypoint.sh base_output_directory=<output directory> dataset_path=<dataset path> per_device_batch_size=1 enable_checkpointing=false enable_profiler=false remat_policy=full global_parameter_scale=4 steps=300 max_target_length=2048 use_iota_embed=true reuse_example_batch=1 dataset_type=synthetic attention=flash gcs_metrics=True run_name=$(USER)-pw-xpk-test-1' |
There was a problem hiding this comment.
Curious if you have a simpler example. This seems like a complicated README case example to maintain.
Like in the non-pathways workload command example we just print hello world. Is that possible here too?
There was a problem hiding this comment.
Thanks for the question, Victor! Let me check if other dependencies are up-to-date. If so, I can simplify it to more like "python3 Maxtext...", however, we really want the workload to actually train!
There was a problem hiding this comment.
however, we really want the workload to actually train!
Agreed this is the ultimately goal but we should go one step at a time. How about we have both a simple workload and a MaxText example?
I want users to be able to verify they have xpk + system working before trying to train.
| --docker-name='user-workload' \ | ||
| --docker-image=<maxtext docker image> \ |
There was a problem hiding this comment.
--base-docker-image flow is more recommended for xpk users so that local changes can be iterated on. Is that possible for pathways?
There was a problem hiding this comment.
Currently, the docker-image provided by the users is a fixed docker image, but if the others deps (mentioned above) are merged, I can update docker-image to base-docker-image! I would also prefer that!
| @@ -4857,17 +4896,15 @@ def directory_path_type(value): | |||
| workload_pathways_workload_arguments.add_argument( | |||
| '--proxy-server-image', | |||
| type=str, | |||
| default='gcr.io/cloud-tpu-v2-images/pathways/pathways-demo:proxy_server', | |||
There was a problem hiding this comment.
lmk if you do want to keep the ability to have a default and we can think through how this would look!
There was a problem hiding this comment.
I don't think it is super important, as folks would need to replace it during iteration!
There was a problem hiding this comment.
during iteration
Can you go into this more? will customers need to replace this often? curious if we should make it simpler if so.
Removed a Pathways arg.
RoshaniN
force-pushed
the
roshanin-check-use-pathways
branch
from
86bdd91 to
24b9a11
Compare
|
I will be punting this PR for later. With the way parsers are configured, it is not straightfoward to ensure --use-pathways is provided when any of the other Pathways args are provided. I will resume when the requirements are clearer and this edge case takes priority. |
Fixes / Features
Testing / Documentation
Testing details.
Additionally, I did some manual tests to see that the user gets the right error message for -
Also manually tested that invalid docker images are detected and user is guided with an error message.