Extend TGI integration tests #561
Conversation
dacorvo
force-pushed
the
extend_docker_tests
branch
2 times, most recently
from
60f7870
to
70cb0d0
Compare
This improves logs readability.
dacorvo
force-pushed
the
extend_docker_tests
branch
2 times, most recently
from
44986bf
to
4f24388
Compare
dacorvo
force-pushed
the
extend_docker_tests
branch
from
9a44aaf
to
c5bb24e
Compare
| """Expose a pre-trained neuron model | ||
|
|
||
| The fixture first makes sure the following model artifacts are present on the hub: | ||
| - exported neuron model under optimum/neuron-testing-<version>-<name>, |
There was a problem hiding this comment.
Why do you want to both cache it under optimum/neuron-testing-cache repo and store in a repo optimum/neuron-testing-<version>-<name> instead of just the cache repo?
There was a problem hiding this comment.
Because I want to test both fetching a neuron model from the hub and exporting it "on-the_fly" from the cached artifacts.
| For each exposed model, the local directory is maintained for the duration of the | ||
| test session and cleaned up afterwards. | ||
| The hub model artifacts are never cleaned up and persist accross sessions. | ||
| They must be cleaned up manually when the optimum-neuron version changes. |
There was a problem hiding this comment.
btw do we also need to clean the cache in the official neuron cache repo?
There was a problem hiding this comment.
The official cache is not affected by these files. It is however sometimes affected when running tests locally: the registry is updated for dev versions for instance.
I usually remove the registry files afterwards.
We could also remove older compiler trees.
| They must be cleaned up manually when the optimum-neuron version changes. | ||
|
|
||
| """ | ||
| config_name = request.param |
There was a problem hiding this comment.
it's "model_name" / "model_arch"?
There was a problem hiding this comment.
It is a free-form string used as the key in the model_configurations dictionary (here gpt2, llama).
What does this PR do?
This extends the TGI integration tests by running all tests on not only llama but also gpt2 model configurations.