Enable selection of GPU on inference page #1511
Conversation
rodrigoberriel
force-pushed
the
select_gpu_inference
branch
from
6d019e1
to
a1fa842
Compare
digits/model/forms.py
Outdated
| else get_device(index).totalGlobalMem) | ||
| ), | ||
| ) for index in config_value('gpu_list').split(',') if index], | ||
| default='next', |
There was a problem hiding this comment.
What happens here if there are no GPUs on the system?
There was a problem hiding this comment.
@lukeyeager the same behavior that you'd get if you were to start a training job without a GPU:
- If all setup was made with CPU only, you'd get the same page you currently get (without the multi-gpu form)
- If you manage to have Caffe compiled with CUDA and has no GPU on the system (suppose you removed it afterwards -- I tested it masking all my GPUs using an invalid id on
CUDA_VISIBLE_DEVICES), then you'd get the page without multi-gpu form and your job will fail exactly as the train job fails currently:
Check failed: error == cudaSuccess (38 vs. 0) no CUDA-capable device is detected
digits/model/images/generic/views.py
Outdated
| return flask.render_template( | ||
| 'models/images/generic/show.html', | ||
| form=generic_form, |
There was a problem hiding this comment.
I am a bit uneasy about passing the form here when all you need is the list of GPUs. Wouldn't it more explicit and self-explanatory to pass the list instead?
There was a problem hiding this comment.
I agree. Do you think this way is okay?
| {% if show_multi_gpu_form %} | ||
| <div class="col-sm-6"> | ||
| <div class="form-group{{mark_errors([form.select_one_of_gpus])}}"> | ||
| {{form.select_one_of_gpus.label}} |
There was a problem hiding this comment.
it's indeed better to use the WTForm fields here but it's inconsistent with the rest of the file... so I'm not sure about that...
There was a problem hiding this comment.
I've changed that too. What you think?
|
@gheinrich I'm on a deadline (March 27), but I could work on the unit tests after that. I have to say that I am not familiar with unit tests using flask, but I can give it a try. |
|
Hi, any update on the merge for this PR? |
This PR attempts to fix #1418, see #1418 (comment). I tried to minimize changes to avoid side-effects.
If the user doesn't have more than one GPU (or only one CUDA_VISIBLE_DEVICES for DIGITS), nothing is going to change:
But if the user has multiple GPUs (2 or more), it is going to look like this:
where
Next Availableis the default choice and has the same behavior as the current implementation. If the user selects a GPU (only one can be selected), then this GPU will be used.PS: I modified images/generic/show.html the same way I did for images/classification/show.html, but I didn't test the former as much as I did on the latter.