gpu assignment scaling #46
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
How about a unit test for the benchmarking? just to make sure the process doesn't take inordinate amounts of time for small to mid-size specs, and doesn't result in final penalty that are too crazy. |
|
Tagging @shaoxiongji for documentation purposes; this PR would add two flags to the config-config. The behavior, if I unerstand correctly, is as follows:
|
TimotheeMickus
approved these changes
Feb 6, 2024
If the heuristic only penalizes for the existence of empty GPUs, then reducing the number of empty GPUs by one will not be visible in the loss unless it is the last one. The penalty should be applied for each empty GPU individually.
Start by assigning a ready task to the first slot of each GPU
In this optmization iteration, one task per GPU is selected as a candidate. The selected slot is the one that is locally optimal to get rid of: i.e. not considering the destination of the swap.
It flickers past too fast anyhow.
Waino
force-pushed
the
feat/gpu_assignment_scaling
branch
from
fbb3cd5 to
735b096
Compare
|
Still LGTM, can merge |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Instead of considering all swaps of all gpu slots, make some educated guesses interleaved with random subsetting
Parameter
--time_budget_sallows setting a fixed time budget for the GPU assignmentFix a bug in the extra_empty_penalty: apply extra_empty_penalty for each empty GPU
If the heuristic only penalizes for the existence of empty GPUs, then
reducing the number of empty GPUs by one will not be visible in the loss
unless it is the last one. The penalty should be applied for each empty
GPU individually.
Some misc improvements to config-config, e.g. use tqdm for progress indicator instead of spewing garbage all over stdout.