[QEff Finetune] Adding dataset padding changes #478
+133
−31
Conversation
This file contains hidden or 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
quic-swatia
requested review from
quic-rishinr,
ochougul,
quic-hemagnih and
quic-amitraj
as code owners
Signed-off-by: Swati Allabadi <sallabad@qti.qualcomm.com>
quic-swatia
force-pushed
the
dataset_padding
branch
from
b8104e1 to
adebe02
Compare
Signed-off-by: Swati Allabadi <sallabad@qti.qualcomm.com>
added 2 commits
June 27, 2025 18:27
Signed-off-by: Swati Allabadi <sallabad@qti.qualcomm.com>
Signed-off-by: Swati Allabadi <sallabad@qti.qualcomm.com>
quic-meetkuma
suggested changes
Jun 30, 2025
| ) | ||
| if num_dummy_samples_per_batch > 0: | ||
| num_dummy_samples += num_dummy_samples_per_batch | ||
| loss = loss * train_config.val_batch_size / num_dummy_samples_per_batch |
There was a problem hiding this comment.
Why train_config.val_batch_size?
Signed-off-by: Swati Allabadi <sallabad@qti.qualcomm.com>
quic-swatia
force-pushed
the
dataset_padding
branch
from
6dc7950 to
8d725ab
Compare
quic-mamta
requested changes
Jun 30, 2025
quic-meetkuma
suggested changes
Jul 1, 2025
| kwargs["collate_fn"] = DataCollatorForSeq2Seq(dataset_processer) | ||
| return kwargs | ||
|
|
||
|
|
||
| def get_num_ddp_devices(): |
There was a problem hiding this comment.
This is generic function, not specific to dataset_utils. This should be ideally in helper.py.
| ) | ||
| if num_dummy_samples_per_batch > 0: | ||
| num_dummy_samples += num_dummy_samples_per_batch | ||
| loss = loss * train_config.val_batch_size / num_dummy_samples_per_batch |
| num_replicas = 1 | ||
| if train_config.enable_ddp: | ||
| num_replicas = dist.get_world_size() | ||
| remainder = len(dataset) % (num_replicas * train_config.train_batch_size) |
There was a problem hiding this comment.
We are padding based on train bs. What happens in case of test dataset with valid_batch_size?
| if num_dummy_samples_per_batch > 0: | ||
| num_dummy_samples += num_dummy_samples_per_batch | ||
| loss = loss * train_config.val_batch_size / num_dummy_samples_per_batch | ||
|
|
There was a problem hiding this comment.
Can we not generalize the logic in L492 to L502?
I mean else case considers bs>1 use case with 'labels' shape of [batch, seq]. The same logic should be true for bs=1 use case. Why we need separate logic.
This same can be done in train case. Try to avoid unnecessary if conditions.
Signed-off-by: Swati Allabadi <sallabad@qti.qualcomm.com>
…c to steps. Signed-off-by: Swati Allabadi <sallabad@qti.qualcomm.com>
Signed-off-by: Swati Allabadi <sallabad@qti.qualcomm.com>
quic-swatia
pushed a commit
to quic-swatia/efficient-transformers
that referenced
this pull request
Jul 4, 2025
…h code before PR quic#478 Signed-off-by: Swati Allabadi <sallabad@qti.qualcomm.com>
quic-swatia
force-pushed
the
dataset_padding
branch
from
1da94fd to
2239119
Compare
quic-mamta
approved these changes
Jul 4, 2025
quic-meetkuma
approved these changes
Jul 4, 2025
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.
Padding the dataset with dummy samples (they won't contribute in total_loss) to make the #samples a multiple of degree of ddp*batch_size) in case of