Streamline KD & QAD transformers Trainers #708
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
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #708 +/- ##
=======================================
Coverage 74.69% 74.69%
=======================================
Files 192 192
Lines 18946 18946
=======================================
Hits 14152 14152
Misses 4794 4794 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
AAnoosheh
requested review from
ChenhanYu,
kevalmorabia97,
kinjalpatel27,
mxinO and
realAsma
examples/llm_qat/README.md
Outdated
| > **_NOTE:_** `launch.sh` defaults to use `LlamaDecoderLayer` as the transformer layer class. If your model uses a different class, you need to pass `--fsdp_transformer_layer_cls_to_wrap <your_layer_class>` to the `launch.sh` script. For example, for `Qwen/Qwen3-8B`, specify `--fsdp_transformer_layer_cls_to_wrap Qwen3DecoderLayer` as an additional argument. | ||
| > **_NOTE:_** The script defaults to using FSDP1. To use FSDP2, pass "--use_fsdp2 True" to the `launch.sh` script. Note that FSDP2 is less stable than FSDP1 currently. Use it with caution. | ||
| > **_NOTE:_** The script defaults to using FSDP1. To use FSDP2, pass "--backend=fsdp2" to the `launch.sh` script. Note that FSDP2 is less stable than FSDP1 currently. Use it with caution. |
Collaborator
There was a problem hiding this comment.
Is this statement still valid? Note that FSDP2 is less stable than FSDP1 currently. Use it with caution.
Contributor
Author
There was a problem hiding this comment.
I doubt it, but I don't have proof. I don't have proof that it is less stable either.
kevalmorabia97
approved these changes
Dec 19, 2025
Signed-off-by: Asha Anoosheh <aanoosheh@nvidia.com>
Signed-off-by: Asha Anoosheh <aanoosheh@nvidia.com>
Signed-off-by: Asha Anoosheh <aanoosheh@nvidia.com>
Signed-off-by: Asha Anoosheh <aanoosheh@nvidia.com>
Signed-off-by: Asha Anoosheh <aanoosheh@nvidia.com>
Signed-off-by: Asha Anoosheh <aanoosheh@nvidia.com>
AAnoosheh
force-pushed
the
aanoosheh/kd-trainer-streamline
branch
from
bde5788 to
190e4d2
Compare
realAsma
reviewed
Dec 24, 2025
| model = model.export() if export_student else model | ||
| super().save_model(output_dir, _internal_call, *args, **kwargs) | ||
| with model.hide_teacher_model(), model.hide_loss_modules(enable=not _internal_call): | ||
| if _internal_call: |
Contributor
There was a problem hiding this comment.
Why do we need to save teacher if _internal_call = True? Can we avoid saving teacher irrespective of _internal_call value? This will accelerate checkpoint save/load (you would likely need to hide teach during the final checkpoint load as well)
realAsma
reviewed
Dec 24, 2025
| ``` | ||
|
|
||
| > **_NOTE:_** QAD currently requires quantization to be applied before the FSDP wrapper. Training is not supported for models that exceed single GPU memory capacity. | ||
| > **_NOTE:_** QAD doesn't support original FSDP - only FSDP2. It also requires quantization to be applied before the FSDP wrapper. |
Contributor
There was a problem hiding this comment.
Suggested change
| > **_NOTE:_** QAD doesn't support original FSDP - only FSDP2. It also requires quantization to be applied before the FSDP wrapper. | |
| > **_NOTE:_** QAD doesn't support FSDP1 (https://docs.pytorch.org/docs/stable/fsdp.html) - only FSDP2. It also requires quantization to be applied before the FSDP wrapper. |
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.
What does this PR do?
Type of change: ? Refactor and stabilization
Overview:
Usage
# Add a code snippet demonstrating how to use thisTesting
Before your PR is "Ready for review"
Additional Information