Context Parallel SFT Support for dataset in THD format #10688
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
Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com>
Signed-off-by: tomlifu <tomlifu@users.noreply.github.com>
tomlifu
changed the title
xrennvidia
reviewed
Oct 3, 2024
nemo/collections/nlp/models/language_modeling/megatron_gpt_model.py
Outdated
Show resolved
Hide resolved
nemo/collections/nlp/data/language_modeling/megatron/gpt_sft_dataset.py
Outdated
Show resolved
Hide resolved
nemo/collections/nlp/models/language_modeling/megatron_gpt_model.py
Outdated
Show resolved
Hide resolved
nemo/collections/nlp/data/language_modeling/megatron/gpt_sft_dataset.py
Outdated
Show resolved
Hide resolved
xrennvidia
reviewed
Oct 3, 2024
Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com>
… thd_cp_support
Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com>
Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com>
Signed-off-by: tomlifu <tomlifu@users.noreply.github.com>
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
nemo/collections/nlp/models/language_modeling/megatron_gpt_model.py
Fixed
Show resolved
Hide resolved
xrennvidia
reviewed
Oct 18, 2024
nemo/collections/nlp/models/language_modeling/megatron_gpt_model.py
Fixed
Show resolved
Hide resolved
nemo/collections/nlp/data/language_modeling/megatron/gpt_sft_dataset.py
Outdated
Show resolved
Hide resolved
nemo/collections/nlp/data/language_modeling/megatron/gpt_sft_dataset.py
Outdated
Show resolved
Hide resolved
Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com>
Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com>
|
Please fix DCO also. |
Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com>
tomlifu
force-pushed
the
thd_cp_support
branch
from
0339a8f to
90feb4a
Compare
Signed-off-by: tomlifu <tomlifu@users.noreply.github.com>
Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com>
… thd_cp_support
Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com>
Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com>
Signed-off-by: tomlifu <tomlifu@users.noreply.github.com>
tomlifu
force-pushed
the
thd_cp_support
branch
from
6e8ac3b to
463a478
Compare
|
Could you let me know when it will be completed? I’ve been really looking forward to this feature. It works in pretrain, but it’s really strange that it doesn’t work in SFT. |
cuichenx
reviewed
Nov 22, 2024
Signed-off-by: root <root@cw-dfw-h100-001-074-012.cm.cluster>
Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com>
xrennvidia
reviewed
Nov 26, 2024
Signed-off-by: tomlifu <tomzhanglf@gmail.com>
Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com>
Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com>
Signed-off-by: tomlifu <tomlifu@users.noreply.github.com>
|
beep boop 🤖: 🙏 The following files have warnings. In case you are familiar with these, please try helping us to improve the code base. Your code was analyzed with PyLint. The following annotations have been identified: Thank you for improving NeMo's documentation! |
8 tasks
XuesongYang
pushed a commit
to paarthneekhara/NeMo
that referenced
this pull request
Jan 18, 2025
* Add context parallel support for packed dataset in THD format * add changes with debug print * remove debug print Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * Apply isort and black reformatting Signed-off-by: tomlifu <tomlifu@users.noreply.github.com> * fix cu_seqlens and cu_seqlens_padded Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * cu_seqlens and cu_seqlens_padded fix Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * more fix on cu_seqlens and cu_seqlens_padded Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * Apply isort and black reformatting Signed-off-by: tomlifu <tomlifu@users.noreply.github.com> * addressing Xiaowei's review Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * addressing more review comments Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * fix for the case where cp=1 Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * Apply isort and black reformatting Signed-off-by: tomlifu <tomlifu@users.noreply.github.com> * more fix to address Xiaowei's comment Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * fix the loss_mask for THD formated data Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * Apply isort and black reformatting Signed-off-by: tomlifu <tomlifu@users.noreply.github.com> * fixed eos_idx[0][0] out of bounds issue Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * fixed CP=1 case Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * fixed thd_get_partitioned_indices assertion issue when pp=1 Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * Apply isort and black reformatting Signed-off-by: tomlifu <tomlifu@users.noreply.github.com> * fixed data packing issue Signed-off-by: root <root@cw-dfw-h100-001-074-012.cm.cluster> * fixed an issue where cp>1 has different loss curves Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * remove redudant check for cu_seqlens Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * fixed NeMo CI failure issue due to old TE version in CI Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * Apply isort and black reformatting Signed-off-by: tomlifu <tomlifu@users.noreply.github.com> --------- Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> Signed-off-by: tomlifu <tomlifu@users.noreply.github.com> Signed-off-by: root <root@cw-dfw-h100-001-074-012.cm.cluster> Signed-off-by: tomlifu <tomzhanglf@gmail.com> Co-authored-by: tomlifu <tomlifu@users.noreply.github.com> Co-authored-by: root <root@cw-dfw-h100-001-074-012.cm.cluster> Co-authored-by: Xiaowei Ren <103958965+xrennvidia@users.noreply.github.com>
youngeunkwon0405
pushed a commit
to youngeunkwon0405/NeMo
that referenced
this pull request
Feb 10, 2025
* Add context parallel support for packed dataset in THD format * add changes with debug print * remove debug print Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * Apply isort and black reformatting Signed-off-by: tomlifu <tomlifu@users.noreply.github.com> * fix cu_seqlens and cu_seqlens_padded Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * cu_seqlens and cu_seqlens_padded fix Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * more fix on cu_seqlens and cu_seqlens_padded Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * Apply isort and black reformatting Signed-off-by: tomlifu <tomlifu@users.noreply.github.com> * addressing Xiaowei's review Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * addressing more review comments Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * fix for the case where cp=1 Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * Apply isort and black reformatting Signed-off-by: tomlifu <tomlifu@users.noreply.github.com> * more fix to address Xiaowei's comment Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * fix the loss_mask for THD formated data Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * Apply isort and black reformatting Signed-off-by: tomlifu <tomlifu@users.noreply.github.com> * fixed eos_idx[0][0] out of bounds issue Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * fixed CP=1 case Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * fixed thd_get_partitioned_indices assertion issue when pp=1 Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * Apply isort and black reformatting Signed-off-by: tomlifu <tomlifu@users.noreply.github.com> * fixed data packing issue Signed-off-by: root <root@cw-dfw-h100-001-074-012.cm.cluster> * fixed an issue where cp>1 has different loss curves Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * remove redudant check for cu_seqlens Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * fixed NeMo CI failure issue due to old TE version in CI Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> * Apply isort and black reformatting Signed-off-by: tomlifu <tomlifu@users.noreply.github.com> --------- Signed-off-by: Lifu Zhang <tomzhanglf@gmail.com> Signed-off-by: tomlifu <tomlifu@users.noreply.github.com> Signed-off-by: root <root@cw-dfw-h100-001-074-012.cm.cluster> Signed-off-by: tomlifu <tomzhanglf@gmail.com> Co-authored-by: tomlifu <tomlifu@users.noreply.github.com> Co-authored-by: root <root@cw-dfw-h100-001-074-012.cm.cluster> Co-authored-by: Xiaowei Ren <103958965+xrennvidia@users.noreply.github.com> Signed-off-by: Youngeun Kwon <youngeunk@nvidia.com>
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 ?
This PR adds CP support for THD format and is compatible with cu_seqlen_padded in the latest CUDNN fused attention.
Steps to run SFT + CP + THD format:
scripts/nlp_language_modeling/prepare_packed_ft_dataset.pyto pack the dataset into THD format in desired sequence length. For example:PR Type:
If you haven't finished some of the above items you can still open "Draft" PR.