TP support for reverse KL loss #400
Open
+297
−76
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
jlamypoirier
reviewed
Dec 3, 2025
fast_llm/functional/cross_entropy.py
Outdated
| else: | ||
| predicted_logits = (target * logits_norm).sum(dim=-1, keepdim=True) | ||
|
|
||
| # shouldn't the predicted_logits be scaled by the number of ranks so that the average loss is correct? i.e. |
Collaborator
There was a problem hiding this comment.
There is no averaging. We calculate log Z - sum_ranks (sum_i t_i * z_i_rank * mask) which is the same as log Z - sum_i t_i * z_i.
| MAX_DROPLESS_BLOCK_SIZE_ROW = 128 | ||
|
|
||
|
|
||
| class ReverseKLImpl(str, enum.Enum): |
Collaborator
There was a problem hiding this comment.
Not needed, there is only one implementation.
| "beta_fast": config["beta_fast"], | ||
| "beta_slow": config["beta_slow"], | ||
| "original_context_length": config["original_max_position_embeddings"], | ||
| "attention_factor": config["rope_scaling"]["attention_factor"], |
Collaborator
There was a problem hiding this comment.
Should be the same for llama 3 above?
|
|
||
| @pytest.mark.parametrize("use_mask", [True, False]) | ||
| def test_cross_entropy_vocab_tp_two_ranks(use_mask): | ||
| _spawn_dist(2, _ce_vocab_tp_worker, use_mask) |
Collaborator
There was a problem hiding this comment.
Need to mark as slow (how long does this take?). Does it work on cpu? Also it would be better to run all tests in the same spawn call because of the huge distributed overhead.
jlamypoirier
reviewed
Dec 3, 2025
| # then we average: 1/K sum_ranks (log Z - sum_i t_i * z_i) | ||
| # = log Z - 1/K sum_ranks (sum_i t_i * z_i) | ||
| # but sum_ranks (sum_i t_i * z_i) = sum_i t_i * z_i (over all vocab) | ||
| predicted_logits = predicted_logits * group.size() |
Collaborator
There was a problem hiding this comment.
This looks wrong, see previous comment. The previous version was tested and confirmed to work.
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.
TP support for reverse KL loss.
TODO:
test_rkl_loss.pyshould probably be integrated in the testing framework?🔍 Type of change
Select all that apply:
📝 Changes
✅ Checklist
Make sure the following tasks are completed before submitting the PR:
General
Dependencies and Configuration
Testing
Performance Impact
📊 Performance Impact Details
If there is any impact on performance, describe it and provide benchmark results, if applicable:
🗒️ Additional Notes
Include any additional context, information, or considerations here, such as known issues, follow-up tasks, or backward compatibility concerns.