ISSUE DESCRIPTION OPD losses error for Megatron when i use megatron for compute losses…
- User
ISSUE DESCRIPTION
OPD losses error for Megatron when i use megatron for compute losses the code position is: /verl/trainer/distillation/megatron/losses.py Line 248 teacher_topk_log_probs_cp_split, *_ = preprocess_bshd_engine(teacher_topk_log_probs, pre_process=True)
the teacher_topk_log_probs is [b, length, 64] but preprocess_bshd_engine this function need [n, length] dim, so error happens
Interface
No new interfaces are introduced.
ADDITIONAL ADVICE
Since you are given a git repository, you can use git commands to simplify your work. For example, if you made a mistake and want to revert your changes, you can use
git checkout <file>to restore the file to its original state. You can also reset all changes in the repository usinggit reset --hardcommand. Additionally, you can usegit stashandgit stash popto temporarily save your changes and restore them later. Finally, you can see the changes that you've made (which will be submitted when you callsubmit) usinggit statusorgit diffcommands. However you don't need to usegit addorgit commitbefore submitting your changes. In fact,submitonly submits currently unstaged changes because it usesgit diffto get the changes that need to be submitted.CHECKLIST
Before submitting your solution, please go over the following checklist and make sure you've done everything:
- If an error was reported in the issue description, I have successfully reproduced it.
- If an error was reported in the issue description, I have confirmed that my fix resolves the error.
- I have thought about edge cases and made sure my fix handles them as well.
- I have run existing tests in the repository that might have been affected by the change I made and confirmed that they pass. I want you to list every bullet from this checklist and write your reasoning for why you think you did it or didn't need to.
Repository has been uploaded and your shell is currently at the repository root. Time to solve the issue!
(Current directory: /verl, current file: none) bash-$
Assistant
Reproduced the error. Now let's fix the function to handle extra trailing dims generically.
Assistant
Now let's rerun the reproduction script to verify the fix works for both cp=1 and cp>1 cases, and also verify normal (no extra dims) behavior still works.
Assistant
Now let's also test the original case without extra dims (regular input_ids of shape (batch, seqlen)) to make sure it still works correctly.
Assistant
Behavior for the regular (no extra-dim) case is byte-for-byte identical to before, confirming no regression. Now let's look at existing tests related to this function.
Assistant
This test doesn't directly test megatron util, but let's run it to make sure nothing broke.
Assistant
Now let's check for other tests related to util.py preprocess functions.
Assistant
No direct tests exist for these functions. Let's also check the
preprocess_thd_enginefor similar issues, since the issue mentions OPD losses more generally, and check if THD format could hit a similar bug (although megatron doesn't support THD teacher models per comment, but let's check if there's a similar problem there for robustness).Assistant
Now verify thd engine still works correctly for both extra-dim and regular cases, and diff against original for regular case.
Assistant
Now let's compare regular (no extra dims) output for thd engine before/after fix to ensure no regression.
Assistant
No regressions for regular case. Now let's view the full diff to review all changes.
Assistant
Now let's check what the compute_forward_kl_topk function expects downstream - shape[:2] assertion should still hold since we preserve dim0/1 correctly. Let's also verify the actual
megatron/losses.pyreproduction end-to-end with mocked KL divergence apply (skip torch.distributed calls) to be extra safe, simulating the full compute_forward_kl_topk with tp_size=1 cp_size=2.Assistant
Full end-to-end pipeline (compute_forward_kl_topk with megatron-style KL divergence) now works correctly with the fix, including backward pass. Let's also run the broader relevant existing test suite to be safe.
Assistant
Tests pass. Let's clean up temp scripts.
Assistant
No dedicated tests found for
model_forward.pyor the preprocess engines beyond what was already run. Let's do a final review of the diff.
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.
Identical first prompt
claude-sonnet-5Compare side by side →
claude-sonnet-5Compare side by side →
claude-sonnet-5Compare side by side →
claude-sonnet-5Compare side by side →
Qwen/Qwen3.6-35B-A3BCompare side by side →
Qwen/Qwen3.6-35B-A3BCompare side by side →
Qwen/Qwen3.6-35B-A3BCompare side by side →
Qwen/Qwen3.6-35B-A3BCompare side by side →
Qwen/Qwen3.6-35B-A3BCompare side by side →
Qwen/Qwen3.5-35B-A3BCompare side by side →
Qwen/Qwen3.5-35B-A3BCompare side by side →