Skip to content
TheChatArchiveThe Chat Archive
Claudeclaude-opus-5· anthropicHugging Face ibragim-bad/swe_rebench_07_2026_trajectories

ISSUE DESCRIPTION OPD losses error for Megatron when i use megatron for compute losses…

  1. 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 using git reset --hard command. Additionally, you can use git stash and git stash pop to temporarily save your changes and restore them later. Finally, you can see the changes that you've made (which will be submitted when you call submit) using git status or git diff commands. However you don't need to use git add or git commit before submitting your changes. In fact, submit only submits currently unstaged changes because it uses git diff to 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-$

  2. Claude Assistant

    I'll start by exploring the relevant file.

  3. Claude Assistant

    Now let me apply the fix to support extra trailing dims.

  4. Claude Assistant

    Now fix the position_ids expansion which assumes 2D input.

  5. Claude Assistant

    Checklist review:

    • Reproduced the reported error: Yes. With a stubbed megatron.core.parallel_state (megatron isn't installed here), calling preprocess_bshd_engine with a jagged nested tensor of shape (b, seq, topk) — exactly what verl/trainer/distillation/megatron/losses.py passes for teacher_topk_log_probs/teacher_topk_ids — raised RuntimeError: expand(torch.FloatTensor{[3, 4]}, size=[3]) ..., because the function only allocated a (batch, seqlen) buffer.
    • Confirmed the fix resolves the error: Yes. preprocess_bshd_engine now derives trailing_dims = input_ids.shape[2:] and allocates (batch, local_seqlen, *trailing_dims) buffers (both the output buffer and the per-sequence CP padding buffer), and position_ids is explicitly expanded to (batch, local_seqlen) instead of expand_as(input_ids_bshd). I verified with a script that for cp∈{1,2}×tp∈{1,2} and every cp_rank, the 3-D result equals stacking per-column 2-D results (so zigzag CP splitting, masks and position ids are unchanged/correct), and then ran the full compute_forward_kl_topk (megatron losses) end-to-end on CPU with a gloo process group for both bshd and thd formats — shapes (bsz, seqlen) are produced and backward gives finite gradients.
    • Edge cases: 2-D inputs (normal token ids) behave exactly as before (trailing_dims == ()), need_roll, FP8 padding, and NPU mask paths untouched. I also generalized the analogous CP-alignment padding in preprocess_thd_engine (d.numel() → d.shape[0] with trailing dims preserved), which was silently wrong for N-D nested inputs (padding was skipped when seqlen*topk >= align_size); this was caught by my per-column equivalence test for the thd path. Integer/float dtypes both work (ids vs. logprobs).
    • Ran existing tests: tests/workers/test_distillation_topk_symmetry_on_cpu.py passes (4 passed). Other mcore tests require megatron/GPUs and cannot run in this environment. ruff check/ruff format are clean on the changed file. Temporary scripts were removed; only verl/models/mcore/util.py is modified.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.