Skip to content
TheChatArchiveThe Chat Archive
DeepSeekdeepseek/deepseek-v4-pro· deepseekHugging 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

    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. DeepSeek Assistant

    Let me start by understanding the issue:

    1. There's an error in /verl/trainer/distillation/megatron/losses.py at line 248
    2. teacher_topk_log_probs has shape [b, length, 64] (3D tensor)
    3. But preprocess_bshd_engine expects [n, length] (2D tensor)
    4. So we need to fix the shape mismatch

    Let me first look at the file and understand the code.

  3. DeepSeek Assistant

    Let me look at the specific file mentioned in the issue.

  4. DeepSeek Assistant

    Let me scroll to around line 248.

  5. DeepSeek Assistant

    Now let me look at the preprocess_bshd_engine and preprocess_thd_engine functions to understand the expected input

  6. DeepSeek Assistant

    Let me search for preprocess_bshd_engine and preprocess_thd_engine in this file.

  7. DeepSeek Assistant

    Let me see the preprocess_bshd_engine function.

  8. DeepSeek Assistant

    Let me also look at the FSDP losses to see how it handles the same problem, and the preprocess_thd_engine function:

  9. DeepSeek Assistant

    Let me look at the preprocess_thd_engine function.

  10. DeepSeek Assistant

    Let me see the full preprocess_bshd_engine function.

  11. DeepSeek Assistant

    Let me see the rest of the preprocess_bshd_engine function.

  12. DeepSeek Assistant

    Now let me understand the full picture. Let me look at the FSDP losses to see how it handles the same case, and also understand how these nested tensors are created.

  13. DeepSeek Assistant

    Let me also check the rest of preprocess_thd_engine for the CP case, and understand the nested tensor shapes better

  14. DeepSeek Assistant

    I need to open the util.py file first.

  15. DeepSeek Assistant

    Let me try to reproduce the issue. First, let me understand the exact error by creating a test

  16. DeepSeek Assistant

    Now I understand the nested tensor shapes. Let me now reproduce the actual

  17. DeepSeek Assistant

    I've confirmed the issue. Let me now fix both preprocess_bshd_engine and preprocess_thd_engine to handle 3D inputs. Let me first fix preprocess_bshd_engine

  18. DeepSeek Assistant

    Let me just use the open command correctly.

  19. DeepSeek Assistant

    Now I have a clear picture. I need to fix both preprocess_bshd_engine and preprocess_thd_engine. Let me apply the fixes:

    Fix 1: preprocess_bshd_engine - Line 601 (input_ids_bshd shape)

  20. DeepSeek Assistant

    Fix 2: preprocess_bshd_engine - Line 612 (seq_padded shape

  21. DeepSeek Assistant

    Fix 3: preprocess_bshd_engine - Lines 641-642 and 654 (position_ids handling

  22. DeepSeek Assistant

    Now fix line 654 for the cp_size > 1

  23. DeepSeek Assistant

    Fix 4: preprocess_thd_engine - Lines 409-412 (padding for multi-dim tens

  24. DeepSeek Assistant

    Now fix the preprocess_thd_engine padding issue at lines 409-412

  25. DeepSeek Assistant

    Now let me verify all the changes look correct

  26. DeepSeek Assistant

    Let me verify the changes with a simple test that doesn't require megat

  27. DeepSeek Assistant

    Now let me also verify that the existing 2D behavior is preserved by testing with 2D nested

  28. DeepSeek Assistant

    Good, both 2D and 3D cases work. Let me also check if there are other callers that might be affected

  29. DeepSeek Assistant

    Let me check how callers use position_ids from preprocess_bshd_engine

  30. DeepSeek Assistant

    Let me see the broader context of how position_ids is

  31. DeepSeek Assistant

    The changes look correct and backward-compatible. Let me check if there are any existing tests we

  32. DeepSeek Assistant

    Let me check for tests related to

  33. DeepSeek Assistant

    Let me check the relevant test

  34. DeepSeek Assistant

    Let me check the Megatron-specific test

  35. DeepSeek Assistant

    The changes look correct and backward-compatible. Let me do a final review of all changes.

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