[megatron] fix: dpskv3 convert src and dst mixed up bug

[Yangruipis]

Checklist Before Starting

• Searched for similar PR(s).
• Checked PR Title format
  • In format of: [modules] type: Title
  • modules are in fsdp, megatron, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data
  • type is in feat, fix, refactor, chore
  • can involve multiple modules, seperated by , or space, like [megatron, fsdp, doc] feat: xxx

What does this PR do?

• fix DeepseekV3 convert bug introduced from [megatron] feat: robust and efficient mcore converter with meta device init and numel check for dpsk #1995 which mixed up the src and dst parameters of function safe_copy. appologize for my mistake

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluatuion results, etc.

High-Level Design

Demonstrate the high-level design if this PR is complex.

Specific Changes

List the specific changes.

API

Demonstrate how the API changes if any.

Usage Example

Provide usage example(s) for easier usage.

# Add code snippet or script demonstrating how to use this 

Checklist Before Submitting

• Read the Contribute Guide.
• Apply pre-commit checks.
• Add [BREAKING] to the PR title description if it breaks any API.
• Update the documentation about your changes in the docs.
• New CI unit test(s) are added to cover the code path.
• Rely on existing unit tests on CI that covers the code path.

[Yangruipis]

@ETOgaosion need a review and an emergency merge 🚑 🚑 🚑

[vermouth1992]

Is there any way to implement a unit test for this function?

[Yangruipis]

Is there any way to implement a unit test for this function?

sure, i'm willing to add one in next pr

[ETOgaosion]

@vermouth1992 I think it's hard to judge whether to-megatron's convertion is right, since there is no reference to compare tensor values.

[Yangruipis]

@vermouth1992 I think it's hard to judge whether to-megatron's convertion is right, since there is no reference to compare tensor values.

maybe we can extract the api of from_hf_to_mcore(which is used in model pre conversion) and from_mcore_to_hf(which is used in sharding manager weight sync), and test if the weight is exactly the same after the two conversion

[ETOgaosion]

@Yangruipis Cool! You are smart~ Welcome to contribute.