[model] fix: refactor qwen2vl patches & support no-image input for fsdp

[hiyouga]

What does this PR do?

This PR tries to fix #3491

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include 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
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

Tested with latest transformers

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

⚠️ Breaking

We adopt a new format for Qwen2VL's position ids: (4, batch size, seq len)

Assuming a vision position ids (mrope) has a shape of (3, batch size, seq len) and a text position ids (normal rope) has a shape of (1, batch size, seq len), we concatenate both to obtain the final position ids.

This aligns with the implementation in the Transformers >= 4.54.0 🤗
https://github.com/huggingface/transformers/blob/v4.54.0/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py#L1469

🎤 New

We have refactored the Qwen2VL and Qwen2.5VL patches, supporting no-image input for FSDP by introducing fake ViT inputs. We have also removed some redundant code for better maintainability.

🚨 Changes

We move the ulysses logic into the attention function. So the position ids will be scattered before the language model part.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)

[gemini-code-assist]

Code Review

This pull request refactors the monkey patching for Qwen2-VL models by consolidating logic from qwen2_5_vl.py into qwen2_vl.py and removing the former. It also adds support for text-only inputs with FSDP by creating dummy image tensors. The refactoring is a good step towards reducing code duplication. However, I've found a few critical bugs and areas for improvement. Specifically, there's a NameError in monkey_patch.py due to an incorrect import order, and a TypeError in qwen2_vl.py from subscripting a None value. I've also noted the use of magic numbers and missing fields in return objects which could affect maintainability and functionality like cached generation. Please see my detailed comments for suggestions.

[techkang]

Great job! I have some questions on this bug, can you share you ideas?

1. Do you know the root cause of this bug?
2. It seems that we can set warp_policy to None the same as https://github.com/volcengine/verl/blob/5c98ed1b313be5ae3de10a3e800cbb769374e1a5/verl/workers/fsdp_workers.py#L452-L454 and the training process can continue. Can we fix the bug this way?
3. If both method can fix this, which one is better?

[vermouth1992]

It seems that CI fails :(

[hiyouga]

@vermouth1992 This PR is ready for merge :)

[Clementine24]

Great job! This pr solve my confusion about mixed data training bugs. By the way, it seems that multi-turn sglang rollout generation still use old approach to get position ids?

[hiyouga]

@Clementine24 Possibly I forgot it. Could you submit a PR to fix it?