[fix] remove dead code, fix typos, and eliminate unnecessary GPU-CPU syncs

[choiszt]

Motivation

During code review, identified several issues that hurt training performance and code quality:

• Unnecessary .item() calls causing GPU→CPU synchronization barriers in the forward pass
• Dead code (unused variables, duplicate function definitions, redundant assignments)
• A typo (torch.torch.int32) and a suboptimal comparison pattern

Modifications

Eliminate unnecessary GPU-CPU syncs

• qwen2_5_vl_liger.py: Removed 4 unused variables (tokens_count, n_image_tokens, n_video_tokens, visual_tokens)
  that each called .item(), causing GPU-CPU sync every forward pass
• qwen2_5_vl_ops.py: Removed 3 unused variables (bsz, q_len, kv_seq_len) in attn_forward, where
  torch.max(position_ids).item() also triggered GPU-CPU sync

Remove dead code

• qwen3_vl_liger.py: Removed redundant hidden_states = outputs[0] (already assigned on the line above)
• sequence_packing_utils.py: Removed duplicate _get_unpad_data function definition (identical copy)

Fix typo and optimize comparison

• sequence_packing_utils.py: Fixed torch.torch.int32 → torch.int32 in _unpad_input
• qwen3_vl_ops.py: Changed layer_idx in range(len(deepstack_visual_embeds)) → layer_idx < len(deepstack_visual_embeds)
  (logically equivalent, avoids creating a range object each iteration)

Verification

Verified with actual GPU training on 2x A100 (FSDP2 + Liger + rmpad):

Model	Size	Steps	Result
Qwen2.5-VL-3B	3.75B	10	✅ Training completed without errors
Qwen3-VL-4B	4.44B	10	✅ Training completed without errors

## Checklist

• Follow commit message convention
• Format code with black and isort
• Verified with GPU training (Qwen2.5-VL + Qwen3-VL)
• No functional changes — only dead code removal, typo fix, and perf improvement

[Copilot]

Pull request overview

This PR improves training performance and code quality by removing dead code and avoiding unnecessary GPU→CPU synchronization points, while also fixing a dtype typo and simplifying a small comparison.

Changes:

• Removed unused variables (and associated .item() calls) that introduced GPU→CPU syncs in Qwen2.5-VL and Qwen2.5-VL attention/forward paths.
• Removed redundant/dead assignments and a duplicate helper function definition.
• Fixed an invalid dtype reference (torch.torch.int32 → torch.int32) and simplified a layer-index bounds check.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
src/lmms_engine/models/sequence_packing_utils.py	Removes duplicate _get_unpad_data definition and fixes torch.int32 dtype usage in _unpad_input.
src/lmms_engine/models/qwen3_vl/qwen3_vl_ops.py	Simplifies deepstack layer gating condition to avoid range(...) membership check.
src/lmms_engine/models/qwen3_vl/qwen3_vl_liger.py	Removes a redundant hidden_states = outputs[0] assignment.
src/lmms_engine/models/qwen2_5_vl/qwen2_5_vl_ops.py	Removes unused variables that computed torch.max(position_ids).item() (sync point) inside attn_forward.
src/lmms_engine/models/qwen2_5_vl/qwen2_5_vl_liger.py	Removes unused token counting variables that called .item() in lce_forward.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.