[trainer] fix: handle empty tensors in extract_multi_modal_inputs

[yurekami]

Summary

Fixes #4500

This PR fixes a RuntimeError: torch.cat(): expected a non-empty list of Tensors that occurs when processing mixed text and multi-modal batches where some micro-batches contain only text data.

Problem

When training multi-modal models with mixed datasets (some samples have images, some are text-only), text-only samples may have empty lists or empty tensors for multi-modal keys like pixel_values. The extract_multi_modal_inputs function was collecting these empty values and passing them to torch.cat(), causing the crash.

Error trace:

RuntimeError: torch.cat(): expected a non-empty list of Tensors
  at verl/utils/model.py line 748 (extract_multi_modal_inputs)

Solution

Added validation to skip empty tensors and lists when collecting multi-modal inputs:

1. Skip tensors with numel() == 0
2. Skip empty lists/tuples with len() == 0
3. Add safety check before torch.cat to handle edge cases

Changes

• verl/utils/model.py: Updated extract_multi_modal_inputs() function

Example

# Before: This would crash
batch_data = [
    {"pixel_values": torch.tensor([...])},  # Multi-modal sample
    {"pixel_values": []},  # Text-only sample - empty list
    {"pixel_values": None},  # Text-only sample - None
]
result = extract_multi_modal_inputs(batch_data)  # RuntimeError!

# After: Works correctly
result = extract_multi_modal_inputs(batch_data)  # Returns concatenated tensors

Test plan

• Syntax validation passes
• Unit test with mixed text/multi-modal batch
• Integration test with SFT training on mixed dataset

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

The pull request modifies the extract_multi_modal_inputs function in verl/utils/model.py to enhance input filtering. It now explicitly skips None values, empty torch.Tensor objects, and empty lists/tuples when collecting multi-modal inputs. This change aims to correctly handle mixed batches that may contain text-only samples, preventing empty multi-modal inputs from being processed. Additionally, a new check was added to skip processing any collected multi-modal input key if no valid values were appended to its list.