fix: ensure contiguous tensors after slicing in QwenImage attention

[diodiogod]

Summary

Fixes sporadic RuntimeError: view size is not compatible with input tensor's size and stride that persists even after PR nunchaku-tech/nunchaku#673.

The Problem

Even with PR #673's fix (changing .view() to .reshape() in linear.py), users still experience sporadic crashes with the same error. The issue occurs when:

1. joint_hidden_states is sliced along the sequence dimension in qwenimage.py:319-320
2. These slicing operations create non-contiguous tensor views
3. Despite .reshape() theoretically handling non-contiguous tensors, certain stride patterns still cause failures

This explains why:

• The error is sporadic (depends on memory layout from prior operations)
• Only rebooting fixes it (clears GPU memory fragmentation)

The Fix

Add .contiguous() immediately after slicing operations to ensure tensors are copied to contiguous memory before being passed to nunchaku linear layers.

# Before
txt_attn_output = joint_hidden_states[:, :seq_txt, :]
img_attn_output = joint_hidden_states[:, seq_txt:, :]

# After
txt_attn_output = joint_hidden_states[:, :seq_txt, :].contiguous()
img_attn_output = joint_hidden_states[:, seq_txt:, :].contiguous()

Environment

• PyTorch 2.8.0+cu129
• ComfyUI 0.3.68

Questions for Maintainers

Why does .reshape() in linear.py:182 still fail despite being designed to handle non-contiguous tensors? Is there:

• A PyTorch 2.8 behavior change?
• A specific stride pattern that .reshape() can't handle?
• An issue in the nunchaku linear layer implementation?

This PR is defensive programming that prevents the crash, but the root cause may need investigation.

Closes #535