What

Moves skip_layer_mask to the activations' device at its consumption points in models/ltx_video/models/transformers/attention.py:

• in AttnProcessor2_0.__call__, at the reshape(batch_size, 1, 1) step — covers the AttentionSkip, AttentionValues, and Residual strategies downstream;
• in BasicTransformerBlock.forward, at the view(-1, 1, 1) step — covers the TransformerBlock strategy.

Why

Transformer3DModel.create_skip_layer_mask builds the mask with device=self.device. Under mmgp offloading the transformer's parameters live in CPU RAM between calls, so self.device reports cpu and the mask is created there, while activations run on cuda. Any STG generation that actually skips blocks (an all-ones mask is discarded early) crashes with:

File ".../models/ltx_video/models/transformers/attention.py", line 1132, in __call__
  hidden_states_a *= skip_layer_mask
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

Reproduced with LTX Video 0.9.8 13B (ltxv_0.9.8_13B_dev_quanto_bf16_int8.safetensors) on Windows, torch 2.10.0+cu130, mmgp 3.7.6. Fixing at the consumption points (rather than at mask creation) keeps the fix correct wherever mmgp happens to be keeping the weights at the time.

Notes for review

• fix: AMD gfx1201 (RDNA4/ROCm) — INT8 Triton f32 MFMA, LTX Video device fix, validate_settings KeyError #1822 patches the same crash but only the AttentionSkip and AttentionValues branches, bundled with unrelated AMD/ROCm changes. This PR is the standalone, complete version: the Residual and TransformerBlock branches have the identical latent bug and are covered here.
• The .to() is a no-op when the mask is already on the right device, so non-offloaded setups are unaffected.

🤖 Generated with Claude Code