Keep the transformer and Qwen text encoder off CUDA during initial load/quantization in low-VRAM mode so model startup avoids full-model OOM before offloading and quantization can take effect.
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Jaret Burkett <jaretburkett@gmail.com>
When training Klein models with a `control_path` (edit/kontext-style
paired datasets), `encode_image_refs()` returns tensors that reside on
the VAE's device (CPU, since the VAE weights are loaded via
`load_file(..., device="cpu")` and are never explicitly moved to the
training device). Concatenating those CPU tensors with the training
latents (`packed_latents`) that live on CUDA raises:
RuntimeError: Expected all tensors to be on the same device
Fix: move `img_cond_seq` and `img_cond_seq_ids` to the same device
(and dtype) as `img_input` / `img_input_ids` before concatenation.
Co-authored-by: HuangYuChuh <HuangYuChuh@users.noreply.github.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
