Fix loading of BF16 bare nn.Parameters (LTX-2 learnable_registers)

[hc20k]

Problem

LTX-2 GGUFs (sulphur-2-distilled, ltx-2.3-*) fail to load:

RuntimeError: Error(s) in loading state_dict for LTXAVModel:
  While copying the parameter named "audio_embeddings_connector.learnable_registers",
  whose dimensions in the model are torch.Size([128, 2048]) and whose dimensions in
  the checkpoint are torch.Size([128, 4096]), an exception occurred :
  ('only Tensors of floating point dtype can require gradients',).

Root cause

learnable_registers is a bare nn.Parameter on a plain module (comfy/ldm/lightricks/embeddings_connector.py), not a Linear/Conv/Embedding/LayerNorm. GGMLOps never wraps it, so GGMLLayer._load_from_state_dict never runs and it reaches torch's default path, which builds a Parameter(..., requires_grad=...).

These tensors are BF16. loader.py only reshapes F32/F16, so BF16 stays as the raw uint8 mmap buffer — 2 bytes/element, hence the doubled last dim. The BF16 rescue directly below only covers len(shape) <= 1, and these are 2D. Torch therefore receives uint8 and raises.

The reported dim mismatch is a red herring: GGMLTensor.shape returns the logical shape (so torch's shape check passes), while the error text prints input_param.size() — the raw byte size.

Fix

GGMLLayer can only ever intercept weight/bias — everything else goes to unexpected_keys. So any other key must reach torch as a real float tensor.

Testing

On sulphur-2-distilled-Q4_K_M.gguf (arch ltxv, 4444 tensors):

• Changes exactly 2 tensors. Of 292 non-weight/bias keys, 290 are F32 (skipped by the is_quantized guard) and 2 are the BF16 registers.
• ~3 MB fp32 in a 14.1 GB model; 1772 tensors remain lazily quantized.
• UnetLoaderGGUF now loads LTXAVModel; registers land as bfloat16 at (128, 2048) / (128, 4096).
• Other archs unaffected — their non-weight/bias keys are F32, so is_quantized is False.

Note

#392 addresses this same class of bug for lumina2 by matching key names. Happy to narrow this to learnable_registers specifically if you'd prefer the surgical approach.