LTX-2.5: BF16 non-layer parameters remain GGMLTensor and expose packed storage dimensions

[Dracoform]

Environment

• Windows 11
• ComfyUI via Stability Matrix Portable
• AMD Radeon RX 7900 XT 20 GB
• ComfyUI-GGUF commit:
  6ea2651e7df66d7585f6ffee804b20e92fb38b8a

Model

elix3r/LTX-2.5-22b-distilled-GGUF

ltx-2.5-22b-distilled-transformer-Q5_K_M.gguf

GGUF qtypes:

F32  (2906)
BF16 (3)
Q5_K (1072)
Q6_K (368)

The three BF16 tensors are:

audio_embeddings_connector.learnable_registers
logical: (128, 2048)
raw:     (128, 4096)

video_embeddings_connector.learnable_registers
logical: (128, 4096)
raw:     (128, 8192)

keyframes_abs_pos_embedding
logical: (1, 4096)
raw:     (1, 8192)

These BF16 tensors remain GGMLTensor objects.

Unlike normal GGMLLayer weights, LTX-2.5 uses these parameters directly in ordinary torch operations.

This causes the packed BF16 storage size to leak into runtime operations.

I encountered the following errors:

RuntimeError:
Tensors must have same number of dimensions: got 4 and 3

and later:

RuntimeError:
The size of tensor a (4096) must match tensor b (8192)
at non-singleton dimension 2

Workaround

Immediately after gguf_sd_loader(), I materialized only these three BF16 tensors using dequantize_tensor() before passing the state dict into load_diffusion_model_state_dict().

Result

• Text conditioning works

• video connector works

• audio connector works

• keyframes_abs_pos_embedding works

• first sampler: 8/8

• second sampler: 3/3

• AudioVAE loads

• CausalDiffusionVAE loads

• complete workflow finishes successfully

  Prompt executed in 00:10:58

I have a small proof-of-concept patch that materializes only those three named LTX-2.5 BF16 parameters. I can open a PR if this is the preferred fix location.