[ComfyUI v0.23.0] Sampling hangs at 0/N, GPU pinned 100%, with PromptRelay (LTX-2.3, Windows) — works fine on v0.22.0 #18

Summary

After upgrading ComfyUI v0.22.0 → v0.23.0, my LTX-2.3 image-to-video workflow that uses PromptRelay hangs during sampling:
SamplerCustomAdvanced stays at 0/8, GPU at 100%, and never advances. The exact same workflow completes in ~500s on ComfyUI
v0.22.0. Reverting ComfyUI to v0.22.0 fixes it completely.

Environment

• ComfyUI: 0.23.0 (with the deps it requires: comfy-aimdo 0.4.7, comfy-kitchen 0.2.10, frontend 1.44.19)
• PromptRelay: latest main (commit ca5d4e3, 2026-05-20)
• OS: Windows 11
• GPU: RTX 5090 32GB (Blackwell), driver 595.79
• Python 3.11.8, PyTorch 2.10.0+cu128, CUDA 12.8
• Attention: SageAttention 2.2.0.post4, launched with --use-sage-attention --disable-dynamic-vram
• Model: LTX-2.3 (LTXAV) I2V, 1920x1088, 511 pixel frames (64 latent frames), 8 steps

Symptom (log)

Models load fine, then:
[PromptRelay] Built penalty matrix (scaled): Lq=34170, Lk=1024, nonzero=13041243/34990080
[PromptRelay] Built penalty matrix (scaled): Lq=426, Lk=1024, nonzero=162567/436224
0%| | 0/8 [00:00<?, ?it/s]
…then it sits at 0/8 indefinitely with GPU at 100% (not a deadlock — it's computing something but never advancing). Model loaded
partially: loaded partially; 23237 MB loaded, 799 MB offloaded, lowvram patches: 103.

Expected

Sampling progresses and the video generates, as on v0.22.0 (~500s).

What I've already ruled out

• Not #14250 / pinned memory: models load completely (no HostBuffer.read_file_slice error); adding --disable-pinned-memory
  does not help.
• Not the new SDPA backend: v0.23.0 added a Windows cuDNN-first SDPA priority in comfy/ops.py, and attention_pytorch now
  routes through comfy.ops.scaled_dot_product_attention. I monkey-patched comfy.ops.scaled_dot_product_attention back to plain
  torch.nn.functional.scaled_dot_product_attention (bypassing the cuDNN-first priority) — still hangs. So the cuDNN backend is
  not the cause.

Suspected area (for your investigation)

The hang is in PromptRelay's masked cross-attention path. For LTX, apply_patches() wraps each block's attn2.forward via
add_object_patch, then _make_ltx_mask_wrapper injects the penalty matrix as an additive mask and forces attention_pytorch
via optimized_attention_override. On v0.23.0 this large masked cross-attn (Lq≈34170, Lk≈1024) never completes. v0.23.0 made large
changes to comfy/model_patcher.py (add_object_patch / get_model_object), comfy/ops.py, and comfy/model_management.py
(loader/offload) — any of which could be interacting badly with the patch + additive-mask path.

I have not isolated a minimal repro without PromptRelay (a stock LTX-2.3 workflow), so I can't yet confirm whether this is
PromptRelay-specific or a broader v0.23.0 + LTX masked-attention regression.

Notes

Happy to provide the full workflow JSON, full startup log, or test any patch you suggest.