Fix Hunyuan 3D 2.1 multi-GPU worksplit: use cond_or_uncond instead of hardcoded chunk(2)

[Kosinkadink]

Problem

Hunyuan 3D 2.1 throws ValueError: not enough values to unpack (expected 2, got 1) when using multi-GPU worksplit (MultiGPU CFG Split node).

The model's forward() method hardcodes context.chunk(2, dim=0) assuming both cond and uncond are always batched together. In multi-GPU worksplit mode, a device may only receive one of them, so chunk(2) on a batch of 1 returns only 1 chunk.

Fix

Use transformer_options["cond_or_uncond"] to detect whether both cond and uncond are present in the batch. Only perform the half-swap when both are present (len == 2 and set == {0, 1}). When only one is present, skip the reordering entirely (identity operation).

Changes

• comfy/ldm/hunyuan3dv2_1/hunyuandit.py: Replace hardcoded chunk(2) with conditional swap gated on cond_or_uncond

Performance (dual RTX 4090)

Sampling-only speed (isolating the accelerated portion)

Resolution	Single GPU	Multi GPU	Sampling Speedup
4096	3.58 it/s	6.38 it/s	1.78x
8192	1.62 it/s	3.10 it/s	1.91x

End-to-end wall time (includes non-accelerated VAE decode + mesh generation)

Resolution	Single GPU	Multi GPU	Overall Speedup
4096	22.79s	19.22s	1.19x
8192	35.76s	28.02s	1.28x

Sampling-only speedup is near-theoretical 2x for CFG split across 2 GPUs. Overall speedup is lower because VAE decode and mesh generation are not accelerated by worksplit.

Benchmarks: 2 warmup runs + 5 timed runs, batch_size=1, 30 steps, euler sampler.