Multi-GPU interactive mode hangs during model initialization on some machines

[FVEFWFE]

Running Matrix-Game 3.0 interactive mode with torchrun on 4x A100 SXM 80GB.
Stock generate.py with --interactive works on some machines but silently hangs
during model initialization on others.

Setup:

• torch 2.10.0, FA2, CUDA 12.8, Python 3.12
• torchrun --nproc_per_node=4 with --ulysses_size 4, --use_int8, --t5_cpu
• Docker image based on nvidia/cuda:12.8.0-devel-ubuntu22.04

What happens:

• NCCL init succeeds on all ranks
• T5 encoder loads successfully on all ranks
• All 4 ranks log "Initializing Model (DiT)..."
• Then hangs forever: 0% CPU, 0% GPU VRAM (633MB CUDA context only)
• No error, no timeout, no crash

What we've tried:

• NCCL_P2P_DISABLE=1
• NCCL_DEBUG=INFO (init shows COMPLETE, no errors)
• device_id in init_process_group()
• 2 GPUs instead of 4
• Different volume sizes (75GB, 200GB)

Key observation:

• Works on some RunPod machines (e.g. vqnzttsdu44t, c250s631h6rc)
• Hangs on others (e.g. cmg18urp1kg5, hleuzkw9vyyn)
• Same Docker image, same code, same config

The hang appears to be inside WanModel.from_pretrained() or the
dist.barrier() that precedes it. Has anyone else seen this?

[FVEFWFE]

disabling int8 fixed this