Reset torch default device to cpu after offload.profile(): fixes device mismatches in econdataset/pymaf/smplx code that assumes CPU-default tensor creation, without affecting mmgps explicit hook-based device management

Add GPU memory diagnostic print to confirm whether size=xlarge is actually being granted

Try mmgp VerylowRAM_LowVRAM profile: LowRAM_LowVRAM still OOMs during attention

Fix remaining comment typo

Fix comment typo (apostrophe escaping) from previous commit

Exclude image_normalizer from mmgp offload management (its .scale() method bypasses mmgps forward-hook, stranding buffers on CPU); pin it on GPU directly instead

Fix create_mean_pose() to build its numpy array from CPU tensors explicitly, instead of masking it with a global torch default-device reset that broke mmgp's own device handling

Reset torch default device to cpu after offload.profile(): mmgp leaves it as cuda, which broke SMPLDataset tensor creation later in the script

Use MMGP budget-based offloading/quantization instead of a blanket .to('cuda'): the pipeline genuinely OOMs on this ZeroGPU MIG slice's effective VRAM

Disable expandable_segments CUDA allocator: it hits an NVML assertion on MIG-sliced Blackwell GPUs

Keep PSHuman's custom multiview attention processors but swap their inner xformers.ops.memory_efficient_attention call for torch's native scaled_dot_product_attention (xformers' Hopper kernel crashes on Blackwell/sm_120)

Drop xformers: its Hopper-specific flash-attention kernel crashes on Blackwell (sm_120) with 'CUDA error: invalid argument'. torch 2.8's native SDPA (used automatically by diffusers) replaces it.

Update inference.py

Update inference.py

Migrated from GitHub