Remove verbose= kwarg from ReduceLROnPlateau: removed in newer PyTorch (we are now on 2.8.0)

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

Fix device mismatch: maskrcnn detector lands on cuda under mmgps lingering default device, but torch.from_numpy() input stays on cpu

Bump GPU duration to 600s: memory issue is fixed (attention slicing), task now just needs more time for the full multiview diffusion + mesh reconstruction pipeline

Implement attention slicing: chunk SDPA calls along the batch*heads dim since torch 2.8s fused kernels reject this shape on this GPU, forcing an O(seq_len^2 * batch) math fallback that OOMs even a 96GB card

Diagnose why SDPA auto-select avoids flash/efficient backends: try FLASH_ATTENTION explicitly and log the actual error, plus dtype/contiguity/mask info

Add memory diagnostics at each attention call to see if VRAM usage is growing unbounded or just spiking once

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

Request ZeroGPU xlarge (full 96GB Blackwell card) instead of default large (48GB half-card): the multiview attention activation memory OOMs on the half-card even with mmgp weight offloading

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

Bump diffusers to 0.29.0: mmgp (via optimum-quanto) needs PixArtTransformer2DModel, absent from the old 0.26.0 pin; verified all of PSHumans own diffusers imports (incl. LoRACompatibleConv/AdaLayerNorm) still resolve at 0.29.0

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

Revert forced EFFICIENT_ATTENTION backend (unsupported for these tensor shapes on this GPU); back to auto-selected SDPA

Force SDPA's memory-efficient backend for PSHuman's multiview attention: rule out a math-fallback OOM on Blackwell

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.

Bump warp-lang to 1.17.0: kaolin 0.18.0 physics subsystem imports warp.fem.linalg.inverse_qr, missing from the old 1.4.2 pin

Bump sympy to 1.13.3 and triton to 3.4.0: both hard-required by torch 2.8.0 (sympy>=1.13.3, triton==3.4.0 exactly on linux x86_64)

Bump typing_extensions to 4.12.2: torch 2.8.0 requires >=4.10.0, conflicting with the old 4.9.0 pin

Upgrade to torch 2.8.0+cu128 (adds Blackwell/sm_12x kernels) with matching kaolin 0.18.0, pytorch3d 0.7.8+pt2.8.0cu128, torch_scatter 2.1.2+pt28cu128, xformers 0.0.32.post2

Add nvidia-cuda-runtime + LD_LIBRARY_PATH shim so nvdiffrast's JIT-compiled CUDA extension can find libcudart.so.13 on Blackwell/CUDA-13 ZeroGPU workers

Fix torch pin to 2.1.2 (xformers==0.0.23.post1 requires torch==2.1.2 exactly, conflicting with the 2.1.0 I pinned previously)

Pin torch/torchvision/torchaudio to 2.1.0+cu121: unpinned torch was drifting to latest PyPI (now 2.14.0/cu13x), breaking ABI with the hard-pinned kaolin/pytorch3d/torch_scatter cu121 wheels (libcudart.so.13 not found)

Pin setuptools<82: pkg_resources was fully removed in setuptools 82, breaking nvdiffrast/torch.utils.cpp_extension import

Add setuptools so pkg_resources is available for nvdiffrast/torch.utils.cpp_extension import

Update requirements.txt

Update requirements.txt

increase ZeroGPU needed time for inference

includes ZeroGPU deco for main function

import spaces for ZeroGPU

set xformers 0.0.23.post1

Update requirements.txt

Update requirements.txt

Update requirements.txt

downgradio 5.8.0

Update requirements.txt

back to working properly config, psguman currently not compatible with Zero

Update requirements.txt

Update requirements.txt

Update requirements.txt

upgrade xformers==0.0.29.post1

back to torch modules 2.1.2 v

look for the right putorch3d

upgrade peft>=0.17.0