revert: back to PyTorch nightly cu128 (nightly + xformers) for Blackwell sm_120

fix: remove duplicate 'install' in pip command

fix: use pytorch.org stable cu128 index for guaranteed torch+torchvision ABI compatibility

fix: replace mmengine set_random_seed import with local impl (avoids PyTorch JIT LaTeX bug)

fix: remove cu128 index pin — let pip auto-select best stable PyTorch

fix: switch from PyTorch nightly to stable (2.7+ supports Blackwell, no NVML bug)

fix: switch to expandable_segments allocator (less fragmentation than no-caching)

fix: set PYTORCH_NO_CUDA_MEMORY_CACHING before any torch import (was too late inside run_inference)

fix: disable CUDACachingAllocator to workaround PyTorch nightly NVML bug on MIG GPU

fix: replace manual attention (OOM) with F.scaled_dot_product_attention (memory-efficient)

fix: enable cpu_offload to avoid OOM on 48GB MIG GPU

fix: add _HAS_FLASH_ATTN guard — mirror xformers pattern to prevent NoneType call

fix: default USE_XFORMERS to False (env var unreliable in ZeroGPU context)

fix: disable flash_attn globally (not available on Blackwell)

fix: disable xformers globally (Blackwell sm_120 not supported)

fix: use black placeholder when nuScenes images missing (HF Space doesnt need real images)

fix: set num_workers=0 in inference config files (was overriding script default)

fix: set num_workers=0 for ZeroGPU daemon process compatibility

fix maps download: repo has expansion/*.json not nuscenes/maps/expansion/*.json

add nuScenes maps/expansion download from doradream/magicdrive-data

fix: robustly download ema.pt LFS file + strict check_env

fix: from_pretrained must be a .pt file, not model/ dir

fix: fallback to nn.LayerNorm when apex missing

fix: disable apex FusedLayerNorm (use nn.LayerNorm)

fix: resolve model.from_pretrained placeholder

fix: move inference in-process to avoid ZeroGPU SIGSEGV (-11)

Fix stdout pipe deadlock: merge stderr→stdout, read in thread

Live progress bar: poll subprocess, update every 2s

Check for output video even on inference error

Add CUDA_LAUNCH_BLOCKING=1 to get real error traceback

Fix permission: use home dir instead of /data for marker

Separate PyTorch upgrade and mmcv checks

Create /data directory before writing marker file

Upgrade PyTorch to nightly for Blackwell sm_120 support

Switch dtype from bf16 to fp16

Reduce sampling steps to 15, disable CPU offload for speed

Disable xformers to avoid SIGSEGV from CUDA version mismatch

Set ZeroGPU timeout to 5 minutes (duration=300)

Fix numpy version conflict: numpy<2.0 + pickle compat shim

Bump numpy>=2.0.0 for pickle compatibility

Skip flash-attn: make import lazy (inference uses xformers)

Move flash-attn to runtime install (needs torch in build env)

Add flash-attn and safetensors to requirements.txt

Fix: remove torchvision pin, add transformers, patch write_video import

Pin torchvision==0.19.0 to fix write_video import error

Fix dataset downloads: add repo_type='dataset' for HF datasets

Add try/except and better error messages to download_weights

Add data pkl files to check_env and download list

Fix: always enable Generate button, check weights at runtime

Fix download: use Popen with streaming output instead of silent subprocess.run