Recalibrate GPU duration reservations, drop deprecated warnings

Replace torchvision.io.read_video with a PyAV-based shim

Save output WAVs via soundfile instead of torchaudio.save

Use bare PyPI torch/torchaudio 2.11.0 instead of +cu128 index build

Pin torch to exact cu128 build to match torchaudio

Bump torch/torchaudio to a Blackwell-compatible version, swap mmcv for mmcv-lite

Remove pinned spaces==0.48.2 causing a hard build failure

Add explicit CUDA availability diagnostic inside GPU worker calls

Force local_files_only on open_clip's own hf_hub_download reference

Move offline-mode patch inside each GPU function's own body

Re-assert offline mode inside the ZeroGPU worker itself

Fix HF_HUB_OFFLINE never actually taking effect

Fix spurious "Too many arguments provided for the endpoint" warnings

Fix Generate click never reaching server: watchdog broke the click event

Stop Generate from going silent: merge watchdog-arm into the click event

Fix Generate click going silent after adding the watchdog-arm step

Fix infinite retry loop in video upload transcode

Add client-side watchdog so a crashed generation can't wedge Generate

Prevent GPU OOM from crashing the whole Space container

Replace server-side slot-visibility updates with client-side JS

Disable Generate button and show busy label during generation

Patch Gradio's video-conversion race that kills the generation connection

Fix ffmpeg stream-copy crash in strip_audio_from_video

Bump spaces==0.48.1 -> 0.48.2 to resolve build conflict

Reroute encoders + MMAudio downloads through user-controlled mirror

Fix stereo conformance across all write paths

Fix xregen duration truncation: use full segment span, not model window

Fix xregen truncating total audio duration when clip < segment window

Remove channel-layout safety nets; assert stereo (2,T) everywhere

Normalize all seg wavs to stereo (2,T) at save time

Preserve MMAudio stereo; stereo-wins channel normalization

Fix (1,T) shape crash in _cf_join; bump HunyuanFoley secs_per_step to 0.5

Free GPU memory between HunyuanFoley segments to prevent OOM

Revert to single multi-seg GPU call; bump hunyuan load_overhead to 90s

Fix HunyuanFoley initial gen: one GPU call per segment

Fix HunyuanFoley: pre-download SigLIP2, use local_files_only=True

Fix MMAudio: load BigVGAN from local snapshot dir, not HF network

Fix HunyuanFoley: save text_feats to disk inside GPU worker

Fix HunyuanFoley xregen: register model lib sys.path in main process

Fix HunyuanFoley xregen AttributeError: return torch.device from _get_device_and_dtype

Add try/except traceback logging to _hunyuan_gpu_infer to surface AttributeError

Fix TARO GPU abort: increase load_overhead 15→120s to cover model loading time

Add granular step-by-step logging in _taro_gpu_infer to find exact GPU abort point

Fix pending regen waveform segment boundary misalignment

Fix MMAudio TypeError: cast num_steps to int in _mmaudio_gpu_infer

Fix TARO GPU worker crash: use local_files_only=True for audioldm2 VAE/vocoder

Add xregen error logging: try/except in _xregen_dispatch, logging in xregen_mmaudio._run and _mmaudio_gpu_infer to surface GPU worker crashes

Add verbose debug logging throughout TARO generation pipeline

Fix xregen ZeroGPU TypeError: duration fns must accept all positional args

Fix ZeroGPU tmp file isolation: extract clips inside GPU worker