Add _cuda_init call-stack debug patch to locate ZeroGPU trigger
daKhosa
fdb3e41
Disable Triton in worker to fix ZeroGPU _cuda_init block during generation
daKhosa
faf9489
Triton kernel compilation (used by Quanto INT8) calls torch._C._cuda_init()
mid-generation, which ZeroGPU blocks after the 1-second lease expires.
Fix: stub sys.modules['triton'] = None at the start of _worker_entrypoint
so Quanto falls back to its pure-PyTorch INT8 path — no _cuda_init after
lease expiry. Restore multiprocessing.Process + @spaces.GPU(duration=1)
POC pattern from POC_HF.
Fix ZeroGPU GPU access: switch from subprocess.Popen to multiprocessing.Process
daKhosa
afd702d
The spaces library only patches multiprocessing.Process.start() to register
child PIDs with the GPU broker. subprocess.Popen children are not registered
and get no GPU access (CUDA init fails with "No CUDA GPUs are available").
Changes:
- generate.py: extract run_generation() callable + _worker_entrypoint for
multiprocessing pickling; add immediate torch.zeros(1, device='cuda') right
after signal received to establish CUDA context before lease expires
- app.py: import _worker_entrypoint, use multiprocessing.get_context('spawn')
.Process instead of subprocess.Popen; sleep 0.8s in GPU lease to give
worker time to init CUDA; proc.join() instead of proc.wait()
Switch to ZeroGPU POC pattern: duration=1 billing, subprocess retains CUDA context
daKhosa
b840bd7
- app.py: pre-spawn generate.py from main process, @spaces.GPU(duration=1) only
for CUDA init signal; stream incremental logs from JSON result file
- generate.py: add --signal-path polling, --result-path incremental JSON logging,
add distillation LoRA (ltx-2.3-22b-distilled-lora-384.safetensors at 0.5),
experimental LoRA at 1.0; guidance_scale default 5.0
Capture stderr (tqdm/progress) in log stream alongside stdout
daKhosa
f5b01ae
Persist WanGPSession across generations; eliminate per-call model reload
daKhosa
deab82a
Replace subprocess approach with a cached WanGPSession. Session is created
once on the first @spaces.GPU call and reused for all subsequent generations,
keeping the model in memory. Logs are streamed via a queue + worker thread.
Replace Frames slider with Seconds (1-5s); convert to frames internally at 25fps
daKhosa
36ea2bf
Single-quote 'Experimental' in subtitle
daKhosa
3c276bc
Update subtitle to reflect experimental LoRA v1
daKhosa
1776eeb
Pre-download ltx-2.3-22b-distilled-lora-384 required by experimental LoRA
daKhosa
8d83ec8
Switch to experimental Sulphur LoRA v1
daKhosa
9d642bd
Untrack 10Eros subfolder; ignore it from Sulphur repo
daKhosa
ab07e2b
Add 10Eros Space (TenStrip/LTX2.3-10Eros layer-scaled merge)
daKhosa
4bd6143
Pre-download Gemma at startup; bump guidance to 5.0 for prompt adherence
