Cleaner UI (strip duplicated speaker-name labels, no third-person self-naming) + faster (shorter one-sentence lines, max_new_tokens 140)

Restore reliable streaming + per-line voice (revert combined-generator experiment that ZeroGPU would not dispatch)

Import voice at module load (pre-fork) to avoid import-lock deadlock inside ZeroGPU worker

Move combined text+voice streamer into model.py (importable) so ZeroGPU runs its @spaces.GPU generator body

Stream voice with text: synthesize each sentence as the line is written; VoxCPM2 resident + disk-cached refs

Fix MiniCPM5-1B generate (transformers 5.x dict inputs); disable torch.compile so VoxCPM2 runs eager

Add per-character VoxCPM2 voices; switch actors to MiniCPM5-1B; draining memory meter; save-highlights

Thousand-Token Theater: MiniCPM improv troupe with a hard 1,000-token memory

Thousand-Token Theater: MiniCPM improv troupe with a hard 1,000-token memory

Thousand-Token Theater: MiniCPM improv troupe with a hard 1,000-token memory

Thousand-Token Theater: MiniCPM improv troupe with a hard 1,000-token memory

Thousand-Token Theater: MiniCPM improv troupe with a hard 1,000-token memory

Thousand-Token Theater: MiniCPM improv troupe with a hard 1,000-token memory

Thousand-Token Theater: MiniCPM improv troupe with a hard 1,000-token memory

Thousand-Token Theater: MiniCPM improv troupe with a hard 1,000-token memory

Thousand-Token Theater: MiniCPM improv troupe with a hard 1,000-token memory