Optimized CSM-1B TTS pipeline for RTX 5090 (Blackwell sm_120). CUDA graph replay via patched HF Transformers. ~0.46x RTF. Topics (tags): csm text-to-speech rtx-5090 blackwell cuda-graphs torch-compile sesame streaming pytorch
text-to-speech streaming pytorch tts sesame csm huggingface blackwell torch-compile rtx-5090 sm-120 cuda-graphs
-
Updated
Apr 5, 2026 - Python