Title: RTX 5050 (Blackwell sm_120) Support - When is it coming to stable PyTorch?
Body:
Hello PyTorch community,
I recently purchased an RTX 5050 (Blackwell architecture with CUDA capability sm_120) and I’m having compatibility issues with the current stable PyTorch releases.
Current situation:
-
PyTorch 2.5.1 (stable): Detects CUDA but shows warning that sm_120 is not supported. When trying to run TTS models (XTTS/Coqui), I get:
CUDA error: no kernel image is available for execution on the device -
PyTorch 2.9.1 (nightly): Has pickle/security errors when loading models
-
PyTorch 2.7.0 (nightly): Still doesn’t have sm_120 kernel support
My question: Is there an official timeline for when stable PyTorch will support RTX 50-series (Blackwell/sm_120)? Will it be in PyTorch 2.10 or 3.0?
System specs:
-
GPU: NVIDIA GeForce RTX 5050 (8GB VRAM)
-
RAM: 32GB
-
CUDA: 12.4
-
OS: Windows 11
I’m currently forced to use CPU mode which works but is significantly slower. Any guidance or timeline would be greatly appreciated!
Thanks in advance!