Interdisciplinary · Reference

ML Desktop: Blackwell GPU Fix (RTX 5060 Ti)

First real workload run on the new machine — a genuine PyTorch/GPU compatibility gap, diagnosed and fixed

nvidia-smi and torch.cuda.is_available() both said the GPU was fine. It wasn't — the pinned PyTorch version predated Blackwell architecture support entirely. Detected is not the same as working.

The symptom

UserWarning: NVIDIA GeForce RTX 5060 Ti with CUDA capability sm_120 is not compatible
with the current PyTorch installation. The current PyTorch install supports CUDA
capabilities sm_50 sm_60 sm_70 sm_75 sm_80 sm_86 sm_90.

Root cause

Confirmed via research: torch==2.1.2 (the version pinned by the toolkit being tested) predates NVIDIA's Blackwell architecture entirely. PyTorch only added official Blackwell (sm_120) support starting with version 2.7.0 (April 2025), via cu128-and-later builds.

torch.cuda.is_available() returning True was genuinely misleading — that's driver-level detection only, not proof the actual compute kernels work for this specific architecture. A GPU can be correctly detected by the driver and still be unusable by a too-old PyTorch build.

The fix

Upgraded torch alone, left the rest of whatever toolkit's pinned versions in place:

pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

Result: clean upgrade to torch 2.11.0+cu128, torchvision 0.26.0, torchaudio 2.11.0 — no errors.

Verifying it actually works, not just installs

Confirmed with a real GPU computation, not just version detection:

python3 -c "
import torch
x = torch.randn(1000, 1000, device='cuda')
y = torch.randn(1000, 1000, device='cuda')
z = torch.matmul(x, y)
torch.cuda.synchronize()
print('GPU compute OK:', z.shape, z.device)
"
# → GPU compute OK: torch.Size([1000, 1000]) cuda:0

Confirmed: the GPU is genuinely usable now, not just detected.

General lesson for this machine going forward: any future CUDA-based project that pins an older torch version is likely to hit this exact same wall. Worth checking a new toolkit's pinned torch version against Blackwell support (2.7.0+) before assuming nvidia-smi working means the project will too.