RuntimeError: CUDA error: CUBLAS_STATUS_EXECUTION_FAILED when calling `cublasGemmEx

I have a GTX 1660 SUPER 6 gb vram, Ubuntu 20.04, Python 3.9, Driver Version: 460.91.03 , CUDA Version: 11.2

At this stage:
3%|███▏ | 27/1024
I get an error:
```
.../ru-dalle/main.py", line 31, in
_pil_images, _scores = generate_images(text, tokenizer, dalle, vae, top_k=top_k, images_num=images_num, top_p=top_p)
..........
.../ru-dalle/venv/lib/python3.9/site-packages/torch/nn/functional.py", line 1848, in linear
return torch._C._nn.linear(input, weight, bias)
RuntimeError: CUDA error: CUBLAS_STATUS_EXECUTION_FAILED when calling `cublasGemmEx( handle, opa, opb, m, n, k, &falpha, a, CUDA_R_16F, lda, b, CUDA_R_16F, ldb, &fbeta, c, CUDA_R_16F, ldc, CUDA_R_32F, CUBLAS_GEMM_DFALT_TENSOR_OP)`
```

The video memory used at the time of the error:
4893MiB / 5936MiB

Also, at the very beginning of generation, I get a warning:
`UserWarning: __floordiv__ is deprecated, and its behavior will change in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values. To keep the current behavior, use torch.div(a, b, rounding_mode='trunc'), or for actual floor division, use torch.div(a, b, rounding_mode='floor').`