Open kuraga opened 3 years ago
Well applying https://github.com/sberbank-ai/ru-gpts/issues/42#issuecomment-839830391 let me load the model.
But on model apply I get https://github.com/sberbank-ai/ru-gpts/issues/42#issue-813829451 itself:
<...>
/usr/local/lib/python3.8/dist-packages/deepspeed/ops/sparse_attention/matmul.py in _sdd_matmul(a, b, trans_a, trans_b, trans_c, spdims, block, luts, num_locks, widths, packs, bench, time)
202 total = 0 if bench else None
203 for off_width in range(0, width, max_width):
--> 204 current = kernel(
205 a.data_ptr(),
206 b.data_ptr(),
/usr/local/lib/python3.8/dist-packages/triton/kernel.py in __call__(self, grid, *args)
114 # pack parameters into a byte buffer
115 params = struct.pack(self.tys, *args)
--> 116 kernel = self.fn.autotune(params, grid, self.stream)
117 # run kernel
118 grid = grid(kernel.opt)
RuntimeError: CUDA: Error- invalid ptx
Hi!
My only guess that during uploading of the checkpoint smth went wrong. I uploaded again checkpoint in another source. Please, try here: https://sc.link/Dk6Y
I'm facing https://github.com/sberbank-ai/ru-gpts/issues/68. After applying https://github.com/sberbank-ai/ru-gpts/issues/68#issuecomment-894067337 error dissapears but I'm facing out-of-memory error.
Thanks!
Hello, @Alenush !
According to error, some weights are missing in
.pth
. (Missing keys in state_dict: ...
) Could model change?Thanks!