Looking for PyTorch Binaries for Older GPUs

[VladimirK (Vladimir Kudryashov)]

I have GT 710. It has compute capability of 3.5. Unfortunately, since version 1.3.1 PyTorch does not support cc < 3.7. This means PyTorch > 1.3.0 does not support my GPU. When I try to use PyTorch 1.6.0 I get the error:

RuntimeError: CUDA error: no kernel image is available for execution on the device

In this case, @peterjc123 recommends to build PyTorch from source. I tried to build PyTorch from source, but it turned out to be not so easy. I gave up after a few days. Then I began to look for someone who succeeded and I found it. Nelson Liu built PyTorch Binaries that support old GPUs. Newer PyTorch Binaries for Older GPUs These binaries for Linux but I need Windows binaries.

I didn’t find any binaries for Windows, but I’m sure they exist.

If anyone has built or found PyTorch Binaries for Windows, please share here.

[mszhanyi (Yi Zhang)]

Hope it helps.
Previous PyTorch Versions | PyTorch

[VladimirK (Vladimir Kudryashov)]

Thank you @mszhanyi, but the official distributions of pytorch do not support 3.5 compute capabilities.

[KFrank]

Hi Vladimir!

@VladimirK

the official distributions of pytorch do not support 3.5 compute capabilities.

I had a similar issue a couple of years ago, and was provided a link
to some legacy binaries that I can no longer access.

For what it’s worth, here’s the (dead?) link:

Legacy PyTorch binaries on Google Drive

@ptrblck directed me to this link and I believe that the binaries were
built by @peterjc123.

I was able to find a windows binary on that link that supported my old
(laptop) gpu. But I had to go back to the archaic pytorch version 0.3.0
to do so.

For additional context, here is the original thread:

Pre-built pytorch for CUDA [compute capability] 3.0 on windows?

Hi Forum! Would anybody know of a pre-built pytorch windows / CUDA 3.0 version? (It’s windows 10, if that matters.) I’m aware that pytorch no longer formally supports older CUDA versions, but I have seen older pre-built packages floating around on the internet – just not this configuration. (I will be installing pytorch on an older laptop with a CUDA 3.0 quadro k1100m graphics card. I’ll just be monkeying around, so I don’t need the gpu, but I’d like to monkey around with the gpu, as …

Best.

K. Frank

[VladimirK (Vladimir Kudryashov)]

Hi Frank. Thanks for the information. I have already built pytorch for my GPU with cc 3.5. torch-1.6.0a0+b31f58d-cp37-cp37m-win_amd64.whl on Google Drive Unfortunately, this build does not contain CUDA. In order to use this build, you need to install CUDA 9.2. If I manage to build the full version, I will post the link here.

[VladimirK (Vladimir Kudryashov)]

I added missing files. torch-1.6.0a0+b31f58d-cp37-cp37m-win_amd64.whl
I have been tested this file on my computer in another Windows (Windows 7).

[Kry_Be]

Heyu can You build torch 1.7 and cuda 10.2 ? I have this issue as well, and not a single person shared compiled torch with support for compute 3.5.They should compile version with support for old cards at least once every 6 months or so, dropping support immediately is very bad idea and confusing for people who had no clue they did that.If i had known sooner, i wouldnt bought my GPU, i got gtyx 780cause it has as much cuda cores as gtx 1080 , more than supported gtx `1050 with compute 5…
I hope someone will take care of it properly, dropping backwards compatibility like that is weird stuff, If you guys lack space i can create googledrive account and keep torch on it.
Even better would be to compile torch 1.8 and cuda 11 with old cards support but who knows how ? I spent 2 days on it, full days and still had issues, because every software is changing their backwards compatibility silently and oyu have error after error , you update one thing and then the other ingredient is surprise - not working anymore in that new version… googling for days to find someone who did this before succesfully.