Andrei Bârsan

Any progress on this? I'm noticing ``` with fsspec.open("s3://some-bucket/debug.txt", "wt") as f: f.write("test") ``` Failing with an error in the parent post, but the following workaround ``` with fsspec.filesystem("s3").open("s3://some-bucket/debug.txt", "wt")...

As before, I needed to RTFM. Bumping to the latest `s3fs` (and related packages) fixed this issue! Specifically, here are my related package versions: ``` s3fs==2021.10.0 fsspec=2021.10.0 botocore==1.20.106 aiobotocore==1.4.2 ```

Ah, it may be due to my CPU tensors not being in pinned memory: https://forums.developer.nvidia.com/t/does-cudamemcpyasync-require-pinned-memory/40411

Joining this thread as I have a very relevant question. Thank you for the awesome crate @Twinklebear!! I have hit the exact same problem as the original issue author: I...

Apparently callbacks still aren't working in the current version... Any plans on fixing them?

Alternatively, a dirty solution is to disable cuDNN. In my case, inference still runs reasonably fast wihthout cuDNN. @PangolinLE314 I sadly could not build MNC's fork of Caffe with cuDNN...

@PangolinLE314 awesome! That worked! I could build with cuDNN 5, all the Caffe tests passed, and inference is now considerably faster and uses less memory. Thank you very much!

Thank you for the response! But in this case we would have to generate our own version IDs, and can't rely on S3's generated ones, right? That is, our full...

Could `_strip_protocol` be used in the S3 fs implementation to address this, maybe? Like how in local filesystems it does things like `os.expanduser`? https://github.com/intake/filesystem_spec/blob/e734622e2b837625d5c8f27477d6968d837f68b8/fsspec/implementations/local.py#L148

Yeah, I think `fsspec` crashes before we can actually make the request. There seems to be some logic for handling the magic part of the URI (it figures out something...