Luca Wehrstedt comments

Results 191 comments of


                                            Luca Wehrstedt

Acking the receiver (to read the message) of a completed write operation

I'm not sure I follow. Is there a specific issue you're facing?

Acking the receiver (to read the message) of a completed write operation

> when reading from a pipe, if there is no message in the pipe (e.g., nobody writes to the pipe yet), will the reader be blocked until there is a...

cuda_ipc: Couldn't find GPU with UUID xxx

Interesting. I must admit I had seen this before but then it went away and I didn't think much about it since. TensorPipe should be able to circumvent limits to...

cuda_ipc: Couldn't find GPU with UUID xxx

I looked more into this and indeed it seems there's an NVIDIA Docker integration of sorts that limits the container to only see some of the GPUs by, probably, limiting...

cuda_ipc: Couldn't find GPU with UUID xxx

BTW I was wondering what you were trying to achieve when you hit this problem. Were you separating processes on the same machine using containers on purpose? What's the reason...

[Question]How to detect pipe(obtained from ctx->connect()) is writable?

A pipe is always immediately writable, and the pipe might "hold onto" your write requests for a bit until the connections are ready, but this is invisible to you.

[Question]How to detect pipe(obtained from ctx->connect()) is writable?

A "connection refused" means that the server isn't listening on that address/port, you should check how you're obtaining that address, it should come from the server's Listener instance.

Error: "transport retry counter exceeded" when torch.distributed.rpc.init_rpc between different pods in k8s

Love to hear that you're finding TensorPipe useful (and would be interested in knowing more what you're doing with it). And you're right that the IB device selection of the...

Error: "transport retry counter exceeded" when torch.distributed.rpc.init_rpc between different pods in k8s

> why there is no channel implemented by ibv? * There is one channel, called `cuda_gdr`, which also uses IB, but is independent from the `ibv` transport. Currently `cuda_gdr` only...

Redefinition of 'struct prctl_mm_map'

Indeed I agree with your proposed fix, I don't see any reason to include `linux/prctl.h` over `sys/prctl.h`. If you can submit a PR I'll happily merge it.