gpu

there may be a log error in pytorch/aten/src/THC/THCGenerateByteTypes.h, in line2, #error "You must define THC_GENERIC_FILE before including THGenerateByteTypes.h", the "THGenerateByteTypes.h" should be "THCGenerateByteTypes.h".
bty, this file is a general file to generate files of different scalar_t, so i think, the name is better to be THGenerateTypes.h.

As shown in taichi-dev/taichi#3910, replacing property with simple attributes can speedup python part of taichi a lot.
Lessons learned is that we should avoid using @property when applicable since it's expensive. So let's review the usage of @property in our python codebase and replace them as much as possible.

Here's a list of simple grep in our codebase showing

At this moment relu_layer op doesn't allow threshold configuration, and legacy RELU op allows that.
We should add configuration option to relu_layer.

I am working on creating a WandbCallback for Weights and Biases. I am glad that CatBoost has a callback system in place but it would be great if we can extend the interface.

The current callback only supports after_iteration that takes info. Taking inspiration from XGBoost callback system it would be great if we can have before iteration that takes info, before_training, and `after

Hi ,

I have tried out both loss.backward() and model_engine.backward(loss) for my code. There are several subtle differences that I have observed , for one retain_graph = True does not work for model_engine.backward(loss) . This is creating a problem since buffers are not being retained every time I run the code for some reason.

Please look into this if you could.

I want to preemptively start this thread to survey for suggestions. A cursory search lead me to this promising repository https://github.com/enigo-rs/enigo

Since closing the window is a common point of failure, that will be the focus for the first pass of testing as I learn how to use the library.

Components for testing:

• bridge
• editor
• renderer
• settings
• wind

Somehow some of these names start with fmt_ocl_ while most start with fmt_opencl_. Is this intentional? It causes the fmt_ocl_ to be listed/tested/benchmarked first. Then there's fmt_opencl_1otus5 with a 1 (one) in there. So the first OpenCL formats become:

john_register_one(&fmt_ocl_cryptosafe);
john_register_one(&fmt_ocl_cryptsha1);
john_register_one(&fmt_ocl_KeePass);
john

Description

Change the signature of cupy.{percentile,quantile} to provide exactly the same API as NumPy.

I think it's ok to implement overwrite_input as nop (just ignore the option).

Additional Information

No response

https://github.com/brendangregg/FlameGraph

Add alpha value in the dataframe that is returned when predicting

See this comment link

Based on @karthikeyann's work on this PR rapidsai/cudf#9767 I'm wondering if it makes sense to consider removing the defaults for the stream parameters in various detail functions. It is pretty surprising how often these are getting missed.

The most common case seems to be in factory functions and various ::create functions. Maybe just do it for those?

环境

1.系统环境：

• Ubuntu 20.04.2 LTS x86_64
• GLIBCXX: 3.4.28 (libstdc++.so.6.0.28)
• GPU:
  

2.MegEngine版本：1.6.0rc1
3.python版本：Python 3.8.10

The program stuck at net.load when I was trying to use the MegFlow. I wait for more than 10min and there is no sign of finishing it.

Current implementation of join can be improved by performing the operation in a single call to the backend kernel instead of multiple calls.

This is a fairly easy kernel and may be a good issue for someone getting to know CUDA/ArrayFire internals. Ping me if you want additional info.