dfdx

dfdx copied to clipboard

Please add a split method, to iterate over subtensors of the given tensor

7

Please add a method for splitting on a given axis, returning an iterator over subtensors selected in that axis. For example, suppose `mytensor` is this `2x3` tensor: ```rust [ [1,...

emchristiansen

For some generic types that implement `Dim`, their value need to get subtracted from. Note: I haven't any form of stress test on this, so this PR should be considered...

swfsql

Remove functions calling the deprecated `_MM_SET_FLUSH_ZERO_MODE`

3

From https://github.com/rust-lang/stdarch/pull/1454, it seems that setting `_MM_SET_FLUSH_ZERO_MODE` results in "immediate Undefined Behavior", and as a result these functions were deprecated with 1.75.0. This PR removes the functions that include calls...

rainiwu

CUDA kernels missing __hmin and __hmax

1

This is weird since the exact opposite thing happened in #762, but it appears that now `__hmin` and `__hmax` are not being found. I tested driver versions 525 and 535...

opfromthestart

Adds `unstack`

1

1. Adds an empty `.rustfmt.toml` so formatting for this project will be respected in workspaces with different formatting. 1. Adds `unstack`, which is the inverse of `stack`. Note: I've never...

emchristiansen

`TapeGlobal` and thoughts on the tape variants

3

# TapeGlobal Here's another tape-tracking API to consider, as implemented in client code: ```rust type FloatInner = f32; type DfdxDevice = Cpu; thread_local! { static TAPE_GLOBAL: once_cell::sync::Lazy< Mutex, > =...

emchristiansen

Hi, I'm unable to build this crate without changes for old GPU (Tesla M40, capability level 5.2, driver 462). Most of the problems are related to FP16 and the fact...

nikvoid

feat(webgpu): add tensor operations

1

# What This PR Does Starts a general binary op implementation for `Webgpu`. This code is based on an (old, in-progress WGPU branch)[https://github.com/DonIsaac/dfdx/tree/don/feat/wgpu2] I made a while ago. I've (mostly)...

DonIsaac

I have two models, identical in structure, in which one is meant to be updated periodically from the other. Prior to #854 I did so using `TensorCollection::iter_tensors` as follows:: ```rust...

hovinen

Re-opening per #215: I was working on implementing AlexNet in dfdx per https://github.com/LaurentMazare/tch-rs/blob/main/src/vision/alexnet.rs, but have gotten a bit blocked by not having this layer type available. Would it be possible...

quietlychris