glake

glake copied to clipboard

Compatibility with future python versions

2

Hi, this is a really cool peice of work! I was wondering if this was compatible with pytorch 2.3.0? And if not, if there was any way to make it...

thecheekygeek

question about torch 2.1.0 integration

3

Thanks for your sharing! I'm greatly appreciate your work for reducing the cuda memory fragmentation. Recently I have integrated GMLake into torch2.1.0 and finished compiling without error. I would like...

Pegessi

use glake-torch-1.13.1 whl package error:OSError: libmkl_gnu_thread.so.1: cannot open shared object file: No such file or directory

1

before pip install torch-1.13.0a0+git49444c3-cp38-cp38-linux_x86_64.whl root@worker-0:/workspace#python Python 3.8.13 packaged by conda-forge (default,Mar 25 2022,06:04:10) [GCC 10.3.0]on linux Type "help","copyright","credits"or "license"for more information. >>import torch >>print(torch.__version__) 1.13.1+cu117 after pip install torch-1.13.0a0+git49444c3-cp38-cp38-linux_x86_64.whl root@worker-0:/workspace#...

112sdfsdfsdfasf

Wheel file

1

Could you please provide the wheel file corresponding to PyTorch 1.13.1?

indigoLovee

Could you please provide the libcuda.so or libc10_cuda.so corresponding to pytorch 1.13.1?

1

indigoLovee

Building Pytorch release 2.1 + Glake failed

5

Very cool work, we really hope to use Glake in our LLM training. However, I failed when trying to compile glake on pytorch release 2.1. My system information and error...

SolenoidWGT

How to use this work in the TensorFlow system？

2

fuhailin

Comparison with expandable_segments in pytorch/c10?

3

https://github.com/pytorch/pytorch/pull/96995 https://github.com/pytorch/pytorch/blob/95a86ed9ca107329151e0dc172386d50dd3471c6/c10/cuda/CUDACachingAllocator.cpp#L311-L324 > The expandable_segments:True option is used to enable/disable this behavior. We use cuda's low-level memory APIs, which are similar to mmap, to extend the memory segments. These APIs...

YouJiacheng

# Training - [ ] Multi-stream Memory Reuse: Done, will be released - [ ] Compatible with Expandable Segment - [ ] Memory Pattern Profiling tool - [ ] DoubleOverlapping(for...

ruizhang1230

vAttention said that: if use 2M pageSize, 128M physical memory can be wasted per-request in the worst-case in Llama-3-8B (TP-1), but if use 64KB, 128M would be only 4M Do...

nalinaly