Gpu sharding #19
Open
Gpu sharding #19
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
IndexIDMap was missing from the go bindings, even though is is important to wrap a flat index in if addWithID functionality is required.
IDMapWrapper isn't practical bc I would like to be able to transfer gpuIndex (or any other index for that matter) to gpu using the transferToGpu. So I dont want a specific structure which wraps an idmap index, rather just return an Index interface.
Better to have a test to make sure everything functions properly than an example usage..
In addition added a test for the bindings which show how to use the functionality.
This is potentially necessary for various things, but specifically is critical for us NOW because RemoveIDs isn't implemented for GPUIndexs, so until we implement that functionality in faiss, the only solution is to transfer the index back and forth to the CPU and remove ids which the index is on CPU.
This is so that it will be possible to compile and use the library on a machine which doesn't have a gpu as well.
Didn't know t.Log existed, better to use it.
I discoverd the hard way that after transferring an index to the GPU, the gpu memory isn't freed when the index is deleted. It seems like a gpuResource is being allocated and holding the memory. Its necessary to free the gpuResource in order to free the Gpumemory. In order to free the gpuResource we will need to save it, and call free on it. In the next commit I will add the Free functionality to gpu_bindings.go.
…letes the index TransferToCpu will free the gpu memory after copying the index to CPU. Also added a function to CreateGpuIndex just because it was easy and may be useful for someone.
Its not really the best, but I assume that most Gpus will run out of memory when creating 20 gpu indices, so if the test passes, most likely the free function is working ;)
instead of hold just one gpu resource , hold all the gpus resrouces. this is neccesrry when we want to do sharding on multiple gpus with the same index
* Removed redundant get_sizet function * Changed inside TransferToGpu from gpuResources to gpuResource + made adjustments to work with resources and not resource in the return value * Added TODO with comment out function TransferToAllGPUsWithOptions for future development * Free function was adjusted to support array of gpuReource
this is just to let cpu mode code to compile
Support multiple gpus
currently there is bug on faiss facebookresearch/faiss#2357 to use sharding you have to use this version of faiss for now git clone -b fix-cpi-gpu-faiss_index_cpu_to_gpu_multiple_with_options https://github.com/AviadHAv/faiss.git
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.