Skip to content

Fix HIP memory leaks in RNN kernels #727

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
magaonka-amd wants to merge 1 commit into
base: rocm-jaxlib-v0.8.0
Choose a base branch
from
Draft

Fix HIP memory leaks in RNN kernels #727

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 9 additions & 0 deletions jaxlib/gpu/rnn_kernels.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -172,6 +172,9 @@ DoRnnComputeWorkspaceReserveSpaceSizes(int input_size, int hidden_size,
JAX_RETURN_IF_ERROR(
JAX_AS_STATUS(gpudnnDestroyRNNDataDescriptor(input_data_desc)));
JAX_RETURN_IF_ERROR(JAX_AS_STATUS(gpudnnDestroyRNNDescriptor(rnn_desc)));
#ifdef JAX_GPU_HIP
JAX_RETURN_IF_ERROR(JAX_AS_STATUS(gpuFree(dropout_states_dev)));
#endif

// Round up to nearest multiples of 4 so we can return them as f32 arrays.
workSpaceSize += (workSpaceSize % 4);
Expand Down Expand Up @@ -351,6 +354,9 @@ static absl::Status DnnRNNForward_(gpuStream_t stream, void** buffers,
JAX_RETURN_IF_ERROR(
JAX_AS_STATUS(gpudnnDestroyDropoutDescriptor(dropout_desc)));
JAX_RETURN_IF_ERROR(JAX_AS_STATUS(gpudnnDestroyRNNDescriptor(rnn_desc)));
#ifdef JAX_GPU_HIP
JAX_RETURN_IF_ERROR(JAX_AS_STATUS(gpuFree(dropout_states_dev)));
#endif

return absl::OkStatus();
}
Expand Down Expand Up @@ -536,6 +542,9 @@ static absl::Status DnnRNNBackward_(gpuStream_t stream, void** buffers,
JAX_RETURN_IF_ERROR(
JAX_AS_STATUS(gpudnnDestroyDropoutDescriptor(dropout_desc)));
JAX_RETURN_IF_ERROR(JAX_AS_STATUS(gpudnnDestroyRNNDescriptor(rnn_desc)));
#ifdef JAX_GPU_HIP
JAX_RETURN_IF_ERROR(JAX_AS_STATUS(gpuFree(dropout_states_dev)));
#endif

return absl::OkStatus();
}
Expand Down
1 change: 1 addition & 0 deletions jaxlib/gpu/vendor.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -770,6 +770,7 @@ inline hipsparseStatus_t gpusparseCreate(gpusparseHandle_t* handle) {
#define GPU_STREAM_NON_BLOCKING hipStreamNonBlocking

#define gpuMalloc hipMalloc
#define gpuFree hipFree
Copy link

@Arech8 Arech8 Mar 5, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I can't believe it's not there yet...

Note, a corresponding entry is needed for CUDA too, and this needs to be upstreamed too

Copy link

@Arech8 Arech8 Mar 5, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ok, saw the upstream link in the descipriton, awesome. Then just an update for CUDA is needed.

Copy link
Author

@magaonka-amd magaonka-amd Mar 5, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

yep very funny to me aswell. looks like we have unlimited memory no need to free it. just keep using it :).
okay I'll update upstream aswell. thanks.

#define gpuGetLastError hipExtGetLastError
#define gpuGetErrorString hipGetErrorString
#define gpuMemcpyAsync hipMemcpyAsync
Expand Down