Skip to content

[CUDA] Add decode-optimized LinearAttention (GatedDeltaNet) kernels #28985

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

Merged
kunal-vaishnavi merged 6 commits into from
Jun 18, 2026
Merged

[CUDA] Add decode-optimized LinearAttention (GatedDeltaNet) kernels #28985

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
1ca8d3d
Improve LinarAttention decode
tianleiwu Jun 10, 2026
ebcd859
address feedbacks
tianleiwu Jun 12, 2026
74945fd
refactoring
tianleiwu Jun 12, 2026
d01b895
lintrunner -a
tianleiwu Jun 12, 2026
6ee345a
Address review feedback in CUDA helper and decode comments
Copilot Jun 12, 2026
cdcea5e
address more feedbacks
tianleiwu Jun 18, 2026
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
29 changes: 1 addition & 28 deletions onnxruntime/contrib_ops/cuda/bert/causal_conv_with_state_impl.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,41 +18,14 @@
#include <cuda_runtime.h>
#include <math.h>
#include "contrib_ops/cuda/bert/causal_conv_with_state_impl.h"
#include "core/providers/cuda/cu_inc/cuda_type_helper.cuh"

namespace onnxruntime {
namespace contrib {
namespace cuda {

namespace {

template <typename T>
__device__ __forceinline__ float to_float(T val);

template <>
__device__ __forceinline__ float to_float(float val) { return val; }

template <>
__device__ __forceinline__ float to_float(half val) { return __half2float(val); }

#if __CUDA_ARCH__ >= 800 || !defined(__CUDA_ARCH__)
template <>
__device__ __forceinline__ float to_float(__nv_bfloat16 val) { return __bfloat162float(val); }
#endif

template <typename T>
__device__ __forceinline__ T from_float(float val);

template <>
__device__ __forceinline__ float from_float(float val) { return val; }

template <>
__device__ __forceinline__ half from_float(float val) { return __float2half(val); }

#if __CUDA_ARCH__ >= 800 || !defined(__CUDA_ARCH__)
template <>
__device__ __forceinline__ __nv_bfloat16 from_float(float val) { return __float2bfloat16(val); }
#endif

__device__ __forceinline__ float silu_fn(float x) {
return x / (1.0f + expf(-x));
}
Expand Down
Loading
Loading