Repository navigation

All pull requests

Quacomm AI Engine Direct - Bump QNN SDK version from 2.50 to 2.51
#23918winskuo-quic opened Oct 16, 2026Contributor
2
[cuda] Deflake test_aoti_compile_traces_the_measured_fastest_value
#23917Gasoonjia opened Oct 16, 2026Contributor
2
Qualcomm AI Engine Direct - Keep each node's own LPBQ block size
#23916msluszniak opened Oct 16, 2026Contributor
2
Qualcomm AI Engine Direct - Keep a non-delegated op's weight dequantize on CPU
#23915msluszniak opened Oct 16, 2026Contributor
2
Resize every operand before its consumer in one pass
#23914psiddh opened Oct 16, 2026Contributor
4
Check quantized embedding qparam dtypes against how the kernels read them
#23912Shakhtar-Sankur opened Oct 16, 2026
2
Skip ggml targets on platforms their backends do not support
#23911digantdesai opened Oct 16, 2026Contributor
4
Packed HiFi FC: fused 4-bit kernel on HiFi4
#23910mcremon-meta opened Oct 16, 2026Contributor
3
Update MG README with offgraph cache and batching
#23909metascroy opened Oct 16, 2026Contributor
3
[cadence] Enable all hifi op tests in the CMake build
#23907aliafzal opened Oct 15, 2026Contributor
2
Generalize LowerHFAttention to learned-position cached attention
#23905JakeStevens opened Oct 15, 2026Contributor
3
Add Whisper runner under examples/audio
#23904JakeStevens opened Oct 15, 2026Contributor
3
Support Whisper positional indexing
#23903JakeStevens opened Oct 15, 2026Contributor
3
Lower cache-free SDPA to et_vk.sdpa
#23902JakeStevens opened Oct 15, 2026Contributor
3
Add Whisper HF export (encoder, cross_kv, static-cache decoder) (#23901)
#23901JakeStevens opened Oct 15, 2026Contributor
3
[ExecuTorch][WebGPU] Use 2D-folded dispatch for cat and slice
#23900digantdesai opened Oct 15, 2026Contributor
4
Add CMake targets for HF LLM runners (#23899)
#23899JakeStevens opened Oct 15, 2026Contributor
3
Don't log an error when there are no CUDA backend symbols to promote
#23895edmundshieh opened Oct 15, 2026
4
Copy returned views out of AOTI methods as contiguous tensors
#23894edmundshieh opened Oct 15, 2026
5
[ET-VK][conv2d_dw] Fix depthwise convs that fall back to the generic shader
#23892borys3kk opened Oct 15, 2026
5
Namespace lifted tensor constants per method in to_native (#23891)
#23891JakeStevens opened Oct 15, 2026Contributor
3
Fix adding observers on non-float nodes and Fix incorrect turn of an output buffer into model parameter
#23890robert-kalmar opened Oct 15, 2026Collaborator
9
Add Arm Buck2 coverage and declaration checks to GitHub CI
#23888usamahz opened Oct 15, 2026Collaborator
3
NXP backend: refactor dataset_creators to enable better benchmarking
#23886novak-vaclav opened Oct 15, 2026Collaborator
9
[ExecuTorch][WebGPU] Add a shared-memory direct kernel for groups == 1 conv2d
#23882digantdesai opened Oct 15, 2026Contributor
11

Navigation Menu

Repository navigation

All pull requests

Quacomm AI Engine Direct - Bump QNN SDK version from 2.50 to 2.51

[cuda] Deflake test_aoti_compile_traces_the_measured_fastest_value

Qualcomm AI Engine Direct - Keep each node's own LPBQ block size

Qualcomm AI Engine Direct - Keep a non-delegated op's weight dequantize on CPU

Resize every operand before its consumer in one pass

Check quantized embedding qparam dtypes against how the kernels read them

Skip ggml targets on platforms their backends do not support

Packed HiFi FC: fused 4-bit kernel on HiFi4

Update MG README with offgraph cache and batching

[cadence] Enable all hifi op tests in the CMake build

Generalize LowerHFAttention to learned-position cached attention

Add Whisper runner under examples/audio

Support Whisper positional indexing

Lower cache-free SDPA to et_vk.sdpa

Add Whisper HF export (encoder, cross_kv, static-cache decoder) (#23901)

[ExecuTorch][WebGPU] Use 2D-folded dispatch for cat and slice

Add CMake targets for HF LLM runners (#23899)

Don't log an error when there are no CUDA backend symbols to promote

Copy returned views out of AOTI methods as contiguous tensors

[ET-VK][conv2d_dw] Fix depthwise convs that fall back to the generic shader

Namespace lifted tensor constants per method in `to_native` (#23891)

Fix adding observers on non-float nodes and Fix incorrect turn of an output buffer into model parameter

Add Arm Buck2 coverage and declaration checks to GitHub CI

NXP backend: refactor dataset_creators to enable better benchmarking

[ExecuTorch][WebGPU] Add a shared-memory direct kernel for groups == 1 conv2d

Pull requests