v0.8.20

NNlib v0.8.20

Diff since v0.8.19

Closed issues:

• Bugged fallback in grouped Convs (#369)
• NNlibAMDGPUExt: UndefVarError: MIOpen not defined (#482)

Merged pull requests:

• Add hand written gelu derivative (#480) (@chengchingwen)
• [NNlibAMDGPUExt] Load MIOpen module only if it is available (#483) (@pxl-th)
• Use KernelAbstractions.jl for upsample kernels (#486) (@pxl-th)
• Use KernelAbstractions.jl for gather/scatter kernels (#487) (@pxl-th)
• [AMDGPU] Add dispatch path for FP16 batched mul (#488) (@pxl-th)

v0.8.19

NNlib v0.8.19

Diff since v0.8.18

Merged pull requests:

• Add AMDGPU extension (#470) (@pxl-th)
• Merge NNlibCUDA as subpackage (#471) (@ToucheSir)
• Add dropout & attention tests for AMDGPU (#472) (@pxl-th)
• Allow regular convolution for AMDGPU (#473) (@pxl-th)
• Add downstream tests for Lux.jl (#474) (@avik-pal)
• change AMDGPUExt to NNlibAMDGPUExt (#477) (@CarloLucibello)

v0.8.18

NNlib v0.8.18

Diff since v0.8.17

Closed issues:

• Unstable performance (#461)

Merged pull requests:

• Fix conv with groups when falling in direct backend (#468) (@gabrielpreviato)

v0.8.17

NNlib v0.8.17

Diff since v0.8.16

Closed issues:

• Multi-head attention? (#385)
• Batched multiplication support for ndims > 3 (#391)
• Symmetric Padding (#463)

Merged pull requests:

• implement dot_product_attention (#455) (@CarloLucibello)
• symmetric and circular padding (#465) (@nikopj)
• Reduce lpnormpool unnecessary conversion (#467) (@skyleaworlder)
• Update deprecated actions & change coverage upload entry (#469) (@skyleaworlder)

v0.8.16

NNlib v0.8.16

Diff since v0.8.15

Closed issues:

• batched_vec >1000X slower than batched_mul (#462)

Merged pull requests:

• Add lppool implementation (#447) (@skyleaworlder)
• Refactor batched_vec (#464) (@jondeuce)

v0.8.15

NNlib v0.8.15

Diff since v0.8.14

Merged pull requests:

• add _rng_compat_array (#458) (@mcabbott)

v0.8.14

NNlib v0.8.14

Diff since v0.8.13

Closed issues:

• support arbitrary number of batch dimensions in batched_mul (#451)

Merged pull requests:

• Add within_gradient (#434) (@mcabbott)
• Fixup docs for fold and unfold (#453) (@ToucheSir)
• Add dropout (#454) (@mcabbott)
• fast_maximum also for logsumexp (#456) (@Sleort)

v0.8.13

NNlib v0.8.13

Diff since v0.8.12

Merged pull requests:

• Avoid maximum in softmax (#450) (@mcabbott)

v0.8.12

NNlib v0.8.12

Diff since v0.8.11

Merged pull requests:

• add gather(src, IJK...) (#449) (@CarloLucibello)

v0.8.11

NNlib v0.8.11

Diff since v0.8.10

Closed issues:

• (Flaky?) CI failures on GHA latest + Buildkite (#359)

Merged pull requests:

• Trigger tagbot on issue comments (#440) (@Saransh-cpp)
• Remove threading from all ∇*conv_filter and re-enable old tests (#441) (@ToucheSir)
• Slightly faster softplus (#443) (@Sleort)
• Add fold and unfold (#444) (@nikopj)