inclusionAI/cuLA

[Feature] Integrate Lightning SM90 Implementation via CuTe DSL about 17 hours ago

good first issue help wanted sm90

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

Small B/H/S optimizations about 1 month ago

enhancement help wanted performance

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

[Feature] Add GVA support for Lightning about 2 months ago

enhancement good first issue help wanted

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

[KDA] Grouped Value Attention (GVA) support 2 months ago

enhancement good first issue help wanted

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

WIP-RFC: Multi-GPU Unified Interface for cuLA 3 months ago

enhancement help wanted integration

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

Modular GDN forward / backward kernels (compatible with Kimi CP) 3 months ago

enhancement help wanted

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

More aggressive fusion of small neighboring kernels (cumsum etc.) for inference 3 months ago

enhancement help wanted inference

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

Identify and implement CUDA optimization opportunities for Intracard CP (single-card sequence splitting) 3 months ago

enhancement help wanted distributed

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

KDA Backward pass optimizations 3 months ago

enhancement help wanted sm10x sm90 train

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

Support for bf16 Beta 3 months ago

enhancement good first issue

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

We need a KDA decode kernel 4 months ago

enhancement help wanted inference

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

Lightning-attn supports pretransposed states 4 months ago

good first issue help wanted performance

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

Prefill Output Pretransposed States `[V, K]` for Decode 4 months ago

help wanted performance inference

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

Integrate cuLA into flash-linear-attention via FLA's kernel dispatch mechanism 4 months ago

enhancement good first issue integration

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

KDA MTP (Multi-Token Prediction) support 4 months ago

enhancement help wanted inference

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

Polynomial approximation to mitigate the exponential bottleneck (Flash-Attention-4) 4 months ago

help wanted research performance

inclusionAI/cuLA

534

CUDA kernels for linear attention variants, written in CuTe DSL and CUTLASS C++.

Python

Open Issues Need Help