Bruce-Lee-LY / decoding_attentionLinks
Decoding Attention is specially optimized for MHA, MQA, GQA and MLA using CUDA core for the decoding stage of LLM inference.
☆40Updated last month
Alternatives and similar repositories for decoding_attention
Users that are interested in decoding_attention are comparing it to the libraries listed below
Sorting:
- ☆60Updated 3 months ago
- Multiple GEMM operators are constructed with cutlass to support LLM inference.☆18Updated this week
- ⚡️Write HGEMM from scratch using Tensor Cores with WMMA, MMA and CuTe API, Achieve Peak⚡️ Performance.☆93Updated 2 months ago
- A practical way of learning Swizzle☆22Updated 6 months ago
- Performance of the C++ interface of flash attention and flash attention v2 in large language model (LLM) inference scenarios.☆39Updated 5 months ago
- ☆92Updated 4 months ago
- Quantized Attention on GPU☆44Updated 8 months ago
- Implement Flash Attention using Cute.☆92Updated 7 months ago
- ☆31Updated 6 months ago
- ☆49Updated 2 months ago
- Standalone Flash Attention v2 kernel without libtorch dependency☆111Updated 10 months ago
- A GPU-optimized system for efficient long-context LLMs decoding with low-bit KV cache.☆56Updated last week
- ☆96Updated 10 months ago
- 使用 CUDA C++ 实现的 llama 模型推理框架☆58Updated 8 months ago
- ☆75Updated 2 months ago
- 使用 cutlass 仓库在 ada 架构上实现 fp8 的 flash attention☆74Updated 11 months ago
- ☆79Updated 6 months ago
- SpInfer: Leveraging Low-Level Sparsity for Efficient Large Language Model Inference on GPUs☆50Updated 4 months ago
- GPTQ inference TVM kernel☆40Updated last year
- A standalone GEMM kernel for fp16 activation and quantized weight, extracted from FasterTransformer☆94Updated 2 weeks ago
- DeeperGEMM: crazy optimized version☆71Updated 2 months ago
- ⚡️FFPA: Extend FlashAttention-2 with Split-D, achieve ~O(1) SRAM complexity for large headdim, 1.8x~3x↑ vs SDPA.🎉☆194Updated 2 months ago
- ☆89Updated 2 months ago
- ☆11Updated 6 months ago
- QuTLASS: CUTLASS-Powered Quantized BLAS for Deep Learning☆57Updated 2 weeks ago
- [DAC'25] Official implement of "HybriMoE: Hybrid CPU-GPU Scheduling and Cache Management for Efficient MoE Inference"☆64Updated last month
- TileFusion is an experimental C++ macro kernel template library that elevates the abstraction level in CUDA C for tile processing.☆91Updated last month
- Open deep learning compiler stack for cpu, gpu and specialized accelerators☆19Updated this week
- Several optimization methods of half-precision general matrix vector multiplication (HGEMV) using CUDA core.☆63Updated 10 months ago
- Tile-based language built for AI computation across all scales☆30Updated last week