> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.nvidia.com/cudnn/latest/llms.txt. # Latest ## Docs - [NVIDIA cuDNN](https://docs.nvidia.com/cudnn/latest/overview.md): GPU-accelerated building blocks for deep learning. - [Samples](https://docs.nvidia.com/cudnn/latest/samples.md) - [Model Samples](https://docs.nvidia.com/cudnn/latest/model-samples.md) - [Operations Overview](https://docs.nvidia.com/cudnn/latest/operations/overview.md) - [Python API Naming Conventions](https://docs.nvidia.com/cudnn/latest/operations/api_conventions.md) - [Determinism Support](https://docs.nvidia.com/cudnn/latest/operations/determinism.md) - [Attention](https://docs.nvidia.com/cudnn/latest/operations/Attention.md) - [Ahead-of-time Attention (SM120 C/C++)](https://docs.nvidia.com/cudnn/latest/operations/Attention/aot_sm120_attention.md) - [Gated Attention Block (SM107)](https://docs.nvidia.com/cudnn/latest/operations/Attention/gated_attention_block.md) - [Attention Gate + Quantize Decode Step (SM100-class)](https://docs.nvidia.com/cudnn/latest/operations/Attention/gate_quant_decode.md) - [Compact GQA backward (SM107)](https://docs.nvidia.com/cudnn/latest/operations/Attention/compact_gqa_backward.md) - [NVFP4 Attention QAT Backward](https://docs.nvidia.com/cudnn/latest/operations/Attention/nvfp4_attention_qat_backward.md) - [SDPA Backward (SM100/SM103, d = 128)](https://docs.nvidia.com/cudnn/latest/operations/Attention/sdpa_bwd_sm100_d128.md) - [SDPA Backward (SM107, d = 256–512)](https://docs.nvidia.com/cudnn/latest/operations/Attention/sdpa_bwd_sm107.md) - [SDPA Backward (SM120)](https://docs.nvidia.com/cudnn/latest/operations/Attention/sdpa_bwd_sm120.md) - [Sparse Attention](https://docs.nvidia.com/cudnn/latest/operations/sparse_attention.md) - [Native Sparse Attention (NSA)](https://docs.nvidia.com/cudnn/latest/operations/sparse_attention/nsa.md) - [Simplified Sparse Attention](https://docs.nvidia.com/cudnn/latest/operations/sparse_attention/simplified_sparse_attention.md) - [DeepSeek Sparse Attention (DSA)](https://docs.nvidia.com/cudnn/latest/operations/sparse_attention/dsa.md) - [Query Sparse Attention (QSA)](https://docs.nvidia.com/cudnn/latest/operations/sparse_attention/qsa.md) - [Indexed QSA training](https://docs.nvidia.com/cudnn/latest/operations/sparse_attention/qsa/qsa_indexed.md) - [HSTU Attention (Blackwell SM100/SM103)](https://docs.nvidia.com/cudnn/latest/operations/sparse_attention/hstu_attention.md) - [HSTU LayerNorm-Multiply-SiLU-Dropout (LMSD)](https://docs.nvidia.com/cudnn/latest/operations/sparse_attention/hstu_lmsd.md) - [Block Sparse Attention (BSA)](https://docs.nvidia.com/cudnn/latest/operations/sparse_attention/bsa.md) - [CSA Fused Compressor](https://docs.nvidia.com/cudnn/latest/operations/sparse_attention/csa.md) - [Aligned HCA Backward](https://docs.nvidia.com/cudnn/latest/operations/sparse_attention/csa/aligned_hca_backward.md) - [Flex Attention](https://docs.nvidia.com/cudnn/latest/operations/sparse_attention/flex_attention.md) - [Flex Attention Mask Plan Design](https://docs.nvidia.com/cudnn/latest/operations/sparse_attention/flex_attention/flex_attention_design.md) - [Block Scaling](https://docs.nvidia.com/cudnn/latest/operations/BlockScaling.md) - [NVFP4 block-scale conversion](https://docs.nvidia.com/cudnn/latest/operations/BlockScaling/nvfp4_block_scale_conversion.md) - [Causal Conv1d](https://docs.nvidia.com/cudnn/latest/operations/CausalConv1d.md) - [Causal Conv1d Python API](https://docs.nvidia.com/cudnn/latest/operations/CausalConv1d/causal_conv1d.md) - [Causal Conv1d Decode Update](https://docs.nvidia.com/cudnn/latest/operations/CausalConv1d/causal_conv1d_update.md) - [Concatenate](https://docs.nvidia.com/cudnn/latest/operations/Concatenate.md) - [Convolutions](https://docs.nvidia.com/cudnn/latest/operations/Convolutions.md) - [Fused Conv3D + Bias + Residual + RMSNorm + SiLU](https://docs.nvidia.com/cudnn/latest/operations/Convolutions/conv3d_postops.md) - [Custom Reduction](https://docs.nvidia.com/cudnn/latest/operations/CustomReduction.md) - [How custom reductions work](https://docs.nvidia.com/cudnn/latest/operations/CustomReduction/overview.md) - [FFT Causal Conv1d](https://docs.nvidia.com/cudnn/latest/operations/FFTCausalConv1d.md) - [Linear Attention](https://docs.nvidia.com/cudnn/latest/operations/LinearAttention.md) - [FLA Integration Shims](https://docs.nvidia.com/cudnn/latest/operations/LinearAttention/fla.md) - [Kimi Delta Attention in JAX](https://docs.nvidia.com/cudnn/latest/operations/LinearAttention/kda_jax.md) - [Fused GDN Decode / Verify Steps (FlashInfer Backend)](https://docs.nvidia.com/cudnn/latest/operations/LinearAttention/gdn_decode_flashinfer_backend.md) - [Matmul](https://docs.nvidia.com/cudnn/latest/operations/Matmul.md) - [Discrete Grouped GEMM + dSwiGLU (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/discrete_grouped_gemm_dswiglu.md) - [Discrete Grouped GEMM + SwiGLU (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/discrete_grouped_gemm_swiglu.md) - [FP8 Projection for Decode Rows (SM100-class)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/fp8_projection_decode.md) - [SwiGLU FP4 Gate / Up Product for Decode Rows (SM100-class)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/swiglu_fp4_decode.md) - [GEMM + Amax (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/gemm_amax.md) - [GEMM + dsReLU (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/gemm_dsrelu.md) - [GEMM + RoPE + MXFP8 Projection (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/gemm_proj_rope_mxfp8.md) - [GEMM + sReLU (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/gemm_srelu.md) - [GEMM + SwiGLU (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/gemm_swiglu.md) - [Grouped GEMM (SM100 BF16)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/grouped_gemm.md) - [Grouped GEMM + dGLU (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/grouped_gemm_dglu.md) - [Grouped GEMM + dsReLU (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/grouped_gemm_dsrelu.md) - [Grouped GEMM + dSwiGLU (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/grouped_gemm_dswiglu.md) - [Grouped GEMM + GLU (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/grouped_gemm_glu.md) - [Grouped GEMM + GLU + Hadamard (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/grouped_gemm_glu_hadamard.md) - [Grouped GEMM + GLU + Hadamard + Quant (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/grouped_gemm_glu_hadamard_quant.md) - [Grouped GEMM + Quant (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/grouped_gemm_quant.md) - [Grouped GEMM + Quant -- Unified (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/grouped_gemm_quant_unified.md) - [Grouped GEMM + sReLU (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/grouped_gemm_srelu.md) - [Grouped GEMM + SwiGLU (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/grouped_gemm_swiglu.md) - [Grouped GEMM + WGrad (Unified)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/grouped_gemm_wgrad.md) - [DSv4.1 mHC Projection/RMS Backward (SM100/SM103)](https://docs.nvidia.com/cudnn/latest/operations/Matmul/gemm_fusions/mhc_projection_bwd.md) - [MoE Grouped Matmul](https://docs.nvidia.com/cudnn/latest/operations/MoeGroupedMatmul.md) - [Shared-Expert Decode Pair (SM100-class)](https://docs.nvidia.com/cudnn/latest/operations/MoeGroupedMatmul/shared_expert_decode.md) - [Normalizations](https://docs.nvidia.com/cudnn/latest/operations/Normalizations.md) - [LayerNorm and RMSNorm in JAX](https://docs.nvidia.com/cudnn/latest/operations/Normalizations/norm_jax.md) - [RMSNorm + RHT + Amax (SM100)](https://docs.nvidia.com/cudnn/latest/operations/Normalizations/rmsnorm_rht_amax.md) - [Fused RMSNorm + SiLU](https://docs.nvidia.com/cudnn/latest/operations/Normalizations/rmsnorm_silu.md) - [Add + RMSNorm + Quantize Decode Step (SM100-class)](https://docs.nvidia.com/cudnn/latest/operations/Normalizations/add_norm_quant_decode.md) - [Pointwise and Reduction](https://docs.nvidia.com/cudnn/latest/operations/Pointwise.md) - [Engram Saved-State Gate](https://docs.nvidia.com/cudnn/latest/operations/Pointwise/engram_saved_gate.md) - [Resampling](https://docs.nvidia.com/cudnn/latest/operations/Resampling.md) - [Reshape](https://docs.nvidia.com/cudnn/latest/operations/Reshape.md) - [RoPE (Rotary Position Embedding)](https://docs.nvidia.com/cudnn/latest/operations/RoPE.md) - [Tail RoPE + Microscaled QDQ](https://docs.nvidia.com/cudnn/latest/operations/RoPE/rope-qdq.md) - [Prepared BF16 Tail RoPE](https://docs.nvidia.com/cudnn/latest/operations/RoPE/rope-tail.md) - [DSv4.1 Vision RoPE Backward](https://docs.nvidia.com/cudnn/latest/operations/RoPE/vision_rope_backward.md) - [Slice](https://docs.nvidia.com/cudnn/latest/operations/Slice.md) - [Transpose](https://docs.nvidia.com/cudnn/latest/operations/Transpose.md) - [GNN simple aggregation](https://docs.nvidia.com/cudnn/latest/operations/gnn/agg_simple.md) - [GAT and GATv2 attention](https://docs.nvidia.com/cudnn/latest/operations/gnn/mha_gat.md) - [Compile-time constant tensors](https://docs.nvidia.com/cudnn/latest/utilities/CompileTimeConstants.md) - [Adding Torch Custom Ops in cuDNN Frontend](https://docs.nvidia.com/cudnn/latest/utilities/adding_torch_custom_ops.md) - [CUDA Graphs](https://docs.nvidia.com/cudnn/latest/utilities/cuda-graphs.md) - [Custom Execution Plan](https://docs.nvidia.com/cudnn/latest/utilities/custom-execution-plan.md) - [Deviceless Ahead-of-time Compilation](https://docs.nvidia.com/cudnn/latest/utilities/deviceless-ahead-of-time-compilation.md) - [Dynamic Shapes and Kernel Cache](https://docs.nvidia.com/cudnn/latest/utilities/dynamic-kernel-cache.md) - [Limiting Execution Plan Shared Memory](https://docs.nvidia.com/cudnn/latest/utilities/shared-memory-limit.md) - [Framework integration performance: getting the best per-call performance from cuDNN Frontend](https://docs.nvidia.com/cudnn/latest/utilities/framework_integration_performance.md) - [First-class `cudnn.Handle` — design](https://docs.nvidia.com/cudnn/latest/utilities/handle_first_class_design.md) - [Python-native `cudnn.pygraph` and pluggable execution backends](https://docs.nvidia.com/cudnn/latest/utilities/python_graph_and_execution_backends.md) - [Composing multi-kernel blocks in Python](https://docs.nvidia.com/cudnn/latest/utilities/composing_kernel_blocks.md) - [C++ launcher: repeated calls launched from C++](https://docs.nvidia.com/cudnn/latest/utilities/cpp_launcher.md) - [Overview](https://docs.nvidia.com/cudnn/latest/developer/overview.md) - [Core Concepts](https://docs.nvidia.com/cudnn/latest/developer/core-concepts.md) - [Graphs](https://docs.nvidia.com/cudnn/latest/developer/graph-api.md) - [Hardware Forward Compatibility](https://docs.nvidia.com/cudnn/latest/developer/forward-compatibility.md) - [Odds and Ends](https://docs.nvidia.com/cudnn/latest/developer/misc.md) - [Debugging](https://docs.nvidia.com/cudnn/latest/developer/debugging.md) - [Installing cuDNN Backend](https://docs.nvidia.com/cudnn/latest/installation/backend.md) - [Prerequisites](https://docs.nvidia.com/cudnn/latest/installation/backend/prerequisites.md) - [Installing cuDNN Backend on Linux](https://docs.nvidia.com/cudnn/latest/installation/backend/linux.md) - [Installing cuDNN Backend on Windows](https://docs.nvidia.com/cudnn/latest/installation/backend/windows.md) - [Installing cuDNN Frontend](https://docs.nvidia.com/cudnn/latest/installation/frontend.md) - [cuDNN Frontend Dependencies](https://docs.nvidia.com/cudnn/latest/installation/frontend/frontend-dependencies.md) - [Installing the cuDNN Python Frontend](https://docs.nvidia.com/cudnn/latest/installation/frontend/python-frontend-install.md) - [Installing the cuDNN C++ Frontend](https://docs.nvidia.com/cudnn/latest/installation/frontend/cpp-frontend-install.md) - [Next Steps](https://docs.nvidia.com/cudnn/latest/installation/frontend/frontend-next-steps.md) - [Building and Running a cuDNN Dependent Program](https://docs.nvidia.com/cudnn/latest/installation/build-run-cudnn.md) - [Cross-Compiling cuDNN Samples](https://docs.nvidia.com/cudnn/latest/installation/cross-compiling.md) - [Supported Products](https://docs.nvidia.com/cudnn/latest/reference/support-matrix.md) - [FAQs](https://docs.nvidia.com/cudnn/latest/reference/faq.md) - [Support](https://docs.nvidia.com/cudnn/latest/reference/support.md) - [Software License Agreement](https://docs.nvidia.com/cudnn/latest/reference/eula.md) - [Acknowledgements](https://docs.nvidia.com/cudnn/latest/reference/acknowledgements.md) - [Notices](https://docs.nvidia.com/cudnn/latest/reference/notices.md)