FlyDSL
A Python-native, MLIR-based GPU kernel domain-specific language (DSL) introduced by AMD for generating optimized GPU kernels. It has been integrated into the TorchInductor compilation pipeline.
Timeline (last 90 days)
The difference between FlyDSL and Triton is that Triton targets developers who need to hand-write kernels, while FlyDSL provides out-of-the-box acceleration through automatic routing of the existing operator API, so the two are more likely complementary than substitutes.
The official PyTorch blog announced that users of AMD MI350 series GPUs can use FlyDSL directly through the existing operator API.
FlyDSL covers core operations such as dense and grouped GEMM, fused multi-head attention (FMHA/Attention), MXFP8/MXFP4 scaled GEMM, RMSNorm and TopK.
Developers can obtain accelerated support for FlyDSL operators on AMD's new-generation hardware without rewriting model code or switching frameworks.
Users continue to call familiar interfaces such as torch.matmul and F.scaled_dot_product_attention, and the underlying layer automatically routes to the FlyDSL implementation.
The currently confirmed support scope is the MI350 series.
Users of MI300 or other AMD GPUs cannot directly obtain equivalent acceleration for now and need to wait for subsequent expansion.
The official announcement did not disclose specific performance improvement figures, comparison data with rocBLAS/composable_kernel, or implementation details of FlyDSL itself (for example, whether it is based on MLIR or a custom DSL).
FlyDSL is a Python-native, MLIR-based GPU kernel DSL (Domain-Specific Language) introduced by AMD.
The AMD team integrated FlyDSL into TorchInductor as a new backend.
6 more timeline events
All Facts (16)
The difference between FlyDSL and Triton is that Triton targets developers who need to hand-write kernels, while FlyDSL provides out-of-the-box acceleration through automatic routing of the existing operator API, so the two are more likely complementary than substitutes.
50%UnverifiedThe official announcement did not disclose specific performance improvement figures, comparison data with rocBLAS/composable_kernel, or implementation details of FlyDSL itself (for example, whether it is based on MLIR or a custom DSL).
50%UnverifiedFlyDSL covers core operations such as dense and grouped GEMM, fused multi-head attention (FMHA/Attention), MXFP8/MXFP4 scaled GEMM, RMSNorm and TopK.
50%UnverifiedDevelopers can obtain accelerated support for FlyDSL operators on AMD's new-generation hardware without rewriting model code or switching frameworks.
50%UnverifiedUsers continue to call familiar interfaces such as torch.matmul and F.scaled_dot_product_attention, and the underlying layer automatically routes to the FlyDSL implementation.
50%UnverifiedThe currently confirmed support scope is the MI350 series.
50%UnverifiedUsers of MI300 or other AMD GPUs cannot directly obtain equivalent acceleration for now and need to wait for subsequent expansion.
50%UnverifiedFlyDSL is a Python-native, MLIR-based GPU kernel DSL (Domain-Specific Language) introduced by AMD.
50%UnverifiedFlyDSL is connected to the core stages of Inductor, including lowering, compilation, caching, benchmarking, and the auto-tuning pipeline.
50%UnverifiedAMD has brought optimized kernels including GEMM, Group GEMM, and Flex Attention to PyTorch via FlyDSL.
50%UnverifiedOn the AMD MI355, FlyDSL's FP16 GEMM achieved approximately 19% geometric mean speedup compared to Triton.
50%UnverifiedOn the decode shapes of Minimax M3, FlyDSL Flex Attention achieved up to 5.8x speedup relative to Triton.
50%UnverifiedOn the prefill sizes of DeepSeek V2 and V3, FlyDSL Flex Attention achieved close to a 2x improvement.
50%UnverifiedBy being based on MLIR, FlyDSL can leverage MLIR's existing Pass infrastructure (such as loop transformations, vectorization, and memory hierarchy analysis) to generate code that more closely matches hardware characteristics, without needing to implement these optimization logics from scratch.
50%UnverifiedThe official PyTorch blog announced that users of AMD MI350 series GPUs can use FlyDSL directly through the existing operator API.
50%UnverifiedThe AMD team integrated FlyDSL into TorchInductor as a new backend.
50%