Triton language and compiler GitHub repository

README, MIT, triton-lang org. Triton is an MLIR-based language and compiler for tiled tensor primitives; per-kernel autotuning and an MLIR pass pipeline are documented. TRITON_F32_DEFAULT sets the default input precision of tl.dot on 32-bit floats (ieee, tf32, or tf32x3) - a real, quiet lever on which tensor-core path a Triton kernel takes. FlashAttention's own README points at the Triton fused-attention tutorial as the readable reference implementation. Used here for the 'DSL above the tensor-core instruction' layer, not for any speed claim.