FlagDNN Release Notes

Contents

FlagDNN Release Notes#

v0.2.0#

Note

This is a preview release. The version number shown is a pre-release identifier and may change upon final release. Content in this preview is for reference only and does not constitute a commitment or warranty for the final product.

  • Added Features

    • Graph Execution Engine — New graph-mode execution path with IR capture, kernel fusion, auto-tuning, and multi-node planning. Supports graph-level operator dispatch and memory optimization.

    • Neural Network Operators — conv1d, conv2d, conv3d, conv_fprop, conv_dgrad, conv_wgrad, causal_conv1d, max_pool2d, max_pool3d, avg_pool1d, avg_pool2d, avg_pool3d, adaptive_avg_pool2d, adaptive_avg_pool3d, adaptive_max_pool2d, adaptive_max_pool3d, gelu_approx_tanh, silu, swish, leaky_relu, leaky_relu_, prelu, elu, elu_, rrelu, rrelu_, mish, softplus, softsign, softshrink, softmin, log_softmax, hardswish, relu6, selu, glu, celu, tanh, sigmoid, sigmoid_backward, logsigmoid, hardtanh, hardtanh_, threshold, threshold_.

    • Normalization Operators — batchnorm, batchnorm_inference, layernorm, rmsnorm, group_norm.

    • Linear Algebra Operators — mm, mv, dot, matmul.

    • Math Operators — exp, log, rsqrt, square, positive, isinf, isnan, max, min, scale, ge, gt, le, lt, maximum, minimum, fmax, fmin, bitwise_and, bitwise_or, bitwise_xor, bitwise_not, logical_and, logical_or, logical_not, unary.

    • Reduction Operators — cummin, cummax, any, all, reduction.

    • Loss Operators — kl_div, mse_loss, l1_loss.

    • Tensor Operators — embedding, one_hot, concatenate, gen_index, identity, reshape, slice, transpose, binary_select.

    • Fused Operators — add_square, rmsnorm_rht_amax.

    • Other Operators — interpolate.

    • Attention Operators — sdpa, sdpa_backward (graph mode).

    • Iluvatar Backend — Added Iluvatar GPU backend support with heuristics config and op blacklist.

    • Operator Registry — Added conf/operators.yaml for standardized operator metadata.

    • Graph Benchmark Suite — Comprehensive benchmark framework for graph-mode operators.

  • Enhanced Features

    • Eager-mode operators underwent deep performance tuning and framework restructuring.

    • Triton kernel launch patterns optimized to reduce overhead.

    • Benchmark framework unified with standardized shape configurations.

v0.1.0#

Initial release of FlagDNN.

  • Added Features

    • Deep neural network computing library with multi-backend support.

    • ReLU operator with Triton kernel implementation.

    • Flexible multi-backend support mechanism.

    • PyTorch integration via flag_dnn.ops API.

Contents