best_compile_options.py

BEST_CONFIG_RT_DETR = {
    "triton.cudagraphs": True,
    "triton.cudagraph_trees": True,
    "efficient_conv_bn_eval_fx_passes": True,
    "force_mixed_mm": False,
    "aggressive_fusion": True,
    "max_autotune": True,
    "autotune_local_cache": False,
    "static_weight_shapes": True,
    "epilogue_fusion": True,
    "epilogue_fusion_first": False,
    "pattern_matcher": False,
    "force_fuse_int_mm_with_mul": True,
    "use_mixed_mm": False,
    "reorder_for_compute_comm_overlap": False,
    "max_autotune_pointwise": True,
    "force_same_precision": False,
    "conv_1x1_as_mm": True,
    "split_reductions": False,
    "shape_padding": False,
    "comprehensive_padding": False,
    "pad_channels_last": False,
    "force_shape_pad": True,
    "permute_fusion": True,
    "assume_aligned_inputs": False,
    "cpp.enable_unsafe_math_opt_flag": True,
    "triton.cudagraph_skip_dynamic_graphs": True,
    "triton.slow_path_cudagraph_asserts": False,
    "triton.cudagraph_support_input_mutation": False,
    "triton.dense_indexing": True,
    "triton.autotune_cublasLt": True,
    "triton.tiling_prevents_pointwise_fusion": False,
    "triton.tiling_prevents_reduction_fusion": True,
    "cuda.enable_cuda_lto": False,
    "cuda.use_fast_math": False,
}