FlagTree 0.5.0 Release

• Added Features

  • Officially released TLE feature, supporting 31 core primitives for TLE-Lite and TLE-Struct:

    • TLE-Lite:

      • NVIDIA: tle.load(is_async=True)，tle.extract_tile，tle.insert_tile，tle.device_mesh,tle.sharding，tle.distributed_barrier，tle.remote，tl.load（for local_ptr），tl.store（for local_ptr），tl.atomic_add/and/cas/max/min/or/xchg/xor（for local_ptr）.

      • Tsingmicro: tle.device_mesh, tle.sharding, tle.remote, and tl.store（for local_ptr）.

    • TLE-Struct:

      • NVIDIA: tle.gpu.alloc, tle.gpu.copy, tle.gpu.local_ptr, and tle.gpu.local_ptr (for remote).

      • Huawei Ascend: tle.dsa.alloc, tle.dsa.copy, tle.dsa.local_ptr, tle.dsa.local_ptr (for remote), tle.dsa.to_tensor, tle.dsa.to_buffer, tle.add/sub/mul/div/max/min, tle.dsa.pipeline, tle.dsa.parallel, tle.dsa.hint, tle.dsa.extract_slice, tle.dsa.insert_slice, tle.dsa.extract_element, tle.dsa.subview, tle.dsa.ascend.{UB,L1,L0A,L0B,L0C}.

      • Tsingmicro: tle.dsa.alloc, tle.dsa.local_ptr, and tle.dsa.local_ptr (for remote).

  • Officially released FLIR (FlagTree Intermediate Representation) feature.

    • Supports 76 Triton language primitives and 103 operators.

    • Enabled shared compiler passes across backends

    • Supported hardware platforms: AIPU, Huawei Ascend, and Tsingmicro