Triton 编译器发布: Triton 3.8.0 Release Notes

来源摘要

# Triton 3.8.0 Release Notes ## Table of Contents - [Dialect & Frontend](#dialect--frontend-38) - [Backend & Compiler](#backend--compiler-38) - [AMD/HIP Backend](#amdhip-backend-38) - [NVIDIA Backend](#nvidia-backend-38) - [Gluon & Layout Improvements](#gluon--layout-improvements-38) - [Kernels & Benchmarks](#kernels--benchmarks-38) - [Proton Profiling](#proton-profiling-38) - [Testing & CI](#testing--ci-38) - [Build & Infrastructure](#build--infrastructure-38) - [Documentation](#documentation-38) - [Breaking Changes](#breaking-changes-38) - [Contributors](#contributors-38) --- ## Dialect & Frontend ### New Features - **Aggregate types:** `@triton.aggregate` and `@gluon.aggregate` are now public APIs. Aggregates support inherited fields, default values, generated constructors, immutable instances, and `aggregate_replace()` ([#10095](https://github.com/triton-lang/triton/pull/10095), [#9572](https://github.com/triton-lang/triton/pull/9572)) - **`tl.topk`:** Added a `descending` argument. Set `descending=False` to return the smallest values ([#9355](https://github.com/triton-lang/triton/pull/9355)) - **Tensor descriptors:** Tensor descriptors can be passed inside tuple-valued kernel arguments ([#9422](https://github.com/triton-lang/triton/pull/9422)) - **Interpreter:** Added support for `tl.dot_scaled` ([#10311](https://github.com/triton-lang/triton/pull/10311)) ### Improvements - **Autotuning listener:** Added a listener that reports the selected configuration, measured timings, tuning duration, and disk-cache status ([#10125](https://github.com/triton-lang/triton/pull/10125)) - **JIT cache keys:** Dependency cache keys are now generated deterministically ([#10494](https://github.com/triton-lang/triton/pull/10494)) ### Bug Fixes - **Division and atomics:** `tl.fdiv(...,