文件最后提交记录最后更新时间
refactor(docs): Docs update 4 个月前
docs(install): modify installation instruction docs 2 个月前
fix(autotuner): disable async compilation for functions not directly decorated with @triton.jit Co-authored-by: Xuan Peng<pengxuan9@huawei.com> # message auto-generated for no-merge-commit merge: !1279 merge fix/autotune-issue-0212 into main fix(autotuner): disable async compilation for functions not directly decorated with @triton.jit Created-by: HinPeng Commit-by: Xuan Peng Merged-by: ascend-robot Description: ## Description This PR fixes a bug in autotuner where asynchronous compilation mode was enabled even for functions that are not directly decorated with @triton.jit. Currently, autotuner unconditionally sets async_compile = True when it is constructed, regardless of the nature of the wrapped function. Asynchronous compilation relies on the function being a direct triton.jit kernel (i.e., having the expected attributes and being compiled via the Triton JIT pipeline). However, in some cases the function passed to autotuner is indirectly jitted, for example when it is wrapped by libentry() or other utility decorators. This leads to runtime errors since other decorators may not compatible with AsyncCompileMode yet. ## Solution: Disable async compilation unless self.fn is a direct instance of a triton.jit function. See merge request: Ascend/triton-ascend!12793 个月前
feat(performance):Update the performance charts in README.md 2 个月前
docs(install): modify installation instruction docs 2 个月前
docs(api): add triton op API docs Co-authored-by: wangzhanpeng5<wangzhanpeng5@huawei.com> # message auto-generated for no-merge-commit merge: !1363 merge 0309_main_docs_api into main docs(api): add triton op API docs Created-by: wangzhanpeng5 Commit-by: wangzhanpeng5 Merged-by: ascend-robot Description: add triton op API docs See merge request: Ascend/triton-ascend!13632 个月前
refactor(docs): Docs update 4 个月前
fix:Fix the infinite loop in isUsedWithCondition caused by nested for-loops Co-authored-by: wangzhanpeng5<wangzhanpeng5@huawei.com> # message auto-generated for no-merge-commit merge: !1480 merge 0330_main_zj_isUsedWithCondition into main fix:Fix the infinite loop in isUsedWithCondition caused by nested for-loops Created-by: wangzhanpeng5 Commit-by: wangzhanpeng5 Merged-by: ascend-robot Description: ## 修复ISSUE [https://gitcode.com/Ascend/triton-ascend/issues/376](https://gitcode.com/Ascend/triton-ascend/issues/376) ## 同步release分支修复PR [https://gitcode.com/Ascend/triton-ascend/pull/1431](https://gitcode.com/Ascend/triton-ascend/pull/1431) ## 问题描述 在编译包含循环且循环体内调用复杂排序算子(如 triton_patch_topk)的 Triton 内核时,编译过程卡在 TTIR → TTADAPTER 转换阶段。通过 GDB 定位,发现死循环发生在 mlir::triton::isUsedWithCondition 递归分析循环内值的依赖关系时,由于缺少 visited 集合,遇到循环依赖(如循环迭代参数在循环体内被反复使用)时导致无限递归,最终栈溢出或超时。 ## 问题原因分析 - isUsedWithCondition 函数递归遍历 Value 的所有使用,当遇到 LoopLikeOpInterface(如 scf.for)时,会继续递归分析循环迭代参数,而这些参数最终又指向循环体内的值,形成循环依赖。 - 原始实现未记录已访问的 Value,导致同一 Value 被重复递归,形成无限递归。 ## 解决方法 为 isUsedWithCondition 引入 llvm::DenseSet<Value> visited 记录已访问的 Value,并在递归开始时检查,若已访问则直接返回 false。加入 visited 集合后,递归遍历时记录已访问的 Value,遇到循环依赖(如循环迭代参数重复出现)时直接返回,从而切断无限递归链,确保函数能在有限步内完成分析。 See merge request: Ascend/triton-ascend!14802 个月前
docs(install): modify installation instruction docs 2 个月前
refactor(docs): Docs update 4 个月前
docs(install): modify installation instruction docs 2 个月前
refactor(docs): Docs update 4 个月前
docs(install): modify installation instruction docs 2 个月前
docs(install): modify installation instruction docs 2 个月前
docs(install): modify installation instruction docs 2 个月前
refactor(docs): Docs update 4 个月前
refactor(docs): Docs update 4 个月前