The following PR includes optimizations and enhancements for Inductor's autotuning mechanism. The goal is to reduce the autotuning time by 50%, without degrading kernels latency.
The enhancements in this project are delivered for the FASTAUTOTUNE=1 / ‘Expert’ mode, where the baseline version is v2.7.1.
Final performance verifications were done on Ascend910_957 (90.90.93.32 station) with triton_ascend=3.2.2; torch_npu=2.7.1.post4.dev20260509; cann-9.1.0.

Main features added:

1. Configs Static Pruning: Prune a significant part of the original configs based on technics of UB-aware Filter, MFU Cache, NPU-Aware Config Generation and Diversity Filter
2. Dynamic Filter: Adaptive tiling config selection driven by predictive models.
3. Profiling Enhancement: Use the most optimal profiling method to speed up stats collection time without compromising quality
4. Logging and Stats collection for performance analisys and benchmarking.

Environment params to be set by a user to enable the feature:

1. FASTA_CONFIG_OPTIMIZER=1 (enables static pruning mechanism - default=0)
2. FASTA_DYNAMIC_FILTER=1 (activate the dynamic filter algo - default=0)
3. FASTA_MSPTI_EN=1 (defines mspti as an optimal profiling method - default=0)

Additional feature params

Statistics gathering

1. FASTA_AUTOTUNE_STATS=1 (activate autotune stats gathering - default=0)

Dynamic Filter algo

2. FASTA_R1_PCT (activate autotune stats gathering - default=0.3)
3. FASTA_BASE_BUDGET (Base measurement budget as a fraction of total configs (N) - default=0.35)
4. FASTA_HIGH_BUDGET (Controls the total R2 exploration budget - default=0.4)
5. FASTA_LOW_BUDGET (Trade-off between exploration and measurement cost - default=0.25)
6. FASTA_MAX_ROUNDS (Maximum number of optimization rounds - default=2)

【合入来源】

如有社区issue，请关联issue链接
请勿携带内部流程信息（需求链接、问题单、内部issue等）

• 需求
• 问题单
• issue/工单
• 重构优化
• 资料更新

【修改方案】

请描述修改内容的具体实现，涉及哪些组件之间进行交互，可以用1、2、3、...进行罗列
如果是需求或者重构类的PR，需要补充详细设计文档（说明上下游组件关系、时序图、类图、DFX能力等内容）

【资料变更】

请确认是否涉及资料变更。如涉及，需要在PR中体现，并简要说明修改内容。如不涉及，需填写“不涉及”

【接口变更】

请确认是否涉及跨代码仓或者客户面可见的接口变更。如涉及，需要详细说明接口以及对应的变更内容，同时需要在资料中体现。如不涉及，需填写“不涉及”

【功能验证】

说明测试场景，测试方法。如果本次测试方式与常规单元测试不同，请详细说明您的测试步骤
新增/变更内容是否已新增/适配UT测试用例看护，并补充测试自验证截图

【CheckList】

PR提交人对以下CheckList自检项进行全量自检，自检通过或不涉及，均修改 [ ] 为 [x]

• 代码注释完备，正确记录错误日志
• 代码实现进行了返回值、空指针等校验
• PR标题正确使用类型标签，如：feat、fix、refactor、docs、test等
• PR持续集成流水线（CI）执行通过，代码检查无异常