remove revertUseOfInputCache

[liqiangxl]

Prefer to avoid illegal input cache instead of revertUseOfInputCache,

[liqiangxl]

!test

[github-actions]

Description

• Remove redundant input cache promotion logic

• Avoid illegal caching of global memory inputs

• Simplify memory type promotion workflow

• Fix potential invalid IR graph construction

---

Changes walkthrough 📝

	Relevant files
Bug fix	utils.cpp Remove input cache reversion logic                                              csrc/scheduler/utils.cpp Remove revertUseOfInputCache function that handled input cache reversion Eliminate check for input cache promotion to global memory Simplify promoteProducerMemoryTypes by removing special case handling Prevent invalid caching of global memory inputs by design +0/-65

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 No relevant tests

⚡ Recommended focus areas for review Possible Issue The removal of revertUseOfInputCache may lead to unnecessary global memory caching of input tensors, which could degrade performance. The new code no longer checks for input caches being promoted to global memory, potentially resulting in redundant memory operations. if (it == tvs_to_promote.end() || it->second == producer->getMemoryType()) { continue; } // Required memory type of the producer const auto new_mem_type = it->second; producer->setMemoryType(new_mem_type); } }