UPSTREAM PR #1276: fix: correct sdapi LoRA file handling

[loci-dev]

Note

Source pull request: leejet/stable-diffusion.cpp#1276

• open LoRA files through their discovered absolute paths
• do not resolve symlinks when building the LoRA list

[loci-review]

Overview

Analysis of 48,305 functions across two binaries revealed 63 modified functions (0.13%) from a single commit fixing LoRA file handling. Power consumption increased minimally: build.bin.sd-server (+0.601%, +3,097 nJ) and build.bin.sd-cli (+0.703%, +3,374 nJ).

Function counts: 63 modified, 2 new, 3 removed, 48,237 unchanged.

Function Analysis

The commit expanded LoraEntry from 2 to 3 std::string members (adding fullpath), causing expected regressions in object lifecycle operations:

Memory Operations:

• std::vector<LoraEntry>::_S_max_size: +227ns throughput (+184%), +227ns response (+142%)
• std::__new_allocator<LoraEntry>::allocate: +70ns throughput (+86%), +70ns response (+67%)
• std::__new_allocator<LoraEntry>::deallocate: +16ns throughput (+86%), +17ns response (+62%)

Lifecycle Operations:

• LoraEntry::operator=: +6ns throughput (+27%), +2,738ns response (+50%)
• LoraEntry::LoraEntry (move): +9ns throughput (+40%), +581ns response (+50%)
• LoraEntry::~LoraEntry: +6ns throughput (+28%), +280ns response (+49%)
• std::__copy_move_backward<LoraEntry>: +12ns throughput (+16%), +2,749ns response (+49%)

Application Logic:

• get_lora_full_path lambda: +71ns throughput (+81%), +1,551ns response (+45%)

Iterator Operations:

• __normal_iterator<LoraEntry*>::operator+: +44ns throughput (+53%), +44ns response (+42%)
• __normal_iterator<shared_ptr<LoraModel>*>::operator+: -69ns throughput (-48%), -69ns response (-42%) — improvement from algorithmic optimization

Standard Library Anomalies:

• std::_Hashtable::begin: +187ns throughput (+310%), +187ns response (+181%) — likely compiler optimization differences
• std::_Sp_counted_ptr_inplace<FluxFlowDenoiser>::_M_destroy: +189ns throughput (+180%), +189ns response (+61%) — unrelated to code changes

Other analyzed functions showed consistent patterns aligned with the structural change.

Additional Findings

Justification: The regressions are expected consequences of storing an additional string field. The refactoring eliminates filesystem syscalls (microseconds) by caching absolute paths, replaces double-pass lookups (std::any_of + fetch) with single-pass (std::find_if), and fixes path resolution bugs. All affected functions execute during initialization/cache updates, not in the inference pipeline. The sub-1% power increase and sub-microsecond absolute impacts are negligible compared to image generation workloads (seconds). Zero impact on GPU operations, tensor computations, or ML inference performance.

🔎 Full breakdown: Loci Inspector.
💬 Questions? Tag @loci-dev.