UPSTREAM PR #1270: feat: accept legacy image parameter on v1/images/edits

[loci-dev]

Note

Source pull request: leejet/stable-diffusion.cpp#1270

The OpenAI API originally supported only a single image, passed through 'image' instead of 'image[]'. 'Old' OpenAI modules (as recent as 1.69 from Debian stable), and probably other clients, send edit requests in this format.

[loci-review]

Overview

Analysis of stable-diffusion.cpp following commit 36b3937 ("feat: accept legacy image parameter on v1/images/edits") reveals minimal performance impact. Of 48,313 total functions, 55 were modified (0.11%), with 0 new or removed functions.

Binaries analyzed:

• build.bin.sd-server: 515,491.29 nJ → 519,006.76 nJ (+0.682%)
• build.bin.sd-cli: 480,109.60 nJ → 483,483.03 nJ (+0.703%)

The single-file change adds backward compatibility for legacy API parameters in HTTP request parsing, isolated from inference paths.

Function Analysis

Standard Library Regressions:

• std::swap<_Prime_rehash_policy>: +81ns throughput (+111%), +81ns response (+81%) — hash table rehashing utility
• std::shared_ptr::operator= (GGMLBlock, VAE variants): +80ns throughput (+103%), +80ns response (+8%) — smart pointer assignments during model initialization
• _M_allocate_buckets: +68ns throughput (+62%), +68ns response (+26%) — hash table bucket allocation
• nlohmann::json::create: +133ns throughput (+54%), +133ns response (+5%) — JSON object allocation

Standard Library Improvements:

• std::make_move_iterator: -169ns throughput (-68%), -169ns response (-59%)
• __normal_iterator::operator+: -69ns throughput (-48%), -69ns response (-42%)

Quantization Function:

• make_block_q4_Kx8: +642ns throughput (+8%), +642ns response (+8%) — quantized block repacking during model loading

All regressions occur in standard library functions with no source code changes, suggesting compiler optimization differences. Absolute timing changes remain under 1 microsecond. HTTP processing functions show throughput increases but negligible response time impact, confirming isolation from critical paths.

Additional Findings

The code change affects only HTTP request parsing (examples/server/main.cpp), adding one boolean check and conditional file retrieval. No changes to inference pipeline, GPU operations, quantization algorithms, or ML workloads. The make_block_q4_Kx8 regression affects model loading (one-time cost) rather than inference. For a model with 10,000 quantized blocks, cumulative impact is 6.4ms, negligible compared to typical loading times (5-30 seconds). All performance-critical paths (attention mechanisms, convolutions, VAE operations) remain unchanged.

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