UPSTREAM PR #1316: fix: sd-server memory leak

[loci-dev]

Note

Source pull request: leejet/stable-diffusion.cpp#1316

free results (sd_images_t array) from generate_images() after write_image_to_vector()

[loci-review]

Overview

Analysis of stable-diffusion.cpp compared 49,758 functions across two versions, identifying 56 modified functions, 1 new function, and 49,701 unchanged. The target version introduces a memory leak fix in sd-server (commit 85676b7) with minimal performance impact.

Binaries analyzed:

• build.bin.sd-server: -0.028% power consumption (527,129.70 nJ → 526,980.49 nJ)
• build.bin.sd-cli: 0.0% power consumption change (491,105.58 nJ → 491,105.69 nJ)

Overall impact is positive, with hot-path optimizations offsetting minor regressions in non-critical functions.

Function Analysis

Performance-critical improvements:

• ggml_compute_forward_map_custom3: Response time decreased 32.85% (-76.86ns: 233.99ns → 157.13ns); throughput time decreased 35.05% (-76.85ns: 219.25ns → 142.40ns). This GGML tensor operation is called thousands of times per inference, compounding to meaningful savings.

• GGMLRunner::copy_data_to_backend_tensor: Response time decreased 11.45% (-197.67ns: 1,725.58ns → 1,527.91ns); throughput time decreased 56.89% (-197.69ns: 347.49ns → 149.80ns). Critical for GPU inference, reducing data transfer overhead between host and backend memory.

• std::vector<sd_lora_t>::end: Response time decreased 69.44% (-183.29ns: 263.94ns → 80.65ns); throughput time decreased 75.41% (-183.29ns: 243.07ns → 59.78ns). Benefits LoRA configuration processing during request handling.

Non-critical regressions:

• std::vector::begin and std::vector::begin: Both show ~289% throughput increases (+180.81ns: 62.49ns → 243.30ns), but absolute impact is negligible for these STL accessors used during initialization.

• nlohmann::json::lexer::scan_string: Response time increased 6.01% (+2,530ns: 42,105.09ns → 44,635.09ns) with stable throughput time (-0.42%). Affects request parsing at boundary, not inference hot path.

Other analyzed functions showed minor variations in standard library operations with negligible real-world impact.

Additional Findings

The memory leak fix (free_results() function added at three cleanup sites) prevents unbounded memory growth in long-running servers without measurable energy overhead. Indirect benefits include reduced heap fragmentation improving allocator performance, evident in the 50% throughput improvement for std::_Construct string operations. The combined hot-path optimizations provide approximately 0.79ms savings per image generation, while the memory stability improvements are essential for production deployments handling sustained workloads.

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