UPSTREAM PR #21573: autoparser: fix MiniMax handling

[loci-dev]

Note

Source pull request: ggml-org/llama.cpp#21573

Overview

There was a problem handling the generation prompt from MiniMax because it shares a trailing newline with the non-generation-prompt line.

Additional information

Added extra tests for Minimax.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Nope.

[loci-review]

Overview

Analysis of commit f0b2b56 ("parser: fix MiniMax handling") shows negligible performance impact. Out of 125,224 functions analyzed, 84 were modified (0.067%), with no new or removed functions. All changes are compiler optimization artifacts, not source code modifications—the commit only changed MiniMax template parsing in 2 files.

Power Consumption: Virtually unchanged across all binaries (max change: -0.039%)

Binary	Power Change
libllama.so	-0.0002%
libggml.so	0.0%
libggml-base.so	0.0%
libggml-cpu.so	0.0%
libmtmd.so	-0.0001%
llama-tts	-0.039%
llama-cvector-generator	+0.0004%
llama-bench	0.0%
llama-quantize	0.0%
llama-tokenize	0.0%
llama-gguf-split	0.0%
llama-qwen2vl-cli	0.0%
llama-gemma3-cli	0.0%
llama-llava-cli	0.0%
llama-minicpmv-cli	0.0%

Function Analysis

All modified functions are STL templates, JSON utilities, HTTP handlers, and template engine code—none affect inference hot paths (matrix operations, attention, KV cache). Changes show balanced improvements and regressions:

Improvements (75% of functions):

• STL iterators: 45-74% faster (entry block consolidation, eliminated branches)
• HTTP utilities: 20-27% faster (removed trampoline blocks)
• Template operations: 25-36% faster (optimized entry sequences)

Regressions (25% of functions):

• std::swap, std::function::operator=: 75-110% slower (indirect jumps for PIC/ASLR)
• JSON operations: 14-82% slower (extra entry indirection)

All absolute changes under 200ns. No performance-critical functions affected.

💬 Questions? Tag @loci-dev