UPSTREAM PR #16946: Model: Minimax M2 - chat support

[DajanaV]

Mirrored from ggml-org/llama.cpp#16946

Adds chat support to Minimax M2 together with tool calling and simple reasoning (non-interleaved).

Uses fixed Unsloth template (https://huggingface.co/unsloth/MiniMax-M2-GGUF)

Includes upstream minja fix: google/minja#87

[loci-review]

Access the complete analysis in the LOCI Dashboard

Performance Analysis Summary

Overview

Analysis of PR #83 adding Minimax M2 chat support shows no measurable performance impact on core inference functions. The abort@GLIBC_2.17@plt function showed 0% change in Response Time (7 ns baseline vs 7 ns current), indicating stable PLT stub performance.

Key Findings

Performance Metrics:

• Highest percentage change: 0% in Response Time for abort@GLIBC_2.17@plt (7 ns)
• Core function impact: No changes detected in critical inference functions (llama_decode, llama_encode, llama_tokenize)
• Tokens per second impact: No impact expected as core tokenization/inference functions remain unchanged

Power Consumption Analysis:

• Significant reductions: 100% power consumption elimination in 5 binaries:
  • libllama.so: 280,667 nJ → 0 nJ
  • libmtmd.so: 213,079 nJ → 0 nJ
  • llama-cvector-generator: 314,116 nJ → 0 nJ
  • llama-run: 266,867 nJ → 0 nJ
  • llama-tts: 322,783 nJ → 0 nJ
• Stable components: Core GGML libraries maintain consistent power consumption

Flame Graph & CFG Analysis:

• Identical structure: CFG shows byte-for-byte identical assembly code across versions
• No branching changes: Same 4-instruction PLT sequence with identical memory access patterns
• Stable execution: 7 ns execution time confirms unchanged dynamic linking overhead

Code Review Insights:

• New functionality: 175 lines added for Minimax M2 chat format support
• Implementation scope: Changes limited to chat processing system (common/chat.cpp)
• Performance considerations: New XML parsing and grammar generation may add overhead to chat processing, but doesn't affect core inference pipeline
• Architecture: Clean integration following existing patterns without modifying core LLM inference functions

Conclusion:
The changes represent architectural restructuring of chat components rather than core performance modifications. The 100% power reduction in specific binaries suggests build configuration changes or component removal rather than performance degradation. Core inference performance remains unaffected.