UPSTREAM PR #17063: Add ops needed for new hybrid models: SOFTPLUS, EXPM1, TRI, SOLVE_TRI, CUMSUM

[DajanaV]

Mirrored from ggml-org/llama.cpp#17063

The ops needed for the new hybrid models including Qwen3 Next and Kimi Linear.

Prerequisite to merging ggml-org/llama.cpp#16095

[loci-agentic-ai]

Access the complete analysis in the LOCI Dashboard

Based on the analysis of PR #109 in the llama.cpp repository, this pull request introduces new mathematical operations to the GGML tensor library without impacting core inference performance.

Summary

The PR adds several new mathematical operations including CUMSUM, TRI (triangular matrix operations), SOLVE_TRI (triangular system solver), and unary operations EXPM1 and SOFTPLUS. These additions expand GGML's mathematical capabilities but do not modify existing inference pathways.

Key Findings

Performance Impact: The highest percentage change identified was in the ggml_set_op_params_i32 function with increased Response Time. However, this function is a utility for setting operation parameters and is not part of the critical inference path. Core functions like llama_decode, llama_encode, and llama_tokenize remain unchanged, indicating no impact on tokens per second performance.

Core Function Analysis: The changes do not affect any of the performance-critical components identified in the project summary:

• Model Processing Module functions remain untouched
• Token Processing Module is unaffected
• Memory Management and Batch Processing modules are unchanged
• No modifications to inference-critical paths

Power Consumption: Analysis shows minimal power consumption changes at the binary level, with new operations adding computational overhead only when explicitly used. The main affected binary is the core GGML library.

Code Structure: The implementation follows established GGML patterns by adding new operation types to enums, implementing forward computation functions, and extending backend support. The changes are well-contained within the mathematical operations framework.

Code Review: The GitHub analysis shows proper implementation of new mathematical operations with appropriate test coverage and documentation updates. No critical issues were identified in the implementation.

Conclusion: This PR represents a feature addition rather than a performance modification. The new mathematical operations expand GGML's capabilities without affecting existing inference performance, making it a low-risk enhancement to the codebase's mathematical foundation.