UPSTREAM PR #17289: convert : remove unnecessary chat template patching

[DajanaV]

Mirrored from ggml-org/llama.cpp#17289

Remove chat template patching that is no longer necessary:

• Support GLM 4.6 template ochafik/minja#5
• Enable SmolLM 3 template. google/minja#74

[loci-agentic-ai]

Access the complete analysis in the LOCI Dashboard

Performance Analysis Summary

Based on the comprehensive analysis of project_id=2621b8c0-b5ce-11f0-b333-453f42058aa1 comparing version d92759bb-fa39-48d2-8e30-324d7703c52c against baseline 032e8f46-edb9-425d-b2b4-b3ae82d31e9b, the changes show minimal performance impact with no meaningful functional modifications.

Performance Metrics Overview

The analysis reveals negligible performance variations:

• Highest Response Time Change: can_reuse function (+0.096%, +0.063 ns absolute)
• Highest Throughput Change: llm_ffn_exps_block_regex function (+0.153%, +0.153 ns absolute)
• Power Consumption: No significant changes across all binaries (<0.001% variation)

Technical Analysis

Function-Level Insights: Both functions showing the highest percentage changes were unmodified between versions, indicating the variations represent measurement noise rather than code changes. The can_reuse function operates as a simple computational unit with 65 ns execution time, while llm_ffn_exps_block_regex handles regex processing with minimal self-execution overhead.

CFG Comparison: The control flow graphs for can_reuse are identical between versions, confirming no structural or assembly-level changes. The 0.06 ns timing difference stems from environmental factors rather than code modifications.

GitHub Code Review: The associated PR #221 removes unnecessary chat template patching in Python conversion scripts, affecting only the model conversion process without impacting runtime inference performance.

Impact Assessment

Core Function Impact: None of the critical inference functions (llama_decode, llama_encode, llama_tokenize) show performance changes, indicating no impact on tokens per second throughput.

Power Efficiency: All binaries maintain consistent power consumption profiles with variations below measurement precision.

Overall Assessment: The sub-nanosecond timing variations are within normal system noise levels and do not represent functional regressions or performance concerns. The changes reflect successful removal of conversion-time workarounds without affecting runtime performance.