UPSTREAM PR #16992: chore : fix models indent after refactor

[DajanaV]

Mirrored from ggml-org/llama.cpp#16992

Quite a few model files had incorrect indent after the refactor, this is just a cleanup of that.

[loci-agentic-ai]

Access the complete analysis in the LOCI Dashboard

Performance Analysis Summary: LLaMA.cpp PR #77

Critical Function Analysis

Based on the performance data analysis, the functions with measurable changes are not part of the core LLaMA.cpp critical functions listed in the project summary. The observed changes affect:

• _RegexMask constructor (standard library): +0.08% response time (+0.01 ns)
• llm_build_olmo2 constructor: +0.42% throughput (+5.82 ns)
• llm_build_llada constructor: +3.65% bottleneck (+4.20 ns)

Key Finding: None of the performance-critical functions (llama_decode(), llama_encode(), llama_tokenize(), llama_model_load_from_file(), etc.) show measurable performance changes in this analysis.

KPI Impact Assessment

1. Tokens Per Second

Impact: No measurable impact

• Analysis: Core inference functions (llama_decode, llama_encode, llama_tokenize) show no performance changes
• Reference Baseline: 2ms llama_decode slowdown = 7% tokens/sec reduction on smollm:135m/i7-1255U
• Current Changes: No changes detected in inference pipeline functions
• Conclusion: Tokens per second performance remains unchanged

2. Power Consumption

Impact: Minimal changes in specific binaries

• libllama.so: +0.002% increase (280,667 nJ vs 280,662 nJ)
• llama-run: -0.0% decrease (266,867 nJ vs 266,868 nJ)
• All other binaries: No measurable change (0.0%)
• Root Cause: Minor variations in model constructor execution times
• Assessment: Changes fall within measurement precision limits

3. Quantization Efficiency

Impact: No impact

• Analysis: No changes detected in quantization-related functions
• Functions Checked: llama_model_quantize(), quantization format handlers
• Conclusion: Quantization performance remains unaffected

4. Memory Usage

Impact: No impact

• Analysis: No changes in memory management functions
• Functions Checked: llama_memory_clear(), llama_memory_seq_rm(), KV cache operations
• Conclusion: Memory allocation and management efficiency unchanged

5. Batch Processing

Impact: No impact

• Analysis: No changes in batch processing functions
• Functions Checked: llama_batch_init(), llama_decode() with batches
• Conclusion: Batch processing efficiency remains unchanged

Technical Analysis

Change Attribution

The observed performance variations are attributed to:

• Indentation changes: Affect debug information and binary layout
• Compiler optimization: Minor differences in code structure influence optimization decisions
• Measurement precision: Changes are within typical measurement variance (±0.01-0.1 ns)

Control Flow Analysis

• No functional changes: All modifications are whitespace-only
• Binary layout effects: Debug information changes may affect instruction cache alignment
• Symbol organization: Different code formatting can influence symbol table layout

Action Items

Code Quality Improvements

• Completed: Indentation standardization across 39 model files
• Benefit: Enhanced code maintainability and consistency
• Risk: None - purely cosmetic changes

Performance Monitoring

• Current Status: Performance variations within acceptable thresholds
• Recommendation: No performance-related actions required
• Validation: Core inference pipeline functions remain unaffected

Build Optimization

• Observation: Minor binary layout differences due to debug information changes
• Impact: Negligible effect on runtime performance
• Action: No build system modifications needed

Conclusion

This PR represents a code quality improvement with no functional performance impact. The observed performance variations are measurement artifacts from compilation differences rather than algorithmic changes. All critical LLaMA.cpp functions maintain their performance characteristics, ensuring no impact on inference speed, power efficiency, or memory usage.