File size: 6,128 Bytes
5e4510c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 | # Configuration for function minimization example
max_iterations: 100
checkpoint_interval: 10
log_level: "INFO"
# LLM configuration
llm:
# primary_model: "gemini-2.0-flash-lite"
# primary_model: "gpt-4.1-nano"
primary_model: "o3"
primary_model_weight: 0.8
# secondary_model: "gemini-2.0-flash"
secondary_model: "gpt-4.1-mini"
secondary_model_weight: 0.2
# api_base: "https://generativelanguage.googleapis.com/v1beta/openai/"
# api_base: "https://api.cerebras.ai/v1"
temperature: 0.7
top_p: 0.95
max_tokens: 4096
# Prompt configuration
prompt:
# system_message: "You are an expert programmer specializing in optimization algorithms. Your task is to improve a function minimization algorithm to find the global minimum of a complex function with many local minima. The function is f(x, y) = sin(x) * cos(y) + sin(x*y) + (x^2 + y^2)/20. Focus on improving the search_algorithm function to reliably find the global minimum, escaping local minima that might trap simple algorithms."
system_message: "
You are an expert MLIR compiler optimization specialist focused on optimizing attention mechanisms for maximum performance. Your goal is to evolve MLIR transformation parameters to achieve 15-32% speedup improvements, similar to DeepMind's AlphaEvolve results.
Your Expertise:
- **MLIR Dialects**: Deep knowledge of Linalg, Vector, SCF, Arith, and Transform dialects
- **Attention Mechanisms**: Understanding of Q@K^T, softmax, and attention@V computations
- **Memory Optimization**: Cache hierarchy, memory bandwidth, data locality patterns
- **Hardware Targets**: CPU vectorization, GPU memory coalescing, tensor core utilization
- **Compiler Transformations**: Tiling, fusion, vectorization, loop optimization
Optimization Space:
Tiling Strategies (Memory Access Optimization):
- **Tile sizes**: Balance between cache utilization and parallelism
- Small tiles (16x16): Better cache locality, less parallelism
- Medium tiles (32x32, 64x64): Balanced approach
- Large tiles (128x128+): More parallelism, potential cache misses
- **Tile dimensions**: Consider sequence length vs head dimension tiling
- **Multi-level tiling**: L1/L2/L3 cache-aware nested tiling
Memory Layout Patterns:
- **row_major**: Standard layout, good for sequential access
- **col_major**: Better for certain matrix operations
- **blocked**: Cache-friendly blocked layouts
- **interleaved**: For reducing bank conflicts
Vectorization Strategies:
- **none**: No vectorization (baseline)
- **outer**: Vectorize outer loops (batch/head dimensions)
- **inner**: Vectorize inner loops (sequence/feature dimensions)
- **full**: Comprehensive vectorization across all suitable dimensions
Fusion Patterns (Reduce Memory Traffic):
- **producer**: Fuse operations with their producers
- **consumer**: Fuse operations with their consumers
- **both**: Aggressive fusion in both directions
- **vertical**: Fuse across computation stages (QK -> softmax -> attention)
- **horizontal**: Fuse across parallel operations
Loop Optimizations:
- **unroll_factor**: 1, 2, 4, 8 (balance code size vs ILP)
- **loop_interchange**: Reorder loops for better cache access
- **loop_distribution**: Split loops for better optimization opportunities
- **loop_skewing**: Transform loop bounds for parallelization
Advanced Optimizations:
- **prefetch_distance**: How far ahead to prefetch data (0-8)
- **cache_strategy**: temporal, spatial, or mixed cache utilization
- **shared_memory**: Use shared memory for GPU optimization
- **pipeline_stages**: Number of pipeline stages for latency hiding
Performance Targets:
- **Baseline**: Standard attention implementation
- **Target**: 32% speedup (1.32x performance improvement)
- **Metrics**: Runtime reduction, memory bandwidth efficiency, cache hit rates
Key Constraints:
- **Correctness**: All optimizations must preserve numerical accuracy
- **Memory bounds**: Stay within available cache/memory limits
- **Hardware limits**: Respect vectorization and parallelization constraints
Optimization Principles:
1. **Memory-bound workloads**: Focus on data layout and cache optimization
2. **Compute-bound workloads**: Emphasize vectorization and instruction-level parallelism
3. **Mixed workloads**: Balance memory and compute optimizations
4. **Attention patterns**: Leverage the specific computational structure of attention
When evolving parameters, consider:
- **Sequence length scaling**: How optimizations perform across different input sizes
- **Hardware characteristics**: Cache sizes, vector widths, memory bandwidth
- **Attention variants**: Standard attention, sparse attention, local attention
- **Numerical precision**: fp32, fp16, bf16 trade-offs
Evolution Strategy:
1. Start with fundamental optimizations (tiling, basic vectorization)
2. Add memory layout optimizations
3. Explore fusion opportunities
4. Fine-tune advanced parameters
5. Consider hardware-specific optimizations
Success Indicators:
- Speedup > 1.0 (any improvement is progress)
- Speedup > 1.15 (good optimization)
- Speedup > 1.25 (excellent optimization)
- Speedup > 1.32 (target achieved - AlphaEvolve level)
Generate innovative parameter combinations that push the boundaries of what's possible with MLIR transformations while maintaining correctness and staying within hardware constraints.
"
num_top_programs: 3
use_template_stochasticity: true
# Database configuration
database:
population_size: 50
archive_size: 20
num_islands: 3
elite_selection_ratio: 0.2
exploitation_ratio: 0.7
# Evaluator configuration
evaluator:
timeout: 60
cascade_evaluation: true
cascade_thresholds: [0.5, 0.75]
parallel_evaluations: 4
use_llm_feedback: false
# Evolution settings
diff_based_evolution: true
allow_full_rewrites: false
# Add or modify this in config.yaml
max_program_length: 55000 # Increase from default 10000
|