File size: 1,569 Bytes
af83196 | 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 | # GPU Mode: Float16 Vector Addition
# Usage: uv run skydiscover-run initial_program.py evaluator.py -c config.yaml -s <strategy> -i 50
max_iterations: 100
checkpoint_interval: 10
log_level: INFO
llm:
models:
- name: "gpt-5"
weight: 1.0
api_base: https://api.openai.com/v1
temperature: 0.7
# top_p: 0.95 # omitted by default; some providers (e.g. Anthropic) reject both temperature and top_p
max_tokens: 32000
timeout: 600
prompt:
system_message: |
You are an expert Triton kernel engineer. Output ONLY Python code - no explanations.
REQUIRED OUTPUT STRUCTURE:
1. Imports: torch, triton, triton.language as tl
2. @triton.jit kernel function(s)
3. def custom_kernel(data) wrapper - REQUIRED entry point
Task: Optimize float16 vector addition kernel. C = A + B
Input: Tuple of (A, B) tensors of shape (N, N) and dtype torch.float16
Output: Tensor of shape (N, N) and dtype torch.float16
N can be: 256, 512, 1024, 2048, 4096, 8192
Optimization tips:
- Block size tuning (512, 1024, 2048, 4096)
- Use @triton.autotune for automatic parameter tuning
- Vectorized loads for memory operations
- Grid configuration for occupancy
- Memory coalescing for sequential access patterns
MUST use @triton.jit decorator. MUST return float16 tensor.
Output complete, working code in a single ```python``` block.
evaluator:
timeout: 600
max_retries: 3
cascade_evaluation: true
cascade_thresholds: [0.4, 0.3]
diff_based_generation: true
max_solution_length: 60000
random_seed: 42
|