-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathconfig.yaml
More file actions
50 lines (43 loc) · 1.34 KB
/
Copy pathconfig.yaml
File metadata and controls
50 lines (43 loc) · 1.34 KB
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
# KernelGen Configuration File
# 统一配置文件,控制所有生成和评估参数
# 数据集配置
dataset:
source: "huggingface" # huggingface 或 local
name: "ScalingIntelligence/KernelBench"
level: 2 # 1, 2, 3, 4
problem_ids: [40, 41, 42] # 指定问题ID,空列表表示所有问题
# 生成配置
generation:
max_iterations: 20 # 最大迭代次数
early_stop_threshold: 1.2 # 加速比阈值,达到后提前停止
min_successful_iterations: 3 # 最少成功迭代次数
backend: "triton" # triton, cuda
# LLM配置
llm:
server_type: "deepseek"
model_name: "deepseek-coder"
temperature: 0.0
max_tokens: 4096
# 性能测试配置
performance:
device: "cuda"
warmup_runs: 10
benchmark_runs: 100
timeout: 180 # 超时时间(秒)
# 输出配置
output:
base_dir: "runs" # 基础输出目录
run_name: "kernelgen_run" # 运行名称
save_all_kernels: true # 保存所有生成的kernel
save_best_only: false # 仅保存最佳kernel
save_logs: true
save_performance_data: true
# 评估配置
evaluation:
num_correct_trials: 5 # 正确性测试次数
num_perf_trials: 100 # 性能测试次数
pass_at_k_values: [1, 5, 10] # pass@k计算的k值
# 并发配置
concurrency:
num_workers: 4 # 并发worker数量
api_query_interval: 0.1 # API查询间隔(秒)