KernelAgent/examples/configs/nvidia.yaml at d32198ecba4e8d7bb75f2c2a48c1530f3ea0681d · meta-pytorch/KernelAgent · GitHub

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
# Explicit NVIDIA platform config — all 8 component slots spelled out.
#
# This is equivalent to the default (platform is nvidia when omitted),
# but serves as documentation of the available component keys and as a
# starting point for mixed-platform configs.
#
# Usage:
#   python examples/run_opt_manager.py \
#       --kernel-dir examples/optimize_01_matvec \
#       --config examples/configs/nvidia.yaml

strategy: beam_search
num_workers: 4
strategy_config:
  num_top_kernels: 2
  num_bottlenecks: 2
openai_model: gpt-5
high_reasoning_effort: true

# Worker configuration
benchmark_warmup: 25
benchmark_repeat: 100
divergence_threshold: 50.0
target_platform: cuda
gpu_name: "NVIDIA H100 NVL 94GB"

platform:
  # Manager-level components
  verifier: nvidia
  benchmarker: nvidia
  worker_runner: nvidia
  # Worker-level components
  specs_provider: nvidia
  profiler: nvidia
  roofline_analyzer: nvidia
  bottleneck_analyzer: nvidia
  rag_prescriber: nvidia