LLM-Low-Level-code-optimization/config.yaml at tvm-bench · On-Point-RND/LLM-Low-Level-code-optimization · 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
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
save_onnx: true
save_relax_ir: true
save_relax_json: false
save_tir_ir: true
save_tir_json: false
profile_before_optimization: true
profile_after_relax: true
profile_after_tir: true

skip_prof_report: true


profile:
  warmup_iters: 2
  number: 3
  repeat: 3
  baseline_multiplier: 5  # abort full profiling if quick run > n * baseline_mean_ms
  verbose: true

check_correctness: true

transform_tir_mode: llm_transform_tir

meta_schedule_max_trials_global: 256

meta_schedule_builder_workers: -1

# Number of iterative LLM TIR optimization steps (used when transform_tir_mode: llm_transform_tir)
llm_tir_max_steps: 5

agents:
  tir_code:
    api_key: "EMPTY"
    base_url: "http://localhost:8052/v1"
    model: "Qwen3-Coder-Next"
    temperature: 0.1
    max_tokens: null
    top_p: null
    top_k: null
    timeout: 1000.0

  tir_analysis:
    api_key: "EMPTY"
    base_url: "http://localhost:8052/v1"
    model: "Qwen3-Coder-Next"
    temperature: 0.3
    max_tokens: null
    top_p: null
    top_k: null
    timeout: 1000.0

  tir_summary:
    api_key: "EMPTY"
    base_url: "http://localhost:8052/v1"
    model: "Qwen3-Coder-Next"
    temperature: 0.2
    max_tokens: null
    top_p: null
    top_k: null
    timeout: 1000.0

relax_transforms:
  - "LegalizeOps"           # перевести ops во внутренние примитивы TVM
  - "FoldConstant"          # свернуть константы
  - "CanonicalizeBindings"  # упростить биндинги
  - "Normalize"             # нормализация выражений
  - "FuseOps"               # слить блоки ops
  - "FuseTIR"               # вынести слитые ops в TIR
  - "DeadCodeElimination"   # выкинуть мёртвый код