trainpulse/config.example.json at main · Thezone-1/trainpulse · 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
{
  "addr": "127.0.0.1:9876",
  "interval": "1s",
  "mode": "auto",
  "history_size": 120,
  "log_level": "info",
  "log_format": "json",
  "metrics_namespace": "trainpulse",
  "harvest": {
    "enabled": false,
    "allow_exec": false,
    "idle_utilization_pct": 5,
    "idle_window": "5m",
    "max_lease_ttl": "30m"
  },
  "rules": [
    {
      "name": "custom_low_mfu",
      "field": "training.mfu",
      "operator": "lt",
      "value": 0.35,
      "severity": "warning",
      "score_impact": 12,
      "description": "MFU is below the team-defined efficiency target"
    },
    {
      "name": "custom_high_allreduce",
      "field": "training.all_reduce_wait_ms",
      "operator": "gt",
      "value": 75,
      "severity": "critical",
      "score_impact": 20,
      "description": "All-reduce wait exceeded the configured SLO"
    }
  ],
  "diagnoses": [
    {
      "root_cause": "dataloader_starvation",
      "when_signals": ["dataloader_starvation"],
      "confidence": 0.9,
      "explanation": "GPUs are idling on input; our storage tier is the usual culprit.",
      "actions": [
        "Increase dataloader workers",
        "Move the shard to the fast NVMe cache",
        "Enable prefetching or pinned memory"
      ]
    },
    {
      "root_cause": "team_efficiency_breach",
      "when_signals": ["custom_low_mfu"],
      "confidence": 0.7,
      "explanation": "MFU fell below the team-defined efficiency target set in rules.",
      "actions": [
        "Check kernel fusion and precision mode",
        "Review sequence packing efficiency",
        "Open an efficiency ticket for the run"
      ]
    }
  ]
}