clawft-kernel 0.8.3

WeftOS kernel: boot sequence, process table, service registry, and health subsystem
Documentation 
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
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178

//! Roster import and the residency ledger through the managed adapter.
//! The roster is a fixture copy in this repo; the real `~/llm` is never
//! read, and the launcher is a fake script.

use std::sync::Arc;
use std::time::Duration;

use super::config::RestartPolicy;
use super::fakes::*;
use super::residency::{GB, ResidencyLedger};
use super::roster::{RosterOverlay, flavor_for_backend, import_roster};
use super::spec::{InferFlavor, LatencyClass};
use crate::chain::ChainManager;
use crate::model_manifest::ModelFormat;
use crate::workload_governance::{NodeTrustTier, WorkloadGate};
use crate::workload_runtime::host::{RUNTIME_CHAIN_SOURCE, WorkloadHost};
use crate::workload_runtime::types::{RuntimeError, WorkloadRuntime};

const YAML: &str = include_str!("fixtures/model-lab-roster.yaml");

fn overlay() -> RosterOverlay {
    serde_yaml::from_str(
        r#"
excludes:
  planner: ["role:coder-daily"]
latency_class:
  coder-small: interactive
sticky:
  swarm-micro: false
"#,
    )
    .unwrap()
}

#[test]
fn imports_memory_ports_flavors_and_the_overlay() {
    let r = import_roster(YAML, &overlay()).unwrap();
    let ids: Vec<_> = r.specs.iter().map(|s| s.spec.role.as_str()).collect();
    assert_eq!(
        ids,
        ["coder-daily", "coder-compact", "coder-small", "planner", "swarm-worker", "swarm-micro", "voice-llm"]
    );
    let daily = r.get("coder-daily").unwrap();
    assert_eq!(daily.spec.runtime, InferFlavor::MlxLm);
    assert_eq!(daily.spec.serve.port, Some(8081));
    assert_eq!(daily.spec.model.as_deref(), Some("Qwen3-Coder-Next-80B-A3B"));
    assert_eq!(daily.spec.memory.weights_bytes, 55 * GB);
    assert!(daily.spec.sticky, "inference is sticky by default (ADR-101 section 6)");
    assert_eq!(daily.spec.latency_class, LatencyClass::Batch);
    assert_eq!(daily.status, "live");
    assert_eq!(r.get("coder-small").unwrap().spec.memory.weights_bytes, 9_700_000_000);
    assert_eq!(r.get("coder-compact").unwrap().spec.runtime, InferFlavor::Ollama);
    // Overlay: excludes, latency, sticky.
    assert_eq!(r.get("planner").unwrap().spec.excludes, ["role:coder-daily"]);
    assert_eq!(r.get("coder-small").unwrap().spec.latency_class, LatencyClass::Interactive);
    assert!(!r.get("swarm-micro").unwrap().spec.sticky);
    // The role text can say voice.
    assert_eq!(r.get("voice-llm").unwrap().spec.latency_class, LatencyClass::Interactive);
    // "a or b" backends take the first.
    assert_eq!(r.get("swarm-worker").unwrap().spec.runtime, InferFlavor::Ollama);
    assert_eq!(r.get("swarm-micro").unwrap().spec.runtime, InferFlavor::MlxLm);
    // Every imported spec validates.
    assert!(r.specs.iter().all(|s| s.spec.validate().is_ok()));
}

#[test]
fn what_is_not_imported_is_reported_with_a_reason() {
    let r = import_roster(YAML, &RosterOverlay::default()).unwrap();
    let why = |id: &str| r.skipped.iter().find(|(i, _)| i == id).map(|(_, w)| w.clone()).unwrap_or_default();
    assert!(why("coder-next").contains("alias of coder-daily"), "{}", why("coder-next"));
    assert!(why("vlm").contains("no inference adapter"));
    assert!(why("tts").contains("no inference adapter"));
    assert!(why("broken-mem").contains("ram_gb"));
    assert!(why("bad id!").contains("plain token"));
    assert_eq!(r.specs.len() + r.skipped.len(), 12, "nothing vanishes");
}

#[test]
fn bad_input_is_refused_at_the_boundary() {
    assert!(import_roster("roster: [", &RosterOverlay::default()).is_err());
    assert!(import_roster(&"x".repeat(2 * 1024 * 1024), &RosterOverlay::default()).is_err());
    let many: String = std::iter::once("roster:\n".to_string())
        .chain((0..300).map(|i| format!("- id: r{i}\n  backend: mlx_lm\n  port: 9000\n  ram_gb: 1\n  repo: a{i}\n")))
        .collect();
    assert!(import_roster(&many, &RosterOverlay::default()).is_err());
    // An overlay with an unknown key is refused.
    assert!(serde_yaml::from_str::<RosterOverlay>("bogus: 1").is_err());
    // No roster key at all is an empty import, not an error.
    assert_eq!(import_roster("meta: {}", &RosterOverlay::default()).unwrap().specs.len(), 0);
    assert_eq!(flavor_for_backend("llama.cpp"), Some(InferFlavor::LlamaCpp));
    assert_eq!(flavor_for_backend(""), None);
}

fn fast() -> RestartPolicy {
    RestartPolicy { max_restarts: 1, base: Duration::from_millis(300), cap: Duration::from_secs(1), ..RestartPolicy::default() }
}

fn managed_with_ledger(ledger: &Arc<ResidencyLedger>) -> Managed {
    let l = ledger.clone();
    let m = managed_with(InferFlavor::LlamaCpp, fast(), move |c| c.ledger = Some(l));
    for n in ["Model-A", "Model-B"] {
        adopt_model(&m.reg, m.tmp.path(), n, ModelFormat::Gguf);
    }
    m
}

fn spec_of(role: &str, model: &str, gb: u64, excludes: &[&str]) -> super::InferenceSpec {
    let mut s = spec(role, InferFlavor::LlamaCpp, free_port());
    s.model = Some(model.into());
    s.memory.weights_bytes = gb * GB;
    s.excludes = excludes.iter().map(|x| x.to_string()).collect();
    s
}

#[tokio::test]
async fn the_adapter_refuses_a_start_over_budget_and_frees_it_on_stop() {
    let ledger = Arc::new(ResidencyLedger::new(Some(10 * GB)));
    let (m1, m2) = (managed_with_ledger(&ledger), managed_with_ledger(&ledger));
    let a = load(&m1.rt, spec_of("coder", "Model-A", 6, &[])).await;
    let b = load(&m2.rt, spec_of("planner", "Model-B", 6, &[])).await;
    m1.rt.start(&a).await.unwrap();
    let e = m2.rt.start(&b).await.unwrap_err();
    assert!(matches!(e, RuntimeError::AdmissionRefused(_)), "{e:?}");
    let text = e.to_string();
    assert!(text.contains("unplaceable") && text.contains("'planner' needs 6.0 GB") && text.contains("coder 6.0 GB"), "{text}");
    // Nothing was launched for the refused one.
    assert!(!m2.script_dir.join("pid.txt").exists());
    assert_eq!(ledger.snapshot(), [("coder".to_string(), 6 * GB)]);
    // Stopping the holder gives the memory back.
    m1.rt.stop(&a, Duration::from_secs(3)).await.unwrap();
    assert_eq!(ledger.used(), 0);
    m2.rt.start(&b).await.unwrap();
    // Unload releases too, and a second start of a running instance does not double count.
    assert!(m2.rt.start(&b).await.is_err(), "already running");
    assert_eq!(ledger.used(), 6 * GB, "the failed second start kept the first reservation");
    m2.rt.unload(b).await.unwrap();
    assert_eq!(ledger.used(), 0);
    m1.rt.unload(a).await.unwrap();
}

#[tokio::test]
async fn co_residency_exclusions_hold_across_adapters() {
    let ledger = Arc::new(ResidencyLedger::new(None));
    let (m1, m2) = (managed_with_ledger(&ledger), managed_with_ledger(&ledger));
    let a = load(&m1.rt, spec_of("coder-daily", "Model-A", 1, &[])).await;
    let b = load(&m2.rt, spec_of("planner", "Model-B", 1, &["role:coder-daily"])).await;
    m1.rt.start(&a).await.unwrap();
    let e = m2.rt.start(&b).await.unwrap_err().to_string();
    assert!(e.contains("cannot be resident beside 'coder-daily'") && e.contains("'planner' excludes it"), "{e}");
    m1.rt.unload(a).await.unwrap();
    m2.rt.start(&b).await.unwrap();
    m2.rt.unload(b).await.unwrap();
}

#[tokio::test]
async fn through_the_host_an_unplaceable_start_is_chained_with_its_reason() {
    let ledger = Arc::new(ResidencyLedger::new(Some(5 * GB)));
    let m = managed_with_ledger(&ledger);
    let chain = Arc::new(ChainManager::new(0, 1000));
    let gate = WorkloadGate::exempt(0.8, false, "test")
        .with_chain(chain.clone())
        .with_permit(infer_permit())
        .unwrap();
    let host = WorkloadHost::new(m.rt.clone(), Arc::new(gate), "operator", NodeTrustTier::Paired)
        .with_chain(chain.clone());
    let h = host.load(&workload(spec_of("big", "Model-A", 8, &[])), &wl_cfg()).await.unwrap();
    let e = host.start(&h).await.unwrap_err();
    assert!(e.to_string().contains("unplaceable"), "{e}");
    let refused: Vec<_> = chain
        .tail(0)
        .into_iter()
        .filter(|e| e.source == RUNTIME_CHAIN_SOURCE && e.kind == "workload.refuse")
        .collect();
    assert_eq!(refused.len(), 1);
    let p = refused[0].payload.clone().unwrap().to_string();
    assert!(p.contains("unplaceable") && p.contains("8.0 GB") && p.contains("5.0 GB"), "{p}");
    host.unload(h).await.unwrap();
}