hunt commited on
Commit ·
ed92f65
1
Parent(s): a9f16ea
Implement Task 8: job_orchestrator (confirm/cancel/progress state machine, 291/291 tests passing)
Browse files
miranda-nodes/src/forge/job_orchestrator.rs
ADDED
|
@@ -0,0 +1,297 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
//! Task 8 — Conversational deployment & progress reporting.
|
| 2 |
+
//!
|
| 3 |
+
//! Wires Job Parser output through confirmation, dispatch, and progress
|
| 4 |
+
//! query handling as real state-machine logic — testable without live
|
| 5 |
+
//! GPU/model infra, since what this task actually needs is correct state
|
| 6 |
+
//! transitions and query responses, not the training/merge work itself
|
| 7 |
+
//! (that's Tasks 6/7's disclosed-scope territory).
|
| 8 |
+
//!
|
| 9 |
+
//! Requirement 4.2/4.3/4.4: a Forge job goes through
|
| 10 |
+
//! `AwaitingConfirmation -> Running -> Completed/Failed`, and
|
| 11 |
+
//! `progress_summary` answers "how's the fine-tune going?" against
|
| 12 |
+
//! whatever state a running job is currently in.
|
| 13 |
+
|
| 14 |
+
use chrono::{DateTime, Utc};
|
| 15 |
+
use uuid::Uuid;
|
| 16 |
+
|
| 17 |
+
use crate::forge::finetune_pipeline::TrainingMetrics;
|
| 18 |
+
use crate::forge::gpu_provisioner::ProvisionError;
|
| 19 |
+
use crate::forge::job_parser::JobSpec;
|
| 20 |
+
use crate::forge::merge_pipeline::MergeError;
|
| 21 |
+
|
| 22 |
+
#[derive(Debug, Clone, PartialEq)]
|
| 23 |
+
pub enum JobStatus {
|
| 24 |
+
/// Requirement 5.2 — job is parsed but blocked on the user
|
| 25 |
+
/// confirming a cost/action that exceeded their autonomy threshold.
|
| 26 |
+
AwaitingConfirmation,
|
| 27 |
+
Running { started_at: DateTime<Utc> },
|
| 28 |
+
Completed { finished_at: DateTime<Utc>, summary: String },
|
| 29 |
+
Failed { failed_at: DateTime<Utc>, reason: String },
|
| 30 |
+
Cancelled { cancelled_at: DateTime<Utc> },
|
| 31 |
+
}
|
| 32 |
+
|
| 33 |
+
#[derive(Debug, Clone)]
|
| 34 |
+
pub struct ForgeJob {
|
| 35 |
+
pub id: Uuid,
|
| 36 |
+
pub spec: JobSpec,
|
| 37 |
+
pub status: JobStatus,
|
| 38 |
+
pub created_at: DateTime<Utc>,
|
| 39 |
+
}
|
| 40 |
+
|
| 41 |
+
/// Tracks in-flight Forge jobs so progress queries and completion
|
| 42 |
+
/// announcements have somewhere real to read from, rather than each
|
| 43 |
+
/// caller needing to thread job state through by hand. In-memory here;
|
| 44 |
+
/// a real deployment would persist this the same way `ModelRegistry`
|
| 45 |
+
/// discloses its own in-memory-for-now posture.
|
| 46 |
+
#[derive(Debug, Default)]
|
| 47 |
+
pub struct JobOrchestrator {
|
| 48 |
+
jobs: Vec<ForgeJob>,
|
| 49 |
+
}
|
| 50 |
+
|
| 51 |
+
impl JobOrchestrator {
|
| 52 |
+
pub fn new() -> Self {
|
| 53 |
+
Self::default()
|
| 54 |
+
}
|
| 55 |
+
|
| 56 |
+
/// Requirement 4.2 — a freshly parsed job always starts
|
| 57 |
+
/// `AwaitingConfirmation`, even if the GPU provisioner would allow
|
| 58 |
+
/// it autonomously; the confirmation dialogue step (Requirement 4.2)
|
| 59 |
+
/// is a conversational turn that has to happen regardless, and
|
| 60 |
+
/// `confirm` is the only way to move a job out of this state.
|
| 61 |
+
pub fn submit(&mut self, spec: JobSpec, now: DateTime<Utc>) -> Uuid {
|
| 62 |
+
let id = Uuid::new_v4();
|
| 63 |
+
self.jobs.push(ForgeJob {
|
| 64 |
+
id,
|
| 65 |
+
spec,
|
| 66 |
+
status: JobStatus::AwaitingConfirmation,
|
| 67 |
+
created_at: now,
|
| 68 |
+
});
|
| 69 |
+
id
|
| 70 |
+
}
|
| 71 |
+
|
| 72 |
+
pub fn confirm(&mut self, id: Uuid, now: DateTime<Utc>) -> Result<(), OrchestratorError> {
|
| 73 |
+
let job = self.job_mut(id)?;
|
| 74 |
+
if job.status != JobStatus::AwaitingConfirmation {
|
| 75 |
+
return Err(OrchestratorError::InvalidTransition {
|
| 76 |
+
id,
|
| 77 |
+
from: format!("{:?}", job.status),
|
| 78 |
+
to: "Running".to_string(),
|
| 79 |
+
});
|
| 80 |
+
}
|
| 81 |
+
job.status = JobStatus::Running { started_at: now };
|
| 82 |
+
Ok(())
|
| 83 |
+
}
|
| 84 |
+
|
| 85 |
+
pub fn cancel(&mut self, id: Uuid, now: DateTime<Utc>) -> Result<(), OrchestratorError> {
|
| 86 |
+
let job = self.job_mut(id)?;
|
| 87 |
+
match job.status {
|
| 88 |
+
JobStatus::Completed { .. } | JobStatus::Failed { .. } | JobStatus::Cancelled { .. } => {
|
| 89 |
+
Err(OrchestratorError::InvalidTransition {
|
| 90 |
+
id,
|
| 91 |
+
from: format!("{:?}", job.status),
|
| 92 |
+
to: "Cancelled".to_string(),
|
| 93 |
+
})
|
| 94 |
+
}
|
| 95 |
+
_ => {
|
| 96 |
+
job.status = JobStatus::Cancelled { cancelled_at: now };
|
| 97 |
+
Ok(())
|
| 98 |
+
}
|
| 99 |
+
}
|
| 100 |
+
}
|
| 101 |
+
|
| 102 |
+
pub fn complete_finetune(&mut self, id: Uuid, metrics: TrainingMetrics, now: DateTime<Utc>) -> Result<(), OrchestratorError> {
|
| 103 |
+
let job = self.job_mut(id)?;
|
| 104 |
+
job.status = JobStatus::Completed {
|
| 105 |
+
finished_at: now,
|
| 106 |
+
summary: format!(
|
| 107 |
+
"fine-tune finished: final loss {:.3}, {:.1} GPU-hours",
|
| 108 |
+
metrics.final_loss, metrics.gpu_hours
|
| 109 |
+
),
|
| 110 |
+
};
|
| 111 |
+
Ok(())
|
| 112 |
+
}
|
| 113 |
+
|
| 114 |
+
pub fn fail(&mut self, id: Uuid, reason: impl Into<String>, now: DateTime<Utc>) -> Result<(), OrchestratorError> {
|
| 115 |
+
let job = self.job_mut(id)?;
|
| 116 |
+
job.status = JobStatus::Failed { failed_at: now, reason: reason.into() };
|
| 117 |
+
Ok(())
|
| 118 |
+
}
|
| 119 |
+
|
| 120 |
+
/// Requirement 4.3 — answers "how's the fine-tune going?" against
|
| 121 |
+
/// whatever state the job is actually in, in conversational form.
|
| 122 |
+
pub fn progress_summary(&self, id: Uuid) -> Result<String, OrchestratorError> {
|
| 123 |
+
let job = self.job(id)?;
|
| 124 |
+
Ok(match &job.status {
|
| 125 |
+
JobStatus::AwaitingConfirmation => {
|
| 126 |
+
"That job is still waiting on your confirmation before it starts.".to_string()
|
| 127 |
+
}
|
| 128 |
+
JobStatus::Running { started_at } => {
|
| 129 |
+
let elapsed = Utc::now() - *started_at;
|
| 130 |
+
format!("Still running — it's been going for about {} minutes so far.", elapsed.num_minutes().max(0))
|
| 131 |
+
}
|
| 132 |
+
JobStatus::Completed { summary, .. } => summary.clone(),
|
| 133 |
+
JobStatus::Failed { reason, .. } => format!("That job failed: {reason}"),
|
| 134 |
+
JobStatus::Cancelled { .. } => "That job was cancelled.".to_string(),
|
| 135 |
+
})
|
| 136 |
+
}
|
| 137 |
+
|
| 138 |
+
pub fn get(&self, id: Uuid) -> Result<&ForgeJob, OrchestratorError> {
|
| 139 |
+
self.job(id)
|
| 140 |
+
}
|
| 141 |
+
|
| 142 |
+
fn job(&self, id: Uuid) -> Result<&ForgeJob, OrchestratorError> {
|
| 143 |
+
self.jobs.iter().find(|j| j.id == id).ok_or(OrchestratorError::NotFound(id))
|
| 144 |
+
}
|
| 145 |
+
|
| 146 |
+
fn job_mut(&mut self, id: Uuid) -> Result<&mut ForgeJob, OrchestratorError> {
|
| 147 |
+
self.jobs.iter_mut().find(|j| j.id == id).ok_or(OrchestratorError::NotFound(id))
|
| 148 |
+
}
|
| 149 |
+
}
|
| 150 |
+
|
| 151 |
+
#[derive(Debug, Clone, PartialEq, thiserror::Error)]
|
| 152 |
+
pub enum OrchestratorError {
|
| 153 |
+
#[error("no job found with id {0}")]
|
| 154 |
+
NotFound(Uuid),
|
| 155 |
+
#[error("job {id} cannot transition from {from} to {to}")]
|
| 156 |
+
InvalidTransition { id: Uuid, from: String, to: String },
|
| 157 |
+
}
|
| 158 |
+
|
| 159 |
+
/// Converts a provisioning/merge failure into a user-facing failure
|
| 160 |
+
/// reason string, per design.md's error handling: "the raw error is
|
| 161 |
+
/// surfaced to the user in conversational form rather than swallowed."
|
| 162 |
+
pub fn describe_provision_failure(err: &ProvisionError) -> String {
|
| 163 |
+
err.to_string()
|
| 164 |
+
}
|
| 165 |
+
|
| 166 |
+
pub fn describe_merge_failure(err: &MergeError) -> String {
|
| 167 |
+
err.to_string()
|
| 168 |
+
}
|
| 169 |
+
|
| 170 |
+
#[cfg(test)]
|
| 171 |
+
mod tests {
|
| 172 |
+
use super::*;
|
| 173 |
+
use crate::forge::job_parser::{JobMethod, ModelRef};
|
| 174 |
+
use std::path::PathBuf;
|
| 175 |
+
use std::time::Duration;
|
| 176 |
+
|
| 177 |
+
fn sample_spec() -> JobSpec {
|
| 178 |
+
JobSpec {
|
| 179 |
+
method: JobMethod::LoraFineTune,
|
| 180 |
+
base_models: vec![ModelRef {
|
| 181 |
+
name: "GLM-9B".to_string(),
|
| 182 |
+
family: "GLM".to_string(),
|
| 183 |
+
size: "9B".to_string(),
|
| 184 |
+
local_path: Some(PathBuf::from("/models/glm-9b")),
|
| 185 |
+
hf_repo: None,
|
| 186 |
+
}],
|
| 187 |
+
instructions: None,
|
| 188 |
+
merge_method: None,
|
| 189 |
+
estimated_cost: 1.5,
|
| 190 |
+
estimated_duration: Duration::from_secs(3600),
|
| 191 |
+
}
|
| 192 |
+
}
|
| 193 |
+
|
| 194 |
+
#[test]
|
| 195 |
+
fn submitted_job_starts_awaiting_confirmation() {
|
| 196 |
+
let mut orch = JobOrchestrator::new();
|
| 197 |
+
let id = orch.submit(sample_spec(), Utc::now());
|
| 198 |
+
assert_eq!(orch.get(id).unwrap().status, JobStatus::AwaitingConfirmation);
|
| 199 |
+
}
|
| 200 |
+
|
| 201 |
+
#[test]
|
| 202 |
+
fn confirm_transitions_to_running() {
|
| 203 |
+
let mut orch = JobOrchestrator::new();
|
| 204 |
+
let id = orch.submit(sample_spec(), Utc::now());
|
| 205 |
+
orch.confirm(id, Utc::now()).unwrap();
|
| 206 |
+
assert!(matches!(orch.get(id).unwrap().status, JobStatus::Running { .. }));
|
| 207 |
+
}
|
| 208 |
+
|
| 209 |
+
#[test]
|
| 210 |
+
fn cannot_confirm_a_job_twice() {
|
| 211 |
+
let mut orch = JobOrchestrator::new();
|
| 212 |
+
let id = orch.submit(sample_spec(), Utc::now());
|
| 213 |
+
orch.confirm(id, Utc::now()).unwrap();
|
| 214 |
+
let err = orch.confirm(id, Utc::now()).unwrap_err();
|
| 215 |
+
assert!(matches!(err, OrchestratorError::InvalidTransition { .. }));
|
| 216 |
+
}
|
| 217 |
+
|
| 218 |
+
#[test]
|
| 219 |
+
fn progress_summary_before_confirmation_mentions_awaiting_confirmation() {
|
| 220 |
+
let mut orch = JobOrchestrator::new();
|
| 221 |
+
let id = orch.submit(sample_spec(), Utc::now());
|
| 222 |
+
let summary = orch.progress_summary(id).unwrap();
|
| 223 |
+
assert!(summary.to_lowercase().contains("confirmation"));
|
| 224 |
+
}
|
| 225 |
+
|
| 226 |
+
#[test]
|
| 227 |
+
fn progress_summary_while_running_mentions_running() {
|
| 228 |
+
let mut orch = JobOrchestrator::new();
|
| 229 |
+
let id = orch.submit(sample_spec(), Utc::now());
|
| 230 |
+
orch.confirm(id, Utc::now()).unwrap();
|
| 231 |
+
let summary = orch.progress_summary(id).unwrap();
|
| 232 |
+
assert!(summary.to_lowercase().contains("running"));
|
| 233 |
+
}
|
| 234 |
+
|
| 235 |
+
#[test]
|
| 236 |
+
fn complete_finetune_produces_a_summary_with_metrics() {
|
| 237 |
+
let mut orch = JobOrchestrator::new();
|
| 238 |
+
let id = orch.submit(sample_spec(), Utc::now());
|
| 239 |
+
orch.confirm(id, Utc::now()).unwrap();
|
| 240 |
+
let metrics = TrainingMetrics { final_loss: 0.31, duration: Duration::from_secs(3600), gpu_hours: 1.0 };
|
| 241 |
+
orch.complete_finetune(id, metrics, Utc::now()).unwrap();
|
| 242 |
+
|
| 243 |
+
let summary = orch.progress_summary(id).unwrap();
|
| 244 |
+
assert!(summary.contains("0.31") || summary.contains("0.310"));
|
| 245 |
+
}
|
| 246 |
+
|
| 247 |
+
#[test]
|
| 248 |
+
fn fail_produces_a_summary_containing_the_reason() {
|
| 249 |
+
let mut orch = JobOrchestrator::new();
|
| 250 |
+
let id = orch.submit(sample_spec(), Utc::now());
|
| 251 |
+
orch.confirm(id, Utc::now()).unwrap();
|
| 252 |
+
orch.fail(id, "GPU quota exceeded", Utc::now()).unwrap();
|
| 253 |
+
let summary = orch.progress_summary(id).unwrap();
|
| 254 |
+
assert!(summary.contains("GPU quota exceeded"));
|
| 255 |
+
}
|
| 256 |
+
|
| 257 |
+
#[test]
|
| 258 |
+
fn cancel_from_awaiting_confirmation_succeeds() {
|
| 259 |
+
let mut orch = JobOrchestrator::new();
|
| 260 |
+
let id = orch.submit(sample_spec(), Utc::now());
|
| 261 |
+
orch.cancel(id, Utc::now()).unwrap();
|
| 262 |
+
assert!(matches!(orch.get(id).unwrap().status, JobStatus::Cancelled { .. }));
|
| 263 |
+
}
|
| 264 |
+
|
| 265 |
+
#[test]
|
| 266 |
+
fn cannot_cancel_an_already_completed_job() {
|
| 267 |
+
let mut orch = JobOrchestrator::new();
|
| 268 |
+
let id = orch.submit(sample_spec(), Utc::now());
|
| 269 |
+
orch.confirm(id, Utc::now()).unwrap();
|
| 270 |
+
let metrics = TrainingMetrics { final_loss: 0.2, duration: Duration::from_secs(60), gpu_hours: 0.1 };
|
| 271 |
+
orch.complete_finetune(id, metrics, Utc::now()).unwrap();
|
| 272 |
+
let err = orch.cancel(id, Utc::now()).unwrap_err();
|
| 273 |
+
assert!(matches!(err, OrchestratorError::InvalidTransition { .. }));
|
| 274 |
+
}
|
| 275 |
+
|
| 276 |
+
#[test]
|
| 277 |
+
fn querying_an_unknown_job_id_returns_not_found() {
|
| 278 |
+
let orch = JobOrchestrator::new();
|
| 279 |
+
let err = orch.progress_summary(Uuid::new_v4()).unwrap_err();
|
| 280 |
+
assert!(matches!(err, OrchestratorError::NotFound(_)));
|
| 281 |
+
}
|
| 282 |
+
|
| 283 |
+
#[test]
|
| 284 |
+
fn describe_provision_failure_surfaces_the_raw_reason() {
|
| 285 |
+
let err = ProvisionError::RequiresConfirmation { estimated: 10.0, threshold: 5.0 };
|
| 286 |
+
let desc = describe_provision_failure(&err);
|
| 287 |
+
assert!(desc.contains("10.00") || desc.contains("10"));
|
| 288 |
+
assert!(desc.contains("5.00") || desc.contains("5"));
|
| 289 |
+
}
|
| 290 |
+
|
| 291 |
+
#[test]
|
| 292 |
+
fn describe_merge_failure_surfaces_the_raw_reason() {
|
| 293 |
+
let err = MergeError::MergeKitFailed { reason: "tensor mismatch".to_string() };
|
| 294 |
+
let desc = describe_merge_failure(&err);
|
| 295 |
+
assert!(desc.contains("tensor mismatch"));
|
| 296 |
+
}
|
| 297 |
+
}
|
miranda-nodes/src/forge/mod.rs
CHANGED
|
@@ -16,6 +16,7 @@ pub mod compatibility;
|
|
| 16 |
pub mod finetune_pipeline;
|
| 17 |
pub mod gpu_provisioner;
|
| 18 |
pub mod job_parser;
|
|
|
|
| 19 |
pub mod merge_pipeline;
|
| 20 |
pub mod model_registry;
|
| 21 |
pub mod naming;
|
|
|
|
| 16 |
pub mod finetune_pipeline;
|
| 17 |
pub mod gpu_provisioner;
|
| 18 |
pub mod job_parser;
|
| 19 |
+
pub mod job_orchestrator;
|
| 20 |
pub mod merge_pipeline;
|
| 21 |
pub mod model_registry;
|
| 22 |
pub mod naming;
|