max / everycycle

3.2 KB · 94 lines History Blame Raw

1	//! Daemon → client events.
2
3	use crate::request::RequestId;
4
5	/// One event in the daemon's response stream for some request.
6	///
7	/// Streams are unordered across requests but ordered within a
8	/// single `RequestId`. Every terminal event sets some semantic
9	/// equivalent of "finished"; clients should not assume the stream
10	/// continues after a terminal event for a given request id.
11	#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
12	pub enum ServerEvent {
13	/// The supervisor has admitted the request and queued it.
14	/// `queue_position` is zero when work begins immediately.
15	Accepted {
16	request_id: RequestId,
17	queue_position: u32,
18	},
19	/// One chunk of generated tokens. Terminal when `finished`.
20	Tokens {
21	request_id: RequestId,
22	chunk: TokenChunk,
23	finished: bool,
24	},
25	/// Response to a `DescribeFleet` request. Terminal for that
26	/// request.
27	FleetDescription(FleetSnapshot),
28	/// Terminal error for a request.
29	Error {
30	request_id: RequestId,
31	kind: ErrorKind,
32	message: String,
33	},
34	}
35
36	/// Generated tokens plus the rendered text. The text is what most
37	/// clients display; the IDs let token-counting clients (rate
38	/// limiters, billing surfaces) operate without re-tokenizing.
39	#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
40	pub struct TokenChunk {
41	/// UTF-8 text the tokenizer produced for these IDs.
42	pub text: String,
43	/// Token IDs in this chunk, in generation order. Opaque to the
44	/// client; the value space is the model's vocabulary.
45	pub ids: Vec<u32>,
46	}
47
48	/// Snapshot of the fleet as the daemon sees it right now.
49	/// Deliberately minimal at A1; expanded as Thread D (audit) and
50	/// Thread G (operator UX) grow.
51	#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
52	pub struct FleetSnapshot {
53	pub hostname: String,
54	pub executors: Vec<ExecutorSummary>,
55	pub loaded_models: Vec<LoadedModel>,
56	pub active_requests: u32,
57	}
58
59	/// One executor's externally visible summary.
60	#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
61	pub struct ExecutorSummary {
62	pub name: String,
63	pub vendor: String,
64	pub device_count: u32,
65	}
66
67	/// A model the daemon has resident.
68	#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
69	pub struct LoadedModel {
70	pub name: String,
71	pub revision: Option<String>,
72	pub assigned_executor: String,
73	}
74
75	/// Closed enum of error kinds at the API layer. Vendor-specific
76	/// detail goes in the accompanying `message`. Adding a variant is
77	/// semver-minor; renaming or removing is semver-major.
78	#[derive(Clone, Copy, Debug, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
79	pub enum ErrorKind {
80	/// Request payload failed validation (shape, range, dtype).
81	Invalid,
82	/// Referenced model is not loaded and the daemon will not
83	/// auto-load it for this request.
84	ModelNotLoaded,
85	/// Supervisor refused the request — out of capacity, over
86	/// quota, or rate-limited.
87	Rejected,
88	/// Request was cancelled, either by the client or by the
89	/// supervisor enforcing a deadline.
90	Cancelled,
91	/// Internal daemon error. `message` carries the detail.
92	Internal,
93	}
94