max / everycycle

3.1 KB · 102 lines History Blame Raw

1	//! Model identifiers, prompt shape, sampling.
2	//!
3	//! Deliberately small: the daemon's model registry is canonical;
4	//! clients refer to models by name and don't carry weights over the
5	//! wire. The format is chosen by the operator at load time, not by
6	//! the client.
7
8	/// Reference to a model the daemon has (or can) load.
9	#[derive(Clone, Debug, PartialEq, Eq, Hash, serde::Serialize, serde::Deserialize)]
10	pub struct ModelRef {
11	/// Operator-chosen name. Stable across reloads of the same
12	/// weights.
13	pub name: String,
14	/// Optional revision tag — lets a client pin a specific weights
15	/// version if the operator has multiple. `None` means "whatever
16	/// is current."
17	pub revision: Option<String>,
18	}
19
20	/// Prompt input to an inference request.
21	///
22	/// Token IDs are not on the wire at this layer — clients send text
23	/// or chat-message structure; the daemon owns the tokenizer for the
24	/// referenced model. Sending pre-tokenized input would couple
25	/// clients to the daemon's tokenizer choice, which is exactly the
26	/// kind of coupling the v1 freeze is meant to prevent.
27	#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
28	pub enum PromptSpec {
29	/// Single text completion.
30	Text(String),
31	/// Chat-style message list. The daemon applies the model's
32	/// chat template.
33	Chat(Vec<ChatMessage>),
34	}
35
36	/// One message in a chat-style prompt.
37	#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
38	pub struct ChatMessage {
39	pub role: ChatRole,
40	pub content: String,
41	}
42
43	/// Conversation role. Mirrors the OpenAI-style triad — the J1 shim
44	/// translates one-to-one without information loss.
45	#[derive(Clone, Copy, Debug, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
46	pub enum ChatRole {
47	System,
48	User,
49	Assistant,
50	}
51
52	/// Sampling parameters. Conservative defaults; the daemon may
53	/// further constrain values it can't honor for the selected model.
54	#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)]
55	pub struct SamplingParams {
56	pub max_tokens: u32,
57	pub temperature: f32,
58	pub top_p: f32,
59	/// Stop strings. Generation halts on the first match.
60	pub stop: Vec<String>,
61	/// Seed for reproducible sampling. `None` means daemon-chosen.
62	pub seed: Option<u64>,
63	}
64
65	impl Default for SamplingParams {
66	fn default() -> Self {
67	Self {
68	max_tokens: 512,
69	temperature: 0.7,
70	top_p: 0.95,
71	stop: Vec::new(),
72	seed: None,
73	}
74	}
75	}
76
77	/// Scheduling priority class. The supervisor's fair-share policy
78	/// uses these to weight admission and preemption between
79	/// concurrent clients (Thread C, A2).
80	#[derive(
81	Clone,
82	Copy,
83	Debug,
84	Default,
85	PartialEq,
86	Eq,
87	PartialOrd,
88	Ord,
89	serde::Serialize,
90	serde::Deserialize,
91	)]
92	pub enum Priority {
93	/// Background work; preemptible. Batch jobs, embeddings runs.
94	Background,
95	/// Default. Interactive but not latency-critical.
96	#[default]
97	Normal,
98	/// Latency-critical interactive sessions. Preempts background
99	/// work but not other interactive sessions.
100	Interactive,
101	}
102