| 1 |
|
| 2 |
|
| 3 |
use core::time::Duration; |
| 4 |
|
| 5 |
use crate::model::{ModelRef, Priority, PromptSpec, SamplingParams}; |
| 6 |
|
| 7 |
|
| 8 |
|
| 9 |
|
| 10 |
#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash, serde::Serialize, serde::Deserialize)] |
| 11 |
pub struct RequestId(pub u64); |
| 12 |
|
| 13 |
|
| 14 |
#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)] |
| 15 |
pub struct ClientRequest { |
| 16 |
pub id: RequestId, |
| 17 |
pub kind: RequestKind, |
| 18 |
} |
| 19 |
|
| 20 |
|
| 21 |
|
| 22 |
|
| 23 |
#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)] |
| 24 |
pub enum RequestKind { |
| 25 |
|
| 26 |
|
| 27 |
Inference(InferenceRequest), |
| 28 |
|
| 29 |
|
| 30 |
DescribeFleet, |
| 31 |
|
| 32 |
|
| 33 |
LoadModel(ModelRef), |
| 34 |
|
| 35 |
|
| 36 |
|
| 37 |
|
| 38 |
UnloadModel(ModelRef), |
| 39 |
|
| 40 |
|
| 41 |
|
| 42 |
Cancel(RequestId), |
| 43 |
} |
| 44 |
|
| 45 |
|
| 46 |
#[derive(Clone, Debug, serde::Serialize, serde::Deserialize)] |
| 47 |
pub struct InferenceRequest { |
| 48 |
pub model: ModelRef, |
| 49 |
pub prompt: PromptSpec, |
| 50 |
pub sampling: SamplingParams, |
| 51 |
pub stream: bool, |
| 52 |
pub priority: Priority, |
| 53 |
|
| 54 |
|
| 55 |
|
| 56 |
pub deadline: Option<Duration>, |
| 57 |
} |
| 58 |
|