bb_runtime/contracts/model.rs
1//! `bb::Model` — Contract trait for ML models.
2//!
3//! Each method takes the engine's `&mut RuntimeResourceRef<'_>` ctx
4//! plus a [`CompletionHandle`] AND returns [`ContractResponse`]. See
5//! [`crate::contracts::index`] for the sync (Now) vs async (Later)
6//! semantics.
7//!
8//! ## Associated type: `Tensor`
9//!
10//! One associated type covers input tensors, output tensors, parameter
11//! vectors, gradients, and deltas. Mixed-precision (e.g. f32 input +
12//! f16 weights + f32 output) is handled by wiring [`Codec`] nodes
13//! around the model in the Module body — not by multiplying associated
14//! types per port.
15//!
16//! [`Codec`]: crate::contracts::Codec
17
18use crate::completion::{CompletionHandle, ContractResponse};
19use crate::runtime::RuntimeResourceRef;
20
21/// User-facing Contract trait for an ML model.
22pub trait Model: Send + Sync {
23 /// Tensor storage type. One associated type covers
24 /// input/output/params/grad/delta. Implement as `[f32]` for
25 /// flat f32 tensors.
26 type Tensor: ?Sized + bb_ir::types::Storage;
27 /// Library-maker-defined error type.
28 type Error: std::error::Error + std::fmt::Display + Send + Sync + 'static;
29
30 /// Forward pass: `input → output`. `ctx` is the per-dispatch
31 /// runtime surface; impls reach their declared `#[depends(...)]`
32 /// siblings through [`RuntimeResourceRef::dependency`].
33 fn forward(
34 &mut self,
35 ctx: &mut RuntimeResourceRef<'_>,
36 input: &Self::Tensor,
37 completion: CompletionHandle<Box<Self::Tensor>, Self::Error>,
38 ) -> ContractResponse<Box<Self::Tensor>, Self::Error>;
39
40 /// Load parameters wholesale.
41 fn load_parameters(
42 &mut self,
43 ctx: &mut RuntimeResourceRef<'_>,
44 params: &Self::Tensor,
45 completion: CompletionHandle<(), Self::Error>,
46 ) -> ContractResponse<(), Self::Error>;
47
48 /// Backward pass: accumulate gradients given upstream gradient.
49 fn backward(
50 &mut self,
51 ctx: &mut RuntimeResourceRef<'_>,
52 grad: &Self::Tensor,
53 completion: CompletionHandle<(), Self::Error>,
54 ) -> ContractResponse<(), Self::Error>;
55
56 /// Apply a parameter delta in-place.
57 fn apply_delta(
58 &mut self,
59 ctx: &mut RuntimeResourceRef<'_>,
60 delta: &Self::Tensor,
61 completion: CompletionHandle<(), Self::Error>,
62 ) -> ContractResponse<(), Self::Error>;
63
64 /// Compute loss: `(input, target) → scalar score`. Returns `f32`
65 /// regardless of the tensor element type — loss is always a
66 /// framework-fixed scalar.
67 fn compute_loss(
68 &mut self,
69 ctx: &mut RuntimeResourceRef<'_>,
70 input: &Self::Tensor,
71 target: &Self::Tensor,
72 completion: CompletionHandle<f32, Self::Error>,
73 ) -> ContractResponse<f32, Self::Error>;
74
75 /// Snapshot the current parameter tensor (owned — async
76 /// serialization needs owned values).
77 fn params(
78 &self,
79 ctx: &mut RuntimeResourceRef<'_>,
80 completion: CompletionHandle<Box<Self::Tensor>, Self::Error>,
81 ) -> ContractResponse<Box<Self::Tensor>, Self::Error>;
82}