ml-runtime-inference 0.2.3

Typed inference requests, results, and provenance for rust-ml-runtime
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
use bytes::Bytes;
use ml_runtime_model::ModelReference;
use serde::{Deserialize, Serialize};
use serde_json::Value;
use std::{collections::BTreeMap, path::PathBuf, time::Duration};

/// Determines where the runtime attempts to execute a request and whether it may fall back.
#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq, Eq)]
pub enum ExecutionPolicy {
    #[default]
    LocalOnly,
    RemoteOnly,
    PreferLocal,
    PreferRemote,
    LocalThenRemote,
    RemoteThenLocal,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct Tensor {
    pub shape: Vec<usize>,
    pub values: Vec<f32>,
}

impl Tensor {
    pub fn new(shape: impl Into<Vec<usize>>, values: impl Into<Vec<f32>>) -> Self {
        Self {
            shape: shape.into(),
            values: values.into(),
        }
    }
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
pub struct Image {
    pub mime_type: String,
    pub data: Bytes,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
pub struct Audio {
    pub mime_type: String,
    pub data: Bytes,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
pub struct InferenceOptions {
    pub timeout_ms: Option<u64>,
    pub max_output_tokens: Option<u64>,
    pub batch_size: Option<usize>,
    pub require_local: bool,
    pub require_remote: bool,
    pub stream: bool,
    #[serde(default)]
    pub execution: ExecutionPolicy,
}

impl Default for InferenceOptions {
    fn default() -> Self {
        Self {
            timeout_ms: None,
            max_output_tokens: None,
            batch_size: None,
            require_local: false,
            require_remote: false,
            stream: false,
            execution: ExecutionPolicy::LocalOnly,
        }
    }
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub enum Input {
    Text(String),
    Tokens(Vec<u32>),
    Tensor(Tensor),
    Image(Image),
    Audio(Audio),
    Binary(Bytes),
    Structured(Value),
}

impl Input {
    pub fn batch_size(&self) -> usize {
        match self {
            Self::Tokens(_) => 1,
            Self::Tensor(tensor) => tensor.shape.first().copied().unwrap_or(1),
            _ => 1,
        }
    }
}

impl From<String> for Input {
    fn from(value: String) -> Self {
        Self::Text(value)
    }
}

impl From<&str> for Input {
    fn from(value: &str) -> Self {
        Self::Text(value.to_owned())
    }
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub enum Output {
    Text(String),
    Tokens(Vec<u32>),
    Tensor(Tensor),
    Embedding(Vec<f32>),
    Structured(Value),
    Binary(Bytes),
}

#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
pub struct ExecutionMetadata {
    #[serde(default)]
    pub request_id: Option<String>,
    pub model: String,
    pub model_version: Option<String>,
    /// Execution location or transport selected by this runtime (`local`, `remote`, etc.).
    pub provider: String,
    /// Computation implementation reported by the execution target (`onnx`, `cpu`, etc.).
    pub backend: String,
    /// Observable target selected by this runtime. For a client this may be remote even when
    /// `backend` reports the server's local computation backend.
    pub execution_target: String,
    pub routing_policy: ExecutionPolicy,
    pub fallback: bool,
    pub fallback_from: Option<String>,
    pub fallback_reason: Option<String>,
    pub latency_ms: Option<f64>,
    pub input_tokens: Option<u64>,
    pub output_tokens: Option<u64>,
    pub hardware: Option<String>,
    #[serde(default)]
    pub cache_hit: bool,
    #[serde(default)]
    pub batch_size: usize,
    #[serde(default)]
    pub queue_wait_ms: Option<f64>,
    #[serde(default)]
    pub model_load_ms: Option<f64>,
    #[serde(default)]
    pub execution_ms: Option<f64>,
    #[serde(default)]
    pub streaming_requested: bool,
    #[serde(default)]
    pub streaming_supported: bool,
    #[serde(default)]
    pub output_event_count: usize,
    #[serde(default)]
    pub completion_state: Option<String>,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct InferenceRequest {
    pub model: ModelReference,
    pub input: Input,
    pub options: InferenceOptions,
}

impl InferenceRequest {
    /// Creates a request with default local-only execution options.
    pub fn new(model: impl Into<ModelReference>, input: impl Into<Input>) -> Self {
        Self {
            model: model.into(),
            input: input.into(),
            options: InferenceOptions::default(),
        }
    }

    /// Replaces the request's execution options.
    pub fn with_options(mut self, options: InferenceOptions) -> Self {
        self.options = options;
        self
    }
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct InferenceResult {
    pub output: Output,
    pub metadata: ExecutionMetadata,
}

impl InferenceResult {
    pub fn output(&self) -> &Output {
        &self.output
    }

    pub fn metadata(&self) -> &ExecutionMetadata {
        &self.metadata
    }

    pub fn into_output(self) -> Output {
        self.output
    }
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct InferenceChunk {
    pub output: Output,
    pub done: bool,
    pub metadata: Option<ExecutionMetadata>,
}

pub type InferenceOutput = Output;

/// A model-neutral request for one or more typed decisions about immutable input state.
#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct DecisionRequest {
    pub state: Value,
    pub decisions: Vec<DecisionQuestion>,
}

/// Backend/model execution facts used by the structured-decision planner.
#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
pub struct DecisionModelCapabilities {
    pub backend: String,
    pub device: DeviceKind,
    pub model_architecture: ModelArchitecture,
    pub supports_batching: bool,
    pub max_batch_size: Option<usize>,
    pub batch_size: Option<usize>,
    pub supports_async: bool,
    /// True only when an in-flight native operation can actually be interrupted.
    pub supports_cancellation: bool,
    pub supports_parallel_execution: bool,
    pub recommended_parallelism: Option<usize>,
    pub supports_structured_decisions: bool,
}

impl DecisionModelCapabilities {
    pub fn conservative(backend: impl Into<String>) -> Self {
        Self {
            backend: backend.into(),
            device: DeviceKind::Unknown,
            model_architecture: ModelArchitecture::StructuredDecision,
            supports_batching: false,
            max_batch_size: Some(1),
            batch_size: Some(1),
            supports_async: false,
            supports_cancellation: false,
            supports_parallel_execution: false,
            recommended_parallelism: Some(1),
            supports_structured_decisions: true,
        }
    }
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "snake_case")]
pub enum DeviceKind {
    Cpu,
    Gpu,
    NeuralEngine,
    WebAssembly,
    Unknown,
    Other(String),
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "snake_case")]
pub enum ModelArchitecture {
    StructuredDecision,
    Tensor,
    Generative,
    Other(String),
}

/// Application constraints. These bound execution without prescribing a backend strategy.
#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
pub struct DecisionExecutionPolicy {
    #[serde(default)]
    pub latency_target: Option<Duration>,
    #[serde(default)]
    pub max_parallelism: Option<usize>,
    #[serde(default)]
    pub max_batch_size: Option<usize>,
    #[serde(default = "default_true")]
    pub allow_parallel: bool,
    #[serde(default = "default_true")]
    pub allow_batching: bool,
    #[serde(default = "default_true")]
    pub allow_selective_execution: bool,
}

fn default_true() -> bool {
    true
}

impl Default for DecisionExecutionPolicy {
    fn default() -> Self {
        Self {
            latency_target: None,
            max_parallelism: None,
            max_batch_size: None,
            allow_parallel: true,
            allow_batching: true,
            allow_selective_execution: true,
        }
    }
}

/// One decision node and the predicates that control whether it is required.
#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct DecisionNode {
    pub question: DecisionQuestion,
    #[serde(default)]
    pub dependencies: Vec<DecisionDependency>,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct DecisionDependency {
    pub decision: String,
    #[serde(default)]
    pub condition: DecisionDependencyCondition,
}

#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
#[serde(tag = "type", content = "value", rename_all = "snake_case")]
pub enum DecisionDependencyCondition {
    #[default]
    Completed,
    ChoiceEquals(String),
    NoulEquals(bool),
    ScoreAtLeast(f64),
    ScoreAtMost(f64),
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct DecisionGraphRequest {
    pub state: Value,
    pub nodes: Vec<DecisionNode>,
    #[serde(default)]
    pub policy: DecisionExecutionPolicy,
}

impl DecisionGraphRequest {
    pub fn independent(request: DecisionRequest, policy: DecisionExecutionPolicy) -> Self {
        Self {
            state: request.state,
            nodes: request
                .decisions
                .into_iter()
                .map(|question| DecisionNode {
                    question,
                    dependencies: Vec::new(),
                })
                .collect(),
            policy,
        }
    }
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
#[serde(tag = "type", rename_all = "snake_case")]
pub enum DecisionExecutionStrategy {
    Single,
    Sequential,
    Parallel { concurrency: usize },
    Batched { batch_size: usize },
    Selective,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
pub struct DecisionPlanStage {
    pub nodes: Vec<String>,
    pub batches: Vec<Vec<String>>,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
pub struct DecisionExecutionPlan {
    pub strategy: DecisionExecutionStrategy,
    pub reason: String,
    pub stages: Vec<DecisionPlanStage>,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
pub struct DecisionExecutionDiagnostics {
    pub capabilities: DecisionModelCapabilities,
    pub strategy: DecisionExecutionStrategy,
    pub reason: String,
    pub requested_nodes: usize,
    pub executed_nodes: usize,
    pub skipped_nodes: usize,
    pub batches: Vec<Vec<String>>,
    pub max_concurrency: usize,
    pub cancelled: bool,
}

impl DecisionRequest {
    pub fn new(state: impl Into<Value>, decisions: Vec<DecisionQuestion>) -> Self {
        Self {
            state: state.into(),
            decisions,
        }
    }
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct DecisionQuestion {
    pub name: String,
    pub instructions: String,
    pub kind: DecisionType,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
#[serde(tag = "type", rename_all = "snake_case")]
pub enum DecisionType {
    Choice {
        options: Vec<DecisionOption>,
    },
    Score {
        levels: Vec<Value>,
    },
    /// Laya's binary yes/no decision head ("noul" in the model contract).
    Noul {
        #[serde(default)]
        false_description: Option<Value>,
        #[serde(default)]
        true_description: Option<Value>,
    },
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct DecisionOption {
    pub label: String,
    #[serde(default)]
    pub description: Option<Value>,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct DecisionResult {
    pub model: ModelIdentity,
    pub backend: String,
    pub decisions: Vec<TypedDecision>,
    pub execution: DecisionExecution,
    pub provenance: DecisionProvenance,
}

/// Result returned by the capability-aware executor. Flattening preserves the
/// existing JSON shape while keeping `DecisionResult` source-compatible.
#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct PlannedDecisionResult {
    #[serde(flatten)]
    pub result: DecisionResult,
    pub planning: DecisionExecutionDiagnostics,
}

impl std::ops::Deref for PlannedDecisionResult {
    type Target = DecisionResult;

    fn deref(&self) -> &Self::Target {
        &self.result
    }
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
pub struct ModelIdentity {
    pub identifier: String,
    pub revision: Option<String>,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct TypedDecision {
    pub name: String,
    pub kind: String,
    pub value: DecisionValue,
    pub probabilities: BTreeMap<String, f64>,
    pub confidence: f64,
    pub action_probability: f64,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
#[serde(tag = "type", content = "value", rename_all = "snake_case")]
pub enum DecisionValue {
    Choice(String),
    Score(f64),
    Noul(bool),
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
pub struct DecisionExecution {
    pub latency: Duration,
    pub input_tokens: u64,
    pub output_tokens: u64,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
pub struct DecisionProvenance {
    pub artifact_path: PathBuf,
    pub artifact_sha256: String,
    pub runtime_version: String,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
pub struct ModelDescription {
    pub identifier: String,
    pub revision: Option<String>,
    pub backend: String,
    pub artifact_path: PathBuf,
    pub artifact_sha256: String,
    pub decision_types: Vec<String>,
}

#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub enum InferenceStreamEvent {
    Started(ExecutionMetadata),
    Output(InferenceOutput),
    Completed(ExecutionMetadata),
}