incin-core 0.1.0

Core tensor operations and static shape checking for incin
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
513
514
515
516
//! Checked collective descriptors, plan construction, and the plan builder.

use super::*;

/// Monotonic position of a collective in one plan.
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
pub struct SequenceToken(u64);

impl SequenceToken {
    /// Zero-based sequence number.
    #[must_use]
    pub const fn get(self) -> u64 {
        self.0
    }
}

/// Stable semantic identity attached to one collective.
///
/// Counts, dtypes, and placements are not enough to distinguish two
/// same-shaped parameters. A higher-level planner such as data parallelism
/// assigns a tag so swapping those parameters changes the preflight hash
/// before either rank launches.
#[derive(Debug, Default, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
pub struct CollectiveTag(u64);

impl CollectiveTag {
    /// An unlabelled collective, used by the generic planner methods.
    pub const ANONYMOUS: Self = Self(0);

    /// Build a stable caller-defined identity.
    #[must_use]
    pub const fn new(value: u64) -> Self {
        Self(value)
    }

    /// Numeric identity included in the plan hash.
    #[must_use]
    pub const fn get(self) -> u64 {
        self.0
    }
}

/// One fully checked transport launch.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct CollectiveDescriptor {
    tag: CollectiveTag,
    group: GroupId,
    kind: CollectiveKind,
    input_elements: usize,
    output_elements: usize,
    input_bytes: usize,
    output_bytes: usize,
    dtype: DTypeId,
    source: PlacementKind,
    destination: PlacementKind,
    sequence: SequenceToken,
    stream: StreamId,
    depends_on: Option<SequenceToken>,
}

impl CollectiveDescriptor {
    /// Higher-level semantic identity included in preflight agreement.
    #[must_use]
    pub const fn tag(&self) -> CollectiveTag {
        self.tag
    }

    /// Ordered communicator identity and cardinality.
    #[must_use]
    pub const fn group(&self) -> GroupId {
        self.group
    }

    /// Collective operation and reduction semantics.
    #[must_use]
    pub const fn kind(&self) -> CollectiveKind {
        self.kind
    }

    /// Number of logical input elements at each rank.
    #[must_use]
    pub const fn input_elements(&self) -> usize {
        self.input_elements
    }

    /// Number of logical output elements at each rank.
    #[must_use]
    pub const fn output_elements(&self) -> usize {
        self.output_elements
    }

    /// Checked input byte count at each rank.
    #[must_use]
    pub const fn input_bytes(&self) -> usize {
        self.input_bytes
    }

    /// Checked output byte count at each rank.
    #[must_use]
    pub const fn output_bytes(&self) -> usize {
        self.output_bytes
    }

    /// Runtime projection of the static or `Dyn` dtype.
    #[must_use]
    pub const fn dtype(&self) -> DTypeId {
        self.dtype
    }

    /// Placement consumed by this movement.
    #[must_use]
    pub const fn source(&self) -> PlacementKind {
        self.source
    }

    /// Placement produced by this movement.
    #[must_use]
    pub const fn destination(&self) -> PlacementKind {
        self.destination
    }

    /// Monotonic launch position.
    #[must_use]
    pub const fn sequence(&self) -> SequenceToken {
        self.sequence
    }

    /// Logical communication stream.
    #[must_use]
    pub const fn stream(&self) -> StreamId {
        self.stream
    }

    /// Earlier collective that must complete before this launch.
    pub const fn depends_on(&self) -> Option<SequenceToken> {
        self.depends_on
    }
}

/// Immutable ordered plan tied to one physical mesh identity.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct CollectivePlan {
    mesh: MeshId,
    descriptors: Vec<CollectiveDescriptor>,
    hash: u64,
}

impl CollectivePlan {
    /// Physical/logical mesh this plan was built against.
    #[must_use]
    pub const fn mesh_id(&self) -> MeshId {
        self.mesh
    }

    /// Collectives in launch order.
    #[must_use]
    pub fn descriptors(&self) -> &[CollectiveDescriptor] {
        &self.descriptors
    }

    /// Stable hash compared before any rank launches.
    #[must_use]
    pub const fn hash(&self) -> u64 {
        self.hash
    }

    /// Compact value exchanged during preflight.
    #[must_use]
    pub fn summary(&self) -> PlanSummary {
        PlanSummary {
            mesh: self.mesh,
            hash: self.hash,
            collectives: self.descriptors.len(),
        }
    }
}

/// Builder that assigns sequence tokens and derives all message counts.
#[derive(Debug)]
pub struct CollectivePlanBuilder<'a, M: ValidMesh> {
    mesh: &'a DeviceMesh<M>,
    descriptors: Vec<CollectiveDescriptor>,
}

/// Static placement/dtype pair that denotes an actual collective.
///
/// Local-only identity and replicated-to-sharded selection intentionally have
/// no implementation. A static caller cannot submit them to a collective
/// planner and wait for a runtime "no collective required" error.
#[diagnostic::on_unimplemented(
    message = "`{Self}` to `{To}` is not a collective supported for dtype `{K}`",
    label = "no statically valid collective transition",
    note = "integer mean reductions require a floating dtype; local-only placement changes do not belong in a collective plan"
)]
pub trait PlannedCollectiveTransition<K: CollectiveDType, To: Placement>:
    LegalTransition<To>
{
}

impl<K, M, Axis> PlannedCollectiveTransition<K, Replicated<M>> for Sharded<M, Axis>
where
    K: CollectiveDType,
    M: ValidMesh,
    Axis: PlacementAxis,
{
}

impl<K, M, R> PlannedCollectiveTransition<K, Replicated<M>> for Partial<M, R>
where
    K: CollectiveReductionDType<R>,
    M: ValidMesh,
    R: PartialReduction,
{
}

impl<K, M, R, Axis> PlannedCollectiveTransition<K, Sharded<M, Axis>> for Partial<M, R>
where
    K: CollectiveReductionDType<R>,
    M: ValidMesh,
    R: PartialReduction,
    Axis: PlacementAxis,
{
}

impl<'a, M: ValidMesh> CollectivePlanBuilder<'a, M> {
    /// Begin a plan for a bound mesh.
    #[must_use]
    pub fn new(mesh: &'a DeviceMesh<M>) -> Self {
        Self {
            mesh,
            descriptors: Vec::new(),
        }
    }

    /// Append a transition whose dtype and placements are compile-time known.
    ///
    /// Unsupported static dtypes and illegal transitions have no satisfying
    /// trait implementation and therefore fail to compile.
    pub fn push_static<K, From, To>(
        &mut self,
        axis: MeshAxis,
        rank: usize,
        input_elements: usize,
        stream: StreamId,
        depends_on: Option<SequenceToken>,
    ) -> Result<SequenceToken, PlanError>
    where
        K: ConstDType + BuiltinDType + CollectiveDType,
        From: ConstPlacement + PlannedCollectiveTransition<K, To> + PlacementOn<M>,
        To: ConstPlacement + PlacementOn<M>,
    {
        self.push_static_tagged::<K, From, To>(
            CollectiveTag::ANONYMOUS,
            axis,
            rank,
            input_elements,
            stream,
            depends_on,
        )
    }

    /// Append a statically checked transition with a semantic identity.
    ///
    /// The tag has no transport meaning; it exists so higher-level planners
    /// can make same-shaped operation reordering visible to preflight.
    #[allow(clippy::too_many_arguments)]
    pub fn push_static_tagged<K, From, To>(
        &mut self,
        tag: CollectiveTag,
        axis: MeshAxis,
        rank: usize,
        input_elements: usize,
        stream: StreamId,
        depends_on: Option<SequenceToken>,
    ) -> Result<SequenceToken, PlanError>
    where
        K: ConstDType + BuiltinDType + CollectiveDType,
        From: ConstPlacement + PlannedCollectiveTransition<K, To> + PlacementOn<M>,
        To: ConstPlacement + PlacementOn<M>,
    {
        self.push_checked(
            tag,
            axis,
            rank,
            input_elements,
            K::DTYPE,
            From::PLACEMENT,
            To::PLACEMENT,
            From::TRANSITION,
            stream,
            depends_on,
        )
    }

    /// Append the runtime-selected counterpart of [`push_static`](Self::push_static).
    #[allow(clippy::too_many_arguments)]
    pub fn push_dyn(
        &mut self,
        axis: MeshAxis,
        rank: usize,
        input_elements: usize,
        dtype: DTypeId,
        source: PlacementKind,
        destination: PlacementKind,
        stream: StreamId,
        depends_on: Option<SequenceToken>,
    ) -> Result<SequenceToken, PlanError> {
        self.push_dyn_tagged(
            CollectiveTag::ANONYMOUS,
            axis,
            rank,
            input_elements,
            dtype,
            source,
            destination,
            stream,
            depends_on,
        )
    }

    /// Append a runtime-checked transition with a semantic identity.
    #[allow(clippy::too_many_arguments)]
    pub fn push_dyn_tagged(
        &mut self,
        tag: CollectiveTag,
        axis: MeshAxis,
        rank: usize,
        input_elements: usize,
        dtype: DTypeId,
        source: PlacementKind,
        destination: PlacementKind,
        stream: StreamId,
        depends_on: Option<SequenceToken>,
    ) -> Result<SequenceToken, PlanError> {
        validate_collective_dtype(dtype)?;
        let transition = crate::dist::validate_transition(source, destination)?;
        self.push_checked(
            tag,
            axis,
            rank,
            input_elements,
            dtype,
            source,
            destination,
            transition,
            stream,
            depends_on,
        )
    }

    #[allow(clippy::too_many_arguments)]
    fn push_checked(
        &mut self,
        tag: CollectiveTag,
        axis: MeshAxis,
        rank: usize,
        input_elements: usize,
        dtype: DTypeId,
        source: PlacementKind,
        destination: PlacementKind,
        transition: PlacementTransition,
        stream: StreamId,
        depends_on: Option<SequenceToken>,
    ) -> Result<SequenceToken, PlanError> {
        validate_collective_dtype(dtype)?;
        let runtime_transition = crate::dist::validate_transition(source, destination)?;
        if runtime_transition != transition {
            return Err(PlanError::TransitionMismatch {
                typed: transition,
                runtime: runtime_transition,
            });
        }
        if let Some(dependency) = depends_on
            && dependency.0 >= self.descriptors.len() as u64
        {
            return Err(PlanError::UnknownDependency {
                dependency,
                next: self.descriptors.len(),
            });
        }

        let members = self
            .mesh
            .groups()
            .group(axis, rank)
            .ok_or(PlanError::RankOutOfRange {
                rank,
                world: M::WORLD,
            })?;
        let group = GroupId::new(group_token(self.mesh.id(), axis, &members), members.len())?;
        let kind = kind_for_transition(transition, source)?;
        if let CollectiveKind::AllReduce(op) | CollectiveKind::ReduceScatter(op) = kind {
            validate_collective_reduction(dtype, op)?;
        }
        if matches!(
            kind,
            CollectiveKind::AllGather | CollectiveKind::ReduceScatter(_)
        ) && axis != MeshAxis::Tensor
        {
            return Err(PlanError::WrongAxis {
                kind,
                expected: MeshAxis::Tensor,
                found: axis,
            });
        }
        let output_elements = output_elements(kind, input_elements, group.ranks())?;
        let input_bytes = dtype.size_bytes(input_elements, OperationKind::Storage)?;
        let output_bytes = dtype.size_bytes(output_elements, OperationKind::Storage)?;
        let sequence = SequenceToken(
            u64::try_from(self.descriptors.len()).map_err(|_| PlanError::SequenceOverflow)?,
        );

        self.descriptors.push(CollectiveDescriptor {
            tag,
            group,
            kind,
            input_elements,
            output_elements,
            input_bytes,
            output_bytes,
            dtype,
            source,
            destination,
            sequence,
            stream,
            depends_on,
        });
        Ok(sequence)
    }

    /// Append one globally described point-to-point transfer.
    ///
    /// Higher-level pipeline planning owns the static source/destination-stage
    /// proof. This crate-visible entry point owns the shared runtime
    /// invariants and descriptor hashing so static and dynamic pipeline paths
    /// cannot accidentally mint different wire contracts.
    #[allow(clippy::too_many_arguments)]
    pub(crate) fn push_send_recv_tagged(
        &mut self,
        tag: CollectiveTag,
        rank: usize,
        input_elements: usize,
        dtype: DTypeId,
        source_rank: usize,
        destination_rank: usize,
        source_stage: usize,
        destination_stage: usize,
        stream: StreamId,
        depends_on: Option<SequenceToken>,
    ) -> Result<SequenceToken, PlanError> {
        validate_collective_dtype(dtype)?;
        if let Some(dependency) = depends_on
            && dependency.0 >= self.descriptors.len() as u64
        {
            return Err(PlanError::UnknownDependency {
                dependency,
                next: self.descriptors.len(),
            });
        }

        let members = self.mesh.groups().group(MeshAxis::Pipeline, rank).ok_or(
            PlanError::RankOutOfRange {
                rank,
                world: M::WORLD,
            },
        )?;
        let group = GroupId::new(
            group_token(self.mesh.id(), MeshAxis::Pipeline, &members),
            members.len(),
        )?;
        validate_peer("source", source_rank, group.ranks())?;
        validate_peer("destination", destination_rank, group.ranks())?;
        if source_rank == destination_rank {
            return Err(CollectiveError::SamePeer { rank: source_rank }.into());
        }

        let input_bytes = dtype.size_bytes(input_elements, OperationKind::Storage)?;
        let sequence = SequenceToken(
            u64::try_from(self.descriptors.len()).map_err(|_| PlanError::SequenceOverflow)?,
        );
        self.descriptors.push(CollectiveDescriptor {
            tag,
            group,
            kind: CollectiveKind::SendRecv {
                source: source_rank,
                destination: destination_rank,
            },
            input_elements,
            output_elements: input_elements,
            input_bytes,
            output_bytes: input_bytes,
            dtype,
            source: PlacementKind::PipelineStage {
                index: source_stage,
            },
            destination: PlacementKind::PipelineStage {
                index: destination_stage,
            },
            sequence,
            stream,
            depends_on,
        });
        Ok(sequence)
    }

    /// Freeze ordering and compute the stable preflight hash.
    #[must_use]
    pub fn finish(self) -> CollectivePlan {
        let hash = plan_hash(self.mesh.id(), &self.descriptors);
        CollectivePlan {
            mesh: self.mesh.id(),
            descriptors: self.descriptors,
            hash,
        }
    }
}