vyre-driver-cuda 0.6.2

CUDA/PTX backend for vyre through the CUDA driver API.
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
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
//! Replay helpers for captured CUDA graphs.

use std::ptr::NonNull;
use std::sync::Arc;

use smallvec::SmallVec;
use vyre_driver::BackendError;

use super::allocations::cuda_check;
use super::cuda_graph::{CachedCudaGraph, GraphExecGuard, StreamGuard};
use super::dispatch::CudaBackend;
use super::staging_reserve::{reserve_smallvec, reserve_vec, reserved_vec, resize_vec_slots};
use crate::input_identity::{exact_input_key, ExactInputKey};

impl CachedCudaGraph {
    pub(crate) fn input_shape_matches(&self, inputs: &[&[u8]]) -> bool {
        inputs.len() == self.expected_input_lens.len()
            && self.input_indices.len() == self.expected_input_lens.len()
            && self
                .input_indices
                .iter()
                .zip(self.expected_input_lens.iter())
                .all(|(input_index, expected)| {
                    inputs
                        .get(*input_index)
                        .is_some_and(|input| input.len() == *expected)
                })
    }

    pub(crate) fn materialized_output_cache_matches(
        &self,
        inputs: &[&[u8]],
    ) -> Result<bool, BackendError> {
        let input_state = prepare_cuda_graph_replay_input_state(self, inputs)?;
        self.materialized_output_cache_matches_with_input_state(inputs, &input_state)
    }

    pub(crate) fn materialized_output_cache_matches_with_input_state(
        &self,
        inputs: &[&[u8]],
        input_state: &CudaGraphReplayInputState,
    ) -> Result<bool, BackendError> {
        if !(self.resident_input_replay_safe && self.host_outputs_initialized) {
            return Ok(false);
        }
        cached_input_bytes_match_with_key(self, inputs, &input_state.input_key)
    }
}

#[derive(Clone, Copy, Debug, Default)]
pub(crate) struct CudaGraphReplayStats {
    input_bytes: u64,
    output_bytes: u64,
    host_upload_operations: u64,
    device_readback_operations: u64,
}

#[derive(Clone, Copy, Debug)]
pub(crate) struct CudaGraphReplayInputState {
    input_key: ExactInputKey,
}

#[derive(Clone, Copy, Debug)]
struct PreparedCudaGraphReplayLaunch {
    stats: CudaGraphReplayStats,
    resident_input_replay: bool,
}

const CUDA_GRAPH_REPLAY_SPIN_QUERY_LIMIT: usize = 4096;

fn launch_cuda_graph_exec(
    graph_exec: &GraphExecGuard,
    stream: &StreamGuard,
    label: &'static str,
) -> Result<(), BackendError> {
    let graph_exec = graph_exec.ptr();
    if graph_exec == NonNull::dangling() {
        return Err(BackendError::InvalidProgram {
            fix: format!(
                "Fix: CUDA graph replay received a dangling CUgraphExec sentinel before {label}. Re-record the graph before replay."
            ),
        });
    }
    let stream = stream.ptr();
    if stream == NonNull::dangling() {
        return Err(BackendError::InvalidProgram {
            fix: format!(
                "Fix: CUDA graph replay received a dangling CUstream sentinel before {label}. Re-record the graph before replay."
            ),
        });
    }
    // SAFETY: FFI to libcuda.so. `GraphExecGuard` and `StreamGuard` own
    // non-null CUDA handles and the dangling sentinels are rejected above.
    unsafe {
        cuda_check(
            cudarc::driver::sys::cuGraphLaunch(graph_exec.as_ptr(), stream.as_ptr()),
            label,
        )
    }
}

fn synchronize_cuda_graph_replay_stream(cached: &CachedCudaGraph) -> Result<(), BackendError> {
    for _ in 0..CUDA_GRAPH_REPLAY_SPIN_QUERY_LIMIT {
        if crate::stream::query_raw_stream_ready(
            cached.stream.ptr().as_ptr(),
            "cuStreamQuery (cuda_graph)",
        )? {
            return Ok(());
        }
        std::hint::spin_loop();
    }
    crate::stream::synchronize_raw_stream(
        cached.stream.ptr().as_ptr(),
        "cuStreamSynchronize (cuda_graph fallback)",
    )
}

fn cached_input_bytes_match(
    cached: &CachedCudaGraph,
    inputs: &[&[u8]],
) -> Result<bool, BackendError> {
    let input_key = exact_input_key(inputs)?;
    cached_input_bytes_match_with_key(cached, inputs, &input_key)
}

fn cached_input_bytes_match_with_key(
    cached: &CachedCudaGraph,
    inputs: &[&[u8]],
    input_key: &ExactInputKey,
) -> Result<bool, BackendError> {
    if cached.cached_input_key != *input_key {
        return Ok(false);
    }
    cached_input_bytes_match_after_key_match(cached, inputs)
}

fn cached_input_bytes_match_after_key_match(
    cached: &CachedCudaGraph,
    inputs: &[&[u8]],
) -> Result<bool, BackendError> {
    if cached.input_host_bufs.len() != inputs.len() {
        return Err(BackendError::InvalidProgram {
            fix: format!(
                "Fix: cached cuda graph has {} pinned input buffer(s) for {} caller input(s). Re-record the graph; zip-based replay would skip input uploads.",
                cached.input_host_bufs.len(),
                inputs.len()
            ),
        });
    }
    for (slot_index, (slot, input_index)) in cached
        .input_host_bufs
        .iter()
        .zip(cached.input_indices.iter())
        .enumerate()
    {
        let src = cached_graph_input(inputs, *input_index, slot_index, "cached input compare")?;
        if src.len() > slot.byte_len {
            return Err(BackendError::InvalidProgram {
                fix: format!(
                    "Fix: CUDA graph cached input comparison saw {} byte(s) for a {} byte pinned allocation. Re-record the graph for this input shape.",
                    src.len(),
                    slot.byte_len
                ),
            });
        }
        if src.is_empty() {
            continue;
        }
        let cached_bytes = {
            // SAFETY: `slot` owns a pinned allocation of at least `slot.byte_len`
            // bytes, and the length check above proves `src.len() <= slot.byte_len`.
            unsafe { std::slice::from_raw_parts(slot.as_ptr().cast::<u8>(), src.len()) }
        };
        if cached_bytes != src {
            return Ok(false);
        }
    }
    Ok(true)
}

impl CudaBackend {
    pub(crate) fn try_cuda_graph_materialized_cache_into(
        &self,
        cached: &mut CachedCudaGraph,
        inputs: &[&[u8]],
        outputs: &mut Vec<Vec<u8>>,
    ) -> Result<bool, BackendError> {
        let input_state = self.prepare_cuda_graph_replay_input_state(cached, inputs)?;
        self.try_cuda_graph_materialized_cache_with_input_state_into(
            cached,
            inputs,
            &input_state,
            outputs,
        )
    }

    pub(crate) fn try_cuda_graph_materialized_cache_with_input_state_into(
        &self,
        cached: &mut CachedCudaGraph,
        inputs: &[&[u8]],
        input_state: &CudaGraphReplayInputState,
        outputs: &mut Vec<Vec<u8>>,
    ) -> Result<bool, BackendError> {
        if cached.materialized_output_cache_matches_with_input_state(inputs, input_state)? {
            collect_cuda_graph_outputs(cached, outputs)?;
            self.telemetry.record_cuda_graph_materialized_cache_hit();
            return Ok(true);
        }
        Ok(false)
    }

    pub(crate) fn enqueue_cuda_graph_replay(
        &self,
        cached: &mut CachedCudaGraph,
        inputs: &[&[u8]],
    ) -> Result<CudaGraphReplayStats, BackendError> {
        let input_state = self.prepare_cuda_graph_replay_input_state(cached, inputs)?;
        self.enqueue_cuda_graph_replay_with_input_state(cached, inputs, &input_state)
    }

    pub(crate) fn enqueue_cuda_graph_replay_with_input_state(
        &self,
        cached: &mut CachedCudaGraph,
        inputs: &[&[u8]],
        input_state: &CudaGraphReplayInputState,
    ) -> Result<CudaGraphReplayStats, BackendError> {
        let prepared = prepare_cuda_graph_replay_launch(cached, inputs, input_state)?;
        launch_prepared_cuda_graph_replay(cached, &prepared, "cuGraphLaunch")?;
        self.telemetry.record_cuda_graph_launch();
        Ok(prepared.stats)
    }

    pub(crate) fn finish_cuda_graph_replay_into(
        &self,
        cached: &mut CachedCudaGraph,
        stats: CudaGraphReplayStats,
        outputs: &mut Vec<Vec<u8>>,
    ) -> Result<(), BackendError> {
        synchronize_cuda_graph_replay_stream(cached)?;
        cached.device_inputs_initialized = true;
        self.telemetry.record_sync_point();
        self.record_cuda_graph_replay_stats(stats);
        collect_cuda_graph_outputs(cached, outputs)?;
        cached.host_outputs_initialized = true;
        Ok(())
    }

    pub(crate) fn record_cuda_graph_batched_replay_chunk(&self, lanes: u64) {
        self.telemetry.record_cuda_graph_batched_replay(lanes);
    }

    pub(crate) fn prepare_cuda_graph_replay_input_state(
        &self,
        cached: &CachedCudaGraph,
        inputs: &[&[u8]],
    ) -> Result<CudaGraphReplayInputState, BackendError> {
        prepare_cuda_graph_replay_input_state(cached, inputs)
    }

    pub(crate) fn prepare_cuda_graph_replay_input_state_with_key(
        &self,
        cached: &CachedCudaGraph,
        inputs: &[&[u8]],
        input_key: ExactInputKey,
    ) -> Result<CudaGraphReplayInputState, BackendError> {
        prepare_cuda_graph_replay_input_state_with_key(cached, inputs, input_key)
    }

    /// Replay a cached CUDA graph with new input bytes.
    pub fn dispatch_via_cuda_graph_into(
        &self,
        cached: &mut CachedCudaGraph,
        inputs: &[&[u8]],
        outputs: &mut Vec<Vec<u8>>,
    ) -> Result<(), BackendError> {
        let input_state = self.prepare_cuda_graph_replay_input_state(cached, inputs)?;
        self.dispatch_via_cuda_graph_with_input_state_into(cached, inputs, &input_state, outputs)
    }

    pub(crate) fn dispatch_via_cuda_graph_with_input_state_into(
        &self,
        cached: &mut CachedCudaGraph,
        inputs: &[&[u8]],
        input_state: &CudaGraphReplayInputState,
        outputs: &mut Vec<Vec<u8>>,
    ) -> Result<(), BackendError> {
        if self.try_cuda_graph_materialized_cache_with_input_state_into(
            cached,
            inputs,
            &input_state,
            outputs,
        )? {
            return Ok(());
        }
        let stats =
            self.enqueue_cuda_graph_replay_with_input_state(cached, inputs, &input_state)?;
        self.finish_cuda_graph_replay_into(cached, stats, outputs)
    }

    /// Replay a cached CUDA graph with CUDA event timing.
    pub(crate) fn dispatch_via_cuda_graph_timed_into(
        &self,
        cached: &mut CachedCudaGraph,
        inputs: &[&[u8]],
        outputs: &mut Vec<Vec<u8>>,
    ) -> Result<u64, BackendError> {
        let input_state = self.prepare_cuda_graph_replay_input_state(cached, inputs)?;
        self.dispatch_via_cuda_graph_timed_with_input_state_into(
            cached,
            inputs,
            &input_state,
            outputs,
        )
    }

    pub(crate) fn dispatch_via_cuda_graph_timed_with_input_state_into(
        &self,
        cached: &mut CachedCudaGraph,
        inputs: &[&[u8]],
        input_state: &CudaGraphReplayInputState,
        outputs: &mut Vec<Vec<u8>>,
    ) -> Result<u64, BackendError> {
        if self.try_cuda_graph_materialized_cache_with_input_state_into(
            cached,
            inputs,
            &input_state,
            outputs,
        )? {
            return Ok(0);
        }
        self.warmup()?;
        let prepared = prepare_cuda_graph_replay_launch(cached, inputs, &input_state)?;

        let mut timing_events =
            crate::stream::CudaTimingEventPairLease::acquire(Arc::clone(&self.launch_resources))?;
        {
            let (start, end) = timing_events.events()?;
            start.record(cached.stream.ptr().as_ptr())?;
            launch_prepared_cuda_graph_replay(cached, &prepared, "cuGraphLaunch")?;
            self.telemetry.record_cuda_graph_launch();
            end.record(cached.stream.ptr().as_ptr())?;
            end.synchronize()?;
        }
        timing_events.mark_synchronized();
        cached.device_inputs_initialized = true;
        self.telemetry.record_sync_point();
        let device_ns = {
            let (start, end) = timing_events.events()?;
            start.elapsed_time_ns(end)?
        };
        self.record_cuda_graph_replay_stats(prepared.stats);
        collect_cuda_graph_outputs(cached, outputs)?;
        cached.host_outputs_initialized = true;
        Ok(device_ns)
    }

    /// Replay a cached CUDA graph with CUDA event timing and allocated outputs.
    pub fn dispatch_via_cuda_graph_timed(
        &self,
        cached: &mut CachedCudaGraph,
        inputs: &[&[u8]],
    ) -> Result<vyre_driver::TimedDispatchResult, BackendError> {
        let started = std::time::Instant::now();
        let mut outputs = reserved_vec(
            cached.output_host_bufs.len(),
            "timed cuda graph replay output vector",
        )?;
        let device_ns = self.dispatch_via_cuda_graph_timed_into(cached, inputs, &mut outputs)?;
        let wall_ns = crate::numeric::CUDA_NUMERIC
            .elapsed_nanos_u64(started, "timed cuda graph replay wall latency")?;
        self.telemetry
            .record_timed_dispatch(wall_ns, Some(device_ns), None, None);
        Ok(vyre_driver::TimedDispatchResult {
            outputs,
            wall_ns,
            device_ns: Some(device_ns),
            enqueue_ns: None,
            wait_ns: None,
        })
    }

    /// Convenience wrapper that allocates the output `Vec` internally.
    pub fn dispatch_via_cuda_graph(
        &self,
        cached: &mut CachedCudaGraph,
        inputs: &[&[u8]],
    ) -> Result<Vec<Vec<u8>>, BackendError> {
        let mut outputs = reserved_vec(
            cached.output_host_bufs.len(),
            "cuda graph replay output vector",
        )?;
        self.dispatch_via_cuda_graph_into(cached, inputs, &mut outputs)?;
        Ok(outputs)
    }
}

impl CudaGraphReplayStats {
    fn from_cached(cached: &CachedCudaGraph) -> Self {
        Self {
            input_bytes: cached.replay_input_bytes,
            output_bytes: cached.replay_output_bytes,
            host_upload_operations: cached.replay_host_upload_operations,
            device_readback_operations: cached.replay_device_readback_operations,
        }
    }
}

fn prepare_cuda_graph_replay(
    cached: &mut CachedCudaGraph,
    inputs: &[&[u8]],
    input_state: &CudaGraphReplayInputState,
) -> Result<(CudaGraphReplayStats, bool), BackendError> {
    let resident_input_replay = cached.resident_input_replay_safe
        && cached.device_inputs_initialized
        && cached_input_bytes_match_with_key(cached, inputs, &input_state.input_key)?;

    if !resident_input_replay {
        for (slot_index, ((slot, input_index), transfer_len)) in cached
            .input_host_bufs
            .iter_mut()
            .zip(cached.input_indices.iter())
            .zip(cached.input_transfer_lens.iter())
            .enumerate()
        {
            let src = cached_graph_input(inputs, *input_index, slot_index, "input replay staging")?;
            slot.copy_from_slice(src)?;
            if *transfer_len > src.len() {
                slot.zero_range(src.len(), transfer_len - src.len())?;
            }
        }
        cached.cached_input_key = input_state.input_key;
        cached.device_inputs_initialized = false;
        cached.host_outputs_initialized = false;
    }
    let mut stats = CudaGraphReplayStats::from_cached(cached);
    if resident_input_replay {
        stats.input_bytes = 0;
        stats.host_upload_operations = 0;
    }
    Ok((stats, resident_input_replay))
}

fn prepare_cuda_graph_replay_launch(
    cached: &mut CachedCudaGraph,
    inputs: &[&[u8]],
    input_state: &CudaGraphReplayInputState,
) -> Result<PreparedCudaGraphReplayLaunch, BackendError> {
    let (stats, resident_input_replay) = prepare_cuda_graph_replay(cached, inputs, input_state)?;
    Ok(PreparedCudaGraphReplayLaunch {
        stats,
        resident_input_replay,
    })
}

fn launch_prepared_cuda_graph_replay(
    cached: &mut CachedCudaGraph,
    prepared: &PreparedCudaGraphReplayLaunch,
    label: &'static str,
) -> Result<(), BackendError> {
    let graph_exec = if prepared.resident_input_replay {
        &cached.resident_input_graph_exec
    } else {
        &cached.graph_exec
    };
    launch_cuda_graph_exec(graph_exec, &cached.stream, label)
}

fn prepare_cuda_graph_replay_input_state(
    cached: &CachedCudaGraph,
    inputs: &[&[u8]],
) -> Result<CudaGraphReplayInputState, BackendError> {
    prepare_cuda_graph_replay_input_state_with_key(cached, inputs, exact_input_key(inputs)?)
}

fn prepare_cuda_graph_replay_input_state_with_key(
    cached: &CachedCudaGraph,
    inputs: &[&[u8]],
    input_key: ExactInputKey,
) -> Result<CudaGraphReplayInputState, BackendError> {
    validate_cached_graph_inputs(cached, inputs)?;
    Ok(CudaGraphReplayInputState { input_key })
}

fn cached_graph_input<'a>(
    inputs: &[&'a [u8]],
    input_index: usize,
    slot_index: usize,
    context: &'static str,
) -> Result<&'a [u8], BackendError> {
    inputs
        .get(input_index)
        .copied()
        .ok_or_else(|| BackendError::InvalidProgram {
            fix: format!(
                "Fix: cached cuda graph {context} slot {slot_index} maps to logical input {input_index}, but replay received only {} input(s). Re-record the graph from a valid BindingPlan.",
                inputs.len()
            ),
        })
}

fn validate_cached_graph_slot_index_map(
    indices: &[usize],
    expected_len: usize,
    slot_kind: &'static str,
    action: &'static str,
) -> Result<(), BackendError> {
    if indices.len() != expected_len {
        return Err(BackendError::InvalidProgram {
            fix: format!(
                "Fix: cached cuda graph has {} logical {slot_kind} index(es) but {expected_len} expected {slot_kind} length(s). Re-record the graph; descriptor-ordered graph {slot_kind}s must map back to Program::buffers {slot_kind} slots.",
                indices.len(),
            ),
        });
    }
    let mut sorted_indices = SmallVec::<[usize; 8]>::new();
    reserve_smallvec(
        &mut sorted_indices,
        indices.len(),
        "cuda graph slot index validation",
    )?;
    sorted_indices.extend(indices.iter().copied());
    crate::backend::ordering::sort_unstable_if_needed(sorted_indices.as_mut_slice());
    for (expected_index, index) in sorted_indices.iter().copied().enumerate() {
        if index != expected_index {
            return Err(BackendError::InvalidProgram {
                fix: format!(
                    "Fix: cached cuda graph logical {slot_kind} index {index} at sorted slot {expected_index} is not dense over 0..{expected_len}. Re-record the graph from Program::buffers logical {slot_kind} order before {action}.",
                ),
            });
        }
    }
    Ok(())
}

fn validate_cached_graph_input_index_map(
    input_indices: &[usize],
    expected_len: usize,
) -> Result<(), BackendError> {
    validate_cached_graph_slot_index_map(input_indices, expected_len, "input", "replay")
}

fn validate_cached_graph_output_index_map(
    output_indices: &[usize],
    expected_len: usize,
) -> Result<(), BackendError> {
    validate_cached_graph_slot_index_map(output_indices, expected_len, "output", "collection")
}

fn validate_cached_graph_inputs(
    cached: &CachedCudaGraph,
    inputs: &[&[u8]],
) -> Result<(), BackendError> {
    if cached.input_host_bufs.len() != cached.expected_input_lens.len() {
        return Err(BackendError::InvalidProgram {
            fix: format!(
                "Fix: cached cuda graph has {} pinned input buffer(s) but {} expected input length(s). Re-record the graph before replay.",
                cached.input_host_bufs.len(),
                cached.expected_input_lens.len()
            ),
        });
    }
    if cached.input_transfer_lens.len() != cached.expected_input_lens.len() {
        return Err(BackendError::InvalidProgram {
            fix: format!(
                "Fix: cached cuda graph has {} input transfer length(s) but {} expected input length(s). Re-record the graph; zip-based replay would skip or truncate input uploads.",
                cached.input_transfer_lens.len(),
                cached.expected_input_lens.len()
            ),
        });
    }
    validate_cached_graph_input_index_map(&cached.input_indices, cached.expected_input_lens.len())?;
    if inputs.len() != cached.expected_input_lens.len() {
        return Err(BackendError::InvalidProgram {
            fix: format!(
                "Fix: cached cuda graph expects {} inputs but received {}.",
                cached.expected_input_lens.len(),
                inputs.len()
            ),
        });
    }
    for (idx, ((input_index, expected_len), transfer_len)) in cached
        .input_indices
        .iter()
        .zip(cached.expected_input_lens.iter())
        .zip(cached.input_transfer_lens.iter())
        .enumerate()
    {
        let input = cached_graph_input(inputs, *input_index, idx, "shape validation")?;
        let received_len = input.len();
        if received_len != *expected_len {
            return Err(BackendError::InvalidProgram {
                fix: format!(
                    "Fix: cached cuda graph input {idx} expects {expected_len} bytes but \
                     received {}  -  re-record the graph for this input shape.",
                    received_len
                ),
            });
        }
        if *transfer_len < *expected_len {
            return Err(BackendError::InvalidProgram {
                fix: format!(
                    "Fix: cached cuda graph input {idx} expects {expected_len} bytes but its captured transfer length is {transfer_len}. Re-record the graph before replay; truncated graph memcpy would leave stale device input bytes.",
                ),
            });
        }
    }
    Ok(())
}

fn collect_cuda_graph_outputs(
    cached: &CachedCudaGraph,
    outputs: &mut Vec<Vec<u8>>,
) -> Result<(), BackendError> {
    if cached.output_host_bufs.len() != cached.output_lens.len()
        || cached.output_indices.len() != cached.output_lens.len()
    {
        return Err(BackendError::InvalidProgram {
            fix: format!(
                "Fix: cached cuda graph has {} pinned output buffer(s), {} logical output index(es), and {} output length(s). Re-record the graph before collecting outputs.",
                cached.output_host_bufs.len(),
                cached.output_indices.len(),
                cached.output_lens.len()
            ),
        });
    }
    validate_cached_graph_output_index_map(&cached.output_indices, cached.output_lens.len())?;
    resize_vec_slots(
        outputs,
        cached.output_lens.len(),
        "cuda graph replay output vector",
    )?;
    reserve_cuda_graph_output_slots(&cached.output_indices, &cached.output_lens, outputs)?;
    let output_count = outputs.len();
    for (slot_index, (buf, (output_index, byte_len))) in cached
        .output_host_bufs
        .iter()
        .zip(cached.output_indices.iter().zip(cached.output_lens.iter()))
        .enumerate()
    {
        let output = outputs
            .get_mut(*output_index)
            .ok_or_else(|| BackendError::InvalidProgram {
                fix: format!(
                    "Fix: cached cuda graph output slot {slot_index} maps to logical output {output_index}, but collection has only {} output slot(s). Re-record the graph from a valid BindingPlan.",
                    output_count
                ),
            })?;
        buf.copy_prefix_into(*byte_len, output)?;
    }
    Ok(())
}

fn reserve_cuda_graph_output_slots(
    output_indices: &[usize],
    output_lens: &[usize],
    outputs: &mut [Vec<u8>],
) -> Result<(), BackendError> {
    if output_indices.len() != output_lens.len() || output_lens.len() != outputs.len() {
        return Err(BackendError::InvalidProgram {
            fix: format!(
                "Fix: cached cuda graph output preflight expected {} logical output index(es), {} output length(s), and {} caller output slot(s). Re-record the graph before collecting outputs.",
                output_indices.len(),
                output_lens.len(),
                outputs.len()
            ),
        });
    }
    let output_count = outputs.len();
    for (slot_index, (output_index, byte_len)) in
        output_indices.iter().zip(output_lens.iter()).enumerate()
    {
        let output = outputs
            .get_mut(*output_index)
            .ok_or_else(|| BackendError::InvalidProgram {
                fix: format!(
                    "Fix: cached cuda graph output preflight slot {slot_index} maps to logical output {output_index}, but collection has only {} output slot(s). Re-record the graph from a valid BindingPlan.",
                    output_count
                ),
            })?;
        reserve_vec(output, *byte_len, "cuda graph replay output bytes")?;
    }
    Ok(())
}

impl CudaBackend {
    fn record_cuda_graph_replay_stats(&self, stats: CudaGraphReplayStats) {
        self.telemetry
            .record_host_to_device_bytes(stats.input_bytes);
        self.telemetry
            .record_device_to_host_readback(stats.output_bytes);
        self.telemetry
            .record_host_upload_operations(stats.host_upload_operations);
        self.telemetry
            .record_device_readback_operations(stats.device_readback_operations);
    }
}

#[cfg(test)]
mod source_contract_tests {
    use super::{
        cached_graph_input, validate_cached_graph_input_index_map,
        validate_cached_graph_output_index_map,
    };

    #[test]
    fn cached_graph_replay_input_index_map_accepts_reordered_descriptor_inputs() {
        validate_cached_graph_input_index_map(&[2, 0, 1], 3).expect(
            "Fix: descriptor-ordered CUDA graph inputs may map to reordered logical slots.",
        );

        let first = [0xA1, 0xA2];
        let second = [0xB1];
        let third = [0xC1, 0xC2, 0xC3];
        let inputs: &[&[u8]] = &[first.as_slice(), second.as_slice(), third.as_slice()];

        assert_eq!(
            cached_graph_input(inputs, 2, 0, "test replay")
                .expect("Fix: graph replay should resolve logical input 2 for descriptor slot 0."),
            third.as_slice()
        );
        assert_eq!(
            cached_graph_input(inputs, 0, 1, "test replay")
                .expect("Fix: graph replay should resolve logical input 0 for descriptor slot 1."),
            first.as_slice()
        );
    }

    #[test]
    fn cached_graph_replay_input_index_map_rejects_stale_or_non_dense_maps() {
        assert!(
            validate_cached_graph_input_index_map(&[0, 0, 2], 3).is_err(),
            "Fix: duplicate CUDA graph logical input indexes must fail before replay can alias an input slot."
        );
        assert!(
            validate_cached_graph_input_index_map(&[0, 2, 3], 3).is_err(),
            "Fix: sparse CUDA graph logical input indexes must fail before replay can skip an input slot."
        );
        assert!(
            validate_cached_graph_input_index_map(&[0, 1], 3).is_err(),
            "Fix: truncated CUDA graph logical input maps must fail before zip-based replay staging."
        );

        let only = [0xAA];
        let inputs: &[&[u8]] = &[only.as_slice()];
        assert!(
            cached_graph_input(inputs, 1, 0, "test replay").is_err(),
            "Fix: stale CUDA graph logical input indexes must become BackendError, not a panic or wrong-slot replay."
        );
    }

    #[test]
    fn cached_graph_replay_output_index_map_accepts_reordered_descriptor_outputs() {
        validate_cached_graph_output_index_map(&[1, 0, 2], 3).expect(
            "Fix: descriptor-ordered CUDA graph outputs may map to reordered logical slots.",
        );
        assert!(
            validate_cached_graph_output_index_map(&[0, 0, 2], 3).is_err(),
            "Fix: duplicate CUDA graph logical output indexes must fail before collection can alias an output slot."
        );
        assert!(
            validate_cached_graph_output_index_map(&[0, 2, 3], 3).is_err(),
            "Fix: sparse CUDA graph logical output indexes must fail before collection can skip an output slot."
        );
        assert!(
            validate_cached_graph_output_index_map(&[0, 1], 3).is_err(),
            "Fix: truncated CUDA graph logical output maps must fail before positional collection can drop a slot."
        );
    }

    #[test]
    fn cuda_graph_replay_uses_fallible_output_staging_reservation() {
        let source = include_str!("cuda_graph_replay.rs");
        assert!(source.contains(
            "use super::staging_reserve::{reserve_smallvec, reserve_vec, reserved_vec, resize_vec_slots};"
        ));
        assert!(source.contains("fn collect_cuda_graph_outputs("));
        assert!(source.contains(") -> Result<(), BackendError>"));
        assert!(!source.contains(concat!(
            "Vec::with_capacity",
            "(cached.output_host_bufs.len())"
        )));
        assert!(
            source.contains("resize_vec_slots(")
                && !source.contains(concat!("outputs", ".extend("))
                && !source.contains(concat!("outputs", ".resize_with(")),
            "Fix: CUDA graph replay output staging must use the shared fallible resize helper instead of bespoke growth."
        );
        let collector = source
            .split("fn collect_cuda_graph_outputs(")
            .nth(1)
            .and_then(|tail| tail.split("fn reserve_cuda_graph_output_slots(").next())
            .expect("Fix: CUDA graph replay must expose output collection before output-slot preflight.");
        let preflight = collector
            .find(
                "reserve_cuda_graph_output_slots(&cached.output_indices, &cached.output_lens, outputs)?",
            )
            .expect(
                "Fix: CUDA graph replay output collection must preflight every caller output slot.",
            );
        let output_lookup = collector
            .find(".get_mut(*output_index)")
            .expect("Fix: CUDA graph replay output collection must route by logical output index.");
        let copy = collector
            .find("buf.copy_prefix_into(*byte_len, output)?")
            .expect("Fix: CUDA graph replay output collection must copy pinned graph outputs.");
        assert!(
            preflight < output_lookup && output_lookup < copy,
            "Fix: CUDA graph replay must reserve every caller output before copying any pinned graph output bytes."
        );
        let preflight_helper = source
            .split("fn reserve_cuda_graph_output_slots(")
            .nth(1)
            .and_then(|tail| tail.split("impl CudaBackend").next())
            .expect("Fix: CUDA graph replay must expose output-slot preflight before backend telemetry.");
        assert!(
            preflight_helper.contains("output_indices.len() != output_lens.len()")
                && preflight_helper.contains(".get_mut(*output_index)")
                && preflight_helper.contains("reserve_vec(output, *byte_len, \"cuda graph replay output bytes\")?"),
            "Fix: CUDA graph replay output preflight must validate cardinality, route by logical output index, and reserve every destination byte capacity."
        );
        assert!(
            source.contains("cached.input_host_bufs.len() != cached.expected_input_lens.len()")
                && source.contains("cached.input_transfer_lens.len() != cached.expected_input_lens.len()")
                && source.contains("validate_cached_graph_input_index_map(&cached.input_indices")
                && source.contains("cached_graph_input(inputs, *input_index")
                && source.contains("descriptor-ordered graph inputs must map back to Program::buffers input slots")
                && source.contains("zip-based replay would skip or truncate input uploads")
                && source.contains("*transfer_len < *expected_len")
                && source.contains("truncated graph memcpy would leave stale device input bytes")
                && source.contains("zip-based replay would skip input uploads")
                && source.contains(".zip(cached.input_indices.iter())")
                && source.contains(".zip(cached.input_transfer_lens.iter())")
                && !source.contains(concat!("inputs", "[idx]"))
                && source.contains("cached.output_host_bufs.len() != cached.output_lens.len()")
                && source.contains("cached.output_indices.len() != cached.output_lens.len()")
                && source.contains("validate_cached_graph_output_index_map(&cached.output_indices")
                && source.contains(".zip(cached.output_indices.iter().zip(cached.output_lens.iter()))"),
            "Fix: CUDA graph replay must validate cached graph input/output metadata before zip-based staging."
        );
        assert_eq!(
            source
                .matches(concat!("cudarc::driver::sys::", "cuGraphLaunch("))
                .count(),
            1,
            "Fix: CUDA graph replay must keep raw cuGraphLaunch behind one checked helper."
        );
        assert!(
            source.contains("fn launch_cuda_graph_exec(")
                && source.contains("dangling CUgraphExec sentinel")
                && source.contains("dangling CUstream sentinel"),
            "Fix: CUDA graph replay launch helper must validate graph and stream handles before FFI."
        );
    }

    #[test]
    fn timed_and_untimed_graph_replay_share_resident_input_skip_copy_path() {
        let source = include_str!("cuda_graph_replay.rs");
        assert!(
            source.contains("fn prepare_cuda_graph_replay(")
                && source.matches("prepare_cuda_graph_replay(cached, inputs,").count() >= 2,
            "Fix: timed and untimed CUDA graph replay must share one resident-input preparation path."
        );
        assert!(
            source.contains("fn prepare_cuda_graph_replay_launch(")
                && source.contains("fn launch_prepared_cuda_graph_replay(")
                && source
                    .matches(
                        "launch_prepared_cuda_graph_replay(cached, &prepared, \"cuGraphLaunch\")"
                    )
                    .count()
                    == 2,
            "Fix: timed and untimed CUDA graph replay must share prepared launch graph selection."
        );
        let launch_helper = source
            .split("fn launch_prepared_cuda_graph_replay(")
            .nth(1)
            .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - prepared CUDA graph launch helper must exist")
            .split("fn prepare_cuda_graph_replay_input_state(")
            .next()
            .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - prepared launch helper must precede input-state preparation");
        assert!(
            !launch_helper.contains("cached.device_inputs_initialized = true;"),
            "Fix: CUDA graph replay must not mark device inputs initialized immediately after cuGraphLaunch; the stream/timing fence must complete first."
        );
        assert!(
            source.contains("synchronize_cuda_graph_replay_stream(cached)?;\n        cached.device_inputs_initialized = true;")
                && source.contains("end.synchronize()?;")
                && source.contains("timing_events.mark_synchronized();\n        cached.device_inputs_initialized = true;"),
            "Fix: CUDA graph replay must mark resident device inputs initialized only after successful untimed and timed completion fences."
        );
        let timed_section = source
            .split("pub(crate) fn dispatch_via_cuda_graph_timed_into")
            .nth(1)
            .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - timed CUDA graph replay entrypoint must exist")
            .split("/// Convenience wrapper")
            .next()
            .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - timed replay section must precede convenience wrapper");
        assert!(
            timed_section.contains("prepare_cuda_graph_replay_launch(cached, inputs, &input_state)?")
                && timed_section.contains("launch_prepared_cuda_graph_replay(cached, &prepared, \"cuGraphLaunch\")")
                && !timed_section.contains("for (slot, src) in cached.input_host_bufs"),
            "Fix: timed CUDA graph replay must use resident-input graph replay when safe instead of always copying host inputs."
        );
        let timed_sync_pos = timed_section
            .find("end.synchronize()?;")
            .expect("Fix: timed CUDA graph replay must fence the end timing event before trusting resident inputs.");
        let timing_release_pos = timed_section
            .find("timing_events.mark_synchronized();")
            .expect("Fix: timed CUDA graph replay must mark timing events reusable only after end-event synchronization.");
        let initialized_pos = timed_section
            .find("cached.device_inputs_initialized = true;")
            .expect("Fix: timed CUDA graph replay must promote resident inputs after completion is proven.");
        assert!(
            timed_sync_pos < timing_release_pos && timing_release_pos < initialized_pos,
            "Fix: timed CUDA graph replay must prove timing-event completion before timing lease reuse or resident-input promotion."
        );
    }

    #[test]
    fn materialized_graph_cache_is_shared_by_single_and_batched_replay_paths() {
        let replay_source = include_str!("cuda_graph_replay.rs");
        let compiled_dispatch = include_str!("../pipeline/compiled_dispatch.rs");
        assert!(
            replay_source
                .contains("pub(crate) fn try_cuda_graph_materialized_cache_with_input_state_into(")
                && replay_source.contains(
                    "pub(crate) fn dispatch_via_cuda_graph_with_input_state_into("
                )
                && replay_source.contains(
                    "pub(crate) fn dispatch_via_cuda_graph_timed_with_input_state_into("
                ),
            "Fix: single CUDA graph replay must route materialized output cache hits through the shared helper."
        );
        assert!(
            compiled_dispatch.contains("materialized_output_batch_cache_partition_into")
                && compiled_dispatch.contains("let miss_entries =")
                && compiled_dispatch.contains("for (chunk_index, chunk) in miss_entries.chunks(lane_count).enumerate()")
                && compiled_dispatch.contains("chunk_index")
                && compiled_dispatch.contains(".checked_mul(lane_count)")
                && compiled_dispatch.contains("prepare_cuda_graph_replay_input_state")
                && compiled_dispatch.contains("try_cuda_graph_materialized_cache_with_input_state_into")
                && compiled_dispatch.contains("enqueue_cuda_graph_replay_with_input_state")
                && compiled_dispatch.contains("continue;")
                && compiled_dispatch.contains("[LaunchedMaterializedBatch; MAX_GRAPH_CACHE_ENTRIES_PER_PIPELINE]")
                && compiled_dispatch.contains("input_key: miss.input_key"),
            "Fix: batched CUDA graph replay must partition materialized exact-input cache hits before lane planning, reuse precomputed input keys, and only finish lanes that actually launched."
        );
    }

    #[test]
    fn cached_graph_input_key_gates_byte_compare_and_rewrites_invalidate_host_outputs() {
        let replay_source = include_str!("cuda_graph_replay.rs");
        let graph_source = include_str!("cuda_graph.rs");
        assert!(
            replay_source.contains("use crate::input_identity::{exact_input_key, ExactInputKey};")
                && replay_source.contains("fn cached_input_bytes_match_with_key(")
                && replay_source.contains("if cached.cached_input_key != *input_key")
                && replay_source.contains("cached_input_bytes_match_after_key_match"),
            "Fix: raw CUDA graph exact-input checks must use the shared tuple key as a fast reject before expensive pinned-host byte comparison."
        );
        assert!(
            replay_source.contains("let input_key = exact_input_key(inputs)?;")
                && replay_source.contains("cached.cached_input_key = input_state.input_key;")
                && replay_source.contains("cached.device_inputs_initialized = false;")
                && replay_source.contains("cached.host_outputs_initialized = false;"),
            "Fix: rewriting cached graph host inputs must update the exact-input key and immediately invalidate resident device inputs plus materialized host outputs before graph launch/finish can fail."
        );
        let prepare_replay = replay_source
            .split("fn prepare_cuda_graph_replay(")
            .nth(1)
            .expect("Fix: CUDA graph replay preparation must stay centralized.")
            .split("fn prepare_cuda_graph_replay_launch(")
            .next()
            .expect("Fix: replay preparation must precede prepared-launch construction.");
        let key_update = prepare_replay
            .find("cached.cached_input_key = input_state.input_key;")
            .expect("Fix: replay preparation must update the cached input key after rewriting host inputs.");
        let device_invalidate = prepare_replay
            .find("cached.device_inputs_initialized = false;")
            .expect("Fix: replay preparation must invalidate resident device inputs before launch can fail.");
        let host_invalidate = prepare_replay
            .find("cached.host_outputs_initialized = false;")
            .expect("Fix: replay preparation must invalidate materialized host outputs before launch can fail.");
        assert!(
            key_update < device_invalidate && device_invalidate < host_invalidate,
            "Fix: rewritten CUDA graph inputs must invalidate resident device state before the graph can be re-used after an enqueue failure."
        );
        assert!(
            graph_source.contains("pub(crate) cached_input_key: ExactInputKey")
                && graph_source.contains("let cached_input_key = exact_input_key(sample_inputs)?;"),
            "Fix: recorded CUDA graphs must initialize cached_input_key from the captured sample inputs."
        );
    }

    #[test]
    fn raw_graph_replay_prepares_input_state_once_per_dispatch_path() {
        let replay_source = include_str!("cuda_graph_replay.rs");
        assert!(
            replay_source.contains("pub(crate) struct CudaGraphReplayInputState")
                && replay_source.contains("fn prepare_cuda_graph_replay_input_state(")
                && replay_source.contains("validate_cached_graph_inputs(cached, inputs)?;")
                && replay_source.contains("input_key: exact_input_key(inputs)?"),
            "Fix: CUDA graph replay must centralize shape validation and exact-input key creation in a reusable input-state object."
        );
        let untimed_section = replay_source
            .split("pub fn dispatch_via_cuda_graph_into")
            .nth(1)
            .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - untimed CUDA graph replay entrypoint must exist")
            .split("/// Replay a cached CUDA graph with CUDA event timing.")
            .next()
            .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - untimed replay section must precede timed replay");
        assert_eq!(
            untimed_section
                .matches("prepare_cuda_graph_replay_input_state(cached, inputs)?")
                .count(),
            1,
            "Fix: untimed raw CUDA graph replay must validate/hash inputs once and reuse that state for materialized-cache check plus launch preparation."
        );
        assert!(
            untimed_section.contains("try_cuda_graph_materialized_cache_with_input_state_into")
                && untimed_section.contains("enqueue_cuda_graph_replay_with_input_state"),
            "Fix: untimed raw CUDA graph replay must pass the prepared input state through both cache and launch paths."
        );
        assert!(
            replay_source.contains("dispatch_via_cuda_graph_with_input_state_into(cached, inputs, &input_state, outputs)")
                && replay_source.contains("pub(crate) fn dispatch_via_cuda_graph_with_input_state_into("),
            "Fix: untimed raw CUDA graph replay must expose a with-input-state entrypoint so compiled pipelines can reuse precomputed exact-input keys."
        );
        let timed_section = replay_source
            .split("pub(crate) fn dispatch_via_cuda_graph_timed_into")
            .nth(1)
            .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - timed CUDA graph replay entrypoint must exist")
            .split("/// Replay a cached CUDA graph with CUDA event timing and allocated outputs.")
            .next()
            .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - timed replay section must precede timed wrapper");
        assert_eq!(
            timed_section
                .matches("prepare_cuda_graph_replay_input_state(cached, inputs)?")
                .count(),
            1,
            "Fix: timed raw CUDA graph replay must validate/hash inputs once and reuse that state before event timing."
        );
        assert!(
            timed_section.contains("try_cuda_graph_materialized_cache_with_input_state_into")
                && timed_section.contains("prepare_cuda_graph_replay_launch(cached, inputs, &input_state)?")
                && timed_section.contains("launch_prepared_cuda_graph_replay(cached, &prepared, \"cuGraphLaunch\")"),
            "Fix: timed raw CUDA graph replay must reuse the prepared input state for materialized and resident-input replay decisions."
        );
        assert!(
            replay_source.contains("dispatch_via_cuda_graph_timed_with_input_state_into(\n            cached,\n            inputs,\n            &input_state,\n            outputs,\n        )")
                && replay_source.contains("pub(crate) fn dispatch_via_cuda_graph_timed_with_input_state_into("),
            "Fix: timed raw CUDA graph replay must expose a with-input-state entrypoint so compiled pipelines can reuse precomputed exact-input keys."
        );
    }

    #[test]
    fn compiled_batch_graph_misses_reuse_materialized_cache_input_keys() {
        let replay_source = include_str!("cuda_graph_replay.rs");
        let compiled_dispatch = include_str!("../pipeline/compiled_dispatch.rs");
        assert!(
            replay_source.contains("prepare_cuda_graph_replay_input_state_with_key")
                && replay_source.contains("input_key: ExactInputKey")
                && replay_source.contains("validate_cached_graph_inputs(cached, inputs)?;")
                && replay_source.contains("Ok(CudaGraphReplayInputState { input_key })"),
            "Fix: raw CUDA graph replay must accept a precomputed exact-input key while still validating graph shape."
        );
        assert!(
            compiled_dispatch.contains("struct MaterializedBatchMiss")
                && compiled_dispatch.contains("input_key: MaterializedInputKey")
                && compiled_dispatch.contains("materialized_input_key(inputs)?")
                && compiled_dispatch.contains("cache.snapshot_with_key(inputs, &input_key)")
                && compiled_dispatch.contains("prepare_cuda_graph_replay_input_state_with_key")
                && compiled_dispatch.contains("take_cached_graph_with_key(")
                && compiled_dispatch.contains("&first_miss.input_key")
                && compiled_dispatch.contains("miss.input_key"),
            "Fix: compiled batched CUDA graph replay must reuse materialized-cache exact-input keys for graph-cache selection and graph miss replay instead of hashing each miss again."
        );
        let partition_section = compiled_dispatch
            .split("fn materialized_output_batch_cache_partition_into")
            .nth(1)
            .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - compiled materialized batch partition function must exist")
            .split("fn materialized_output_cache_hit_into")
            .next()
            .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - batch partition section must precede single cache helper");
        let key_position = partition_section
            .find("materialized_input_key(inputs)?")
            .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - batch partition must compute exact-input keys");
        let lock_position = partition_section
            .find("let cache = self.lock_materialized_output_cache")
            .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - batch partition must acquire materialized cache lock");
        let resize_position = partition_section
            .find("resize_vec_slots(\n            outputs,\n            batches.len(),\n            \"cuda graph materialized batch output\",")
            .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - batch partition must resize output slots before hit copy");
        let hit_copy_position = partition_section
            .find("snapshot.copy_into(compiled_graph_output_mut(")
            .expect("Fix: replace expect with fallible API or document caller precondition; panic only on programmer error - batch partition must copy materialized cache hits");
        assert!(
            partition_section.contains("for (batch_index, inputs) in batches.iter().enumerate()")
                && partition_section.contains("input_keys.push((batch_index, materialized_input_key(inputs)?));")
                && partition_section.contains("let cache = self.lock_materialized_output_cache")
                && key_position < lock_position,
            "Fix: compiled materialized batch replay must compute exact-input keys before acquiring the materialized-output cache lock."
        );
        assert!(
            key_position < resize_position
                && lock_position < resize_position
                && resize_position < hit_copy_position,
            "Fix: compiled materialized batch replay must finish exact-input key and cache-snapshot partitioning before resizing caller-owned output slots, then resize before copying cache hits."
        );
    }
}