File size: 24,572 Bytes
fa05bf1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
diff --git a/python/ray/data/_internal/block_batching/pinned_staging.py b/python/ray/data/_internal/block_batching/pinned_staging.py
new file mode 100644
index 0000000000..8a3c285f02
--- /dev/null
+++ b/python/ray/data/_internal/block_batching/pinned_staging.py
@@ -0,0 +1,219 @@
+"""One-batch pinned-memory prefetch for synchronous map_batches GPU actors."""
+
+from concurrent.futures import Future, ThreadPoolExecutor
+from threading import RLock
+
+
+class PinnedPrefetch:
+    """Stage concrete NumPy batches on one producer thread, in FIFO order.
+
+    Only the consumer advances the upstream iterator. Each returned dictionary
+    owns fresh CUDA storage on actor-local cuda:0. There is at most one pending
+    batch in addition to the batch held by the consumer.
+
+    staging_collate_fn, if supplied, receives owned, writable NumPy arrays on
+    the producer thread. It must return a nonempty dict of numeric NumPy arrays
+    or dense CPU tensors. It must not capture the actor/model, launch CUDA work,
+    or call this iterator's methods.
+
+    Consume on one thread and use the returned tensors on that thread's
+    current CUDA stream. A UDF using additional streams must manage their
+    synchronization and record_stream calls itself. Call close() on early exit;
+    the map-task ExitStack and transform generator both do this.
+    """
+
+    def __init__(self, batches, staging_collate_fn=None):
+        if staging_collate_fn is not None and not callable(staging_collate_fn):
+            raise TypeError("staging_collate_fn must be callable or None")
+
+        # Importing this module does not import or initialize torch.
+        import torch
+
+        self._torch = torch
+        self._batches = iter(batches)
+        self._collate_fn = staging_collate_fn
+        self._device = torch.device("cuda", 0)
+        self._stream = torch.cuda.Stream(device=self._device)
+        self._pool = ThreadPoolExecutor(
+            max_workers=1, thread_name_prefix="ray-pinned-staging"
+        )
+        self._pending = None
+        self._started = False
+        self._exhausted = False
+        self._closed = False
+        # Serialize close against next and other close calls. The producer
+        # never acquires this lock; joining it while holding the lock is safe.
+        self._lock = RLock()
+
+    def __iter__(self):
+        return self
+
+    def _submit_next(self):
+        """Called only by the consumer, with the iterator lock held."""
+        if self._exhausted:
+            return None
+        try:
+            batch = next(self._batches)
+        except StopIteration:
+            self._exhausted = True
+            return None
+        except Exception as error:
+            self._exhausted = True
+            failed = Future()
+            failed.set_exception(error)
+            return failed
+
+        try:
+            # Never pass the iterator or next(upstream) to the executor.
+            return self._pool.submit(self._stage, batch)
+        except Exception as error:
+            # Even submission failure during lookahead must preserve batch N.
+            self._exhausted = True
+            failed = Future()
+            failed.set_exception(error)
+            return failed
+
+    @staticmethod
+    def _check_batch(batch):
+        if not isinstance(batch, dict) or not batch:
+            raise TypeError("Pinned staging requires a nonempty dict")
+        if not all(isinstance(key, str) for key in batch):
+            raise TypeError("Pinned staging requires string column names")
+
+    @staticmethod
+    def _copy_array(value):
+        import numpy as np
+
+        if not isinstance(value, np.ndarray) or value.dtype.kind not in "biufc":
+            raise TypeError("Pinned staging requires numeric or boolean arrays")
+        # An owned, writable, native-endian, C-contiguous copy also handles
+        # read-only object-store views and arrays with negative strides.
+        return np.array(
+            value, dtype=value.dtype.newbyteorder("="), order="C", copy=True
+        )
+
+    def _stage(self, batch):
+        import numpy as np
+
+        torch = self._torch
+        host, device = {}, {}
+        with torch.cuda.device(self._device), torch.no_grad():
+            try:
+                with torch.cuda.nvtx.range("ray::collate_pin"):
+                    self._check_batch(batch)
+                    if self._collate_fn is not None:
+                        batch = self._collate_fn(
+                            {
+                                key: self._copy_array(value)
+                                for key, value in batch.items()
+                            }
+                        )
+                        self._check_batch(batch)
+
+                    for key, value in batch.items():
+                        if isinstance(value, np.ndarray):
+                            cpu = torch.from_numpy(self._copy_array(value))
+                        elif (
+                            isinstance(value, torch.Tensor)
+                            and value.device.type == "cpu"
+                            and value.layout == torch.strided
+                            and not value.is_quantized
+                            and not value.is_nested
+                        ):
+                            # pin_memory() may alias an already-pinned tensor.
+                            # Start with fresh pageable storage even if the
+                            # collator keeps or reuses its pinned output.
+                            cpu = torch.empty(
+                                tuple(value.shape),
+                                dtype=value.dtype,
+                                device="cpu",
+                                pin_memory=False,
+                            )
+                            cpu.copy_(value.detach())
+                        else:
+                            raise TypeError(
+                                f"Column {key!r}: expected a numeric NumPy "
+                                "array or a dense, non-quantized CPU tensor"
+                            )
+                        host[key] = cpu.pin_memory()
+
+                with torch.cuda.stream(self._stream):
+                    with torch.cuda.nvtx.range("ray::H2D"):
+                        for key, pinned in host.items():
+                            device[key] = pinned.to(
+                                self._device, non_blocking=True
+                            )
+                        ready = torch.cuda.Event()
+                        ready.record(self._stream)
+
+                # The future completes only AFTER H2D completes. This wait is
+                # on the producer thread; compute for batch N can continue
+                # while the producer stages N+1. Keep all pinned sources alive.
+                ready.synchronize()
+                return device, ready
+            except BaseException as error:
+                # A later column's copy, event creation, or event recording
+                # can fail after earlier H2D work has already been issued.
+                # Drain before releasing any pinned source.
+                self._stream.synchronize()
+                if isinstance(error, StopIteration):
+                    raise RuntimeError(
+                        "staging_collate_fn raised StopIteration"
+                    ) from error
+                raise
+            finally:
+                host.clear()
+
+    def __next__(self):
+        with self._lock:
+            if self._closed:
+                raise StopIteration
+            try:
+                if not self._started:
+                    self._started = True
+                    self._pending = self._submit_next()
+                if self._pending is None:
+                    self.close()
+                    raise StopIteration
+
+                # result() waits for the producer's ready.synchronize().
+                device, ready = self._pending.result()
+                self._pending = None
+                stream = self._torch.cuda.current_stream(self._device)
+                stream.wait_event(ready)
+                for value in device.values():
+                    value.record_stream(stream)
+
+                # Start N+1 before handing N to the UDF. An upstream exception
+                # is stored in a Future and raised on the next consumption.
+                self._pending = self._submit_next()
+                return device
+            except BaseException:
+                self.close()
+                raise
+
+    def close(self):
+        """Join staging, drain issued H2D, and release unconsumed resources.
+
+        Idempotent, including calls through both finally and ExitStack. Does
+        not advance upstream or surface errors from unconsumed lookahead.
+        Returned CUDA tensors remain owned by the UDF and are never overwritten.
+        """
+        with self._lock:
+            if self._closed:
+                return
+            self._closed = True
+            try:
+                if self._pending is not None:
+                    self._pending.cancel()
+                # Running jobs cannot be cancelled. _stage drains both normal
+                # and partially-issued H2D before returning or raising, so
+                # joining the producer also drains all issued copies.
+                self._pool.shutdown(wait=True, cancel_futures=True)
+            finally:
+                self._pending = None
+                self._batches = iter(())
+                self._collate_fn = None
+                self._pool = None
+                self._stream = None
+                self._torch = None
diff --git a/python/ray/data/_internal/compute.py b/python/ray/data/_internal/compute.py
index d8eb354c0e..e120814bd4 100644
--- a/python/ray/data/_internal/compute.py
+++ b/python/ray/data/_internal/compute.py
@@ -116,6 +116,8 @@ class ActorPoolStrategy(ComputeStrategy):
         max_tasks_in_flight_per_actor: Optional[int] = None,
         max_concurrent_calls_per_actor: Optional[int] = None,
         enable_true_multi_threading: Optional[bool] = None,
+        pinned_staging: bool = False,
+        staging_collate_fn: Optional[Callable] = None,
     ):
         """Construct ActorPoolStrategy for a Dataset transform.
 
@@ -137,6 +139,16 @@ class ActorPoolStrategy(ComputeStrategy):
                 than 1 UDF runs per actor. Otherwise, respects the `max_concurrent_calls_per_actor` argument.
                 By default, this flag is `None`, which gets translated to `False`.
                 For more details, see the `ActorPoolStrategy` class docstring.
+            pinned_staging: Experimental map_batches-only CUDA staging. Defaults
+                to False. Requires num_gpus=1, a positive integer batch_size,
+                batch_format="numpy", zero_copy_batch=False and a synchronous
+                UDF. Sets actor call concurrency to 1 and disables fusion.
+                The UDF receives a dict of CUDA tensors on local cuda:0 instead
+                of NumPy arrays. Prefetches one batch within each actor task.
+            staging_collate_fn: Optional deterministic CPU-only callable,
+                executed on the staging thread with an owned NumPy batch.
+                Return a nonempty dict of dense CPU tensors or numeric arrays.
+                Must not capture the actor/model or launch CUDA work.
         """
         if size is not None:
             if size < 1:
@@ -172,6 +184,19 @@ class ActorPoolStrategy(ComputeStrategy):
                 max_concurrent_calls_per_actor,
             )
 
+        if staging_collate_fn is not None and (
+            not pinned_staging or not callable(staging_collate_fn)
+        ):
+            raise ValueError("staging_collate_fn requires pinned_staging=True")
+        if pinned_staging:
+            if enable_true_multi_threading or max_concurrent_calls_per_actor not in (
+                None, 1
+            ):
+                raise ValueError("pinned_staging requires one synchronous actor call")
+            max_concurrent_calls_per_actor = 1
+        self.pinned_staging = pinned_staging
+        self.staging_collate_fn = staging_collate_fn
+
         self.min_size = min_size or 1
         self.max_size = max_size or float("inf")
 
@@ -213,6 +238,10 @@ class ActorPoolStrategy(ComputeStrategy):
             and self.max_size == other.max_size
             and self.initial_size == other.initial_size
             and self.enable_true_multi_threading == other.enable_true_multi_threading
+            and getattr(self, "pinned_staging", False)
+            == getattr(other, "pinned_staging", False)
+            and getattr(self, "staging_collate_fn", None)
+            == getattr(other, "staging_collate_fn", None)
             and self.max_tasks_in_flight_per_actor
             == other.max_tasks_in_flight_per_actor
             and self.max_concurrent_calls_per_actor
@@ -226,6 +255,7 @@ class ActorPoolStrategy(ComputeStrategy):
             f"initial_size={self.initial_size}, "
             f"max_tasks_in_flight_per_actor={self.max_tasks_in_flight_per_actor}, "
             f"max_concurrent_calls_per_actor={self.max_concurrent_calls_per_actor}, "
+            f"pinned_staging={getattr(self, 'pinned_staging', False)}, "
             f"num_workers={self.num_workers}, "
             f"enable_true_multi_threading={self.enable_true_multi_threading}, "
             f"ready_to_total_workers_ratio={self.ready_to_total_workers_ratio})"
diff --git a/python/ray/data/_internal/execution/interfaces/task_context.py b/python/ray/data/_internal/execution/interfaces/task_context.py
index 35ce6506ea..1701595883 100644
--- a/python/ray/data/_internal/execution/interfaces/task_context.py
+++ b/python/ray/data/_internal/execution/interfaces/task_context.py
@@ -47,6 +47,11 @@ class TaskContext:
     # Override of the target max-block-size for the task
     target_max_block_size_override: Optional[int] = None
 
+    # Worker-local cleanup, installed by _map_task (never sent from the driver).
+    _pinned_staging_cleanup: Optional[contextlib.ExitStack] = field(
+        default=None, init=False, repr=False, compare=False
+    )
+
     # Additional keyword arguments passed to the task.
     kwargs: Dict[str, Any] = field(default_factory=dict)
 
diff --git a/python/ray/data/_internal/execution/operators/map_operator.py b/python/ray/data/_internal/execution/operators/map_operator.py
index 818c62cf31..dcd43c335c 100644
--- a/python/ray/data/_internal/execution/operators/map_operator.py
+++ b/python/ray/data/_internal/execution/operators/map_operator.py
@@ -6,6 +6,7 @@ import logging
 import math
 import time
 from abc import ABC, abstractmethod
+from contextlib import ExitStack
 from dataclasses import replace
 from typing import (
     TYPE_CHECKING,
@@ -818,7 +819,12 @@ def _map_task(
 
     ctx.kwargs.update(kwargs)
 
-    with DataContext.current(data_context), TaskContext.current(ctx):
+    with (
+        DataContext.current(data_context),
+        TaskContext.current(ctx),
+        ExitStack() as staging_cleanup,
+    ):
+        ctx._pinned_staging_cleanup = staging_cleanup
         map_transformer.override_target_max_block_size(
             ctx.target_max_block_size_override
         )
@@ -843,6 +849,9 @@ def _map_task(
         udf_time_scope = UDFTimeScope()
 
         def transform_iter_factory():
+            # Close the previous attempt before creating a fresh producer.
+            # The outer ExitStack also handles cancellation/GeneratorExit.
+            staging_cleanup.close()
             # Clear any per-task custom stats before each attempt (the reporter
             # is reused across retries of this task), so a prior attempt's stats
             # can't leak into this one. A producing transform repopulates it
diff --git a/python/ray/data/_internal/planner/plan_udf_map_op.py b/python/ray/data/_internal/planner/plan_udf_map_op.py
index 8d0af2b4d4..9d5e740a78 100644
--- a/python/ray/data/_internal/planner/plan_udf_map_op.py
+++ b/python/ray/data/_internal/planner/plan_udf_map_op.py
@@ -297,6 +297,32 @@ def plan_udf_map_op(
     )
 
     compute = get_compute(op.compute)
+    pinned_staging = getattr(compute, "pinned_staging", False)
+    if pinned_staging:
+        from ray.data._internal.utils.torch_inference import (
+            _BaseTorchInferenceUDFWrapper,
+        )
+
+        user_fn = op.fn.__call__ if isinstance(op.fn, CallableClass) else op.fn
+        if (
+            not isinstance(op, MapBatches)
+            or op.zero_copy_batch
+            or op.batch_format != "numpy"
+            or type(op.batch_size) is not int
+            or op.batch_size < 1
+            or op.ray_remote_args.get("num_gpus") != 1
+            or op.ray_remote_args_fn is not None
+            or _is_async_udf(user_fn)
+            or (
+                isinstance(op.fn, type)
+                and issubclass(op.fn, _BaseTorchInferenceUDFWrapper)
+            )
+        ):
+            raise ValueError(
+                "pinned_staging requires synchronous map_batches, "
+                "zero_copy_batch=False, batch_format='numpy', integer batch_size, "
+                "num_gpus=1, and no TorchInference wrapper or ray_remote_args_fn"
+            )
     udf_is_callable_class = isinstance(op.fn, CallableClass)
     fn, init_fn = _get_udf(
         op.fn,
@@ -309,10 +335,15 @@ def plan_udf_map_op(
 
     if isinstance(op, MapBatches):
         transform_fn = BatchMapTransformFn(
-            _generate_transform_fn_for_map_batches(fn),
+            _generate_transform_fn_for_map_batches(
+                fn, pinned_staging=pinned_staging,
+                staging_collate_fn=getattr(compute, "staging_collate_fn", None),
+            ),
             batch_size=op.batch_size,
             batch_format=op.batch_format,
-            zero_copy_batch=op.zero_copy_batch,
+            # Private CPU views are never given to the UDF. Staging makes
+            # owned pinned/device copies, and copies collator input if needed.
+            zero_copy_batch=True if pinned_staging else op.zero_copy_batch,
             is_udf=True,
             output_block_size_option=output_block_size_option,
         )
@@ -343,6 +374,7 @@ def plan_udf_map_op(
         ray_remote_args_fn=op.ray_remote_args_fn,
         ray_remote_args=op.ray_remote_args,
         per_block_limit=op.per_block_limit,
+        supports_fusion=not pinned_staging,
     )
 
 
@@ -376,6 +408,9 @@ def _get_udf(
             not is_async_udf
             and isinstance(compute, ActorPoolStrategy)
             and not compute.enable_true_multi_threading
+            # Staging fixes actor call concurrency at one; keep the UDF on
+            # the thread/stream where staged inputs are handed off.
+            and not getattr(compute, "pinned_staging", False)
         ):
             # NOTE: By default Actor-based UDFs are restricted to run within a
             # single-thread (when enable_true_multi_threading=False).
@@ -621,9 +656,25 @@ class _TransformingBatchIterator(Iterator[DataBatch]):
 
 def _generate_transform_fn_for_map_batches(
     fn: UserDefinedFunction,
+    *,
+    pinned_staging: bool = False,
+    staging_collate_fn: Optional[Callable] = None,
 ) -> MapTransformCallable[DataBatch, DataBatch]:
 
-    if _is_async_udf(fn):
+    if pinned_staging:
+        def transform_fn(batches, ctx):
+            from ray.data._internal.block_batching.pinned_staging import PinnedPrefetch
+
+            if ctx._pinned_staging_cleanup is None:
+                raise RuntimeError("Pinned staging requires a map task cleanup scope")
+            staged = PinnedPrefetch(batches, staging_collate_fn)
+            ctx._pinned_staging_cleanup.callback(staged.close)
+            try:
+                yield from _TransformingBatchIterator(staged, fn)
+            finally:
+                staged.close()
+
+    elif _is_async_udf(fn):
         transform_fn = _generate_transform_fn_for_async_map(
             fn,
             _validate_batch_output,
diff --git a/python/ray/data/tests/block_batching/test_pinned_staging.py b/python/ray/data/tests/block_batching/test_pinned_staging.py
new file mode 100644
index 0000000000..4718d0ffb6
--- /dev/null
+++ b/python/ray/data/tests/block_batching/test_pinned_staging.py
@@ -0,0 +1,133 @@
+import sys
+import threading
+from concurrent.futures import ThreadPoolExecutor
+from contextlib import ExitStack
+from types import SimpleNamespace
+
+import pytest
+
+from ray.data._internal.block_batching.pinned_staging import PinnedPrefetch
+
+
+@pytest.fixture
+def fake_prefetch(monkeypatch):
+    stream = SimpleNamespace(wait_event=lambda event: None)
+    torch = SimpleNamespace(
+        device=lambda *args: "cuda:0",
+        cuda=SimpleNamespace(
+            Stream=lambda **kwargs: stream,
+            current_stream=lambda device: stream,
+        ),
+    )
+    monkeypatch.setitem(sys.modules, "torch", torch)
+
+    class Prefetch(PinnedPrefetch):
+        def _stage(self, batch):
+            return {
+                "x": SimpleNamespace(value=batch, record_stream=lambda stream: None)
+            }, None
+
+    return Prefetch
+
+
+def test_fifo_thread_boundary_and_lookahead(fake_prefetch):
+    owner = threading.get_ident()
+    produced = []
+    second_started = threading.Event()
+
+    class Prefetch(fake_prefetch):
+        def _stage(self, batch):
+            assert threading.get_ident() != owner
+            produced.append(batch)
+            if batch == 1:
+                second_started.set()
+            return super()._stage(batch)
+
+    def upstream():
+        for i in range(4):
+            assert threading.get_ident() == owner
+            yield i
+
+    p = Prefetch(upstream())
+    try:
+        assert next(p)["x"].value == 0
+        # N+1 staging is running while the consumer still owns N.
+        assert second_started.wait(3)
+        assert [b["x"].value for b in p] == [1, 2, 3]
+        assert produced == [0, 1, 2, 3]
+    finally:
+        p.close()
+
+
+def test_lookahead_error_preserves_good_batch(fake_prefetch):
+    def upstream():
+        yield 7
+        raise ValueError("upstream failed")
+
+    p = fake_prefetch(upstream())
+    assert next(p)["x"].value == 7
+    with pytest.raises(ValueError, match="upstream failed"):
+        next(p)
+    assert p._closed
+
+
+def test_cleanup_joins_running_producer_before_retry(fake_prefetch):
+    started, release, drained = (threading.Event() for _ in range(3))
+
+    class Prefetch(fake_prefetch):
+        def _stage(self, batch):
+            if batch == 1:
+                started.set()
+                assert release.wait(3)
+                drained.set()
+            return super()._stage(batch)
+
+    p = Prefetch(iter([0, 1, 2]))
+    cleanup = ExitStack()
+    cleanup.callback(p.close)
+    assert next(p)["x"].value == 0
+    assert started.wait(3)
+    with ThreadPoolExecutor(max_workers=1) as closer:
+        closing = closer.submit(cleanup.close)
+        try:
+            assert not drained.is_set()
+            assert not closing.done()
+        finally:
+            release.set()
+        closing.result(timeout=3)
+    assert drained.is_set()
+    assert p._closed and p._pending is None
+    retry = fake_prefetch(iter([0, 1, 2]))
+    assert [b["x"].value for b in retry] == [0, 1, 2]
+
+
+def test_cuda_owned_buffers_mutating_collator_and_tail():
+    import numpy as np
+
+    torch = pytest.importorskip("torch")
+    if not torch.cuda.is_available():
+        pytest.skip("CUDA required")
+
+    source = np.arange(8, dtype=np.float32).reshape(4, 2)
+    source.flags.writeable = False
+
+    def collate(batch):
+        batch["x"] += 1
+        return batch
+
+    compute = torch.cuda.Stream()
+    retained = []
+    with torch.cuda.stream(compute):
+        p = PinnedPrefetch(iter([{"x": source}, {"x": source[:1]}]), collate)
+        try:
+            for batch in p:
+                retained.append(batch["x"])
+                # Exercise mutation outside torch.inference_mode.
+                batch["x"].add_(2)
+        finally:
+            p.close()
+    compute.synchronize()
+    np.testing.assert_array_equal(source, np.arange(8).reshape(4, 2))
+    np.testing.assert_array_equal(retained[0].cpu().numpy(), source + 3)
+    np.testing.assert_array_equal(retained[1].cpu().numpy(), source[:1] + 3)
+    assert retained[0].data_ptr() != retained[1].data_ptr()