test_cuda.py

# Owner(s): ["module: cuda"]
# ruff: noqa: F841

import contextlib
import ctypes
import functools
import gc
import json
import os
import pickle
import random
import subprocess
import sys
import tempfile
import threading
import time
import unittest
import warnings
from collections import defaultdict
from copy import deepcopy
from itertools import product
from random import randint
from unittest.mock import patch

import psutil

import torch
import torch.cuda
import torch.nn as nn
from torch import inf, nan
from torch.cuda._memory_viz import (
    _profile_to_snapshot,
    profile_plot,
    segment_plot,
    trace_plot,
)
from torch.testing._internal.autocast_test_lists import AutocastTestLists, TestAutocast
from torch.testing._internal.common_cuda import (
    _create_scaling_case,
    _get_torch_cuda_version,
    blas_library_context,
    PLATFORM_SUPPORTS_GREEN_CONTEXT,
    PLATFORM_SUPPORTS_WORKQUEUE_CONFIG,
    SM70OrLater,
    SM89OrLater,
    TEST_CUDNN,
    TEST_MULTIGPU,
    tf32_on_and_off,
    xfailCUDAIfSM89OrLaterOnWindows,
)
from torch.testing._internal.common_device_type import (
    instantiate_device_type_tests,
    largeTensorTest,
    onlyCUDA,
    onlyNativeDeviceTypes,
    skipCUDAIf,
)
from torch.testing._internal.common_optimizers import (
    _get_optim_inputs_including_global_cliquey_kwargs,
    optim_db,
    optims,
    TensorTracker,
)
from torch.testing._internal.common_utils import (
    cuda_python_error_check,
    EXPANDABLE_SEGMENTS,
    freeze_rng_state,
    gcIfJetson,
    get_cycles_per_ms,
    instantiate_parametrized_tests,
    IS_ARM64,
    IS_FBCODE,
    IS_JETSON,
    IS_LINUX,
    IS_SANDCASTLE,
    IS_WINDOWS,
    IS_X86,
    load_tests,
    MI200_ARCH,
    MI300_ARCH,
    parametrize,
    recover_orig_fp32_precision,
    run_tests,
    serialTest,
    setBlasBackendsToDefaultFinally,
    skipCUDAMemoryLeakCheckIf,
    skipCUDANonDefaultStreamIf,
    skipIfRocm,
    skipIfRocmArch,
    skipIfRocmVersionLessThan,
    slowTest,
    subtest,
    TemporaryFileName,
    TEST_CUDA,
    TEST_CUDA_GRAPH,
    TEST_CUDA_PYTHON_BINDINGS,
    TEST_NUMPY,
    TEST_WITH_ROCM,
    TestCase,
)
from torch.utils._triton import has_triton
from torch.utils.checkpoint import checkpoint_sequential
from torch.utils.viz._cycles import observe_tensor_cycles


requiresCppContext = unittest.skipUnless(
    (IS_X86 or IS_ARM64) and IS_LINUX, "cpp contexts are linux x86/aarch64 only"
)

# load_tests from common_utils is used to automatically filter tests for
# sharding on sandcastle. This line silences flake warnings
load_tests = load_tests  # noqa: PLW0127

try:
    import torchvision.models  # noqa: F401

    # from torchvision.models import resnet18

    HAS_TORCHVISION = True
except ImportError:
    HAS_TORCHVISION = False
skipIfNoTorchVision = unittest.skipIf(not HAS_TORCHVISION, "no torchvision")

TEST_CUDAMALLOCASYNC = TEST_CUDA and (
    torch.cuda.get_allocator_backend() == "cudaMallocAsync"
)
TEST_LARGE_TENSOR = TEST_CUDA
TEST_MEDIUM_TENSOR = TEST_CUDA
TEST_BF16 = False
TEST_PYNVML = (
    torch.cuda._HAS_PYNVML and not IS_JETSON
)  # nvml not fully supported on jetson
if TEST_CUDA:
    TEST_LARGE_TENSOR = torch.cuda.get_device_properties(0).total_memory >= 12e9
    TEST_MEDIUM_TENSOR = torch.cuda.get_device_properties(0).total_memory >= 6e9
    TEST_BF16 = torch.cuda.is_bf16_supported()

_cycles_per_ms = None

_wait_for_cpu_kernel = None


def skip_background_threads_on_windows(f):
    @functools.wraps(f)
    def wrapped(self, **kwargs):
        if IS_WINDOWS and SM89OrLater and kwargs.get("use_background_threads"):
            raise unittest.SkipTest("using background threads fails on Windows")
        return f(self, **kwargs)

    return wrapped


def get_wait_for_cpu_kernel():
    """Returns a compiled CUDA/HIP spin-wait kernel that blocks the GPU stream
    until the host sets a pinned int32 flag to non-zero. On NVIDIA, requires
    SM70+ for the relaxed-system-scope PTX load. On AMD, uses
    __hip_atomic_load with __HIP_MEMORY_SCOPE_SYSTEM (the AMDGCN equivalent).

    Usage::

        kernel = get_wait_for_cpu_kernel()
        flag = torch.zeros(1, dtype=torch.int32, device="cpu").pin_memory()
        with torch.cuda.stream(s):
            kernel(grid=(1, 1, 1), block=(1, 1, 1), args=[flag])
        # stream s is now blocked until:
        flag[0] = 1
    """
    global _wait_for_cpu_kernel
    if _wait_for_cpu_kernel is None:
        from torch.cuda import _compile_kernel

        _wait_for_cpu_kernel = _compile_kernel(
            r"""
            __global__ void wait_for_cpu(int *pinned_cpu_flag) {
                int flag = 0;
                do {
            #ifdef __HIP_PLATFORM_AMD__
                    flag = __hip_atomic_load(pinned_cpu_flag, __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_SYSTEM);
            #else
                    asm volatile("ld.relaxed.sys.global.s32 %0, [%1];" : "=r"(flag) : "l"(pinned_cpu_flag) : "memory");
            #endif
                } while (flag == 0);
            }
            """,
            "wait_for_cpu",
        )
    return _wait_for_cpu_kernel


@unittest.skipIf(not TEST_CUDA, "CUDA not available, skipping tests")
@torch.testing._internal.common_utils.markDynamoStrictTest
class TestCuda(TestCase):
    _do_cuda_memory_leak_check = True
    _do_cuda_non_default_stream = True
    FIFTY_MIL_CYCLES = 50000000

    def setUp(self):
        super().setUp()

    def tearDown(self):
        super().tearDown()

    @property
    def expandable_segments(self):
        return EXPANDABLE_SEGMENTS

    def test_pinned_memory_with_cudaregister(self):
        try:
            torch.cuda.memory._set_allocator_settings(
                "pinned_use_cuda_host_register:True,pinned_num_register_threads:8"
            )
            t = torch.ones(20)
            self.assertFalse(t.is_pinned())
            try:
                pinned_t = torch.ones(1 << 21).pin_memory()
                self.assertTrue(pinned_t.is_pinned())
                pinned_t = torch.ones(1 << 24).pin_memory()
                self.assertTrue(pinned_t.is_pinned())
            except RuntimeError as e:
                # Some GPUs don't support same address space on host and device side
                pass
        finally:
            torch.cuda.memory._set_allocator_settings(
                "pinned_use_cuda_host_register:False"
            )

    def test_pinned_memory_with_cudaregister_multithread(self):
        num_threads = 4
        threads = [
            threading.Thread(target=self.test_pinned_memory_with_cudaregister)
            for t in range(num_threads)
        ]
        for thread in threads:
            thread.start()
        for thread in threads:
            thread.join()

    @serialTest()
    def test_host_memory_stats(self):
        # Helper functions
        def empty_stats():
            return {
                "allocated_bytes.allocated": 0,
                "allocated_bytes.current": 0,
                "allocated_bytes.freed": 0,
                "allocated_bytes.peak": 0,
                "allocations.allocated": 0,
                "allocations.current": 0,
                "allocations.freed": 0,
                "allocations.peak": 0,
                "host_alloc_time.count": 0,
                "host_free_time.count": 0,
                "num_host_alloc": 0,
                "num_host_free": 0,
                "active_bytes.allocated": 0,
                "active_bytes.current": 0,
                "active_bytes.freed": 0,
                "active_bytes.peak": 0,
                "active_requests.allocated": 0,
                "active_requests.current": 0,
                "active_requests.freed": 0,
                "active_requests.peak": 0,
            }

        def check_stats(expected):
            stats = torch.cuda.host_memory_stats()
            for k, v in expected.items():
                if v != stats[k]:
                    print(f"key: {k}, expected: {v}, stats: {stats[k]}")
                self.assertEqual(v, stats[k])

        # Setup the test cleanly
        alloc1 = 10
        alloc1_aligned = 16
        alloc2 = 20
        alloc2_aligned = 32
        expected = empty_stats()

        # Reset any lingering state
        gc.collect()
        torch._C._host_emptyCache()

        # Check that stats are empty
        check_stats(expected)

        # Make first allocation and check stats
        t1 = torch.ones(alloc1 * 1024, pin_memory=True)
        self.assertTrue(t1.is_pinned())
        for prefix in ["active_requests", "allocations"]:
            for suffix in ["allocated", "current", "peak"]:
                expected[prefix + "." + suffix] += 1

        allocation_size1 = alloc1_aligned * 1024 * 4
        for prefix in ["allocated_bytes", "active_bytes"]:
            for suffix in ["allocated", "current", "peak"]:
                expected[prefix + "." + suffix] += allocation_size1

        expected["num_host_alloc"] += 1
        expected["host_alloc_time.count"] += 1

        check_stats(expected)

        # Make second allocation and check stats
        t2 = torch.ones(alloc2 * 1024, pin_memory=True)
        self.assertTrue(t2.is_pinned())
        for prefix in ["active_requests", "allocations"]:
            for suffix in ["allocated", "current", "peak"]:
                expected[prefix + "." + suffix] += 1

        allocation_size2 = alloc2_aligned * 1024 * 4
        for prefix in ["allocated_bytes", "active_bytes"]:
            for suffix in ["allocated", "current", "peak"]:
                expected[prefix + "." + suffix] += allocation_size2

        expected["num_host_alloc"] += 1
        expected["host_alloc_time.count"] += 1

        check_stats(expected)

        # Empty cache and check stats
        torch._C._host_emptyCache()

        check_stats(expected)

        # Finally, check the reset of peak and accumulated stats
        torch.cuda.reset_peak_host_memory_stats()
        torch.cuda.reset_accumulated_host_memory_stats()

        expected = empty_stats()

    @serialTest()
    def test_host_memory_stats_active_after_free_no_events(self):
        """Regression test for free()'s no-events fast path.

        When a pinned tensor is freed without any CUDA stream having been
        recorded against its block (i.e. it was never used as the source of
        an async H2D copy), the allocator pushes the block back onto the
        free list directly. That path must decrement active_requests.current
        and active_bytes.current; otherwise the counters grow monotonically
        on every plain pin-and-drop cycle even though the underlying memory
        is correctly recycled.

        Two cycles exercise both alloc paths: the first triggers a new slab
        via add_allocated_block, the second reuses the cached block via
        get_free_block.
        """
        # host_memory_stats() returns an empty dict until the host caching
        # allocator has been initialized; force init so the stat keys exist
        # regardless of test ordering.
        torch.cuda.init()
        gc.collect()
        torch._C._host_emptyCache()
        torch.cuda.reset_peak_host_memory_stats()
        torch.cuda.reset_accumulated_host_memory_stats()

        for _ in range(2):
            t = torch.empty(1024 * 1024, dtype=torch.float32, pin_memory=True)
            self.assertTrue(t.is_pinned())
            del t
            gc.collect()
            stats = torch.cuda.host_memory_stats()
            self.assertEqual(
                stats["active_bytes.current"],
                0,
                "active_bytes.current should be 0 after a pin-and-drop cycle "
                "with no async copy: the block is returned to the cache, so "
                "the active counter must drop back to zero.",
            )
            self.assertEqual(
                stats["active_requests.current"],
                0,
                "active_requests.current should be 0 after a pin-and-drop cycle.",
            )

    @serialTest()
    def test_host_memory_stats_active_after_event_drain(self):
        """Regression test for the event-drain path through empty_cache.

        When a pinned block is used as the source of a non_blocking H2D copy
        and then freed, the block sits on the events queue until
        process_events_for_specific_size drains it. That function takes a
        size parameter that is -1 when invoked from process_events (the
        all-sizes case, reached via _host_emptyCache). The decrement of
        active_bytes_bucket_stats must use the block's actual size, not the
        -1 parameter, otherwise active_bytes.current is incremented by 1
        byte per drained event and active_bytes.freed goes negative.
        """
        torch.cuda.init()
        gc.collect()
        torch._C._host_emptyCache()
        torch.cuda.reset_peak_host_memory_stats()
        torch.cuda.reset_accumulated_host_memory_stats()

        for _ in range(2):
            t = torch.empty(1024 * 1024, dtype=torch.float32, pin_memory=True)
            t.cuda(non_blocking=True)
            torch.cuda.synchronize()
            del t
            gc.collect()
            # Drain via process_events_for_specific_size(-1, ...).
            torch._C._host_emptyCache()

            stats = torch.cuda.host_memory_stats()
            self.assertEqual(
                stats["active_bytes.current"],
                0,
                "active_bytes.current should be 0 after empty_cache drains "
                "the events queue.",
            )
            self.assertEqual(
                stats["active_requests.current"],
                0,
                "active_requests.current should be 0 after drain.",
            )
            self.assertGreaterEqual(
                stats["active_bytes.freed"],
                0,
                "active_bytes.freed must be non-negative; a negative value "
                "means the drain decremented by the wrong size.",
            )
            self.assertGreaterEqual(
                stats["active_requests.freed"],
                0,
                "active_requests.freed must be non-negative.",
            )

    def test_pinned_memory_empty_cache(self):
        try:
            for alloc_settings in (True, False):
                torch.cuda.memory._set_allocator_settings(
                    f"pinned_use_cuda_host_register:{alloc_settings}"
                )
                try:
                    t = torch.ones(1024 * 1024, pin_memory=True)
                    self.assertTrue(t.is_pinned())
                    del t
                    torch._C._host_emptyCache()
                except RuntimeError as e:
                    # Some GPUs don't support same address space on host and device side
                    pass
        finally:
            torch.cuda.memory._set_allocator_settings(
                "pinned_use_cuda_host_register:False"
            )

    # Pinned allocator background thread does not shut down cleanly on Windows
    # Python process hangs
    @unittest.skipIf(IS_WINDOWS and SM89OrLater, "Fails on windows with SM89+")
    def test_pinned_memory_use_background_threads(self):
        script = """
import torch

torch.cuda.memory._set_allocator_settings(
    f"pinned_use_background_threads:True"
)
t = torch.ones(1024 * 1024, pin_memory=True)
print(t.is_pinned())
"""
        proc = subprocess.run([sys.executable, "-c", script], capture_output=True)
        self.assertEqual(proc.returncode, 0)

    def test_cudart_register(self):
        t = torch.ones(20)
        self.assertFalse(t.is_pinned())
        cudart = torch.cuda.cudart()
        r = cudart.cudaHostRegister(t.data_ptr(), t.numel() * t.element_size(), 0)
        self.assertEqual(r, 0)
        self.assertTrue(t.is_pinned())
        r = cudart.cudaHostUnregister(t.data_ptr())
        self.assertEqual(r, 0)
        self.assertFalse(t.is_pinned())

    def test_memory_allocation(self):
        gc.collect()
        torch.cuda.empty_cache()
        mem = None
        size = 1
        prev = 0
        try:
            prev = torch.cuda.memory_allocated()
            mem = torch.cuda.caching_allocator_alloc(size)
            self.assertGreater(torch.cuda.memory_allocated(), prev)
        finally:
            if mem is not None:
                torch.cuda.caching_allocator_delete(mem)
                self.assertEqual(torch.cuda.memory_allocated(), prev)

    def test_caching_allocator_alloc_negative_size(self):
        with self.assertRaisesRegex(ValueError, "Invalid memory size"):
            torch.cuda.memory.caching_allocator_alloc(-1024)

    def test_memory_stats(self):
        gc.collect()
        torch.cuda.empty_cache()
        torch.cuda.reset_peak_memory_stats()
        torch.cuda.reset_accumulated_memory_stats()
        prev_allocated = torch.accelerator.memory_allocated()
        prev_reserved = torch.accelerator.memory_reserved()
        prev_max_allocated = torch.accelerator.max_memory_allocated()
        prev_max_reserved = torch.accelerator.max_memory_reserved()
        self.assertEqual(prev_allocated, prev_max_allocated)
        self.assertEqual(prev_reserved, prev_max_reserved)
        # Activate 1kB memory
        prev_active_current = torch.accelerator.memory_stats()[
            "active_bytes.all.current"
        ]
        tmp = torch.randn(256, device="cuda")
        # Detect if the current active memory is 1kB
        self.assertEqual(
            torch.accelerator.memory_stats()["active_bytes.all.current"],
            1024 + prev_active_current,
        )
        self.assertEqual(torch.accelerator.memory_stats()["active_bytes.all.freed"], 0)
        del tmp
        gc.collect()
        torch.accelerator.empty_cache()
        self.assertEqual(
            torch.accelerator.memory_stats()["active_bytes.all.current"],
            prev_active_current,
        )
        self.assertEqual(
            torch.accelerator.memory_stats()["active_bytes.all.freed"], 1024
        )
        torch.accelerator.reset_peak_memory_stats()
        self.assertEqual(torch.accelerator.max_memory_allocated(), prev_max_allocated)
        self.assertEqual(torch.accelerator.max_memory_reserved(), prev_max_reserved)

    def test_check_error(self):
        # Assert this call doesn't raise.
        torch.cuda.check_error(0)

        with self.assertRaisesRegex(
            torch.cuda.CudaError, "out of memory|hipErrorOutOfMemory"
        ):
            torch.cuda.check_error(2)

    def test_cuda_get_device_name(self):
        # Testing the behaviour with None as an argument
        current_device = torch.cuda.current_device()
        current_device_name = torch.cuda.get_device_name(current_device)
        device_name_None = torch.cuda.get_device_name(None)
        self.assertEqual(current_device_name, device_name_None)

        # Testing the behaviour for No argument
        device_name_no_argument = torch.cuda.get_device_name()
        self.assertEqual(current_device_name, device_name_no_argument)

    def test_cuda_get_device_capability(self):
        # Testing the behaviour with None as an argument
        current_device = torch.cuda.current_device()
        current_device_capability = torch.cuda.get_device_capability(current_device)
        device_capability_None = torch.cuda.get_device_capability(None)
        self.assertEqual(current_device_capability, device_capability_None)

        # Testing the behaviour for No argument
        device_capability_no_argument = torch.cuda.get_device_capability()
        self.assertEqual(current_device_capability, device_capability_no_argument)

    def test_cuda_get_device_properties(self):
        # Testing the behaviour with None as an argument
        current_device = torch.cuda.current_device()
        current_device_properties = torch.cuda.get_device_properties(current_device)
        device_properties_None = torch.cuda.get_device_properties(None)
        self.assertEqual(current_device_properties, device_properties_None)

        # Testing the behaviour for No argument
        device_properties_no_argument = torch.cuda.get_device_properties()
        self.assertEqual(current_device_properties, device_properties_no_argument)

    @unittest.skipIf(
        IS_JETSON, "oom reporting has issues on jetson igx due to partial nvml support"
    )
    def test_out_of_memory(self):
        tensor = torch.zeros(1024, device="cuda")

        oom_regex = (
            "would exceed allowed memory"
            if TEST_CUDAMALLOCASYNC
            else f"Tried to allocate 800000000.00 GiB. GPU {tensor.device.index} has a total capacity of"
        )
        with self.assertRaisesRegex(RuntimeError, oom_regex):
            torch.empty(1024 * 1024 * 1024 * 800000000, dtype=torch.int8, device="cuda")

        with self.assertRaisesRegex(
            RuntimeError, "Tried to allocate more than 1EB memory"
        ):
            torch.empty(
                1024 * 1024 * 1024 * 8000000000, dtype=torch.int8, device="cuda"
            )

        # ensure out of memory error doesn't disturb subsequent kernel
        tensor.fill_(1)
        self.assertTrue((tensor == 1).all())

    # CUDA memory allocations on windows do not OOM on rtx even when they cross allowed memory
    # Skip test until this is investigated
    @unittest.skipIf(IS_WINDOWS and SM89OrLater, "Fails on windows with SM89+")
    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC or IS_JETSON, "Segmentation fault (core dumped)"
    )
    @serialTest()
    def test_out_of_memory_retry(self):
        torch.cuda.empty_cache()
        total_memory = torch.cuda.get_device_properties(0).total_memory
        oom_regex = (
            "would exceed allowed memory"
            if TEST_CUDAMALLOCASYNC
            else "Tried to allocate"
        )
        size = int(total_memory * 0.5)
        a = torch.empty(size, dtype=torch.int8, device="cuda")
        with self.assertRaisesRegex(RuntimeError, oom_regex):
            b = torch.empty(size, dtype=torch.int8, device="cuda")
        del a
        b = torch.empty(size, dtype=torch.int8, device="cuda")
        del b
        # We used a lot of memory here, clean up so we don't affect other tests too much
        torch.cuda.empty_cache()
        torch.cuda.reset_peak_memory_stats()

    @serialTest()
    @unittest.skipIf(
        IS_JETSON, "oom reporting has issues on jetson igx due to partial nvml support"
    )
    def test_set_per_process_memory_fraction(self):
        torch.cuda.empty_cache()
        orig = torch.cuda.get_per_process_memory_fraction(0)
        torch.cuda.reset_peak_memory_stats(0)
        try:
            # test invalid fraction value.
            with self.assertRaisesRegex(TypeError, "Invalid type"):
                torch.cuda.set_per_process_memory_fraction(1)
            with self.assertRaisesRegex(ValueError, "Invalid fraction value"):
                torch.cuda.set_per_process_memory_fraction(-0.1)
            with self.assertRaisesRegex(ValueError, "Invalid fraction value"):
                torch.cuda.set_per_process_memory_fraction(2.0)

            tensor = torch.zeros(1024, device="cuda")
            torch.cuda.empty_cache()
            total_memory = torch.cuda.get_device_properties(0).total_memory
            torch.cuda.set_per_process_memory_fraction(0.5, 0)

            # test 0.499 allocation is ok.
            application = int(total_memory * 0.499) - torch.cuda.max_memory_reserved()
            tmp_tensor = torch.empty(application, dtype=torch.int8, device="cuda")
            del tmp_tensor
            torch.cuda.empty_cache()

            application = int(total_memory * 0.5)
            # it will get OOM when try to allocate more than half memory.
            oom_regex = (
                "would exceed allowed memory"
                if TEST_CUDAMALLOCASYNC
                else "out of memory"
            )
            with self.assertRaisesRegex(RuntimeError, oom_regex):
                torch.empty(application, dtype=torch.int8, device="cuda")

            # ensure out of memory error doesn't disturb subsequent kernel
            tensor.fill_(1)
            self.assertTrue((tensor == 1).all())
        finally:
            torch.cuda.set_per_process_memory_fraction(orig, 0)

    @unittest.skipIf(
        IS_JETSON, "oom reporting has issues on jetson igx due to partial nvml support"
    )
    @serialTest()
    def test_get_per_process_memory_fraction(self):
        # get the initial memory fraction
        init_fraction = torch.cuda.get_per_process_memory_fraction()

        # set and get the limiting cases
        torch.cuda.set_per_process_memory_fraction(1.0)
        self.assertEqual(torch.cuda.get_per_process_memory_fraction(), 1.0)
        torch.cuda.set_per_process_memory_fraction(0.0)
        self.assertEqual(torch.cuda.get_per_process_memory_fraction(), 0.0)

        # test a few random cases
        for val in torch.rand(3):
            torch.cuda.set_per_process_memory_fraction(float(val))
            self.assertEqual(torch.cuda.get_per_process_memory_fraction(), float(val))

        # restore the initial memory fraction
        torch.cuda.set_per_process_memory_fraction(init_fraction)

    def test_uuid(self):
        uuid = torch.cuda.get_device_properties(0).uuid
        self.assertEqual(len(str(uuid)), 36)  # xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx
        self.assertEqual(len(uuid.bytes), 16)

    def test_copy_non_blocking(self):
        def _test_copy_non_blocking(a, b):
            event = torch.cuda.Event()
            a.copy_(b, non_blocking=True)
            event.record()
            event.synchronize()
            self.assertEqual(a, b)

        # 10MB copies
        x = torch.ones(10000000, dtype=torch.uint8).cuda()
        y = torch.zeros(10000000, dtype=torch.uint8).pin_memory()
        _test_copy_non_blocking(x, y)

        x = torch.zeros(10000000, dtype=torch.uint8).pin_memory()
        y = torch.ones(10000000, dtype=torch.uint8).cuda()
        _test_copy_non_blocking(x, y)

        # Test the case where the pinned data_ptr is not equal to the storage data_ptr.
        x_base = torch.zeros(10000000, dtype=torch.uint8).pin_memory()
        x = x_base[1:]
        self.assertTrue(x.is_pinned())
        self.assertTrue(x_base.is_pinned())
        self.assertNotEqual(x_base.data_ptr(), x.data_ptr())
        self.assertEqual(x_base.storage().data_ptr(), x.storage().data_ptr())
        y = torch.ones(10000000 - 1, dtype=torch.uint8).cuda()
        _test_copy_non_blocking(x, y)

    def test_copy_non_blocking_type_conversion(self):
        a = torch.ones(1, device="cuda")
        b = torch.zeros(1, device="cpu", pin_memory=True)
        c = torch.empty(1, device="cuda", dtype=torch.long)
        torch.cuda._sleep(int(100 * get_cycles_per_ms()))
        b.copy_(a, non_blocking=True)
        c.copy_(b, non_blocking=True)
        self.assertEqual(a, c, exact_dtype=False)

    @serialTest()
    def test_to_non_blocking(self):
        stream = torch.cuda.current_stream()

        def _test_to_non_blocking(a, non_blocking, dst):
            torch.cuda.synchronize()
            # Pushes an 0.1 second spin to stream so if the copy is non blocking,
            # stream will almost surely be active when we query().
            torch.cuda._sleep(int(100 * get_cycles_per_ms()))
            b = a.to(device=dst, non_blocking=non_blocking)
            self.assertEqual(stream.query(), not non_blocking)
            stream.synchronize()
            self.assertEqual(a, b)
            self.assertTrue(b.is_pinned() == (non_blocking and dst == "cpu"))

        for dst, try_non_blocking in product(("cuda", "cpu"), (True, False)):
            # Creates source on the opposite device from destination.
            src = torch.randn(
                1000000,
                device="cuda" if dst == "cpu" else "cpu",
                pin_memory=dst == "cuda",
            )
            _test_to_non_blocking(src, try_non_blocking, dst)

    def test_to_cpu_blocking_by_default(self):
        src = torch.randn(1000000, device="cuda")
        torch.cuda.synchronize()
        torch.cuda._sleep(int(100 * get_cycles_per_ms()))
        dst = src.to(device="cpu")
        self.assertEqual(torch.cuda.current_stream().query(), True)
        self.assertEqual(src, dst)
        self.assertFalse(dst.is_pinned())

    def test_serialization_array_with_storage(self):
        x = torch.randn(5, 5).cuda()
        y = torch.IntTensor(2, 5).fill_(0).cuda()
        q = [x, y, x, y.storage()]
        with tempfile.NamedTemporaryFile() as f:
            torch.save(q, f)
            f.seek(0)
            q_copy = torch.load(f)
        self.assertEqual(q_copy, q, atol=0, rtol=0)
        q_copy[0].fill_(5)
        self.assertEqual(q_copy[0], q_copy[2], atol=0, rtol=0)
        self.assertTrue(isinstance(q_copy[0], torch.cuda.FloatTensor))
        self.assertTrue(isinstance(q_copy[1], torch.cuda.IntTensor))
        self.assertTrue(isinstance(q_copy[2], torch.cuda.FloatTensor))
        self.assertTrue(isinstance(q_copy[3], torch.storage.TypedStorage))
        self.assertTrue(isinstance(q_copy[3]._untyped_storage, torch.UntypedStorage))
        q_copy[1].fill_(10)
        self.assertEqual(q_copy[3], torch.cuda.IntStorage(10).fill_(10))

    @unittest.skipIf(
        IS_WINDOWS and SM89OrLater, "preferred_blas_library not supported on Windows"
    )
    @unittest.skipIf(IS_FBCODE or IS_SANDCASTLE, "Does not work in fbcode yet")
    @setBlasBackendsToDefaultFinally
    def test_preferred_blas_library_settings(self):
        def _check_default():
            default = torch.backends.cuda.preferred_blas_library()
            if torch.version.cuda:
                # CUDA logic is easy, it's always cublas
                self.assertTrue(default == torch._C._BlasBackend.Cublas)
            else:
                # ROCm logic is less so, it's cublaslt for some Instinct, cublas for all else
                gcn_arch = str(
                    torch.cuda.get_device_properties(0).gcnArchName.split(":", 1)[0]
                )
                if gcn_arch in ["gfx90a", "gfx942", "gfx950", "gfx1200", "gfx1201"]:
                    self.assertTrue(default == torch._C._BlasBackend.Cublaslt)
                else:
                    self.assertTrue(default == torch._C._BlasBackend.Cublas)

        _check_default()
        # "Default" can be set but is immediately reset internally to the actual default value.
        self.assertTrue(
            torch.backends.cuda.preferred_blas_library("default")
            != torch._C._BlasBackend.Default
        )
        _check_default()
        self.assertTrue(
            torch.backends.cuda.preferred_blas_library("cublas")
            == torch._C._BlasBackend.Cublas
        )
        self.assertTrue(
            torch.backends.cuda.preferred_blas_library("hipblas")
            == torch._C._BlasBackend.Cublas
        )
        # check bad strings
        with self.assertRaisesRegex(
            RuntimeError,
            "Unknown input value. Choose from: default, cublas, hipblas, cublaslt, hipblaslt, ck.",
        ):
            torch.backends.cuda.preferred_blas_library("unknown")
        # check bad input type
        with self.assertRaisesRegex(RuntimeError, "Unknown input value type."):
            torch.backends.cuda.preferred_blas_library(1.0)
        # check env var override
        custom_envs = [
            ({"TORCH_BLAS_PREFER_CUBLASLT": "1"}, "_BlasBackend.Cublaslt"),
            ({"TORCH_BLAS_PREFER_HIPBLASLT": "1"}, "_BlasBackend.Cublaslt"),
            ({"TORCH_BLAS_PREFER_CUBLASLT": "0"}, "_BlasBackend.Cublas"),
            ({"TORCH_BLAS_PREFER_HIPBLASLT": "0"}, "_BlasBackend.Cublas"),
        ]
        test_script = "import torch;print(torch.backends.cuda.preferred_blas_library())"
        for env_config, expected in custom_envs:
            env = os.environ.copy()
            for key, value in env_config.items():
                env[key] = value
            r = (
                subprocess.check_output([sys.executable, "-c", test_script], env=env)
                .decode("ascii")
                .strip()
            )
            self.assertEqual(expected, r)

        # explicitly check default when no env vars are set
        if not any(
            os.environ.get(v)
            for v in ("TORCH_BLAS_PREFER_CUBLASLT", "TORCH_BLAS_PREFER_HIPBLASLT")
        ):
            torch.backends.cuda.preferred_blas_library("default")
            _check_default()

    @unittest.skipIf(
        IS_WINDOWS and SM89OrLater, "preferred_blas_library not supported on Windows"
    )
    @unittest.skipIf(TEST_CUDAMALLOCASYNC, "temporarily disabled for async")
    @serialTest()
    @blas_library_context("cublas")
    def test_cublas_workspace_explicit_allocation(self):
        a = torch.randn(7, 7, device="cuda", requires_grad=False)
        if torch.version.hip:
            default_workspace_size = 1024 * 32 * 1024  # :1024:32  32MiB
            # different size (128 MiB) expected on MI300 GPU
            gcn_arch = str(
                torch.cuda.get_device_properties(0).gcnArchName.split(":", 1)[0]
            )
            if "gfx94" in gcn_arch or "gfx95" in gcn_arch:
                default_workspace_size = 1024 * 128 * 1024  # :1024:128
        else:
            default_workspace_size = (
                4096 * 2 * 1024 + 16 * 8 * 1024
            )  # :4096:2:16:8  8MiB
            # different size (32 MiB) expected on Hopper/Blackwell GPU
            if torch.cuda.get_device_capability()[0] in (9, 10, 12):
                default_workspace_size = 4096 * 8 * 1024

        def check_workspace_size(inp):
            torch._C._cuda_clearCublasWorkspaces()
            start = torch.cuda.memory_stats()["active_bytes.all.allocated"]
            with torch.no_grad():
                torch.matmul(inp, inp)
            finish = torch.cuda.memory_stats()["active_bytes.all.allocated"]
            return finish - start

        # check default
        self.assertLess(abs(check_workspace_size(a) - default_workspace_size), 524288)

        # check explicit size via API
        explicit_size = 3072 * 1024
        torch.backends.cuda.cublas_workspace_size(explicit_size)
        self.assertLess(abs(check_workspace_size(a) - explicit_size), 524288)

        # check invalid size rejected
        with self.assertRaisesRegex(
            RuntimeError, "cublas workspace size must be non-negative"
        ):
            torch.backends.cuda.cublas_workspace_size(-1)

        # restore default
        torch._C._cuda_resetCublasWorkspaceSize()
        self.assertLess(abs(check_workspace_size(a) - default_workspace_size), 524288)

        torch._C._cuda_clearCublasWorkspaces()

    @unittest.skipIf(TEST_CUDAMALLOCASYNC, "temporarily disabled for async")
    @unittest.skipIf(IS_FBCODE, "not enabled by default on fbcode")
    @unittest.skipIf(TEST_WITH_ROCM, "not enabled by default on rocm")
    @setBlasBackendsToDefaultFinally
    @serialTest()
    def test_cublas_unified_workspace(self):
        torch.cuda.reset_peak_memory_stats()
        # We run a cuBLAS matmul, guaranteeing a workspace present on the current stream
        torch.backends.cuda.preferred_blas_library("cublas")
        x = torch.randn(128, 128, device="cuda")
        torch.matmul(x, x)
        # Stash the maximum memory allocated after running matmuls
        warmed_alloc = torch.cuda.max_memory_allocated()
        torch.backends.cuda.preferred_blas_library("cublaslt")
        torch.matmul(x, x)
        lt_alloc = torch.cuda.max_memory_allocated()
        # With unified workspaces, the peak memory allocation should not increase after
        # switching to Lt, otherwise the temporary allocation would bump the peak
        self.assertEqual(warmed_alloc, lt_alloc)

    @setBlasBackendsToDefaultFinally
    def test_cublas_workspace_size_api(self):
        # Test getter returns a positive default
        original_size = torch.backends.cuda.cublas_workspace_size()
        self.assertGreater(original_size, 0)

        original_lt_size = torch.backends.cuda.cublaslt_workspace_size()
        self.assertGreater(original_lt_size, 0)

        # Test setter changes the value and getter reflects it
        new_size = 64 * 1024 * 1024  # 64 MiB
        result = torch.backends.cuda.cublas_workspace_size(new_size)
        self.assertEqual(result, new_size)
        self.assertEqual(torch.backends.cuda.cublas_workspace_size(), new_size)

        new_lt_size = 2 * 1024 * 1024  # 2 MiB
        result_lt = torch.backends.cuda.cublaslt_workspace_size(new_lt_size)
        self.assertEqual(result_lt, new_lt_size)
        self.assertEqual(torch.backends.cuda.cublaslt_workspace_size(), new_lt_size)

        # Test validation rejects negative values
        with self.assertRaisesRegex(
            RuntimeError, "cublas workspace size must be non-negative"
        ):
            torch.backends.cuda.cublas_workspace_size(-1)
        with self.assertRaisesRegex(
            RuntimeError, "cublaslt workspace size must be non-negative"
        ):
            torch.backends.cuda.cublaslt_workspace_size(-1)

    @setBlasBackendsToDefaultFinally
    def test_blas_workspace_size_api(self):
        # Dispatches to the correct backend based on preferred_blas_library()
        pref = torch.backends.cuda.preferred_blas_library()
        if pref == torch._C._BlasBackend.Cublaslt:
            expected = torch.backends.cuda.cublaslt_workspace_size()
        else:
            # Default and Cublas both map to cuBLAS
            expected = torch.backends.cuda.cublas_workspace_size()
        self.assertEqual(torch.backends.cuda.blas_workspace_size(), expected)

        # Explicit backend= parameter
        self.assertEqual(
            torch.backends.cuda.blas_workspace_size(backend="cublas"),
            torch.backends.cuda.cublas_workspace_size(),
        )
        self.assertEqual(
            torch.backends.cuda.blas_workspace_size(backend="cublaslt"),
            torch.backends.cuda.cublaslt_workspace_size(),
        )
        self.assertEqual(
            torch.backends.cuda.blas_workspace_size(
                backend=torch._C._BlasBackend.Cublas
            ),
            torch.backends.cuda.cublas_workspace_size(),
        )

        # Setting via the dispatcher updates the underlying function
        new_size = 16 * 1024 * 1024
        torch.backends.cuda.blas_workspace_size(new_size, backend="cublas")
        self.assertEqual(torch.backends.cuda.cublas_workspace_size(), new_size)

        torch.backends.cuda.blas_workspace_size(new_size, backend="cublaslt")
        self.assertEqual(torch.backends.cuda.cublaslt_workspace_size(), new_size)

        # CK backend has no workspace
        with self.assertRaisesRegex(
            RuntimeError, "CK backend does not use a workspace."
        ):
            torch.backends.cuda.blas_workspace_size(backend="ck")

        # Invalid string
        with self.assertRaisesRegex(
            RuntimeError,
            "Unknown backend string. Choose from: default, cublas, hipblas, cublaslt, hipblaslt, ck.",
        ):
            torch.backends.cuda.blas_workspace_size(backend="invalid")

        # Invalid type
        with self.assertRaisesRegex(RuntimeError, "Unknown backend type."):
            torch.backends.cuda.blas_workspace_size(backend=42)

    @unittest.skipIf(TEST_CUDAMALLOCASYNC, "temporarily disabled for async")
    @setBlasBackendsToDefaultFinally
    def test_cublas_workspace_lazy_reallocation(self):
        torch.backends.cuda.preferred_blas_library("cublas")

        original_size = torch.backends.cuda.cublas_workspace_size()
        torch._C._cuda_clearCublasWorkspaces()

        # Trigger initial allocation with matmul
        a = torch.randn(7, 7, device="cuda", requires_grad=False)
        with torch.no_grad():
            torch.matmul(a, a)

        mem_after_first = torch.cuda.memory_stats()["active_bytes.all.allocated"]

        # Increase workspace size
        bigger_size = original_size + 32 * 1024 * 1024  # +32 MiB
        torch.backends.cuda.cublas_workspace_size(bigger_size)

        # No immediate memory change (lazy reallocation)
        mem_after_set = torch.cuda.memory_stats()["active_bytes.all.allocated"]
        self.assertEqual(mem_after_first, mem_after_set)

        # Next matmul triggers reallocation
        with torch.no_grad():
            torch.matmul(a, a)

        mem_after_realloc = torch.cuda.memory_stats()["active_bytes.all.allocated"]
        self.assertGreater(mem_after_realloc, mem_after_first)

    def test_cublas_allow_tf32_get_set(self):
        skip_tf32_cublas = "TORCH_ALLOW_TF32_CUBLAS_OVERRIDE" in os.environ and int(
            os.environ["TORCH_ALLOW_TF32_CUBLAS_OVERRIDE"]
        )
        if skip_tf32_cublas:
            self.assertTrue(torch.backends.cuda.matmul.allow_tf32)
            return

        orig = torch.backends.cuda.matmul.allow_tf32
        self.assertEqual(torch._C._get_cublas_allow_tf32(), orig)
        torch.backends.cuda.matmul.allow_tf32 = not orig
        self.assertEqual(torch._C._get_cublas_allow_tf32(), not orig)
        torch.backends.cuda.matmul.allow_tf32 = orig

    def test_float32_matmul_precision_get_set(self):
        orig = torch.get_float32_matmul_precision()
        skip_tf32_cublas = "TORCH_ALLOW_TF32_CUBLAS_OVERRIDE" in os.environ and int(
            os.environ["TORCH_ALLOW_TF32_CUBLAS_OVERRIDE"]
        )
        # this is really just checking that the environment variable is respected during testing
        # and not overwritten by another function that doesn't revert it to the initial value
        if not skip_tf32_cublas:
            self.assertFalse(torch.backends.cuda.matmul.allow_tf32)
            self.assertEqual(torch.get_float32_matmul_precision(), "highest")
        else:
            self.assertTrue(torch.backends.cuda.matmul.allow_tf32)

        for p in ("medium", "high"):
            torch.set_float32_matmul_precision(p)
            self.assertEqual(torch.get_float32_matmul_precision(), p)
            self.assertTrue(torch.backends.cuda.matmul.allow_tf32)
        torch.set_float32_matmul_precision("highest")
        self.assertEqual(torch.get_float32_matmul_precision(), "highest")
        self.assertFalse(torch.backends.cuda.matmul.allow_tf32)
        torch.set_float32_matmul_precision(orig)

    def test_cublas_allow_fp16_reduced_precision_reduction_get_set(self):
        orig = torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction
        orig_splitk = (
            torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction_split_k
        )
        self.assertEqual(
            torch._C._get_cublas_allow_fp16_reduced_precision_reduction(),
            (orig, orig_splitk),
        )
        torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = not orig
        self.assertEqual(
            torch._C._get_cublas_allow_fp16_reduced_precision_reduction(),
            (not orig, True),
        )
        torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = (
            False,
            False,
        )
        self.assertEqual(
            torch._C._get_cublas_allow_fp16_reduced_precision_reduction(),
            (False, False),
        )
        with self.assertRaisesRegex(RuntimeError, "allow_splitk=False"):
            torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = (
                True,
                False,
            )
        torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = (
            orig,
            orig_splitk,
        )

    def test_cublas_allow_bf16_reduced_precision_reduction_get_set(self):
        orig = torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction
        orig_splitk = (
            torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction_split_k
        )
        self.assertEqual(
            torch._C._get_cublas_allow_bf16_reduced_precision_reduction(),
            (orig, orig_splitk),
        )
        torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = not orig
        self.assertEqual(
            torch._C._get_cublas_allow_bf16_reduced_precision_reduction(),
            (not orig, True),
        )
        torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = (
            False,
            False,
        )
        self.assertEqual(
            torch._C._get_cublas_allow_bf16_reduced_precision_reduction(),
            (False, False),
        )
        with self.assertRaisesRegex(RuntimeError, "allow_splitk=False"):
            torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = (
                True,
                False,
            )
        torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = (
            orig,
            orig_splitk,
        )

    def test_cublas_allow_fp16_accumulation_get_set(self):
        orig = torch.backends.cuda.matmul.allow_fp16_accumulation
        self.assertEqual(torch._C._get_cublas_allow_fp16_accumulation(), orig)
        torch.backends.cuda.matmul.allow_fp16_accumulation = not orig
        self.assertEqual(torch._C._get_cublas_allow_fp16_accumulation(), not orig)
        torch.backends.cuda.matmul.allow_fp16_accumulation = orig

    def test_cudnn_allow_tf32_get_set(self):
        with torch.backends.cudnn.flags(
            enabled=None, benchmark=None, deterministic=None, allow_tf32=False
        ):
            self.assertFalse(torch.backends.cudnn.allow_tf32)
        with torch.backends.cudnn.flags(
            enabled=None, benchmark=None, deterministic=None, allow_tf32=True
        ):
            self.assertTrue(torch.backends.cudnn.allow_tf32)

    def test_cudnn_depthwise_kernel_get_set(self):
        self.assertEqual(torch.backends.cudnn.depthwise_kernel, "auto")

        # Test all valid values via the flags() context manager
        for mode in ("auto", "cudnn", "native"):
            with torch.backends.cudnn.flags(
                enabled=None,
                benchmark=None,
                deterministic=None,
                allow_tf32=None,
                depthwise_kernel=mode,
            ):
                self.assertEqual(torch.backends.cudnn.depthwise_kernel, mode)

        # Invalid value should raise
        with self.assertRaises(RuntimeError):
            torch._C._set_cudnn_depthwise_kernel("invalid")

        # Verify the flags() context manager restores the previous value
        with torch.backends.cudnn.flags(
            enabled=None,
            benchmark=None,
            deterministic=None,
            allow_tf32=None,
            depthwise_kernel="native",
        ):
            self.assertEqual(torch.backends.cudnn.depthwise_kernel, "native")
        self.assertEqual(torch.backends.cudnn.depthwise_kernel, "auto")

    @recover_orig_fp32_precision
    def test_fp32_precision_with_tf32(self):
        with torch.backends.cudnn.flags(
            enabled=None,
            benchmark=None,
            benchmark_limit=None,
            deterministic=None,
            allow_tf32=True,
            fp32_precision="none",
        ):
            self.assertEqual(torch.backends.cudnn.conv.fp32_precision, "tf32")
            self.assertEqual(torch.backends.cudnn.rnn.fp32_precision, "tf32")

        with torch.backends.cudnn.flags(
            enabled=None,
            benchmark=None,
            benchmark_limit=None,
            deterministic=None,
            allow_tf32=False,
            fp32_precision="none",
        ):
            self.assertEqual(torch.backends.cudnn.conv.fp32_precision, "none")
            self.assertEqual(torch.backends.cudnn.rnn.fp32_precision, "none")

    @recover_orig_fp32_precision
    @serialTest()
    def test_fp32_precision_with_float32_matmul_precision(self):
        torch.set_float32_matmul_precision("highest")
        self.assertEqual(torch.backends.cuda.matmul.fp32_precision, "ieee")
        torch.set_float32_matmul_precision("high")
        self.assertEqual(torch.backends.cuda.matmul.fp32_precision, "tf32")
        torch.set_float32_matmul_precision("medium")
        self.assertEqual(torch.backends.cuda.matmul.fp32_precision, "tf32")

    @recover_orig_fp32_precision
    @serialTest()
    def test_invalid_status_for_legacy_api(self):
        torch.backends.cudnn.conv.fp32_precision = "none"
        torch.backends.cudnn.rnn.fp32_precision = "tf32"
        with self.assertRaisesRegex(RuntimeError, "mix of the legacy and new APIs"):
            print(torch.backends.cudnn.allow_tf32)

        torch.set_float32_matmul_precision("highest")
        torch.backends.cuda.matmul.fp32_precision = "tf32"
        with self.assertRaisesRegex(RuntimeError, "mix of the legacy and new APIs"):
            print(torch.get_float32_matmul_precision())

        if not TEST_WITH_ROCM:
            with self.assertRaisesRegex(RuntimeError, "mix of the legacy and new APIs"):
                print(torch.backends.cuda.matmul.allow_tf32)

    def test_type_conversions(self):
        x = torch.randn(5, 5)
        self.assertIsInstance(x.float(), torch.FloatTensor)
        self.assertIsInstance(x.cuda().double(), torch.cuda.DoubleTensor)
        self.assertIsInstance(x.cuda().float(), torch.cuda.FloatTensor)
        self.assertIsInstance(x.cuda().float().cpu(), torch.FloatTensor)
        self.assertIsInstance(x.cuda().float().cpu().int(), torch.IntTensor)

        y = x.storage()
        self.assertIsInstance(y.float(), torch.FloatStorage)
        self.assertIsInstance(y.cuda().double(), torch.cuda.DoubleStorage)
        self.assertIsInstance(y.cuda().float(), torch.cuda.FloatStorage)
        self.assertIsInstance(y.cuda().float().cpu(), torch.FloatStorage)
        self.assertIsInstance(y.cuda().float().cpu().int(), torch.IntStorage)

    @unittest.skip("was disabled due to not enough memory, but actually it always fail")
    def test_arithmetic_large_tensor(self):
        x = torch.empty(2**30, device="cuda")

        x.fill_(1)
        self.assertEqual(x.sum(), 2**30)

        x += 1
        self.assertEqual(x.sum(), 2**31)

        x.fill_(1)
        x -= 0.5
        self.assertEqual(x.sum(), 2**29)

        x.fill_(1)
        x *= 2
        self.assertEqual(x.sum(), 2**31)

        x.fill_(1)
        x /= 2
        self.assertEqual(x.sum(), 2**29)

    def test_gather_bool(self):
        t = torch.tensor([[False, True], [True, True]], device="cuda")
        self.assertEqual(
            torch.gather(t, 1, torch.tensor([[0, 0], [1, 0]], device="cuda")),
            torch.tensor([[False, False], [True, True]], device="cuda"),
        )

    def test_torch_manual_seed_seeds_cuda_devices(self):
        with freeze_rng_state():
            x = torch.zeros(4, 4).float().cuda()
            torch.manual_seed(2)
            self.assertEqual(torch.cuda.initial_seed(), 2)
            x.uniform_()
            torch.manual_seed(2)
            y = x.clone().uniform_()
            self.assertEqual(x, y)
            self.assertEqual(torch.cuda.initial_seed(), 2)

    def test_manual_seed(self):
        with freeze_rng_state():
            x = torch.zeros(4, 4).float().cuda()
            torch.cuda.manual_seed(2)
            self.assertEqual(torch.cuda.initial_seed(), 2)
            x.uniform_()
            a = torch.bernoulli(torch.full_like(x, 0.5))
            torch.cuda.manual_seed(2)
            y = x.clone().uniform_()
            b = torch.bernoulli(torch.full_like(x, 0.5))
            self.assertEqual(x, y)
            self.assertEqual(a, b)
            self.assertEqual(torch.cuda.initial_seed(), 2)

    def test_specify_improper_device_name(self):
        with tempfile.TemporaryDirectory() as tmpdir:
            fname = os.path.join(tmpdir, "tempfile.pt")
            with self.assertRaisesRegex(RuntimeError, "Invalid device string"):
                torch.save(
                    [torch.nn.Parameter(torch.randn(10, 10))],
                    fname,
                    _use_new_zipfile_serialization=True,
                )
                torch.load(fname, "cuda0")

    def test_get_device_index(self):
        from torch.cuda._utils import _get_device_index

        with self.assertRaisesRegex(RuntimeError, "Invalid device string"):
            _get_device_index("cuda0", optional=True)

        with self.assertRaisesRegex(ValueError, "Expected a cuda device"):
            cpu_device = torch.device("cpu")
            _get_device_index(cpu_device, optional=True)

    def test_serialization_array_with_empty(self):
        x = [torch.randn(4, 4).cuda(), torch.cuda.FloatTensor()]
        with tempfile.NamedTemporaryFile() as f:
            torch.save(x, f)
            f.seek(0)
            x_copy = torch.load(f)
        for original, copy in zip(x, x_copy):
            self.assertEqual(copy, original)
            self.assertIs(type(copy), type(original))
            self.assertEqual(copy.get_device(), original.get_device())

    @skipCUDANonDefaultStreamIf(True)
    def test_streams(self):
        default_stream = torch.cuda.current_stream()
        user_stream = torch.cuda.Stream()
        self.assertEqual(torch.cuda.current_stream(), default_stream)
        self.assertNotEqual(default_stream, user_stream)
        self.assertEqual(default_stream.cuda_stream, 0)
        self.assertNotEqual(user_stream.cuda_stream, 0)
        with torch.cuda.stream(user_stream):
            self.assertEqual(torch.cuda.current_stream(), user_stream)
        self.assertTrue(user_stream.query())
        tensor1 = torch.ByteTensor(5).pin_memory()
        default_stream.synchronize()
        self.assertTrue(default_stream.query())

    def test_stream_event_repr(self):
        s = torch.cuda.current_stream()
        self.assertTrue("torch.cuda.Stream" in s.__repr__())
        e = torch.cuda.Event()
        self.assertTrue("torch.cuda.Event" in e.__repr__())
        s.record_event(e)
        self.assertTrue("torch.cuda.Event" in e.__repr__())

    def test_cuda_stream_protocol(self):
        stream = torch.cuda.Stream()

        self.assertTrue(hasattr(stream, "__cuda_stream__"))

        result = stream.__cuda_stream__()

        self.assertIsInstance(result, tuple)
        self.assertEqual(len(result), 2)
        self.assertEqual(result[0], 0)  # Protocol version
        self.assertEqual(result[1], stream.cuda_stream)  # Stream handle

        external_stream = torch.cuda.ExternalStream(stream.cuda_stream)
        external_result = external_stream.__cuda_stream__()

        self.assertEqual(external_result[0], 0)
        self.assertEqual(external_result[1], external_stream.cuda_stream)

    def test_events(self):
        stream = torch.cuda.current_stream()
        event = torch.cuda.Event(enable_timing=True)
        self.assertTrue(event.query())
        start_event = torch.cuda.Event(enable_timing=True)
        stream.record_event(start_event)
        torch.cuda._sleep(int(50 * get_cycles_per_ms()))
        stream.record_event(event)
        self.assertFalse(event.query())
        event.synchronize()
        self.assertTrue(event.query())
        self.assertGreater(start_event.elapsed_time(event), 0)

        event = torch.cuda.Event(enable_timing=True)
        self.assertEqual(event.cuda_event, 0)
        self.assertEqual(event.event_id, 0)

        event.record()
        self.assertNotEqual(event.cuda_event, 0)
        self.assertNotEqual(event.event_id, 0)
        self.assertEqual(event.cuda_event, event.event_id)

    def test_events_elapsedtime(self):
        event1 = torch.cuda.Event(enable_timing=False)
        event2 = torch.cuda.Event(enable_timing=False)
        with self.assertRaisesRegex(
            ValueError,
            "Both events must be created with argument 'enable_timing=True'",
        ):
            event1.elapsed_time(event2)

        event1 = torch.cuda.Event(enable_timing=True)
        event2 = torch.cuda.Event(enable_timing=True)
        with self.assertRaisesRegex(
            ValueError, "Both events must be recorded before calculating elapsed time"
        ):
            event1.elapsed_time(event2)

        # check default value of enable_timing: False
        event1 = torch.cuda.Event()
        event2 = torch.cuda.Event()
        with self.assertRaisesRegex(
            ValueError,
            "Both events must be created with argument 'enable_timing=True'",
        ):
            event1.elapsed_time(event2)

    def test_generic_stream_event(self):
        stream = torch.Stream("cuda")
        self.assertEqual(stream.device_index, torch.cuda.current_device())
        cuda_stream = torch.cuda.Stream(
            stream_id=stream.stream_id,
            device_index=stream.device_index,
            device_type=stream.device_type,
        )
        self.assertIsInstance(cuda_stream, torch.Stream)
        self.assertTrue(issubclass(type(cuda_stream), torch.Stream))
        self.assertTrue(torch.Stream in type(cuda_stream).mro())
        self.assertEqual(stream.stream_id, cuda_stream.stream_id)
        self.assertNotEqual(stream.stream_id, torch.cuda.current_stream().stream_id)

        event1 = torch.Event("cuda", enable_timing=True)
        event2 = torch.Event("cuda", enable_timing=True)
        self.assertEqual(event1.event_id, 0)
        a = torch.randn(1000)
        b = torch.randn(1000)
        with torch.cuda.stream(cuda_stream):
            a_cuda = a.to("cuda", non_blocking=True)
            b_cuda = b.to("cuda", non_blocking=True)
            self.assertEqual(stream.stream_id, torch.cuda.current_stream().stream_id)
        event1.record(stream)
        event1.synchronize()
        self.assertTrue(event1.query())
        c_cuda = a_cuda + b_cuda
        event2.record()
        event2.synchronize()
        self.assertTrue(event2.query())
        self.assertNotEqual(event1.event_id, event2.event_id)
        self.assertEqual(c_cuda.cpu(), a + b)
        self.assertTrue(event1.elapsed_time(event2) > 0)
        cuda_event = torch.cuda.Event()
        self.assertIsInstance(cuda_event, torch.Event)
        self.assertTrue(issubclass(type(cuda_event), torch.Event))
        self.assertTrue(torch.Event in type(cuda_event).mro())

    @unittest.skip("Fails with Triton 3.7")
    def test_stream_event_compatibility(self):
        s1 = torch.cuda.Stream()
        s2 = torch.cuda.Stream()
        torch.accelerator.set_stream(s1)
        self.assertEqual(torch.accelerator.current_stream().stream_id, s1.stream_id)
        self.assertEqual(
            torch.accelerator.current_stream().native_handle, s1.cuda_stream
        )
        self.assertEqual(
            torch.accelerator.current_stream().native_handle, s1.native_handle
        )
        torch.accelerator.set_stream(s2)
        self.assertEqual(torch.accelerator.current_stream().stream_id, s2.stream_id)
        self.assertEqual(
            torch.accelerator.current_stream().native_handle, s2.cuda_stream
        )
        self.assertEqual(
            torch.accelerator.current_stream().native_handle, s2.native_handle
        )
        with self.assertRaisesRegex(
            RuntimeError, "Device index value .* is out of index range"
        ):
            torch.accelerator.current_stream(torch.accelerator.device_count())
        e1 = torch.cuda.Event(enable_timing=True)
        e2 = torch.cuda.Event(enable_timing=True)
        e3 = torch.Event(enable_timing=True)
        s3 = torch.Stream()
        with s3:
            self.assertEqual(torch.accelerator.current_stream(), s3)
            e1.record(s3)
            a = torch.randn(1000)
            a_cuda = a.to("cuda")
            del a_cuda
            e2.record(s3)
            e2.synchronize()
            self.assertGreater(e1.elapsed_time(e2), 0)
            e3.record(s3)
        with self.assertRaisesRegex(
            RuntimeError, "expected other to be a torch.cuda.Event object"
        ):
            e1.elapsed_time(e3)
        with self.assertRaisesRegex(
            RuntimeError, "expected other to be a torch.Event object"
        ):
            e3.elapsed_time(e1)
        with self.assertRaisesRegex(
            RuntimeError, "expected event to be a torch.Event object"
        ):
            s3.record_event(e1)
        with self.assertRaisesRegex(
            RuntimeError,
            "expected stream to be a torch.Stream or torch.cuda.Stream object",
        ):
            e2.record(e2)
        with self.assertRaisesRegex(
            RuntimeError,
            "expected stream to be a torch.Stream or torch.cuda.Stream object",
        ):
            e2.wait(e2)

    def test_record_stream(self):
        cycles_per_ms = get_cycles_per_ms()

        t = torch.FloatTensor([1, 2, 3, 4]).pin_memory()
        result = torch.cuda.FloatTensor(t.size())
        stream = torch.cuda.Stream()
        ptr = [None]

        # Performs the CPU->GPU copy in a background stream
        def perform_copy():
            with torch.cuda.stream(stream):
                tmp = t.cuda(non_blocking=True)
                ptr[0] = tmp.data_ptr()
            torch.cuda.current_stream().wait_stream(stream)
            tmp.record_stream(torch.cuda.current_stream())
            torch.cuda._sleep(int(50 * cycles_per_ms))  # delay the copy
            result.copy_(tmp)

        perform_copy()
        with torch.cuda.stream(stream):
            tmp2 = torch.cuda.FloatTensor(t.size())
            tmp2.zero_()
            self.assertNotEqual(
                tmp2.data_ptr(), ptr[0], msg="allocation reused to soon"
            )

        self.assertEqual(result.tolist(), [1, 2, 3, 4])

        if not TEST_CUDAMALLOCASYNC:
            # In the native allocator, we expect "tmp"'s side-stream-tagged block will be reused
            # in that side stream after result.copy_(tmp) in the main stream finishes.
            torch.cuda.current_stream().synchronize()
            with torch.cuda.stream(stream):
                tmp3 = torch.cuda.FloatTensor(t.size())
                self.assertEqual(tmp3.data_ptr(), ptr[0], msg="allocation not reused")

    def test_record_stream_on_shifted_view(self):
        # See issue #27366

        # This test detects unexpected block reallocation. For reliable test,
        # the stream to allocate tensors is isolated. The allocator will not
        # reuse free blocks which were allocated from another stream.
        stream_alloc = torch.cuda.Stream()
        with torch.cuda.stream(stream_alloc):
            base = torch.cuda.FloatTensor([10, 10])

        # Record another stream on a shifted view tensor.
        view = base[5:]
        self.assertTrue(view.storage_offset() > 0)

        stream_record = torch.cuda.Stream()
        with torch.cuda.stream(stream_record):
            torch.cuda._sleep(int(50 * get_cycles_per_ms()))

        view.record_stream(stream_record)

        # Delete those tensors to make the block free soon.
        data_ptr = base.data_ptr()
        del base, view

        # A new tensor should not be allocated to the block above.
        stream_alloc.synchronize()

        with torch.cuda.stream(stream_alloc):
            try_realloc = torch.cuda.FloatTensor([10, 10])

        self.assertNotEqual(try_realloc.data_ptr(), data_ptr)

    def test_device_context_manager(self):
        prev_device = torch.cuda.current_device()
        with torch.accelerator.device_index(None):
            self.assertEqual(torch.cuda.current_device(), prev_device)
        self.assertEqual(torch.cuda.current_device(), prev_device)
        with torch.accelerator.device_index(0):
            self.assertEqual(torch.cuda.current_device(), 0)
        self.assertEqual(torch.cuda.current_device(), prev_device)

    @unittest.skipIf(not TEST_MULTIGPU, "only one GPU detected")
    def test_multi_device_context_manager(self):
        src_device = 0
        dst_device = 1
        torch.cuda.set_device(src_device)
        with torch.accelerator.device_index(dst_device):
            self.assertEqual(torch.cuda.current_device(), 1)
        self.assertEqual(torch.cuda.current_device(), src_device)

    def test_stream_context_manager(self):
        prev_stream = torch.cuda.current_stream()
        with torch.cuda.Stream() as stream:
            self.assertEqual(stream, torch.cuda.current_stream())
        self.assertEqual(prev_stream, torch.cuda.current_stream())

    @unittest.skipIf(not TEST_MULTIGPU, "only one GPU detected")
    def test_multi_device_stream_context_manager(self):
        src_device = 0
        dst_device = 1
        torch.cuda.set_device(src_device)
        src_prev_stream = torch.cuda.current_stream(src_device)
        dst_prev_stream = torch.cuda.current_stream(dst_device)
        with torch.cuda.Stream(dst_device) as dst_stream:
            self.assertEqual(dst_device, torch.cuda.current_device())
            self.assertEqual(dst_stream, torch.cuda.current_stream())
            self.assertEqual(src_prev_stream, torch.cuda.current_stream(src_device))
        self.assertEqual(src_device, torch.cuda.current_device())
        self.assertEqual(src_prev_stream, torch.cuda.current_stream())
        self.assertEqual(dst_prev_stream, torch.cuda.current_stream(dst_device))

    def test_noncontiguous_pinned_memory(self):
        # See issue #3266
        x = torch.arange(0, 10).view((2, 5))
        self.assertEqual(x.t(), x.t().pin_memory())

    def test_caching_pinned_memory(self):
        cycles_per_ms = get_cycles_per_ms()

        # check that allocations are reused after deletion
        t = torch.FloatTensor([1]).pin_memory()
        ptr = t.data_ptr()
        del t
        t = torch.FloatTensor([1]).pin_memory()
        self.assertEqual(t.data_ptr(), ptr, msg="allocation not reused")

        # check that the allocation is not reused if it's in-use by a copy
        gpu_tensor = torch.cuda.FloatTensor([0])
        torch.cuda._sleep(int(1000 * cycles_per_ms))  # delay the copy by 1s
        gpu_tensor.copy_(t, non_blocking=True)
        del t
        t = torch.FloatTensor([1]).pin_memory()
        self.assertNotEqual(t.data_ptr(), ptr, msg="allocation reused too soon")
        self.assertEqual(list(gpu_tensor), [1])

    def test_caching_allocator_record_stream_oom(self):
        """allocations delayed by a record_stream call should still be freed on
        an out-of-memory in cuda_malloc_retry. see issue #19219"""
        stream = torch.cuda.Stream()

        with torch.cuda.stream(stream):
            y = torch.zeros(40 * 1024 * 1024, device="cuda")

        for _ in range(100):
            x = torch.empty(40 * 1024 * 1024, device="cuda")
            with torch.cuda.stream(stream):
                y += x
            # delays reuse of `x` until after all operations in `stream`
            x.record_stream(stream)
            del x

        # we've made a mess by allocating up to the device capacity. free any
        # cached blocks in case it affects future tests.
        torch.cuda.empty_cache()

    # Tests for historic illegal memory access, see #17040.
    def test_reduction_gpu_memory_accessing(self):
        x = torch.ones(512, 8, dtype=torch.float32, device="cuda")
        torch.sum(x, 0)

    def test_sum_fp16(self):
        x = torch.zeros(10, device="cuda", dtype=torch.float16)
        self.assertEqual(x.sum(), 0)

        x = torch.ones(65504, device="cuda", dtype=torch.float16)
        self.assertEqual(x.sum(), 65504)
        self.assertEqual(x.sum(dtype=torch.float32), 65504)

        x = torch.ones(65536, device="cuda", dtype=torch.float16)
        self.assertEqual(x.sum(dtype=torch.float32), 65536)

        a = torch.zeros(1203611).bernoulli_(0.0005)
        x = a.to(device="cuda", dtype=torch.float16)
        self.assertEqual(x.sum().item(), a.sum().item())

        a = torch.zeros(100, 121, 80).bernoulli_(0.0005)
        x = a.to(device="cuda", dtype=torch.float16)
        self.assertEqual(x.sum((0, 2)).float().cpu(), a.sum((0, 2)))

    def test_mean_fp16(self):
        x = torch.ones(65536, device="cuda", dtype=torch.float16)
        self.assertEqual(x.mean(), 1)

        x = torch.ones(65536, device="cuda", dtype=torch.float16)
        self.assertEqual(x.mean(dtype=torch.float32), 1)

    def test_prod_large(self):
        # tests global reduction (should_global_reduce = true) in case of non-zero identity element
        x = torch.ones(240000, device="cuda", dtype=torch.float32)
        self.assertEqual(x.prod(), 1)

        # test for complex types. Note 240k is divisible by 4
        for dtype in [torch.cfloat, torch.cdouble]:
            x = torch.ones(240000, device="cuda", dtype=dtype) * (0 + 1j)
            self.assertEqual(x.prod(), 1)

    def test_multinomial_ext(self):
        # Test two corner cases from older PyTorch (Issue #4858)
        freqs = torch.cuda.FloatTensor(
            [
                0.0,
                0.0,
                0.0,
                0.0,
                0.0,
                0.0,
                0.0,
                0.0,
                0.0,
                0.03178183361887932,
                0.027680952101945877,
                0.033176131546497345,
                0.046052902936935425,
                0.07742464542388916,
                0.11543981730937958,
                0.14148041605949402,
                0.15784293413162231,
                0.13180233538150787,
                0.08271478116512299,
                0.049702685326337814,
                0.027557924389839172,
                0.018125897273421288,
                0.011851548217236996,
                0.010252203792333603,
                0.007422595750540495,
                0.005372154992073774,
                0.0045109698548913,
                0.0036087757907807827,
                0.0035267581697553396,
                0.0018864056328311563,
                0.0024605290964245796,
                0.0022964938543736935,
                0.0018453967059031129,
                0.0010662291897460818,
                0.0009842115687206388,
                0.00045109697384759784,
                0.0007791675161570311,
                0.00020504408166743815,
                0.00020504408166743815,
                0.00020504408166743815,
                0.00012302644609007984,
                0.0,
                0.00012302644609007984,
                4.100881778867915e-05,
                0.0,
                0.0,
                0.0,
                0.0,
                0.0,
                0.0,
            ]
        )

        torch.cuda.manual_seed(11042)
        sample = torch.multinomial(freqs, 1000, True)
        self.assertNotEqual(freqs[sample].min(), 0)

        p = torch.zeros(3421, 2, device="cuda", dtype=torch.float)
        p[:, 1] = 1
        torch.cuda.manual_seed(5214)
        r = torch.multinomial(p, 1)
        self.assertNotEqual(r.min().item(), 0)

        # test corner case from Issue #13867
        torch.cuda.manual_seed(33)
        probs = torch.randn(1000000, device="cuda").clamp(min=0) * 3e-5
        samples = probs.multinomial(1000000, replacement=True)
        self.assertGreater(probs[samples].min().item(), 0)

    def _spawn_test_multinomial_invalid_probs_cuda(self, probs):
        import subprocess

        try:
            p = subprocess.Popen(
                [
                    sys.executable,
                    "-c",
                    f"""\
import sys
import torch
from torch import inf, nan
try:
    with torch.random.fork_rng(devices=[0]):
        torch.multinomial(torch.tensor({probs}).to('cuda'), 2, replacement=True)
        torch.cuda.synchronize()
    sys.exit(-1) # Should not be reached
except RuntimeError as e:
    sys.exit(-2)
""",
                ],
                stdout=subprocess.PIPE,
                stderr=subprocess.PIPE,
                universal_newlines=True,
            )
            out, err = p.communicate(timeout=10)
            p.wait(timeout=10)
        except subprocess.TimeoutExpired:
            p.kill()
            out, err = p.communicate()
        expected_messages = [
            "device-side assert triggered",  # CUDA
            "Assertion",  # CUDA
            "HSA_STATUS_ERROR_EXCEPTION",  # ROCm with TORCH_USE_HIP_DSA
            "Device-side assertion",  # ROCm with TORCH_USE_HIP_DSA
            # ROCm without TORCH_USE_HIP_DSA returns a launch failure instead
            # of a proper device-side assertion, but still catches the error
            "hipErrorLaunchFailure",
            "unspecified launch failure",
        ]
        self.assertTrue(any(msg in out or msg in err for msg in expected_messages))

    @slowTest
    def test_multinomial_invalid_probs_cuda(self):
        self._spawn_test_multinomial_invalid_probs_cuda([1.0, -1.0, 1.0])
        self._spawn_test_multinomial_invalid_probs_cuda([1.0, inf, 1.0])
        self._spawn_test_multinomial_invalid_probs_cuda([1.0, -inf, 1.0])
        self._spawn_test_multinomial_invalid_probs_cuda([1.0, 1.0, nan])

    @slowTest
    def test_index_out_of_bounds_exception_cuda(self):
        # Test that indexing out of bounds causes assert
        stderr = TestCase.runWithPytorchAPIUsageStderr("""\
#!/usr/bin/env python3

import torch
from torch.testing._internal.common_utils import (TestCase, run_tests, slowTest)

class TestThatContainsCUDAAssertFailure(TestCase):

    @slowTest
    def test_index_bounds_cuda(self):
        x = torch.arange(10, device="cuda")
        y = x[torch.tensor([11])].cpu()

if __name__ == '__main__':
    run_tests()
""")
        # CUDA raises cudaErrorAssert (710) with "device-side assert triggered"
        # ROCm with TORCH_USE_HIP_DSA raises a proper device-side assertion
        # ROCm without TORCH_USE_HIP_DSA raises hipErrorLaunchFailure (719)
        # which still catches the error but with less specific messaging
        is_cuda_assert = "device-side assert triggered" in stderr
        is_hip_assert = "hipErrorLaunchFailure" in stderr
        is_hip_assert = is_hip_assert or "unspecified launch failure" in stderr
        is_hip_assert = is_hip_assert or "HSA_STATUS_ERROR_EXCEPTION" in stderr
        self.assertTrue(
            is_cuda_assert or is_hip_assert,
            f"Expected device assert error in stderr, got: {stderr}",
        )

    @slowTest
    @unittest.skipIf(not TEST_LARGE_TENSOR, "not enough memory")
    @serialTest()
    def test_huge_index(self):
        src = torch.empty(15000000, 45, device="cuda", dtype=torch.long).random_(
            0, 2**22
        )
        idx = torch.randperm(src.shape[0], device="cuda")
        res = src[idx]
        res_cpu = src.cpu()[idx.cpu()]
        self.assertEqual(res.cpu(), res_cpu)

    def test_randint_randomness_for_large_range(self) -> None:
        # For large ranges, randint generation is slightly different. This lead to a subtle bug where some Philox
        # offsets were not calculated correctly, resulting in reused random states.
        # See https://github.com/pytorch/pytorch/issues/125224
        size = 1_000_000
        high = 6_000_000_000  # Keep this above 2**32

        def run(dev: torch.device) -> int:
            # Measure how many unique numbers are generated in 2 consecutive calls to randint. If random states are
            # reused, this will yield fewer unique numbers.
            gen = torch.Generator(device=dev)
            gen.manual_seed(0)
            t1 = torch.randint(
                0, high, [size], device=dev, generator=gen, dtype=torch.int64
            )
            t2 = torch.randint(
                0, high, [size], device=dev, generator=gen, dtype=torch.int64
            )
            return torch.stack([t1, t2]).unique().shape[0]

        # Use CPU as reference. The results should not deviate too much.
        self.assertTrue(
            abs(run(torch.device("cuda")) - run(torch.device("cpu"))) < 10_000
        )

    @largeTensorTest("20GB", "cuda")
    @serialTest()
    def test_randint_generation_for_large_numel(self) -> None:
        numel = 2**31 + 1
        s = torch.randint(2, (numel,), device="cuda", dtype=torch.int8).sum()
        self.assertTrue(s > 0, "expected randint in [0, 1] to generate nonzero values")

    @parametrize("dtype", [torch.float32, torch.double])
    def test_random_no_reused_random_states(self, dtype: torch.dtype) -> None:
        # Test if random states do not overlap between consecutive rand/randn calls.
        # See https://github.com/pytorch/pytorch/issues/125224

        def run(func, dev: torch.device, dtype: torch.dtype) -> int:
            # Measure how many unique numbers are generated in 2 consecutive calls. If random states are
            # reused, this will yield fewer unique numbers.
            size = 1000000
            gen = torch.Generator(device=dev)
            gen.manual_seed(0)
            t1 = func((size,), device=dev, generator=gen, dtype=dtype)
            t2 = func((size,), device=dev, generator=gen, dtype=dtype)
            return torch.stack([t1, t2]).unique().shape[0]

        # Use CPU as reference. The results should not deviate too much.
        for func in [torch.rand, torch.randn]:
            deviation = abs(
                run(func, torch.device("cuda"), dtype)
                - run(func, torch.device("cpu"), dtype)
            )
            self.assertTrue(deviation < 50_000, deviation)

    def test_min_max_inits(self):
        # Testing if THC_reduceAll received the correct index initialization.
        # This affects the result of THC_reduceAll operations at extreme values
        x = torch.cuda.ByteTensor([0])
        y = torch.cuda.ByteTensor([255])
        expected = torch.cuda.LongTensor([0])[0]

        _, v = x.max(dim=0)
        self.assertEqual(v, expected)

        _, v = y.min(dim=0)
        self.assertEqual(v, expected)

    def test_nvtx(self):
        # Just making sure we can see the symbols
        torch.cuda.nvtx.range_push("foo")
        torch.cuda.nvtx.mark("bar")
        torch.cuda.nvtx.range_pop()
        range_handle = torch.cuda.nvtx.range_start("range_start")
        torch.cuda.nvtx.range_end(range_handle)

    def test_bincount_ext(self):
        # ensure CUDA code coverage
        input_size = (100000,)
        w = torch.randn(input_size, dtype=torch.double, device="cuda")
        w_cpu = w.cpu()
        # test shared memory impl
        t = torch.randint(50, input_size, dtype=torch.int8, device="cuda")
        self.assertEqual(t.cpu().bincount(), t.bincount())
        self.assertEqual(t.cpu().bincount(w_cpu), t.bincount(w))
        # test global memory impl
        #   see `CUDAHistogramMemoryType` in SummaryOps.cu
        #   50000 * sizeof(int64_t) == 390 KiB, which should exceed smem of any known GPU
        t = torch.randint(50000, input_size, dtype=torch.int64, device="cuda")
        self.assertEqual(t.cpu().bincount(), t.bincount())
        self.assertEqual(t.cpu().bincount(w_cpu), t.bincount(w))

        t = torch.zeros([10], dtype=torch.int32, device="cuda")
        # 35488 * 65536 as int32 would cause overflow to negative value
        # giving negative bin offset
        t[0] = 35488
        counted = t.bincount(minlength=65536)
        self.assertEqual(torch.sum(counted), 10)

    def test_tiny_half_norm_(self):
        a = torch.arange(25).cuda().float()
        a /= 100000000
        b = a.half()
        self.assertGreater(b.norm().item(), 0)

    def test_norm_type_conversion(self):
        a = torch.ones(65536).cuda().half()
        self.assertEqual(a.norm(p=0, dtype=torch.float32), 65536)

    @unittest.skipIf(not TEST_MEDIUM_TENSOR, "not enough memory")
    @serialTest()
    def test_cuda_opaque_type(self):
        x = torch.ones(600_000_000, dtype=torch.int32, device="cuda")
        y = torch.where(x > 0, x, x)
        self.assertEqual(y, x)

    def test_cuda_memory_leak_detection_propagates_errors(self):
        with self.assertRaisesRegex(
            RuntimeError, r"The size of tensor a \(3\) must match"
        ):
            with self.assertLeaksNoCudaTensors():
                x = torch.randn(3, 1, device="cuda")
                y = torch.randn(2, 1, device="cuda")
                x + y

    @unittest.skipIf(not TEST_MEDIUM_TENSOR, "not enough memory")
    @serialTest()
    def test_cuda_kernel_loop_overflow(self):
        # Issue #24309: In extreme cases, the loop variable could overflow and continue
        # the kernel loop with a negative index, causing a RuntimeError (invalid write):
        x = torch.randn(1, 1, 1, 2**30 + 1, dtype=torch.float16, device="cuda")
        expected = x[0, 0, 0, 2**30]
        y = torch.nn.functional.avg_pool2d(x, kernel_size=1)
        torch.cuda.synchronize()
        self.assertEqual(y[0, 0, 0, 2**30], expected)

    @unittest.skipIf(not TEST_LARGE_TENSOR, "not enough memory")
    @gcIfJetson
    @serialTest()
    def test_cuda_kernel_loop_overflow_large(self):
        # Make sure input.numel() > INT_MAX is handled:
        x = torch.randn(1, 1, 1, 2**31, dtype=torch.float16, device="cuda")
        with self.assertRaisesRegex(RuntimeError, "value cannot be converted to type"):
            y = torch.nn.functional.avg_pool2d(x, kernel_size=1)

        # Issue #24309: In extreme cases, the loop variable could overflow and continue
        # the kernel loop with a negative index, causing a RuntimeError (invalid write):
        x = torch.randn(1, 1, 1, 2**31 - 1, dtype=torch.float16, device="cuda")
        expected = x[0, 0, 0, 2**31 - 2]
        y = torch.nn.functional.avg_pool2d(x, kernel_size=1)
        torch.cuda.synchronize()
        self.assertEqual(y[0, 0, 0, 2**31 - 2], expected)

    # this might create a reference cycle on self...
    def _make_multiply_in_stream(self):
        class MultiplyInStream(torch.autograd.Function):
            @staticmethod
            def forward(ctx, x, val):
                ctx.val = val
                ctx.stream = torch.cuda.current_stream()
                return x * val

            @staticmethod
            def backward(ctx, grad):
                self.assertEqual(torch.cuda.current_stream(), ctx.stream)
                # delays the operation in the background stream
                torch.cuda._sleep(1000 * 5000)
                return grad * ctx.val, None

        return MultiplyInStream

    @skipCUDANonDefaultStreamIf(True)
    def test_streaming_backwards_sync(self):
        default_stream = torch.cuda.current_stream()
        stream = torch.cuda.Stream()

        MultiplyInStream = self._make_multiply_in_stream()

        # Tests using grads outside the backward() stream context
        # See "Stream semantics of backward passes" on https://pytorch.org/docs/stable/notes/cuda.html
        x = torch.randn(5, 5, device="cuda", requires_grad=True)
        with torch.cuda.stream(stream):
            stream.wait_stream(default_stream)
            output = MultiplyInStream.apply(x, 2)
            output.sum().backward()
        # sync needed
        default_stream.wait_stream(stream)
        self.assertEqual(x.grad, torch.ones_like(x) * 2)
        self.assertEqual(torch.cuda.current_stream(), default_stream)

        # Tests that using grads in the same stream context as backward()
        # is safe regardless what streams bwd ops ran on
        bwd_ambient_stream = torch.cuda.Stream()
        x = torch.randn(5, 5, device="cuda", requires_grad=True)
        with torch.cuda.stream(stream):
            stream.wait_stream(default_stream)
            output = MultiplyInStream.apply(x, 3)
        with torch.cuda.stream(bwd_ambient_stream):
            bwd_ambient_stream.wait_stream(stream)
            output.sum().backward()
            # x was first used on "stream" so its AccumulateGrad leaf should run on "stream".
            # The end of backward() should have synced "bwd_ambient_stream" with "stream"
            # so it should be safe to use x.grad here without any syncs.
            self.assertEqual(x.grad, torch.ones_like(x) * 3)
            self.assertEqual(torch.cuda.current_stream(), bwd_ambient_stream)

    def test_streaming_backwards_multiple_streams(self):
        MultiplyInStream = self._make_multiply_in_stream()

        class StreamModel(torch.nn.Module):
            def __init__(self) -> None:
                super().__init__()
                self.event = torch.cuda.Event()
                self.stream0 = torch.cuda.Stream()
                self.stream1 = torch.cuda.Stream()

            def forward(self, x, x_first_use_on_ambient):
                if x_first_use_on_ambient:
                    x0 = x.clone()
                self.stream0.wait_stream(torch.cuda.current_stream())
                self.stream1.wait_stream(torch.cuda.current_stream())
                with torch.cuda.stream(self.stream0):
                    if not x_first_use_on_ambient:
                        x0 = x.clone()
                    y0 = MultiplyInStream.apply(x0, 2)
                    self.event.record(stream=torch.cuda.current_stream())

                with torch.cuda.stream(self.stream1):
                    y1 = MultiplyInStream.apply(x, 3)
                    self.stream1.wait_event(self.event)
                    return y0 + y1

        stream = torch.cuda.Stream()

        for x_first_use_on_ambient in (True, False):
            # the out_of_place=False, iters=1 case stresses if proper syncs are inserted
            # when grads are initially None and stolen by backward ops.
            for out_of_place, iters in ((True, 1), (False, 1), (False, 5)):
                with torch.cuda.stream(stream):
                    x = torch.randn(5, 5, device="cuda", requires_grad=True)
                    model = StreamModel().cuda()
                    x.register_hook(
                        lambda grad: self.assertEqual(
                            torch.cuda.current_stream(),
                            stream if x_first_use_on_ambient else model.stream0,
                        )
                    )
                    for p in model.parameters():
                        self.assertTrue(p.grad is None)
                    for _ in range(iters):
                        loss = model(x, x_first_use_on_ambient).sum()
                        if out_of_place:
                            x_grad = torch.autograd.grad((loss,), (x,))[0]
                        else:
                            loss.backward()
                # See "Stream semantics of backward passes" on https://pytorch.org/docs/stable/notes/cuda.html
                torch.cuda.current_stream().wait_stream(stream)

                if out_of_place:
                    self.assertEqual(x_grad, torch.ones_like(x) * 5 * iters)
                else:
                    self.assertEqual(x.grad, torch.ones_like(x) * 5 * iters)

    def test_streaming_backwards_sync_graph_root(self):
        # This function tests if bwd ops running on a side stream properly sync with the GraphRoot.
        # The potential bug it targets is a race condition. The test uses multiple trials and
        # torch.cuda._sleep such that if the race condition exists, the test will almost certainly fail,
        # but there's a chance it may spuriously pass. Passing does not guarantee the backend is bug-free,
        # but failure does guarantee there is a bug.
        fwd_bwd_op_stream = torch.cuda.Stream()
        bwd_ambient_stream = torch.cuda.Stream()
        # We need these streams to be different otherwise the test is meaningless.
        self.assertTrue(fwd_bwd_op_stream != bwd_ambient_stream)

        size = int(1e3)

        a = torch.full((size,), 2.0, device="cuda", requires_grad=True)
        b = torch.full((size,), 3.0, device="cuda", requires_grad=True)

        # I don't think we need any manual record_streams below.
        # a and b remain in scope for the entire test.
        # c and grad remain in scope for each iteration, and there's a full sync between iterations.
        for trial in range(5):
            torch.cuda.synchronize()
            a.grad = b.grad = None
            with torch.cuda.stream(fwd_bwd_op_stream):
                c = a * b

            with torch.cuda.stream(bwd_ambient_stream):
                torch.cuda.synchronize()
                # Long-running dummy kernel on bwd_ambient_stream delays filling of grad
                torch.cuda._sleep(int(50 * get_cycles_per_ms()))
                # Fills grad on bwd_ambient_stream
                grad = torch.full((size,), float(trial + 1), device="cuda")

                # Bwd ops still run on fwd_bwd_ops_stream, so the following will likely fail if
                # bwd ops don't sync with bwd_ambient_stream before consuming grad.
                torch.autograd.backward(tensors=c, grad_tensors=grad)

                # See https://github.com/pytorch/pytorch/issues/47028
                # assertEquals below run on bwd_ambient_stream, so this test may also fail
                # if backward() fails to sync with bwd_ambient_stream at the end.
                # Synchronizing here works around the issue until a proper fix can be made.
                torch.cuda.synchronize()
                with torch.no_grad():
                    self.assertEqual(a.grad, grad * b)
                    self.assertEqual(b.grad, grad * a)

    def test_streaming_backwards_callback(self):
        # Tests if autograd callbacks sync properly with respect to leaf streams and
        # the user-facing stream surrounding backward(). If it fails, first suspect is
        # sync logic where  "final_callbacks_" are called in torch/csrc/autograd/engine.cpp
        MultiplyInStream = self._make_multiply_in_stream()

        size = int(1e3)
        a = torch.full((size,), 1, device="cuda", dtype=torch.float, requires_grad=True)
        b = torch.full((size,), 1, device="cuda", dtype=torch.float, requires_grad=True)

        s0 = torch.cuda.Stream()
        s1 = torch.cuda.Stream()
        s2 = torch.cuda.Stream()

        stash = []

        # sets up a nontrivial structure of leaf streams
        s0.wait_stream(torch.cuda.current_stream())
        with torch.cuda.stream(s0):
            c = MultiplyInStream.apply(a, 2)

        s1.wait_stream(torch.cuda.current_stream())
        with torch.cuda.stream(s1):
            d = MultiplyInStream.apply(b, 3)
            s1.wait_stream(s0)
            e = c * d

            def clone_leaf_grads():
                stash.append(a.grad.clone())
                stash.append(b.grad.clone())

            # Use a hook on e to install the callback
            e.register_hook(
                lambda grad: torch.autograd.Variable._execution_engine.queue_callback(
                    clone_leaf_grads
                )
            )

        s2.wait_stream(s1)
        with torch.cuda.stream(s2):
            e.sum().backward()
            # The autograd engine should sync s2 with all leaf streams then run the callback clone_leaf_grads on s2.
            # If those things happened properly, checking the values of the cloned grads on s2 should be safe:
            self.assertEqual(stash[0], torch.full_like(a, 6))
            self.assertEqual(stash[1], torch.full_like(a, 6))

    def test_fixed_cuda_assert_async(self):
        with self.assertRaisesRegex(
            RuntimeError, "Boolean value of Tensor with no values is ambiguous"
        ):
            torch._assert_async(torch.tensor([], device="cuda"))
        with self.assertRaisesRegex(
            RuntimeError,
            "Boolean value of Tensor with more than one value is ambiguous",
        ):
            torch._assert_async(torch.tensor([0, 0], device="cuda"))

        torch._assert_async(torch.tensor(1, device="cuda"))
        torch._assert_async(torch.tensor(0.1, device="cuda"))
        torch._assert_async(torch.tensor(-0.1, device="cuda"))
        torch._assert_async(torch.tensor(True, device="cuda"))
        torch._assert_async(torch.tensor(0 + 0.1j, device="cuda"))

        fail_stmts = [
            "torch._assert_async(torch.tensor(0, device='cuda'))",
            "torch._assert_async(torch.tensor(0.0, device='cuda'))",
            "torch._assert_async(torch.tensor(False, device='cuda'))",
            "torch._assert_async(torch.tensor(0 + 0j, device='cuda'))",
        ]

        import subprocess

        for stmt in fail_stmts:
            with self.subTest(stmt=stmt):
                r = subprocess.call(
                    [
                        sys.executable,
                        "-c",
                        f"""\
import torch

{stmt}
torch.cuda.synchronize()
""",
                    ]
                )
                self.assertTrue(r != 0)

    @unittest.skipIf(TEST_CUDAMALLOCASYNC, "FAIL")
    def test_cublas_multiple_threads_same_device(self):
        # Note, these parameters should be very carefully tuned
        # Too small number makes it hard for the racing condition
        # to happen, while too large number sometimes cause hang
        size = 1024
        num_threads = 2
        trials = 3
        test_iters = 100

        weight = torch.ones((size, size), device="cuda")
        results = {}
        barrier = threading.Barrier(num_threads)

        def _worker(t):
            my_stream = torch.cuda.Stream()
            # Hard sync so we don't need to worry about creating and using tensors
            # across streams or the fact that default streams are thread-local.
            # Those issues are not the target of this test.
            torch.cuda.synchronize()
            # Line up threads to increase likelihood of race conditions.
            barrier.wait()
            with torch.cuda.stream(my_stream):
                for _ in range(test_iters):
                    # If all threads are sharing the same cublas handle,
                    # the following sequence may occur:
                    # thread 0 calls cublasSetStream()
                    # thread 1 calls cublasSetStream()
                    # thread 0 launches its raw gemm, which it thinks is in
                    #          its own stream, but is actually in thread 1's stream.
                    # thread 0 enqueues its div_, which IS in its own stream,
                    #          but actually now races with its gemm.
                    results[t] = torch.mm(results[t], weight)
                    results[t].div_(float(size))
            torch.cuda.synchronize()

        for _ in range(trials):
            for t in range(num_threads):
                results[t] = torch.ones((size, size), device="cuda")

            threads = [
                threading.Thread(target=_worker, args=(t,)) for t in range(num_threads)
            ]

            for thread in threads:
                thread.start()
            for thread in threads:
                thread.join()

            for t in range(num_threads):
                self.assertEqual(results[t].sum().item(), size * size)

    # Test is flaky on Windows (https://github.com/pytorch/pytorch/issues/57401)
    @unittest.skipIf(IS_WINDOWS, "Test is flaky on Windows (see issue 57401)")
    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_cudnn_multiple_threads_same_device(self):
        # This function is intended to test the lazy creation and reuse of per-thread
        # cudnn handles on each device in aten/src/ATen/cudnn/Handles.cpp.
        # Failure here likely indicates something wrong with that logic.
        weight = torch.ones((1, 1, 2, 2), device="cuda")

        results = {}

        num_threads = 2
        trials = 3
        test_iters = 1000
        barrier = threading.Barrier(num_threads)

        with torch.backends.cudnn.flags(enabled=True):

            def _worker(t):
                my_stream = torch.cuda.Stream()
                # Hard sync so we don't need to worry about creating and using tensors
                # across streams or the fact that default streams are thread-local.
                # Those issues are not the target of this test.
                torch.cuda.synchronize()
                # Line up threads to increase likelihood of race conditions.
                barrier.wait()
                with torch.cuda.stream(my_stream):
                    for _ in range(test_iters):
                        # If all threads are sharing the same cudnn handle,
                        # the following sequence may occur:
                        # thread 0 calls setCuDNNStreamToCurrent()
                        # thread 1 calls setCuDNNStreamToCurrent()
                        # thread 0 launches its raw convolution, which it thinks is in
                        #          its own stream, but is actually in thread 1's stream.
                        # thread 0 enqueues its div_, which IS in its own stream,
                        #          but now races with its convolution.
                        results[t] = torch.nn.functional.conv2d(
                            results[t], weight, padding=0
                        )
                        results[t].div_(4.0)
                torch.cuda.synchronize()

            for _ in range(trials):
                for t in range(num_threads):
                    results[t] = torch.ones((1, 1, 2048, 2048), device="cuda")

                threads = [
                    threading.Thread(target=_worker, args=(t,))
                    for t in range(num_threads)
                ]

                for thread in threads:
                    thread.start()
                for thread in threads:
                    thread.join()

                for t in range(num_threads):
                    self.assertEqual(
                        results[t].sum().item(),
                        (2048 - test_iters) * (2048 - test_iters),
                    )

    def test_cusparse_multiple_threads_same_device(self):
        size = 1024
        num_threads = 2
        trials = 3
        test_iters = 500

        def ones_sparse(size):
            a = torch.arange(size, device="cuda")
            indices = torch.cartesian_prod(a, a).t()
            values = torch.ones(size * size, device="cuda")
            return torch.sparse_coo_tensor(indices, values)

        weight = ones_sparse(size)
        results = {}
        barrier = threading.Barrier(num_threads)

        def _worker(t):
            my_stream = torch.cuda.Stream()
            # Hard sync so we don't need to worry about creating and using tensors
            # across streams or the fact that default streams are thread-local.
            # Those issues are not the target of this test.
            torch.cuda.synchronize()
            # Line up threads to increase likelihood of race conditions.
            barrier.wait()
            with torch.cuda.stream(my_stream):
                for _ in range(test_iters):
                    # If all threads are sharing the same cublas handle,
                    # the following sequence may occur:
                    # thread 0 calls cublasSetStream()
                    # thread 1 calls cublasSetStream()
                    # thread 0 launches its raw gemm, which it thinks is in
                    #          its own stream, but is actually in thread 1's stream.
                    # thread 0 enqueues its div_, which IS in its own stream,
                    #          but actually now races with its gemm.
                    results[t] = weight.mm(results[t])
                    results[t].div_(float(size))
            torch.cuda.synchronize()

        for _ in range(trials):
            for t in range(num_threads):
                results[t] = torch.ones((size, size), device="cuda")

            threads = [
                threading.Thread(target=_worker, args=(t,)) for t in range(num_threads)
            ]

            for thread in threads:
                thread.start()
            for thread in threads:
                thread.join()

            for t in range(num_threads):
                self.assertEqual(results[t].sum().item(), size * size)

    @slowTest
    @unittest.skipIf(not TEST_LARGE_TENSOR, "not enough memory")
    @serialTest()
    def test_max_large_axis(self):
        x = torch.zeros(2**32, device="cuda", dtype=torch.int8)
        x[-1] = 1
        val, idx = x.max(0)
        self.assertEqual(val, 1)
        self.assertEqual(idx, x.shape[0] - 1)

    @unittest.skipIf(not TEST_NUMPY, "Numpy not found")
    def test_to_numpy(self):
        self.assertRaises(TypeError, lambda: torch.empty(1, device="cuda").numpy())

    def test_graph_is_current_stream_capturing(self):
        self.assertFalse(torch.cuda.is_current_stream_capturing())

        if TEST_CUDA and (not TEST_WITH_ROCM):
            s = torch.cuda.Stream()
            with torch.cuda.stream(s):
                g = torch.cuda.CUDAGraph()
                self.assertFalse(torch.cuda.is_current_stream_capturing())
                self.assertFalse(s.is_capturing())
                g.capture_begin()
                self.assertTrue(torch.cuda.is_current_stream_capturing())
                self.assertTrue(s.is_capturing())
                g.capture_end()

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_capture_simple(self):
        s = torch.cuda.Stream()

        with torch.cuda.stream(s):
            a = torch.full((1000,), 1, device="cuda")
            g = torch.cuda.CUDAGraph()
            torch.cuda.empty_cache()
            g.capture_begin()
            b = a
            for _ in range(10):
                b = b + 1
            g.capture_end()
        torch.cuda.current_stream().wait_stream(s)

        g.replay()

        self.assertEqual(b.sum().item(), 11000.0)

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_capture_stale_default_stream_error(self):
        """Default-stream warmup + side-stream capture raises a clear error
        (not an opaque CUDA crash) when override is not enabled."""
        default_stream = torch.cuda.default_stream()
        with torch.cuda.stream(default_stream):
            model = torch.nn.Linear(32, 32).cuda()
            x = torch.ones(4, 32, device="cuda")
            loss = model(x).sum()
            loss.backward()
            model.zero_grad()

        s = torch.cuda.Stream()
        static_x = torch.ones(4, 32, device="cuda")

        with torch.cuda.stream(s):
            static_loss = model(static_x).sum()
            static_loss.backward()
            model.zero_grad()
            torch.cuda.synchronize()
            g = torch.cuda.CUDAGraph()
            g.capture_begin(capture_error_mode="relaxed")
            try:
                with self.assertRaisesRegex(
                    RuntimeError, "stale reference to the default stream"
                ):
                    static_loss = model(static_x).sum()
                    static_loss.backward()
            finally:
                g.capture_end()
        torch.cuda.synchronize()

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_capture_override_stale_stream_default_warmup(self):
        """override_stale_capture_stream fixes default-stream warmup +
        side-stream capture."""
        default_stream = torch.cuda.default_stream()
        with torch.cuda.stream(default_stream):
            model = torch.nn.Linear(32, 32).cuda()
            x = torch.ones(4, 32, device="cuda")
            loss = model(x).sum()
            loss.backward()
            model.zero_grad()

        s = torch.cuda.Stream()
        static_x = torch.ones(4, 32, device="cuda")

        torch.autograd.graph.set_override_stale_capture_stream(True)
        try:
            with torch.cuda.stream(s):
                static_loss = model(static_x).sum()
                static_loss.backward()
                model.zero_grad()
                torch.cuda.synchronize()
                g = torch.cuda.CUDAGraph()
                g.capture_begin()
                static_loss = model(static_x).sum()
                static_loss.backward()
                g.capture_end()
            torch.cuda.current_stream().wait_stream(s)
            g.replay()
            torch.cuda.synchronize()
            for p in model.parameters():
                self.assertIsNotNone(p.grad)
        finally:
            torch.autograd.graph.set_override_stale_capture_stream(False)

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_capture_override_stale_stream_side_warmup(self):
        """override_stale_capture_stream fixes side-stream warmup +
        different side-stream capture."""
        warmup_stream = torch.cuda.Stream()
        capture_stream = torch.cuda.Stream()

        with torch.cuda.stream(warmup_stream):
            model = torch.nn.Linear(32, 32).cuda()
            x = torch.ones(4, 32, device="cuda")
            loss = model(x).sum()
            loss.backward()
            model.zero_grad()
            static_x = torch.ones(4, 32, device="cuda")

        torch.cuda.synchronize()

        torch.autograd.graph.set_override_stale_capture_stream(True)
        try:
            with torch.cuda.stream(capture_stream):
                static_loss = model(static_x).sum()
                static_loss.backward()
                model.zero_grad()
                torch.cuda.synchronize()
                g = torch.cuda.CUDAGraph()
                g.capture_begin()
                static_loss = model(static_x).sum()
                static_loss.backward()
                g.capture_end()

            torch.cuda.current_stream().wait_stream(capture_stream)
            g.replay()
            torch.cuda.synchronize()

            for p in model.parameters():
                self.assertIsNotNone(p.grad)
        finally:
            torch.autograd.graph.set_override_stale_capture_stream(False)

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graphsafe_set_get_rng_state(self):
        # Define a function to create generator states, with optional graph registration
        def create_states(generator):
            """Initializes generator states and registers them with a CUDA graph if provided."""
            # Ensure the CUDA generator is initialized
            torch.rand(1, device="cuda")
            generator.manual_seed(0)

            # Save the current state of the generator
            old_state = generator.graphsafe_get_state()
            # Create and save a cloned state of the generator
            new_state = generator.clone_state()
            # Return the original generator and its two states
            return generator, old_state, new_state

        def register_states_to_graph(generator_state, graph):
            _, old_state, new_state = generator_state
            graph.register_generator_state(old_state)
            graph.register_generator_state(new_state)

        # Define a function to perform specific RNG actions using the generator's states
        def perform_random_generation_steps(generator_state):
            generator, old_state, new_state = generator_state
            random_values = []

            # Generate random numbers with the new generator state
            generator.graphsafe_set_state(new_state)
            random_values.append(torch.rand(5, device="cuda", generator=generator))

            # Generate random numbers twice with the old generator state
            generator.graphsafe_set_state(old_state)
            random_values.extend(
                [torch.rand(5, device="cuda", generator=generator) for _ in range(2)]
            )

            return random_values

        # Define a function to retrieve the final offsets of the original and new generator states
        def get_final_offsets_of_states(generator_state):
            _, old_state, new_state = generator_state
            old_state_offset = old_state.get_offset()
            new_state_offset = new_state.get_offset()
            return old_state_offset, new_state_offset

        # Set up and test a new CUDA generator
        generator = torch.Generator(device="cuda")
        generator_state = create_states(generator)

        # Set up and test the default CUDA generator with a CUDA Graph
        g = torch.cuda.CUDAGraph()
        s = torch.cuda.Stream()
        default_generator = torch.cuda.default_generators[0]
        default_generator_state = create_states(default_generator)
        register_states_to_graph(default_generator_state, g)

        # Perform random number generation within a CUDA graph
        with torch.cuda.stream(s):
            g.capture_begin()
            graphed_random_values = perform_random_generation_steps(
                default_generator_state
            )
            g.capture_end()

        # Synchronize the streams and replay the graph
        torch.cuda.current_stream().wait_stream(s)
        for _ in range(3):
            random_values = perform_random_generation_steps(generator_state)
            g.replay()
            offset = get_final_offsets_of_states(generator_state)
            graph_offset = get_final_offsets_of_states(default_generator_state)

            # Compare the final offsets of states for both generators to ensure consistency
            self.assertEqual(offset, graph_offset)
            # Compare the states generated outside and inside the graph
            self.assertEqual(random_values, graphed_random_values)

    @xfailCUDAIfSM89OrLaterOnWindows
    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_rng_after_failed_capture(self):
        """Test that a stream can be captured again for RNG after a failed capture."""
        if TEST_WITH_ROCM and self.expandable_segments:
            self.skipTest(
                "ROCm expandable segments has known issue with graph capture recovery - #179911"
            )
        stream = torch.cuda.Stream()
        graph = torch.cuda.CUDAGraph()
        x = torch.ones(1, device="cuda")

        with torch.cuda.stream(stream):
            graph.capture_begin()
            with self.assertRaises(RuntimeError):
                (x + 1).item()
            with self.assertRaises(RuntimeError):
                graph.capture_end()

        torch.cuda.current_stream().wait_stream(stream)

        result = torch.randn(4, device="cuda")
        self.assertEqual(result.shape, (4,))

        new_graph = torch.cuda.CUDAGraph()
        buf = torch.empty(4, device="cuda")
        with torch.cuda.stream(stream):
            new_graph.capture_begin()
            buf.copy_(torch.randn_like(buf))
            new_graph.capture_end()
        torch.cuda.current_stream().wait_stream(stream)
        buf.zero_()
        new_graph.replay()
        torch.cuda.synchronize()
        self.assertFalse(torch.allclose(buf, torch.zeros_like(buf)))

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_rng_concurrent_replay_on_different_streams(self):
        """Concurrent replay of two graphs sharing a generator on different streams.

        With per-generator state (old code), this would race on the shared
        rng_state_offset_extragraph_ tensor. With per-(generator, capture_id)
        state, each graph has its own tensors.
        """
        seed = 1234
        shape = (64,)

        torch.manual_seed(seed)
        ref0 = torch.randn(shape, device="cuda")
        ref1 = torch.randn(shape, device="cuda")

        torch.manual_seed(seed)
        g0 = torch.cuda.CUDAGraph()
        g1 = torch.cuda.CUDAGraph()
        s_cap = torch.cuda.Stream()
        buf0 = torch.empty(shape, device="cuda")
        buf1 = torch.empty(shape, device="cuda")

        with torch.cuda.stream(s_cap):
            g0.capture_begin()
            buf0.copy_(torch.randn_like(buf0))
            g0.capture_end()

        torch.cuda.current_stream().wait_stream(s_cap)

        with torch.cuda.stream(s_cap):
            g1.capture_begin()
            buf1.copy_(torch.randn_like(buf1))
            g1.capture_end()

        torch.cuda.current_stream().wait_stream(s_cap)

        s0 = torch.cuda.Stream()
        s1 = torch.cuda.Stream()

        buf0.zero_()
        buf1.zero_()

        s0.wait_stream(torch.cuda.current_stream())
        s1.wait_stream(torch.cuda.current_stream())

        with torch.cuda.stream(s0):
            g0.replay()
        with torch.cuda.stream(s1):
            g1.replay()

        torch.cuda.synchronize()

        self.assertEqual(buf0, ref0)
        self.assertEqual(buf1, ref1)

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_memory_stats_of_multiple_generators_and_graphs(self):
        # Function to clear CUDA cache and collect garbage
        def clear_cuda_cache():
            gc.collect()
            torch.cuda.empty_cache()

        # Executes a simple graph task which includes capturing and executing a random number generation within a CUDA graph.
        def simple_graph_task(graph, default_generator, generator_states):
            s = torch.cuda.Stream()
            with torch.cuda.stream(s):
                graph.capture_begin()
                for generator_state in generator_states:
                    default_generator.graphsafe_set_state(generator_state)
                    torch.rand(1, device="cuda")
                graph.capture_end()
            torch.cuda.current_stream().wait_stream(s)
            graph.replay()

        def get_memory_stats():
            stats = torch.cuda.memory_stats()
            num_blocks = stats["active.all.current"]
            total_size = stats["active_bytes.all.current"]
            return num_blocks, total_size

        def test(num_graphs, num_generators):
            baseline = get_memory_stats()
            baseline_num_blocks, baseline_total_size = baseline

            # Allocate CUDA graphs
            graphs = [torch.cuda.CUDAGraph() for _ in range(num_graphs)]

            # Allocate and manage generator states
            default_generator = torch.cuda.default_generators[0]
            generator_states = [default_generator.graphsafe_get_state()]

            for _ in range(1, num_generators):
                generator_states.append(default_generator.clone_state())

            for graph in graphs:
                for generator_state in generator_states:
                    graph.register_generator_state(generator_state)
                simple_graph_task(graph, default_generator, generator_states)

            # Assert conditions after graph tasks
            num_blocks, total_size = get_memory_stats()
            # Each (generator, graph) pair gets 2 tensors (seed + offset)
            expected_captured_states = num_generators * num_graphs
            expected_blocks_diff = 2 * expected_captured_states
            expected_size_diff = 2 * 512 * expected_captured_states

            self.assertEqual(
                (num_blocks - baseline_num_blocks),
                expected_blocks_diff,
                "Unexpected number of active blocks.",
            )
            self.assertEqual(
                (total_size - baseline_total_size),
                expected_size_diff,
                "Unexpected total memory size.",
            )

            # Cleanup graphs and clear CUDA cache
            while graphs:
                graph = graphs.pop()
                del graph
            clear_cuda_cache()

            # Assert that memory stats return to baseline after cleanup
            self.assertEqual(
                get_memory_stats(),
                baseline,
                "Memory stats do not match baseline after cleanup.",
            )

        # Running the test function with different parameters
        test(1, 1)
        test(3, 2)
        test(10, 20)

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_capture_reset_recapture(self):
        s = torch.cuda.Stream()

        with torch.cuda.stream(s):
            a = torch.full((1000,), 1, device="cuda")
            g = torch.cuda.CUDAGraph()
            torch.cuda.empty_cache()
            g.capture_begin()
            b = a
            for _ in range(10):
                b = b + 1
            g.capture_end()
        torch.cuda.current_stream().wait_stream(s)

        g.replay()

        self.assertEqual(b.sum().item(), 11000.0)

        g.reset()

        with torch.cuda.stream(s):
            g.capture_begin()
            b.fill_(2.0)
            for _ in range(10):
                b = b + 2
            g.capture_end()
        torch.cuda.current_stream().wait_stream(s)

        g.replay()
        self.assertEqual(b.sum().item(), 22000.0)

        g.reset()
        del g

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_debugdump(self):
        torch.cuda.empty_cache()
        x = torch.randn(10240000, device="cuda")
        y = torch.rand_like(x)
        g = torch.cuda.CUDAGraph()
        g.enable_debug_mode()
        s0 = torch.cuda.Stream()
        s1 = torch.cuda.Stream()
        s0.wait_stream(torch.cuda.current_stream())
        with torch.cuda.stream(s0):
            g.capture_begin()
            z = x + y
            with torch.cuda.stream(s1):
                s1.wait_stream(s0)
                z + y
            s0.wait_stream(s1)
            g.capture_end()
        s0.synchronize()
        torch.cuda.synchronize()
        with tempfile.TemporaryDirectory() as tempdir:
            g.debug_dump(os.path.join(tempdir, "out_multi_stream.dot"))

    @unittest.skipIf(
        not TEST_CUDA_GRAPH,
        "CUDA >= 11.0 / ROCm >= 7.0 required for external events in cuda graphs",
    )
    @skipIfRocmVersionLessThan((7, 0))
    def test_graph_timing(self):
        torch.cuda.empty_cache()
        x = torch.randn(10240000, device="cuda")
        y = torch.rand_like(x)
        g = torch.cuda.CUDAGraph()
        start_event = torch.cuda.Event(enable_timing=True, external=True)
        end_event = torch.cuda.Event(enable_timing=True, external=True)
        with torch.cuda.graph(g):
            start_event.record()
            z = x + y
            end_event.record()
        torch.cuda.synchronize()
        g.replay()
        torch.cuda.synchronize()
        self.assertTrue(start_event.elapsed_time(end_event) > 0)

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_error(self):
        # We need to run this test in a separate thread as the error we trigger
        # puts the cuda context in a bad state
        script = """
import torch

g = torch.cuda.CUDAGraph()
try:
    g.capture_begin()
except RuntimeError as e:
    if "CUDA graphs must be captured on a non-default stream." in str(e):
        exit(0)
    else:
        exit(1)
exit(2)
"""
        try:
            subprocess.check_output(
                [sys.executable, "-c", script],
                stderr=subprocess.STDOUT,
                # On Windows, opening the subprocess with the default CWD makes `import torch`
                # fail, so just set CWD to this script's directory
                cwd=os.path.dirname(os.path.realpath(__file__)),
            )
        except subprocess.CalledProcessError as e:
            if e.returncode == 1:
                self.assertTrue(
                    False,
                    "Error raise by starting capture without a stream is not the expected one",
                )
            elif e.returncode == 2:
                self.assertTrue(
                    False,
                    "Error raised by starting capture without a stream was not caught",
                )

    @unittest.skipIf(
        (not TEST_CUDA),
        "CUDA >= 11.0 required for graphs",
    )
    def test_graph_warn_if_has_zero_nodes(self):
        with warnings.catch_warnings(record=True) as caught:
            g = torch.cuda.CUDAGraph()
            s = torch.cuda.Stream()
            with torch.cuda.stream(s):
                g.capture_begin()
                g.capture_end()
        self.assertTrue(
            any("The CUDA Graph is empty" in str(w.message) for w in caught)
        )

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    @unittest.skipIf(
        IS_JETSON, "oom reporting has issues on jetson igx due to partial nvml support"
    )
    def test_graph_capture_oom(self):
        oom_regex = (
            "would exceed allowed memory" if TEST_CUDAMALLOCASYNC else "out of memory"
        )
        with self.assertRaisesRegex(RuntimeError, oom_regex):
            with torch.cuda.graph(torch.cuda.CUDAGraph()):
                torch.zeros(2**40, device="cuda")

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    @serialTest()
    @blas_library_context("cublas")
    def test_repeat_graph_capture_cublas_workspace_memory(self):
        (x, y, z) = 1024, 512, 64
        a = torch.rand((x, y), device="cuda")
        b = torch.rand((y, z), device="cuda")

        # warmup
        torch.mm(a, b)

        for i in range(100):
            torch_graph = torch.cuda.CUDAGraph()
            with torch.cuda.graph(torch_graph):
                torch.mm(a, b)
            torch_graph.replay()
            if i == 0:
                free_bytes_before, total_bytes = torch.cuda.mem_get_info()
                used_gb_before = (total_bytes - free_bytes_before) / 1e9

        free_bytes_after, _ = torch.cuda.mem_get_info()
        used_gb_after = (total_bytes - free_bytes_after) / 1e9

        self.assertGreater(0.005 + used_gb_before, used_gb_after)

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_rng_functional(self):
        ops_with_kwargs = (
            (torch.nn.functional.dropout, {"p": 0.1}),
            (torch.nn.functional.rrelu, {"training": True}),
        )
        size = 10000

        def run(op, kwargs):
            a = torch.randn((size,), device="cuda", dtype=torch.float)

            # Control
            torch.cuda.manual_seed(5)
            eager_out = a
            for _ in range(6):
                eager_out = op(eager_out, **kwargs)

            graph_in = a.clone()
            stream = torch.cuda.Stream()
            stream.wait_stream(torch.cuda.current_stream())
            with torch.cuda.stream(stream):
                torch.cuda.manual_seed(5)

                g = torch.cuda.CUDAGraph()
                torch.cuda.empty_cache()
                g.capture_begin()
                graph_out = graph_in
                for _ in range(2):
                    graph_out = op(graph_out, **kwargs)
                g.capture_end()
            torch.cuda.current_stream().wait_stream(stream)

            # Runs a graphed->eager->graphed sequence of RNG ops.
            # replay() plays 2 invocations of the op, so the sequence has 6
            # invocations total, matching Control.
            # replay() reads from graph_in and writes to graph_out.
            g.replay()
            out = op(graph_out, **kwargs)
            out = op(out, **kwargs)
            graph_in.copy_(out)
            g.replay()

            # If replay() updated RNG state correctly, graph_out
            # should now hold data equal to eager_out.
            try:
                self.assertEqual(eager_out, graph_out)
            except Exception as e:
                raise RuntimeError("Failed on ", op) from e

            # Do the same operations varying seeds
            seeds = [6, 128, 9999]

            for seed in seeds:
                torch.cuda.manual_seed(seed)
                graph_in.copy_(a)
                for _ in range(3):
                    g.replay()

                # If the random seed was not updated then the graph would
                # generate the same output as in previous check.
                try:
                    self.assertNotEqual(eager_out, graph_out)
                except Exception as e:
                    raise RuntimeError("Failed on ", op) from e

                # Now repeat the same operations in non-graphed mode.
                torch.cuda.manual_seed(seed)
                for _ in range(3):
                    eager_out.copy_(a)
                    eager_out = op(eager_out, **kwargs)
                    eager_out = op(eager_out, **kwargs)

                # In the end, graph_out and eager_out must be equal
                # as they went under the same set of operations.
                try:
                    self.assertEqual(eager_out, graph_out)
                except Exception as e:
                    raise RuntimeError("Failed on ", op) from e

            # We hold references to all tensors used across streams up til this sync,
            # so no need to call record_stream on those tensors.
            torch.cuda.synchronize()

        for op, kwargs in ops_with_kwargs:
            run(op, kwargs)

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_rng_distributions(self):
        size = 10000
        input = torch.rand((size,), device="cuda", dtype=torch.float)
        alloc = torch.empty((size,), device="cuda", dtype=torch.float)

        # Torch ops to test with sample args (tuple) and kwargs (dict)
        torch_with_args = (
            ("bernoulli", (input.clone(),), {}),
            # multinomial uses some uncapturable CUDA calls.
            # TODO: reenable multinomial tests if/when the implementation is capturable.
            # ("multinomial", (input.clone(), size, True), {}),
            # ("multinomial", (input.clone(), size // 2, False), {}),
            # TODO: reenable normal test, where std is a device
            # tensor, when graph test failures are fixed
            # ("normal", (input.clone() + 1, input.clone()), {}),
            ("normal", (input.clone() + 1, 1.0), {}),
            ("poisson", (input.clone(),), {}),
            ("rand", (size,), {"device": "cuda", "dtype": torch.float}),
            ("randint", (0, 3, (size,)), {"device": "cuda", "dtype": torch.float}),
            ("randn", (size,), {"device": "cuda", "dtype": torch.float}),
        )

        # Tensor methods to test with sample args (tuple)
        tensor_with_args = (
            ("bernoulli_", (input.clone(),)),
            ("cauchy_", ()),
            ("exponential_", ()),
            ("geometric_", (0.3,)),
            ("log_normal_", ()),
            ("normal_", ()),
            ("random_", ()),
            ("uniform_", ()),
        )

        def run(module, op, args, kwargs):
            torch.cuda.manual_seed(5)

            # Each path runs a dummy op to increment the state a bit before creating controls.
            if module == "torch":
                dummy = getattr(torch, op)(*args, **kwargs)
                control1 = getattr(torch, op)(*args, **kwargs)
                control2 = getattr(torch, op)(*args, **kwargs)
            else:
                dummy = alloc.clone()
                control1 = alloc.clone()
                control2 = alloc.clone()
                getattr(dummy, op)(*args)
                getattr(control1, op)(*args)
                getattr(control2, op)(*args)

            stream = torch.cuda.Stream()
            stream.wait_stream(torch.cuda.current_stream())
            with torch.cuda.stream(stream):
                torch.cuda.manual_seed(5)

                g = torch.cuda.CUDAGraph()
                torch.cuda.empty_cache()
                if module == "torch":
                    g.capture_begin()
                    t1 = getattr(torch, op)(*args, **kwargs)
                    t2 = getattr(torch, op)(*args, **kwargs)
                    g.capture_end()
                else:
                    t1 = alloc.clone()
                    t2 = alloc.clone()
                    g.capture_begin()
                    getattr(t1, op)(*args)
                    getattr(t2, op)(*args)
                    g.capture_end()
            torch.cuda.current_stream().wait_stream(stream)

            if not TEST_CUDAMALLOCASYNC:
                # Makes sure values haven't been populated yet
                # (in other words, makes sure capture didn't actually run ops).
                # We can only try this with the native allocator, for which captured
                # addresses are already backed by cudaMalloced memory.
                # If we try it with cudaMallocAsync, CUDA won't event consider
                # the captured addresses allocated until replay(), and if we
                # access them before replay() we get IMAs.
                try:
                    self.assertNotEqual(control1, t1)
                    self.assertNotEqual(control2, t2)
                except Exception as e:
                    raise RuntimeError("Failed on " + module + "." + op) from e

            # Set a new seed to check if graph would use it
            for seed in [6, 314, 271]:
                torch.cuda.manual_seed(seed)
                # Runs a dummy op prelude, as for controls, to make sure replay()
                # picks up the dummy op's state increment.
                if module == "torch":
                    dummy = getattr(torch, op)(*args, **kwargs)
                    control1 = getattr(torch, op)(*args, **kwargs)
                    control2 = getattr(torch, op)(*args, **kwargs)
                else:
                    getattr(dummy, op)(*args)
                    getattr(control1, op)(*args)
                    getattr(control2, op)(*args)

                torch.cuda.manual_seed(seed)
                if module == "torch":
                    dummy = getattr(torch, op)(*args, **kwargs)
                else:
                    getattr(dummy, op)(*args)

                # see above comment on TEST_CUDAMALLOCASYNC
                if not TEST_CUDAMALLOCASYNC:
                    t1.copy_(alloc)
                    t2.copy_(alloc)

                # Runs RNG ops that fill t1 and t2.
                g.replay()

                try:
                    self.assertEqual(control1, t1)
                    self.assertEqual(control2, t2)
                except Exception as e:
                    raise RuntimeError("Failed on " + module + "." + op) from e

            # We hold references to all tensors used across streams up til this sync,
            # so no need to call record_stream on those tensors.
            torch.cuda.synchronize()

        for op_with_args in torch_with_args:
            run("torch", *op_with_args)

        for meth_with_args in tensor_with_args:
            # Adds an empty dict for kwargs, which none of the Tensor methods use
            run("Tensor", *(meth_with_args + ({},)))

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_two_successive(self):
        torch.cuda.empty_cache()

        size = 1000
        kSmallBuffer = 2097152

        def func_with_temps(t, val):
            x = t.clone() + val
            y = t.clone() + val
            return x + y

        s = torch.cuda.Stream()

        for share_mem in ("Don't share", "via pool()", "via graph_pool_handle()"):
            g0 = torch.cuda.CUDAGraph()
            g1 = torch.cuda.CUDAGraph()

            a = torch.ones((size,), device="cuda")

            s.wait_stream(torch.cuda.current_stream())
            with torch.cuda.stream(s):
                g0_args = (
                    (torch.cuda.graph_pool_handle(),)
                    if share_mem == "via graph_pool_handle()"
                    else ()
                )
                g0.capture_begin(*g0_args)
                b = a.clone()
                for _ in range(5):
                    b = func_with_temps(b, 1)
                g0.capture_end()

                g1_args = (g0.pool(),) if share_mem == "via pool()" else g0_args
                g1.capture_begin(*g1_args)
                for _ in range(5):
                    b = func_with_temps(b, 1)
                g1.capture_end()
            torch.cuda.current_stream().wait_stream(s)

            # mixes unrelated eager ops with replays
            c = a.clone()
            for _ in range(2):
                c = func_with_temps(c, 3)
            g0.replay()
            for _ in range(2):
                c = func_with_temps(c, 3)
            g1.replay()
            for _ in range(2):
                c = func_with_temps(c, 3)

            self.assertEqual(b.sum().item(), size * 3070)
            self.assertEqual(c.sum().item(), size * 442)

            if not TEST_CUDAMALLOCASYNC:
                # These stat checks are specific to the native allocator.
                if share_mem != "Don't share":
                    self.assertEqual(
                        reserved_no_sharing  # noqa: F821
                        - torch.cuda.memory_stats()["reserved_bytes.all.current"],
                        kSmallBuffer,
                    )
                else:
                    reserved_no_sharing = torch.cuda.memory_stats()[
                        "reserved_bytes.all.current"
                    ]

            del a, b, c, g0, g1
            # Tensors used across streams (a and b) were held until just now, so no need to call record_stream on them.
            torch.cuda.synchronize()
            torch.cuda.empty_cache()

    @unittest.skipIf(
        (not TEST_CUDA_GRAPH)
        or IS_WINDOWS
        or (  # appears to still be broken on Windows as of 11.4+
            torch.version.cuda
            and int(torch.version.cuda.split(".")[0]) == 11
            and int(torch.version.cuda.split(".")[1]) < 4
        ),
        "Graph bindings disallow concurrent replay for CUDA < 11.4, see "
        + "https://github.com/pytorch/pytorch/pull/57556",
    )
    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_concurrent_replay(self):
        torch.cuda.empty_cache()

        size = 1000000  # largeish to help expose race conditions

        def func_with_temps(t, val):
            x = t.clone() + val
            y = t.clone() + val
            return x + y

        s = torch.cuda.Stream()

        for share_mem in ("Don't share", "via pool()", "via graph_pool_handle()"):
            g0 = torch.cuda.CUDAGraph()
            g1 = torch.cuda.CUDAGraph()

            s0 = torch.cuda.Stream()
            s1 = torch.cuda.Stream()

            a = torch.ones((size,), device="cuda")

            s.wait_stream(torch.cuda.current_stream())
            with torch.cuda.stream(s):
                g0_args = (
                    (torch.cuda.graph_pool_handle(),)
                    if share_mem == "via graph_pool_handle()"
                    else ()
                )
                g0.capture_begin(*g0_args)
                b = a.clone()
                for _ in range(5):
                    b = func_with_temps(b, 1)
                g0.capture_end()

                g1_args = (g0.pool(),) if share_mem == "via pool()" else g0_args
                g1.capture_begin(*g1_args)
                c = a.clone()
                for _ in range(5):
                    c = func_with_temps(c, 2)
                g1.capture_end()

            # To reproduce data corruption, I need g0 and g1's kernels to run concurrently.
            # But replay() (especially cudaGraphLaunch) can incur significant CPU overhead.
            # The following pattern helps align device-side execution of g0 and g1's kernels.
            torch.cuda.synchronize()
            with torch.cuda.stream(s0):
                torch.cuda._sleep(1000000)
                s1.wait_stream(s0)
                g0.replay()
            with torch.cuda.stream(s1):
                g1.replay()
            torch.cuda.current_stream().wait_stream(s0)
            torch.cuda.current_stream().wait_stream(s1)

            if (not TEST_CUDAMALLOCASYNC) and (share_mem != "Don't share"):
                # If we used the native allocator and shared mempools,
                # we expect the concurrent replays corrupted each other.
                self.assertNotEqual(b.sum().item(), size * 94)
                self.assertNotEqual(c.sum().item(), size * 156)
            else:
                # If we EITHER
                #   - used the native allocator without sharing mempools, OR
                #   - used cudaMallocAsync, which ignores graph pool-sharing hints and should always be safe
                # we don't expect memory corruption.
                self.assertEqual(b.sum().item(), size * 94)
                self.assertEqual(c.sum().item(), size * 156)

            del a, b, c, g0, g1
            # Tensors used across streams (a, b, c) were held until just now, so no need to call record_stream on them.
            torch.cuda.synchronize()
            torch.cuda.empty_cache()

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_three_successive(self):
        torch.cuda.empty_cache()

        size = 1000

        s = torch.cuda.Stream()

        for share_mem in ("Don't share", "via pool()", "via graph_pool_handle()"):
            a = torch.ones((size,), device="cuda")

            g0 = torch.cuda.CUDAGraph()
            g1 = torch.cuda.CUDAGraph()
            g2 = torch.cuda.CUDAGraph()

            s.wait_stream(torch.cuda.current_stream())
            with torch.cuda.stream(s):
                g0_args = (
                    (torch.cuda.graph_pool_handle(),)
                    if share_mem == "via graph_pool_handle()"
                    else ()
                )
                g0.capture_begin(*g0_args)
                b = a.clone()
                c = b + 1
                d = b + 2
                g0.capture_end()

                args = (g0.pool(),) if share_mem == "via pool()" else g0_args

                g1.capture_begin(*args)
                e = c + 3
                del c
                g1.capture_end()

                g2.capture_begin(*args)
                f = d + 4
                g2.capture_end()
            torch.cuda.current_stream().wait_stream(s)

            # Tests that replaying in capture order is valid
            g0.replay()
            g1.replay()
            g2.replay()

            self.assertEqual(e.sum().item(), size * 5)
            self.assertEqual(f.sum().item(), size * 7)

            # Tests that replaying as g0, g2, g1 is only valid if they don't share a pool
            g0.replay()
            g2.replay()
            g1.replay()

            expect_corruption = (not TEST_CUDAMALLOCASYNC) and (
                share_mem != "Don't share"
            )
            # If we used the native allocator and shared mempools, g2's capture should have reused c's memory for f.
            # We replayed g2 then g1, so we expect g1's captured "e = c + 3" mistakenly filled e with "f's vals + 3".
            self.assertEqual(
                e.sum().item(), size * (7 + 3) if expect_corruption else size * 5
            )
            self.assertEqual(f.sum().item(), size * 7)

            del a, b, d, e, f, g0, g1, g2
            # Tensors used across streams (a, e, f) were held until just now, so no need to call record_stream on them.
            torch.cuda.synchronize()
            torch.cuda.empty_cache()

    @unittest.skipIf(
        (not TEST_CUDA_GRAPH) or TEST_CUDAMALLOCASYNC,
        "CUDA >= 11.0 or ROCM >= 5.3 required for graphs",
    )
    def test_graph_memory_stats_and_use_result_after_destroy_graph(self):
        kSmallSize = 1048576
        kSmallBuffer = 2097152
        kLargeBuffer = 20971520
        kMinLargeAlloc = 10485760
        kRoundLarge = 2097152

        elem = 4

        # this was annoying to write but stresses the expectations pretty rigorously
        # For small_pool cases, delta_cudaMallocs and delta_cudaMalloc_bytes include
        # an extra kSmallBuffer segment for the per-capture RNG state tensors, which
        # are allocated on the default stream (separate from stream s).
        cases = (
            (512 // elem, 2, 2 * kSmallBuffer, kSmallBuffer, "small_pool"),
            (kSmallSize // elem, 3, 3 * kSmallBuffer, kSmallBuffer, "small_pool"),
            ((kSmallSize + 512) // elem, 1, kLargeBuffer, kLargeBuffer, "large_pool"),
            (
                (kMinLargeAlloc - 512) // elem,
                2,
                2 * kLargeBuffer,
                kLargeBuffer,
                "large_pool",
            ),
            (
                (kMinLargeAlloc + 512) // elem,
                3,
                3
                * (
                    kRoundLarge
                    * ((kMinLargeAlloc + 512 + kRoundLarge - 1) // kRoundLarge)
                ),
                kRoundLarge * ((kMinLargeAlloc + 512 + kRoundLarge - 1) // kRoundLarge),
                "large_pool",
            ),
        )

        stats_to_check = ("segment.", "reserved_bytes.", "active.", "active_bytes.")

        gc.collect()
        torch.cuda.empty_cache()

        s = torch.cuda.Stream()

        for (
            numel,
            delta_cudaMallocs,
            delta_cudaMalloc_bytes,
            delta_cudaMalloc_bytes_post_del_g,
            pool_string,
        ) in cases:
            if pool_string == "small_pool":
                delta_active_blocks = 3  # one from "b" plus a sneaky two from CUDAGraph's one-element rng seed and offset holders
                delta_active_bytes = (
                    numel * elem + 1024
                )  # + 1024 for CUDAGraph's rng seed and offset holders each
            else:
                delta_active_blocks = 1  # We only check the large pool, which isn't affected by rng offset holder
                delta_active_bytes = numel * elem

            g = torch.cuda.CUDAGraph()
            s.wait_stream(torch.cuda.current_stream())
            with torch.cuda.stream(s):
                # Per-capture RNG state tensors are allocated on the default stream
                # (not the capture stream), so they occupy a separate segment from
                # user tensors created here on stream s.
                a = torch.ones((numel,), device="cuda")

                precapture_stats = torch.cuda.memory_stats()

                g.capture_begin()
                b = a.clone()
                for _ in range(5):
                    b = b.clone() + 1
                g.capture_end()
            torch.cuda.current_stream().wait_stream(s)

            gc.collect()

            postcapture_stats = torch.cuda.memory_stats()

            expecteds = (
                delta_cudaMallocs,
                delta_cudaMalloc_bytes,
                delta_active_blocks,
                delta_active_bytes,
            )
            # Double checks replay and stats before and after a call to empty_cache
            for i in range(2):
                for stat, expected in zip(stats_to_check, expecteds):
                    stat = stat + pool_string + ".current"
                    current = postcapture_stats[stat] - precapture_stats[stat]

                    # There will only ever be one expandable segment in each of the small and large pools. The way the
                    # bookkeeping is done in the allocator means that we never increment the number of segments.
                    if self.expandable_segments and "segment" in stat:
                        expected = 0
                    # These two cases hit an edge case where the PyTorch allocator won't immediately unmap part of an
                    # expandable segment (and as a result reduce the number of reserved bytes) if the block to unmap is
                    # smaller than the page size
                    if (
                        self.expandable_segments
                        and "reserved" in stat
                        and (numel == cases[3][0] or numel == cases[4][0])
                    ):
                        expected = 2 * kLargeBuffer

                    self.assertEqual(
                        current,
                        expected,
                        "Pre to post capture delta of "
                        + stat
                        + f" = {current}, expected = {expected}, numel = {numel}",
                    )

                g.replay()
                self.assertEqual(b.sum().item(), 6 * numel)
                if i == 0:
                    torch.cuda.empty_cache()

            del g
            gc.collect()
            torch.cuda.empty_cache()
            postdel_stats = torch.cuda.memory_stats()

            # Uses graph result b after graph has been deleted
            self.assertEqual(b.sum().item(), 6 * numel)

            # b should be the only live reference remaining from the graph's private pool
            expecteds = (1, delta_cudaMalloc_bytes_post_del_g, 1, numel * elem)
            for stat, expected in zip(stats_to_check, expecteds):
                stat = stat + pool_string + ".current"
                current = postdel_stats[stat] - precapture_stats[stat]

                # There will only ever be one expandable segment in each of the small and large pools. The way the
                # bookkeeping is done in the allocator means that we never increment the number of segments.
                if self.expandable_segments and "segment" in stat:
                    expected = 0
                # These two cases hit an edge case where the PyTorch allocator won't immediately unmap part of an
                # expandable segment (and as a result reduce the number of reserved bytes) if the block to unmap is
                # smaller than the page size
                if (
                    self.expandable_segments
                    and "reserved" in stat
                    and numel == cases[3][0]
                ):
                    expected = 2 * kLargeBuffer
                if (
                    self.expandable_segments
                    and "reserved" in stat
                    and numel == cases[4][0]
                ):
                    expected = kLargeBuffer

                self.assertEqual(
                    current,
                    expected,
                    "Pre capture to post graph delete delta of "
                    + stat
                    + f" = {current}, expected = {expected}, numel = {numel}",
                )

            # del a, b before the next case is essential, otherwise overwriting a and b in the next case
            # can throw off its allocation/deallocation counts.
            del a, b
            # Tensors used across streams (a and b) were held until just now, so no need to call record_stream on them.
            torch.cuda.synchronize()
            torch.cuda.empty_cache()

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_record_stream(self):
        # Makes sure graph capture defers attempting to reclaim allocations used across streams. See
        # "Q. Why skip process_events if a capture might be underway?" in c10/cuda/CUDACachingAllocator.cpp
        torch.cuda.empty_cache()

        potential_problem = torch.zeros((3,), device="cuda")
        a = torch.zeros((3,), device="cuda")
        s0 = torch.cuda.Stream()
        s1 = torch.cuda.Stream()
        s2 = torch.cuda.Stream()
        g = torch.cuda.CUDAGraph()

        torch.cuda.synchronize()
        with torch.cuda.stream(s0):
            potential_problem.record_stream(s0)
            torch.cuda._sleep(TestCuda.FIFTY_MIL_CYCLES)
            potential_problem.fill_(1.0)
        del potential_problem

        with torch.cuda.stream(s1):
            g.capture_begin()
            # potential_problem's allocation should still be outstanding. if DeviceCachingAllocator::malloc
            # mistakenly calls process_events, it will trigger cudaEventQueries on potential_problem's end-of-life
            # event, which will cause the capture to error.
            b = a.clone()

            # Let's also see what happens if we record_stream on a tensor during capture.
            s2.wait_stream(s1)
            with torch.cuda.stream(s2):
                b.fill_(1.0)
                b.record_stream(s2)  # dummy record_stream
                del b
            s1.wait_stream(s2)
            g.capture_end()
        torch.cuda.synchronize()

        # dummy allocation triggers process_events, Hopefully successfully processes b's end-of-life event.
        torch.zeros((3,), device="cuda")

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    # If this test is the first in the process to try cudnn rnns with dropout, it'll initialize
    # DropoutState's long-lived internal buffer. Calling code perceives this (correct) behavior
    # as a memory leak unless we skip the leak check.
    @skipCUDAMemoryLeakCheckIf(True)
    @serialTest()
    def test_graph_cudnn_dropout(self):
        # Tests the interaction of cuda graph capture with DropoutState's syncs in ATen/native/cudnn/RNN.cpp.
        # In particular, if user runs a sequence of captured and noncaptured cudnn rnns, DropoutState should
        # avoid syncing noncapturing streams with captured events or vice versa.
        torch.cuda.empty_cache()

        model = torch.nn.LSTM(512, 512, 2, dropout=0.5).cuda()
        x = torch.ones(100, 192, 512, device="cuda")

        model(x)

        g = torch.cuda.CUDAGraph()
        s = torch.cuda.Stream()
        s.wait_stream(torch.cuda.current_stream())
        with torch.cuda.stream(s):
            g.capture_begin()
            model(x)
            g.capture_end()
        torch.cuda.current_stream().wait_stream(s)

        g.replay()

        model(x)

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    @serialTest()
    def test_graph_checkpoint_preserve_rng_state(self):
        torch.cuda.manual_seed(42)

        def fn(x):
            return x * torch.sigmoid(torch.randn(1, device="cuda"))

        fn(torch.ones(1, device="cuda"))

        torch.cuda.manual_seed(42)
        eager_in = torch.ones(1, device="cuda", requires_grad=True)
        eager_out = torch.utils.checkpoint.checkpoint(
            fn, eager_in, use_reentrant=False, preserve_rng_state=True
        )
        (eager_in_grad,) = torch.autograd.grad(eager_out, eager_in)

        g = torch.cuda.CUDAGraph()
        with torch.cuda.graph(g):
            graph_in = torch.ones(1, device="cuda", requires_grad=True)
            graph_out = torch.utils.checkpoint.checkpoint(
                fn, graph_in, use_reentrant=False, preserve_rng_state=True
            )
            (graph_in_grad,) = torch.autograd.grad(graph_out, graph_in)

        torch.cuda.manual_seed(42)
        g.replay()

        self.assertEqual(eager_in_grad, graph_in_grad, rtol=0.0, atol=0.0)

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    @serialTest()
    def test_graph_manual_seed_mismatch_raises(self):
        torch.cuda.manual_seed(0)
        g = torch.cuda.CUDAGraph()
        with self.assertRaisesRegex(
            RuntimeError,
            "CUDAGeneratorImpl::set_current_seed can be called during stream capture only if new seed is the same as the original seed.",
        ):
            with torch.cuda.graph(g):
                torch.cuda.manual_seed(1)

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    @parametrize(
        "with_amp,cache_enabled,allow_unused_input",
        [
            subtest((True, True, True), decorators=[unittest.expectedFailure]),
            subtest((False, False, False), decorators=[unittest.expectedFailure]),
        ],
        name_fn=lambda x, y, z: "{}{}{}".format(
            {True: "with_amp", False: "without_amp"}[x],
            {True: "_cache_enabled", False: "_cache_disabled"}[y] if x else "",
            {True: "_allow_unused_input", False: "_not_allow_unused_input"}[z],
        ),
    )
    @serialTest()
    def test_graph_make_graphed_callables(
        self, with_amp, cache_enabled, allow_unused_input
    ):
        torch.manual_seed(5)
        torch.cuda.manual_seed(5)

        N, D_in, H, D_out = 640, 4096, 2048, 1024

        class MLP1(torch.nn.Module):
            def __init__(self, D_in: int, H: int, D_out: int):
                super().__init__()
                self.net_1 = torch.nn.Sequential(
                    torch.nn.Linear(D_in, H), torch.nn.Dropout(p=0.1)
                ).cuda()
                self.net_2 = torch.nn.Sequential(
                    torch.nn.Linear(H, D_out), torch.nn.Dropout(p=0.2)
                ).cuda()

            def forward(self, input_dict: dict):
                x = input_dict["x"]
                return self.net_2(self.net_1(x))

        class MLP2(torch.nn.Module):
            def __init__(self, D_in: int, H: int, D_out: int):
                super().__init__()
                self.net_1 = torch.nn.Sequential(
                    torch.nn.Linear(D_in, H), torch.nn.Dropout(p=0.1)
                ).cuda()
                self.net_2 = torch.nn.Sequential(
                    torch.nn.Linear(H, D_out), torch.nn.Dropout(p=0.2)
                ).cuda()

            def forward(self, x):
                return self.net_2(self.net_1(x))

        class ParameterlessModule(torch.nn.Module):
            def forward(self, x):
                idx = (
                    torch.arange(x.size(0), device=x.device)
                    .view(-1, 1)
                    .repeat(1, x.size(1))
                )
                return {"output": torch.gather(x, 0, idx)}

        models = []
        for _ in range(2):
            model_section1 = MLP1(D_in, H, H).cuda()
            model_section2 = MLP2(H, H, D_out).cuda()
            model_section3 = ParameterlessModule().cuda()
            models.append(
                torch.nn.Sequential(model_section1, model_section2, model_section3)
            )

        model_graphed = models[0]
        model_control = models[1]

        model_graphed.load_state_dict(model_control.state_dict())

        opt_graphed = torch.optim.SGD(model_graphed.parameters(), lr=0.1)
        opt_control = torch.optim.SGD(model_control.parameters(), lr=0.1)

        x = torch.randn(N, D_in, device="cuda")
        h = torch.randn(N, H, device="cuda", requires_grad=True)
        h2 = torch.randn(N, D_out, device="cuda", requires_grad=True)
        unused_input = torch.randn(N, H, device="cuda", requires_grad=True)
        y_pred = torch.randn(N, D_out, device="cuda", requires_grad=True)
        y = torch.randn(N, D_out, device="cuda")

        loss_fn_control = torch.nn.functional.mse_loss
        relu_control = torch.nn.functional.relu

        # This is a good stress test. It graphs four callables: two Modules and two python functions.
        with torch.amp.autocast(
            device_type="cuda", enabled=with_amp, cache_enabled=cache_enabled
        ):
            (
                model_graphed[0],
                model_graphed[1],
                model_graphed[2],
                relu_graphed,
                loss_fn_graphed,
            ) = torch.cuda.make_graphed_callables(
                (
                    model_graphed[0],
                    model_graphed[1],
                    model_graphed[2],
                    relu_control,
                    loss_fn_control,
                ),
                (
                    ({"x": x, "unused_input": unused_input},),
                    (h,),
                    (h2,),
                    (y_pred,),
                    (y_pred, y),
                ),
                allow_unused_input=allow_unused_input,
            )

        real_inputs = [torch.rand_like(x) for _ in range(10)]
        real_targets = [torch.rand_like(y) for _ in range(10)]

        for m, opt, relu, loss_fn in zip(
            (model_graphed, model_control),
            (opt_graphed, opt_control),
            (relu_graphed, relu_control),
            (loss_fn_graphed, loss_fn_control),
        ):
            # Resets RNC states before iterations for graphed and ungraphed models,
            # so dropout math should be bitwise identical for both.
            torch.manual_seed(5)
            torch.cuda.manual_seed(5)
            for data, target in zip(real_inputs, real_targets):
                opt.zero_grad(set_to_none=True)
                with torch.amp.autocast(
                    device_type="cuda", enabled=with_amp, cache_enabled=cache_enabled
                ):
                    y_pred = m({"x": data, "unused_input": unused_input})["output"]
                    y_pred = relu(y_pred)
                    loss = loss_fn(y_pred, target)
                    loss.backward()
                opt.step()

        for p, pc in zip(model_graphed.parameters(), model_control.parameters()):
            self.assertEqual(p, pc)

        # We graphed the models in training mode. Eval should still run ungraphed.
        model_graphed.eval()
        model_control.eval()
        self.assertEqual(
            model_graphed({"x": real_inputs[0]}), model_control({"x": real_inputs[0]})
        )

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    @parametrize(
        "with_amp,cache_enabled,allow_unused_input",
        [
            subtest((False, False, True)),
            subtest((True, False, True)),
            subtest((True, True, True), decorators=[unittest.expectedFailure]),
            subtest((False, False, False)),
        ],
        name_fn=lambda x, y, z: "{}{}{}".format(
            {True: "with_amp", False: "without_amp"}[x],
            {True: "_cache_enabled", False: "_cache_disabled"}[y] if x else "",
            {True: "_allow_unused_input", False: "_not_allow_unused_input"}[z],
        ),
    )
    @serialTest()
    def test_graph_make_graphed_callables_parameterless_nograd_module(
        self, with_amp, cache_enabled, allow_unused_input
    ):
        torch.manual_seed(5)
        torch.cuda.manual_seed(5)

        N, D_in, H, _ = 640, 4096, 2048, 1024

        class ParameterlessModule(torch.nn.Module):
            def forward(self, input_dict: dict):
                x = input_dict["x"]
                idx = (
                    torch.arange(x.size(0), device=x.device)
                    .view(-1, 1)
                    .repeat(1, x.size(1))
                )
                return {"output": torch.gather(x, 0, idx)}

        models = []
        for _ in range(2):
            model_section1 = ParameterlessModule().cuda()
            models.append(torch.nn.Sequential(model_section1))

        model_graphed = models[0]
        model_control = models[1]

        model_graphed.load_state_dict(model_control.state_dict())

        x = torch.randn(N, D_in, device="cuda", requires_grad=False)
        unused_input = torch.randn(N, H, device="cuda", requires_grad=False)
        y = torch.randn(N, D_in, device="cuda")

        # This is a good stress test. It graphs four callables: two Modules and two python functions.
        with torch.amp.autocast(
            device_type="cuda", enabled=with_amp, cache_enabled=cache_enabled
        ):
            model_graphed[0] = torch.cuda.make_graphed_callables(
                model_graphed[0],
                ({"x": x, "unused_input": unused_input},),
                allow_unused_input=allow_unused_input,
            )

        real_inputs = [torch.rand_like(x, requires_grad=True) for _ in range(10)]
        real_targets = [torch.rand_like(y) for _ in range(10)]

        for m in (model_graphed, model_control):
            # Resets RNC states before iterations for graphed and ungraphed models,
            # so dropout math should be bitwise identical for both.
            torch.manual_seed(5)
            torch.cuda.manual_seed(5)
            for data, _ in zip(real_inputs, real_targets):
                with torch.amp.autocast(
                    device_type="cuda", enabled=with_amp, cache_enabled=cache_enabled
                ):
                    m({"x": data, "unused_input": unused_input})["output"]

        # We graphed the models in training mode. Eval should still run ungraphed.
        model_graphed.eval()
        model_control.eval()
        self.assertEqual(
            model_graphed({"x": real_inputs[0]}), model_control({"x": real_inputs[0]})
        )

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_make_graphed_callables_same_pool(self):
        torch.manual_seed(5)
        torch.cuda.manual_seed(5)
        models = []
        num_models = 3
        for _ in range(num_models):
            models.append(
                torch.nn.Sequential(
                    torch.nn.Linear(32, 128),
                    torch.nn.ReLU(),
                    torch.nn.Linear(128, 128),
                ).cuda()
            )
        # we will reuse the same pool for all graph captures
        mempool = torch.cuda.graph_pool_handle()
        graphed_models = []
        for model in models:
            x = torch.randn([64, 32], device="cuda")
            graphed_model = deepcopy(model)
            graphed_model = torch.cuda.make_graphed_callables(
                graphed_model, (x,), pool=mempool
            )
            graphed_models.append(graphed_model)

        for model, graphed_model in zip(models, graphed_models):
            x = torch.randn([64, 32], device="cuda")
            y = model(x)
            yg = graphed_model(x)
            l = y.norm()
            lg = yg.norm()
            l.backward()
            lg.backward()

            self.assertEqual(y, yg)
            self.assertEqual(l, lg)
            for p, pg in zip(model.parameters(), graphed_model.parameters()):
                self.assertEqual(p, pg)
                self.assertEqual(p.grad, pg.grad)
                self.assertNotEqual(p.data_ptr(), pg.data_ptr())
                self.assertNotEqual(p.grad.data_ptr(), pg.grad.data_ptr())

    def test_cuda_graph_inference_mode(self):
        # This test is to verify that capturing a CUDAGraph in inference mode
        # doesn't create RNG State tensors as inference tensors which can't
        # be inplace modified later.

        def fn(x):
            return x + 1

        x = torch.randn(10, 10, device="cuda")

        with torch.inference_mode():
            captured_fn = torch.cuda.make_graphed_callables(fn, (x,))
            inp = torch.ones(10, 10, device="cuda")
            self.assertEqual(captured_fn(inp), inp + 1)

        torch.cuda.make_graphed_callables(fn, (x,))
        inp = torch.ones(10, 10, device="cuda") * 10
        self.assertEqual(fn(inp), inp + 1)

    def _test_graphed_optimizer(
        self, steps_warmup, steps_train, optimizer_ctor, kwargs
    ):
        for actually_do_graphs in (True, False):
            params = [torch.randn((i + 5, i + 5), device="cuda") for i in range(2)] + [
                torch.randn((), device="cuda")
            ]
            params_control = [p.clone().requires_grad_() for p in params]
            params_graphed = [p.clone().requires_grad_() for p in params]

            grads = [
                [torch.randn_like(p) for p in params]
                for _ in range(steps_warmup + steps_train)
            ]

            # Control (capturable=False)

            opt = optimizer_ctor(params_control, capturable=False, **kwargs)

            for i in range(steps_warmup + steps_train):
                for j, p in enumerate(params_control):
                    p.grad = grads[i][j]
                opt.step()

            # capturable=True

            opt = optimizer_ctor(params_graphed, capturable=True, **kwargs)

            for i in range(steps_warmup):
                for j, p in enumerate(params_graphed):
                    p.grad = grads[i][j]
                opt.step()

            if actually_do_graphs:
                g = torch.cuda.CUDAGraph()
                with torch.cuda.graph(g):
                    opt.step()

            for i in range(steps_train):
                if actually_do_graphs:
                    for j, p in enumerate(params_graphed):
                        p.grad.copy_(grads[i + steps_warmup][j])
                    g.replay()
                else:
                    # Passing capturable=True to the constructor and running without graphs should still be
                    # numerically correct, even if it's not ideal for performance.
                    for j, p in enumerate(params_graphed):
                        p.grad = grads[i + steps_warmup][j]
                    opt.step()

            for p_control, p_graphed in zip(params_control, params_graphed):
                self.assertEqual(p_control, p_graphed)

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_optims_with_explicitly_capturable_param_groups(self):
        # mimicking `_test_graphed_optimizer` maladroitly to pass two param_groups to optimizer.__init__
        n_warmup, n_replay = 3, 2
        for optimizer, second_param_group_capturable in product(
            (
                torch.optim.Adam,
                torch.optim.AdamW,
                torch.optim.ASGD,
                torch.optim.Adamax,
                torch.optim.NAdam,
                torch.optim.RAdam,
                torch.optim.Adadelta,
                torch.optim.RMSprop,
                torch.optim.Rprop,
            ),
            (True, False),
        ):
            ref_p1, param1 = (
                torch.nn.Parameter(torch.ones(1, device="cuda")) for _ in range(2)
            )
            ref_p2, param2 = (
                torch.nn.Parameter(torch.ones(1, device="cuda")) for _ in range(2)
            )
            grads1, grads2 = (
                [torch.randn_like(param1) for _ in range(n_warmup + n_replay)]
                for _ in range(2)
            )
            ref_grads1, ref_grads2 = (
                [t.clone() for t in tensors] for tensors in (grads1, grads2)
            )
            params = [
                {"params": [param1], "capturable": True},
                {"params": [param2], "capturable": second_param_group_capturable},
            ]
            opt = optimizer(params)
            opt_ = optimizer(
                [
                    {"params": [ref_p1], "capturable": False},
                    {"params": [ref_p2], "capturable": False},
                ]
            )

            for i in range(n_warmup + n_replay):
                ref_p1.grad = ref_grads1[i]
                ref_p2.grad = ref_grads2[i]
                opt_.step()

            for i in range(n_warmup):
                param1.grad = grads1[i]
                param2.grad = grads2[i]
                opt.step()

            g = torch.cuda.CUDAGraph()
            if not second_param_group_capturable:
                with self.assertRaisesRegex(RuntimeError, "Attempting CUDA graph"):
                    with torch.cuda.graph(g):
                        opt.step()
            else:
                with torch.cuda.graph(g):
                    opt.step()

                for i in range(n_replay):
                    param1.grad.copy_(grads1[n_warmup + i])
                    param2.grad.copy_(grads2[n_warmup + i])
                    g.replay()
                self.assertEqual(ref_p1, param1)
                self.assertEqual(ref_p2, param2)

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_cuda_graph_error_options(self):
        def fn():
            x = torch.zeros([2000], device="cuda")
            y = x + x + x
            return y

        mem = None

        def raw_malloc():
            global mem
            mem = None
            stream = torch.cuda.Stream()
            try:
                with torch.cuda.stream(stream):
                    mem = torch.cuda.caching_allocator_alloc(1024)
            except BaseException:
                if mem is None:
                    return
            try:
                torch.cuda.caching_allocator_delete(mem)
                mem = None
                return None
            except BaseException:
                pass

        def throws_on_cuda_event(capture_error_mode):
            graph = torch.cuda.CUDAGraph()
            torch.cuda.synchronize()
            stream = torch.cuda.Stream()
            stream.wait_stream(torch.cuda.current_stream())
            with torch.cuda.stream(stream):
                fn()
            stream.synchronize()
            torch.cuda.current_stream().wait_stream(stream)
            torch.cuda.synchronize()
            try:
                with torch.cuda.graph(
                    graph, stream=stream, capture_error_mode=capture_error_mode
                ):
                    out = fn()
                    thread = threading.Thread(target=raw_malloc)
                    thread.start()
                    thread.join()
            except Exception:
                if mem is not None:
                    torch.cuda.caching_allocator_delete(mem)
                return True

            return False

        self.assertFalse(throws_on_cuda_event("thread_local"))
        self.assertFalse(throws_on_cuda_event("relaxed"))

        # Exception would Corrupt Process and make other tests fail
        # self.assertTrue(throws_on_cuda_event("global"))

    @unittest.skipIf(
        not TEST_CUDA_GRAPH,
        "CUDA >= 11.0 or ROCM >= 5.3 required for graphs, cuda-python must be installed",
    )
    def test_cuda_graph_raw_graph_keep_graph_false(self):
        graph = torch.cuda.CUDAGraph(keep_graph=False)
        x = torch.zeros([2000], device="cuda")
        y = torch.ones([2000], device="cuda")
        with torch.cuda.graph(graph, capture_error_mode="relaxed"):
            z = x + y

        with self.assertRaisesRegex(
            RuntimeError,
            r"instantiate\(\) is intended to be called by the user only when keep_graph=true",
        ):
            raw_pointer = graph.instantiate()

        with self.assertRaisesRegex(
            RuntimeError,
            r"You cannot access the raw (cuda|hip)Graph_t instance unless CUDAGraph was initialized with keep_graph=true",
        ):
            raw_pointer = graph.raw_cuda_graph()

    @unittest.skipIf(
        not TEST_CUDA_GRAPH or not TEST_CUDA_PYTHON_BINDINGS,
        "CUDA >= 11.0 or ROCM >= 5.3 required for graphs, cuda-bindings must be installed",
    )
    def test_cuda_graph_raw_graph(self):
        import cuda.bindings.runtime as cudart

        graph = torch.cuda.CUDAGraph(keep_graph=True)
        x = torch.zeros([2000], device="cuda")
        y = torch.ones([2000], device="cuda")
        with torch.cuda.graph(graph, capture_error_mode="relaxed"):
            z = x + y

        raw_pointer = graph.raw_cuda_graph()

        cudart_cuda_graph = cudart.cudaGraph_t(init_value=raw_pointer)
        _, num_nodes = cuda_python_error_check(
            cudart.cudaGraphGetNodes(cudart_cuda_graph)
        )
        nodes, _ = cuda_python_error_check(
            cudart.cudaGraphGetNodes(cudart_cuda_graph, num_nodes)
        )
        for node in nodes:
            cuda_python_error_check(cudart.cudaGraphNodeGetType(node))

        graph.replay()

    @unittest.skipIf(
        not TEST_CUDA_GRAPH or not TEST_CUDA_PYTHON_BINDINGS,
        "CUDA >= 11.0 or ROCM >= 5.3 required for graphs, cuda-bindings must be installed",
    )
    @parametrize("keep_graph", [True, False])
    def test_cuda_graph_raw_graph_exec(self, keep_graph):
        import cuda.bindings.runtime as cudart

        graph = torch.cuda.CUDAGraph(keep_graph=keep_graph)
        x = torch.zeros([2000], device="cuda")
        y = torch.ones([2000], device="cuda")
        with torch.cuda.graph(graph, capture_error_mode="relaxed"):
            z = x + y

        if keep_graph:
            with self.assertRaisesRegex(
                RuntimeError,
                r"You cannot access the raw (cuda|hip)GraphExec_t instance until instantiate\(\) has been called",
            ):
                graph.raw_cuda_graph_exec()

            graph.instantiate()
        raw_pointer = graph.raw_cuda_graph_exec()

        cudart_cuda_graph_exec = cudart.cudaGraphExec_t(init_value=raw_pointer)
        cuda_python_error_check(cudart.cudaGraphExecGetFlags(cudart_cuda_graph_exec))

        graph.replay()

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_cuda_graph_raw_graph_reset_and_recapture(self):
        graph = torch.cuda.CUDAGraph(keep_graph=True)
        x = torch.zeros([2000], device="cuda")
        with torch.cuda.graph(graph, capture_error_mode="relaxed"):
            x += 1.0

        graph.instantiate()
        graph.replay()
        self.assertTrue(torch.all(x == 1.0))
        # Exercise the code path where you reinstantiate the cuda graph twice.
        graph.instantiate()
        graph.replay()
        self.assertTrue(torch.all(x == 2.0))
        graph.replay()
        self.assertTrue(torch.all(x == 3.0))

        # Check that graph capture can succeed after resetting.
        graph.reset()

        # Don't do x[:] = 0.0 because we want to capture a new address
        # in the next cuda graph, to make sure we are running a new
        # cuda graph.
        x = torch.zeros([2000], device="cuda")
        with torch.cuda.graph(graph, capture_error_mode="relaxed"):
            x += 2.0

        graph.instantiate()
        graph.replay()
        self.assertTrue(torch.all(x == 2.0))
        # Exercise the code path where you reinstantiate the cuda graph twice.
        graph.instantiate()
        graph.replay()
        self.assertTrue(torch.all(x == 4.0))
        graph.replay()
        self.assertTrue(torch.all(x == 6.0))

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_cuda_graph_allocator_propagates_stream(self):
        segments = torch.cuda.memory_snapshot()
        existing_pools = {s["segment_pool_id"] for s in segments}
        x = torch.randn(10240000, device="cuda")
        y = torch.rand_like(x)
        g = torch.cuda.CUDAGraph()
        s0 = torch.cuda.Stream()
        s1 = torch.cuda.Stream()
        s0.wait_stream(torch.cuda.current_stream())
        with torch.cuda.stream(s0):
            g.capture_begin()
            z = x + y
        with torch.cuda.stream(s1):
            s1.wait_stream(s0)
            z + y
        s0.wait_stream(s1)
        with torch.cuda.stream(s0):
            g.capture_end()
        segments = torch.cuda.memory_snapshot()
        x = [
            s["segment_pool_id"]
            for s in segments
            if s["segment_pool_id"] not in existing_pools
        ]
        self.assertEqual(len(x), 2)
        self.assertEqual(x[0], x[1])

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_cuda_graph_tensor_item_not_allowed(self):
        test_script = """\
import torch
import sys
# Tensor.item() calls a synchronize which is not allowed in a cudagraph
# Valid for CUDA and ROCm
def my_func(a: torch.Tensor, b: torch.Tensor, perm: torch.Tensor):
    idx = perm[0]
    a[0] *= b[idx]  # should raise an error during capture
    return a

a = torch.rand(500, 500, device="cuda")
b = torch.rand(500, 500, device="cuda")
perm = torch.randint(0, 500, (500,), device="cuda")

g = torch.cuda.CUDAGraph()

with torch.cuda.graph(g):
    output = my_func(a, b, perm)
"""
        with self.assertRaisesRegex(
            subprocess.CalledProcessError,
            "calls a synchronize which is not allowed in a cudagraph",
        ):
            r = (
                subprocess.check_output([sys.executable, "-c", test_script])
                .decode("ascii")
                .strip()
            )

    def test_batch_norm_gather_stats(self):
        input = torch.randn(1, 3, 3, 3, device="cuda")
        mean, invstd = torch.batch_norm_gather_stats(
            input,
            mean=torch.ones(2, 3, device="cuda"),
            invstd=torch.ones(2, 3, device="cuda"),
            running_mean=None,
            running_var=None,
            momentum=0.1,
            eps=1e-5,
            count=2,
        )
        self.assertEqual(mean, torch.ones(3, device="cuda"))
        self.assertEqual(invstd, torch.ones(3, device="cuda"))

    def test_matmul_memory_use(self):
        def get_max_used():
            torch.cuda.synchronize()
            val = torch.cuda.max_memory_allocated()
            torch.cuda.reset_peak_memory_stats()
            return val

        a = torch.rand(1, 32, 32, device="cuda")
        b = torch.rand(24, 32, 1, device="cuda")

        get_max_used()

        torch.matmul(a, b)

        matmul_mem = get_max_used()

        a = a.expand(24, 32, 32)
        torch.matmul(a, b)

        matmul_expand_mem = get_max_used()

        torch.bmm(a, b)

        bmm_mem = get_max_used()

        self.assertEqual(matmul_expand_mem, matmul_mem)
        self.assertEqual(bmm_mem, matmul_mem)

    @unittest.skipIf(not TEST_WITH_ROCM, "ROCm-only test")
    def test_rocm_backward_pass_guard(self):
        # The test exercises a ROCm-specific feature.

        class MyFunction(torch.autograd.Function):
            @staticmethod
            def forward(ctx, tensor, constant):
                self.assertFalse(torch._C._rocm_is_backward_pass())
                ctx.constant = constant
                return tensor * constant

            @staticmethod
            def backward(ctx, grad_output):
                self.assertTrue(torch._C._rocm_is_backward_pass())
                return grad_output * ctx.constant, None

        class MyModule(torch.nn.Module):
            def __init__(self) -> None:
                super().__init__()
                self.a = torch.nn.Parameter(torch.randn(()))

            def forward(self, x):
                return MyFunction.apply(x, self.a)

        model = MyModule()
        criterion = torch.nn.MSELoss(reduction="sum")
        optimizer = torch.optim.SGD(model.parameters(), lr=1e-6)

        x = torch.randn(5, 5)
        result = model(x)
        loss = criterion(result, x)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

    def test_matmul_device_mismatch(self):
        cpu = torch.rand((10, 10))
        cuda = cpu.cuda()
        with self.assertRaisesRegex(
            RuntimeError, "Expected all tensors to be on the same device"
        ):
            cpu @ cuda
        with self.assertRaisesRegex(
            RuntimeError, "Expected all tensors to be on the same device"
        ):
            cuda @ cpu

        for s, m1, m2 in product((cpu, cuda), repeat=3):
            if s.device == m1.device == m2.device:
                torch.addmm(s, m1, m2)
            else:
                with self.assertRaisesRegex(
                    RuntimeError, "Expected all tensors to be on the same device"
                ):
                    torch.addmm(s, m1, m2)

    @unittest.skipIf(TEST_MULTIGPU, "Testing on one GPU is sufficient")
    def test_lazy_init(self):
        """Validate that no CUDA calls are made during `import torch` call"""

        def check_output(script: str) -> str:
            return (
                subprocess.check_output([sys.executable, "-c", script])
                .decode("ascii")
                .strip()
            )

        VISIBLE_DEVICES = (
            "HIP_VISIBLE_DEVICES" if TEST_WITH_ROCM else "CUDA_VISIBLE_DEVICES"
        )
        test_script = f"import os; import torch;os.environ['{VISIBLE_DEVICES}']='32';print(torch.cuda.device_count())"
        rc = check_output(test_script)
        self.assertEqual(rc, "0")
        if not TEST_WITH_ROCM:
            # Check that `cuInit` was not called during the import
            # By using ctypes and calling cuDeviceCountGet() and expect CUDA_ERROR_NOT_INITIALIZED == 3
            # See https://github.com/pytorch/pytorch/issues/116276 for more details
            libcuda_name = "libcuda.so.1" if not IS_WINDOWS else "nvcuda.dll"
            cuda_driver_api_call = (
                f"ctypes.CDLL('{libcuda_name}').cuDeviceGetCount(ctypes.byref(x))"
            )
            rc = check_output(
                f"import torch; import ctypes;x=ctypes.c_int(-1);print({cuda_driver_api_call})"
            )
            self.assertEqual(rc, "3")

    @unittest.skipIf(not TEST_WITH_ROCM, "not relevant for CUDA testing")
    def test_hip_device_count(self):
        """Validate device_count works with both CUDA/HIP visible devices"""
        test_script = """\
import torch
import os
print(f"{torch.cuda.device_count()}")
"""
        custom_envs = [
            {"CUDA_VISIBLE_DEVICES": "0", "HIP_VISIBLE_DEVICES": None},
            {"CUDA_VISIBLE_DEVICES": None, "HIP_VISIBLE_DEVICES": "0"},
            {"CUDA_VISIBLE_DEVICES": "0,1,2,3", "HIP_VISIBLE_DEVICES": "0"},
            {"ROCR_VISIBLE_DEVICES": "0", "HIP_VISIBLE_DEVICES": None},
        ]

        if torch.cuda.device_count() >= 2:
            custom_envs.extend(
                [
                    {"ROCR_VISIBLE_DEVICES": "1,2,3", "HIP_VISIBLE_DEVICES": "0"},
                ]
            )

        for env_config in custom_envs:
            env = os.environ.copy()
            for key, value in env_config.items():
                if value is None:
                    env.pop(key, None)
                else:
                    env[key] = value
            r = (
                subprocess.check_output([sys.executable, "-c", test_script], env=env)
                .decode("ascii")
                .strip()
            )
            self.assertEqual("1", r)

    @unittest.skipIf(not TEST_MULTIGPU, "requires multiple devices")
    def test_device_count_not_cached_pre_init(self):
        visible_devices = (
            "HIP_VISIBLE_DEVICES" if torch.version.hip else "CUDA_VISIBLE_DEVICES"
        )
        test_script = f"""\
import torch
import os
r1 = torch.cuda.device_count()
os.environ['{visible_devices}'] = '0'
r2 = torch.cuda.device_count()
torch.empty(10, device='cuda')
print(f"{{r1}}, {{r2}}")
"""

        r = (
            subprocess.check_output([sys.executable, "-c", test_script])
            .decode("ascii")
            .strip()
        )

        x = torch.cuda.device_count()
        self.assertEqual(f"{x}, 1", r)

    @unittest.skipUnless("CI" in os.environ, "Only run on CI")
    @unittest.skipIf(
        _get_torch_cuda_version() >= (13, 1),
        "This test does not fail on CUDA 13.1 or newer",
    )
    def test_gds_fails_in_ci(self):
        if IS_WINDOWS or TEST_WITH_ROCM:
            error_msg = "is not supported on this platform"
        else:
            error_msg = "cuFileHandleRegister failed"
        with TemporaryFileName() as f:
            with self.assertRaisesRegex(RuntimeError, error_msg):
                torch.cuda.gds.GdsFile(f, os.O_CREAT | os.O_RDWR)

    @unittest.skipIf(
        IS_WINDOWS, "test relies on fork; Windows multiprocessing uses spawn"
    )
    def test_is_pinned_no_context(self):
        test_script = """\
import torch
import multiprocessing


def fork_and_check_is_pinned():
    # Create a pipe to communicate between parent and child processes
    parent_conn, child_conn = multiprocessing.Pipe()

    def worker(conn):
        try:
            x = torch.randn(10)
            x.is_pinned()
            dev = torch.accelerator.current_accelerator()
            x = torch.ones(10, device=dev)[0].item()
            conn.send(x)
        except Exception as e:
            conn.send(str(e))
        finally:
            conn.close()
    # Fork a new process
    p = multiprocessing.Process(target=worker, args=(child_conn,))
    p.start()
    # Receive the result from the child process
    result = parent_conn.recv()
    parent_conn.close()
    # Wait for the child process to finish
    p.join()
    if isinstance(result, str) and result.startswith("Error"):
        raise RuntimeError(result)
    return result

x = torch.randn(10)
# check that is_pinned won't poison future fork
x.is_pinned()
ret = fork_and_check_is_pinned()
print(ret)

"""
        r = (
            subprocess.check_output([sys.executable, "-c", test_script])
            .decode("ascii")
            .strip()
        )
        self.assertEqual(r, "1.0")


@unittest.skipIf(not TEST_CUDA, "CUDA not available, skipping tests")
@torch.testing._internal.common_utils.markDynamoStrictTest
class TestCudaAllocator(TestCase):
    def tearDown(self):
        super().tearDown()
        # Regression check: no test in this class should leave the runtime
        # expandable_segments knob mismatched against the suite's env-derived
        # baseline. This is the only class that toggles it.
        md = torch.cuda.memory._snapshot()["allocator_settings"]
        self.assertEqual(md["expandable_segments"], EXPANDABLE_SEGMENTS)

    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    def test_memory_snapshot(self):
        try:
            torch.cuda.memory.empty_cache()
            torch.cuda.memory._record_memory_history("state", stacks="python")
            # make x the second block in a segment
            torch.rand(2 * 311, 411, device="cuda")
            unused = torch.rand(310, 410, device="cuda")
            x = torch.rand(311, 411, device="cuda")

            # create a bunch of tensors that all will tile into the
            # same segment to  exercise the history merging code
            # 512B is the minimum block size,
            # so we allocate all the tensors to this size to make sure
            # they tile evenly
            tensors = [torch.rand(128, device="cuda") for _ in range(1000)]
            while tensors:
                del tensors[randint(0, len(tensors) - 1)]

            # exercise the history trimming code
            torch.rand(128 * 5, device="cuda")

            ss = torch.cuda.memory._snapshot()
            found_it = False
            for seg in ss["segments"]:
                self.assertTrue("frames" in seg)
                for b in seg["blocks"]:
                    if b["requested_size"] == 311 * 411 * 4:
                        self.assertTrue("test_cuda" in b["frames"][0]["filename"])
                        found_it = True
                        self.assertEqual(x.untyped_storage().data_ptr(), b["address"])
            self.assertTrue(found_it)

            if not IS_WINDOWS:
                with tempfile.NamedTemporaryFile() as f:
                    torch.cuda.memory._save_segment_usage(f.name)
                    with open(f.name) as f2:
                        self.assertTrue("test_cuda.py" in f2.read())
            del unused
            del x
            torch.cuda.empty_cache()
            ss = torch.cuda.memory._snapshot()
            self.assertTrue(
                ss["device_traces"][0][-1]["action"]
                in ("segment_free", "segment_unmap")
            )

        finally:
            torch.cuda.memory._record_memory_history(None)

    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    def test_allocation_traceback(self):
        def this_is_a_funky_allocator():
            return torch.rand(311, 411, device="cuda")

        try:
            torch.cuda.memory.empty_cache()
            torch.cuda.memory._record_memory_history("all", stacks="python")

            x = this_is_a_funky_allocator()
            frames = torch.cuda.memory._allocation_traceback(x.data_ptr())
            self.assertIsInstance(frames, list)
            self.assertGreater(len(frames), 0)
            func_names = [f["name"] for f in frames]
            self.assertIn("this_is_a_funky_allocator", func_names)

            # freed pointer should return None
            ptr = x.data_ptr()
            del x
            torch.cuda.empty_cache()
            self.assertIsNone(torch.cuda.memory._allocation_traceback(ptr))
        finally:
            torch.cuda.memory._record_memory_history(None)

    def test_allocation_traceback_no_recording(self):
        torch.cuda.memory._record_memory_history(None)
        x = torch.rand(64, device="cuda")
        self.assertIsNone(torch.cuda.memory._allocation_traceback(x.data_ptr()))

    @unittest.skipUnless(
        (IS_X86 or IS_ARM64) and IS_LINUX, "linux x86/aarch64 only cpp unwinding"
    )
    def test_direct_traceback(self):
        from torch._C._profiler import gather_traceback, symbolize_tracebacks  # @manual

        c = gather_traceback(True, True, True)
        (r,) = symbolize_tracebacks([c])
        r = str(r)
        self.assertTrue("test_cuda.py" in r)
        self.assertTrue("unwind" in r)

    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    @requiresCppContext
    def test_memory_snapshot_with_cpp(self):
        try:
            torch.cuda.memory.empty_cache()
            torch.cuda.memory._record_memory_history("state", stacks="all")
            x = torch.rand(311, 411, device="cuda")

            ss = torch.cuda.memory._snapshot()["segments"]
            found_it = False
            for seg in ss:
                for b in seg["blocks"]:
                    if b["requested_size"] == 311 * 411 * 4:
                        self.assertTrue("::rand" in str(b["frames"]))
                        found_it = True
            self.assertTrue(found_it)

        finally:
            torch.cuda.memory._record_memory_history(None)

    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    def test_memory_snapshot_forward_trace(self):
        """Test that forward traces from anomaly mode are captured in memory snapshots."""
        try:
            torch.cuda.memory.empty_cache()
            torch.cuda.memory._record_memory_history("all", stacks="python")

            # Create a simple model with a backward pass
            # Use detect_anomaly(check_nan=False) to capture forward traces
            x = torch.randn(10, 10, device="cuda", requires_grad=True)

            with torch.autograd.detect_anomaly(check_nan=False):
                y = x * 2  # Forward pass - creates autograd node
                z = y.sum()
                z.backward()  # Backward pass - allocations here should have forward traces

            ss = torch.cuda.memory._snapshot()

            # Check that we have device traces with forward_frames
            found_forward_frames = False
            if "device_traces" in ss:
                for device_trace in ss["device_traces"]:
                    for entry in device_trace:
                        if isinstance(entry, dict) and "forward_frames" in entry:
                            # forward_frames should be a list of strings
                            self.assertIsInstance(entry["forward_frames"], list)
                            if len(entry["forward_frames"]) > 0:
                                found_forward_frames = True
                                # The forward frames should be strings
                                self.assertIsInstance(entry["forward_frames"][0], str)

            # forward_frames must be present when using detect_anomaly
            self.assertTrue(
                found_forward_frames,
                "forward_frames should be present in memory snapshot when using detect_anomaly",
            )

        finally:
            torch.cuda.memory._record_memory_history(None)

    @skipIfRocm(msg="ROCTracer does not capture Python stack frames in profiler output")
    def test_memory_profiler_viz(self):
        with torch.profiler.profile(
            with_stack=True, profile_memory=True, record_shapes=True
        ) as prof:
            x = torch.rand(128, 128, device="cuda")
            x * x + x * x
        plot = profile_plot(prof)
        plot = json.dumps(_profile_to_snapshot(prof))
        self.assertTrue("test_cuda.py" in plot)
        self.assertTrue("test_memory_profiler_viz" in plot)
        self.assertTrue("category" in plot)

    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    @requiresCppContext
    def test_cycles(self):
        fired = False

        def observer(html):
            nonlocal fired
            fired = True
            self.assertTrue("torch.Tensor" in html)
            self.assertTrue("test_cuda" in html)
            self.assertTrue("cell_contents" in html)

        disarm = observe_tensor_cycles(observer)

        def noop():
            pass

        try:

            def create():
                x = torch.empty(3, 4, device="cuda")

                def foo(p):
                    if p:
                        return foo(not p)
                    else:
                        return x

                return foo

            create()
            gc.collect()
            # the callback has to run outside of the collect
            # call so it doesn't actual fire until the next
            # method call after a gc.collect
            noop()
            self.assertTrue(fired)
        finally:
            disarm()

    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    @requiresCppContext
    def test_memory_plots(self):
        for context, stacks in (
            ("all", "all" if IS_LINUX else "python"),
            ("all", "python"),
            (None, "python"),
        ):
            try:
                torch.cuda.memory.empty_cache()
                torch.cuda.memory._record_memory_history(
                    "all", context=context, stacks=stacks
                )

                def run():
                    x = torch.rand(128, 128, device="cuda")
                    x * x + x * x

                run()
                cpp = stacks == "all"
                record_context = context is not None
                ss = torch.cuda.memory._snapshot()

                trace_plot(ss)
                trace_plot(ss, filter_freed=True)
                segment_plot(ss)
                text = json.dumps(ss)

                self.assertTrue(record_context == ("test_memory_plots" in text))
                self.assertTrue(cpp == ("::rand" in text))
                self.assertTrue(str(128 * 128 * 4) in text)

            finally:
                torch.cuda.memory._record_memory_history(None)

    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    @requiresCppContext
    def test_memory_plots_free_stack(self):
        for context in ["alloc", "all", "state"]:
            try:
                torch.cuda.memory.empty_cache()
                torch.cuda.memory._record_memory_history(context=context)
                x = None

                def thealloc():
                    nonlocal x
                    x = torch.rand(3, 4, device="cuda")

                def thefree():
                    nonlocal x
                    del x

                thealloc()
                thefree()
                ss = json.dumps(torch.cuda.memory._snapshot())
                self.assertEqual(("thefree" in ss), (context == "all"))
                self.assertEqual(("thealloc" in ss), (context != "state"))
            finally:
                torch.cuda.memory._record_memory_history(None)

    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    @unittest.skipIf(not has_triton(), "test needs triton")
    @requiresCppContext
    @serialTest()
    def test_memory_compile_regions(self):
        expected_allocation_sequence = [
            "Torch-Compiled Region: 0/0",
            "Torch-Compiled Region: 1/0",
            "Torch-Compiled Region: 0/0",
        ]

        class MyModel(nn.Module):
            def __init__(self):
                super().__init__()
                self.linear1 = nn.Linear(10, 10)
                self.linear2 = nn.Linear(10, 10)

            def forward(self, x):
                x = self.linear1(x)

                if x.sum() > 0:
                    x = x + 1
                else:
                    x = x - 1

                x = self.linear2(x)

                return x

        try:
            torch.cuda.memory.empty_cache()
            input_tensor = torch.randn(1, 10, device="cuda")
            # Create an instance of the model
            model = MyModel()
            model.to("cuda")
            # Compile the model using torch.compile
            compiled_model = torch.compile(model)
            # Create a sample input tensor
            torch.cuda.memory._record_memory_history(
                context="all", compile_context=True
            )
            compiled_model(input_tensor)
            ss = torch.cuda.memory._snapshot()["device_traces"]
            device_idx = 0
            allocation_sequence = []
            while len(ss[device_idx]) == 0:
                device_idx = device_idx + 1
            for s in ss[device_idx]:
                context = s["compile_context"]
                if context == "N/A":
                    continue
                if len(allocation_sequence) > 0 and allocation_sequence[-1] == context:
                    continue
                allocation_sequence.append(context)
            self.assertTrue(allocation_sequence == expected_allocation_sequence)
        except RuntimeError as e:
            pass
        finally:
            torch.cuda.memory._record_memory_history(None)

    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    @requiresCppContext
    def test_memory_plots_history_context(self):
        try:
            torch.cuda.memory.empty_cache()
            x = None

            def should_capture1():
                nonlocal x
                x = torch.rand(4, 4, device="cuda")

            def should_not_capture():
                nonlocal x
                x = torch.rand(3, 4, device="cuda")

            def should_capture2():
                nonlocal x
                x = torch.rand(4, 4, device="cuda")

            # Recording with context and python call stacks should capture the call stack.
            torch.cuda.memory._record_memory_history(context="all", stacks="python")
            should_capture1()
            # Recording with context=None should not capture the call stack.
            torch.cuda.memory._record_memory_history(context=None)
            should_not_capture()
            # Recording with context and python call stacks should capture the call stack.
            torch.cuda.memory._record_memory_history(context="all", stacks="python")
            should_capture2()

            ss = json.dumps(torch.cuda.memory._snapshot())
            self.assertTrue("should_capture1" in ss)
            self.assertTrue("should_not_capture" not in ss)
            self.assertTrue("should_capture2" in ss)
        finally:
            torch.cuda.memory._record_memory_history(None)

    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    @requiresCppContext
    def test_memory_plots_free_segment_stack(self):
        for context in ["alloc", "all", "state"]:
            try:
                torch._C._cuda_clearCublasWorkspaces()
                torch.cuda.memory.empty_cache()
                torch.cuda.memory._record_memory_history(context=context)
                x = torch.rand(3, 4, device="cuda")
                del x
                torch.cuda.memory.empty_cache()

                ss = json.dumps(torch.cuda.memory._snapshot())
                self.assertEqual(("empty_cache" in ss), (context == "all"))
            finally:
                torch.cuda.memory._record_memory_history(None)

    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    @requiresCppContext
    def test_memory_plots_metadata(self):
        for context in ["alloc", "all", "state"]:
            try:
                torch._C._cuda_clearCublasWorkspaces()
                torch.cuda.memory.empty_cache()
                torch.cuda.memory._set_memory_metadata("metadata test")
                torch.cuda.memory._record_memory_history(context=context)
                x = torch.rand(3, 4, device="cuda")
                del x
                torch.cuda.memory.empty_cache()
                torch.cuda.memory._set_memory_metadata("")

                ss = torch.cuda.memory._snapshot()
                for event in ss["device_traces"][0]:
                    self.assertTrue(event["user_metadata"] == "metadata test")
            finally:
                torch.cuda.memory._record_memory_history(None)

    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    def test_memory_snapshot_script(self):
        try:
            torch._C._cuda_clearCublasWorkspaces()
            torch.cuda.memory.empty_cache()
            torch.cuda.memory._record_memory_history("state", stacks="python")

            @torch.jit.script
            def foo():
                return torch.rand(311, 411, device="cuda")

            x = foo()

            ss = torch.cuda.memory._snapshot()["segments"]
            found_it = False
            for seg in ss:
                for b in seg["blocks"]:
                    if b["requested_size"] == 311 * 411 * 4:
                        self.assertEqual(b["frames"][0]["name"], "foo")
                        found_it = True
            self.assertTrue(found_it)

        finally:
            torch.cuda.memory._record_memory_history(None)

    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    def test_memory_snapshot_skip_actions(self):
        """Test that skip_actions parameter correctly filters out specified action types."""
        # Test cases: (skip_actions, description)
        test_cases = [
            (["free_requested"], "single action"),
            (["free_requested", "free_completed"], "multiple actions"),
            (["alloc"], "alloc action"),
            (["segment_alloc", "segment_free"], "segment actions"),
            ([], "empty list"),
            (None, "None default"),
        ]

        for skip_actions, description in test_cases:
            with self.subTest(skip_actions=skip_actions, description=description):
                try:
                    torch.cuda.memory.empty_cache()
                    torch.cuda.memory._record_memory_history(
                        context="all", stacks="python", skip_actions=skip_actions
                    )

                    # Perform allocations and frees to generate various actions
                    x = torch.rand(128, 128, device="cuda")
                    del x
                    torch.cuda.synchronize()

                    # Take snapshot to potentially generate snapshot event
                    ss = torch.cuda.memory._snapshot()
                    device_idx = 0
                    while len(ss["device_traces"][device_idx]) == 0:
                        device_idx = device_idx + 1
                    device_trace = ss["device_traces"][device_idx]

                    # Verify traces are not empty (other actions are still recorded)
                    self.assertGreater(
                        len(device_trace),
                        0,
                        f"Trace should not be empty for skip_actions={skip_actions}",
                    )

                    if skip_actions:
                        all_actions = {event["action"] for event in device_trace}

                        # Verify skipped actions are not in traces
                        for sa in skip_actions:
                            self.assertNotIn(
                                sa,
                                all_actions,
                                f"Action {sa} should be skipped with skip_actions={skip_actions}",
                            )
                        # Check that we have actions that are NOT in the skip list
                        non_skipped_actions = all_actions - set(skip_actions)
                        self.assertGreater(
                            len(non_skipped_actions),
                            0,
                            f"Should have non-skipped actions for skip_actions={skip_actions}",
                        )

                finally:
                    torch.cuda.memory._record_memory_history(None)

    @serialTest()
    def test_max_split_expandable(self):
        try:
            orig = torch.cuda.get_per_process_memory_fraction()
            torch.cuda.memory.empty_cache()
            mb = 1024 * 1024
            _, all_memory = torch.cuda.memory.mem_get_info()
            pre_reserved = torch.cuda.memory_reserved()
            total_allowed = 120 * mb + pre_reserved
            fraction_allowed = total_allowed / all_memory
            self.assertEqual(int(round(fraction_allowed * all_memory)), total_allowed)
            torch.cuda.memory.set_per_process_memory_fraction(fraction_allowed)

            def alloc(n):
                return torch.ones(n * mb, dtype=torch.int8, device="cuda")

            torch.cuda.memory._set_allocator_settings(
                "expandable_segments:False,max_split_size_mb:40"
            )
            a = alloc(40)
            torch.cuda.memory._set_allocator_settings(
                "expandable_segments:True,max_split_size_mb:40"
            )
            b = alloc(40)
            torch.cuda.memory._set_allocator_settings(
                "expandable_segments:False,max_split_size_mb:40"
            )
            c = alloc(40)
            with self.assertRaises(torch.OutOfMemoryError):
                alloc(40)
            del a, b, c
            # force release_cached_blocks to run with some expandable segments in the free list
            alloc(120)
        finally:
            torch.cuda.memory.set_per_process_memory_fraction(orig)
            # Test toggles expandable_segments internally; restore the
            # suite's baseline so subsequent tests see consistent state.
            torch.cuda.memory._set_allocator_settings(
                f"expandable_segments:{EXPANDABLE_SEGMENTS}"
            )

    @serialTest()
    def test_garbage_collect_expandable(self):
        try:
            orig = torch.cuda.get_per_process_memory_fraction(0)
            torch.cuda.memory.empty_cache()
            mb = 1024 * 1024
            _, all_memory = torch.cuda.memory.mem_get_info()
            pre_reserved = torch.cuda.memory_reserved()
            total_allowed = 120 * mb + pre_reserved
            fraction_allowed = total_allowed / all_memory
            self.assertEqual((fraction_allowed * all_memory), total_allowed)
            torch.cuda.memory.set_per_process_memory_fraction(fraction_allowed)

            def alloc(n):
                return torch.ones(n * mb, dtype=torch.int8, device="cuda")

            torch.cuda.memory._set_allocator_settings(
                "expandable_segments:False,garbage_collection_threshold:0.5"
            )
            a = alloc(40)
            torch.cuda.memory._set_allocator_settings(
                "expandable_segments:True,garbage_collection_threshold:0.5"
            )
            b = alloc(40)
            del a, b
            # causes GC to run. The expandable segment block will be split
            # so GC would not attempt to free it anyway, but this at least makes sure
            # expandable_segment blocks can be in the free list when this is called.
            alloc(80)
        finally:
            torch.cuda.memory.set_per_process_memory_fraction(orig)
            # Test toggles expandable_segments internally; restore the
            # suite's baseline so subsequent tests see consistent state.
            torch.cuda.memory._set_allocator_settings(
                f"expandable_segments:{EXPANDABLE_SEGMENTS}"
            )

    def test_allocator_settings(self):
        def power2_div(size, div_factor):
            pow2 = 1
            while pow2 < size:
                pow2 = pow2 * 2
            if pow2 == size:
                return pow2
            step = pow2 / 2 / div_factor
            ret = pow2 / 2
            while ret < size:
                ret = ret + step
            return ret

        torch.cuda.memory.empty_cache()
        key_allocated = (
            "active_bytes.all.allocated"
            if not TEST_CUDAMALLOCASYNC
            else "allocated_bytes.all.current"
        )
        key_requested = "requested_bytes.all.allocated"

        nelems = 21 * 1024 * 1024
        nbytes = 4 * nelems  # floats are 4 bytes

        nelems_big = 100 * 1024 * 1024
        nbytes_big = 4 * nelems_big  # floats are 4 bytes

        start_mem = torch.cuda.memory_stats()[key_allocated]
        torch.cuda.memory._set_allocator_settings("")
        x = torch.rand(nelems, device="cuda")

        # test roundup_power2_divisions single value syntax
        reg_mem = torch.cuda.memory_stats()[key_allocated]
        start_requested = torch.cuda.memory_stats()[key_requested]
        torch.cuda.memory._set_allocator_settings("roundup_power2_divisions:4")
        y = torch.rand(nelems, device="cuda")

        pow2_div4_mem = torch.cuda.memory_stats()[key_allocated]
        current_requested = torch.cuda.memory_stats()[key_requested]

        self.assertEqual(reg_mem - start_mem, nbytes)
        if not TEST_CUDAMALLOCASYNC:
            # not supported with the cudaMallocAsync backend
            self.assertEqual(pow2_div4_mem - reg_mem, power2_div(nbytes, 4))
            self.assertEqual(current_requested - start_requested, nbytes)

        torch.cuda.memory._set_allocator_settings("garbage_collection_threshold:0.5")
        torch.cuda.memory._set_allocator_settings(
            "garbage_collection_threshold:0.5,max_split_size_mb:40"
        )

        # should have reset the power2 divisions now
        torch.cuda.memory.empty_cache()
        start_mem = torch.cuda.memory_stats()[key_allocated]
        z = torch.rand(nelems, device="cuda")
        reg_mem = torch.cuda.memory_stats()[key_allocated]
        self.assertEqual(reg_mem - start_mem, nbytes)

        # roundup_power2_divisions knob array syntax
        torch.cuda.memory.empty_cache()
        torch.cuda.memory._set_allocator_settings(
            "garbage_collection_threshold:0.5,roundup_power2_divisions:[64:8,128:2,256:2,512:2,1024:1,>:1]"
        )
        start_mem = torch.cuda.memory_stats()[key_allocated]
        w = torch.rand(nelems, device="cuda")

        pow2_div8_mem = torch.cuda.memory_stats()[key_allocated]
        if not TEST_CUDAMALLOCASYNC:
            # not supported with the cudaMallocAsync backend
            self.assertEqual(pow2_div8_mem - start_mem, power2_div(nbytes, 8))

        torch.cuda.memory.empty_cache()
        start_mem = torch.cuda.memory_stats()[key_allocated]
        v = torch.rand(nelems_big, device="cuda")

        pow2_div2_mem = torch.cuda.memory_stats()[key_allocated]
        if not TEST_CUDAMALLOCASYNC:
            # not supported with the cudaMallocAsync backend
            self.assertEqual(pow2_div2_mem - start_mem, power2_div(nbytes_big, 2))

        torch.cuda.memory.empty_cache()
        torch.cuda.memory._set_allocator_settings("release_lock_on_cudamalloc:True")
        start_mem = torch.cuda.memory_stats()[key_allocated]
        w = torch.rand(nelems, device="cuda")
        reg_mem = torch.cuda.memory_stats()[key_allocated]
        self.assertEqual(reg_mem - start_mem, nbytes)

        with self.assertRaises(ValueError):
            torch._C._accelerator_setAllocatorSettings("foo:1,bar:2")

        with self.assertRaises(ValueError):
            torch._C._accelerator_setAllocatorSettings(
                "garbage_collection_threshold:1.2"
            )

        with self.assertRaises(ValueError):
            torch._C._accelerator_setAllocatorSettings("max_split_size_mb:2")

        with self.assertRaises(ValueError):
            torch._C._accelerator_setAllocatorSettings(
                "release_lock_on_cudamalloc:none"
            )

        with self.assertRaises(ValueError):
            torch._C._accelerator_setAllocatorSettings(
                "pinned_use_cuda_host_register:none"
            )

        with self.assertRaises(ValueError):
            torch._C._accelerator_setAllocatorSettings(
                "pinned_num_register_threads:none"
            )

        with self.assertRaises(ValueError):
            torch._C._accelerator_setAllocatorSettings(
                "pinned_num_register_threads:1024"
            )

        # Test throw_on_cudamalloc_oom config parsing - valid formats
        torch.cuda.memory._set_allocator_settings("throw_on_cudamalloc_oom:True")
        torch.cuda.memory._set_allocator_settings("throw_on_cudamalloc_oom:False")

        # Test throw_on_cudamalloc_oom config parsing - invalid formats
        with self.assertRaises(ValueError):
            torch._C._accelerator_setAllocatorSettings("throw_on_cudamalloc_oom:maybe")

    @unittest.skipIf(TEST_CUDAMALLOCASYNC, "throw_on_cudamalloc_oom not supported")
    @serialTest()
    def test_throw_on_cudamalloc_oom(self):
        """Test that throw_on_cudamalloc_oom + per_process_memory_fraction works correctly."""
        torch.cuda.empty_cache()
        device = torch._C._cuda_getDevice()
        torch._C._cuda_resetAccumulatedMemoryStats(device)

        try:
            # Test 1: With rejection disabled (default), allocations should succeed
            torch._C._accelerator_setAllocatorSettings("throw_on_cudamalloc_oom:False")
            x = torch.empty(10 * 1024 * 1024, dtype=torch.int8, device="cuda")
            del x
            torch.cuda.empty_cache()

            # Test 2: With throw_on_cudamalloc_oom enabled and a tight memory
            # fraction, allocations that exceed the fraction limit should be
            # preemptively rejected with OutOfMemoryError.
            # Both settings must go through _accelerator_setAllocatorSettings so
            # they are read from CUDAAllocatorConfig.
            fraction = 0.005
            torch._C._accelerator_setAllocatorSettings(
                f"throw_on_cudamalloc_oom:True,per_process_memory_fraction:{fraction}"
            )

            total_mem = torch.cuda.get_device_properties(0).total_memory
            # Allocate the allowed threshold + 1 MiB to guarantee rejection
            alloc_bytes = int(total_mem * fraction) + 1024 * 1024
            with self.assertRaises(torch.cuda.OutOfMemoryError):
                torch.empty(alloc_bytes, dtype=torch.int8, device="cuda")

            # Check that rejection counter was incremented
            stats = torch.cuda.memory_stats()
            self.assertGreater(stats["num_oom_rejections"], 0)

        finally:
            torch.cuda.empty_cache()
            torch._C._accelerator_setAllocatorSettings(
                "throw_on_cudamalloc_oom:False,per_process_memory_fraction:1.0"
            )

    def test_allocator_backend(self):
        def check_output(script: str) -> str:
            return (
                subprocess.check_output([sys.executable, "-c", script])
                .decode("ascii")
                .strip()
            )

        test_script = """\
import os
os.environ["PYTORCH_ALLOC_CONF"] = "max_split_size_mb:20,backend:cudaMallocAsync,release_lock_on_cudamalloc:none"
import torch
torch.cuda.init()
print(torch.cuda.get_allocator_backend())
"""
        rc = check_output(test_script)
        self.assertEqual(rc, "cudaMallocAsync")

    def test_allocator_memory_fraction_setting(self):
        def make_env(fraction):
            env = os.environ.copy()
            var = "PYTORCH_CUDA_ALLOC_CONF"
            key = "per_process_memory_fraction"
            value = [
                x
                for x in env.get(var, "").split(",")
                if len(x) > 0 and not x.startswith(f"{key}:")
            ]
            value.append(f"{key}:{fraction}")
            env[var] = ",".join(value)
            return env

        def run_test(value):
            test_script = """\
import os
import torch
device = torch._C._cuda_getDevice()
value = torch.cuda.memory.get_per_process_memory_fraction(device)
print(value, end="")
            """
            return subprocess.run(
                [sys.executable, "-c", test_script],
                env=make_env(value),
                text=True,
                check=True,
                capture_output=True,
            )

        self.assertEqual(run_test(0.0).stdout, "0.0")
        self.assertEqual(run_test(0.5).stdout, "0.5")
        self.assertEqual(run_test(1.0).stdout, "1.0")

        with self.assertRaises(subprocess.CalledProcessError) as e:
            run_test(-0.1)
        if "per_process_memory_fraction is invalid" not in e.exception.stderr:
            raise AssertionError(e.exception.stderr)

        with self.assertRaises(subprocess.CalledProcessError) as e:
            run_test(1.1)
        if "per_process_memory_fraction is invalid" not in e.exception.stderr:
            raise AssertionError(e.exception.stderr)

    def test_cachingAllocator_raw_alloc(self):
        # Test that raw_alloc respects the setting that
        # activates/deactivates the caching allocator

        # Helper function that calls raw_alloc and returns
        # relevant field in data structure
        def requested_bytes_alloc_stats(raw_alloc_size, stream):
            start = torch.cuda.memory_stats()["requested_bytes.all.allocated"]
            mem_ptr = torch._C._cuda_cudaCachingAllocator_raw_alloc(
                raw_alloc_size, stream
            )
            finish = torch.cuda.memory_stats()["requested_bytes.all.allocated"]
            torch._C._cuda_cudaCachingAllocator_raw_delete(mem_ptr)
            return finish - start

        torch.cuda.empty_cache()
        device = torch._C._cuda_getDevice()
        stream = torch._C._cuda_getCurrentRawStream(device)
        torch._C._cuda_resetAccumulatedMemoryStats(device)

        # size of allocation
        raw_alloc_size = 1024 * 1024  # 1 MB

        try:
            # Deactivate the caching allocator
            torch.cuda.caching_allocator_enable(False)

            # For a deactivated caching allocator, result is zero
            cuda_alloc_size = requested_bytes_alloc_stats(raw_alloc_size, stream)
            self.assertEqual(cuda_alloc_size, 0)

        finally:
            # Make sure we get back to the default state that is
            # an activated caching allocator
            torch.cuda.caching_allocator_enable(True)

            # For an active caching allocator, result matches raw_alloc_size
            cuda_alloc_size = requested_bytes_alloc_stats(raw_alloc_size, stream)
            self.assertEqual(cuda_alloc_size, raw_alloc_size)

    @unittest.skipIf(
        IS_JETSON, "oom reporting has issues on jetson igx due to partial nvml support"
    )
    @parametrize("max_split_size_mb_setting", [False, True])
    def test_raises_oom(self, max_split_size_mb_setting):
        if max_split_size_mb_setting:
            # CudaCachingAllocator does early return when searching available blocks
            # if max_split_size_mb is not set
            # Setting this triggers more parts of the code
            torch.cuda.memory._set_allocator_settings("max_split_size_mb:1024")
            torch.cuda.memory.empty_cache()
        with self.assertRaises(torch.cuda.OutOfMemoryError):
            torch.empty(1024 * 1024 * 1024 * 1024, device="cuda")

    @unittest.skipIf(
        not ((IS_X86 or IS_ARM64) and IS_LINUX), "cpp traces are linux x86/aarch64 only"
    )
    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    def test_cpp_memory_snapshot_pickle(self):
        from torch.utils.cpp_extension import load_inline

        source = """
        #include <torch/csrc/cuda/memory_snapshot.h>
        py::object do_snapshot() {
            std::string data = torch::cuda::_memory_snapshot_pickled();
            return py::bytes(data);
        }
        void record(bool e, bool ctx) {
            torch::cuda::_record_memory_history(e, ctx, 10, ctx, ctx);
        }
        """
        m = load_inline(
            name="snapshot", cpp_sources=[source], functions=["do_snapshot", "record"]
        )
        for ctx in (False, True):
            try:
                m.record(True, ctx)

                @torch.jit.script
                def the_script_fn():
                    return torch.rand(311, 411, device="cuda")

                def run():
                    t = the_script_fn()
                    return pickle.loads(m.do_snapshot())

                mem = run()
                found = False
                for s in mem["segments"]:
                    for b in s["blocks"]:
                        if b["state"] == "active_allocated":
                            if b["requested_size"] == 311 * 411 * 4:
                                if ctx:
                                    frame_text = str(b["frames"])
                                    # C++ frame
                                    self.assertTrue("::rand" in frame_text)
                                    # script frame
                                    self.assertTrue("the_script_fn" in frame_text)
                                    # python frame
                                    self.assertTrue("case.py" in frame_text)
                                found = True
                last_action = mem["device_traces"][0][-1]
                self.assertEqual(last_action["action"], "alloc")
                self.assertEqual(last_action["size"], 311 * 411 * 4)
                self.assertTrue(found)
            finally:
                m.record(False, False)

    @unittest.skipIf(TEST_CUDAMALLOCASYNC, "temporarily disabled")
    @unittest.skipIf(
        IS_JETSON, "oom reporting has issues on jetson igx due to partial nvml support"
    )
    def test_notifies_oom(self):
        x = False

        def cb(device, alloc, device_alloc, device_free):
            nonlocal x
            x = True

        torch._C._cuda_attach_out_of_memory_observer(cb)
        with self.assertRaises(torch.cuda.OutOfMemoryError):
            torch.empty(1024 * 1024 * 1024 * 1024, device="cuda")
        self.assertTrue(x)

    def test_allocator_fuzz(self):
        # fuzz
        if (
            torch.version.hip
            and "expandable_segments:True"
            in torch._C._accelerator_getAllocatorSettings()
        ):
            raise unittest.SkipTest(
                "ROCm needs https://github.com/ROCm/rocm-systems/pull/3023"
            )
        state = random.getstate()
        random.seed(123)
        N = 10000
        try:
            mem = []
            total = 0
            c = 0

            def alloc():
                nonlocal total, c
                b = random.randrange(2 * 1024 * 1024 // 4, 20 * 1024 * 1024 // 4)
                mem.append((c, torch.full((b,), c, dtype=torch.int32, device="cuda")))
                c += 1
                total += b

            def free():
                nonlocal total
                idx = random.randrange(0, len(mem))
                v, x = mem.pop(idx)
                self.assertTrue(torch.all(v == x))
                total -= x.numel()

            choices = [alloc, free, torch.cuda.memory.empty_cache]
            for _ in range(N):
                while total >= 1024 * 1024 * 1024 / (4 * 10):
                    free()
                (action,) = random.choices(choices, weights=[1, 1 if mem else 0, 0.1])
                action()
        finally:
            random.setstate(state)

    @unittest.skipIf(not TEST_PYNVML, "pynvml/amdsmi is not available")
    def test_nvml_get_handler(self):
        if not torch.version.hip:
            self.assertTrue(torch.cuda._get_pynvml_handler() is not None)
        else:
            self.assertTrue(torch.cuda._get_amdsmi_handler() is not None)

    @unittest.skipIf(not TEST_PYNVML, "pynvml/amdsmi is not available")
    def test_temperature(self):
        self.assertTrue(0 <= torch.cuda.temperature() <= 150)

    @unittest.skipIf(not TEST_PYNVML, "pynvml/amdsmi is not available")
    def test_device_memory_used(self):
        """
        Verify used device memory in bytes.
        On Windows the NVML used value has been observed not to increase after
        a CUDA allocation (delta 0); we only assert API sanity there (non-negative,
        non-decreasing after alloc, <= total memory). Need to investigate expected behavior
        with Windows WDDM
        """
        torch.cuda.synchronize()
        gc.collect()
        torch.cuda.empty_cache()
        a = torch.cuda.device_memory_used()
        num_bytes = 512 * 1024**2
        _ = torch.empty(num_bytes, dtype=torch.int8, device="cuda")
        torch.cuda.synchronize()
        torch.cuda.empty_cache()
        b = torch.cuda.device_memory_used()
        if IS_WINDOWS:
            # NVML used memory does not reflect CUDA allocations on WDDM; only check API sanity
            self.assertGreaterEqual(a, 0, "device_memory_used should be non-negative")
            self.assertGreaterEqual(b, 0, "device_memory_used should be non-negative")
            self.assertGreaterEqual(
                b, a, "used memory should not decrease after allocation"
            )
            total = torch.cuda.get_device_properties(0).total_memory
            self.assertLessEqual(a, total, "used should not exceed total device memory")
            self.assertLessEqual(b, total, "used should not exceed total device memory")
        else:
            mem_bytes = b - a
            # test the order of magnitude
            self.assertTrue(num_bytes // 32 <= mem_bytes <= num_bytes * 32)

    @unittest.skipIf(not TEST_PYNVML, "pynvml/amdsmi is not available")
    def test_power_draw(self):
        self.assertTrue(torch.cuda.power_draw() >= 0)

    @unittest.skipIf(not TEST_PYNVML, "pynvml/amdsmi is not available")
    def test_clock_speed(self):
        self.assertTrue(torch.cuda.clock_rate() >= 0)

    @unittest.skipIf(not TEST_PYNVML, "pynvml/amdsmi is not available")
    @unittest.skipIf(not TEST_WITH_ROCM, "amdsmi specific test")
    def test_raw_amdsmi_device_count(self):
        """
        This unit test will verify if the number of GPUs shown in `amd-smi
        list` is equivalent to the count returned by `_raw_device_count_amdsmi`.
        This should be unaffected by visible device settings.
        """
        raw_device_cnt = int(
            subprocess.check_output(
                "amd-smi list | grep 'GPU' | wc -l", shell=True
            ).strip()
        )
        self.assertEqual(torch.cuda._raw_device_count_amdsmi(), raw_device_cnt)

    @unittest.skipIf(not TEST_PYNVML, "pynvml/amdsmi is not available")
    @unittest.skipIf(not TEST_WITH_ROCM, "amdsmi specific test")
    def test_raw_amdsmi_device_uuids(self):
        """
        This unit test will extract a list of UUIDs for each GPU using
        rocminfo information, and check whether each UUID is present in
        the output from `_raw_device_uuid_amdsmi` this allows us to test
        that the pytorch call is returning a correct list of UUIDs.
        """
        cmd = "rocminfo | grep -o 'Uuid:.*GPU-.*' | sed 's/Uuid:.*GPU-//'"
        uuids = subprocess.check_output(cmd, shell=True, text=True).strip().split("\n")
        uuids = [s.strip() for s in uuids]
        raw_uuids = torch.cuda._raw_device_uuid_amdsmi()
        for uuid in uuids:
            matching = True
            if not any(uuid in raw_id for raw_id in raw_uuids):
                matching = False
        self.assertEqual(True, matching)

    @unittest.skipIf(not TEST_PYNVML, "pynvml/amdsmi is not available")
    @unittest.skipIf(not TEST_WITH_ROCM, "amdsmi specific test")
    def test_uuid_visible_devices(self):
        """
        This unit test will simulate an environment where a UUID is passed
        via CUDA/HIP_VISIBLE_DEVICES and ensure that the correct device count
        is returned. This allows us to test that the visible device functionality
        is operating as expected.
        """
        test_script = """\
import torch
import os
print(f"{torch.cuda.device_count()}")
        """
        cmd = "rocminfo | grep -o 'Uuid:.*GPU-.*' | sed 's/Uuid://'"
        uuids = subprocess.check_output(cmd, shell=True, text=True).strip().split("\n")
        uuids = [s.strip() for s in uuids]

        custom_envs = []
        for uuid in uuids:
            custom_envs.append(
                {"CUDA_VISIBLE_DEVICES": f"{uuid}", "HIP_VISIBLE_DEVICES": None}
            )
            custom_envs.append(
                {"HIP_VISIBLE_DEVICES": f"{uuid}", "CUDA_VISIBLE_DEVICES": None}
            )

        for env_config in custom_envs:
            env = os.environ.copy()
            for key, value in env_config.items():
                if value is None:
                    env.pop(key, None)
                else:
                    env[key] = value
            r = (
                subprocess.check_output([sys.executable, "-c", test_script], env=env)
                .decode("ascii")
                .strip()
            )
            self.assertEqual("1", r)


MIN_BLOCK_SIZE = 512
SMALL_SIZE = 1048576
SMALL_BUFFER = 2097152
LARGE_BUFFER = 20971520


def get_cudagraph_segments(pool_id):
    segments = torch.cuda.memory_snapshot()
    return [segment for segment in segments if segment["segment_pool_id"] == pool_id]


def get_all_cudagraph_segments():
    segments = torch.cuda.memory_snapshot()
    return [segment for segment in segments if segment["segment_pool_id"] != (0, 0)]


def cudagraphify(fn, inputs, pool=None, stream=None):
    if not TEST_CUDA_GRAPH:
        raise unittest.SkipTest("cuda graph test is skipped")

    torch.cuda.synchronize()
    if stream is None:
        stream = torch.cuda.Stream()

    stream.wait_stream(torch.cuda.current_stream())
    with torch.cuda.stream(stream):
        fn(*inputs)
    stream.synchronize()
    torch.cuda.current_stream().wait_stream(stream)
    torch.cuda.synchronize()

    graph = torch.cuda.CUDAGraph()
    with torch.cuda.graph(graph, stream=stream, pool=pool):
        static_outputs = fn(*inputs)

    return graph, static_outputs


def int8_cuda(size):
    return torch.ones([size], device="cuda", dtype=torch.uint8)


def live_blocks(pool_id):
    blocks = 0
    seg = get_cudagraph_segments(pool_id)
    for segment in get_cudagraph_segments(pool_id):
        for block in segment["blocks"]:
            blocks += block["state"] == "active_allocated"
    return blocks


def tensor_metadata(x):
    return {
        "nbytes": x.untyped_storage().nbytes(),
        "data_ptr": x.untyped_storage().data_ptr(),
        "size": x.shape,
        "stride": x.stride(),
        "dtype": x.dtype,
        "device": x.device,
        "storage_offset": x.storage_offset(),
    }


def reconstruct_from_tensor_metadata(metadata):
    s = torch._C._construct_storage_from_data_pointer(
        metadata["data_ptr"], metadata["device"], metadata["nbytes"]
    )
    t = torch.empty([0], device=metadata["device"], dtype=metadata["dtype"])
    t.set_(
        source=s,
        storage_offset=metadata["storage_offset"],
        size=metadata["size"],
        stride=metadata["stride"],
    )
    return t


@unittest.skipIf(
    not TEST_CUDA or TEST_CUDAMALLOCASYNC,
    "CUDA required, not supported with CUDAMallocAsync",
)
@torch.testing._internal.common_utils.markDynamoStrictTest
class TestBlockStateAbsorption(TestCase):
    @property
    def expandable_segments(self):
        return EXPANDABLE_SEGMENTS

    def checkCheckpointedBlock(self, before_block, after_block):
        for field in ("size", "state"):
            self.assertEqual(before_block[field], after_block[field])

    def checkCheckpointedState(self, before_segments, after_segments):
        # after may contain additional segments, but all of the segments in before
        # should be exactly equivalent to after
        after_ptr_to_segment = {
            segment["address"]: segment for segment in after_segments
        }

        for before_segment in before_segments:
            self.assertTrue(before_segment["address"] in after_ptr_to_segment)
            after_segment = after_ptr_to_segment[before_segment["address"]]

            for field in (
                "device",
                "total_size",
                "allocated_size",
                "active_size",
                "segment_type",
                "segment_pool_id",
            ):
                self.assertEqual(before_segment[field], after_segment[field])

            self.assertEqual(
                len(before_segment["blocks"]), len(after_segment["blocks"])
            )
            for before_block, after_block in zip(
                before_segment["blocks"], after_segment["blocks"]
            ):
                self.checkCheckpointedBlock(before_block, after_block)

    @staticmethod
    def setCheckpointPoolState(
        device, state, stale_storages_ptr, storages_deleters=None
    ):
        stale_storages_ptr = [t.untyped_storage()._cdata for t in stale_storages_ptr]
        storages_deleters = (
            []
            if not storages_deleters
            else [t.untyped_storage()._cdata for t in storages_deleters]
        )
        torch._C._cuda_setCheckpointPoolState(
            device, state, stale_storages_ptr, storages_deleters
        )

    def checkFunction(self, fn, inputs, pool=None):
        graph, outputs = cudagraphify(fn, inputs, pool=pool)

        pool_id = graph.pool()
        device = outputs[0].device.index

        segments_before_checkpoint = get_cudagraph_segments(pool_id)

        state = torch._C._cuda_getCheckpointState(device, pool_id)
        self.setCheckpointPoolState(device, state, [], [])

        self.checkCheckpointedState(
            segments_before_checkpoint, get_cudagraph_segments(pool_id)
        )

    def setUp(self):
        super().setUp()
        self.segment_length = len(get_all_cudagraph_segments())

    def tearDown(self):
        torch.cuda.synchronize()
        gc.collect()
        torch.cuda.empty_cache()

        self.assertEqual(len(get_all_cudagraph_segments()), self.segment_length)

        super().tearDown()

    def test_simple(self):
        def foo():
            x = torch.zeros([SMALL_SIZE * 8], device="cuda", dtype=torch.uint8)
            x = x + x
            x1 = int8_cuda(SMALL_SIZE) + int8_cuda(SMALL_SIZE) + int8_cuda(SMALL_SIZE)
            y = int8_cuda(SMALL_SIZE) + x1
            z = int8_cuda(SMALL_SIZE)
            return x, y, z

        self.checkFunction(foo, [])

    def test_allocated_in_middle_of_segment(self):
        def foo():
            small_buffers = [int8_cuda(MIN_BLOCK_SIZE) for _ in range(11)]
            return small_buffers[5].add_(2)

        self.checkFunction(foo, [])

    def test_multiple_middle_allocations(self):
        def foo():
            small_buffers = [int8_cuda(MIN_BLOCK_SIZE) for _ in range(11)]
            return small_buffers[5], small_buffers[8]

        self.checkFunction(foo, [])

    def test_middle_allocations_contiguous(self):
        def foo():
            small_buffers = [int8_cuda(MIN_BLOCK_SIZE) for _ in range(11)]
            return small_buffers[5], small_buffers[6]

        self.checkFunction(foo, [])

    def test_additional_free_following_checkpoint(self):
        def foo():
            return (int8_cuda(MIN_BLOCK_SIZE),)

        def foo2():
            return (int8_cuda(MIN_BLOCK_SIZE),)

        graph, outputs = cudagraphify(foo, [])
        pool_id = graph.pool()

        segments_before_checkpoint = get_cudagraph_segments(pool_id)

        state = torch._C._cuda_getCheckpointState(outputs[0].device.index, pool_id)

        graph2, outputs2 = cudagraphify(foo2, [], pool=graph.pool())

        self.setCheckpointPoolState(outputs[0].device.index, state, outputs2, [])

        del outputs2

        self.checkCheckpointedState(
            segments_before_checkpoint, get_cudagraph_segments(pool_id)
        )

    # TODO: re-enable
    # def test_additional_free_error(self):
    #     def foo():
    #         return int8_cuda(MIN_BLOCK_SIZE),

    #     def foo2():
    #         return int8_cuda(MIN_BLOCK_SIZE),

    #     graph, outputs = cudagraphify(foo, [])
    #     pool_id = graph.pool()

    #     segments_before_checkpoint = get_cudagraph_segments(pool_id)

    #     state = torch._C._cuda_getCheckpointState(outputs[0].device.index, pool_id)

    # graph2, outputs2 = cudagraphify(foo2, [], pool=graph.pool())
    # with self.assertRaisesRegex(Exception, "being manually freed must be passed"):
    #     self.setCheckpointPoolState(outputs[0].device.index, state, [], [])

    def test_tensor_dies_after_checkpoint(self):
        def foo():
            return int8_cuda(MIN_BLOCK_SIZE), int8_cuda(MIN_BLOCK_SIZE)

        graph, outputs = cudagraphify(foo, [])
        pool_id = graph.pool()
        device = outputs[0].device.index

        segments_before_checkpoint = get_cudagraph_segments(pool_id)
        state = torch._C._cuda_getCheckpointState(outputs[0].device.index, pool_id)

        output_data_ptrs = [output.data_ptr() for output in outputs]

        del outputs

        self.setCheckpointPoolState(device, state, [], [])

        self.assertEqual(live_blocks(pool_id), 2)
        torch._C._cuda_cudaCachingAllocator_raw_delete(output_data_ptrs[0])
        self.assertEqual(live_blocks(pool_id), 1)
        torch._C._cuda_cudaCachingAllocator_raw_delete(output_data_ptrs[1])
        self.assertEqual(live_blocks(pool_id), 0)

    def test_assigning_back_deleter_fns_to_tensor(self):
        def foo(x):
            return (
                int8_cuda(SMALL_BUFFER) + x,
                int8_cuda(SMALL_BUFFER) + x,
                int8_cuda(LARGE_BUFFER) + x,
            )

        inp = torch.tensor([1], device="cuda")
        graph, outputs = cudagraphify(foo, [inp])
        pool_id = graph.pool()
        graph.replay()

        device = outputs[0].device.index

        for i in range(len(outputs)):
            self.assertEqual(outputs[i].mean(dtype=torch.float), 2)

        state = torch._C._cuda_getCheckpointState(outputs[0].device.index, pool_id)

        output_ptrs = [output.untyped_storage().data_ptr() for output in outputs]
        ten_metadata = [tensor_metadata(t) for t in outputs]

        self.assertEqual(live_blocks(pool_id), 3)

        del outputs

        self.assertEqual(live_blocks(pool_id), 0)

        reconstructed_tensors = [
            reconstruct_from_tensor_metadata(metadata) for metadata in ten_metadata
        ]

        for i in range(len(reconstructed_tensors)):
            self.assertEqual(reconstructed_tensors[i].mean(dtype=torch.float), 2)

        inp.add_(1)
        graph.replay()

        for i in range(len(reconstructed_tensors)):
            self.assertEqual(reconstructed_tensors[i].mean(dtype=torch.float), 3)

        self.setCheckpointPoolState(
            device, state, [], [reconstructed_tensors[0], reconstructed_tensors[1]]
        )

        self.assertEqual(live_blocks(pool_id), 3)

        reconstructed_tensors[0] = None
        self.assertEqual(live_blocks(pool_id), 2)

        reconstructed_tensors[1] = None
        self.assertEqual(live_blocks(pool_id), 1)

        # should not change, we did not pass it in to swap data ptrs
        reconstructed_tensors[2] = None
        self.assertEqual(live_blocks(pool_id), 1)

        torch._C._cuda_cudaCachingAllocator_raw_delete(output_ptrs[2])

        self.assertEqual(live_blocks(pool_id), 0)

    @skipIfNoTorchVision
    def test_resnet(self):
        import torchvision

        m = torchvision.models.resnet50()
        m.eval()
        m = m.cuda()

        inp = torch.rand([1, 3, 255, 255], device="cuda")
        self.checkFunction(m, [inp])

    def test_check_pool_live_allocations(self):
        def foo():
            return torch.ones([4], device="cuda")

        pool = torch.cuda.graph_pool_handle()
        graph, outputs = cudagraphify(foo, [], pool=pool)

        index = outputs[0].device.index

        def check(live_dps):
            return torch._C._cuda_checkPoolLiveAllocations(index, pool, live_dps)

        self.assertTrue(check({outputs[0].data_ptr()}))

        self.assertFalse(check({outputs[0].data_ptr(), 0}))
        self.assertFalse(check(set()))

        del outputs
        self.assertTrue(check(set()))

    def test_allocate_in_thread_to_pool(self):
        def foo():
            return torch.rand([4], device="cuda")

        pool = torch.cuda.graph_pool_handle()
        graph, outputs = cudagraphify(foo, [], pool=pool)
        device = outputs[0].device.index
        del outputs

        @contextlib.contextmanager
        def _use_cuda_memory_pool_manager(device, mem_pool):
            """
            Context manager to use cuda graph pool for new allocations. If you use this manager
            all cudagraph tensors in use should be reflected in the allocator or they will be overwritten.
            existing_graph should already have been used in a capture, and the mem_pool must already exist.
            """
            torch.cuda.synchronize()
            stream = torch.cuda.Stream()
            stream.wait_stream(torch.cuda.current_stream())

            with torch.cuda.stream(stream), torch.device(device):
                torch._C._cuda_beginAllocateCurrentThreadToPool(device, mem_pool)
                try:
                    yield
                finally:
                    torch._C._cuda_endAllocateToPool(device, mem_pool)
                    torch._C._cuda_releasePool(device, mem_pool)

            torch.cuda.current_stream().wait_stream(stream)

        segments = get_cudagraph_segments(pool)
        self.assertEqual(len(get_cudagraph_segments(pool)), 1)

        def use_pool():
            def alloc_three():
                a = int8_cuda(LARGE_BUFFER)
                b = int8_cuda(LARGE_BUFFER)
                c = a + b

            with _use_cuda_memory_pool_manager(device, pool):
                # three allocations
                for _ in range(10):
                    alloc_three()

            # three more allocations not in pool
            alloc_three()

        def no_pool():
            # two allocations
            for _ in range(10):
                a = int8_cuda(LARGE_BUFFER)
                b = int8_cuda(LARGE_BUFFER)
                del a, b

        graph_thread = threading.Thread(target=use_pool)
        no_graph_thread = threading.Thread(target=no_pool)
        graph_thread.start()
        no_graph_thread.start()

        graph_thread.join()
        no_graph_thread.join()

        self.assertEqual(
            len(get_cudagraph_segments(pool)), 2 if self.expandable_segments else 4
        )

        del graph

        torch.cuda.synchronize()
        gc.collect()
        torch.cuda.empty_cache()

        self.assertEqual(len(get_cudagraph_segments(pool)), 0)

    @unittest.skipIf(not EXPANDABLE_SEGMENTS, "requires expandable_segments")
    def test_expandable_segment_checkpoint_growth(self):
        # Test restoring checkpoint when expandable segment has grown.
        # Captures graph1 (small), checkpoints, then captures graph2 (larger)
        # which grows the segment. Restoring to the smaller checkpoint should
        # handle the extra mapped memory correctly.
        pool_id = torch.cuda.graph_pool_handle()
        stream = torch.cuda.Stream()
        device = torch.cuda.current_device()

        def get_large_mapped_blocks():
            # memory_snapshot only includes mapped blocks
            snapshot_segments = torch.cuda.memory_snapshot()
            res = []
            for segment in snapshot_segments:
                if segment["segment_type"] != "large":
                    continue
                blocks = segment["blocks"]
                for block in blocks:
                    res.append((block["size"], block["state"]))
            return res

        with torch.cuda.stream(stream):
            g = torch.cuda.CUDAGraph()
            g.capture_begin(pool=pool_id)
            g.capture_end()
        original_state = torch._C._cuda_getCheckpointState(device, pool_id)

        # 2MB allocation
        def small_fn():
            return (torch.empty(2 * 1024 * 1024, device="cuda", dtype=torch.int8),)

        graph1, out1 = cudagraphify(small_fn, [], pool=pool_id, stream=stream)

        large_mapped_blocks = get_large_mapped_blocks()
        # small_fn allocates 2 MB, which leads to a 20 MB segment in large blocks. This becomes a 2 MB (or 2097152 bytes)
        # active allocated block and a 18 MB (or 18874368 bytes) inactive block.
        self.assertEqual(
            large_mapped_blocks, [(2097152, "active_allocated"), (18874368, "inactive")]
        )

        small_state = torch._C._cuda_getCheckpointState(device, pool_id)
        out1_metadata = [tensor_metadata(t) for t in out1]
        del out1

        self.setCheckpointPoolState(device, original_state, [], [])

        # 32MB allocation
        def large_fn():
            return (torch.empty(32 * 1024 * 1024, device="cuda", dtype=torch.int8),)

        graph2, out2 = cudagraphify(large_fn, [], pool=pool_id, stream=stream)

        large_mapped_blocks = get_large_mapped_blocks()
        # large_fn allocates 32 MB, which leads to a 40 MB segment in large blocks. This becomes a 32 MB (or 33554432 bytes)
        # active allocated block and a 8 MB (or 8388608 bytes) inactive block.
        self.assertEqual(
            large_mapped_blocks, [(33554432, "active_allocated"), (8388608, "inactive")]
        )

        self.setCheckpointPoolState(device, original_state, out2, [])

        graph1.replay()
        reconstructed = [reconstruct_from_tensor_metadata(m) for m in out1_metadata]
        out1_storage = [o.untyped_storage()._cdata for o in reconstructed]
        torch._C._cuda_setCheckpointPoolState(device, small_state, [], out1_storage)

        large_mapped_blocks = get_large_mapped_blocks()
        # private memory pool never decreases size so we still have 40 MB segment. When we set checkpoint pool state for
        # small_state, we allocate a 2 MB (or 2097152 bytes) active block and a 18 MB (or 18874368 bytes) inactive block.
        # Since the total segment size is 40 MB, we still have a 20 MB inactive block. These two inactive blocks
        # automatically merge into a single 38 MB (or 39845888 bytes) inactive block. As a result, we have a 2 MB active
        # block and a 38 MB inactive block in the memory snapshot.
        self.assertEqual(
            large_mapped_blocks, [(2097152, "active_allocated"), (39845888, "inactive")]
        )

    def test_no_triton_on_import(self):
        """Test that Triton is not imported on first GPU use"""
        script = "import sys; import torch; torch.rand(2, device='cuda'); print('triton' in sys.modules)"

        rc = (
            subprocess.check_output(
                [sys.executable, "-c", script],
                # On Windows, opening the subprocess with the default CWD makes `import torch`
                # fail, so just set CWD to this script's directory
                cwd=os.path.dirname(os.path.realpath(__file__)),
            )
            .strip()
            .decode("ascii")
        )
        self.assertEqual(rc, "False", "Triton was imported when importing torch!")


@contextlib.contextmanager
def caching_host_allocator_use_host_register(use_cuda_host_register: bool):
    if use_cuda_host_register:
        torch.cuda.memory._set_allocator_settings(
            "pinned_use_cuda_host_register:True,pinned_num_register_threads:8"
        )
    try:
        yield
    finally:
        if use_cuda_host_register:
            torch.cuda.memory._set_allocator_settings(
                "pinned_use_cuda_host_register:False"
            )


@contextlib.contextmanager
def caching_host_allocator_use_background_threads(use_background_threads: bool):
    if use_background_threads:
        torch.cuda.memory._set_allocator_settings("pinned_use_background_threads:True")
    try:
        yield
    finally:
        if use_background_threads:
            torch.cuda.memory._set_allocator_settings(
                "pinned_use_background_threads:False"
            )


@contextlib.contextmanager
def caching_host_allocator_max_round_threshold_and_max_cached_size(
    max_round_threshold_mb: int | None = None, max_cached_size_mb: int | None = None
):
    allocator_settings = torch._C._cuda_memorySnapshot(None)["allocator_settings"]
    cur_max_round_threshold_mb = (
        allocator_settings["max_round_threshold"] // 1024 // 1024
    )
    cur_max_cached_size_mb = allocator_settings["max_cached_size"] // 1024 // 1024

    parts = []
    if max_round_threshold_mb is not None:
        parts.append(f"pinned_max_round_threshold_mb:{max_round_threshold_mb}")
    if max_cached_size_mb is not None:
        parts.append(f"pinned_max_cached_size_mb:{max_cached_size_mb}")
    torch._C._accelerator_setAllocatorSettings(",".join(parts))
    try:
        yield
    finally:
        torch._C._accelerator_setAllocatorSettings(
            f"pinned_max_round_threshold_mb:{cur_max_round_threshold_mb},pinned_max_cached_size_mb:{cur_max_cached_size_mb}"
        )


@unittest.skipIf(not TEST_CUDA, "CUDA not available, skipping tests")
class TestCachingHostAllocatorConfig(TestCase):
    def setUp(self):
        super().setUp()
        gc.collect()
        torch._C._host_emptyCache()

    def tearDown(self):
        torch._C._host_emptyCache()
        super().tearDown()

    def test_max_round_threshold(self):
        size_bytes = 129 * 1024 * 1024

        with caching_host_allocator_max_round_threshold_and_max_cached_size(128, None):
            t = torch.empty(size_bytes, dtype=torch.uint8, pin_memory=True)
            stats = torch.cuda.host_memory_stats()

            allocated = stats["allocated_bytes.current"]
            self.assertLess(allocated, 256 * 1024 * 1024)
            self.assertGreaterEqual(allocated, size_bytes)

    def test_max_cached_size(self):
        size_bytes = 64 * 1024 * 1024

        with caching_host_allocator_max_round_threshold_and_max_cached_size(None, 32):
            t = torch.empty(size_bytes, dtype=torch.uint8, pin_memory=True)
            del t
            gc.collect()

            stats = torch.cuda.host_memory_stats()
            self.assertEqual(stats["allocations.current"], 0)

    def test_both_thresholds(self):
        size_bytes = 150 * 1024 * 1024

        with caching_host_allocator_max_round_threshold_and_max_cached_size(128, 256):
            t = torch.empty(size_bytes, dtype=torch.uint8, pin_memory=True)
            stats = torch.cuda.host_memory_stats()

            self.assertLess(stats["allocated_bytes.current"], 256 * 1024 * 1024)
            del t
            gc.collect()

            stats = torch.cuda.host_memory_stats()
            self.assertEqual(stats["allocations.current"], 1)

    def test_round_threshold_larger_than_cached_size(self):
        # When round_threshold > cached_size, allocations above the cache
        # limit should be neither rounded up nor cached.
        size_bytes = 100 * 1024 * 1024

        with caching_host_allocator_max_round_threshold_and_max_cached_size(256, 64):
            t = torch.empty(size_bytes, dtype=torch.uint8, pin_memory=True)
            stats = torch.cuda.host_memory_stats()

            # 100 MB is within the 256 MB round threshold but above the
            # 64 MB cache limit — should NOT be rounded to 128 MB.
            allocated = stats["allocated_bytes.current"]
            self.assertGreaterEqual(allocated, size_bytes)
            self.assertLess(allocated, 128 * 1024 * 1024)

            del t
            gc.collect()

            # Block should have been freed, not cached.
            stats = torch.cuda.host_memory_stats()
            self.assertEqual(stats["allocations.current"], 0)


@unittest.skipIf(not TEST_CUDA, "CUDA not available, skipping tests")
class TestCachingHostAllocatorCudaGraph(TestCase):
    # As soon as pinned host memory allocated by a private pool is
    # used (by copy_ in this case) during stream capture, it can never
    # be recycled.
    @parametrize("use_cuda_host_register", [True, False])
    def test_pin_memory_no_use(self, use_cuda_host_register):
        # A pinned host memory block cannot be reused if it is not deleted
        with caching_host_allocator_use_host_register(use_cuda_host_register):
            graph = torch.cuda.CUDAGraph()
            with torch.cuda.graph(graph, capture_error_mode="thread_local"):
                data = torch.empty(8, pin_memory=True)
                data2 = torch.empty(8, pin_memory=True)
            if data.data_ptr() == data2.data_ptr():
                raise AssertionError("data and data2 should have different data_ptr")
            del data2

    @parametrize("use_cuda_host_register", [True, False])
    def test_pin_memory_no_use2(self, use_cuda_host_register):
        # A pinned host memory block can be reused if it is deleted
        # and has never been used by copy_
        with caching_host_allocator_use_host_register(use_cuda_host_register):
            graph = torch.cuda.CUDAGraph()
            with torch.cuda.graph(graph, capture_error_mode="thread_local"):
                data = torch.randn(8).pin_memory()
                data_ptr = data.data_ptr()
                del data
                data2 = torch.randn(8).pin_memory()
                if data2.data_ptr() != data_ptr:
                    raise AssertionError(
                        "data2 should have same data_ptr as deleted data"
                    )

    @parametrize("use_cuda_host_register", [True, False])
    def test_pin_memory_use(self, use_cuda_host_register):
        # A pinned host memory block cannot be reused if it has been used by copy_
        with caching_host_allocator_use_host_register(use_cuda_host_register):
            graph = torch.cuda.CUDAGraph()
            with torch.cuda.graph(graph, capture_error_mode="thread_local"):
                data = torch.randn(8).pin_memory()
                data_gpu = torch.randn(8, device="cuda")
                data_gpu.copy_(data, non_blocking=True)
                old_data_ptr = data.data_ptr()
                del data
                data2 = torch.randn(8).pin_memory()
            if data2.data_ptr() == old_data_ptr:
                raise AssertionError(
                    "data2 should have different data_ptr than old_data_ptr"
                )

    @parametrize("use_cuda_host_register", [True, False])
    @parametrize("use_background_threads", [True, False])
    @parametrize(
        "use_memory, delete_memory",
        [(True, True), (True, False), (False, True), (False, False)],
    )
    @skip_background_threads_on_windows
    def test_two_graphs(
        self, use_background_threads, use_cuda_host_register, use_memory, delete_memory
    ):
        # Pinned host memory belongs to private pools, not to
        # graphs. Therefore, using two graphs that share a pool
        # instead of one graph does not change any of our invariants.
        with (
            caching_host_allocator_use_background_threads(use_background_threads),
            caching_host_allocator_use_host_register(use_cuda_host_register),
        ):
            shared_pool = torch.cuda.graph_pool_handle()
            graph1 = torch.cuda.CUDAGraph()
            graph2 = torch.cuda.CUDAGraph()

            with torch.cuda.graph(
                graph1, pool=shared_pool, capture_error_mode="thread_local"
            ):
                data = torch.randn(8).pin_memory()
                if use_memory:
                    data_gpu = torch.randn(8, device="cuda")
                    data_gpu.copy_(data, non_blocking=True)

                old_data_ptr = data.data_ptr()
                if delete_memory:
                    del data

            with torch.cuda.graph(
                graph2, pool=shared_pool, capture_error_mode="thread_local"
            ):
                data2 = torch.randn(8).pin_memory()
                if use_memory:
                    data_gpu = torch.randn(8, device="cuda")
                    data_gpu.copy_(data2, non_blocking=True)

                new_data_ptr = data2.data_ptr()
                if delete_memory:
                    del data2

            if delete_memory and not use_memory:
                if new_data_ptr != old_data_ptr:
                    raise AssertionError("new_data_ptr should equal old_data_ptr")
            else:
                if new_data_ptr == old_data_ptr:
                    raise AssertionError("new_data_ptr should differ from old_data_ptr")

    def test_unpinned_memory_use(self):
        # Copying between CPU and CUDA tensors during graph capture
        # with unpinned CPU memory should raise an error.
        graph = torch.cuda.CUDAGraph()
        with self.assertRaisesRegex(
            RuntimeError,
            "Cannot copy between CPU and CUDA tensors during CUDA graph capture",
        ):
            with torch.cuda.graph(graph):
                data = torch.empty(8)
                data_gpu = torch.randn(8, device="cuda")
                data_gpu.copy_(data, non_blocking=True)

    def test_unpinned_memory_use_device_to_host(self):
        # Copying from CUDA to unpinned CPU during graph capture should also error.
        graph = torch.cuda.CUDAGraph()
        with self.assertRaisesRegex(
            RuntimeError,
            "Cannot copy between CPU and CUDA tensors during CUDA graph capture",
        ):
            with torch.cuda.graph(graph):
                data_gpu = torch.randn(8, device="cuda")
                data = torch.empty(8)
                data.copy_(data_gpu, non_blocking=True)


@unittest.skipIf(not TEST_CUDA, "CUDA not available, skipping tests")
class TestMemPool(TestCase):
    def _setup_mempool_limited_memory_test(self, additional_allowed_memory_in_mb):
        device = torch.device("cuda:0")

        self.init_fraction = torch.cuda.get_per_process_memory_fraction()
        torch.cuda.memory.empty_cache()
        mb = 1024 * 1024
        _, all_memory = torch.cuda.memory.mem_get_info(device)
        pre_reserved = torch.cuda.memory_reserved(device)
        total_allowed = additional_allowed_memory_in_mb * mb + pre_reserved
        fraction_allowed = total_allowed / all_memory
        torch.cuda.memory.set_per_process_memory_fraction(fraction_allowed, device)

        dtype = torch.int8
        return device, dtype

    def _teardown_mempool_limited_memory_test(self):
        torch.cuda.memory.empty_cache()
        torch.cuda.memory.set_per_process_memory_fraction(self.init_fraction)

    def test_mempool_id(self):
        pool1 = torch.cuda.graph_pool_handle()
        pool2 = torch.cuda.MemPool().id

        # first value of id in a user created pool is always zero
        self.assertEqual(pool1[0] == 0, pool2[0] == 0)

        # each call to torch.cuda.graph_pool_handle() or torch.cuda.MemPool()
        # increments the id
        self.assertTrue(abs(pool2[1] - pool1[1]) > 0)

    @unittest.skipIf(
        TEST_CUDAMALLOCASYNC, "setContextRecorder not supported by CUDAMallocAsync"
    )
    def test_pool_id_in_snapshot(self):
        try:
            torch.cuda.memory.empty_cache()
            torch.cuda.memory._record_memory_history("all")

            pool = torch.cuda.MemPool()
            with torch.cuda.use_mem_pool(pool):
                x = torch.rand(64, device="cuda")

            ss = torch.cuda.memory._snapshot()

            # segment_pool_id should match the MemPool id
            found_segment = False
            for seg in ss["segments"]:
                if seg["segment_pool_id"] == pool.id:
                    found_segment = True
                    break
            self.assertTrue(found_segment)

            # trace entries for this allocation should carry pool_id
            found_trace = False
            for trace in ss["device_traces"]:
                for te in trace:
                    if "pool_id" not in te:
                        continue
                    if te["pool_id"] == pool.id and te["action"] == "alloc":
                        found_trace = True
                        break
            self.assertTrue(found_trace)

            del x
        finally:
            torch.cuda.memory._record_memory_history(None)

    def get_dummy_allocator(self, check_vars):
        dummy_allocator_source_vars = """
        #include <torch/extension.h>
        #include <ATen/cuda/Exceptions.h>
        #include <cuda_runtime_api.h>

        extern "C" {
            C10_EXPORT int called_dummy_alloc = 0;
            C10_EXPORT int called_dummy_free = 0;

            // Note that windows needs __declspec(dllexport): https://stackoverflow.com/a/24575865
            C10_EXPORT void* dummy_alloc(size_t size, int device, void* stream) {
            called_dummy_alloc = 123;
            void* ptr;
            C10_CUDA_CHECK(cudaMallocManaged(&ptr, size));
            return ptr;
            }

            C10_EXPORT void dummy_free(void* ptr, size_t size, int device, void* stream) {
            called_dummy_free = 321;
            C10_CUDA_CHECK(cudaFree(ptr));
            }
        }
        """
        dummy_allocator_source_no_vars = """
        #include <torch/extension.h>
        #include <ATen/cuda/Exceptions.h>
        #include <cuda_runtime_api.h>

        extern "C" {
          // Note that windows needs __declspec(dllexport): https://stackoverflow.com/a/24575865
          C10_EXPORT void* dummy_alloc(size_t size, int device, void* stream) {
            void* ptr;
            C10_CUDA_CHECK(cudaMallocManaged(&ptr, size));
            return ptr;
          }

          C10_EXPORT void dummy_free(void* ptr, size_t size, int device, void* stream) {
            C10_CUDA_CHECK(cudaFree(ptr));
          }
        }
        """

        from torch.utils.cpp_extension import load_inline

        dummy_allocator_libname = "dummy_allocator"
        dummy_allocator = load_inline(
            name=dummy_allocator_libname,
            cpp_sources=dummy_allocator_source_vars
            if check_vars
            else dummy_allocator_source_no_vars,
            is_python_module=False,
            keep_intermediates=False,
            verbose=True,
            with_cuda=True,
        )
        allocator = torch.cuda.memory.CUDAPluggableAllocator(
            dummy_allocator,
            "dummy_alloc",
            "dummy_free",
        )
        return allocator, dummy_allocator

    def test_mempool_empty_cache(self):
        torch.cuda.empty_cache()
        pool = torch.cuda.MemPool()
        x = torch.empty(1024, 1024, device="cuda")

        with torch.cuda.use_mem_pool(pool):
            y = torch.empty(1024, 1024, device="cuda")

        del y
        del x
        del pool
        segments = torch.cuda.memory._snapshot()["segments"]
        self.assertTrue(len(segments) > 0, "expected more than one segment")

    @serialTest()
    def test_mempool_empty_cache_inactive(self):
        torch.cuda.empty_cache()
        allocator, dummy_allocator = self.get_dummy_allocator(check_vars=True)
        alloc_lib = ctypes.CDLL(dummy_allocator)
        called_dummy_alloc = ctypes.c_int.in_dll(alloc_lib, "called_dummy_alloc")
        called_dummy_free = ctypes.c_int.in_dll(alloc_lib, "called_dummy_free")
        self.assertEqual(called_dummy_alloc.value, 0)
        self.assertEqual(called_dummy_free.value, 0)

        def f():
            pool = torch.cuda.MemPool(allocator.allocator())

            # allocate memory with ncclMemAlloc
            with torch.cuda.use_mem_pool(pool):
                x = torch.arange(1024 * 1024 * 2, device="cuda")
            # Note: pool will be destroyed upon function return, but x, which
            # was allocated via the pool is still alive.
            return x

        x = f()
        self.assertEqual(called_dummy_alloc.value, 123)
        self.assertEqual(called_dummy_free.value, 0)

        del x
        torch.cuda.empty_cache()
        self.assertEqual(called_dummy_free.value, 321)

        # reset dummy allocator indicator variables because they are global
        called_dummy_free.value = 0
        called_dummy_alloc.value = 0

    @serialTest()
    def test_mempool_with_allocator(self):
        pool = torch.cuda.MemPool()

        allocator, dummy_allocator = self.get_dummy_allocator(check_vars=True)

        pool = torch.cuda.MemPool(allocator.allocator())

        # pool's use count should be 1 at this point as MemPool object
        # holds a reference
        self.assertEqual(pool.use_count(), 1)

        # no allocations happened yet, so called_dummy_alloc and
        # called_dummy_free should be 0
        alloc_lib = ctypes.CDLL(dummy_allocator)
        called_dummy_alloc = ctypes.c_int.in_dll(alloc_lib, "called_dummy_alloc")
        called_dummy_free = ctypes.c_int.in_dll(alloc_lib, "called_dummy_free")
        self.assertEqual(called_dummy_alloc.value, 0)
        self.assertEqual(called_dummy_free.value, 0)

        nelem_1mb = 1024 * 1024 // 4

        with torch.cuda.use_mem_pool(pool):
            out_0 = torch.randn(nelem_1mb, device="cuda")

            # pool's use count should be 2 at this point as use_mem_pool
            # holds a reference
            self.assertEqual(pool.use_count(), 2)

        # pool's use count should be back to 1 at this point as use_mem_pool
        # released its reference
        self.assertEqual(pool.use_count(), 1)

        # called_dummy_alloc should be 123 if dummy_alloc was used to allocate
        # out tensor
        self.assertEqual(called_dummy_alloc.value, 123)

        out_non_pool = torch.empty(nelem_1mb, device="cuda")

        with torch.cuda.use_mem_pool(pool):
            # pool should have 1 segment since we made a small allocation (1 MB)
            # above and so the CUDACachingAllocator packed it into a 2 MB buffer
            self.assertEqual(len(pool.snapshot()), 1)

            out_1 = torch.randn(nelem_1mb, device="cuda")

            # pool should still have 1 segment since we made another small allocation
            # (1 MB) that got packed into the existing 2 MB buffer
            self.assertEqual(len(pool.snapshot()), 1)

            out_2 = torch.randn(nelem_1mb, device="cuda")

            # pool now should have 2 segments since the CUDACachingAllocator had
            # to make a new 2 MB buffer to accommodate out_2
            self.assertEqual(len(pool.snapshot()), 2)

        self.assertEqual(len(pool.snapshot()), 2)

        del out_0, out_1, out_2

        # pool's destructor calls emptyCache()
        del pool

        # called_dummy_free should be 321 if dummy_free was used to deallocate
        # out tensor
        self.assertEqual(called_dummy_free.value, 321)

        # reset dummy allocator indicator variables because they are global
        called_dummy_free.value = 0
        called_dummy_alloc.value = 0

    @serialTest()
    def test_tensor_delete_after_allocator_delete(self):
        allocator, dummy_allocator = self.get_dummy_allocator(check_vars=True)
        pool = torch.cuda.MemPool(allocator.allocator())

        with torch.cuda.use_mem_pool(pool):
            data = torch.empty(4, device="cuda")

        alloc_lib = ctypes.CDLL(dummy_allocator)
        called_dummy_alloc = ctypes.c_int.in_dll(alloc_lib, "called_dummy_alloc")
        called_dummy_free = ctypes.c_int.in_dll(alloc_lib, "called_dummy_free")

        self.assertEqual(called_dummy_alloc.value, 123)
        self.assertEqual(called_dummy_free.value, 0)

        # Explicitly drop the allocator before releasing the
        # tensor. CUDACachingAllocator's PrivatePool instance should
        # keep the allocator live until all tensors die.

        # N.B.: Deleting the pool doesn't actually do anything, since
        # it doesn't own the allocator object. But we do it anyway
        # because it is the situation that users are likely to face.
        del pool
        del allocator

        self.assertEqual(called_dummy_free.value, 0)

        del data
        torch.cuda.empty_cache()

        self.assertEqual(called_dummy_free.value, 321)

        # reset dummy allocator indicator variables because they are global
        called_dummy_free.value = 0
        called_dummy_alloc.value = 0

    @serialTest()
    def test_mempool_limited_memory_with_allocator(self):
        allocator, _ = self.get_dummy_allocator(check_vars=False)
        pool_do_not_use = torch.cuda.MemPool(allocator.allocator())
        pool_use = torch.cuda.MemPool(allocator.allocator(), use_on_oom=True)

        nelem_1mb = 1024 * 1024 // 4

        self._setup_mempool_limited_memory_test(80)
        try:
            # remaining free mem: 80 mb
            # mempool_use [] 0 mb
            # mempool_do_not_use [] 0 mb
            # default pool [] 0 mb
            with torch.cuda.use_mem_pool(pool_do_not_use):
                a = torch.randn(40 * nelem_1mb, device="cuda")
            with torch.cuda.use_mem_pool(pool_use):
                b = torch.randn(40 * nelem_1mb, device="cuda")
            a_dataptr = a.data_ptr()
            b_dataptr = b.data_ptr()
            # remaining free mem: 0 mb
            # mempool_do_not_use [aaaa] 40 mb
            # mempool_use [bbbb] 40 mb
            # default pool [] 0 mb
            with self.assertRaises(torch.OutOfMemoryError):
                # out of memory
                c = torch.randn(40 * nelem_1mb, device="cuda")

            del a, b
            # remaining free mem: 0 mb
            # mempool_do_not_use [____] 40 mb
            # mempool_use [____] 40 mb
            # default pool [] 0 mb

            # c should not oom and instead can use mempool_use as fallback
            c = torch.randn(30 * nelem_1mb, device="cuda")
            c_dataptr = c.data_ptr()
            # remaining free mem: 0 mb
            # mempool_do_not_use [____] 40 mb
            # mempool_use [ccc_] 40 mb
            # default pool [] 0 mb
            with self.assertRaises(torch.OutOfMemoryError):
                # out of memory since can't use mempool_do_not_use
                d = torch.randn(30 * nelem_1mb, device="cuda")

            del c
            # remaining free mem: 0 mb
            # mempool_do_not_use [____] 40 mb
            # mempool_use [____] 40 mb
            # default pool [] 0 mb

            # expect that we used same memory address for both a and c
            self.assertEqual(b_dataptr, c_dataptr)

            # make sure we can still use mempool_use as intended after c is deleted
            with torch.cuda.use_mem_pool(pool_use):
                e = torch.randn(20 * nelem_1mb, device="cuda")
            # remaining free mem: 0 mb
            # mempool_do_not_use [____] 40 mb
            # mempool_use [ee__] 40 mb
            # default pool [] 0 mb

            e_dataptr = e.data_ptr()
            del e

            self.assertEqual(e_dataptr, c_dataptr)

            # pool's destructor calls emptyCache()
            del pool_use, pool_do_not_use
        finally:
            self._teardown_mempool_limited_memory_test()

    @serialTest()
    def test_mempool_release_cached_blocks_during_diversion(self):
        # Regression test for the bug where allocations failing inside
        # `torch.cuda.use_mem_pool(...)` would skip the OOM-time
        # release_cached_blocks() retry, because the gate read
        # `captures_underway.empty()` — which is non-empty during any
        # private-pool diversion, even when no real cudaStreamBeginCapture
        # is active. The fix uses a separate `num_active_captures_` counter
        # tracked by CUDAGraph::capture_begin/end, so default-pool cached
        # blocks can still be reclaimed during plain mempool diversion.
        nelem_1mb = 1024 * 1024
        device, dtype = self._setup_mempool_limited_memory_test(80)
        try:
            # Reserve 60 MB on the default pool, then free → cached but not
            # returned to the driver. memory_reserved stays at 60 MB.
            a = torch.empty(60 * nelem_1mb, device=device, dtype=dtype)
            del a

            pool = torch.cuda.MemPool()
            with torch.cuda.use_mem_pool(pool):
                # 60 MB into the private pool. Reservation budget remaining
                # is only 20 MB (80 cap − 60 cached), so cudaMalloc must
                # fail. The OOM retry needs to call release_cached_blocks()
                # on the default pool to free the 60 MB and retry. Pre-fix,
                # that path was gated off and this would raise.
                b = torch.empty(60 * nelem_1mb, device=device, dtype=dtype)
            self.assertEqual(b.numel(), 60 * nelem_1mb)
            del b
            del pool
        finally:
            self._teardown_mempool_limited_memory_test()

    @serialTest()
    def test_mempool_no_split(self):
        torch.cuda.empty_cache()

        # Create two pools - one with no_split=True, one without
        pool_split = torch.cuda.MemPool()  # default: no_split=False
        pool_no_split = torch.cuda.MemPool(no_split=True)

        # 1 MB in number of float32 elements
        nelem_1mb = 1024 * 1024 // 4

        # Allocate 4 MB in each pool - this should allocate a 20 MB segment
        with torch.cuda.use_mem_pool(pool_split):
            a_split = torch.randn(4 * nelem_1mb, device="cuda")
        with torch.cuda.use_mem_pool(pool_no_split):
            a_no_split = torch.randn(4 * nelem_1mb, device="cuda")

        # Now allocate another 4 MB - this should cause splitting in the normal
        # pool but not in the no_split pool. Instead, the no_split pool should
        # have a new segment.
        with torch.cuda.use_mem_pool(pool_split):
            b_split = torch.randn(4 * nelem_1mb, device="cuda")
        with torch.cuda.use_mem_pool(pool_no_split):
            b_no_split = torch.randn(4 * nelem_1mb, device="cuda")

        # assert the number of segments in the no_split pool is larger than that
        # of the split pool
        if len(pool_no_split.snapshot()) <= len(pool_split.snapshot()):
            raise AssertionError(
                f"Expected no_split pool to have more segments, "
                f"but got {len(pool_no_split.snapshot())} vs {len(pool_split.snapshot())}"
            )

        # Specifically, the no_split pool should have exactly 1 block per segment
        for seg in pool_no_split.snapshot():
            if len(seg["blocks"]) != 1:
                raise AssertionError(
                    f"Expected 1 block in no_split segment, got {len(seg['blocks'])}"
                )

        # Count blocks in each pool
        def count_blocks(pool):
            total = 0
            for seg in pool.snapshot():
                total += len(seg["blocks"])
            return total

        blocks_split = count_blocks(pool_split)
        blocks_no_split = count_blocks(pool_no_split)

        # The pool with no_split should have fewer blocks because it doesn't
        # have a remaining block representing the remaining memory in the segment.
        self.assertLess(
            blocks_no_split,
            blocks_split,
            f"Expected no_split pool to have fewer blocks, "
            f"but got {blocks_no_split} vs {blocks_split}",
        )

    @skipIfRocmArch(MI200_ARCH)
    @serialTest()
    def test_deleted_mempool_not_used_on_oom(self):
        """
        Test that a deleted mempool with use_on_oom=True is properly removed from use_on_oom_pools.
        """
        allocator, _ = self.get_dummy_allocator(check_vars=False)

        nelem_1mb = 1024 * 1024 // 4

        # set 40 mb total available memory
        self._setup_mempool_limited_memory_test(40)
        try:
            # Create many pools with use_on_oom=True, allocate memory, then delete the pools
            for _ in range(10):
                pool_use_on_oom = torch.cuda.MemPool(
                    allocator.allocator(), use_on_oom=True
                )
                with torch.cuda.use_mem_pool(pool_use_on_oom):
                    a = torch.randn(40 * nelem_1mb, device="cuda")
                del a
                del pool_use_on_oom

            # create new pool that we want to use_on_oom, all other pools should be deleted
            # all available 40mb in use by mempool
            new_pool_use_on_oom = torch.cuda.MemPool(
                allocator.allocator(), use_on_oom=True
            )
            with torch.cuda.use_mem_pool(new_pool_use_on_oom):
                a = torch.randn(40 * nelem_1mb, device="cuda")
            del a

            # allocate tensors that will fallback to use_on_oom pool since all available 40mb in use by mempool
            # tensors should only use valid pool and not deleted pools
            b = torch.randn(20 * nelem_1mb, device="cuda")
            c = torch.randn(20 * nelem_1mb, device="cuda")

            del b
            del c
            del new_pool_use_on_oom
        finally:
            self._teardown_mempool_limited_memory_test()

    def test_mempool_multithread(self):
        pool_ids = []

        def create_mempool_and_make_active():
            pool = torch.cuda.MemPool()
            pool_ids.extend([pool.id])

        num_threads = 4
        threads = [
            threading.Thread(target=create_mempool_and_make_active)
            for t in range(num_threads)
        ]
        for thread in threads:
            thread.start()
        for thread in threads:
            thread.join()

        # each thread should create a unique mempool, since
        # mempool id creation is atomic
        self.assertEqual(len(set(pool_ids)), 4)

    def test_mempool_emptycache_multithread(self):
        num_threads = 4

        def my_function(pool):
            with torch.cuda.use_mem_pool(pool):
                x = torch.randn(4, device="cuda")
                del x
                torch.cuda.empty_cache()

        pools = [torch.cuda.MemPool() for _ in range(num_threads)]
        threads = [
            threading.Thread(target=my_function, args=(pools[i],))
            for i in range(num_threads)
        ]

        for t in threads:
            t.start()
        for t in threads:
            t.join()

        # empty_cache should have done nothing under mempool context
        for p in pools:
            s = p.snapshot()
            self.assertEqual(len(s), 1, "Expected to have a single segment")

    @serialTest()
    def test_nested_mempool(self):
        torch.cuda.empty_cache()
        pool1 = torch.cuda.MemPool()
        pool2 = torch.cuda.MemPool()
        pool3 = torch.cuda.MemPool()

        data = []
        nelem_1mb = 1024 * 1024 // 4

        def allocate_data():
            x = torch.empty(nelem_1mb * 20, device="cuda")
            data.append(x)

        with torch.cuda.use_mem_pool(pool1):
            allocate_data()
            with torch.cuda.use_mem_pool(pool2):
                allocate_data()
                with torch.cuda.use_mem_pool(pool3):
                    allocate_data()
                allocate_data()
            allocate_data()

        pool1_segments = torch.cuda.memory.memory_snapshot(pool1.id)
        pool2_segments = torch.cuda.memory.memory_snapshot(pool2.id)
        pool3_segments = torch.cuda.memory.memory_snapshot(pool3.id)

        self.assertEqual(len(pool1_segments), 2)
        self.assertEqual(len(pool2_segments), 2)
        self.assertEqual(len(pool3_segments), 1)

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_capture_reclaim_2_streams(self):
        torch.cuda.memory._set_allocator_settings(
            "graph_capture_record_stream_reuse:True"
        )
        torch.cuda.empty_cache()

        s1, s2 = torch.cuda.Stream(), torch.cuda.Stream()
        g = torch.cuda.CUDAGraph(keep_graph=True)

        torch.cuda.synchronize()

        with torch.cuda.stream(s1):
            g.capture_begin()

            # A sink node allocated up-front so it doesn't steal data1's block later.
            sink1 = torch.empty(8, device="cuda")

            # Source tensor on s1; this block is the reuse candidate.
            data1 = torch.empty(8, device="cuda")
            data1_ptr = data1.data_ptr()

            # Fork: do real work on s2 that READS data1 and writes to its own buffer.
            s2.wait_stream(s1)
            with torch.cuda.stream(s2):
                buf2 = torch.empty_like(data1)
                torch.add(data1, 2.0, out=buf2)
                data1.record_stream(s2)

            del data1

            # BEFORE JOIN: must NOT reuse
            data2 = torch.empty(8, device="cuda")
            data2_ptr = data2.data_ptr()

            # Join s2 -> s1 and add a sink node on s1.
            s1.wait_stream(s2)
            sink1.fill_(1.0)

            # AFTER JOIN: now reuse is allowed
            data3 = torch.empty(8, device="cuda")
            data3_ptr = data3.data_ptr()

            g.capture_end()

        torch.cuda.synchronize()

        # No reuse before join; reuse after join.
        self.assertNotEqual(data1_ptr, data2_ptr)
        self.assertEqual(data1_ptr, data3_ptr)

        torch.cuda.memory._set_allocator_settings(
            "graph_capture_record_stream_reuse:False"
        )

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_capture_reclaim_4_streams(self):
        torch.cuda.memory._set_allocator_settings(
            "graph_capture_record_stream_reuse:True"
        )

        torch.cuda.empty_cache()
        s1, s2, s3, s4 = (
            torch.cuda.Stream(),
            torch.cuda.Stream(),
            torch.cuda.Stream(),
            torch.cuda.Stream(),
        )
        g = torch.cuda.CUDAGraph(keep_graph=True)

        torch.cuda.synchronize()

        with torch.cuda.stream(s1):
            g.capture_begin()

            # Source tensor allocated on s1. This block is the candidate for reuse.
            data1 = torch.ones(8, device="cuda")
            data1_ptr = data1.data_ptr()
            sink1 = torch.empty_like(data1)
            sink3 = torch.empty_like(data1)

            s2.wait_stream(s1)
            with torch.cuda.stream(s2):
                buf2 = torch.empty_like(data1)
                torch.add(data1, 2.0, out=buf2)
                data1.record_stream(s2)

            s3.wait_stream(s1)
            with torch.cuda.stream(s3):
                buf3 = torch.empty_like(data1)
                torch.add(data1, 3.0, out=buf3)
                data1.record_stream(s3)

            s4.wait_stream(s1)
            with torch.cuda.stream(s4):
                buf4 = torch.empty_like(data1)
                torch.add(data1, 4.0, out=buf4)
                data1.record_stream(s4)

            # Free data1 inside capture; allocator may reuse later when it's safe.
            del data1

            # PARTIAL JOINS: should NOT allow reuse yet
            # Join s2 -> s1 and add a sink node on s1.
            s1.wait_stream(s2)
            sink1.fill_(1.0)

            # Join s4 -> s3 and add a sink node on s3.
            s3.wait_stream(s4)
            with torch.cuda.stream(s3):
                sink3.fill_(3.0)
                sink3.record_stream(s3)

            # At this point, s1 and s3 subgraphs are NOT yet joined together.
            # Allocating data2 here must NOT reuse data1's block.
            data2 = torch.empty(8, device="cuda")
            data2_ptr = data2.data_ptr()

            # FINAL JOIN: now reuse is allowed
            # Join s3 -> s1 and add a sink node on s1.
            s1.wait_stream(s3)
            sink1.add_(sink3)

            # Now allocator should safely reuse data1's block.
            data3 = torch.empty(8, device="cuda")
            data3_ptr = data3.data_ptr()

            g.capture_end()

        torch.cuda.synchronize()

        # No reuse before full join; reuse after full join.
        self.assertNotEqual(data1_ptr, data2_ptr)
        self.assertEqual(data1_ptr, data3_ptr)

        torch.cuda.memory._set_allocator_settings(
            "graph_capture_record_stream_reuse:False"
        )

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    def test_graph_capture_reclaim_shared_pool(self):
        torch.cuda.memory._set_allocator_settings(
            "graph_capture_record_stream_reuse:True"
        )
        torch.cuda.empty_cache()
        torch.cuda.synchronize()

        shared_pool = torch.cuda.graph_pool_handle()
        cap_stream = torch.cuda.Stream()
        side_stream = torch.cuda.Stream()

        g1 = torch.cuda.CUDAGraph()
        g2 = torch.cuda.CUDAGraph()

        numel = (8 * 1024 * 1024) // 4

        with torch.cuda.stream(cap_stream):
            g1.capture_begin(pool=shared_pool)
            data = torch.empty(numel, device="cuda")
            data_ptr = data.data_ptr()

            side_stream.wait_stream(cap_stream)
            with torch.cuda.stream(side_stream):
                data.add_(1.0)
                data.record_stream(side_stream)

            cap_stream.wait_stream(side_stream)

            del data
            g1.capture_end()

        torch.cuda.current_stream().wait_stream(cap_stream)
        torch.cuda.synchronize()

        with torch.cuda.stream(cap_stream):
            g2.capture_begin(pool=shared_pool)
            data2 = torch.empty(numel, device="cuda")
            data2.fill_(42.0)
            data2_ptr = data2.data_ptr()
            g2.capture_end()

        torch.cuda.current_stream().wait_stream(cap_stream)
        torch.cuda.synchronize()

        self.assertEqual(data_ptr, data2_ptr)

        torch.cuda.memory._set_allocator_settings(
            "graph_capture_record_stream_reuse:False"
        )

    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    @unittest.skipIf(TEST_WITH_ROCM, "ROCM does not support nvrtc")
    @unittest.skipIf(
        not SM70OrLater, "Compute capability >= SM70 required for relaxed ptx flag"
    )
    def test_graph_capture_pre_capture_stream_use(self):
        # Tests that a block with pre-capture stream uses is correctly handled
        # when freed during a subsequent capture on the same pool.
        # Exercises the insert_events path in endAllocateToPool.
        spin_wait_kernel = get_wait_for_cpu_kernel()

        torch.cuda.memory._set_allocator_settings(
            "graph_capture_record_stream_reuse:True"
        )
        torch.cuda.empty_cache()
        torch.cuda.synchronize()

        shared_pool = torch.cuda.graph_pool_handle()
        cap_stream = torch.cuda.Stream()
        side_stream = torch.cuda.Stream()
        flag_cpu = torch.zeros(1, dtype=torch.int32, device="cpu").pin_memory()

        g1 = torch.cuda.CUDAGraph()
        g2 = torch.cuda.CUDAGraph()
        g3 = torch.cuda.CUDAGraph()
        g4 = torch.cuda.CUDAGraph()

        numel = (8 * 1024 * 1024) // 4

        # First capture: allocate data in the shared pool, keep it alive.
        with torch.cuda.stream(cap_stream):
            g1.capture_begin(pool=shared_pool)
            data = torch.empty(numel, device="cuda")
            data_ptr = data.data_ptr()
            g1.capture_end()

        torch.cuda.synchronize()

        # Between captures: block side_stream with a spin-wait kernel
        # (pre-capture stream use). The kernel holds the stream busy until
        # we explicitly set the flag from the host.
        with torch.cuda.stream(side_stream):
            spin_wait_kernel(grid=(1, 1, 1), block=(1, 1, 1), args=[flag_cpu])
            data.record_stream(side_stream)

        # Second capture: free data during capture.
        with torch.cuda.stream(cap_stream):
            g2.capture_begin(pool=shared_pool)
            del data
            g2.capture_end()

        # Trigger process_events. The spin kernel is still holding side_stream,
        # so cudaEventQuery returns cudaErrorNotReady and the block stays pending.
        torch.empty(1, device="cuda")

        # Allocate from the same pool: block must NOT be reused yet.
        with torch.cuda.stream(cap_stream):
            g3.capture_begin(pool=shared_pool)
            not_reused = torch.empty(numel, device="cuda")
            not_reused_ptr = not_reused.data_ptr()
            g3.capture_end()

        self.assertNotEqual(data_ptr, not_reused_ptr)

        # Release the spin kernel so side_stream can finish.
        flag_cpu[0] = 1
        torch.cuda.synchronize()

        # Trigger process_events to reclaim the block.
        torch.empty(1, device="cuda")

        # Fourth capture: the block should now be reusable.
        with torch.cuda.stream(cap_stream):
            g4.capture_begin(pool=shared_pool)
            reused = torch.empty(numel, device="cuda")
            reused_ptr = reused.data_ptr()
            g4.capture_end()

        self.assertEqual(data_ptr, reused_ptr)

        torch.cuda.memory._set_allocator_settings(
            "graph_capture_record_stream_reuse:False"
        )

    @unittest.skipIf(
        not EXPANDABLE_SEGMENTS,
        "requires expandable_segments mode (run via test_cuda_expandable_segments.py)",
    )
    # expandable_segments not supported (PYTORCH_C10_DRIVER_API_SUPPORTED not defined for windows builds)
    @unittest.skipIf(
        IS_WINDOWS and SM89OrLater,
        "expandable_segments not supported (PYTORCH_C10_DRIVER_API_SUPPORTED not defined for windows builds)",
    )
    @skipIfRocm(msg="expandable_segments mode is not supported on ROCm")
    @unittest.skipIf(IS_FBCODE or IS_SANDCASTLE, "Load_inline doesn't work in fbcode")
    def test_mempool_expandable(self):
        torch.cuda.empty_cache()
        allocator, _ = self.get_dummy_allocator(check_vars=False)
        pool = torch.cuda.MemPool(allocator.allocator())

        data = []
        nelem = 1024 * 1024 // 4
        with torch.cuda.use_mem_pool(pool):
            data.append(torch.empty(nelem, device="cuda"))

        # the second allocation should be in expandable segment
        data.append(torch.empty(nelem, device="cuda"))

        segments = torch.cuda.memory.memory_snapshot()

        num_expandable_segments = 0
        for segment in segments:
            if segment["is_expandable"]:
                num_expandable_segments += 1

        self.assertEqual(len(segments), 2, "Expected to have 2 segment")
        self.assertEqual(
            num_expandable_segments, 1, "Expected to have 1 expandable segment only"
        )

    @serialTest()
    def test_mempool_ctx_multithread(self):
        torch.cuda.empty_cache()
        segments = torch.cuda.memory._snapshot()["segments"]
        self.assertEqual(len(segments), 0, "Expected empty pool in the beginning")

        nelem = 1024 * 1024
        trigger_alloc = threading.Event()
        done_allocation = threading.Event()

        def main_thread_fn():
            pool = torch.cuda.MemPool()
            out1 = torch.empty(nelem, dtype=torch.int8, device="cuda")
            with torch.cuda.use_mem_pool(pool):
                out = torch.empty(nelem, dtype=torch.int8, device="cuda")
                del out
                trigger_alloc.set()
                done_allocation.wait()

        def side_thread_fn(segments):
            trigger_alloc.wait()
            out = torch.empty(nelem, dtype=torch.int8, device="cuda")
            s = torch.cuda.memory._snapshot()["segments"]
            segments.append(s)
            done_allocation.set()

        segments = []
        main_thread = threading.Thread(target=main_thread_fn)
        side_thread = threading.Thread(target=side_thread_fn, args=(segments,))

        main_thread.start()
        side_thread.start()
        main_thread.join(timeout=10)
        side_thread.join(timeout=10)

        if main_thread.is_alive() or side_thread.is_alive():
            # release threads so that they don't hang forever
            trigger_alloc.set()
            done_allocation.set()
            self.fail(
                "Test timed out - threads did not complete within the allowed time"
            )

        self.assertEqual(len(segments), 1, "Expected to have memory snapshot")
        self.assertEqual(len(segments[0]), 2, "Expected to have 2 segments allocated")
        active = defaultdict(int)
        for s in segments[0]:
            active[s["segment_pool_id"]] += s["active_size"]
        for k, v in active.items():
            if k == (0, 0):
                self.assertEqual(
                    v, 2097152, "Expected to have 2MB allocated in the default pool"
                )
            else:
                self.assertEqual(
                    v, 0, "Expected to have 0 bytes allocated in the custom pool"
                )

    def test_snapshot_include_traces(self):
        """Test that snapshot() include_traces parameter works correctly"""
        import time

        torch.cuda.empty_cache()
        torch.cuda.memory._record_memory_history()

        pool = torch.cuda.MemPool()

        # Generate trace entries
        with torch.cuda.use_mem_pool(pool):
            tensors = []
            for i in range(1000):
                tensors.append(torch.randn(1024, device="cuda"))
            del tensors

        NUM_RUNS = 10
        times_pool_full = []
        times_pool_notrace = []
        times_global_full = []
        times_global_notrace = []

        # Measure mempool snapshot without traces
        for _ in range(NUM_RUNS):
            # warmup
            snapshot = pool.snapshot(include_traces=False)
        for _ in range(NUM_RUNS):
            start = time.perf_counter()
            snapshot = pool.snapshot(include_traces=False)
            times_pool_notrace.append((time.perf_counter() - start) * 1000)

        # Measure global snapshot without traces
        for _ in range(NUM_RUNS):
            # warmup
            snapshot = torch.cuda.memory_snapshot(include_traces=False)
        for _ in range(NUM_RUNS):
            start = time.perf_counter()
            snapshot = torch.cuda.memory_snapshot(include_traces=False)
            times_global_notrace.append((time.perf_counter() - start) * 1000)

        # Measure mempool snapshot with traces
        for _ in range(NUM_RUNS):
            # warmup
            snapshot = pool.snapshot()
        for _ in range(NUM_RUNS):
            start = time.perf_counter()
            snapshot = pool.snapshot()
            times_pool_full.append((time.perf_counter() - start) * 1000)

        # Measure global snapshot with traces
        for _ in range(NUM_RUNS):
            # warmup
            snapshot = torch.cuda.memory_snapshot()
        for _ in range(NUM_RUNS):
            start = time.perf_counter()
            snapshot = torch.cuda.memory_snapshot()
            times_global_full.append((time.perf_counter() - start) * 1000)

        self.assertTrue(len(snapshot) > 0)

        print(f"Mempool with traces:    {sum(times_pool_full) / NUM_RUNS:.1f} ms")
        print(f"Mempool without traces: {sum(times_pool_notrace) / NUM_RUNS:.1f} ms")
        print(f"Global with traces:     {sum(times_global_full) / NUM_RUNS:.1f} ms")
        print(f"Global without traces:  {sum(times_global_notrace) / NUM_RUNS:.1f} ms")
        print()
        print(f"Mempool speedup: {sum(times_pool_full) / sum(times_pool_notrace):.1f}x")
        print(
            f"Global speedup: {sum(times_global_full) / sum(times_global_notrace):.1f}x"
        )

        torch.cuda.memory._record_memory_history(enabled=None)

    def test_snapshot_include_traces_correctness(self):
        """Test that snapshot with include_traces=False still returns correct memory state"""
        torch.cuda.empty_cache()

        tensor_size = 1024 * 1024
        dtype = torch.uint8
        t1 = torch.ones(tensor_size, device="cuda", dtype=dtype)
        t2 = torch.ones(tensor_size, device="cuda", dtype=dtype)
        t3 = torch.ones(tensor_size, device="cuda", dtype=dtype)

        snapshot_with_traces = torch.cuda.memory_snapshot(include_traces=True)
        snapshot_without_traces = torch.cuda.memory_snapshot(include_traces=False)

        # Both should have segment information
        self.assertTrue(
            len(snapshot_with_traces) > 0, "Snapshot with traces should have segments"
        )
        self.assertTrue(
            len(snapshot_without_traces) > 0,
            "Snapshot without traces should have segments",
        )

        # Count total allocated memory in each snapshot
        def get_total_allocated(segments):
            total = 0
            for segment in segments:
                for block in segment.get("blocks", []):
                    if block.get("state") == "active_allocated":
                        total += block.get("size", 0)
            return total

        allocated_with_traces = get_total_allocated(snapshot_with_traces)
        allocated_without_traces = get_total_allocated(snapshot_without_traces)

        # Allocated memory should match between both snapshots
        self.assertEqual(
            allocated_with_traces,
            allocated_without_traces,
            "Allocated memory should be same with or without traces",
        )

    @unittest.skipIf(TEST_WITH_ROCM, "not enabled by default on rocm")
    @serialTest()
    def test_multi_threads_alloc_in_same_order(self):
        def alloc_tensors(
            stream: torch.cuda.Stream,
            sizes: [int],
            device: str,
            alloc_results: [],
            lock,
            thread_id=0,
        ):
            while len(sizes) > 0:
                with lock:
                    if len(sizes) == 0:
                        return
                    sz = sizes.pop()
                    alloc_results.append(
                        (
                            torch.empty(sz, dtype=torch.int8, device=device),
                            thread_id,
                            sz,
                        )
                    )

        def alloc_with_threads(
            thread_count: int, stream: torch.cuda.Stream, mem_sizes: [int], device: str
        ):
            tensors_list = []
            threads = []
            mem_sizes = mem_sizes[:]
            lock = threading.Lock()
            for i in range(thread_count):
                t = threading.Thread(
                    target=alloc_tensors,
                    args=(stream, mem_sizes, "cuda", tensors_list, lock, i),
                )
                t.start()
                threads.append(t)
            for t in threads:
                t.join()
            tensors_ptrs = [
                (t.data_ptr(), thread_id, sz) for (t, thread_id, sz) in tensors_list
            ]
            return tensors_ptrs

        stream = torch.cuda.current_stream()
        thread_count = 4
        mem_sizes = [
            s * 2 * 1024 * 1024 for s in range(1, 5) for _ in range(thread_count)
        ]
        tensor_ptrs_round_1 = alloc_with_threads(
            thread_count, stream, mem_sizes, "cuda"
        )
        # Force a split-merge cycle: allocate half-sized tensors (splits the
        # freed blocks), then free them (triggers merges back). With per-block
        # counters this changes the free-set ordering; with segment-level
        # counters it stays stable.
        splitters = [
            torch.empty(s, dtype=torch.int8, device="cuda")
            for s in [s * 1024 * 1024 for s in range(1, 5) for _ in range(thread_count)]
        ]
        del splitters
        tensor_ptrs_round_2 = alloc_with_threads(
            thread_count, stream, mem_sizes, "cuda"
        )
        ptrs_round_1 = torch.tensor([ptr for ptr, _, _ in tensor_ptrs_round_1])
        ptrs_round_2 = torch.tensor([ptr for ptr, _, _ in tensor_ptrs_round_2])
        self.assertTrue((ptrs_round_1 == ptrs_round_2).all().item())

    @unittest.skipIf(TEST_WITH_ROCM, "not enabled by default on rocm")
    @serialTest()
    def test_nccl_mem_alloc_addresses_in_random_order(self):
        """
        Test NCCL mem allocator with non-increasing allocation addresses.

        This test uses a custom allocator to simulate ncclMemAlloc scenario where
        allocated memory addresses may not be monotonically increasing across
        different ranks due to fragmentation, OS allocation policies, or prior
        allocations.

        The registration counter-based comparator ensures consistent block ordering
        regardless of actual memory addresses, which is critical for NCCL
        symmetric memory alignment.
        """

        from cuda.bindings import runtime

        ALLOC_FN = ctypes.CFUNCTYPE(
            ctypes.c_void_p, ctypes.c_size_t, ctypes.c_int, ctypes.c_void_p
        )
        FREE_FN = ctypes.CFUNCTYPE(
            None, ctypes.c_void_p, ctypes.c_size_t, ctypes.c_int, ctypes.c_void_p
        )

        class AllocState:
            def __init__(self, test_instance):
                self.first_stream = None
                self.second_stream = None
                self.allocated_addrs = []
                self.buffer_size = 1 * 1024 * 1024 * 1024  # 1GB
                self.base_ptr = -1
                err, base_ptr = runtime.cudaMalloc(self.buffer_size)
                test_instance.assertEqual(
                    err,
                    runtime.cudaError_t.cudaSuccess,
                    "init allocation for test_nccl_mem_alloc_addresses_in_random_order should be successful",
                )
                self.base_ptr = base_ptr
                self.head = base_ptr
                self.tail = base_ptr + self.buffer_size

            def __del__(self):
                if self.base_ptr > 0:
                    runtime.cudaFree(self.base_ptr)

        state = AllocState(self)

        def my_alloc(size, device, stream, _runtime=runtime):
            nonlocal state
            if state.first_stream is None:
                state.first_stream = stream
            elif state.second_stream is None:
                state.second_stream = stream

            if state.first_stream == stream:
                ptr = state.head
                state.head += size
            else:
                state.tail -= size
                ptr = state.tail
            state.allocated_addrs.append(ptr)
            return ptr

        def my_free(ptr, size, device, stream, _runtime=runtime):
            pass

        # Must keep these alive for the lifetime of the allocator
        c_alloc = ALLOC_FN(my_alloc)
        c_free = FREE_FN(my_free)
        alloc_ptr = ctypes.cast(c_alloc, ctypes.c_void_p).value
        free_ptr = ctypes.cast(c_free, ctypes.c_void_p).value
        allocator = torch._C._cuda_customAllocator(alloc_ptr, free_ptr)
        pool = torch.cuda.MemPool(allocator)

        first_stream = torch.cuda.Stream()
        second_stream = torch.cuda.Stream()

        tensor_sizes = [24 * 1024 * 1024, 32 * 1024 * 1024]
        alloc_cases = []
        case_no = 0
        for idx in range(3):
            for stream_idx, stream in enumerate([first_stream, second_stream]):
                # for second stream in reverse order
                reverse_order = stream_idx % 2 == 1
                tensor_sizes.sort(reverse=reverse_order)
                for size in tensor_sizes:
                    alloc_cases.append(
                        {"stream": stream, "size": size, "case_no": case_no}
                    )
                    case_no += 1

        # Test allocation and deallocation patterns
        def alloc_tensors():
            all_tensors = []
            for case in alloc_cases:
                with torch.cuda.stream(case["stream"]):
                    all_tensors.append(
                        torch.empty(case["size"], dtype=torch.uint8, device="cuda")
                    )
            return all_tensors

        with torch.cuda.use_mem_pool(pool):
            first_round_tensors = alloc_tensors()
        tensor_ptrs = [t.data_ptr() for t in first_round_tensors]
        del first_round_tensors

        # Split-merge cycle on ONE block per stream. Each stream has
        # three 24 MB blocks with counters [c0, c1, c2]. Splitting only
        # the first (lowest-counter) block gives it a new, higher
        # counter after the merge, reversing its position among the
        # three. With per-block counters the reallocation order changes
        # (wrong); with segment-level counters it stays stable.
        with torch.cuda.use_mem_pool(pool):
            with torch.cuda.stream(first_stream):
                s1 = torch.empty(tensor_sizes[0] // 2, dtype=torch.uint8, device="cuda")
            with torch.cuda.stream(second_stream):
                s2 = torch.empty(tensor_sizes[0] // 2, dtype=torch.uint8, device="cuda")
            del s1, s2

        with torch.cuda.use_mem_pool(pool):
            second_round_tensors = alloc_tensors()
        second_round_tensor_ptrs = [t.data_ptr() for t in second_round_tensors]
        del second_round_tensors

        mem_snapshot = pool.snapshot()
        self.assertEqual(
            len(mem_snapshot),
            len(tensor_ptrs),
            f"expected to have {len(tensor_ptrs)} segments, but actually got {len(mem_snapshot)}",
        )

        for idx, first_addr in enumerate(tensor_ptrs):
            second_addr = second_round_tensor_ptrs[idx]
            self.assertEqual(
                second_addr,
                first_addr,
                "mem addr allocated for second round should be same with first round",
            )
            self.assertEqual(
                second_addr,
                state.allocated_addrs[idx],
                f"{second_round_tensor_ptrs[idx]=} != {state.allocated_addrs[idx]=}",
            )
        del pool
        del state
        torch.cuda.empty_cache()

    def _alloc_record_free_sync(self, pool_ctx, s2, nbytes):
        """Allocate a block, record it on a second stream, free, synchronize,
        then try to reallocate.  Returns (original_ptr, new_ptr)."""
        with pool_ctx:
            a = torch.empty(nbytes, dtype=torch.uint8, device="cuda")
            original_ptr = a.data_ptr()

            with torch.cuda.stream(s2):
                a.record_stream(s2)
                _ = a + 1

            del a
            torch.cuda.synchronize()

            b = torch.empty(nbytes, dtype=torch.uint8, device="cuda")
            new_ptr = b.data_ptr()
            del b
        return original_ptr, new_ptr

    @serialTest()
    def test_mempool_oom_recovery_releases_cached_blocks(self):
        """Test that the allocator can release default-pool cached blocks to
        satisfy a new allocation when a user mempool is active.

        We fill the default pool's free list with cached blocks (by allocating
        and freeing outside use_mem_pool), then try to allocate inside
        use_mem_pool.  Two paths:
          1. Default pool -- OOM recovery releases cached blocks, succeeds.
          2. use_mem_pool -- same recovery should work (the fix).
        """
        MB = 1024 * 1024
        device = torch.device("cuda:0")

        def align_down_2mb(n):
            return n & ~(2 * MB - 1)

        for label, make_ctx in [
            ("default pool", lambda: contextlib.nullcontext()),
            (
                "user mempool",
                lambda: torch.cuda.use_mem_pool(torch.cuda.MemPool()),
            ),
        ]:
            with self.subTest(label=label):
                torch.cuda.empty_cache()
                free_before = torch.cuda.mem_get_info(device)[0]

                fill_size = align_down_2mb(free_before // 2)
                if fill_size < 64 * MB:
                    self.skipTest("Not enough GPU memory for this test")

                filler = torch.empty(fill_size, dtype=torch.uint8, device=device)
                del filler

                alloc_size = align_down_2mb(free_before - free_before // 8)
                oom = False
                try:
                    with make_ctx():
                        big = torch.empty(alloc_size, dtype=torch.uint8, device=device)
                        del big
                except torch.cuda.OutOfMemoryError:
                    oom = True

                self.assertFalse(
                    oom,
                    f"[{label}] OOM even though the default pool had "
                    f"{fill_size // MB} MiB of freeable cached blocks "
                    "-- release_cached_blocks was likely skipped",
                )

    @serialTest()
    def test_mempool_block_free_not_deferred(self):
        """Test that multi-stream block frees are NOT deferred when only a
        user mempool is active (no graph capture).

        During graph capture, block frees must be deferred (cudaEventRecord
        is illegal).  For user mempools, insert_events should be called
        immediately.  Two paths:
          1. Default pool -- block recycled (baseline).
          2. use_mem_pool -- block recycled (the fix).
        """
        torch.cuda.empty_cache()
        s2 = torch.cuda.Stream()
        nbytes = 1024 * 1024

        for label, pool_ctx in [
            ("default pool", contextlib.nullcontext()),
            ("user mempool", torch.cuda.use_mem_pool(torch.cuda.MemPool())),
        ]:
            with self.subTest(label=label):
                torch.cuda.empty_cache()
                orig, new = self._alloc_record_free_sync(pool_ctx, s2, nbytes)
                self.assertEqual(
                    new,
                    orig,
                    f"[{label}] Block not reused after multi-stream free "
                    "-- free was likely deferred as if under graph capture",
                )


@unittest.skipIf(not TEST_CUDA, "CUDA not available, skipping tests")
@torch.testing._internal.common_utils.markDynamoStrictTest
class TestCudaOptims(TestCase):
    # These tests will be instantiate with instantiate_device_type_tests
    # to apply the new OptimizerInfo structure.

    @onlyCUDA
    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >=5.3 required for graphs"
    )
    @optims(
        [optim for optim in optim_db if optim.has_capturable_arg],
        dtypes=[torch.float32],
    )
    def test_graph_optims(self, device, dtype, optim_info):
        optim_cls = optim_info.optim_cls
        all_optim_inputs = _get_optim_inputs_including_global_cliquey_kwargs(
            device, dtype, optim_info, skip=("differentiable",)
        )

        steps_warmup = 3
        steps_train = 2

        for optim_input in all_optim_inputs:
            kwargs = optim_input.kwargs

            # lr and betas as a Tensor is not supported when capturable=False and foreach=True for torch.optim.adam
            # and torch.optim.adamw
            kwargs["lr"] = 0.1
            if optim_cls in (torch.optim.Adam, torch.optim.AdamW):
                kwargs["betas"] = (0.9, 0.99)

            for actually_do_graphs in (True, False):
                params = [
                    torch.randn((i + 5, i + 5), device=device) for i in range(2)
                ] + [torch.randn((), device=device)]
                params_control = [p.clone().requires_grad_() for p in params]
                params_graphed = [p.clone().requires_grad_() for p in params]

                grads = [
                    [torch.randn_like(p) for p in params]
                    for _ in range(steps_warmup + steps_train)
                ]

                # Control (capturable=False)
                kwargs["capturable"] = False

                opt = optim_cls(params_control, **kwargs)
                for i in range(steps_warmup + steps_train):
                    for j, p in enumerate(params_control):
                        p.grad = grads[i][j]
                    opt.step()

                # capturable=True
                kwargs["capturable"] = True
                opt = optim_cls(params_graphed, **kwargs)

                for i in range(steps_warmup):
                    for j, p in enumerate(params_graphed):
                        p.grad = grads[i][j]
                    opt.step()

                if actually_do_graphs:
                    g = torch.cuda.CUDAGraph()
                    with torch.cuda.graph(g):
                        opt.step()

                for i in range(steps_train):
                    if actually_do_graphs:
                        for j, p in enumerate(params_graphed):
                            p.grad.copy_(grads[i + steps_warmup][j])
                        g.replay()
                    else:
                        # Passing capturable=True to the constructor and running without graphs should still be
                        # numerically correct, even if it's not ideal for performance.
                        for j, p in enumerate(params_graphed):
                            p.grad = grads[i + steps_warmup][j]
                        opt.step()

                for p_control, p_graphed in zip(params_control, params_graphed):
                    self.assertEqual(p_control, p_graphed)

    @onlyCUDA
    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    @optims(
        [
            optim
            for optim in optim_db
            if "fused" in optim.supported_impls and "cuda" in optim.supports_fused_on
        ],
        dtypes=[torch.float32],
    )
    def test_graph_scaling_fused_optimizers(self, device, dtype, optim_info):
        optim_cls = optim_info.optim_cls

        steps_warmup = 3
        steps_train = 2

        optim_inputs = optim_info.optim_inputs_func(device=device)

        for optim_input in optim_inputs:
            kwargs = optim_input.kwargs
            kwargs["fused"] = True

            for actually_do_graphs in (
                (True, False) if optim_info.has_capturable_arg else (True,)
            ):
                params = [torch.randn((i + 5, i + 5), device=device) for i in range(2)]
                params_control = [p.clone().requires_grad_() for p in params]
                params_graphed = [p.clone().requires_grad_() for p in params]

                # `GradScaler` in-place updates gradients thus it's necessary to duplicate gradients.
                grads = [
                    [torch.randn_like(p) for p in params]
                    for _ in range(steps_warmup + steps_train)
                ]
                with torch.no_grad():
                    grads_control = [[g.clone() for g in gs] for gs in grads]
                    grads_graphed = [[g.clone() for g in gs] for gs in grads]

                # Gradient Scaler
                scaler_for_control = torch.cuda.amp.GradScaler(init_scale=128.0)
                with torch.no_grad():
                    scaler_for_control._lazy_init_scale_growth_tracker(device)

                scaler_for_graphed = torch.cuda.amp.GradScaler()
                scaler_for_graphed.load_state_dict(scaler_for_control.state_dict())
                with torch.no_grad():
                    scaler_for_graphed._lazy_init_scale_growth_tracker(device)

                # Control (capturable=False)
                if optim_info.has_capturable_arg:
                    kwargs["capturable"] = False
                opt = optim_cls(params_control, **kwargs)

                for i in range(steps_warmup + steps_train):
                    for j, p in enumerate(params_control):
                        p.grad = grads_control[i][j]
                    scaler_for_control.step(opt)
                    scaler_for_control.update()

                # capturable=True
                if optim_info.has_capturable_arg:
                    kwargs["capturable"] = True
                opt = optim_cls(params_graphed, **kwargs)

                for i in range(steps_warmup):
                    for j, p in enumerate(params_graphed):
                        p.grad = grads_graphed[i][j]
                    scaler_for_graphed.step(opt)
                    scaler_for_graphed.update()

                if actually_do_graphs:
                    g = torch.cuda.CUDAGraph()
                    with torch.cuda.graph(g):
                        scaler_for_graphed.step(opt)
                        scaler_for_graphed.update()

                for i in range(steps_train):
                    if actually_do_graphs:
                        for j, p in enumerate(params_graphed):
                            p.grad.copy_(grads_graphed[i + steps_warmup][j])
                        g.replay()
                    else:
                        # Passing capturable=True to the constructor and running without graphs should still be
                        # numerically correct, even if it's not ideal for performance.
                        for j, p in enumerate(params_graphed):
                            p.grad = grads_graphed[i + steps_warmup][j]
                        scaler_for_graphed.step(opt)
                        scaler_for_graphed.update()

                for p_control, p_graphed in zip(params_control, params_graphed):
                    self.assertEqual(p_control, p_graphed)

    # This test cases compares autocast scaling on a fused optim vs autocast scaling
    # on a forloop optim.
    @onlyNativeDeviceTypes
    @optims(
        [optim for optim in optim_db if "fused" in optim.supported_impls],
        dtypes=[torch.float32],
    )
    def test_grad_scaling_autocast_fused_optimizers(self, device, dtype, optim_info):
        device = device.split(":")[0]
        if device not in optim_info.supports_fused_on:
            self.skipTest(
                f"{device} is not supported for fused on {optim_info.optim_cls.__name__}"
            )
        optim_inputs = optim_info.optim_inputs_func(device=device)
        optim_cls = optim_info.optim_cls
        for optim_input in optim_inputs:
            for _separate_unscale in (True, False):
                kwargs = optim_input.kwargs
                kwargs["fused"] = True
                torch.manual_seed(20)
                (
                    mod_control,
                    mod_scaling,
                    opt_control,
                    opt_scaling,
                    data,
                    loss_fn,
                    _,
                ) = _create_scaling_case(
                    optimizer_ctor=optim_cls, optimizer_kwargs=kwargs, device=device
                )
                optimizer_kwargs = deepcopy(kwargs)
                optimizer_kwargs["fused"] = False
                if "lr" not in kwargs:
                    # _create_scaling_case will set lr = 1.0 if optimizer_kwargs do not set lr
                    optimizer_kwargs["lr"] = 1.0
                opt_control = optim_cls(mod_control.parameters(), **optimizer_kwargs)
                scaler_scaling = torch.amp.GradScaler(device, init_scale=128.0)
                scaler_control = torch.amp.GradScaler(device, init_scale=128.0)

                tracker = TensorTracker()
                # Increase the tolerance for param and (max_)exp_avg_sq when betas aren't tensors
                # cuz the discrepancy between double vs float betas becomes too big. When Tensor
                # betas are used, fused and forloop both would have float betas and the default
                # tolerances are fine.
                assert_eq_kwargs = {}
                if "betas" in opt_control.param_groups[0] and not torch.is_tensor(
                    opt_control.param_groups[0]["betas"][0]
                ):
                    assert_eq_kwargs = {"atol": 2e-5, "rtol": 2e-5}
                for input, target in data:
                    opt_control.zero_grad()
                    with torch.autocast(device_type=device, dtype=torch.half):
                        output_control = mod_control(input)
                        loss_control = loss_fn(output_control, target)
                    scaler_control.scale(loss_control).backward()
                    scaler_control.step(opt_control)
                    scaler_control.update()

                    opt_scaling.zero_grad()
                    with torch.autocast(device_type=device, dtype=torch.half):
                        output_scaling = mod_scaling(input)
                        loss_scaling = loss_fn(output_scaling, target)
                    scaler_scaling.scale(loss_scaling).backward()
                    if _separate_unscale:
                        scaler_scaling.unscale_(opt_scaling)
                    scaler_scaling.step(opt_scaling)
                    scaler_scaling.update()

                    tracker.add(loss_control)
                    tracker.pop_check_set(loss_scaling, self)
                    for param_control, param_scaling in zip(
                        mod_control.parameters(), mod_scaling.parameters()
                    ):
                        tracker.add(param_control.grad)
                        tracker.pop_check_set(param_scaling.grad, self)
                        tracker.add(param_control)
                        tracker.pop_check_set(param_scaling, self, assert_eq_kwargs)

                        state_control, state_scaling = (
                            opt_control.state[param_control],
                            opt_scaling.state[param_scaling],
                        )

                        for k in state_control:
                            actual = state_scaling[k]
                            if k == "step":
                                actual = actual.squeeze()
                            tracker.add(state_control[k])
                            tracker.pop_check_set(
                                actual,
                                self,
                                assert_eq_kwargs
                                if k == "exp_avg_sq" or k == "max_exp_avg_sq"
                                else {},
                            )

    @onlyCUDA
    @parametrize("in_place_unscale", [False, True])
    @optims(
        [optim for optim in optim_db if "cuda" in optim.supports_fused_on],
        dtypes=[torch.float32],
    )
    def test_grad_scaler_with_preset_grad_scale(
        self, device, dtype, optim_info, in_place_unscale
    ):
        weight = torch.ones((5, 5), device="cuda", requires_grad=True)
        weight.grad = torch.full_like(weight, fill_value=15)
        opt = optim_info.optim_cls([weight], lr=0.1, fused=True)
        scaler = torch.amp.GradScaler(init_scale=5)

        # simulate scaling a loss
        scaler.scale(torch.ones(5))

        if in_place_unscale:
            scaler.unscale_(opt)
            # the gradient should have been divided in-place
            self.assertEqual(weight.grad, torch.full_like(weight, fill_value=3))

        # the user sets a `grad_scale` value which should be fused with the optimizer step
        opt.grad_scale = torch.Tensor([3]).cuda()
        scaler.step(opt)

        # check that the user's grad_scale was respected (i.e. the gradient was divided by 5 * 3)
        self.assertEqual(weight.grad, torch.full_like(weight, fill_value=1))

    @onlyCUDA
    @unittest.skipIf(
        not TEST_CUDA_GRAPH, "CUDA >= 11.0 or ROCM >= 5.3 required for graphs"
    )
    @parametrize("foreach, fused", [(False, False), (True, False), (False, True)])
    @optims(
        [
            optim
            for optim in optim_db
            if "foreach" in optim.supported_impls and "cuda" in optim.supports_fused_on
        ],
        dtypes=[torch.float32],
    )
    def test_graph_grad_scaling(self, device, dtype, optim_info, foreach, fused):
        torch.cuda.empty_cache()

        scaler = torch.amp.GradScaler(device="cuda", init_scale=4.0)
        g = torch.cuda.CUDAGraph()
        s = torch.cuda.Stream()

        weight = torch.ones((100,), device="cuda", requires_grad=True)
        opt = optim_info.optim_cls([weight], lr=0.1, foreach=foreach, fused=fused)
        static_input = torch.ones_like(weight)
        static_grad = torch.ones_like(weight)

        # warmup
        s = torch.cuda.Stream()
        s.wait_stream(torch.cuda.current_stream())
        with torch.cuda.stream(s):
            loss = (weight.half() * static_input).sum()
            scaler.scale(loss).backward()
        torch.cuda.current_stream().wait_stream(s)

        opt.zero_grad(set_to_none=True)

        # capture
        with torch.cuda.stream(s):
            g.capture_begin()
            loss = (weight.half() * static_input).sum()
            scaler.scale(loss).backward()
            g.capture_end()

        input_vals = [5, 20000, 5, 40000]
        # If the scale gets updated properly, these are the scale, growth tracker,
        # and grad values we expect.
        expected_scales = [4, 2, 2, 1]
        expected_growth_trackers = [1, 0, 1, 0]
        expected_grad_vals = [5 * 4, float("inf"), 5 * 2, float("inf")]

        for data, scale, growth_tracker, grad_val in zip(
            input_vals, expected_scales, expected_growth_trackers, expected_grad_vals
        ):
            static_input.fill_(data)
            g.replay()
            self.assertEqual(weight.grad, torch.full_like(weight.grad, grad_val))
            scaler.step(opt)
            scaler.update()
            self.assertEqual(scaler._scale, scale)
            self.assertEqual(scaler._growth_tracker, growth_tracker)


@unittest.skipIf(not TEST_CUDA, "CUDA not available, skipping tests")
class TestGDS(TestCase):
    def _get_tmp_dir_fs_type(self):
        my_path = os.path.realpath(tempfile.gettempdir())
        root_type = ""
        for part in psutil.disk_partitions():
            if part.mountpoint == "/":
                root_type = part.fstype
                continue
            if part.mountpoint == my_path:
                return part.fstype
        return root_type

    @unittest.skip("Disabling as USE_CUFILE=0 by default in builds")
    def test_gds_read_write_tensors(self):
        if self._get_tmp_dir_fs_type() not in ("ext4", "xfs"):
            self.skipTest("GPUDirect Storage requires ext4/xfs for local filesystem")
        src1 = torch.randn(1024, device="cuda")
        src2 = torch.randn(2, 1024, device="cuda")
        torch.cuda.gds.gds_register_buffer(src1.untyped_storage())
        torch.cuda.gds.gds_register_buffer(src2.untyped_storage())
        dest1 = torch.empty(1024, device="cuda")
        dest2 = torch.empty(2, 1024, device="cuda")
        with TemporaryFileName() as f:
            file = torch.cuda.gds.GdsFile(f, os.O_CREAT | os.O_RDWR)
            file.save_storage(src1.untyped_storage(), offset=0)
            file.save_storage(src2.untyped_storage(), offset=src1.nbytes)
            file.load_storage(dest1.untyped_storage(), offset=0)
            file.load_storage(dest2.untyped_storage(), offset=src1.nbytes)
        self.assertEqual(src1, dest1)
        self.assertEqual(src2, dest2)
        torch.cuda.gds.gds_deregister_buffer(src1.untyped_storage())
        torch.cuda.gds.gds_deregister_buffer(src2.untyped_storage())


@unittest.skipIf(not TEST_CUDA, "CUDA not available, skipping tests")
class TestCudaAutocast(TestAutocast):
    def setUp(self):
        super().setUp()
        self.autocast_lists = AutocastTestLists(torch.device("cuda:0"))

    def tearDown(self):
        del self.autocast_lists
        super().tearDown()

    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_autocast_torch_fp16(self):
        with torch.backends.cudnn.flags(enabled=True, deterministic=True):
            for op_with_args in self.autocast_lists.torch_fp16:
                skip_test = False
                op, args = op_with_args[0], op_with_args[1]
                if len(op_with_args) == 3:
                    skip_test = op_with_args[2]  # TEST_WITH_ROCM
                if not skip_test:
                    self._run_autocast_outofplace(
                        op, args, torch.float16, device="cuda", amp_dtype=torch.float16
                    )

    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_autocast_torch_bf16(self):
        with torch.backends.cudnn.flags(enabled=True, deterministic=True):
            for op_with_args in self.autocast_lists.torch_fp16:
                skip_test = False
                op, args = op_with_args[0], op_with_args[1]
                if len(op_with_args) == 3:
                    skip_test = op_with_args[2]  # TEST_WITH_ROCM
                should_error_from_cudnn = "cudnn" in op and (
                    "TORCH_CUDNN_V8_API_DISABLED" in os.environ
                    and int(os.environ["TORCH_CUDNN_V8_API_DISABLED"])
                    or torch.cuda.get_device_capability() < (8, 0)
                )
                should_error_from_not_implemented = should_error_from_cudnn
                if not skip_test:
                    if should_error_from_not_implemented:
                        with self.assertRaises(
                            RuntimeError,
                            msg=str(op) + " should not be supported for bfloat16!",
                        ):
                            self._run_autocast_outofplace(
                                op, args, torch.bfloat16, device="cuda"
                            )
                    else:
                        if torch.cuda.is_bf16_supported():
                            self._run_autocast_outofplace(
                                op, args, torch.bfloat16, device="cuda"
                            )
                        else:
                            with self.assertRaisesRegex(
                                RuntimeError, "Device does not support bfloat16"
                            ):
                                self._run_autocast_outofplace(
                                    op, args, torch.bfloat16, device="cuda"
                                )

    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_autocast_torch_fp32(self):
        for op_with_args in self.autocast_lists.torch_fp32:
            op, args, maybe_kwargs = self.args_maybe_kwargs(op_with_args)
            self._run_autocast_outofplace(
                op,
                args,
                torch.float32,
                device="cuda",
                add_kwargs=maybe_kwargs,
                amp_dtype=torch.float16,
            )

    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_autocast_torch_need_autocast_promote(self):
        for op, args in self.autocast_lists.torch_need_autocast_promote:
            self._run_autocast_outofplace(
                op, args, torch.float32, device="cuda", amp_dtype=torch.float16
            )

    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_autocast_torch_expect_builtin_promote(self):
        for op, args, out_type in self.autocast_lists.torch_expect_builtin_promote:
            self._run_autocast_outofplace(
                op,
                args,
                torch.float32,
                device="cuda",
                out_type=out_type,
                amp_dtype=torch.float16,
            )

    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_autocast_nn_fp16(self):
        with torch.backends.cudnn.flags(enabled=True, deterministic=True):
            for op, args in self.autocast_lists.nn_fp16:
                self._run_autocast_outofplace(
                    op,
                    args,
                    torch.float16,
                    device="cuda",
                    module=torch._C._nn,
                    amp_dtype=torch.float16,
                )

    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_autocast_nn_bf16(self):
        with torch.backends.cudnn.flags(enabled=True, deterministic=True):
            for op, args in self.autocast_lists.nn_fp16:
                if torch.cuda.is_bf16_supported():
                    self._run_autocast_outofplace(
                        op, args, torch.bfloat16, device="cuda", module=torch._C._nn
                    )
                else:
                    with self.assertRaisesRegex(
                        RuntimeError, "Device does not support bfloat16"
                    ):
                        self._run_autocast_outofplace(
                            op, args, torch.bfloat16, device="cuda", module=torch._C._nn
                        )

    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_autocast_nn_fp32(self):
        for op, args in self.autocast_lists.nn_fp32:
            self._run_autocast_outofplace(
                op,
                args,
                torch.float32,
                device="cuda",
                module=torch._C._nn,
                amp_dtype=torch.float16,
            )

    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_autocast_linalg_fp16(self):
        with torch.backends.cudnn.flags(enabled=True, deterministic=True):
            for op, args in self.autocast_lists.linalg_fp16:
                self._run_autocast_outofplace(
                    op,
                    args,
                    torch.float16,
                    device="cuda",
                    module=torch._C._linalg,
                    amp_dtype=torch.float16,
                )

    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_autocast_methods_fp16(self):
        with torch.backends.cudnn.flags(enabled=True, deterministic=True):
            for op, args in self.autocast_lists.methods_fp16:
                self._run_autocast_outofplace(
                    op,
                    args,
                    torch.float16,
                    device="cuda",
                    module=None,
                    amp_dtype=torch.float16,
                )

    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_autocast_methods_fp32(self):
        for op, args in self.autocast_lists.methods_fp32:
            self._run_autocast_outofplace(
                op,
                args,
                torch.float32,
                device="cuda",
                module=None,
                amp_dtype=torch.float16,
            )

    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_autocast_methods_expect_builtin_promote(self):
        for op, args, out_type in self.autocast_lists.methods_expect_builtin_promote:
            self._run_autocast_outofplace(
                op,
                args,
                torch.float32,
                device="cuda",
                module=None,
                out_type=out_type,
                amp_dtype=torch.float16,
            )

    def test_autocast_banned(self):
        with torch.autocast("cuda"):
            for op, args, module in self.autocast_lists.banned:
                with self.assertRaises(RuntimeError):
                    getattr(module, op)(*args)

    def test_autocast_ignored_types(self):
        with torch.autocast("cuda"):
            for ignore_type in (torch.double, torch.int32):
                a_ignore = torch.ones((8, 8), dtype=ignore_type, device="cuda:0")
                b_ignore = torch.ones((8, 8), dtype=ignore_type, device="cuda:0")
                c_16 = torch.ones((8, 8), dtype=torch.float16, device="cuda:0")

                # Tests if CastPolicy::fp16 ops ignore double and int
                # Currently, no ops belonging to this policy support integer inputs.
                if ignore_type is torch.double:
                    with self.assertRaises(RuntimeError):
                        torch.mm(a_ignore, c_16)
                    with torch.autocast("cuda", enabled=False):
                        type_no_autocast = torch.mm(a_ignore, b_ignore).dtype
                    self.assertTrue(
                        torch.mm(a_ignore, b_ignore).dtype is type_no_autocast
                    )

                # Tests if CastPolicy::fp32 ops ignore double and int
                with torch.autocast("cuda", enabled=False):
                    type_no_autocast = torch.pow(a_ignore, 2.0).dtype
                self.assertTrue(torch.pow(a_ignore, 2.0).dtype is type_no_autocast)

                # Tests if CastPolicy::fp32_set_opt_dtype ops ignore double and int
                with torch.autocast("cuda", enabled=False):
                    type_no_autocast = torch.sum(a_ignore).dtype
                self.assertTrue(torch.sum(a_ignore).dtype is type_no_autocast)

                # Tests if CastPolicy::fp32_append_dtype ops ignore double and int
                # Currently, no ops belonging to this policy support integer inputs.
                if ignore_type is torch.double:
                    with torch.autocast("cuda", enabled=False):
                        type_no_autocast = torch.norm(a_ignore).dtype
                    self.assertTrue(torch.norm(a_ignore).dtype is type_no_autocast)

    def test_autocast_custom_enabled(self):
        class MyMM(torch.autograd.Function):
            @staticmethod
            @torch.amp.custom_fwd(device_type="cuda")
            def forward(ctx, a, b):
                self.assertTrue(a.dtype is torch.float32)
                self.assertTrue(b.dtype is torch.float32)
                self.assertTrue(torch.is_autocast_enabled())
                ctx.save_for_backward(a, b)
                return a.mm(b)

            @staticmethod
            @torch.amp.custom_bwd(device_type="cuda")
            def backward(ctx, grad):
                self.assertTrue(torch.is_autocast_enabled())
                a, b = ctx.saved_tensors
                a_grad, b_grad = grad.mm(b.t()), a.t().mm(grad)
                self.assertTrue(a_grad.dtype is dtype and b_grad.dtype is dtype)
                return a_grad, b_grad

        mymm = MyMM.apply

        x = torch.randn((8, 8), device="cuda", dtype=torch.float32, requires_grad=True)
        y = torch.randn((8, 8), device="cuda", dtype=torch.float32, requires_grad=True)

        dtypes = (torch.float16, torch.bfloat16) if TEST_BF16 else (torch.float16,)
        for dtype in dtypes:
            with torch.autocast(device_type="cuda", dtype=dtype):
                output = mymm(x, y)
                self.assertTrue(output.dtype is dtype)
                loss = output.sum()
            loss.backward()

    def test_autocast_custom_cast_inputs(self):
        class MyMM(torch.autograd.Function):
            @staticmethod
            @torch.amp.custom_fwd(device_type="cuda", cast_inputs=torch.float32)
            def forward(ctx, a, container, expect_type):
                b = container[1][0]
                self.assertTrue(a.dtype is expect_type)
                self.assertTrue(b.dtype is expect_type)
                self.assertFalse(torch.is_autocast_enabled())
                ctx.save_for_backward(a, b)
                return a.mm(b)

            @staticmethod
            @torch.amp.custom_bwd(device_type="cuda")
            def backward(ctx, grad):
                self.assertFalse(torch.is_autocast_enabled())
                a, b = ctx.saved_tensors
                return grad.mm(b.t()), None, None

        mymm = MyMM.apply

        x = torch.randn((8, 8), device="cuda", dtype=torch.float16, requires_grad=True)
        # Puts one input tensor in a nested container.  y's contained Tensor won't receive a gradient,
        # because torch.autograd.Function can't hand gradients back to non-Tensor forward arguments.
        # Sets requires_grad=False explicitly so we don't lie about expecting a gradient.
        y = (
            0,
            {
                0: torch.randn(
                    (8, 8), device="cuda", dtype=torch.float16, requires_grad=False
                )
            },
        )

        with torch.autocast("cuda"):
            output = mymm(x, y, torch.float32)
            self.assertTrue(output.dtype is torch.float32)
            loss = output.sum()
        loss.backward()

        # Tests if custom_fwd becomes a no-op when mymm runs outside an autocast-enabled region.
        output = mymm(x, y, torch.float16)
        self.assertTrue(output.dtype is torch.float16)
        loss = output.sum()
        loss.backward()

    def test_autocast_custom_deprecated_warning(self):
        with warnings.catch_warnings(record=True) as w:

            class MyMM(torch.autograd.Function):
                @staticmethod
                @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32)
                def forward(ctx, x, y):
                    ctx.save_for_backward(x, y)
                    self.assertFalse(torch.is_autocast_enabled())
                    return x + y

                @staticmethod
                @torch.cuda.amp.custom_bwd
                def backward(ctx, grad):
                    _, _ = ctx.saved_tensors
                    self.assertFalse(torch.is_autocast_enabled())
                    return grad, grad

        self.assertRegex(
            str(w[0].message), r"`torch.cuda.amp.custom_fwd\(args...\)` is deprecated."
        )
        self.assertRegex(
            str(w[1].message), r"`torch.cuda.amp.custom_bwd\(args...\)` is deprecated."
        )

        mymm = MyMM.apply
        x = torch.randn(3, 3, requires_grad=True)
        y = torch.randn(3, 3, requires_grad=True)
        with torch.amp.autocast("cuda"):
            output = mymm(x, y)
            loss = output.sum()
        loss.backward()

    def test_autocast_cat_jit(self):
        # Reported at https://github.com/pytorch/pytorch/issues/38958

        class Model(torch.nn.Module):
            def forward(self):
                a = torch.randn(1)
                b = torch.randn(1)
                c = torch.cat((a, b), 0)
                d = torch.stack([c, c], 0)
                return d

        # The JIT here doesn't really matter, we just need to call
        # cat via the boxed API
        model = Model()
        model_jit_script = torch.jit.script(model)

        with torch.autocast("cuda", enabled=True):
            model()
            model_jit_script()

    # cudnn RNNs require special backend handling (weights are cast to FP16 and reflattened)
    # so they get a dedicated test.
    # Despite the large number of RNN cases it tries, the test takes < 15 seconds on a Titan V (similar to V100).
    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_autocast_rnn(self):
        with torch.backends.cudnn.flags(enabled=True, deterministic=True):
            # seq, batch, features, hidden size
            clses = ("RNN", "GRU", "LSTM")
            T, B, F, H = 3, 4, 5, 6
            dtypes = (torch.float16, torch.float32)
            input_layouts = ("seq_first", "batch_first", "packed")

            for (
                cls,
                num_layers,
                bias,
                input_layout,
                bidirectional,
                try_nonpreflattened_weights,
                input_dtype,
                hidden_dtype,
                weight_dtype,
            ) in product(
                clses,
                (1, 2),
                (True, False),
                input_layouts,
                (True, False),
                (True, False),
                dtypes,
                dtypes,
                dtypes,
            ):
                if input_layout == "seq_first":
                    batch_first = False
                    x = torch.randn((T, B, F), device="cuda", dtype=input_dtype)
                elif input_layout == "batch_first":
                    batch_first = True
                    x = torch.randn((B, T, F), device="cuda", dtype=input_dtype)
                elif input_layout == "packed":
                    batch_first = False
                    x = torch.nn.utils.rnn.pack_padded_sequence(
                        torch.randn((T, B, F), device="cuda", dtype=input_dtype),
                        lengths=(3, 2, 1, 3),
                        enforce_sorted=False,
                    )

                rnn = (
                    getattr(torch.nn, cls)(
                        F,
                        H,
                        num_layers=num_layers,
                        bidirectional=bidirectional,
                        bias=bias,
                        batch_first=batch_first,
                    )
                    .cuda()
                    .to(dtype=weight_dtype)
                )

                if try_nonpreflattened_weights:
                    for p in rnn.parameters():
                        with torch.no_grad():
                            p.set_(p.clone())

                h = torch.randn(
                    (num_layers * (2 if bidirectional else 1), B, H),
                    device="cuda",
                    dtype=hidden_dtype,
                )
                if cls == "LSTM":
                    c = torch.randn(
                        (num_layers * (2 if bidirectional else 1), B, H),
                        device="cuda",
                        dtype=hidden_dtype,
                    )
                    h = (h, c)

                with torch.autocast("cuda"):
                    out, h_out = rnn(x, h)
                out = out.data if input_layout == "packed" else out
                self.assertEqual(out.dtype, torch.float16)
                # Autocast wrapper requires at::_cudnn_rnn is autograd-exposed.  This check can't guarantee
                # at::_cudnn_rnn is autograd-exposed, but if it fires, it indicates some funny business has
                # occurred and we should double check that at::_cudnn_rnn remains autograd-exposed.
                self.assertEqual(
                    out.grad_fn.name(),
                    "MiopenRnnBackward0" if torch.version.hip else "CudnnRnnBackward0",
                )
                out.sum().backward()
                grads = [p.grad.clone() for p in rnn.parameters()]

                rnn.zero_grad()

                if cls == "LSTM":
                    out_control, h_out_control = rnn.to(dtype=torch.float16)(
                        x.half(), (h[0].half(), h[1].half())
                    )
                else:
                    out_control, h_out_control = rnn.to(dtype=torch.float16)(
                        x.half(), h.half()
                    )
                out_control = (
                    out_control.data if input_layout == "packed" else out_control
                )
                out_control.sum().backward()
                grads_control = [p.grad.clone() for p in rnn.parameters()]

                # Compares with default tolerances, even for FP16 execution.  Barring nondeterminism,
                # autocast and control results should be bitwise identical.
                self.assertEqual(out, out_control)

                if cls == "LSTM":
                    self.assertTrue(
                        h_out[0].dtype is torch.float16
                        and h_out[1].dtype is torch.float16
                    )
                    self.assertEqual(h_out[0], h_out_control[0])
                    self.assertEqual(h_out[1], h_out_control[1])
                else:
                    self.assertEqual(h_out.dtype, torch.float16)
                    self.assertEqual(h_out, h_out_control)
                for grad, grad_control in zip(grads, grads_control):
                    self.assertEqual(grad.half(), grad_control)

    @unittest.skipIf(not TEST_CUDNN, "CUDNN not available")
    def test_rnn_packed_sequence_batch_sizes_must_be_cpu(self):
        # Verify that passing batch_sizes on CUDA raises an error instead of segfaulting
        data = torch.randn([18, 16], dtype=torch.float32, device="cuda")
        batch_sizes = torch.tensor([8, 6, 4], dtype=torch.int64, device="cuda")
        hx = torch.randn([1, 8, 32], dtype=torch.float32, device="cuda")
        params = [
            torch.randn([32, 16], dtype=torch.float32, device="cuda"),
            torch.randn([32, 32], dtype=torch.float32, device="cuda"),
        ]

        with self.assertRaisesRegex(
            RuntimeError, "batch_sizes tensor should be on CPU"
        ):
            torch.ops.aten.rnn_relu(
                data,
                batch_sizes=batch_sizes,
                hx=hx,
                params=params,
                has_biases=False,
                num_layers=1,
                dropout=0.5,
                train=False,
                bidirectional=False,
            )

    @serialTest()
    def test_autocast_cache_leak(self):
        # Reported at https://github.com/pytorch/pytorch/issues/48049
        # Test is used to check, if autocast recaches the same parameters
        # when executed in a `torch.no_grad()` block.

        linear = torch.nn.Linear(10, 10).to("cuda")
        data = torch.randn(1, 10, device="cuda")

        with torch.autocast("cuda"):
            with torch.no_grad():
                out = linear(data)
                first_iter_mem = torch.cuda.memory_allocated()
                for _ in range(3):
                    out = linear(data)
                self.assertEqual(first_iter_mem, torch.cuda.memory_allocated())

    def test_autocast_checkpointing(self):
        model = torch.nn.Sequential(
            torch.nn.Linear(8, 8), torch.nn.Linear(8, 8), torch.nn.Linear(8, 8)
        ).cuda()
        input = torch.rand(
            (8, 8), device="cuda", dtype=torch.float16, requires_grad=True
        )
        for reentrant in (True, False):
            with torch.autocast("cuda"):
                output = checkpoint_sequential(model, 2, input, use_reentrant=reentrant)
            self.assertTrue(output.requires_grad)
            self.assertTrue(output.dtype is torch.float16)
            output.sum().backward()

    def test_cuda_autocast_deprecated_warning(self):
        with self.assertWarnsRegex(
            FutureWarning,
            r"`torch.cuda.amp.autocast\(args...\)` is deprecated. Please use `torch.amp.autocast\('cuda', args...\)` instead.",
        ):
            with torch.cuda.amp.autocast():
                _ = torch.ones(10)


class TestCompileKernel(TestCase):
    @unittest.skipIf(not TEST_CUDA, "No CUDA")
    def test_compile_kernel(self):
        # Simple vector addition kernel
        kernel_source = """
        __global__ void add_tensors(const float* a, const float* b, float* c, int n) {
            int i = threadIdx.x + blockIdx.x * blockDim.x;
            if (i < n)
                c[i] = a[i] + b[i];
        }
        """

        # Compile the kernel
        from torch.cuda import _compile_kernel

        add_kernel = _compile_kernel(kernel_source, "add_tensors")

        # Prepare data
        N = 1024
        a = torch.rand(N, device="cuda")
        b = torch.rand(N, device="cuda")
        c = torch.empty_like(a)

        # Calculate grid and block dimensions
        threads_per_block = 256
        blocks_per_grid = (N + threads_per_block - 1) // threads_per_block

        # Launch kernel
        add_kernel(
            grid=(blocks_per_grid, 1, 1),
            block=(threads_per_block, 1, 1),
            args=[a, b, c, N],
        )

        # Verify results
        expected = a + b
        self.assertEqual(c, expected)

        # Test with different tensor types
        a_int = torch.randint(0, 100, (N,), device="cuda", dtype=torch.int32)
        b_int = torch.randint(0, 100, (N,), device="cuda", dtype=torch.int32)
        c_int = torch.empty_like(a_int)

        # Integer addition kernel
        int_kernel_source = """
        __global__ void add_int_tensors(const int* a, const int* b, int* c, int n) {
            int i = threadIdx.x + blockIdx.x * blockDim.x;
            if (i < n)
                c[i] = a[i] + b[i];
        }
        """
        from torch.cuda import _compile_kernel

        add_int_kernel = _compile_kernel(int_kernel_source, "add_int_tensors")

        # Launch kernel
        add_int_kernel(
            grid=(blocks_per_grid, 1, 1),
            block=(threads_per_block, 1, 1),
            args=[a_int, b_int, c_int, N],
        )

        # Verify results
        expected_int = a_int + b_int
        self.assertEqual(c_int, expected_int)

        # Test with header code
        scale_kernel_source = """
        #define SCALE_FACTOR 2.0f

        __device__ float scale_value(float val) {
            return val * SCALE_FACTOR;
        }

        __global__ void scale_tensors(const float* input, float* output, int n) {
            int i = threadIdx.x + blockIdx.x * blockDim.x;
            if (i < n)
                output[i] = scale_value(input[i]);
        }
        """

        scale_kernel = _compile_kernel(scale_kernel_source, "scale_tensors")

        input_tensor = torch.rand(N, device="cuda")
        output_tensor = torch.empty_like(input_tensor)

        scale_kernel(
            grid=(blocks_per_grid, 1, 1),
            block=(threads_per_block, 1, 1),
            args=[input_tensor, output_tensor, N],
        )

        # Verify scaling
        expected_scaled = input_tensor * 2.0
        self.assertEqual(output_tensor, expected_scaled)

        # Test error handling with invalid kernel
        invalid_kernel_source = """
        __global__ void invalid_kernel(float* a) {
            undeclared_variable = 10; // This will cause a compilation error
        }
        """

        with self.assertRaises(RuntimeError):
            _compile_kernel(invalid_kernel_source, "invalid_kernel")

    @unittest.skipIf(not TEST_CUDA, "No CUDA")
    def test_compile_kernel_large_shared_memory(self):
        kernel_source = """
        __global__ void large_shared_memory_kernel(const float* input, float* output, int n) {
            extern __shared__ float shared_data[];

            int tid = threadIdx.x;
            int idx = blockIdx.x * blockDim.x + threadIdx.x;

            // Load data into shared memory
            if (idx < n) {
                shared_data[tid] = input[idx];
            } else {
                shared_data[tid] = 0.0f;
            }
            __syncthreads();

            // Perform reduction in shared memory
            for (int stride = blockDim.x / 2; stride > 0; stride >>= 1) {
                if (tid < stride) {
                    shared_data[tid] += shared_data[tid + stride];
                }
                __syncthreads();
            }

            // Write result
            if (tid == 0) {
                output[blockIdx.x] = shared_data[0];
            }
        }
        """

        from torch.cuda import _compile_kernel, get_device_properties

        kernel = _compile_kernel(kernel_source, "large_shared_memory_kernel")

        threads_per_block = 1024  # 1024 threads * 4 bytes = 4KB, but we'll request 64KB
        shared_mem_size = 64 * 1024  # 64KB

        kernel.set_shared_memory_config(shared_mem_size)

        N = 4096
        input_data = torch.ones(N, device="cuda", dtype=torch.float32)
        output_data = torch.zeros(4, device="cuda", dtype=torch.float32)  # 4 blocks

        kernel(
            grid=(4, 1, 1),
            block=(threads_per_block, 1, 1),
            args=[input_data, output_data, N],
            shared_mem=shared_mem_size,
        )

        # Each block should sum 1024 ones = 1024
        expected = torch.full((4,), 1024.0, dtype=torch.float32)
        self.assertEqual(output_data.cpu(), expected)

        # Test error handling with more than supported shared memory size
        if torch.version.hip:
            max_smem = (
                65536 if get_device_properties().gcnArchName != "gfx950" else 160 * 1024
            )
        else:
            max_smem = get_device_properties().shared_memory_per_block_optin
        excessive_shared_mem = max_smem * 2

        with self.assertRaises(RuntimeError):
            kernel.set_shared_memory_config(excessive_shared_mem)

    @skipIfRocmArch(MI300_ARCH)
    @tf32_on_and_off(0.005)
    @unittest.skipIf(not TEST_CUDA, "No CUDA")
    def test_compile_kernel_advanced(self):
        # Test matrix multiplication
        matmul_kernel_source = """
        __global__ void matrix_multiply(const float* A, const float* B, float* C, int M, int N, int K) {
            int row = blockIdx.y * blockDim.y + threadIdx.y;
            int col = blockIdx.x * blockDim.x + threadIdx.x;

            if (row < M && col < N) {
                float sum = 0.0f;
                for (int i = 0; i < K; i++) {
                    sum += A[row * K + i] * B[i * N + col];
                }
                C[row * N + col] = sum;
            }
        }
        """
        from torch.cuda import _compile_kernel

        matmul_kernel = _compile_kernel(matmul_kernel_source, "matrix_multiply")

        # Matrix dimensions
        M, K, N = 64, 32, 48

        # Create matrices
        A = torch.rand((M, K), device="cuda")
        B = torch.rand((K, N), device="cuda")
        C = torch.zeros((M, N), device="cuda")

        # Calculate grid and block dimensions
        block_dim = (16, 16, 1)
        grid_dim = (
            (N + block_dim[0] - 1) // block_dim[0],
            (M + block_dim[1] - 1) // block_dim[1],
            1,
        )

        # Launch kernel
        matmul_kernel(
            grid=grid_dim,
            block=block_dim,
            args=[A.contiguous(), B.contiguous(), C, M, N, K],
        )

        # Verify results
        expected = torch.matmul(A, B)
        self.assertEqual(C, expected)

        # Test with different compute capability if specified
        device_props = torch.cuda.get_device_properties(torch.cuda.current_device())
        if not torch.version.hip:
            compute_cap = f"{device_props.major}{device_props.minor}"
        else:
            compute_cap = f"{device_props.gcnArchName}"

        # Recompile with explicit compute capability
        matmul_kernel_explicit = _compile_kernel(
            matmul_kernel_source, "matrix_multiply", compute_capability=compute_cap
        )

        C_explicit = torch.zeros((M, N), device="cuda")

        # Launch kernel
        matmul_kernel_explicit(
            grid=grid_dim,
            block=block_dim,
            args=[A.contiguous(), B.contiguous(), C_explicit, M, N, K],
        )

        # Verify results
        self.assertEqual(C_explicit, expected)

    @unittest.skipIf(not TEST_CUDA, "No CUDA")
    def test_compile_kernel_as_custom_op(self):
        # Define a simple vector addition kernel
        kernel_source = """
        __global__ void vector_add(const float* a, const float* b, float* c, int n) {
            int idx = blockIdx.x * blockDim.x + threadIdx.x;
            if (idx < n) {
                c[idx] = a[idx] + b[idx];
            }
        }
        """

        @torch.library.custom_op("test_compile_kernel::vector_add", mutates_args=())
        def vector_add_op(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
            from torch.cuda import _compile_kernel

            # Validate that tensors are 1-dimensional and have the same size
            torch._check(
                a.dim() == 1,
                lambda: f"Expected tensor 'a' to be 1-dimensional, but got {a.dim()} dimensions",
            )
            torch._check(
                b.dim() == 1,
                lambda: f"Expected tensor 'b' to be 1-dimensional, but got {b.dim()} dimensions",
            )
            torch._check(
                a.size() == b.size(),
                lambda: f"Expected tensors to have the same size, but got a.size()={a.size()} and b.size()={b.size()}",
            )
            compiled_kernel = _compile_kernel(kernel_source, "vector_add")

            c = torch.empty_like(a)
            n = a.numel()

            threads_per_block = 256
            blocks_per_grid = (n + threads_per_block - 1) // threads_per_block
            compiled_kernel(
                grid=(blocks_per_grid, 1, 1),
                block=(threads_per_block, 1, 1),
                args=[a, b, c, n],
            )

            return c

        @vector_add_op.register_fake
        def _(a, b):
            return torch.empty_like(a)

        device = torch.device("cuda:0")
        size = (1024,)

        a = torch.randn(size, device=device, dtype=torch.float32)
        b = torch.randn(size, device=device, dtype=torch.float32)

        result = vector_add_op(a, b)

        expected = a + b
        torch.testing.assert_close(result, expected, rtol=1e-5, atol=1e-5)

    @unittest.skipIf(not TEST_CUDA, "No CUDA")
    def test_compile_kernel_custom_op_validation(self):
        kernel_source = """
        __global__ void add_scalar(const float* input, float* output, double scalar, int n) {
            int idx = blockIdx.x * blockDim.x + threadIdx.x;
            if (idx < n) {
                output[idx] = input[idx] + scalar;
            }
        }
        """

        @torch.library.custom_op("test_compile_kernel::add_scalar", mutates_args=())
        def add_scalar_op(input_tensor: torch.Tensor, scalar: float) -> torch.Tensor:
            from torch.cuda import _compile_kernel

            compiled_kernel = _compile_kernel(kernel_source, "add_scalar")

            output = torch.empty_like(input_tensor)
            n = input_tensor.numel()

            threads_per_block = 256
            blocks_per_grid = (n + threads_per_block - 1) // threads_per_block
            compiled_kernel(
                grid=(blocks_per_grid, 1, 1),
                block=(threads_per_block, 1, 1),
                args=[input_tensor, output, scalar, n],
            )

            return output

        @add_scalar_op.register_fake
        def _(input_tensor, scalar):
            return torch.empty_like(input_tensor)

        # Test with opcheck
        device = torch.device("cuda:0")
        input_data = torch.randn((64,), device=device, dtype=torch.float32)
        scalar_val = 3.14

        # Run opcheck validation
        torch.library.opcheck(add_scalar_op, (input_data, scalar_val), {})

        # Also test the actual functionality
        result = add_scalar_op(input_data, scalar_val)
        expected = input_data + scalar_val
        torch.testing.assert_close(result, expected, rtol=1e-5, atol=1e-5)

    @unittest.skipIf(not TEST_CUDA, "No CUDA")
    def test_compile_kernel_double_precision(self):
        """Test that Python floats are correctly handled as doubles in kernels."""
        kernel_source = """
        __global__ void test_double_precision(double* output, double value, int n) {
            int idx = blockIdx.x * blockDim.x + threadIdx.x;
            if (idx < n) {
                output[idx] = value;
            }
        }
        """

        from torch.cuda import _compile_kernel

        compiled_kernel = _compile_kernel(kernel_source, "test_double_precision")

        # Test with high precision value that would lose precision if cast to float32
        # float32 has 7 digits of precision, so we use a value with 15 digits
        high_precision_value = 1.23456789012345
        n = 10

        output = torch.zeros(n, device="cuda", dtype=torch.float64)
        compiled_kernel(
            grid=(1, 1, 1),
            block=(256, 1, 1),
            args=[output, high_precision_value, n],
        )

        # Verify high precision is preserved (would fail with old float32 casting)
        expected = torch.full(
            (n,), high_precision_value, device="cuda", dtype=torch.float64
        )
        torch.testing.assert_close(output, expected, rtol=1e-14, atol=1e-14)

    @unittest.skipIf(not TEST_CUDA, "No CUDA")
    def test_compile_kernel_cuda_headers(self):
        """Test that kernels can include and use CUDA headers like cuda_fp16.h."""
        kernel_source = """
        #ifndef __HIPCC__
        #include <cuda_fp16.h>
        #endif

        extern "C"
        __global__ void half_precision_kernel(__half* output, double input_value, int n) {
            int idx = blockIdx.x * blockDim.x + threadIdx.x;
            if (idx < n) {
                output[idx] = __float2half((float)input_value);
            }
        }
        """

        from torch.cuda import _compile_kernel

        compiled_kernel = _compile_kernel(kernel_source, "half_precision_kernel")

        n = 100
        test_value = 3.14159
        output = torch.zeros(n, device="cuda", dtype=torch.float16)

        compiled_kernel(
            grid=(1, 1, 1),
            block=(256, 1, 1),
            args=[output, test_value, n],
        )

        expected = torch.full((n,), test_value, device="cuda", dtype=torch.float16)
        torch.testing.assert_close(output, expected, rtol=1e-3, atol=1e-3)

    @unittest.skipIf(not TEST_CUDA, "No CUDA")
    def test_compile_kernel_template(self):
        kernel_source = """
        template<typename T>
        __global__ void add_tensors(const T* a, const T* b, T* c, int n) {
            int i = threadIdx.x + blockIdx.x * blockDim.x;
            if (i < n)
                c[i] = a[i] + b[i];
        }
        """

        # Compile the kernel
        from torch.cuda import _compile_kernel

        add_kernel_float = _compile_kernel(kernel_source, "add_tensors<float>")

        # Prepare data
        N = 1024
        a = torch.rand(N, device="cuda")
        b = torch.rand(N, device="cuda")
        c = torch.empty_like(a)

        # Calculate grid and block dimensions
        threads_per_block = 256
        blocks_per_grid = (N + threads_per_block - 1) // threads_per_block

        # Launch kernel
        add_kernel_float(
            grid=(blocks_per_grid, 1, 1),
            block=(threads_per_block, 1, 1),
            args=[a, b, c, N],
        )

        # Verify results
        expected = a + b
        self.assertEqual(c, expected)

        # do again with different dtype
        add_kernel_int = _compile_kernel(kernel_source, "add_tensors<int>")

        # Prepare data
        N = 1024
        a = torch.randint(-1000, 1000, size=(N,), dtype=torch.int, device="cuda")
        b = torch.randint(-1000, 1000, size=(N,), dtype=torch.int, device="cuda")
        c = torch.empty_like(a)

        # Calculate grid and block dimensions
        threads_per_block = 256
        blocks_per_grid = (N + threads_per_block - 1) // threads_per_block

        # Launch kernel
        add_kernel_int(
            grid=(blocks_per_grid, 1, 1),
            block=(threads_per_block, 1, 1),
            args=[a, b, c, N],
        )

        # Verify results
        expected = a + b
        self.assertEqual(c, expected)

    @unittest.skipIf(not TEST_CUDA, "No CUDA")
    def test_compile_kernel_dlpack(self):
        """Test that compile_kernel works with tensors created via DLPack."""
        kernel_source = """
        __global__ void add_tensors(const float* a, const float* b, float* c, int n) {
            int i = threadIdx.x + blockIdx.x * blockDim.x;
            if (i < n)
                c[i] = a[i] + b[i];
        }
        """

        from torch.cuda import _compile_kernel

        add_kernel = _compile_kernel(kernel_source, "add_tensors")

        N = 512
        a = torch.rand(N, device="cuda", dtype=torch.float32)
        b = torch.rand(N, device="cuda", dtype=torch.float32)

        a_dlpack = torch.utils.dlpack.from_dlpack(torch.utils.dlpack.to_dlpack(a))
        b_dlpack = torch.utils.dlpack.from_dlpack(torch.utils.dlpack.to_dlpack(b))
        c = torch.empty_like(a)

        threads_per_block = 256
        blocks_per_grid = (N + threads_per_block - 1) // threads_per_block

        add_kernel(
            grid=(blocks_per_grid, 1, 1),
            block=(threads_per_block, 1, 1),
            args=[a_dlpack, b_dlpack, c, N],
        )

        self.assertEqual(c, a + b)
        a_dlpack[0] = 42.0
        self.assertEqual(a[0].item(), 42.0, "DLPack tensors should share memory")


@unittest.skipIf(not TEST_CUDA, "CUDA not available, skipping tests")
class TestCudaDeviceParametrized(TestCase):
    @skipIfRocmVersionLessThan((7, 0))
    @skipCUDAIf(
        not SM70OrLater, "Compute capability >= SM70 required for relaxed ptx flag"
    )
    def test_graph_external_wait_and_record(self):
        torch.cuda.empty_cache()

        spin_wait_kernel = get_wait_for_cpu_kernel()

        x = torch.ones(4, device="cuda")
        x_cpu = torch.zeros(x.shape, device="cpu").pin_memory()
        flag_cpu = torch.zeros(1, dtype=torch.int32, device="cpu").pin_memory()
        start_event = torch.cuda.Event(external=True)
        end_event = torch.cuda.Event(external=True)

        signalling_graph = torch.cuda.CUDAGraph()
        with torch.cuda.graph(signalling_graph, capture_error_mode="relaxed"):
            spin_wait_kernel(grid=(1, 1, 1), block=(1, 1, 1), args=[flag_cpu])
            start_event.record()

        # This is counter-intuitive, but a cudaEventRecord() during
        # stream capture does not count as the first call to
        # cudaEventRecord(). Rather, cudaGraphLaunch() counts as the
        # first call to cudaEventRecord(). Therefore, all calls to
        # cudaEventQuery() will succeed before that happens.

        # See:
        # "Before the first call to cudaEventRecord(), an event represents an empty set of work, so for example cudaEventQuery() would return cudaSuccess."
        # https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__EVENT.html
        self.assertTrue(start_event.query(), "Start event's work should be empty")

        work_graph = torch.cuda.CUDAGraph()
        with torch.cuda.graph(work_graph, capture_error_mode="relaxed"):
            start_event.wait()
            x_cpu.copy_(x, non_blocking=True)
            end_event.record()

        self.assertTrue(
            torch.all(x_cpu == 0.0), "Copy cannot occur until start_event is recorded"
        )

        try:
            signalling_stream = torch.cuda.Stream()
            with torch.cuda.stream(signalling_stream):
                signalling_graph.replay()

            work_stream = torch.cuda.Stream()
            with torch.cuda.stream(work_stream):
                work_graph.replay()

            self.assertFalse(
                end_event.query(), "Event record node cannot run until flag_cpu[0]=1"
            )

            # Sleep for a little to make sure that work doesn't proceed until we set flag_cpu[0]=1
            time.sleep(1)
            self.assertTrue(
                torch.all(x_cpu == 0.0),
                "Copy cannot occur until start_event is recorded",
            )
        finally:
            # In case an assertion fails, we still need to empty out
            # the GPU queue of work. Therefore, we do this write
            # unconditionally, even if an exception is thrown.

            # This writes allows wait_for_cpu to proceed
            # This is an atomic store at system scope according to this rule:
            # "the scope is thread_scope_system and it is a load or store that affects a naturally-aligned object of sizes 1, 2, 4, 8, or 16 bytes on mapped memory"
            # https://nvidia.github.io/cccl/libcudacxx/extended_api/memory_model.html#atomicity

            # Note that every CPU store is implicitly system scope,
            # even if we don't use C++ atomics like this:
            # std::atomic_ref<int32_t>::store(1);
            flag_cpu[0] = 1

        end_event.synchronize()
        self.assertTrue(
            torch.all(x_cpu == 1.0),
            "Copy should be done once end_event is synchronized",
        )
        # On HIP graphs, the event-record node can become visible to
        # event.synchronize() a few microseconds before the host-side
        # stream state is updated; poll briefly so we exercise the
        # invariant ("after sync the stream is drained") without flaking.
        deadline = time.monotonic() + 0.1
        while not work_stream.query() and time.monotonic() < deadline:
            pass
        self.assertTrue(
            work_stream.query(),
            "end_event.synchronize() completing should imply that work_stream is done",
        )


class TestFXMemoryProfiler(TestCase):
    """Tests for memory profiler augmentation with original stack traces."""

    class MLPModule(nn.Module):
        def __init__(self, device):
            super().__init__()
            torch.manual_seed(5)
            self.net1 = nn.Linear(10, 16, bias=True, device=device)
            self.relu = nn.ReLU()
            self.net2 = nn.Linear(16, 10, bias=True, device=device)

        def forward(self, x):
            a = self.net1(x)
            b = self.relu(a)
            c = self.net2(b)
            return c

    class MLPModule2(nn.Module):
        def __init__(self, device):
            super().__init__()
            torch.manual_seed(5)
            self.net1 = nn.Linear(10, 16, bias=True, device=device)
            self.relu = nn.ReLU()
            self.net2 = nn.Linear(16, 10, bias=True, device=device)

        def forward(self, x):
            d = self.net1(x)
            e = self.relu(d)
            f = self.net2(e)
            return f

    def collect_frames(
        self, augmented_snapshot, collect_device_traces=True, collect_segments=True
    ):
        """Collects all frames that has node metadata from a memory snapshot."""
        # Collect all frames with FX metadata
        fx_frames = []

        # Check device traces for FX debug fields
        if collect_device_traces and "device_traces" in augmented_snapshot:
            for trace_list in augmented_snapshot["device_traces"]:
                for trace_entry in trace_list:
                    if isinstance(trace_entry, dict) and "frames" in trace_entry:
                        for frame in trace_entry["frames"]:
                            if isinstance(frame, dict):
                                # Check for FX debug fields
                                if "fx_node_op" in frame or "fx_node_name" in frame:
                                    fx_frames.append(frame)

        # Check segments/blocks for FX debug fields
        if collect_segments and "segments" in augmented_snapshot:
            for segment in augmented_snapshot["segments"]:
                if "blocks" in segment:
                    for block in segment["blocks"]:
                        if "frames" in block:
                            for frame in block["frames"]:
                                if isinstance(frame, dict):
                                    if "fx_node_op" in frame or "fx_node_name" in frame:
                                        fx_frames.append(frame)
        return fx_frames

    @unittest.skipIf(not torch.cuda.is_available(), "CUDA not available")
    @torch.fx.experimental._config.patch("enrich_profiler_metadata", True)
    def test_fx_memory_profiler_augmentation(self):
        """Test that memory snapshots are augmented with FX debug information."""

        device = "cuda"
        mod = self.MLPModule(device)
        # reset cache to start fresh
        torch.cuda.memory.empty_cache()
        torch.cuda.memory._record_memory_history()
        compiled = torch.compile(mod, backend="aot_eager", fullgraph=True)
        result = compiled(torch.randn(10, 10, device=device))
        augmented_snapshot = torch.cuda.memory._snapshot(augment_with_fx_traces=True)
        torch.cuda.memory._record_memory_history(enabled=None, clear_history=True)
        torch.cuda.empty_cache()

        fx_frames = self.collect_frames(augmented_snapshot)
        self.assertGreater(len(fx_frames), 2)

        for frame in fx_frames:
            # Every FX frame should have both node_op and node_name
            self.assertIn("fx_node_op", frame)
            self.assertIn("fx_node_name", frame)
            self.assertIn("fx_node_target", frame)
            self.assertIn("fx_original_trace", frame)

            self.assertIn(frame["fx_node_name"], ["addmm", "relu", "addmm_1"])
            fx_node_name = frame["fx_node_name"]
            if fx_node_name == "addmm":
                self.assertIn("a = self.net1(x)", frame["fx_original_trace"])
            elif fx_node_name == "addmm_1":
                self.assertIn("c = self.net2(b)", frame["fx_original_trace"])
            elif fx_node_name == "relu":
                self.assertIn("b = self.relu(a)", frame["fx_original_trace"])

        # Test that when we have two graphs with the same src_code, they're not hashed
        # to the same metadata
        mod = self.MLPModule2(device)
        torch.cuda.memory._record_memory_history()
        compiled = torch.compile(mod, backend="aot_eager", fullgraph=True)
        result = compiled(torch.randn(10, 10, device=device))
        augmented_snapshot = torch.cuda.memory._snapshot(augment_with_fx_traces=True)
        torch.cuda.memory._record_memory_history(enabled=None, clear_history=True)

        # avoid collecting segments from previous run for unit test purpose
        fx_frames = self.collect_frames(augmented_snapshot, collect_segments=False)
        self.assertGreater(len(fx_frames), 0)

        for frame in fx_frames:
            # Every FX frame should have both node_op and node_name
            self.assertIn("fx_node_op", frame)
            self.assertIn("fx_node_name", frame)
            self.assertIn("fx_node_target", frame)
            self.assertIn("fx_original_trace", frame)

            self.assertIn(frame["fx_node_name"], ["addmm", "relu", "addmm_1"])
            fx_node_name = frame["fx_node_name"]
            if fx_node_name == "addmm":
                self.assertIn("d = self.net1(x)", frame["fx_original_trace"])
            elif fx_node_name == "addmm_1":
                self.assertIn("f = self.net2(e)", frame["fx_original_trace"])
            elif fx_node_name == "relu":
                self.assertIn("e = self.relu(d)", frame["fx_original_trace"])


@unittest.skipIf(
    not PLATFORM_SUPPORTS_GREEN_CONTEXT, "Green contexts are not supported"
)
class TestCudaGreenContexts(TestCase):
    def setUp(self):
        super().setUp()

    def tearDown(self):
        super().tearDown()

    def test_greencontext_restores_stream(self):
        # need to start on a side stream as we are comparing pointers and want to avoid
        # two NULL streams...
        s = torch.cuda.Stream()
        with torch.cuda.stream(s):
            start_stream = torch.cuda.current_stream()
            ctx = torch.cuda.green_contexts.GreenContext.create(num_sms=1)
            ctx.set_context()
            context_stream = torch.cuda.current_stream()
            ctx.pop_context()
            end_stream = torch.cuda.current_stream()
            self.assertEqual(start_stream.cuda_stream, end_stream.cuda_stream)
            self.assertNotEqual(start_stream.cuda_stream, context_stream.cuda_stream)

    @serialTest()
    def test_greencontext_carveout(self):
        # By default, everything is performed on the current device
        a = torch.randn(4096, 4096, device="cuda", dtype=torch.bfloat16)
        ctx = torch.cuda.green_contexts.GreenContext.create(num_sms=1)
        ctx.set_context()
        torch.matmul(a, a)
        torch.cuda.synchronize()
        t0 = time.perf_counter()
        partial_res = torch.matmul(a, a)
        torch.cuda.synchronize()
        t1 = time.perf_counter()
        ctx.pop_context()
        torch.matmul(a, a)
        torch.cuda.synchronize()
        t2 = time.perf_counter()
        full_res = torch.matmul(a, a)
        torch.cuda.synchronize()
        t3 = time.perf_counter()
        self.assertEqual(partial_res, full_res)
        self.assertGreater(t1 - t0, t3 - t2)

    @serialTest()
    def test_greencontext_stream_carveout(self):
        # By default, everything is performed on the current device
        a = torch.randn(4096, 4096, device="cuda", dtype=torch.bfloat16)
        ctx = torch.cuda.green_contexts.GreenContext.create(num_sms=1)
        ctx_stream = ctx.Stream()
        with torch.cuda.stream(ctx_stream):
            torch.matmul(a, a)
            torch.cuda.synchronize()
            t0 = time.perf_counter()
            partial_res = torch.matmul(a, a)
            torch.cuda.synchronize()
            t1 = time.perf_counter()
        torch.matmul(a, a)
        torch.cuda.synchronize()
        t2 = time.perf_counter()
        full_res = torch.matmul(a, a)
        torch.cuda.synchronize()
        t3 = time.perf_counter()
        self.assertEqual(partial_res, full_res)
        self.assertGreater(t1 - t0, t3 - t2)

    @serialTest()
    def test_greencontext_graphs(self):
        # By default, everything is performed on the current device
        a = torch.randn(4096, 4096, device="cuda", dtype=torch.bfloat16)
        ctx = torch.cuda.green_contexts.GreenContext.create(num_sms=1)
        ctx.set_context()
        partial_res = torch.matmul(a, a)
        ctx.pop_context()
        full_res = torch.matmul(a, a)
        full_res.zero_()
        partial_res.zero_()
        torch.cuda.synchronize()

        g = torch.cuda.CUDAGraph()
        with torch.cuda.graph(g):
            ctx.set_context()
            partial_res = torch.matmul(a, a)
            ctx.pop_context()
            full_res = torch.matmul(a, a)
        g.replay()
        self.assertEqual(partial_res, full_res)

    @unittest.skipIf(
        not PLATFORM_SUPPORTS_WORKQUEUE_CONFIG, "Workqueue config is not supported"
    )
    @serialTest()
    def test_greencontext_workqueue_concurrency_limit(self):
        # By default, everything is performed on the current device
        n_streams = 4
        GreenContext = torch.cuda.green_contexts.GreenContext
        max_wq = GreenContext.max_workqueue_concurrency()
        if max_wq < n_streams:
            self.skipTest(
                f"Device has {max_wq} workqueue(s), need >{n_streams} to test concurrency limiting"
            )

        wq_events = [torch.cuda.Event(enable_timing=True) for _ in range(n_streams * 2)]

        def run_multi_stream_sleep(streams):
            for i, s in enumerate(streams):
                with torch.cuda.stream(s):
                    # note we need to record timing events to ensure that the
                    # workqueue is actually used
                    wq_events[i * 2].record(s)
                    # 100M cycles is enough on all currently supported GPUs
                    # to ensure that the test runs significantly longer than
                    # host overhead to 1) activate the different streams,
                    # 2) record timing events, and 3) synchronize with the GPU.
                    torch.cuda._sleep(100_000_000)
                    wq_events[i * 2 + 1].record(s)
            torch.cuda.synchronize()

        # Note: in case we have lazy module loading, ensure that we called the
        # sleep kernel at least once s.t. it is loaded in memory.
        torch.cuda._sleep(1_000)

        # Baseline: n_streams streams from default context, full workqueue concurrency
        baseline_streams = [torch.cuda.Stream() for _ in range(n_streams)]
        # note: torch.cuda.synchronize() will wait for all kernels on all streams
        # so we can safely use CPU based timing here.
        t0 = time.perf_counter()
        run_multi_stream_sleep(baseline_streams)
        t1 = time.perf_counter()
        baseline_time = t1 - t0

        # Green context with workqueue concurrency limited to 1
        ctx = GreenContext.create(
            workqueue_scope="balanced",
            workqueue_concurrency_limit=1,
        )
        ctx.set_context()
        green_streams = [ctx.Stream() for _ in range(n_streams)]
        t2 = time.perf_counter()
        run_multi_stream_sleep(green_streams)
        t3 = time.perf_counter()
        ctx.pop_context()
        limited_time = t3 - t2

        self.assertGreater(limited_time, baseline_time)


class TestCudaArchList(TestCase):
    def test_get_arch_list_empty_when_cuda_not_compiled(self):
        if torch.cuda._is_compiled():
            self.skipTest("CUDA is compiled")
        self.assertEqual(torch.cuda.get_arch_list(), [])

    @unittest.skipIf(not torch.cuda._is_compiled(), "CUDA not compiled")
    def test_get_arch_list_does_not_require_cuda_is_available(self):
        expected = (torch._C._cuda_getArchFlags() or "").split()
        with patch("torch.cuda.is_available", return_value=False):
            self.assertEqual(torch.cuda.get_arch_list(), expected)


instantiate_parametrized_tests(TestCuda)
instantiate_parametrized_tests(TestCudaAllocator)
instantiate_parametrized_tests(TestCompileKernel)
instantiate_parametrized_tests(TestCachingHostAllocatorCudaGraph)
instantiate_parametrized_tests(TestCachingHostAllocatorConfig)
instantiate_device_type_tests(TestCudaOptims, globals())
instantiate_device_type_tests(TestCudaDeviceParametrized, globals())
instantiate_device_type_tests(TestCudaGreenContexts, globals(), except_for="cpu")


# Tests for fp32_precision flag propagation that don't require an actual CUDA
# device — they only exercise C++ context state management.
class TestFP32PrecisionFlags(TestCase):
    @recover_orig_fp32_precision
    @serialTest()
    def test_generic_fp32_precision_propagates_to_cudnn_conv_rnn(self):
        # Regression test: setting torch.backends.fp32_precision = "ieee"
        # must propagate to cudnn.conv and cudnn.rnn.
        self.assertEqual(torch.backends.cudnn.conv.fp32_precision, "tf32")
        self.assertEqual(torch.backends.cudnn.rnn.fp32_precision, "tf32")
        with torch.backends.flags(fp32_precision="ieee"):
            self.assertEqual(torch.backends.cudnn.conv.fp32_precision, "ieee")
            self.assertEqual(torch.backends.cudnn.rnn.fp32_precision, "ieee")
            self.assertEqual(torch.backends.cudnn.fp32_precision, "ieee")
            self.assertEqual(torch.backends.cuda.matmul.fp32_precision, "ieee")


if __name__ == "__main__":
    run_tests()