ValueError: Pointer argument (at 0) doesn't reference XPU device memory (cpu tensor?)

[dvrogozh]

@vlad-penkin : that's a reproducer associated with PyTorch PR #135567 which changes how pytorch returns tensor's data_ptr. The following test in Huggingface accelerate fails after this PR:

git clone https://github.com/huggingface/accelerate.git && cd accelerate
python3 -m pytest --pspec tests/test_utils.py::UtilsTester::test_dynamo

See details in pytorch/pytorch#135567 (review).

# AOT ID: ['0_forward']
from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, grid_combo_kernels, start_graph, end_graph
from torch._C import _xpu_getCurrentRawStream as get_raw_stream

aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()


# kernel path: /tmp/torchinductor_dvrogozh/hg/chgfabyslvbgrlhijrr6hjw4sjcxtg6f2q7tfxxuhivdvcf2rwij.py
# Topologically Sorted Source Nodes: [mul, add], Original ATen: [aten.mul, aten.add]
# Source node to ATen node mapping:
#   add => add
#   mul => mul
# Graph fragment:
#   %mul : [num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%primals_2, %primals_1), kwargs = {})
#   %add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul, %primals_3), kwargs = {})
triton_poi_fused_add_mul_0 = async_compile.triton('triton_', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor

from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties

@triton_heuristics.pointwise(
    size_hints=[64], 
    filename=__file__,
    triton_meta={'signature': {0: '*fp32', 1: 'fp32', 2: 'fp32', 3: '*fp32', 4: 'i32'}, 'device': DeviceProperties(type='xpu', index=0, cc={'driver_version': '1.3.30049', 'gpu_eu_count': 448, 'gpu_subslice_count': 56, 'has_atomic64': True, 'has_bfloat16_conversions': True, 'has_fp16': True, 'has_fp64': True, 'has_subgroup_2d_block_io': True, 'has_subgroup_matrix_multiply_accumulate': True, 'has_subgroup_matrix_multiply_accumulate_tensor_float32': False, 'max_compute_units': 448, 'max_num_sub_groups': 64, 'max_work_group_size': 1024, 'name': 'Intel(R) Data Center GPU Max 1100', 'platform_name': 'Intel(R) Level-Zero', 'sub_group_sizes': [16, 32], 'total_memory': 51539607552, 'type': 'gpu', 'vendor': 'Intel(R) Corporation', 'version': '1.3'}, major=None, regs_per_multiprocessor=None, max_threads_per_multi_processor=None, multi_processor_count=None), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 3), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_mul_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 0, 'backend_hash': '2D2AD0B55A7DC439E8979A8831462B3B9FC5AD26CB3B7F3AFDEFF5FB7681E7F1', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
    min_elem_per_thread=0
)
@triton.jit
def triton_(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
    xnumel = 40
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x0 = xindex
    tmp0 = tl.load(in_ptr0 + (x0), xmask)
    tmp1 = in_ptr1
    tmp3 = in_ptr2
    tmp2 = tmp0 * tmp1
    tmp4 = tmp2 + tmp3
    tl.store(out_ptr0 + (x0), tmp4, xmask)
''', device_str='xpu')


async_compile.wait(globals())
del async_compile

def call(args):
    primals_1, primals_2, primals_3 = args
    args.clear()
    assert_size_stride(primals_1, (), ())
    assert_size_stride(primals_2, (4, 10), (10, 1))
    assert_size_stride(primals_3, (), ())
    with torch.xpu._DeviceGuard(0):
        torch.xpu.set_device(0)
        buf0 = empty_strided_xpu((4, 10), (10, 1), torch.float32)
        # Topologically Sorted Source Nodes: [mul, add], Original ATen: [aten.mul, aten.add]
        stream0 = get_raw_stream(0)
        triton_poi_fused_add_mul_0.run(primals_2, primals_1.item(), primals_3.item(), buf0, 40, grid=grid(40), stream=stream0)
        del primals_1
        del primals_3
    return (buf0, primals_2, )


def benchmark_compiled_module(times=10, repeat=10):
    from torch._dynamo.testing import rand_strided
    from torch._inductor.utils import print_performance
    primals_1 = rand_strided((), (), device='cpu', dtype=torch.float32)
    primals_2 = rand_strided((4, 10), (10, 1), device='xpu:0', dtype=torch.float32)
    primals_3 = rand_strided((), (), device='cpu', dtype=torch.float32)
    fn = lambda: call([primals_1, primals_2, primals_3])
    return print_performance(fn, times=times, repeat=repeat)


if __name__ == "__main__":
    from torch._inductor.wrapper_benchmark import compiled_module_main
    compiled_module_main('None', benchmark_compiled_module)

[dvrogozh]

@vlad-penkin : I've debugged this on my side and posted PR #2192. The issue was that triton calls PyLong_AsLongLong() to cast device memory pointers instead of PyLong_AsVoidPtr(). This fix did work for my failing test both with and without @guangyey change pytorch/pytorch#135567 on pytorch side. I hope this won't regress any tests on your side. Let's see...