Expanding subset of tensor reads wrong memory

[martenlienen]

🐛 Describe the bug

I have derived the following minimal failing example:

import torch

def expand(x, n):
    return x.expand((n,))

@torch.compile()
def f(n: int, device: str):
    numbers = torch.arange(10, device=device)
    for i in range(len(numbers)):
        expanded = expand(numbers[i], n)
        print(expanded[0])

device = "cuda"
f(1, device)
print()
f(2, device)

This should print the integers from 0 to 9 twice, but what you get instead is

tensor(0, device='cuda:0')
tensor(1, device='cuda:0')
tensor(2, device='cuda:0')
tensor(3, device='cuda:0')
tensor(4, device='cuda:0')
tensor(5, device='cuda:0')
tensor(6, device='cuda:0')
tensor(7, device='cuda:0')
tensor(8, device='cuda:0')
tensor(9, device='cuda:0')

tensor(0, device='cuda:0')
tensor(0, device='cuda:0')
tensor(2, device='cuda:0')
tensor(0, device='cuda:0')
tensor(4, device='cuda:0')
tensor(0, device='cuda:0')
tensor(6, device='cuda:0')
tensor(0, device='cuda:0')
tensor(8, device='cuda:0')
tensor(0, device='cuda:0')

The specific values of n are not important, only that they differ. If you use a linspace instead of an arange, the pattern is different. Then it prints the first value of the linspace in every iteration except every 5th, where it prints the correct value (at least with dtype=torch.float32). If I inline the definition of expand, the bug disappears. It only happens on CUDA devices. If you set device = "cpu", it does not happen. If you don't compile f, it also does not happen. If we .clone() numbers[i], it also does not happen.

While this example prints to show the bug, I have also observed it without print in my sampling code (only every 5th generated sample was not trash).

Error logs

dedicated_log_torch_trace_1jeap82o.log

tl_out.tar.gz

Versions

I have confirmed this bug on 2.5.1, 2.6.0 and today's nightly.

Collecting environment information...
PyTorch version: N/A
Is debug build: N/A
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.5 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.22.1
Libc version: glibc-2.35

Python version: 3.10.12 (main, Jan 17 2025, 14:35:34) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.15.0-131-generic-x86_64-with-glibc2.35
Is CUDA available: N/A
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: 
GPU 0: NVIDIA A100-SXM4-80GB
  MIG 3g.40gb     Device  0:

Nvidia driver version: 535.230.02
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: N/A

CPU:
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        46 bits physical, 57 bits virtual
Byte Order:                           Little Endian
CPU(s):                               96
On-line CPU(s) list:                  0-95
Vendor ID:                            GenuineIntel
Model name:                           Intel(R) Xeon(R) Gold 6336Y CPU @ 2.40GHz
CPU family:                           6
Model:                                106
Thread(s) per core:                   2
Core(s) per socket:                   24
Socket(s):                            2
Stepping:                             6
CPU max MHz:                          2400.0000
CPU min MHz:                          800.0000
BogoMIPS:                             4800.00
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 ds_cpl smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 invpcid_single intel_ppin ssbd mba ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect wbnoinvd dtherm arat pln pts avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid fsrm md_clear pconfig flush_l1d arch_capabilities
L1d cache:                            2.3 MiB (48 instances)
L1i cache:                            1.5 MiB (48 instances)
L2 cache:                             60 MiB (48 instances)
L3 cache:                             72 MiB (2 instances)
NUMA node(s):                         4
NUMA node0 CPU(s):                    0-11,48-59
NUMA node1 CPU(s):                    12-23,60-71
NUMA node2 CPU(s):                    24-35,72-83
NUMA node3 CPU(s):                    36-47,84-95
Vulnerability Gather data sampling:   Mitigation; Microcode
Vulnerability Itlb multihit:          Not affected
Vulnerability L1tf:                   Not affected
Vulnerability Mds:                    Not affected
Vulnerability Meltdown:               Not affected
Vulnerability Mmio stale data:        Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed:               Not affected
Vulnerability Spec rstack overflow:   Not affected
Vulnerability Spec store bypass:      Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:             Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:             Mitigation; Enhanced / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS SW sequence; BHI SW loop, KVM SW loop
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

Versions of relevant libraries:
[pip3] Could not collect
[conda] Could not collect

cc @chauhang @penguinwu @ezyang @bobrenjc93

[eellison]

This does not repro with dynamic=True. Maybe cc @bobrenjc93 or @laithsakka any interest ?

import torch
device = "cuda"


def expand(x, n):
    return x.expand((n,))

def f(n: int, device: str):
    numbers = torch.arange(10, device=device)
    out = []
    for i in range(len(numbers)):
        expanded = expand(numbers[i], n)
        out.append(expanded[0])
        torch._dynamo.graph_break()
    return out

def test_equal():
    torch._dynamo.reset()

    f_c = torch.compile(dynamic=True, backend="inductor")(f)

    try:
        torch.testing.assert_close(f_c(2, "cuda"), f(2, "cuda"))
    except:
        return False
    return True

from torch._inductor.compiler_bisector import CompilerBisector
print(CompilerBisector.do_bisect(test_equal))

[martenlienen]

The important part is to call f_c with two different values for n. Your example uses 2 twice.

[eellison]

The two different values for n triggers dynamic. See:

import torch

def expand(x, n):
    return x.expand((n,))

@torch.compile(dynamic=True)
def f(n: int, device: str):
    numbers = torch.arange(10, device=device)
    for i in range(len(numbers)):
        expanded = expand(numbers[i], n)
        print(expanded[0])

device = "cuda"
f(2, device)

tensor(0, device='cuda:0')
tensor(0, device='cuda:0')
tensor(2, device='cuda:0')
tensor(0, device='cuda:0')
tensor(4, device='cuda:0')
tensor(0, device='cuda:0')
tensor(6, device='cuda:0')
tensor(0, device='cuda:0')
tensor(8, device='cuda:0')
tensor(0, device='cuda:0')

[martenlienen]

Ah, thanks for the explanation.

[aishwaryar12309]

for n=2, isn't it supposed to produce the same output (first elem of expanded) as n=1? I am just confused because when I try it with dynamic=True, I am still not getting that answer and it seems to be printing only the even indices. I was only able to get the expected answer for n=2 by allocating memory to the scalars

def expand(x, n):
    return x.expand((n,))

@torch.compile(dynamic=True)
def f(n: int, device: str):
    numbers = torch.arange(10, device=device)
    for i in range(len(numbers)):
        scalar = numbers[i].clone()
        expanded = scalar.expand((n,))
        print(expanded[0])
        print(expanded)

device = "cuda"
print()
f(2, device)

tensor(0, device='cuda:0')
tensor([0, 0], device='cuda:0')
tensor(1, device='cuda:0')
tensor([1, 1], device='cuda:0')
tensor(2, device='cuda:0')
tensor([2, 2], device='cuda:0')
tensor(3, device='cuda:0')
tensor([3, 3], device='cuda:0')
tensor(4, device='cuda:0')
tensor([4, 4], device='cuda:0')
tensor(5, device='cuda:0')
tensor([5, 5], device='cuda:0')
tensor(6, device='cuda:0')
tensor([6, 6], device='cuda:0')
tensor(7, device='cuda:0')
tensor([7, 7], device='cuda:0')
tensor(8, device='cuda:0')
tensor([8, 8], device='cuda:0')
tensor(9, device='cuda:0')
tensor([9, 9], device='cuda:0')

[martenlienen]

I have expanded the script a bit to highlight the dependence on the datatype.

import torch

def expand(x, n):
    return x.expand((n,))

@torch.compile(dynamic=True)
def f(n: int, dtype: torch.dtype, device: str):
    numbers = torch.arange(18, dtype=dtype, device=device)
    out = []
    for i in range(len(numbers)):
        out.append(expand(numbers[i], n)[0])
        torch._dynamo.graph_break()
    return out

device = "cuda"
for dtype in [torch.int8, torch.int16, torch.int32, torch.int64, torch.float16, torch.float32, torch.float64]:
    print(f"{dtype!s: >13}: {' '.join([f'{int(x.item()): >2d}' for x in f(2, dtype, device)])}")

This produces

   torch.int8:  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 16  0
  torch.int16:  0  0  0  0  0  0  0  0  8  0  0  0  0  0  0  0 16  0
  torch.int32:  0  0  0  0  4  0  0  0  8  0  0  0 12  0  0  0 16  0
  torch.int64:  0  0  2  0  4  0  6  0  8  0 10  0 12  0 14  0 16  0
torch.float16:  0  0  0  0  0  0  0  0  8  0  0  0  0  0  0  0 16  0
torch.float32:  0  0  0  0  4  0  0  0  8  0  0  0 12  0  0  0 16  0
torch.float64:  0  0  2  0  4  0  6  0  8  0 10  0 12  0 14  0 16  0

[bobrenjc93]

This does not repro with dynamic=True. Maybe cc @bobrenjc93 or @laithsakka any interest ?

You mean dynamic=False right? I tried dynamic=True and it does reproduce, dynamic=False doesn't. Taking a look now.

import torch

def expand(x, n):
    return x.expand((n,))

@torch.compile(dynamic=False)
def f(n: int, device: str):
    numbers = torch.arange(10, device=device)
    for i in range(len(numbers)):
        expanded = expand(numbers[i], n)
        print(expanded[0])

device = "cuda"
f(1, device)
print()
f(2, device)