pytorch - 💡(How to fix) Fix torch.export in PyTorch 2.10 emits INT64_MAX instead of None for aten.slice.Tensor end bounds [1 participants]

pytorch2026-03-27 18:05:58

ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

GitHub issue URL

Helpful · Quick feedback

GitHub stats

pytorch/pytorch#178618•Fetched 2026-04-08 01:40:28

View on GitHub

Comments

Participants

Timeline

Reactions

Author

ali-nrn

Participants

ali-nrn

Assignees

tugsbayasgalan

Timeline (top)

mentioned ×9subscribed ×9labeled ×3cross-referenced ×2

Code Example

import torch
import torch.nn as nn
from torch.export import export, Dim

class Model(nn.Module):
    def forward(self, x, y):
        return x[::6, :], y

model = Model()
model.eval()
batch = Dim("batch", min=1, max=1000)
x = torch.randn(30, 6)
y = torch.randn(1000, 3)

prog = export(model, (x, y), dynamic_shapes={"x": {}, "y": {0: batch}})
graph_str = str(prog.graph_module.graph)
print(graph_str)

if "9223372036854775807" in graph_str:
    print("\nBUG: INT64_MAX (9223372036854775807) found in slice args")
else:
    print("\nOK: No INT64_MAX in slice args")

---

PyTorch version: 2.10.0+cpu
Is debug build: False
CUDA used to build PyTorch: Could not collect
ROCM used to build PyTorch: N/A

OS: Debian GNU/Linux 12 (bookworm) (x86_64)
GCC version: (Debian 12.2.0-14+deb12u1) 12.2.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.36

Python version: 3.11.2 (main, Apr 28 2025, 14:11:48) [GCC 12.2.0] (64-bit runtime)
Python platform: Linux-6.1.0-44-amd64-x86_64-with-glibc2.36
Is CUDA available: False
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: GPU 0: NVIDIA A16-4Q
Nvidia driver version: 570.133.20
cuDNN version: Could not collect
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

Versions of relevant libraries:
[pip3] numpy==2.3.4
[pip3] torch==2.10.0+cpu
[pip3] torchaudio==2.10.0+cpu
[pip3] torchvision==0.25.0+cpu

RAW_BUFFERClick to expand / collapse

🐛 Describe the bug

torch.export in PyTorch 2.10.0 emits aten.slice.Tensor with explicit 0 and 9223372036854775807 (INT64_MAX) as start/end arguments, instead of None which was used in PyTorch 2.8.0.

This is a regression from the behavior in PyTorch 2.8.0, where slice operations with step > 1 used None for unbounded start/end:

PyTorch 2.8.0: aten.slice.Tensor(x, 0, None, None, 6) PyTorch 2.10.0: aten.slice.Tensor(x, 0, 0, 9223372036854775807, 6)

While semantically equivalent in Python/PyTorch, the INT64_MAX value causes issues for downstream compilers (e.g., torch-mlir / MLIR pipelines) that materialize this constant as an index type — it overflows when the target uses 32-bit indices.

Minimal reproducer

import torch
import torch.nn as nn
from torch.export import export, Dim

class Model(nn.Module):
    def forward(self, x, y):
        return x[::6, :], y

model = Model()
model.eval()
batch = Dim("batch", min=1, max=1000)
x = torch.randn(30, 6)
y = torch.randn(1000, 3)

prog = export(model, (x, y), dynamic_shapes={"x": {}, "y": {0: batch}})
graph_str = str(prog.graph_module.graph)
print(graph_str)

if "9223372036854775807" in graph_str:
    print("\nBUG: INT64_MAX (9223372036854775807) found in slice args")
else:
    print("\nOK: No INT64_MAX in slice args")

Expected behavior (PyTorch 2.8.0)

slice_1 = torch.ops.aten.slice.Tensor(x, 0, None, None, 6)

None is used for unbounded start/end, which downstream compilers can correctly interpret as "slice to end."

Actual behavior (PyTorch 2.10.0)

slice_1 = torch.ops.aten.slice.Tensor(x, 0, 0, 9223372036854775807, 6)

Versions

PyTorch version: 2.10.0+cpu
Is debug build: False
CUDA used to build PyTorch: Could not collect
ROCM used to build PyTorch: N/A

OS: Debian GNU/Linux 12 (bookworm) (x86_64)
GCC version: (Debian 12.2.0-14+deb12u1) 12.2.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.36

Python version: 3.11.2 (main, Apr 28 2025, 14:11:48) [GCC 12.2.0] (64-bit runtime)
Python platform: Linux-6.1.0-44-amd64-x86_64-with-glibc2.36
Is CUDA available: False
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: GPU 0: NVIDIA A16-4Q
Nvidia driver version: 570.133.20
cuDNN version: Could not collect
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

Versions of relevant libraries:
[pip3] numpy==2.3.4
[pip3] torch==2.10.0+cpu
[pip3] torchaudio==2.10.0+cpu
[pip3] torchvision==0.25.0+cpu

cc @chauhang @penguinwu @avikchaudhuri @zhxchen17 @tugsbayasgalan @angelayi @suo @ydwu4

extent analysis

Fix Plan

To fix the regression issue in PyTorch 2.10.0, we need to modify the torch.export function to use None for unbounded start/end in slice operations.

Here are the steps:

Modify the torch.export function to check if the step size is greater than 1.
If the step size is greater than 1, use None for the end argument in the slice operation.
Update the export function to handle this change.

Example code:

import torch
import torch.nn as nn
from torch.export import export, Dim

# Define a custom export function
def custom_export(model, args, dynamic_shapes=None):
    # Get the graph module
    graph_module = export(model, args, dynamic_shapes=dynamic_shapes)
    
    # Modify the graph to use None for unbounded start/end
    for node in graph_module.graph.nodes:
        if node.kind() == 'aten::slice':
            # Check if the step size is greater than 1
            step_size = node.inputsAt(4)
            if step_size > 1:
                # Use None for the end argument
                node.inputsAt(3) = None
    
    return graph_module

class Model(nn.Module):
    def forward(self, x, y):
        return x[::6, :], y

model = Model()
model.eval()
batch = Dim("batch", min=1, max=1000)
x = torch.randn(30, 6)
y = torch.randn(1000, 3)

# Use the custom export function
prog = custom_export(model, (x, y), dynamic_shapes={"x": {}, "y": {0: batch}})
graph_str = str(prog.graph_module.graph)
print(graph_str)

if "9223372036854775807" in graph_str:
    print("\nBUG: INT64_MAX (9223372036854775807) found in slice args")
else:
    print("\nOK: No INT64_MAX in slice args")

Verification

To verify that the fix worked, run the modified code and check if the INT64_MAX value is still present in the graph string. If the fix is successful, the INT64_MAX value should be replaced with None.

Extra Tips

This fix assumes that the downstream compilers can correctly interpret None as "slice to end."
If the downstream compilers do not support None as an index type, additional modifications may be necessary.
It is recommended to test the modified code thoroughly to ensure that it works as expected in all scenarios.

Vote matrix · Quick signals

Works

Did the solution work? Tap to confirm.

Easy Fix

Was it a quick fix?

Time Saver

Did it save you time?

Blocking

Was it severely blocking?

Common Issue

Are others likely hitting this too?

Flaky / Intermittent

Is it intermittent?

Verified / Reproducible

Can you reproduce it reliably?

#parallel task #integration issue #index setup #retrieval issue #search optimization

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Data

Security

Network

Code

UI/UX

Text

System

Multimedia

Protocol

API

Engineering

pytorch - 💡(How to fix) Fix torch.export in PyTorch 2.10 emits INT64_MAX instead of None for aten.slice.Tensor end bounds [1 participants]

Recommended Tools

GitHub issue graph ai analysis

Code Example

🐛 Describe the bug

Versions

extent analysis

Fix Plan

Verification

Extra Tips

Still need to ship something?

TRENDING

pytorch - 💡(How to fix) Fix torch.export in PyTorch 2.10 emits INT64_MAX instead of None for aten.slice.Tensor end bounds [1 participants]

Recommended Tools

GitHub issue graph ai analysis

Code Example

🐛 Describe the bug

Versions

extent analysis

Fix Plan

Verification

Extra Tips

Still need to ship something?

RELATED_DISCOVERY

TRENDING