pytorch - ✅(Solved) Fix [MPS] torch.conj(x).imag returns un-conjugated imag on conjugate view (CPU correct) [1 pull requests, 1 comments, 2 participants]

pytorch2026-05-19 14:36:58

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#184379•Fetched 2026-05-20 03:39:00

View on GitHub

Comments

Participants

Timeline

Reactions

Author

Participants

Assignees

Timeline (top)

mentioned ×11subscribed ×11labeled ×6assigned ×1

Fix Action

Fixed

Fixed by PR: [MPS] Materialize neg bit in copy_kernel_mps (https://github.com/pytorch/pytorch/pull/184403)

PR fix notes

PR #184403: [MPS] Materialize neg bit in copy_kernel_mps

Repository: pytorch/pytorch
Author: malfet
State: open | merged: False
Link: https://github.com/pytorch/pytorch/pull/184403

Description (problem / solution / changelog)

Stack from ghstack (oldest at bottom):

-> #184403

MPS copy kernel handles is_conj but not is_neg, so cloning or printing a tensor with the negative bit set (x.conj().imag for example) returns the un-negated underlying data.

Fix by mirroring existing conjugation handling logic Fixes https://github.com/pytorch/pytorch/issues/184379

Authored with Claude

Changed files

aten/src/ATen/native/mps/operations/Copy.mm (modified, +13/-5)
test/test_mps.py (modified, +18/-0)

Code Example

import torch

x = torch.complex(
    torch.tensor([1.0, 3.0], device="mps"),
    torch.tensor([2.0, 4.0], device="mps"),
)
c = torch.conj(x)

print("is_conj():       ", c.is_conj())              # True
print("c (repr):        ", c)                        # tensor([1.-2.j, 3.-4.j], device='mps:0')  (resolved for display)
print("c.real:          ", c.real)                   # tensor([1., 3.])   correct
print("c.imag:          ", c.imag)                   # tensor([2., 4.])   ← WRONG, expected [-2., -4.]
print("c.resolve_conj().imag:", c.resolve_conj().imag)  # tensor([-2., -4.])   correct
print("CPU baseline:    ", torch.conj(x.cpu()).imag) # tensor([-2., -4.])   correct

---

is_conj():        True
c (repr):         tensor([1.-2.j, 3.-4.j], device='mps:0')
c.real:           tensor([1., 3.], device='mps:0')
c.imag:           tensor([2., 4.], device='mps:0')
c.resolve_conj().imag: tensor([-2., -4.], device='mps:0')
CPU baseline:     tensor([-2., -4.])

---

PyTorch version: 2.10.0
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: macOS 26.4.1 (arm64)
GCC version: Could not collect
Clang version: 21.0.0 (clang-2100.1.1.101)
CMake version: version 4.3.2
Libc version: N/A

Python version: 3.13.13 (main, Apr  7 2026, 18:19:01) [Clang 21.0.0 (clang-2100.0.123.102)] (64-bit runtime)
Python platform: macOS-26.4.1-arm64-arm-64bit-Mach-O
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
Caching allocator config: N/A

CPU:
Apple M3 Max

Versions of relevant libraries:
[pip3] diffractorch==0.2.0
[pip3] mypy_extensions==1.1.0
[pip3] numpy==2.4.3
[pip3] torch==2.10.0
[pip3] torchvision==0.25.0
[conda] Could not collect

RAW_BUFFERClick to expand / collapse

🐛 Describe the bug

On the MPS backend, .imag accessed on the lazy conjugate view returned by torch.conj() ignores the conjugate bit and returns the original imaginary part instead of its negation. The same code is correct on CPU, and calling .resolve_conj() first restores correctness — so the bug is specifically in the .imag accessor's handling of is_conj() views on MPS.

.real is not affected (it's the same with or without conjugation, so it doesn't expose the bug), and the printed repr of the conjugate view goes through a materialization path that displays the correct values — which makes this easy to miss in interactive use.

Minimal repro:

import torch

x = torch.complex(
    torch.tensor([1.0, 3.0], device="mps"),
    torch.tensor([2.0, 4.0], device="mps"),
)
c = torch.conj(x)

print("is_conj():       ", c.is_conj())              # True
print("c (repr):        ", c)                        # tensor([1.-2.j, 3.-4.j], device='mps:0')  (resolved for display)
print("c.real:          ", c.real)                   # tensor([1., 3.])   correct
print("c.imag:          ", c.imag)                   # tensor([2., 4.])   ← WRONG, expected [-2., -4.]
print("c.resolve_conj().imag:", c.resolve_conj().imag)  # tensor([-2., -4.])   correct
print("CPU baseline:    ", torch.conj(x.cpu()).imag) # tensor([-2., -4.])   correct

Output (torch 2.10.0, macOS 26.4.1, Apple M3 Max):

is_conj():        True
c (repr):         tensor([1.-2.j, 3.-4.j], device='mps:0')
c.real:           tensor([1., 3.], device='mps:0')
c.imag:           tensor([2., 4.], device='mps:0')
c.resolve_conj().imag: tensor([-2., -4.], device='mps:0')
CPU baseline:     tensor([-2., -4.])

Versions

PyTorch version: 2.10.0
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: macOS 26.4.1 (arm64)
GCC version: Could not collect
Clang version: 21.0.0 (clang-2100.1.1.101)
CMake version: version 4.3.2
Libc version: N/A

Python version: 3.13.13 (main, Apr  7 2026, 18:19:01) [Clang 21.0.0 (clang-2100.0.123.102)] (64-bit runtime)
Python platform: macOS-26.4.1-arm64-arm-64bit-Mach-O
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
Caching allocator config: N/A

CPU:
Apple M3 Max

Versions of relevant libraries:
[pip3] diffractorch==0.2.0
[pip3] mypy_extensions==1.1.0
[pip3] numpy==2.4.3
[pip3] torch==2.10.0
[pip3] torchvision==0.25.0
[conda] Could not collect

cc @ezyang @anjali411 @dylanbespalko @mruberry @nikitaved @amjames @kulinseth @malfet @DenisVieriu97 @jhavukainen @aditvenk

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?

#conversation history #tool integration #LLM response #prompt template #agent execution

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 - ✅(Solved) Fix [MPS] torch.conj(x).imag returns un-conjugated imag on conjugate view (CPU correct) [1 pull requests, 1 comments, 2 participants]

Recommended Tools

GitHub issue graph ai analysis

Fix Action

Fixed

PR fix notes

PR #184403: [MPS] Materialize neg bit in copy_kernel_mps

Description (problem / solution / changelog)

Changed files

Code Example

🐛 Describe the bug

Versions

Still need to ship something?

TRENDING

pytorch - ✅(Solved) Fix [MPS] torch.conj(x).imag returns un-conjugated imag on conjugate view (CPU correct) [1 pull requests, 1 comments, 2 participants]

Recommended Tools

GitHub issue graph ai analysis

Fix Action

Fixed

PR fix notes

PR #184403: [MPS] Materialize neg bit in copy_kernel_mps

Description (problem / solution / changelog)

Changed files

Code Example

🐛 Describe the bug

Versions

Still need to ship something?

RELATED_DISCOVERY

TRENDING