Avoid double-counting CPU RAM for integrated CUDA

[arpansahu]

Problem

get_max_memory() adds a separate "cpu" memory entry even when CUDA reports an integrated GPU. Integrated CUDA devices share host RAM, so downstream device-map planning can treat one physical memory pool as independent CUDA and CPU capacities. Closes #4183.

Reproducer

from types import SimpleNamespace
from unittest.mock import patch
import torch
from accelerate.utils import modeling
from accelerate.utils.modeling import get_max_memory

with patch.object(modeling, "is_mps_available", return_value=False), \
     patch.object(torch.cuda, "device_count", return_value=1), \
     patch.object(torch.cuda, "mem_get_info", return_value=(1234, 5678)), \
     patch.object(torch.cuda, "get_device_properties", return_value=SimpleNamespace(is_integrated=True)), \
     patch.object(torch, "tensor"):
    print(get_max_memory())
# before: {0: 1234, "cpu": ...}
# after:  {0: 1234}

Fix

Track whether any available CUDA device reports is_integrated and skip adding a separate CPU memory budget in that case. The attribute is read with getattr(..., False) so older PyTorch/device properties keep the existing behavior.

Testing

Added tests/test_modeling_utils.py::ModelingUtilsTester::test_get_max_memory_integrated_cuda_does_not_add_cpu.

Before the fix:

FAILED tests/test_modeling_utils.py::ModelingUtilsTester::test_get_max_memory_integrated_cuda_does_not_add_cpu
1 failed in 7.37s

After the fix:

2 passed, 1 skipped in 6.56s
7 passed, 6 skipped in 69.63s (0:01:09)

Quality:

ruff check ... --fix: All checks passed!
ruff format ...: 2 files left unchanged
ruff check ...: All checks passed!
ruff format --check ...: 2 files already formatted

[arpansahu]

Disclosure: this contribution was prepared with AI assistance (GitHub Copilot CLI), and
I am flagging that explicitly rather than leaving you to guess.

The defect, the fix and the regression test were verified locally: the new test fails on
unmodified upstream and passes with the change applied, and the surrounding suite is green.

If your project would prefer not to take AI-assisted contributions, or you would rather this
were reworked and resubmitted by hand, please just close it -- no objection at all, and
apologies for the noise.