Added torchaudio installation to setup.py

[tzielinski-habana]

This pull request introduces changes to the setup.py installation logic, specifically addressing the handling of the torchaudio dependency to avoid inadvertently installing CUDA-enabled PyTorch when only CPU support is desired. The changes ensure a safer and more predictable installation process for users.

Dependency management improvements:

• Excluded torchaudio from install_requires in setup.py, as its installation requires special handling to avoid pulling CUDA torch dependencies.
• Added logic to install torchaudio separately using pip install --no-deps with the correct version matching the installed torch, and only if not running in metadata generation mode (dist_info or egg_info).

We need torchaudio because it's imported from upstream vllm as a result of this PR:
vllm-project/vllm#33247

[Copilot]

Pull request overview

This PR adjusts setup.py dependency handling to avoid pulling CUDA-enabled PyTorch when adding torchaudio, by removing torchaudio from install_requires and attempting to install it separately with pip --no-deps using a version derived from the installed torch.

Changes:

• Filter torchaudio out of install_requires derived from requirements.txt.
• Add post-setup() logic that imports torch, derives a torchaudio==x.y.z pin from torch.__version__, and runs pip install --no-deps against the PyTorch CPU wheel index.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The dist_info/egg_info argv guard is not sufficient to prevent this from running during packaging operations: wheel builds typically invoke bdist_wheel (and possibly sdist/build), so this block will still run while producing artifacts. If this logic remains, it should be gated so it only runs for the final install into the target environment (not during build/metadata/wheel creation).

Suggested change

	# Skipped during metadata generation (dist_info / egg_info).
	if "dist_info" not in sys.argv and "egg_info" not in sys.argv:
	# Skipped during metadata generation and build/package creation.
	if not any(cmd in sys.argv for cmd in ("dist_info", "egg_info", "bdist_wheel", "sdist", "build")):

[Copilot]

re.match(...).group(1) will raise an AttributeError if torch.__version__ doesn’t match the expected x.y.z format (e.g., unusual local builds). This would fail the package install even if torchaudio is otherwise optional. Handle the no-match case explicitly (raise a clear error or fall back to a safer version parsing approach, e.g., via packaging.version).

Suggested change

	ver = re.match(r"(\d+\.\d+\.\d+)", torch.__version__).group(1)
	version_match = re.match(r"(\d+\.\d+\.\d+)", torch.__version__)
	if version_match is None:
	raise RuntimeError(
	f"Unable to parse torch version from '{torch.__version__}' "
	"to determine the matching torchaudio version."
	)
	ver = version_match.group(1)

[Copilot]

The torchaudio install runs unconditionally and will re-run on every invocation, potentially downgrading/upgrading an existing torchaudio install. Consider first checking whether torchaudio is already installed and whether its version matches the desired torch version before calling pip.

[Copilot]

Filtering requirements via startswith("torchaudio") can accidentally exclude unrelated packages (e.g., torchaudio-foo) and won’t handle some valid requirement formats reliably. If torchaudio ever appears in requirements files, consider parsing each line as a PEP 508 requirement and filtering by normalized project name instead of a string prefix.

[Copilot]

Running pip install via subprocess.check_call inside setup.py introduces build-time side effects (network access, mutating the build env) under the PEP 517 setuptools.build_meta backend. This will execute during wheel builds (e.g., bdist_wheel) and can break reproducible/offline builds and CI. Prefer moving this to an explicit install step (docs or a dedicated installer script/extra) rather than performing dependency installation from setup.py.

Suggested change

	subprocess.check_call([
	sys.executable,
	"-m",
	"pip",
	"install",
	"--no-deps",
	"--extra-index-url",
	"https://download.pytorch.org/whl/cpu",
	f"torchaudio=={ver}",
	])
	# Building/installing this package should not perform network operations or
	# mutate the environment (e.g., by running `pip install`). Instead, tell the
	# user how to install a matching torchaudio version explicitly.
	cmd = (
	f"{sys.executable} -m pip install --no-deps --extra-index-url "
	"https://download.pytorch.org/whl/cpu "
	f"torchaudio=={ver}"
	)
	raise RuntimeError(
	"torchaudio is required but is not installed.\n\n"
	"To install a CPU-only torchaudio build matching your torch version, run:\n\n"
	f" {cmd}\n\n"
	"Note: This command must be run explicitly by the user; it is not executed\n"
	"automatically during the build to keep builds reproducible and side-effect free."
	)

[github-actions]

🚧 CI Blocked

The main CI workflow was not started for the following reason:

Your branch is behind the base branch. Please merge or rebase to get the latest changes.