comfy-cli install --cpu still installs CUDA version of PyTorch

[Xyc2016]

Describe the bug
When running comfy-cli install with the --cpu flag, it still installs the CUDA version of PyTorch instead of the CPU-only version.

To Reproduce
Steps to reproduce the behavior:

set -ex

rm -rf ~/comfy
rm -rf learn-comfy

uv init --python=3.12 learn-comfy
cd learn-comfy

uv add comfy-cli==1.5.3 pip==25.3

uv run comfy --skip-prompt install --cpu --skip-manager

uv pip show torch

Expected behavior
PyTorch should be installed as the CPU-only version from https://download.pytorch.org/whl/cpu, not the CUDA version.

Example output

Installing for CPU
Installing from repository 'https://github.com/comfyanonymous/ComfyUI' to '/home/xyc/comfy/ComfyUI'
...
Skipping installation of ComfyUI-Manager. (by --skip-manager)
...
ComfyUI is installed at: /home/xyc/comfy/ComfyUI
+ uv pip show torch
Name: torch
Version: 2.9.1
Location: /home/xyc/learn/learn-comfy/.venv/lib/python3.12/site-packages
Requires: filelock, fsspec, jinja2, networkx, nvidia-cublas-cu12, nvidia-cuda-cupti-cu12, ...

Additional context

• Python 3.12.3
• uv 0.9.21
• comfy-cli 1.5.3
• Ubuntu 24.04

[dosubot]

Hi @Xyc2016! I'm Dosu and I’m helping the comfy-cli team.

This is expected behavior in comfy-cli right now: the --cpu flag only prints "Installing for CPU" but does not actually change how PyTorch is installed. It does not trigger installation of the CPU-only PyTorch wheel; instead, the default (often CUDA-enabled) version is installed regardless of the flag. The --cpu flag is meant for launch-time options (e.g., comfy launch -- --cpu), not for dependency selection during install (source).

To ensure you have the CPU-only version of PyTorch, run these commands after installation:

pip uninstall torch torchvision torchaudio -y
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

Be aware that future updates or dependency reinstalls may overwrite this, so you may need to repeat the process if you update or reinstall dependencies (source).

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

   

[Xyc2016]

I found the issue in the following code:

def pip_install_comfyui_dependencies(
    repo_dir,
    gpu: GPU_OPTION,
    plat: constants.OS,
    cuda_version: constants.CUDAVersion,
    skip_torch_or_directml: bool,
    skip_requirement: bool,
):
    os.chdir(repo_dir)

    result = None
    if not skip_torch_or_directml:
        # install torch for AMD Linux
        if gpu == GPU_OPTION.AMD and plat == constants.OS.LINUX:
            pip_url = ["--extra-index-url", "https://download.pytorch.org/whl/rocm6.0"]
            result = subprocess.run(
                [
                    sys.executable,
                    "-m",
                    "pip",
                    "install",
                    "torch",
                    "torchvision",
                    "torchaudio",
                ]
                + pip_url,
                check=False,
            )

        # install torch for NVIDIA
        if gpu == GPU_OPTION.NVIDIA:
            base_command = [
                sys.executable,
                "-m",
                "pip",
                "install",
                "torch",
                "torchvision",
                "torchaudio",
            ]

            if plat == constants.OS.WINDOWS and cuda_version == constants.CUDAVersion.v12_9:
                base_command += [
                    "--extra-index-url",
                    "https://download.pytorch.org/whl/cu129",
                ]
            elif plat == constants.OS.WINDOWS and cuda_version == constants.CUDAVersion.v12_6:
                base_command += [
                    "--extra-index-url",
                    "https://download.pytorch.org/whl/cu126",
                ]
            elif plat == constants.OS.WINDOWS and cuda_version == constants.CUDAVersion.v12_4:
                base_command += [
                    "--extra-index-url",
                    "https://download.pytorch.org/whl/cu124",
                ]
            elif plat == constants.OS.WINDOWS and cuda_version == constants.CUDAVersion.v12_1:
                base_command += [
                    "--extra-index-url",
                    "https://download.pytorch.org/whl/cu121",
                ]
            elif plat == constants.OS.WINDOWS and cuda_version == constants.CUDAVersion.v11_8:
                base_command += [
                    "--extra-index-url",
                    "https://download.pytorch.org/whl/cu118",
                ]
            result = subprocess.run(
                base_command,
                check=False,
            )
        # Update installation to use upstream torch xpu. ipex is no longer needed for Intel Arc GPUs
        # https://github.com/comfyanonymous/ComfyUI/pull/7767
        if gpu == GPU_OPTION.INTEL_ARC:
            pip_url = [
                "--extra-index-url",
                "https://download.pytorch.org/whl/xpu",
            ]

            # TODO: wrap pip install in a function
            result = subprocess.run(
                [
                    sys.executable,
                    "-m",
                    "pip",
                    "install",
                    "torch",
                    "torchvision",
                    "torchaudio",
                ]
                + pip_url,
                check=False,
            )
        if result and result.returncode != 0:
            rprint("Failed to install PyTorch dependencies. Please check your environment (`comfy env`) and try again")
            sys.exit(1)

        # install directml for AMD windows
        if gpu == GPU_OPTION.AMD and plat == constants.OS.WINDOWS:
            result = subprocess.run([sys.executable, "-m", "pip", "install", "torch-directml"], check=True)

        # install torch for Mac M Series
        if gpu == GPU_OPTION.MAC_M_SERIES:
            result = subprocess.run(
                [
                    sys.executable,
                    "-m",
                    "pip",
                    "install",
                    "--pre",
                    "torch",
                    "torchvision",
                    "torchaudio",
                    "--extra-index-url",
                    "https://download.pytorch.org/whl/nightly/cpu",
                ],
                check=True,
            )

    # install requirements.txt
    if skip_requirement:
        return
    result = subprocess.run([sys.executable, "-m", "pip", "install", "-r", "requirements.txt"], check=False)
    if result.returncode != 0:
        rprint("Failed to install ComfyUI dependencies. Please check your environment (`comfy env`) and try again.")
        sys.exit(1)

The problem is that there is no handling for gpu=None, i.e., the CPU-only case. Currently, the code only handles AMD, NVIDIA, Intel Arc, Mac M Series, and installs the corresponding CUDA/ROCm/XPU packages, but if gpu is None, it will still go through the NVIDIA branch and install the CUDA version of PyTorch.

I would like to propose a fix that properly handles the CPU-only installation.

[dosubot]

You're right—there's currently no branch in pip_install_comfyui_dependencies for the CPU-only case, so when gpu=None, the installer defaults to the NVIDIA (CUDA) logic and installs CUDA-enabled PyTorch wheels, even on CPU-only systems. This is confirmed by both the code and user reports that comfy install --cpu still pulls in all NVIDIA dependencies and fails on systems without a GPU (source).

To fix this, add a branch for gpu is None (or gpu == GPU_OPTION.CPU) that installs PyTorch using:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

This will ensure only the CPU wheels are installed. Make sure this branch runs before the NVIDIA logic, and consider how requirements.txt is handled—if it includes torch, it could overwrite your CPU-only install, so you may want to install torch after requirements or pin the version as needed (code reference).

Testing the fix is straightforward: run comfy-cli install --cpu and verify that pip show torch lists no CUDA dependencies. This change would align the install process with user expectations and avoid unnecessary GPU dependencies on CPU-only setups.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}

   

[Xyc2016]

#345