Prebuilt ROCm tarballs package for Windows to install?

[Looong01]

I notice that GIthub Action build of Prebuilt tarballs and ROCm Python packages for Windows platform is passing, but I can only see Linux version here(https://github.com/ROCm/TheRock/blob/main/RELEASES.md#installing-from-tarballs).

Anyone knows how to install C++ Prebuilt tarballs package for Windows?

[ScottTodd]

The https://therock-nightly-tarball.s3.amazonaws.com/ page does have therock-dist-windows-* files. We still need to generate a nice index page for that S3 bucket though: #1131

[RSabbagh52]

Are the files actually there? I just see a large xml file. Is there a pre-built windows version available that is later than the May version ? I seem to keep going in circles... Thanks in advance!

[ScottTodd]

Low tech way:

1. Search the .xml file, find the latest:
   

   Probably want to search for your specific gfx family, like how mine is 1100 (included in 110X):
   

2. Append that URL to the page, like https://therock-nightly-tarball.s3.amazonaws.com/therock-dist-windows-gfx120X-all-7.0.0rc20250806.tar.gz , visit and download

Sorry it's so awkward right now - we definitely plan on adding a proper index page (and user-friendly installers, instead of machine-friendly .tar.gz files)

[RSabbagh52]

Thanks Scott! I just downloaded it. I have a Strix Halo in the gmktec evo-x2. I downloaded the latest gfx1151. I'll let you know how it goes...

[RSabbagh52]

Hi Scott, I opened the compressed file, but not to sure how to install it. I sure I'll figure it out, but if you could give me a little push?

[ScottTodd]

https://github.com/ROCm/TheRock/blob/main/RELEASES.md#using-installed-tarballs

If you want to use the Windows tarballs like how the HIP SDK is used, you could try setting environment variables like HIP_PATH. Though note that this has not been tested much - the current focus has been on testing the Python packages together with PyTorch. General support for other projects and workflows may take some additional engineering work to enable.

[RSabbagh52]

Hi Scott, Sorry to interrupt you so much. I thought this would be the pre-built wheel files. I want to set it up for ai use with comfyui.

…

On Wed, Aug 6, 2025, 2:38 PM Scott Todd ***@***.***> wrote: https://github.com/ROCm/TheRock/blob/main/RELEASES.md#using-installed-tarballs If you want to use the Windows tarballs like how the HIP SDK is used, you could try setting environment variables like HIP_PATH. Though note that this has not been tested much - the current focus has been on testing the Python packages together with PyTorch. General support for other projects and workflows may take some additional engineering work to enable. — Reply to this email directly, view it on GitHub <#1182 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ANE7KUBJ55IUTQ5NI3XR5JD3MJYUPAVCNFSM6AAAAACC63LK3WVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTIMBSGU3DSMI> . You are receiving this because you commented.Message ID: ***@***.***>

[ScottTodd]

Wheel files use https://github.com/ROCm/TheRock/blob/main/RELEASES.md#installing-pytorch-python-packages

You can see my example setup commands here: https://gist.github.com/ScottTodd/f7c72202a3e160baffd715d006238d4a, note the python -m pip install from a virtual environment

[RSabbagh52]

Hi Scott,
I have the new Pytorch files loaded in the environment. The issue I'm having now is starting up comfyui. I didn't think anything would have to change there but it is not starting up. I'm really excited to test the early releases. I'm a long time c++ procrammer but very new to Python so please bare with me. I know your very busy but any help would be appreciated. Here is the output when starting comfyui:

(comfyui-env) C:\Users\Ray\ComfyUI>python main.py
Checkpoint files will always be loaded safely.
HIP Library Path: C:\Users\Ray\miniconda3\envs\comfyui-env\Lib\site-packages_rocm_sdk_core\bin\amdhip64_7.dll
Total VRAM 89977 MB, total RAM 65175 MB
pytorch version: 2.9.0a0+rocm7.0.0rc20250807
AMD arch: gfx1151
ROCm version: (7, 0)
Set vram state to: NORMAL_VRAM
Device: cuda:0 AMD Radeon(TM) 8060S Graphics : native
Using pytorch attention
Traceback (most recent call last):
File "C:\Users\Ray\ComfyUI\main.py", line 147, in
import execution
File "C:\Users\Ray\ComfyUI\execution.py", line 16, in
import nodes
File "C:\Users\Ray\ComfyUI\nodes.py", line 24, in
import comfy.diffusers_load
File "C:\Users\Ray\ComfyUI\comfy\diffusers_load.py", line 3, in
import comfy.sd
File "C:\Users\Ray\ComfyUI\comfy\sd.py", line 29, in
from . import model_detection
File "C:\Users\Ray\ComfyUI\comfy\model_detection.py", line 2, in
import comfy.supported_models
File "C:\Users\Ray\ComfyUI\comfy\supported_models.py", line 5, in
from . import sd1_clip
File "C:\Users\Ray\ComfyUI\comfy\sd1_clip.py", line 3, in
from transformers import CLIPTokenizer
File "C:\Users\Ray\miniconda3\envs\comfyui-env\Lib\site-packages\transformers_init_.py", line 27, in
from . import dependency_versions_check
File "C:\Users\Ray\miniconda3\envs\comfyui-env\Lib\site-packages\transformers\dependency_versions_check.py", line 16, in
from .utils.versions import require_version, require_version_core
File "C:\Users\Ray\miniconda3\envs\comfyui-env\Lib\site-packages\transformers\utils_init_.py", line 24, in
from .auto_docstring import (
File "C:\Users\Ray\miniconda3\envs\comfyui-env\Lib\site-packages\transformers\utils\auto_docstring.py", line 30, in
from .generic import ModelOutput
File "C:\Users\Ray\miniconda3\envs\comfyui-env\Lib\site-packages\transformers\utils\generic.py", line 55, in
from ..model_debugging_utils import model_addition_debugger_context
File "C:\Users\Ray\miniconda3\envs\comfyui-env\Lib\site-packages\transformers\model_debugging_utils.py", line 29, in
import torch.distributed.tensor
File "C:\Users\Ray\miniconda3\envs\comfyui-env\Lib\site-packages\torch\distributed\tensor_init_.py", line 4, in
import torch.distributed.tensor.ops # force import all built-in dtensor ops
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Ray\miniconda3\envs\comfyui-env\Lib\site-packages\torch\distributed\tensor_ops_init.py", line 2, in
from ._conv_ops import * # noqa: F403
^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Ray\miniconda3\envs\comfyui-env\Lib\site-packages\torch\distributed\tensor_ops_conv_ops.py", line 5, in
from torch.distributed.tensor._dtensor_spec import DTensorSpec, TensorMeta
File "C:\Users\Ray\miniconda3\envs\comfyui-env\Lib\site-packages\torch\distributed\tensor_dtensor_spec.py", line 6, in
from torch.distributed.tensor.placement_types import (
File "C:\Users\Ray\miniconda3\envs\comfyui-env\Lib\site-packages\torch\distributed\tensor\placement_types.py", line 8, in
import torch.distributed._functional_collectives as funcol
File "C:\Users\Ray\miniconda3\envs\comfyui-env\Lib\site-packages\torch\distributed_functional_collectives.py", line 9, in
import torch.distributed.distributed_c10d as c10d
File "C:\Users\Ray\miniconda3\envs\comfyui-env\Lib\site-packages\torch\distributed\distributed_c10d.py", line 23, in
from torch._C._distributed_c10d import (
ModuleNotFoundError: No module named 'torch._C._distributed_c10d'; 'torch._C' is not a package

[ScottTodd]

We're getting close to solving that, see #1202

[RSabbagh52]

Ok Thanks! I'll keep my eye out for a solution.