Intel Arc Graphics Thread #476
Replies: 27 comments 88 replies
-
|
UPDATE:
Before Intel update it, the code needs to be modified like this to avoid reporting errors: - match operation:
- case "multiply":
- output[top:bottom, left:right] = destination_portion * source_portion
- case "add":
- output[top:bottom, left:right] = destination_portion + source_portion
- case "subtract":
- output[top:bottom, left:right] = destination_portion - source_portion
+ if operation == "multiply":
+ output[top:bottom, left:right] = destination_portion * source_portion
+ elif operation == "add":
+ output[top:bottom, left:right] = destination_portion + source_portion
+ elif operation == "subtract":
+ output[top:bottom, left:right] = destination_portion - source_portionOf course not modifying it doesn't seem to affect normal usage. |
Beta Was this translation helpful? Give feedback.
-
|
I added this fix to master: 476d543 |
Beta Was this translation helpful? Give feedback.
-
|
About Karras scheduler: Currently using it causes a -996 "uses-fp64-math" error, which may be fixed in the next IPEX release (intel/intel-extension-for-pytorch#285) Until then, can specify to use # comfy/samplers.py
- sigmas = k_diffusion_sampling.get_sigmas_karras(n=steps, sigma_min=self.sigma_min, sigma_max=self.sigma_max, device=self.device)
+ sigmas = k_diffusion_sampling.get_sigmas_karras(n=steps, sigma_min=self.sigma_min, sigma_max=self.sigma_max, device="cpu") |
Beta Was this translation helpful? Give feedback.
-
|
is it Linux only or can you use this on a win10 machine? |
Beta Was this translation helpful? Give feedback.
-
|
You should be able to use WSL, but I haven't tested it. |
Beta Was this translation helpful? Give feedback.
-
|
So again, sharing the good news. An XPU version of their Pytorch extension with Pytorch 2.0 support,
Hats off to ComfyUI for being the only Stable Diffusion UI to be able to do it at the moment but there are a bunch of caveats with running Arc and Stable Diffusion right now from the research I have done. As of the time of posting: 1.) Setup can still be complicated in some respects and can randomly not work because Intel is only really verifying for enterprise Linux distributions and Ubuntu. My base Linux Fedora 38 install has the RPMs installed but I get a cryptic error regarding the runtime saying 2.) From what I have read at intel/intel-extension-for-pytorch#398, there is no native Windows version for some wheels to get 3.) Intel Arc right now has a huge problem with allocating more than 4GB of VRAM in IPEX even though the card has more VRAM in the case of the A750/A770 according to intel/intel-extension-for-pytorch#325. This seems to have mitigations elsewhere in Intel's oneAPI stack like their OpenCL compute runtime where anything you compile, you can work your code to go around the restriction via some flags and changed code according to intel/compute-runtime#627. No workarounds, it seems, until someone gets Intel to fix it. 4.) Because of that, experimenting with other Stable Diffusion UIs, it seems like every once in a while, Arc will occasionally run out of VRAM if you decide to not use lower VRAM flags and throw a 5.) Intel has an equivalent to I'm not even sure if I'm sure that Intel GPU support will not a big priority at the moment for the project, given the big blocker at this point is the 4GB VRAM limit which needs to be fixed by Intel. But that being said, things should be working a lot better than it is at the moment, maybe not average user ready, but should be ready for any mildly technical person. I really want to play with ComfyUI more but I really don't want to restart the application server for every several image I might want to generate even or rolling a dice for SDXL to actually finish a workflow. But the Arc cards are strong. I managed to equal the Nvidia RTX 3070 Ti in certain Stable Diffusion workflows with my 16 GB Intel Arc A770 so I look forward to the future when things are more mature and all the stars are aligned. Also not sure what issues need to be open here but there is potentially 2-3 of them that could be made from my report. Edit: Added a caveat I forgot to mention and filled in some information and fixed some typos. |
Beta Was this translation helpful? Give feedback.
-
|
Maybe it doesn't like the shifting from ram to vram, try running it with --highvram |
Beta Was this translation helpful? Give feedback.
-
|
That's really counterintuitive. But yeah, it works for single image generation, was able to generate 5 images in a row without crashing. It does still crash when I tried to enable a batch of 2 because I decided to push my luck and may still crash like the other UIs when IPEX decides to randomly stop working but maybe that is to be expected given what I mentioned above about VRAM allocation and etc. But yeah, I think that brings this from theoretically useful to possibly useful. I'll file the |
Beta Was this translation helpful? Give feedback.
-
|
@simonlui From what I found
Quote from the last link: |
Beta Was this translation helpful? Give feedback.
-
|
I followed you steps, only used the updated version and then I got this issue can you tell me whats wrong? |
Beta Was this translation helpful? Give feedback.
-
|
So this took me a bit of time, but I have the Docker image I used published here so hopefully someone can find it useful. Some things to note I've found while poking and experimenting with things. 1.) A lot of the issues are gone mentioned before by @kwaa months ago like the Karras scheduler not working, where it is working now even with the new dpmpp_3 schedulers, or noise issues if not using split attention which is gone for the most part unless your graphics driver has crashed too many times and a restart of the computer fixes that. Anyways, hope people have good success with it like I did. I might try and see why ipexrun is failing but for now, I am going to take a break. Edit: Added in a caveat I forgot to mention and fixed some typos. |
Beta Was this translation helpful? Give feedback.
-
|
In the event anyone else missed it: I seem to be missing the file that gets sourced from /opt/intel, too... may edit this post once I figure it out. |
Beta Was this translation helpful? Give feedback.
-
|
Because Arch only packages the basekit and disaggregates all the other toolkits into seperate components which is kinda confusing why. Also, opt/intel/oneapi/setvars.sh should be part of the basekit package as you can see here. You may need to do a reinstall of the basekit if you don't see the file. |
Beta Was this translation helpful? Give feedback.
-
|
I may update this guide in the near future. |
Beta Was this translation helpful? Give feedback.
-
|
Well, short story. After spend whole day to install Arch Linux on Windows 10 WSL (2 in my case). After all procedures, I stuck on
The installer reach a 87% and starts roll back all changes, because intel-oneapi-basekit not contain library libtbb.so.12 that needed to install oneAPI AI Analytics Toolkit v2023. The packgage that contains needed library is intel-oneapi-compiler-shared-runtime but it conflict with intel-oneapi-basekit. Because the istall procedure can't finish with intel-oneapi-basekit, I install intel-oneapi-compiler-shared-runtime. ComfyUI runs: [{user}@pc ArchLinux]$ But if i try to generate image I get an next error: As can see, I tries params like --force-fp16 --bf16-vae --lowvram --use-split-cross-attention --highvram but it's have no effect. So, models are loaded, but KSampler are crashes. And that's my dead end, because I have no idea what's going on. Not in ComfyUI itself, and also there more in Linux Arch. But I suppose this can be a kind of report. |
Beta Was this translation helpful? Give feedback.
-
|
Well, first off, there's no reason for you to use Arch Linux on WSL2 anyways, because you can pick Ubuntu and it's the only supported distribution that Intel supports for non-enterprise users anyways so using anything else is going to be painful and with WSL, it's a fool's errand because it won't gives you any additional benefits like speed and etc. like you may think it does. Secondly, that's what I made my Docker image for because it's very easy to mess up the install and there are a lot of chances to go wrong. Just on a cursory glance, you have the wrong package probably for TBB and the compiler runtime conflict is probably the issue. If you do want to fix the issue, you're going to have to reinstall everything correctly. |
Beta Was this translation helpful? Give feedback.
-
|
Doesn't work on Windows too, unfortunatelly. Do a step-by-step and get "missing startup.sh". Because this error is not specific, without code of error - kind of common for docker using - it's have no specific result in google. Fixing that case basing only on your tutorial not seems possible. |
Beta Was this translation helpful? Give feedback.
-
|
@roflosaur Sorry, it took a while to diagnose because I needed to set up a virtual machine to debug the issue you ran into but I found the root issue which was git was changing the line endings of the files based on OS where startup.sh needed to be tied to Unix line endings only. I have set up a .gitattributes to fix this. I would advise you to do a clean checkout again via deleting and cloning my repository again or if you know the command line, you can re-checkout the repository with the new .gitattributes applied. Thanks for reporting the issue. |
Beta Was this translation helpful? Give feedback.
-
|
Here is an error when build the app. I also downloaded and installed manually Intel One API Base Kit version 2023.1.0.47256 but useless. Can't build an app thus. |
Beta Was this translation helpful? Give feedback.
-
|
Welp, aren't you unlucky on your timing here? Intel has broken their GPG key for OneAPI repository and that means nothing from oneAPI can be downloaded from said repository. I unfortunately can't help you here since my image depends on downloading those packages fresh from Intel for updating reasons. So until they get their stuff together over there, you're stuck as am I. If you want to follow progress on when a fix might come, see this issue |
Beta Was this translation helpful? Give feedback.
-
|
I don't use Windows, but all the pieces should be together to run ComfyUI on Windows now without any horrible downsides like no AOT compilation or missing packages since a new unofficial Intel Extension for Python package has been released without needing any installation of external dependencies and bundling it all together. It remains to be seen whether there will ever be an official package that does this so this package is the best chance anyone has at actually using Arc on Windows natively. According to reports, it is a bit faster than WSL2 but slower than Linux native. The rough steps to do this should be roughly the same as the process outlined in the opening post for Linux minus platform-specific things. 1.) Make sure you install an Intel driver that is 4952 or newer. The latest driver can be found here python -m ensurepip --upgrade3.) Install git from here using the GUI installer or other means. cd <location where to place ComfyUI>
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI5.) Download non-official Intel Extension for Pytorch packages here and install the packages with pip. Assuming they are all put in the root of the ComfyUI repository, run this command line. pip install intel_extension_for_pytorch-2.0.110+git0f2597b-cp310-cp310-win_amd64.whl torch-2.0.0a0+gite9ebda2-cp310-cp310-win_amd64.whl torchvision-0.15.2a0+fa99a53-cp310-cp310-win_amd64.whlAgain, I will need to remind you this is unofficial and does have a degree of risk but it should be mostly safe. 6.) Finish the rest of the installation pip install -r requirements.txtAt this point, you should be done with the installation. To run ComfyUI each time from scratch, open a terminal/command prompt. Then run the following command lines replacing <> with your own input: cd <location of ComfyUI>
python main.py <Any extra ComfyUI arguments you want to use> |
Beta Was this translation helpful? Give feedback.
-
|
When I try to install this node modifications, it's destroy my ComfyUI build. Because just rolling back all changes was not possible, I completely wiped out Python and Anaconda (as I thought, because despite deleting entries from the registry in AppData\Roaming\Python\Python310\site-packages something strange was still going on), and then follow to next procedure. This is not "ture github way", but at least this can help to someone, who want to launch ComfyUI on Intel Arc for Windows.
Then go to step 9, if no error occured. If you have any problems at this step, such messages like: or See the image:
and go to: To figure out for commented names. If names have "~" at start - remove it. Goto step 8 (copy, enter, and wait to end): 8.
or instead last command, if you have less than 16 Gb memory:
At this point, if all works fine, you can use DDU again, and install latest Intel Arc drivers. Should to work. Also you can optional install WAS node suite, by running "install.bat" in the "ComfyUI\custom_nodes\was-node-suite-comfyui-main" folder. REMEMBER: any additions for ComfyUI can destroy your ComfyUI installation for Intel Arc, because them build for CUDA mostly part, and trying to install it can destroy your Python dependences, critical for ComfyUI working. Don't forget download models! |
Beta Was this translation helpful? Give feedback.
-
|
I used IPEX v2.1.10 but it's not working as expected can you tell me whats wrong? |
Beta Was this translation helpful? Give feedback.
-
|
Sorry for the late reply but you are using the wrong version of Python and you didn't read my 2nd instruction point properly.
Do not use the newest 3.11 or 3.12 and use version 3.10 with the non-official packages, this will hold until IPEX updates again but it does not work correctly with the latest ComfyUI and IPEX package at the moment. |
Beta Was this translation helpful? Give feedback.
-
|
Yes, but I also used cp311 version of ipex, which is for cpython 3.11 right? |
Beta Was this translation helpful? Give feedback.
-
|
Yes, which is the core root of the problem. You aren't using the modified packages I have suggested in my instructions and that has not changed from the older v2.0.110+xpu version packages someone else has provided as the newest version of IPEX v2.1.10 does not work with ComfyUI on any OS platform. |
Beta Was this translation helpful? Give feedback.
-
|
Reinstall in forced mode, needed dll in place, but still get this error. Can it be relevant to folder rights or upper or lower letter case in "user"? |
Beta Was this translation helpful? Give feedback.
-
|
I found out about what is causing it. It's a dependency issue with libuv. I will overhaul the tutorial based on that shortly. For you, you'll need to uninstall the oneAPI GUI installed components and then backtrack and follow steps 3 onwards but then the method I will provide should replace that in a much more easy manner. |
Beta Was this translation helpful? Give feedback.
-
|
There really needs to be a better, more updated tutorial for this. Much of the information is stretched out over the thread and mixed up with information between Linux and Windows |
Beta Was this translation helpful? Give feedback.
-
|
I'll consider writing one for Ubuntu because it's the only Linux distro on the consumer side supported officially by Intel unless you pay the big money for Red Hat and SUSE support with their enterprise GPUs. |
Beta Was this translation helpful? Give feedback.
-
|
Thanks, should to try too. |
Beta Was this translation helpful? Give feedback.
-
|
The whole thing about installing the OneAPI AI Analytics toolkit didn't seem necessary though. On a fresh install of Arch, All I did was install Torch 2.0.1 and intel-compute-runtime-bin from Arch repositories, then install the Intel extension for pytorch with the code below
And then from a terminal in the directory where ComfyUI is installed I used
And that line above I added to a start.sh script, which can include launch options for easier usage. But yeah, that was all I had to do and it takes me about 4 seconds to do 20 steps at 512x768. I'm not even sure I needed to get anything from Arch's repository, if so, those packages should be available in any repository, which would make the install fairly simple for any distro. |
Beta Was this translation helpful? Give feedback.
-
|
@burntbyhellfire If you don't use Intel Python, you don't need to install the AI Kit. |
Beta Was this translation helpful? Give feedback.
-
|
If we both have a GPU of the same specs we could both run some benchmark prompts and settings to see if there really are any performance improvements with doing it that way. I just got a Sparkle A770 16GB. |
Beta Was this translation helpful? Give feedback.
-
|
To install on Ubuntu. 1.) Install Linux drivers with the following instructions provided by Intel here sudo apt install python3-pip git3.) Install ComfyUI with the following terminal commands replacing the <> portion with a selection of your choice. cd <Location to put ComfyUI>
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI4.) Install all the Intel Extension for Pytorch pip Python packages first with this terminal command. python -m pip install torch==2.0.1a0 torchvision==0.15.2a0 intel-extension-for-pytorch==2.0.110+xpu --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/5.) Finish the rest of the dependency installation with this terminal command: pip install -r requirements.txtInstallation should be done at this point. To run ComfyUI, one can type in the following terminal commands replacing <> with your own input: cd <location of ComfyUI>
python main.py <Any extra ComfyUI arguments you want to use>For other Linux distros outside of Arch Linux or Ubuntu, one will need to manually install the Intel compute runtime, git, a Python with a version supported by Intel Extension for Pytorch like 3.10 as of the time of this writing or using Intel's Python from the AI Kit which will provide that, pip, and any other dependencies required according to your own Linux distribution's package manager or install scripts but one should be able to then just follow step 3 and onwards without any issue afterwards. |
Beta Was this translation helpful? Give feedback.
-
|
I've updated the Arch Linux setup guide for the latest IPEX and ComfyUI, without oneAPI AI Kit. If anyone wants to uninstall the previously installed AI Kit, it's available: cd /opt/intel/oneapi/installer
sudo ./installer --action remove --product-id intel.oneapi.lin.aikit.product --product-ver 2023.1.0-31760
# if you don't need
paru -Rsc libxcrypt-compat |
Beta Was this translation helpful? Give feedback.
-
New Patch!Thanks to vladmandic/automatic for the code. contributors Basically, you just need to copy Then modify try:
import intel_extension_for_pytorch as ipex
if torch.xpu.is_available():
xpu_available = True
+ from attention import attention_init
+ ok, e = attention_init()
except:
passIt could partially fix intel/intel-extension-for-pytorch#325. (so f**k you, intel) |
Beta Was this translation helpful? Give feedback.
-
|
That's more of a workaround, I've seen his code but I didn't think @comfyanonymous would enjoy special casing his code with a specific attention implementation that would work for Intel GPUs only for working around the issue which is why I didn't submit a variant it when I wrote the patches for enhanced IPEX support. Doing manual memory management like this is a pain in Python and doesn't look particularly nice. I think Intel has to fix it on their end, asking end users of IPEX to do this to every project that implements IPEX is untenable. |
Beta Was this translation helpful? Give feedback.
-
|
I couldn't find any performance gains with Ipex in tests I ran in ComfyUI before, and after setting up an environment and running Comfy from it. |
Beta Was this translation helpful? Give feedback.
-
|
interesting tho, that i didn't do anything with modules/intel/ipex/attention.py, but problem with 4GB limit is gone after patching p.s. error still continue to occur ^_^ |
Beta Was this translation helpful? Give feedback.
-
|
It seems that only older versions of |
Beta Was this translation helpful? Give feedback.
-
|
Cool... Got it working with WSL2 using Arch. Followed the instructions at the top, as well as installed jemalloc paru(root/native environment), openmp via pip(venv). There are issues with the env vars as Arch is pulling my windows env vars through. So I some times have to re run the setvars.sh. I get warnings about libpng when I run via python or ipexrun. Everything still works though. I generally run through ipexrun but add xpu to the command i.e.
Apart from some memory overruns I've had no issues the last few days. |
Beta Was this translation helpful? Give feedback.
-
|
anybody can tell where intel store their patches for pytorch? |
Beta Was this translation helpful? Give feedback.
-
|
You can find these "patches" here but it's really not that much different over just compiling the source tree provided. |
Beta Was this translation helpful? Give feedback.
-
|
thanks! |
Beta Was this translation helpful? Give feedback.
-
|
Yes, and pytorchvision and other associated packages. |
Beta Was this translation helpful? Give feedback.
-
|
new url for downloading extension also i guess new release soon ^_^ i've managed to build on archlinux pytorch with torch patches, but intel-extension-for-pytorch failed a lot with different errors ^_^ |
Beta Was this translation helpful? Give feedback.
-
|
so i was able to compile whl package of intel-extension-for-pytorch -- intel_extension_for_pytorch-2.1.10+git7479afc-cp311-cp311-linux_x86_64: bad part that speed decreased (from 3.4 s/it to 4.3 s/it) |
Beta Was this translation helpful? Give feedback.
-
|
also what is device type of arc a770 - dg2-g12 or acm-g10 or ats-m150? |
Beta Was this translation helpful? Give feedback.
-
|
Getting this error upon installing using "pip install torch==2.0.1a0 torchvision==0.15.2a0 intel_extension_for_pytorch==2.0.110+xpu -f https://developer.intel.com/ipex-whl-stable-xpu", using WSL on Windows with Ubuntu: ERROR: Could not find a version that satisfies the requirement torch==2.0.1a0 (from versions: 1.11.0, 1.12.0, 1.12.1, 1.13.0, 1.13.1, 2.0.0, 2.0.1, 2.1.0, 2.1.1) |
Beta Was this translation helpful? Give feedback.
-
|
There's some kind of redirect happening from https://developer.intel.com/ipex-whl-stable-xpu to an ec2 instance page. My browser pointed to http://ec2-52-27-27-201.us-west-2.compute.amazonaws.com/ipex-release.php?device=xpu&repo=us&release=stable. THE FOLLOWING IS NOT SECURE (but would install the right versions of the packages): I suspect this is a problem on Intel's side. There may not be much any of us can do about it, aside from download the packages from somewhere else. |
Beta Was this translation helpful? Give feedback.
-
|
Pinging @kwaa to update the OP post. Per the updated install instructions at https://intel.github.io/intel-extension-for-pytorch/#installation, you should now use the following pip command instead.
or if you live in China or somewhere close.
Edit: Note that this is only for Linux. Windows users should keep referring to the install procedure I wrote here to install IPEX XPU natively as Intel has yet to publish official Pytorch packages that work for the platform. You will install the IPEX CPU version instead if you use these instructions which is not what you want. |
Beta Was this translation helpful? Give feedback.
-
Ok, updated. |
Beta Was this translation helpful? Give feedback.
-
|
Intel® Extension for PyTorch* v2.1.10+xpu Release Notes Latest intel_extension_for_pytorch-2.1.10+xpu |
Beta Was this translation helpful? Give feedback.
-
|
i've already opened a bug on this error in ComfyUI. i got ComfyUI working with file from #476 (comment) and hack for |
Beta Was this translation helpful? Give feedback.
-
|
This is definitely not a ComfyUI bug, it's an upstream bug with Pytorch 2.1 and Intel's patches for it to make IPEX. Until they fix their end of this, the fact that you have to pile more custom code and hacks to make it work doesn't change the fact that it's broken at the end of the day. It needs to be fixed but you are barking up the wrong tree to get it done. |
Beta Was this translation helpful? Give feedback.
-
|
roger, closed issue for reference intel/intel-extension-for-pytorch#483 |
Beta Was this translation helpful? Give feedback.
-
Even with those changes, it didn't work for me until I used source .venv/bin/activate
# export USE_ONEMKL=OFF
export MKL_DPCPP_ROOT=/opt/intel/oneapi/mkl/latest
source /opt/intel/oneapi/setvars.sh --force
nice -n 19 ipexrun xpu main.py --normalvram --preview taesd --use-pytorch-cross-attention --bf16-unet --bf16-vae --dont-upcast-attention "$@"Pretty sure Installed using: Will require diff --git a/comfy/model_management.py b/comfy/model_management.py
index 3adc427..da7790e 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -49,7 +49,11 @@ try:
import intel_extension_for_pytorch as ipex
if torch.xpu.is_available():
xpu_available = True
-except:
+ from attention import attention_init
+ attention_init()
+ print('*** Intel attn init!')
+except Exception as e:
+ print('*** Intel attn init failed:', e)
pass
try:(The original changes tried to unpack the result from |
Beta Was this translation helpful? Give feedback.
-
|
An update to the situation finally. intel/intel-extension-for-pytorch#483 (comment) mentioned they merged in a fix into the xpu-main branch of IPEX and if you compile that like I did to verify things, everything works out of the box now without hacks. There should be a release of IPEX XPU soon and so I think everything should hopefully then just work out of the box going forward. |
Beta Was this translation helpful? Give feedback.
-
|
Howdy all! Trying to install pytorch to use my ARC a770 with comfyui, but no matter what version I download/install and what args I use for main.py I come up against a "Torch not compiled with CUDA enabled" error. My set up is Fedora 39 with python 3.10.3 used in the venv. Installed torch with the command:
Both the python and ipex commands listed above:
or
produce the following outcome: The assertion is correct. I checked and don't have CUDA support in place: Comfy works in CPU-only mode, but it's stupid slow. There doesn't seem to be an ignore CUDA option for comfy. What am I missing here? |
Beta Was this translation helpful? Give feedback.
-
|
Beta Was this translation helpful? Give feedback.
-
|
The problem is most likely you don't have the correct libraries installed or you aren't setting the magic Intel variables before running it. The way it works currently is to just try IPEX and if that raises an exception the error is hidden and it goes on to try to use CUDA. You can try changing the part at the top of try:
import intel_extension_for_pytorch as ipex
if torch.xpu.is_available():
xpu_available = True
except:
passChange If you're on Linux (or WSL) try something like this before trying to start: export USE_ONEMKL=OFF
export MKL_DPCPP_ROOT=/opt/intel/oneapi/mkl/latest
source /opt/intel/oneapi/setvars.sh --forceExample from my startup script (also may need to activate the Python environment) and based on Linux/Arch. If your Intel/IPEX stuff is installed in different places you'll need to adjust the paths. |
Beta Was this translation helpful? Give feedback.
-
|
Here is an alternative approach for trying to get Arc working. Turn your Arc into fake CUDA: This time we rip off basically the whole CUDA emulation layer is SD.next: First clone the SD.next repo and switch to the Apply this patch: diff --git a/modules/errors.py b/modules/errors.py.bak
similarity index 100%
rename from modules/errors.py
rename to modules/errors.py.bak
diff --git a/modules/intel/ipex/hijacks.py b/modules/intel/ipex/hijacks.py
index 0dbec37393c..fdd18ff6d90 100644
--- a/modules/intel/ipex/hijacks.py
+++ b/modules/intel/ipex/hijacks.py
@@ -1,7 +1,11 @@
import contextlib
import torch
import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
-from modules import devices
+#from modules import devices
+
+class devices:
+ device = torch.device('xpu')
+ dtype = torch.bfloat16
# pylint: disable=protected-access, missing-function-docstring, line-too-long, unnecessary-lambda, no-else-return
@@ -67,6 +71,7 @@ def from_numpy(ndarray):
except Exception: # pylint: disable=broad-exception-caught
original_torch_bmm = torch.bmm
original_scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention
+ raise
# Data Type Errors:Alternatively you can clone my fork with the required hacks: https://github.com/KerfuffleV2/sdnext/tree/comfyhacks — note: I probably won't keep this updated so patching it yourself is probably going to be more reliable. Symlink the Apply this patch to ComfyUI: diff --git a/comfy/model_management.py b/comfy/model_management.py
index fefd3c8..831ce49 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -45,12 +45,9 @@ if args.directml is not None:
# torch_directml.disable_tiled_resources(True)
lowvram_available = False #TODO: need to find a way to get free memory in directml before this can be enabled by default.
-try:
- import intel_extension_for_pytorch as ipex
- if torch.xpu.is_available():
- xpu_available = True
-except:
- pass
+import modules.intel.ipex as _ipex
+print('** IPEX init',_ipex.ipex_init())
+# xpu_available = True
try:
if torch.backends.mps.is_available():
@@ -318,7 +315,7 @@ class LoadedModel:
self.model_accelerated = True
- if is_intel_xpu() and not args.disable_ipex_optimize:
+ if True and not args.disable_ipex_optimize:
self.real_model = torch.xpu.optimize(self.real_model.eval(), inplace=True, auto_kernel_selection=True, graph_mode=True)
return self.real_model
@@ -639,6 +636,7 @@ def pytorch_attention_enabled():
return ENABLE_PYTORCH_ATTENTION
def pytorch_attention_flash_attention():
+ return False
global ENABLE_PYTORCH_ATTENTION
if ENABLE_PYTORCH_ATTENTION:
#TODO: more reliable way of checking for flash attention?
diff --git a/comfy/utils.py b/comfy/utils.py
index f8026dd..05bdd06 100644
--- a/comfy/utils.py
+++ b/comfy/utils.py
@@ -13,7 +13,7 @@ def load_torch_file(ckpt, safe_load=False, device=None):
sd = safetensors.torch.load_file(ckpt, device=device.type)
else:
if safe_load:
- if not 'weights_only' in torch.load.__code__.co_varnames:
+ if not 'weights_only' in torch.load.__code__.co_varnames and False:
print("Warning torch.load doesn't support weights_only on this pytorch version, loading unsafely.")
safe_load = False
if safe_load:(If you knew how long it took me to track down the flash attention thing... Without that part, you'll get random corruption when sampling but only under some conditions.) With these changes it shouldn't be necessary to do stuff like pull in the Intel/IPEX variables or set MKL stuff, I was able to remove those parts from my startup script. The sliced attention from SD.next has two tuneables you can access via environment variable: sdpa_slice_trigger_rate = float(os.environ.get('IPEX_SDPA_SLICE_TRIGGER_RATE', 6))
attention_slice_rate = float(os.environ.get('IPEX_ATTENTION_SLICE_RATE', 4))I had to set So far this seems to be working for me at least as well as the previous way with the latest release of the Intel Torch stuff. I set the fallback types in For reference, here's my startup script: #!/bin/bash
source .venv/bin/activate
export IPEX_ATTENTION_SLICE_RATE=3
nice -n 10 ipexrun xpu main.py --preview-method taesd --use-pytorch-cross-attention --disable-xformers --bf16-vae --bf16-unet "$@" |
Beta Was this translation helpful? Give feedback.
-
|
I'm trying to setup ComfyUI on Debian box with Intel ARC A770 16GB and I found something that I believe could be clarified better in the first post. One of the prerequisites is installing Second thing - that package is absurdly huge. Over 13GB after install, ... Jesus Christ. Do we actually know what parts of the stack SD and ComfyUI uses? Maybe the install size could be a little less. At least on Debian we have the ability to install/remove individual packages from that basekit. |
Beta Was this translation helpful? Give feedback.
-
|
Here is the full list of packages inside basekit meta-package:
Yes, I did run |
Beta Was this translation helpful? Give feedback.
-
|
I've checked the env vars. If I set them from the terminal as a normal user I can get output from However when I try to run ComfyUI with this I get: If I set the env vars from within venv I get: I can then open ComfyUI, but running default graph gives |
Beta Was this translation helpful? Give feedback.
-
|
I've made some progress. After updating compiler and runtime to the newest versions available here the comfy logpython3 main.py --use-pytorch-cross-attention --highvram /media/my_username/STABLE_DIFFUSSION/ComfyUI/venv/lib/python3.11/site-packages/torchvision/io/image.py:13: UserWarning: Failed to load image Python extension: ''If you don't plan on using image functionality from `torchvision.io`, you can ignore this warning. Otherwise, there might be something wrong with your environment. Did you have `libjpeg` or `libpng` installed before building `torchvision` from source? warn( Total VRAM 15474 MB, total RAM 257583 MB Set vram state to: HIGH_VRAM Device: xpu VAE dtype: torch.bfloat16 Using pytorch cross attention Starting serverTo see the GUI go to: http://127.0.0.1:8188 Prompt executed in 26.22 seconds |
Beta Was this translation helpful? Give feedback.
-
|
Which are the same exceptions as detailed earlier in the thread using the latest version of IPEX. You can either downgrade IPEX or hack it with some of the code snippets that have been posted to get things working. |
Beta Was this translation helpful? Give feedback.
-
|
Thanks! I tried hacking around the IPEX problem. I added the script from Vladmandic and patched model management file in Comfy.
ipexrun xpu main.py --use-pytorch-cross-attention --bf16-unet --bf16-vae --dont-upcast-attention --highvram
/media/my_username/STABLE_DIFFUSSION/ComfyUI/venv/lib/python3.11/site-packages/torchvision/io/image.py:13: UserWarning: Failed to load image Python extension: ''If you don't plan on using image functionality from /media/my_username/STABLE_DIFFUSSION/ComfyUI/venv/lib/python3.11/site-packages/torchvision/io/image.py:13: UserWarning: Failed to load image Python extension: ''If you don't plan on using image functionality from Total VRAM 15474 MB, total RAM 257583 MB To see the GUI go to: http://127.0.0.1:8188 |
Beta Was this translation helpful? Give feedback.
-
|
Commenting here because this thread and other available sources was somewhat difficult to follow. On a fresh ubuntu 23 install i did the following git clone comfy
run comfy with this makes comfy run. Atleast i was able to use sdxl and sdxl turbo, control net and some depth map and canny nodes. Then also look at the top of this page and do the 4G patch |
Beta Was this translation helpful? Give feedback.
-
|
correction, after using the correct version of attention.py, it seems that ip adapters, instant id and reactor work also |
Beta Was this translation helpful? Give feedback.
-
|
Okay so I am running in WSL, comfy is installed and runs as intended, seems to be loading models into vram. But the issue is as follows: |
Beta Was this translation helpful? Give feedback.
-
|
I will be attempting to have it work on Windows instead of WSL. |
Beta Was this translation helpful? Give feedback.
-
|
https://drive.google.com/file/d/1G-JMlQfgGlyXWKbF05YXDScrB38amf4Q/view?usp=sharing Windows integrated package |
Beta Was this translation helpful? Give feedback.
-
|
@cunkai I downloaded that, but it's all in... Japanese? I think? Maybe Chinese. I changed the language to USA in settings and restarted, but everything is still moonspeak.
Any way to... fix this? |
Beta Was this translation helpful? Give feedback.
-
|
Just wanted to let everyone know. There is a new version of IPEX for XPU, v2.1.20+xpu, released yesterday. |
Beta Was this translation helpful? Give feedback.
-
|
Sorry for the month long delay, the two issues I mentioned here seemed like they got fixed with the changes made after v2.1.20+xpu and I wanted to make sure that was the case and get my setup verified to work and test the Linux packages before releasing them publicly. You can find them uploaded here as a pre-release to my Docker image for ComfyUI <link removed> which you can use with the image or separately. I have made available Python 3.10 and 3.11 packages. Edit: No longer applies as v2.1.30+xpu has been released so release link has been removed but post has been kept as is for posterity. |
Beta Was this translation helpful? Give feedback.
-
|
I skimmed this entire thread but the information here feels very disjointed and outdated. I have Windows 10. Will Intel Arc work for me with Comfy UI? I used to use SD1111 when I had my RTX 1080, then I switched to Forge, but when I got the Intel Arc I ended up switching to Vlad's fork. The problem is his fork is a mess of UI and it doesn't support Clip Skip 2 for Arc (I have no idea why) so I'm debating trying out Comfy even though I know nothing about it. So what's the verdict? Can Comfy work well on Windows 10 with an A770? |
Beta Was this translation helpful? Give feedback.
-
|
I think the short answer is still No. |
Beta Was this translation helpful? Give feedback.
-
|
To answer, you need a custom build of Intel Extension for Pytorch for ComfyUI to work on Windows. What I said in #476 (comment) still holds for now, someone had a helpful picture guide too there. I didn't build Windows packages since I don't have such a setup but theoretically, if a person replicated what I did in building a custom build on Windows, it would work with the latest stuff. However, you won't get the repackaged all in 1 package the non-official build does and you would need to install a bunch of dependencies on top for it to work properly. And the effort may not be worth the squeeze, the improvement in SDXL is barely .5 of an second saved per iteration, it's close to 1s/it at default 1024x1024 compared to when it first worked with the custom build I mentioned in my comment at around 1.4-1.5s/it on Linux. @chux0519 This is only a problem for single allocations. In practice, this isn't a problem for Arc most of the time except in certain situations like upscaling and in higher resolutions. The main issue is still what I mentioned above. |
Beta Was this translation helpful? Give feedback.
-
|
In addition to updating some code for the recent IPEX release I did for ComfyUI, I've updated the README with a cross-platform installation guide instead of just linking to this thread which should hopefully help first-time users find their way into what they need to do even if the complexity to get ComfyUI running isn't as easy as other vendors. I am hoping someday IPEX functionality finds its way into upstream Pytorch but until then, the install guide is what it is. Do submit a PR to change it if you find anything wrong with what I changed. |
Beta Was this translation helpful? Give feedback.
-
|
In an unusual but welcome move, Intel released a version of Intel Extension for Pytorch for XPU, v2.1.30+xpu, that has the fixes needed to run ComfyUI on Linux and Windows without code modifications, from what wheels I could see in the repository. I was able to verify SD 1.5 is working, but SDXL is being a bit wonky on my end so I need some further time to test it out and see what is up. Do note as well that Linux and the Intel Compute Runtime currently as of writing this has some outstanding issues that are being worked through with the latest versions like this which may be related to my issue so tread carefully. If anyone can verify that this can run ComfyUI without any changes, that would be great so I can move forward with creating a pull request for some code changes I have been meaning to merge into the project. Edit: I verified that SDXL is working for the most part but it is pretty unstable without VRAM reduction built in for higher resolution images and I am getting hit by the slower speed Intel generation speeds by the bug I mentioned above so it's best to stay on a kernel at 6.6.25 LTS/6.8.4 or lower if possible to avoid getting slowdowns at this time. Might change in the future. |
Beta Was this translation helpful? Give feedback.
-
|
Tested with SD1.5 and SDXL. Seems to work. Over 4GB allocations still fail without the attention.py fix. Stable cascade generated either black og garbled image |
Beta Was this translation helpful? Give feedback.
-
|
Stable Cascade and Wuerstchen IPEX issue is being tracked at intel/intel-extension-for-pytorch#529. Over 4GB allocations have been discussed here but the discussion I had months prior in oneapi-src/oneDNN#1638 suggests it needs to be fixed in hardware and Intel decided the hit to translate above 4GB memory addresses takes too much of a performance hit for their liking so Alchemist will probably never see this fixed but Battlemage and up might. |
Beta Was this translation helpful? Give feedback.
-
|
This seems to be the last working recipe for me (Arch Linux): #476 (comment) Since that posting the situation devolved significantly from running ComfyUI out of the box:
Am I holding it wrong or is that what it is at the moment? |
Beta Was this translation helpful? Give feedback.
-
IPEX is fine with the latest version, v2.1.30+xpu, as stated in the prior post to yours.
Correct, this is the only issue currently affecting things that people running rolling Linux distributions will hit, see intel/compute-runtime#726 . The issue is only fixed at the time of this writing in Intel's fork of the kernel in drm-intel-gt-next. And it gives way less performance in the fixed kernel at the moment, 1/4th according to the bug report. Best to stay either on 6.8.4 or 6.6.25 LTS or older if possible.
IPEX has clearly defined Python requirements in its installation page that guides you through the process. The latest Python Intel builds and supports is 3.11 so that is what I stick with in my Docker image and you may possibly want to use that if you don't want to bother with keeping track of and maintaining the correct versioning for the packages needed to run ComfyUI. |
Beta Was this translation helpful? Give feedback.
and others
-
ComfyUI now supports Intel Arc Graphics. (#409)
Since the installation tutorial for Intel Arc Graphics is quite long, I'll write it here first.
Intel Extension for PyTorch is currently only available for Linux, so you will need to have a Linux or WSL environment.
Arch Linux (with
paru) are used here as example operating systems.Install Python and PIP:
Install Intel Compute Runtime and Intel oneAPI Base Kit:
Install ComfyUI:
Install Dependencies (via
venv):python -m venv venv source venv/bin/activate pip install torch==2.0.1a0 torchvision==0.15.2a0 intel-extension-for-pytorch==2.0.120+xpu --extra-index-url https://pytorch-extension.intel.com/release-whl-aitools/ pip install -r requirements.txtFor the second start and beyond, venv needs to be reactivated:
source venv/bin/activateSet oneAPI vars:
source /opt/intel/oneapi/setvars.shRunning ComfyUI (via
python):Running ComfyUI (via
ipexrun):Beta Was this translation helpful? Give feedback.
All reactions