AMD ROCM/MigraphX machile learning #7169
Replies: 7 comments 8 replies
-
|
@mertalev I'd probably need your help to get this something akin to a functional state. Attached the Dockerfile |
Beta Was this translation helpful? Give feedback.
-
|
It looks like there is a prebuilt wheel for it distributed by AMD, but you're right that it supports 3.10 and not 3.11. This is fine - OpenVINO is in the same camp and that image uses 3.10 as a result. I would prefer this to avoid needing to build onnxruntime ourselves. |
Beta Was this translation helpful? Give feedback.
-
|
I’d need to see how to use the 3.10… in any case we’d need to wait until onnxruntime merges my pull request to fix the convolution bug https://github.com/microsoft/onnxruntime/pull/19550On Feb 17, 2024, at 13:29, Mert ***@***.***> wrote:
It looks like there is a prebuilt wheel for it distributed by AMD, but you're right that it supports 3.10 and not 3.11. This is fine - OpenVINO is in the same camp and that image uses 3.10 as a result. I would prefer this to avoid needing to build onnxruntime ourselves.
—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you authored the thread.Message ID: ***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
It would just mean using the FROM python:3.10-bookworm@sha256:b54e76c629a98430ac9c92e4f6bddeb672396a895b44a85022d12ee2f7239144 as builder-rocm
...
FROM builder-rocm as prod-rocm |
Beta Was this translation helpful? Give feedback.
-
|
The other dependencies like |
Beta Was this translation helpful? Give feedback.
-
|
I am going to try to build a image with the python-3.10 and the publicly
available runtime (when the change submits we can get it from the
nightlies).
Do you have any suggestion on how to modify the poetry config to install a
local package? right now I do something dirty like loading the venv and
calling pip uninstall/pit install
…On Sat, 17 Feb 2024 at 15:13, Mert ***@***.***> wrote:
The other dependencies like half would also need to be installed in the
prod image.
—
Reply to this email directly, view it on GitHub
<#7169 (reply in thread)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AANL3WLBIR5D52DPNRREVPTYUE2P3AVCNFSM6AAAAABDNPALN2VHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM4DKMBTHA2TO>
.
You are receiving this because you authored the thread.Message ID:
***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
You can add it like this: [tool.poetry.group.rocm]
optional = true
[tool.poetry.group.rocm.dependencies]
onnxruntime-rocm = {url = "https://repo.radeon.com/rocm/manylinux/rocm-rel-6.0.2/onnxruntime_rocm-inference-1.17.0-cp310-cp310-linux_x86_64.whl"}I actually tried to add it by adding the ROCm repo as a source, but this didn't work. It would have looked like this: [tool.poetry.group.rocm]
optional = true
[tool.poetry.group.rocm.dependencies]
onnxruntime-rocm = {version = "^1.17.0", source = "rocm"}
[[tool.poetry.source]]
name = "rocm"
url = "https://repo.radeon.com/rocm/manylinux/rocm-rel-6.0.2"
priority = "explicit" |
Beta Was this translation helpful? Give feedback.
-
|
Oh well! Dunno what the folks at AMD did, but this is actually working. I am going to clean up the image and send it for review. |
Beta Was this translation helpful? Give feedback.
-
|
oh well, it couldnt be that easy, the image is still using the CPU instead of the ROCM EP. I think it is defaulting that one |
Beta Was this translation helpful? Give feedback.
-
|
If you haven't already, you also need to add the MIGraphX/ROCm providers here |
Beta Was this translation helpful? Give feedback.
-
|
Oh, and you have to pass |
Beta Was this translation helpful? Give feedback.
-
|
i still used the CPU EP. I'll take a look a bit later.
…On Sat, 17 Feb 2024 at 22:39, Mert ***@***.***> wrote:
Oh, and you have to pass --build-arg DEVICE=rocm when building the image,
or else it will target cpu.
—
Reply to this email directly, view it on GitHub
<#7169 (reply in thread)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AANL3WNBDSTGE6JYRZ66TDLYUGO3RAVCNFSM6AAAAABDNPALN2VHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM4DKMBVGE3DK>
.
You are receiving this because you authored the thread.Message ID:
***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
|
How did this turn out I would love to be able to use my 7900xtx to help my server with the ML tasks. |
Beta Was this translation helpful? Give feedback.
-
|
I don't know how, but the Easy Diffusion project has it working great, maybe refer to that? |
Beta Was this translation helpful? Give feedback.
-
|
I couldn't continue the investigation. I parked it for a while. My patch
definitely fixes the issue, but at the cost of disabling algo catching.
…On Thu, 12 Dec 2024 at 10:47, Rena Kunisaki ***@***.***> wrote:
I don't know how, but the Easy Diffusion project has it working great,
maybe refer to that?
—
Reply to this email directly, view it on GitHub
<#7169 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AANL3WMSPKTRTWRBG6Q65D32FHK5JAVCNFSM6AAAAABTQE7QG2VHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTCNJUHAYDINQ>
.
You are receiving this because you authored the thread.Message ID:
***@***.***>
|
Beta Was this translation helpful? Give feedback.
-
The feature
Not for the faint of heart, but following the addition of CUDA, openVINO and ARM based ML addition to immich I am looking into adding that functionality.
I will monitor the progress in this thread, as it is significantly more complex that I though... I got it working, but I am still debugging some of their multiple 'quirks'
As a summary:
Platform
Beta Was this translation helpful? Give feedback.
All reactions