search

justxi / rocm

Ebuilds to install ROCM on Gentoo Linux
38 stars 23 forks source link

tensorflow for rocm #175

Closed perestoronin closed 3 years ago

perestoronin commented 3 years ago

Please tell me where to download the ebuild to build tensorflow for rocm ?

PS. tensorflow requires that the rocm directory is not in /usr, but in /opt/rocm, but if targeted rocm directory at /usr, then the tensorflow build script loops :(

justxi commented 3 years ago

I don´ t know. I don´t have any ebuilds for tensorflow.

justxi commented 3 years ago

Maybe @heroxbd @littlewu2508 have an ebuild for tensorflow?

Or did you mean pytorch? But the ebuilds were not developed by me: https://github.com/justxi/rocm/tree/master/sci-libs/pytorch -> https://github.com/justxi/rocm/issues/133

perestoronin commented 3 years ago

ebuild for tensorflow exists in portage - https://github.com/gentoo/gentoo/blob/master/sci-libs/tensorflow/tensorflow-2.4.0_rc1.ebuild but in ebuild masked rocm by flag internal https://github.com/gentoo/gentoo/blob/master/sci-libs/tensorflow/tensorflow-2.4.0_rc1.ebuild#217 export TF_NEED_ROCM=0

If in this ebuils set TF_NEED_ROCM=1 tensorflow failed to build on some reasons, and first of them consits in illegal location rocm directory /usr instead right /opt/rocm. May be relocating rocm in /opt/rocm in all ebuilds of rocm overlay are cheaper than unsuccessful all my tries to fix tensorflow to work with this overlay (rocm) as is?

PS. also exists old intresting infrastucture of tensorflow for gentoo https://github.com/gridgentoo/gentoo-tensorflow-overlay/blob/master/ai-frameworks/tensorflow/tensorflow-1.4.0.ebuild

perestoronin commented 3 years ago

I see working in progress in portage tree - now added sys-devel/llvm-roc/files/llvm-roc-3.9.0-tensorflow.patch but sci-libs/tensorflow without rocm flag yet.

perestoronin commented 3 years ago

tensorflow new released 2.4.0 14-12-2020

perestoronin commented 3 years ago

rocm 4.0 released https://github.com/RadeonOpenCompute/ROCm/tags

justxi commented 3 years ago

rocm 4.0 released https://github.com/RadeonOpenCompute/ROCm/tags

https://github.com/justxi/rocm/issues/177 ;-)

perestoronin commented 3 years ago

and when try to build tensorflow with flag rocm=1 : gcc: fatal error: cannot execute ‘cc1plus’: execvp: No such file or directory

how to fix with trouble about cc1plus ?