There are several GPUs that ROCm doesn't officially support but
will work correctly if ROCm is directed to treat the GPU as a different
one that is supported and has a similar architecture.
This can be done by setting `HSA_OVERRIDE_GFX_VERSION`.
Ollama has documentation on this topic: https://github.com/ollama/ollama/blob/main/docs/gpu.md#amd-radeon
Verify that all `href` attributes emitted as part of the entrypoint page
after logging in are reachable.
Co-authored-by: Bruno BELANYI <bruno@belanyi.fr>
- rename hardware.opengl to hardware.graphics
- remove hardware.opengl.driSupport, which does nothing
- remove hardware.opengl.setLdLibraryPath, which should never be done
- rename hardware.opengl.driSupport32Bit to hardware.graphics.enable32Bit
- lost of small docs / formatting cleanups
This was meant to make amazon-ssm-agent work "out of the box" on non-NixOS
systems but the feature never really worked.
The problem is that amazon-ssm-agent looks for the files "amazon-ssm-agent.json"
and "seelog.xml" but the files in the package are named
"amazon-ssm-agent.json.template" and "seelog.xml.template". So even with
this overrideEtc = true it would not be able to find the config.
E.g. you'd get an error like
Error occurred fetching the seelog config file path: open /nix/store/pyfxjr0i0hszcj9b6fqly6344zf9zhcb-amazon-ssm-agent-3.3.484.0/etc/amazon/ssm/seelog.xml: no such file or directory
on startup.
Removing this parameter from the from the package doesn't break things as it didn't work in the first place.
This is not a breaking change. Existing setups continue to work as-is.
Users of `cfg.mailerPasswordFile` will get an option rename/deprecation
warning, but that's it (assuming there is no regression).
This adds `cfg.secrets`, which is a wrapper over systemd's
`LoadCredential=` leveraging Forgejo's `environment-to-ini`.
`environment-to-ini` is intended for configuring Forgejo in OCI
containers.
It requires some fairly annoying escaping of the section names to fit
into the allowed environment variable charset.
E.g. `"log.console".COLORIZE = false` becomes
`FORGEJO__LOG_0x2E_CONSOLE__COLORIZE=false`.
- `.` needs to be replaced with `_0X2E_` and
- `-` needs to be replaced with `_0X2D_`
Those are simply the hex representation of each char from an ASCII
table:
. = ASCII 46 = 46 (decimal) = 2E (hex) = 0x2E = _OX2E_
To make interacting with `environment-to-ini` less annoying, we template
and escape the sections/keys in nix:
`cfg.secrets` takes the same free-form sections/keys as `cfg.settings`.
Meaning there is now a generalized abstraction for all keys, not just
those that have been manually implemented in the past.
It goes as far as theoretically allowing one to have `DEFAULT.APP_NAME`
read from a secret file.
I don't know why one would want to do that, but it has been made
possible by this :^)
More reasonable examples are listed in the `cfg.secrets` option example.
We also continue to bootstrap a handful of secrets like
`security.SECRET_KEY`. This is done is a sort of sidecar bootstrap unit
fittingly called `forgejo-secrets.service`.
Overriding those is, just like before, not really intended and requires
the use of `lib.mkForce` and might lead to breakage. But it is, in a
way, more possible than before.
When services.gollum.{user,group} was specified a value other than its
default (i.e. "gollum"), the build failed due to referencing a
non-existing user.
