| Age | Commit message (Collapse) | Author |
|---|
|
Mynion was a prototype and while it was cool and worked well, it is unused and
causing magma to build, which takes forever. I have settled on using ollama for
local inference and a hosted inference API for production.
|
|
Otherwise it might change from underneath me (which it did already once).
|
|
This is basically exactly the client library that I would write myself. Some
parts of it are still beta quality, but it's the sort of thing that I would
contribute to anyway.
Unfortunately I couldn't get the llm-llama-cpp plugin to work because it depends
on llama-cpp-python which is not packaged for nix and is hard to package because
the upstream project vendors a patched version of llama.cpp. So I'm stuck with
ollama for now, but that's fine because it actually works.
|
|
This superceedes exllama and tabbyAPI which I could never get working fully.
Unfortunately I had to switch to NixOS unstable to get all the Go builder stuff
to work, so this is a cause of yet another version drift, but I guess it's
inevitable and I should just learn to mitigate it with my nixpkgs shenanigans.
|
|
Added docs to core libraries and expanded the Example.hs test to test for all
the things I want to support in a build.
|
|
This brings a bunch of improvements. I got rid of some custom packages, I can
now build exllama without using a non-default cuda version. Oh yeah and I get to
use GHC 9.6.2 now, a huge upgrade from 9.4. Unfortunately I also updated ormolu
and some unrelated formatting changed, but that's life I guess.
|
|
Removes my custom llama-cpp build and instead pulls in the upstream build from
nixos-23.11.
|
|
nixfmt is the soon-to-be official formatter for Nix code, as per the NixOS
GitHub group. So I figure I should just adopt it without worrying too much about
the specifics of the formatting. I just formatted everything in one go, hence
the huge diff, oh well.
|
|
This implements a prototype Mynion, my chatbot which will eventually
help me write code here. In fact he's already helping me, and works
pretty well over xmpp.
The prompt is currently not checked in because I'm experimenting with it
a lot, and it should probably be a runtime parameter anyways.
In the course of writing this I added some helper libraries to get me
going, configured black (didn't even know that was possible), and added
'outlines' and its dependencies even though I didn't end up using it.
I'll keep outlines around for now, but I'm not sure how useful it really
is because afaict its just pre-defining some stop conditions. But it
took a while to get it working so I'll just keep it in for now.
|
|
Lots of changes here but the code is much improved. The nix code is clearer and
structured better.
The Haskell code improved in response to the nix changes. I needed to use a
qualified path instead of the abspath because the BIZ_ROOT changes based on
whether bild runs in nix or runs in the user environment.
Rather than passing every argument into Builder.nix, now I just pass the json
from bild and deconstruct it in nix. This is obviously a much better design and
it only came to be after sleeping on it the other night.
|
|
I added 'black' to Biz/Lint.hs, but not the others because they rely on
dependencies being in the PYTHONPATH to work, so they are only relevant
in nix builds and repls.
I also made some other tweaks to the python checkPhase and linted all
the files. Everything should be building and linting correctly now.
|
|
The main change here is 'puts' now returns a value, this enables me to
collect the value from the conduit source while also doing stuff with
it, like printing or logging it as I want.
Previously I was running conduit over the source, *and then* kicking off
the concurrent processes to wait for the process and collect the output.
This would (I think) drain the source before it got to the 'puts'
conduit run, and so I wouldn't be able to get the output streamed in
real time.
It took a lot of refactoring and exploratory programming to get to this
relatively-small diff, but now puts works correctly. At least I think it
does... it seems to work more reliably from ghci than from the shell.
Maybe the shell or TERM is causing nix-store to do some buffering? Maybe
I need to use the threaded runtime in GHC? Not sure, but I will look out
for this issue and try to identify and fix.
Update: yep it was the threaded runtime. I enabled that and now it works
in the shell. I squashed that commit into this one.
|
|
|
|
I had to redo some of my python overrides and crib the bitsandbytes.nix
from upstream.
Ava is failing because:
ValueError: Tokenizer class LlamaTokenizer does not exist or is not
currently imported.
I think this means I need to update my nixpkgs pin, so I'm gonna
snapshot my work in git, do the update, and that might obviate the local
bitsandbytes.nix anyway.
|
|
|
|
|
|
Mostly thid required packaging up some deps, but also had to recompile
stuff with cuda support.
|
|
|
|
I don't care about ghcjs anymore, the most javascript I want to do is jQuery.
|
|
I ended up deleting miso, and consequently all files under Hero/ and Miso/,
because I couldn't get miso to build with GHC 9.2.
Other things:
- Niv has been wrapped by Biz/Bild/Deps.hs, so I can extend it to my liking.
- Apply-refact is gone because I couldn't get it to build.
- Disabled python stuff.
|
|
This means my call to ghc-pkg will look at the full package set from Hoogle.
|
|
This also makes some changes to the build tooling to clean up the environment a
bit, and get us closer to 'bild -s'.
|
|
|
|
Still todo: add authentication. But that can wait.
In re-implementing this, I was able to figure out how to get the Go.mult working
properly as well. The problem is that a tap from a mult channel does not remove
the message from the original channel. I'm not sure if that should be a core
feature or not; for now I'm just draining the channel when it's received in the
Que HTTP handler. (Also, this would be a good place to put persistence: have a
background job read from the original channel, and write the msg to disk via
acid-state; this would obviate the need for a flush to nowhere.)
Also, streaming is working now. The problem was that Scotty closes the
connection after it sees a newline in the body, or something, so streaming over
Scotty doesn't actually work. It's fine, Servant is better anyway.
|
|
Getting me closer to the latest GHC. This release also includes my own packages
that I submitted some time ago.
GHCJS is not present in 21.05 for some reason, but I think it's back in master,
so I might do another upgrade soon, but for now I just disabled my GHCJS
support. I'm not really using it anyway.
I also had to bring it string-quote, update nixos-mailserver, and a few other
things.
|
|
This way I can develop on nixpkgs directly, and directly send patches upstream,
instead of trying to copy files back and forth and work with overlays.
Of course with private stuff that I will never publish, I should use overlays,
but that will probably not be very many things, I imagine.
|
|
I also upstreamed this to nixpkgs-dev
|
|
|
|
|
|
|
|
Eventually I'll just upstream these to nixpkgs-dev. I also want to make z into a
gemini-compatible zettlekasten. All in due time.
|
|
|
|
- structured log messages
- compact logs rewrite the line on bild completion
- using conduit for streaming output (although I think this isn't working quite
right)
- new Proc type for running subprocesses
- general code cleanup and refactoring
|
|
|
|
Part of a larger effort, but doing this iteratively in smaller chunks.
|
|
This makes scanning logs *much* easier. I figure keep it as simple as possible,
just red, yellow, and green.
I also added two spaces between labels in the log messages. It would be nice to
have a more structured logging system, but for now this works.
|
|
Also changes the --test option to a 'test' command. This is because running the
tests for a namespace/exe should never be combined with anything else: you
either want to run the tests, or not.
|
|
Now bild knows how to determine between modules that require ghcjs and ghc. It
also knows what *not* to build, meaning it won't try to build non-buildable nix
targets, for example (unfortunately this is just hardcoded for now), but it also
won't build scm or py targets that I haven't implemented yet. It just silently
fails, which is fine, because it means I can do `bild **/*` and everything just
works.
Of course, if I want to build scm code then I will have to implement that, but
that's not a priority right now.
|
|
I had to refactor Biz/Bild/Rules.nix. I also had to checkin my patched
hoogle.nix file, but I also upstreamed the patch to nixpkgs-dev so it shouldn't
stick around for too long.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
I reorganized in order to debug a problem with the nix overlays. I think
having small, focused overlayes is more understandable than one overlay
that references itself via `rec`. Also the error traces are easier to
follow with smaller overlays.
|
