Customise a model from the portal: a file, a model, a button

A page for anyone in a project, built for the one who knows least: an
Excel, CSV or JSONL file of questions and answers, previewed; a model; a
name; who may use it; and a button. The weights to train on are found
from where the model's came (a GGUF repository's base model), the card's
memory from their size, the node as the one with the most free. When it
is done the adapter is published by itself (a llama.cpp model's replicas
replaced to load it) and is a model of its own in the chat and the API.

The project's by default; its members share it wider (confirmed, and
never wider than the knowledge bases its pairs came from) or delete it.
One trains at a time per project. ctl train start needs only --model and
--data now, and --publish publishes when done.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Ilker ManapandClaude Opus 5.5 committed 2026-11-19 11:31:12 +01:00
1 parent fcda4d3f85
commit a53b2f4be9
33 files changed
+2199 -134

No files matched your search

+9
View File
@@ -2107,6 +2107,15 @@ lent, partition`); the leader tells each named node's agent (root) what to be, t
controller ~20 s later, sbatch ran on n1 as ada. Disabled in AD: suspended in 56 s, the Linux user
gone in 89 s, ssh refused, home kept. Enabled: back in 32 s, same UID, key kept. Taken out of
ml-team: out of the team in 11 s; put back: in again in 56 s.
- P5, replaced 2026-11-17 by the portal (decided 2026-11-15: users of every level, build for the one who
knows least first; any project member may start, no approval, one per project, the project's by default):
**Customise a model** (`/train`): an xlsx/CSV/JSONL of questions and answers, checked and previewed; base
weights from the model's origin (a GGUF repo's card base_model), memory from their size (2x + 1 GiB),
the node with the most free; published by itself when done (llama.cpp replicas replaced to load it);
shared me/project/team/org/everyone (wider than the project only confirmed, never wider than its
knowledge bases), deleted from the page. Left: pairs generated from a knowledge base on the page, a
base-vs-new comparison, the training's GPU time charged to the project, big jobs through Slurm, the
server's own error messages in Turkish.
- P6, done 2026-11-15 but for budgets: each project a Slurm account (org and team accounts above it,
ids keyed by kind: the first install's org, team and project share the id "default"), its members and its
team's and org's on it; AccountingStorageEnforce=associations; the controller's agent keeps them with