Repository navigation
server: POST /lora-adapters/snapshot writes a mind's shadow to an immutable GGUF between steps, and /learn reports the shadow's committed step - #86
Open
joelteply wants to merge 1 commit into
Conversation
…utable GGUF between steps, and /learn reports the shadow's committed step
What a shadow learned on a serving context lived only in memory: a restart
lost it, and nothing could hand it to a gene trial (continuum S4 (d), card
76b05d2a; ONE-RESIDENT-MODEL §9.2 "Snapshot").
- POST /lora-adapters/snapshot {id, name}: on the server loop, between
steps (never a half-applied update), the adapter's current weights are
written beside the name and renamed into place inside --train-dir
(confine_out, the load route's rule). An existing name is refused: a
snapshot is immutable. The reply carries the path, bytes and the steps
the shadow had committed.
- /learn's outcome carries step: the shadow's committed optimizer steps
after this one (each shadow counted on its own; unload forgets it), the
commit the core's cursor stands on.
- learn-every-format.py --snapshot: two steps (step 1, 2), a snapshot at
steps 2, a second snapshot of that name refused, and the snapshot loaded
under a new id shows the forward loss the shadow showed at its third
step (2.526167, both), so the file is the trained weights. Qwen2.5-0.5B,
IntelMac CPU, fresh server.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FRQWzgSfo79JtnwZywHKrE
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What a shadow learned on a serving context lived only in memory: a restart
lost it, and nothing could hand it to a gene trial (continuum S4 (d), card
76b05d2a; ONE-RESIDENT-MODEL §9.2 "Snapshot").
steps (never a half-applied update), the adapter's current weights are
written beside the name and renamed into place inside --train-dir
(confine_out, the load route's rule). An existing name is refused: a
snapshot is immutable. The reply carries the path, bytes and the steps
the shadow had committed.
after this one (each shadow counted on its own; unload forgets it), the
commit the core's cursor stands on.
steps 2, a second snapshot of that name refused, and the snapshot loaded
under a new id shows the forward loss the shadow showed at its third
step (2.526167, both), so the file is the trained weights. Qwen2.5-0.5B,
IntelMac CPU, fresh server.
Co-Authored-By: Claude Opus 5.5 (1M context) noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01FRQWzgSfo79JtnwZywHKrE