Skip to content

validator: on a small testnet card the probe agent is a small model, the reference runs its quantised build, and only declared bases are fetched and run - #173

Merged
BANADDA merged 1 commit into
mainfrom
testnet-fast
Oct 27, 2026
Merged

BANADDA merged 1 commit into
mainfrom
testnet-fast

Conversation

@BANADDA

@BANADDA BANADDA commented Oct 27, 2026

Copy link
Copy Markdown
Member

A single GPU under 40 GB cannot hold the 7B probe agent next to the 7B reference with 32k context. On testnet the validator now places, automatically: the probe agent as Qwen2.5 0.5B Instruct at a 0.15 memory share, the reference SLM as the AWQ build of the reference model (vLLM awq_marlin) at 0.8, and each miner's model at 0.8. It also reads each roster entry's declared base (manifest, then training/recipe.json) and fetches and runs only those bases, instead of every allowed base. Mainnet placement is unchanged and still refuses a card without room.

🤖 Generated with Claude Code

…the reference runs its quantised build, and only declared bases are fetched and run
@BANADDA
BANADDA merged commit 7fb4565 into main Oct 27, 2026
6 checks passed
@BANADDA
BANADDA deleted the testnet-fast branch October 27, 2026 06:15
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant