v0.6.90b420.dev731
This release is AMD's. lilbee ran on an RX 9060 XT and an MI300X, and everything that broke along the way is fixed.
MASSIVE thanks to my friend Jim G. for letting me test on his 9060.
The ROCm flatpak starts
The engine bundle expected libnuma from the host, which the freedesktop runtime does not ship, so the flatpak died at launch. libnuma travels with the bundle now, and a container gate proves the bundled copy resolves on a machine with no ROCm installed at all.
ask stopped spending all its time reranking
Rerank pairs went one per embeddings request, each with its own tokenize round-trip, a constraint left over from an in-process batcher that no longer exists. All pairs go in one batched request now. Measured on hardware: an ask that took 86 seconds finishes that stage in 4.6.
A warm engine gets reused
Context sizing was baked into the engine pin, so a one-shot ask and a running serve computed different targets and refused to share an engine. ctx is a fits-within check now. ask adopts the engine serve already has running instead of restarting it twice.
GPU activity on AMD means something again
amdgpu pins its busy flag at 100% for any process holding a compute context, so the fleet panel bar and the adaptive ingest controller both saw a permanently saturated card whether it was idle or generating. Utilization comes from power draw over the board cap now: 10% at bare idle, 25% with a model resident, 94% during prompt processing. Verified live on an RX 9060 XT and an MI300X, and the amd-smi path was grounded against real captures from datacenter hardware.
No more surprise downloads
A fresh install running a command with no models used to quietly fetch the defaults. Missing models error with guidance now, or swap to an installed model for that run only. Nothing gets persisted from a one-shot command, and an explicit --model is honored as given.
ask prints the answer and nothing else
Warnings and log records were landing in ask's stdout, which broke anyone piping it. Diagnostics go to cli.log, and stdout carries the answer alone. Source links stay clickable on terminals that support them and keep their URL as plain markdown on the ones that don't, so the path survives either way.
The model catalog stopped crashing the TUI
Opening the catalog could crash when a deferred grid mount re-queried a container that had already gone away. The callback uses the container it was scheduled with and bails when that container is no longer running.
Credentials stay hidden
/set openai_api_key sk-... echoed the key back in the clear in the confirmation toast. The echoed value is masked now. The confirmation still names the key, so the command reads the same.
Docs
tested-gpus gains an RDNA4 row and the AMD resident-context finding: the busy flag is pegged, power is the signal.