# Changelog

## v0.1.3

- Running the one-line installer again while onw is running (an update) now restarts the tray on the new version,
  loading the model again, instead of leaving the old tray running until a manual restart. Trays of 0.1.2 and older
  are stopped and the new tray takes over.
- Model download progress comes from the bytes FastFlowLM reports and the model's real file sizes; it used to count
  finished files, so it sat at 0% while a big weights file came down (about 20 GB for qwen3.6-moe-35b-a3b).
- Model files are checked against FastFlowLM's hashes: right after a download (damaged files are fetched again
  once), on demand from the Models tab (... -> "Check the files", then "Download again" for damaged ones), and by size
  before every load (broken weights give garbage such as `<unused49>`).
- The model list no longer goes empty when `flm list` returns nothing because of one damaged model folder: it then
  comes from the `model_list.json` that ships with flm.
- Fixed: installing FastFlowLM could set the chosen model to "engine".
- Gemma 4's thinking (`<|channel>thought`) is split off too when the answer was cut off while thinking.
- Windows: a process that did not get the single-instance lock no longer keeps holding it (it blocked the take-over).

## v0.1.2

- Linux: the installer (`setup.sh`) raises the memlock limit (required by FastFlowLM's Linux guide; Ubuntu's default
  fails `flm validate`) and asks for a restart when it did.
- Models tab: how to fix a memlock failure, and a warning when another FastFlowLM is running (Lemonade Server, a
  system fastflowlm package's `flm serve`): they compete for the NPU.
- `diag.py` also reports running FastFlowLM processes and installed fastflowlm / XRT / Lemonade packages.

## v0.1.1

- FastFlowLM updated to v1.0.7 (v1.0.6 failed every Linux kernel older than 6.17; it is replaced at the next download).
- When FastFlowLM's NPU check (`flm validate`) fails, the Models tab says why (driver version, a link to the driver
  guide, "Check again", flm's own report); downloads stop with that reason.
- New `diag.py`: NPU, driver, firmware, `/dev/accel`, memlock and FastFlowLM state in one file; `--try-model` runs a
  small model on the NPU to tell a wrong check from a real failure.

## v0.1.0

First release of the AMD Ryzen AI NPU edition, based on onw (Intel NPU edition) v0.11.3.

- Inference replaced with the NPU-only engine FastFlowLM v1.0.6 (fetching, starting and proxying ported from
  Lemonade Server), downloaded on first use.
- Models: FastFlowLM's chat models. Nothing runs on the CPU, the GPU or as a hybrid.
- Thinking switches mapped to FastFlowLM's `think`; the thinking comes back as `reasoning_content`.
- Windows: FastFlowLM is tied to onw's server by a job object, so a crashed server never leaves it holding the NPU.
- Detects the XDNA2 NPU and its driver; without one, downloads and loading stop and say why.
- Intel-NPU-only settings removed (device, prompt lookup decoding, 64-token blocks, NPUW, expert memory).
- The check page measures time to first token and generation speed by streaming.
- The OpenVINO engine (graph building, KV cache, converter, measuring scripts) removed.