v34

• New locking library to work on NFSv4 across multiple machines for cluster users.
• Linux autoupdate script.

v33

• Network mirror support.
• Use go modules.
• Blacklist trivial lc0 backend.

v32

• Timeout for painfully slow downloads.
• More network prefetching.

v31

• Don't increase parallelism for GTX16xx GPUs.
• Use special exit code when update needed.

V30

• Support for the cuda lc0 backend, with increased parallelism for cuda-fp16.
• Delegate to lc0 the selection between fp16 and fp32.

v29

• A mechanism was added to support opening book updates.
• The downloaded opening book will now go in more normal user cache locations on mac and linux.
• The --cache option now specifies a top level directory (that should exist) under which the books and client-cache subdirectories will be created.

v28

No changes - but release was accidentally bumped to v28 already, so lets make it official.

v27

• The release appveyro build now creates lc0-training-client instead of client.
• settings.json has been renamed to lc0-training-client-config.json and it will be created in more 'normal' user config locations on mac and linux. The new --config option can be used to override.
• Training network downloads will now go in more normal user cache locations on mac and linux. The new --cache option can be used to override.
• The client can now execute lc0 found in the PATH if there is none found in the current directory.
• dx12 check should hopefully work properly in this version.

v26

New in this release:

• New dx12 backend sanity check with lower impact in performance.

• Infrastructure for training from FEN positions.

• Chess960 support.

v25

New in this release:

• A new flag localhost to change the host name reported to the server. This can also be set in the settings.json file.
• Fix for the changed name of the dx12 backend in lc0.

Also note the new release naming scheme (v25 instead of v0.25.0), hopefully to reduce confusion with lc0 versions.