Skip to content

@gcp gcp released this Apr 4, 2019

  • Implemented batching for GPUs, increasing speed.
  • Added support for Tensor Cores on NVIDIA cards.
  • Optimized Winograd code for CPU and GPU.
  • Improved UCT formula.
  • Improved root move selection by using Lower Confidence Bounds.
  • Improved passing heuristics (disabled in self-play).
  • Improved autotuner behavior in edge cases.
  • The training code now supports mixed precision fp32/fp16 training.
  • AutoGTP can now start from an arbitrary SGF position.
  • AutoGTP can now play matches with different white/black parameters.
  • AutoGTP can now send initial commands to the engine.
  • AutoGTP now provides the full engine options in the SGF comments.
  • AutoGTP now uses batching during self-play.
  • GTP: Add support for gomill-explain_last_move.
  • GTP: lz-analyze and lz-genmove_analyze now support an avoid tag.
  • GTP: lz-analyze and lz-genmove_analyze now support a minmoves tag.
  • GTP: lz-analyze and lz-genmove_analyze now support a (post) interval tag.
  • GTP: loadsgf and printsgf can now save and load time control information.
  • Bugfix: lz-setoption name playouts now works correctly.
  • Bugfix: in analysis mode, provide output even if we did not have to search.
  • Bugfix: detect if the weights file is for the wrong board format.
  • Bugfix: bail out if OpenCL doesn't find a GPU (instead of using OpenCL-on-CPU).
Assets 6

@gcp gcp released this Oct 31, 2018 · 80 commits to master since this release

  • Added support for lz-analyze and lz-genmove_analyze GTP extensions.
    They provide real time information about the search and can be used by GUIs.
  • Added support for lz-setoption to change options without restarting.
    Support setting max memory usage, visits, etc.
  • Implemented Winograd F(4x4, 3x3) for CPU and GPU, speeding up large networks.
  • Symmetrical positions in the opening are now recognized and sped up through mirroring/rotations.
  • Storing coefficients in fp16 format is automatically enabled if it is faster.
  • Support for fp16 compute. Large speedup on cards with fp16 compute support.
  • Improved scaling of MCTS search with large number of threads.
  • Added specific time management for self-play training games.
  • Bugfixes to SGF parsing.
  • Bugfixes to GTP compatibility.
  • AutoGTP: support compressed weights, removing the need to ungzip.
  • AutoGTP: use HTTPS.
  • Validation: support arbitrary GTP commands for each binary.
  • Training code now supports multi-GPU.
  • Bundled OpenBLAS is updated with support for new CPUs.
  • (Open)BLAS is no longer required for compiling.

Note 2018-11-01: ZIP packages have been re-uploaded with missing gzip added.

Assets 6

@gcp gcp released this May 9, 2018 · 199 commits to master since this release

This is a bugfix release for training game generation.

  • Bugfix to Dirichlet noise being more uniform than intended (bug was introduced in v0.13).
  • Tweaks to randomized move selection in training games to reduce needless blunders. Added extra options.
Assets 6

@gcp gcp released this May 4, 2018 · 207 commits to master since this release

  • Support for converted ELF Open Go networks.
  • Network files no longer need to be ungzipped.
  • Improved time usage.
    • Slowed down play, more consistent time usage from opening to middle game.
    • By default no longer exits early in fixed playouts or in Japanese byo yomi.
    • The engine will respond faster on large searches, tree cleanup happens in the background.
  • Memory usage for large searches further reduced.
  • Fix a bug in the locking implementation that could lead to data races.
  • Default tuning is faster and will select faster kernels on most devices.
  • Various other small fixes and improvements.
Assets 6

@gcp gcp released this Mar 24, 2018 · 267 commits to master since this release

  • Added --benchmark option.
  • Improved strength on long/deep searches, reduced memory usage.
  • Improved time management logic.
  • AutoGTP now cleans up old networks.
  • Optimizations for CPU-only version, about 10% speedup.
  • Bugfix: self-play games are now more diverse because "noise" moves are no longer pruned.
  • Bugfix: pondering now respects playout/visit limits.
Assets 6

@gcp gcp released this Feb 19, 2018 · 318 commits to master since this release

  • Optimized, fused Winograd transforms.
  • Optimized final convolutions and OpenCL data transfer.
  • Basic time management logic for non-self-play games.
  • Stronger search due to improved FPU calculation.
  • AutoGTP supports stop/resume (try q+Enter).
  • Replace transposition table by network cache.
  • Bugfixes to time management with tree reuse.
  • Avoid some crashes on buggy OpenCL drivers.
  • Avoid crashes when the user/GUI plays an illegal move.
Assets 6

@gcp gcp released this Jan 19, 2018 · 403 commits to master since this release

  • Faster network evaluation with Winograd transforms.
  • OpenCL code now automatically tunes to the hardware on first run.
  • Search tree is reused between moves.
  • Support for multiple GPU when playing games.
  • Various AutoGTP fixes and enhancements.
Assets 6

@gcp gcp released this Jan 8, 2018 · 466 commits to master since this release

  • Make OpenCL self-check failures more visible.
  • Work around an issue with OpenBLAS & threading.
Assets 6

@gcp gcp released this Jan 7, 2018 · 469 commits to master since this release

  • Speedups and bugfixes to the Leela Zero engine.
  • GPU version now checks OpenCL implementation for correctness.
  • CPU-only version for people without usable GPU.
  • AutoGTP now fetches self-play or match tasks and parameters from the server.
  • Updated OpenBLAS to 0.3.0-dev with support for newer CPUs.
Assets 6

@gcp gcp released this Dec 8, 2017 · 581 commits to master since this release

  • Fixes the noise generation for policy priors.
  • Improved handling of random seeding.
Assets 3
You can’t perform that action at this time.