implement lc0 neural network inference by NripeshN · Pull Request #41 · NripeshN/MetalFish

NripeshN · 2026-02-07T15:32:06Z

Note

Medium Risk
Build and dependency surface changes (protobuf generation, new abseil/zlib links, workflow tweaks) plus significant CMake refactoring could break cross-platform builds or packaging if any dependency resolution differs across runners.

Overview
Adds a new transformer-network inference pipeline (protobuf net.proto generation + src/nn/* sources) and wires it into MCTS/Hybrid via the NNWeights option, with Metal/MPSGraph implementations included for macOS.

Refactors build and runtime GPU integration: removes CUDA backend/build branches and the standalone GPU benchmark, renames/reshapes GPU NNUE integration (gpu_integration.*), and makes AB evaluation always use CPU NNUE (GPU reserved for batched MCTS evaluation). Also updates NNUE weight handling to use a top-level networks/ directory and adjusts download/copy paths.

Updates CI workflows to install protobuf/zlib/abseil across OSes (Windows via vcpkg) and modifies the Elo tournament workflow to be manual-only with optional PR comment posting by input pr_number.

^{Written by Cursor Bugbot for commit e61c002. This will update automatically on new commits. Configure here.}

- [x] Created protobuf format for network weights (adapted from compatible format) - [x] Implemented weight loader with gzip support (.pb/.pb.gz files) - [x] Implemented 112-plane position encoder - [x] Created policy mapping infrastructure (1858 move space) - [x] Implemented network interface with stub backend - [x] Created MCTS evaluator integration - [x] Updated build system (protobuf, zlib dependencies) - [x] Created comprehensive test suite - [x] All copyright headers updated to MetalFish - [x] All lc0/lczero references removed - [ ] TODO: Metal backend for transformer inference (requires ~2000 LOC) - [ ] TODO: Full policy mapping tables - [ ] TODO: Canonicalization transforms - [ ] TODO: Integration with ThreadSafeMCTS - [ ] TODO: Verification tests with actual network Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

- Fix typo: 'Directcopy' -> 'Direct copy' - Add error checking for dup() before gzdopen() - Fix castling rights encoding (use WHITE_OO/BLACK_OO instead of bitwise AND) - Fix policy mapping to stay within bounds (simplified but correct) - Specify minimum protobuf version (3.0) in CMakeLists

Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

…ization - Implemented full 8-position history encoding (was single position only) - Added canonicalization transforms (flip/mirror/transpose) - Implemented ChooseTransform() for optimal board orientation - Added bit manipulation helpers (ReverseBitsInBytes, ReverseBytesInBytes, TransposeBitsInBytes) - Proper auxiliary plane encoding (castling, en passant, rule50, etc.) - Support for all input formats (classical, canonical v1/v2, hectoplies, armageddon) - Fixed castling rights encoding with proper perspective - Added early stopping for canonical formats - Transform application to all piece and auxiliary planes Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

…ence - Create src/nn/metal/metal_network.h and metal_network.mm (~730 LOC) - Implement complete transformer architecture with MPSGraph - Support embedding, multi-head attention, FFN, layer norm - Handle WDL output, batch processing, multiple activations - Integrate with Network interface and CMake build system - Link MetalPerformanceShadersGraph framework Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

- Add comprehensive README.md for Metal backend - Fix manual memory management (disable ARC, explicit release) - Add null checks before releasing Metal objects - Document memory management approach in file header Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

- Implement NNMCTSEvaluator with policy and value extraction - Integrate NN evaluator into ThreadSafeMCTS for policy priors and evaluation - Update test_nn_comparison with comprehensive NN testing - Add fallback to GPU NNUE when NN weights unavailable - Use vector<pair<Move,float>> to avoid std::hash<Move> issues Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

- Replace hardcoded 0.7f, 0.3f, 10000.0f with named constants - Improve code maintainability and make blending strategy clear - Add comments explaining the configuration Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

…e documentation Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

Resolved conflicts: - .gitignore: Kept both entries (_codeql files and network file) - CMakeLists.txt: Merged MCTS sources, NN sources, and framework linking - Combined all MCTS source files from both branches - Added NN_SOURCES section from HEAD - Merged Metal/CUDA framework linking for all targets - Merged GPU benchmark targets with all frameworks - src/mcts/thread_safe_mcts.cpp: Merged NN policy evaluation with softmax - Kept NN policy evaluation code from HEAD - Kept temperature-based softmax normalization from main

…s to queen Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

…l PR number input - Changed workflow trigger to only allow manual execution. - Added input for specifying a PR number to run the tournament on. - Updated concurrency group to use the provided PR number or run ID. - Modified comment handling to check for the PR number input instead of the event type.

…S, Ubuntu, and Windows - Added installation of protobuf and zlib dependencies for macOS and Ubuntu environments. - Introduced a new step for Windows to install dependencies using vcpkg. - Updated CMake configuration steps to accommodate the new dependency installations across different OS.

- Removed stockfish_adapter.cpp, hybrid_search.cpp, enhanced_hybrid_search.cpp, mcts_batch_evaluator.cpp, mcts_tt.cpp, and parallel_search.cpp from the MCTS_SOURCES list. - Streamlined the build configuration for MCTS components.

…tegration - Added custom commands in CMakeLists.txt to generate protobuf files from net.proto, ensuring compatibility with the installed protobuf version. - Updated NN_SOURCES to include generated protobuf files and adjusted include directories accordingly. - Modified CI workflows to install the abseil library alongside protobuf and zlib for macOS and Ubuntu environments. - Ensured that the build configuration links against the absl library where necessary.

- Enhanced CMakeLists.txt to better manage absl library detection and linking, including support for pkg-config on Linux. - Modified CI workflows to remove absl library installation, streamlining dependency management for Ubuntu environments.

…lative castling rights Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

Copilot

Pull request overview

This PR reintroduces an Lc0-compatible (transformer/BT4-style) neural-network inference pipeline into MetalFish, wiring together weight loading, input encoding, policy mapping, a Metal/MPSGraph backend, and MCTS integration for NN-guided search.

Changes:

Added NN infrastructure: protobuf weight schema, loader/decoder, 112-plane encoder, and 1858-move policy mapping.
Implemented/connected a Metal (MPSGraph) NN backend behind a Network interface with a stub fallback.
Integrated NN evaluation and policy priors into ThreadSafeMCTS, plus added a standalone NN comparison test and updated build/CI dependencies.

Reviewed changes

Copilot reviewed 31 out of 32 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
`tests/test_nn_comparison.cpp`	Adds a benchmark-style executable to exercise encoder, loader, network eval, and MCTS best-move checks.
`src/nn/proto/net.proto`	Introduces protobuf schema for NN weight files.
`src/nn/loader.{h,cpp}`	Adds weight loading, gzip decompression, and layer decoding.
`src/nn/encoder.{h,cpp}`	Adds 112-plane position encoding and canonicalization-related helpers.
`src/nn/policy_map.{h,cpp}`	Adds 1858-policy move mapping tables and lookup helpers.
`src/nn/network.{h,cpp}`	Defines abstract NN API + factory; uses Metal backend when enabled, else stub.
`src/nn/network_legacy.{h,cpp}`	Adds parsed/decoded weight structures for (multi)head formats.
`src/nn/metal/**`	Adds Metal/MPSGraph implementation scaffolding, builder, and tables.
`src/mcts/nn_mcts_evaluator.{h,cpp}`	Adds NN→MCTS bridge: encode, evaluate, map policy to legal moves.
`src/mcts/thread_safe_mcts.{h,cpp}`	Wires NN evaluator into MCTS expansion/eval and tracks NN eval stats.
`CMakeLists.txt`	Adds protobuf codegen, zlib/protobuf/abseil linking, Metal NN sources, and NN test target.
`.github/workflows/{ci.yml,elo-tournament.yml}`	Installs protobuf/zlib deps in CI; adjusts Elo workflow to manual trigger with PR number input.
`.gitignore`	Adds CodeQL build artifacts to ignore list.
`IMPLEMENTATION_SUMMARY.md`	Adds a large implementation summary document for the NN work.

src/nn/metal/metal_network.mm

src/nn/encoder.cpp

src/nn/policy_map.cpp

src/nn/encoder.cpp

.github/workflows/ci.yml

IMPLEMENTATION_SUMMARY.md

- Fix encoder missing vertical board flip for black to move: Apply ReverseBytesInBytes to all bitboards when black is the side to move, ensuring the board is always oriented from the side-to-move's perspective - Fix policy map incorrectly mapping knight promotions: Add missing knight promotion entries (a7a8n, b7a8n, etc.) to kMoveStrings table and change knight promotion mapping from 'q' to 'n'. Update kPolicyOutputs from 1858 to 1880 to account for the additional entries - Fix history perspective flip conditional: Remove the incorrect condition 'if (history_idx > 0)' that prevented perspective alternation for the first history position. Perspective should alternate unconditionally - Remove stale contradictory content from IMPLEMENTATION_SUMMARY.md: The file had duplicate content starting at line 240 that contradicted the completed implementation status shown earlier in the file

cursor · 2026-02-07T15:51:27Z

Bugbot Autofix resolved 4 of the 5 bugs found in the latest run.

✅ Fixed: Encoder missing vertical board flip for black to move
- Applied ReverseBytesInBytes to all bitboards when black is to move in both the simple and full encoder overloads, ensuring the board is oriented from the side-to-move's perspective.
✅ Fixed: Policy map incorrectly maps knight promotions to queen
- Added missing knight promotion entries (e.g., a7a8n, b7a8n) to kMoveStrings table, changed the knight promotion mapping from 'q' to 'n', and updated kPolicyOutputs from 1858 to 1880.
✅ Fixed: History perspective flip conditional on wrong check
- Removed the incorrect 'if (history_idx > 0)' condition so that the flip toggle alternates unconditionally for each history step as the NN training expects.
✅ Fixed: Contradictory implementation summary with stale appended content
- Removed the duplicate stale content starting at line 240 that contradicted the completed implementation status described in the first half of the document.

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

src/nn/encoder.h

src/nn/encoder.cpp

src/mcts/thread_safe_mcts.cpp

src/nn/loader.cpp

Bug 1: Policy table size mismatch (high severity) - Changed kPolicyOutputs from 1880 to 1858 (standard Lc0 encoding) - Removed queen promotions from kMoveStrings (they are encoded as regular moves) - Updated MoveToNNIndex to treat queen promotions as regular queen-direction moves - Only underpromotions (r/b/n) use explicit promotion indices 1792-1857 Bug 2: Encoder wrong color perspective (medium severity) - Changed perspective_us from pos.side_to_move() to 'us' (current position's STM) - All history positions now correctly encoded from current STM's perspective - Board flip now consistently uses current STM for all history entries Bug 3: ApplyNNPolicy incorrect logit clamping (medium severity) - Removed incorrect clamping of negative policy logits to 0.0f - Neural network logits can be negative (indicating low-probability moves) - Clamping destroyed the relative ordering among low-probability moves Bug 4: Float16 denormalized conversion (low severity) - Fixed denormalized fp16 to fp32 conversion with proper renormalization - Now correctly finds leading 1 bit and adjusts exponent accordingly - Value = 2^(-14) × (mantissa/1024) is now properly represented in fp32

cursor · 2026-02-07T17:05:02Z

Bugbot Autofix prepared fixes for 4 of the 4 bugs found in the latest run.

✅ Fixed: Policy table size mismatch breaks all promotion moves
- Changed kPolicyOutputs from 1880 to 1858, removed queen promotions from kMoveStrings table, and updated MoveToNNIndex to treat queen promotions as regular queen-direction moves.
✅ Fixed: Encoder uses wrong color perspective for history positions
- Changed perspective_us from pos.side_to_move() to the current position's 'us' variable, ensuring all history positions are encoded from the consistent perspective of the current STM.
✅ Fixed: ApplyNNPolicy clamps negative logits distorting policy distribution
- Removed the incorrect clamping of negative logits to 0.0f, allowing negative logits to properly flow through softmax to preserve the relative ordering of low-probability moves.
✅ Fixed: Incorrect float16 denormalized number conversion
- Implemented proper denormalized fp16 to fp32 conversion by finding the leading 1 bit in the mantissa and correctly computing the fp32 exponent as (103 + leading_bit) with proper fraction alignment.

src/nn/policy_map.cpp

CMakeLists.txt

Bug fixes: 1. MoveToNNIndex: Apply transform before attention policy map lookup - Previously, the attention policy map was queried with raw (untransformed) squares and returned early, bypassing the transform logic entirely - Now transforms are applied first, ensuring canonical move mapping works correctly for non-promotion moves 2. TransformSquare: Fix incorrect transpose implementation - Previously used OR-based conditions that treated kTransposeTransform as both flip and mirror (180° rotation) - Now correctly applies transforms in sequence: flip (file→7-file), mirror (rank→7-rank), transpose (swap file↔rank), matching the encoder's ApplyTransform behavior 3. CMakeLists.txt: Add missing Accelerate framework to test_nn_comparison - The test_nn_comparison target includes MCTS sources that use Accelerate framework functions on macOS, but was missing the framework link

cursor · 2026-02-08T00:00:39Z

Bugbot Autofix prepared fixes for 3 of the 3 bugs found in the latest run.

✅ Fixed: MoveToNNIndex bypasses transform for attention policy map
- Fixed by applying the transform to squares before the attention policy map lookup, ensuring transformed squares are used for canonicalized positions.
✅ Fixed: TransformSquare incorrect for transpose transform flag
- Fixed by implementing transforms in the correct order (flip, mirror, transpose with swap) to match the encoder's ApplyTransform behavior.
✅ Fixed: Missing Accelerate framework link for test_nn_comparison target
- Added ${ACCELERATE_FRAMEWORK} to the target_link_libraries for test_nn_comparison on Apple platforms.

- Add policy softmax temperature (1.359) to match Lc0 default - Simplify TransformSquare logic in policy_map.cpp - Add missing <algorithm> include for std::exp - Add comment clarifying value is from side-to-move perspective

src/mcts/thread_safe_mcts.cpp

src/mcts/tree.cpp

Bug fixes: 1. Remove incorrect nn_policy > 0.0f check in expand_node - Neural network policy outputs are raw logits, not probabilities - Logits can be negative for less preferred moves - The check was incorrectly filtering out valid moves with negative logits - Now all moves are blended with NN policy regardless of logit sign 2. Cache NN evaluation result from expand_node to avoid redundant evaluation - expand_node now accepts an optional EvaluationResult* output parameter - When expand_node performs NN evaluation for policy priors, it caches the result - The cached result is reused for value evaluation in run_iteration - This eliminates redundant NN evaluations that were performing GPU sync

cursor · 2026-02-08T01:22:51Z

Bugbot Autofix prepared fixes for 2 of the 2 bugs found in the latest run.

✅ Fixed: NN policy logit check incorrectly filters negative values
- Removed the incorrect nn_policy > 0.0f check since neural network policy outputs are raw logits that can be negative, and all moves should be blended regardless of logit sign.
✅ Fixed: Redundant NN evaluation in expand_node fallback path
- Modified expand_node to accept an optional EvaluationResult output parameter that caches the NN result, which is then reused in run_iteration for value evaluation, eliminating redundant NN calls.

- Updated CMakeLists.txt to reflect new directory structure for source files: - Changed main.cpp path to src/app/main.cpp - Updated benchmark source paths to src/bench/benchmark_gpu.cpp and src/bench/paper_benchmark.cpp - Introduced new source files for GPU benchmarking and paper benchmark suite, enhancing performance measurement capabilities for the MetalFish engine.

src/nn/metal/metal_network.mm

CMakeLists.txt

src/nn/encoder.cpp

- Fix Metal network to use actual batch size instead of max_batch_size_ for GPU inference, eliminating ~256x performance waste in MCTS - Update CI workflow to download NNUE files to networks/ directory to match CMakeLists.txt paths - Remove incorrect mirror_next/pos.flip() logic in encoder that corrupted piece color extraction for historical positions

cursor · 2026-02-09T12:37:34Z

Bugbot Autofix prepared fixes for 3 of the 3 bugs found in the latest run.

✅ Fixed: Metal network always evaluates max batch size
- Changed eval_batch from max_batch_size_ to batch so GPU inference uses the actual batch size instead of always running 256 positions.
✅ Fixed: NNUE file paths mismatch between CI and CMakeLists
- Updated all CI workflow NNUE download and copy paths from src/ to networks/ to match the CMakeLists.txt configuration.
✅ Fixed: History encoding flip corrupts piece color extraction
- Removed the incorrect mirror_next/pos.flip() logic since Stockfish uses absolute coordinates and the vertical flip for BLACK STM is already handled by ReverseBytesInBytes.

Rewrite README with accurate directory layout, current features, Apple Silicon optimizations, and all three engine modes documented. Remove outdated references and stale performance numbers.

CMakeLists.txt

- Updated CMakeLists.txt to enable ARC for Metal ObjC++ files, enhancing compatibility with MPSGraph. - Revised README to clarify engine modes, UCI options, and Apple Silicon optimizations. - Enhanced MCTS and hybrid search implementations, including better GPU resource management and error handling. - Streamlined evaluator batch processing to reduce overhead and improve efficiency. - Adjusted Metal network initialization to utilize dynamic batch sizes, eliminating unnecessary zero-padding. - Cleaned up UCI engine options related to GPU usage and transformer network weights, ensuring clarity in configuration.

- Introduced GatherBatchEvaluator to enable queue-based cooperative batching of evaluation requests, improving efficiency in handling multiple worker submissions. - Updated ThreadSafeMCTS to utilize GatherBatchEvaluator, allowing workers to submit positions and wait for batch evaluation results. - Enhanced worker context management by assigning unique IDs for gather slot assignment. - Improved cancellation handling for waiting workers during evaluation processes. - Refactored evaluation logic to support batch processing, reducing overhead and improving performance.

cursor

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is ON, but it could not run because the branch was deleted or merged before Autofix could start.}

cursor · 2026-02-09T21:12:55Z

src/eval/nnue/network.cpp

+      std::string msg4 =
+          "The default net can be downloaded from: "
+          "https://github.com/NripeshN/MetalFish/releases/download/nnue/" +
+          std::string(evalFile.defaultName);


NNUE download URL in error message likely invalid

Medium Severity

The NNUE download URL in the fatal error message was changed from the working Stockfish server (tests.stockfishchess.org/api/nn/) to github.com/NripeshN/MetalFish/releases/download/nnue/, but the CI workflows in this same commit still download NNUE files from the old Stockfish URL. This strongly suggests the GitHub releases URL doesn't exist yet. When users hit this exit(EXIT_FAILURE) path because NNUE files are missing, the error message will direct them to a non-working URL, leaving them unable to resolve the issue.

Copilot AI and others added 21 commits January 24, 2026 15:32

Initial plan

6c6d985

Implement full policy tables with 1858 move mappings

1f471ef

Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

Implement Metal/MPSGraph transformer backend for NN inference

6dc2713

Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

Complete verification tests with all 15 benchmark positions and updat…

05311a1

…e documentation Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

Fix PR review issues: remove build artifacts and map knight promotion…

b3f7f9f

…s to queen Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

Fix critical bugs: recalculate max_score after NN blending and use re…

6593b74

…lative castling rights Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

Fix castling plane order consistency and invalid move policy index

c63ee6c

Co-authored-by: NripeshN <86844847+NripeshN@users.noreply.github.com>

Add MetalFish transformer inference

b43f2a6

Copilot AI review requested due to automatic review settings February 7, 2026 15:32

NripeshN changed the title ~~Copilot/implement lc0 neural network inference again~~ implement lc0 neural network inference again Feb 7, 2026

Copilot started reviewing on behalf of NripeshN February 7, 2026 15:32 View session

NripeshN changed the title ~~implement lc0 neural network inference again~~ implement lc0 neural network inference Feb 7, 2026

Copilot AI reviewed Feb 7, 2026

View reviewed changes

src/nn/metal/metal_network.mm Show resolved Hide resolved

cursor bot reviewed Feb 7, 2026

View reviewed changes

src/nn/encoder.cpp Outdated Show resolved Hide resolved

src/nn/policy_map.cpp Outdated Show resolved Hide resolved

src/nn/encoder.cpp Outdated Show resolved Hide resolved

.github/workflows/ci.yml Show resolved Hide resolved

IMPLEMENTATION_SUMMARY.md Outdated Show resolved Hide resolved

Update src/nn/metal/metal_network.mm

919a924

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Handle missing NN history states

a9dde74

cursor bot reviewed Feb 7, 2026

View reviewed changes

src/nn/encoder.h Outdated Show resolved Hide resolved

src/nn/encoder.cpp Show resolved Hide resolved

src/mcts/thread_safe_mcts.cpp Outdated Show resolved Hide resolved

src/nn/loader.cpp Show resolved Hide resolved

NripeshN and others added 2 commits February 7, 2026 17:02

Implement MetalFish Lc0 inference

db792a8

Fix NN comparison mismatch initial

1e18020

cursor bot reviewed Feb 7, 2026

View reviewed changes

src/nn/policy_map.cpp Show resolved Hide resolved

src/nn/policy_map.cpp Show resolved Hide resolved

CMakeLists.txt Show resolved Hide resolved

NripeshN added 2 commits February 8, 2026 00:51

Add policy softmax temperature and simplify transform logic

53ab23a

- Add policy softmax temperature (1.359) to match Lc0 default - Simplify TransformSquare logic in policy_map.cpp - Add missing <algorithm> include for std::exp - Add comment clarifying value is from side-to-move perspective

delete ref

54459a3

cursor bot reviewed Feb 8, 2026

View reviewed changes

src/mcts/thread_safe_mcts.cpp Outdated Show resolved Hide resolved

src/mcts/tree.cpp Show resolved Hide resolved

cursor bot reviewed Feb 9, 2026

View reviewed changes

src/nn/metal/metal_network.mm Outdated Show resolved Hide resolved

CMakeLists.txt Show resolved Hide resolved

src/nn/encoder.cpp Outdated Show resolved Hide resolved

NripeshN added 6 commits February 9, 2026 13:03

refactor codebase

1121bde

Update README to reflect restructured codebase

f97b5cd

Rewrite README with accurate directory layout, current features, Apple Silicon optimizations, and all three engine modes documented. Remove outdated references and stale performance numbers.

fast operations

1051d59

remove cuda tests

0b93f6b

UCI clean up

498a7a8

clean tests

94538b2

cursor bot reviewed Feb 9, 2026

View reviewed changes

CMakeLists.txt Show resolved Hide resolved

CMakeLists.txt Show resolved Hide resolved

NripeshN added 2 commits February 9, 2026 19:30

NripeshN merged commit 45ae05a into main Feb 9, 2026
5 of 8 checks passed

NripeshN deleted the copilot/implement-lc0-neural-network-inference-again branch February 9, 2026 21:01

cursor bot reviewed Feb 9, 2026

View reviewed changes

Conversation

NripeshN commented Feb 7, 2026 • edited by cursor bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

cursor bot commented Feb 7, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

cursor bot commented Feb 7, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

cursor bot commented Feb 8, 2026

Uh oh!

Uh oh!

Uh oh!

cursor bot commented Feb 8, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

cursor bot commented Feb 9, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

cursor bot left a comment

Choose a reason for hiding this comment

Uh oh!

cursor bot Feb 9, 2026

Choose a reason for hiding this comment

NNUE download URL in error message likely invalid

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

NripeshN commented Feb 7, 2026 •

edited by cursor bot

Loading