Issues

• fix: BLIP's native image-text matching and caption decoder never see the image

  ooples/AiDotNet#2297 · Filter by author ooples opened on Oct 6, 2026

• [Tracking] Pull requests run the full ~2-hour CI matrix regardless of their change

  ooples/AiDotNet#2243 · Filter by author ooples opened on Sep 22, 2026
  1

• DINO / RT-DETR: replace dense global encoder attention with multi-scale deformable attention and AIFI

  ooples/AiDotNet#2171 · Filter by author ooples opened on Sep 11, 2026

• Detection/OCR: annotation-based training with each family's standard loss, and metric-improvement tests

  ooples/AiDotNet#2170 · Filter by author ooples opened on Sep 11, 2026

• 96 layer factories accept a NeuralNetworkArchitecture and never read it, silently mis-sizing 68 models

  v1-tier2
  ooples/AiDotNet#2147 · Filter by author ooples opened on Sep 10, 2026

• Model-family coverage: decide whether 224 compositional models should be auto-constructible

  v1-tier2
  ooples/AiDotNet#2139 · Filter by author ooples opened on Sep 9, 2026
  1

• Model-family coverage: 35 models declare a category without implementing its interface

  v1-tier2
  ooples/AiDotNet#2138 · Filter by author ooples opened on Sep 9, 2026
  2

• CI: a cancelled run is indistinguishable from a failing one, so a PR shows dozens of red checks with no cause

  v1-tier2
  ooples/AiDotNet#2131 · Filter by author ooples opened on Sep 9, 2026

• [Epic] Model-family invariants drive one entry path per model (plus: the coverage report is broken, and 345 models have no test class)

  v1-blocker
  ooples/AiDotNet#2091 · Filter by author ooples opened on Sep 5, 2026
  6

• Heavy Timeout Models (nightly) has failed 5 consecutive nights - deferred models are getting zero coverage

  v1-tier2
  ooples/AiDotNet#2087 · Filter by author ooples opened on Sep 5, 2026

• 464 options classes have no copy constructor - decide whether the golden pattern applies to all of them

  v1-tier2
  ooples/AiDotNet#1993 · Filter by author ooples opened on Aug 8, 2026
  2

• PR #1789: CI parallel-execution flakes (GAT, SpiralNet, XDecoder, CpuToWallRatio, MemoryBudget)

  ooples/AiDotNet#1857 · Filter by author ooples opened on Jul 12, 2026
  3

• GaussianSplatting densification (split/prune during training) not exposed through the facade

  ooples/AiDotNet#1835 · Filter by author ooples opened on Jul 9, 2026
  3

• Deep time-series models train single-threaded / underutilize GPU (unbatched per-op tape)

  ooples/AiDotNet#1804 · Filter by author ooples opened on Jul 6, 2026
  7

• ParaformerLarge: BF16 8-bit optimizer throws 'Source array was not long enough' at 661M-param scale

  v1-tier2
  ooples/AiDotNet#1767 · Filter by author ooples opened on Jul 2, 2026
  3

• Timeout models split into overhead-bound (fused-plan, fixable) vs compute-bound (GEMM) — PerfView triage + plan

  ooples/AiDotNet#1743 · Filter by author ooples opened on Jun 29, 2026
  4

• HeavyTimeout: retag to the real-data timeout list (correct #1709)

  ooples/AiDotNet#1714 · Filter by author ooples opened on Jun 28, 2026
  2

• [CI] Umbrella tracker: HeavyTimeout models deferred from the default test gate

  ooples/AiDotNet#1706 · Filter by author ooples opened on Jun 27, 2026
  5

• Foundation-scale CNN training-step perf: non-conv overhead + VGG16 ModelFamily timeouts (#1684 follow-up)

  ooples/AiDotNet#1694 · Filter by author ooples opened on Jun 25, 2026
  2

• epic: grouped/depthwise deformable convolution (DCNv3) — engine kernel across CPU + 6 GPU backends, then faithful InternImage

  ooples/AiDotNet#1691 · Filter by author ooples opened on Jun 25, 2026
  8

• CV segmentation: 7 foundation models OOM on single-model training (UniVS/UNINEXT/MaskDINO/QueryMeldNet/OneFormer/KMaXDeepLab/SegGPT)

  ooples/AiDotNet#1688 · Filter by author ooples opened on Jun 25, 2026
  8

• Playbook: clear remaining #1677 OOM/timeout CI shards — memory-architecture fixes (caching arenas, fused optimizer-in-backward, COW clone, weight streaming) + float as the LAST lever

  ooples/AiDotNet#1684 · Filter by author ooples opened on Jun 24, 2026
  2

• CI: foundation-scale OOM on 16GB runner — triage after arena fix (#1668)

  ooples/AiDotNet#1672 · Filter by author ooples opened on Jun 23, 2026
  4

• Training backward-pass memory/throughput optimizations (optimizer-in-backward, default checkpointing, FlashAttention backward, backward buffer-reuse)

  ooples/AiDotNet#1662 · Filter by author ooples opened on Jun 22, 2026
  2

• Inference stack is built but dormant: enable by default + wire orphaned components (paged-KV, spec-decode, continuous batching, int8) into the inference path

  ooples/AiDotNet#1632 · Filter by author ooples opened on Jun 16, 2026
  9