Issues

• perf: ParaformerLarge full-scale eager BF16 8-bit Adam step time grows 91 s -> 150 s between steps 1 and 2

  ooples/AiDotNet#2322 · Filter by author ooples opened on Oct 13, 2026
  1

• Fifteen wav2vec2-family recognizers share a generic stack instead of their papers' networks

  ooples/AiDotNet#2321 · Filter by author ooples opened on Oct 13, 2026

• Plan: beat the best rival 2x on every model family at the foundation (CPU + GPU, training + inference)

  ooples/AiDotNet#2300 · Filter by author ooples opened on Oct 7, 2026
  2

• 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
  3

• 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
  3

• [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
  2

• 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

• 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

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

  ooples/AiDotNet#1706 · Filter by author ooples opened on Jun 27, 2026
  6

• 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
  10

• Inference/compilation: industry-standard gap audit + exceed-industry levers (value-aware replay verified, verify-then-trust, fused attention, paged KV-cache, ...)

  ooples/AiDotNet#1630 · Filter by author ooples opened on Jun 16, 2026
  4

• Foundation-scale model CI: eliminate all timeouts/OOM via stacked acceleration (tracker)

  ooples/AiDotNet#1622 · Filter by author ooples opened on Jun 15, 2026
  2

• perf: exceed PyTorch-CPU eager on the AIsEval parity benchmark — route Transformer/MLP Predict to fused kernels + tighten the compiled train-step

  ooples/AiDotNet#1478 · Filter by author ooples opened on Jun 2, 2026
  2

• Multi-thread training story: HOGWILD! + in-process DDP + striped-lock trainers (exceed PyTorch DDP)

  enhancement
  ooples/AiDotNet#1369 · Filter by author ooples opened on May 18, 2026
  2

• Verify + document cuDNN/cuBLAS dispatch path on NVIDIA hosts

  ooples/AiDotNet#1159 · Filter by author ooples opened on Apr 18, 2026
  3

• Epic: AiDotNet SDK - Model Metadata & License Verification (Phases 1-2)

  ooples/AiDotNet#425 · Filter by author ooples opened on Nov 7, 2025
  4

• [Gap Analysis] Implement Comprehensive Model Hub Ecosystem

  ooples/AiDotNet#310 · Filter by author ooples opened on Nov 6, 2025
  3