Issues

• [bitnet] planesHeadFromEmbedding takes ~45 s single-threaded at load (tied 128k×2560 head) — parallelise, cache or AOT-convert

  enhancementskill:kotlin-core
  SKaiNET-developers/SKaiNET-transformers#502 · Filter by author michalharakal opened on Oct 17, 2026

• Llama decode wrong for SmolLM2-135M: early <|im_end|>, long-prompt degeneration, '1 2 3…9' → ' 1 1 1 1' — positions/RoPE suspected, tokenizer ruled out

  bug
  SKaiNET-developers/SKaiNET-transformers#483 · Filter by author michalharakal opened on Oct 11, 2026

• [Lane 5 · docs] tokenizer.json parity — maturity gate bullet, one documented entry point, tokenizer.json ↔ GGUF field mapping

  documentationgood first issuesize:xsskill:docssub-issue
  SKaiNET-developers/SKaiNET-transformers#478 · Filter by author michalharakal opened on Oct 10, 2026
  1

• [Lane 4 · kotlin-core] tokenizer.json parity — gated parity test replaying the goldens through both tokenizer.json entry points

  codinggood first issuesize:sskill:kotlin-coresub-issue
  SKaiNET-developers/SKaiNET-transformers#477 · Filter by author michalharakal opened on Oct 10, 2026
  1

• [Lane 4 · numerics] tokenizer.json parity — HF golden encode fixtures for every shipped family that loads a tokenizer.json

  codinggood first issuesize:sskill:numericssub-issue
  SKaiNET-developers/SKaiNET-transformers#476 · Filter by author michalharakal opened on Oct 10, 2026
  1

• moonshine stream: the final encoder window is zero-padded and the encoder has no attention mask

  SKaiNET-developers/SKaiNET-transformers#458 · Filter by author michalharakal opened on Sep 25, 2026

• iree-android: stateful prefill/step contract with a reusable prompt-prefix KV snapshot (and an embeddings-input variant for Vulkan)

  SKaiNET-developers/SKaiNET-transformers#410 · Filter by author michalharakal opened on Sep 3, 2026
  3

• FunctionGemma redecode contract: LM head + argmax run over all SEQ positions (36 % of every step, 1 GB scratch at SEQ 1024 in a 32-bit process); parameter archives are per-SEQ

  SKaiNET-developers/SKaiNET-transformers#406 · Filter by author michalharakal opened on Sep 3, 2026

• [gemma3n] Split into its own family module + conformance (postponed)

  SKaiNET-developers/SKaiNET-transformers#377 · Filter by author michalharakal opened on Sep 1, 2026
  2

• Retire kllama-cli's legacy dense-FP32 LlamaRuntime path; route Llama/Mistral GGUF through the DSL like Qwen already does

  SKaiNET-developers/SKaiNET-transformers#354 · Filter by author michalharakal opened on Aug 29, 2026

• Antora docs site has no version selector — pin per release instead of a single unversioned component

  SKaiNET-developers/SKaiNET-transformers#317 · Filter by author michalharakal opened on Aug 13, 2026

• Validate kllama Apple native-kernel fix (PR #315) on real iOS/macOS hardware

  help wanted
  SKaiNET-developers/SKaiNET-transformers#316 · Filter by author michalharakal opened on Aug 13, 2026
  1

• SmolLM2-135M compiled leg: DSL → StableHLO → IREE showcase (gemma-iree precedent)

  SKaiNET-developers/SKaiNET-transformers#305 · Filter by author michalharakal opened on Aug 12, 2026
  5

• Register the native cinterop kernels on Apple targets in the runtime facades

  SKaiNET-developers/SKaiNET-transformers#298 · Filter by author michalharakal opened on Aug 11, 2026

• Evaluate the existing BERT/MiniLM embedding model (feature/bert-dsl-embeddings): quality + on-device numbers, and a merge path

  SKaiNET-developers/SKaiNET-transformers#273 · Filter by author michalharakal opened on Aug 9, 2026
  1

• FunctionGemma export: convert MLIR-text rewrites to StableHLO/DAG transforms

  SKaiNET-developers/SKaiNET-transformers#248 · Filter by author michalharakal opened on Jul 24, 2026
  1

• feature/qwen35 (DeltaNet) rehab checklist before it can merge

  SKaiNET-developers/SKaiNET-transformers#235 · Filter by author michalharakal opened on Jul 17, 2026

• Phase 4 readiness: validate DSL Qwen output against HF reference on real GGUFs

  SKaiNET-developers/SKaiNET-transformers#118 · Filter by author michalharakal opened on May 4, 2026

• Kotlin/Native inference produces deterministic nonsense across model families and runtime paths (JVM unaffected)

  SKaiNET-developers/SKaiNET-transformers#104 · Filter by author michalharakal opened on May 2, 2026

• Apertus xIELU formula / param-space mismatch — forward pass NaN's on real Apertus-8B Q4_K_S

  SKaiNET-developers/SKaiNET-transformers#103 · Filter by author michalharakal opened on May 2, 2026

• Unified Model Pipeline with Decoupled Tool Calling

  SKaiNET-developers/SKaiNET-transformers#49 · Filter by author michalharakal opened on Apr 11, 2026
  2

• evaluate inference speedup with radix KV cache

  SKaiNET-developers/SKaiNET-transformers#45 · Filter by author michalharakal opened on Apr 9, 2026
  1

• Gemma 4 Multimodal & Advanced Features

  SKaiNET-developers/SKaiNET-transformers#34 · Filter by author michalharakal opened on Apr 8, 2026

• Gemma 4 Text-Only Support

  SKaiNET-developers/SKaiNET-transformers#33 · Filter by author michalharakal opened on Apr 8, 2026

• Add support for Gemma 4

  SKaiNET-developers/SKaiNET-transformers#32 · Filter by author michalharakal opened on Apr 8, 2026