All pull requests

Pull requests

• fix(cuda): widen byte-size products on the weight-upload path (#432)

  #433·jamesburton opened Aug 9, 2026·

  ·
  1

  3

• fix(cpu/matmul): widen GEMV weight row offsets to long (#429)

  #431·jamesburton opened Aug 7, 2026·

  ·
  1

  5

• perf: hoist F16 weight row conversion out of the token loop

  #430·unsafePtr opened Aug 4, 2026·

  ·

• test(unit): report hardware-gated tests as skipped instead of passed (#421)

  #428·jamesburton opened Jul 31, 2026·

  ·
  1

  1

• docs(server): drop the unimplemented lora_adapter request field

  #427·jamesburton opened Jul 31, 2026·

  ·
  1

  1

• server(sse): stop sending Connection: keep-alive on the streaming endpoints

  #426·jamesburton opened Jul 31, 2026·

  ·
  1

  1

• perf(cpu): balanced work partitioning across compute threads (#402)

  #420·jamesburton opened Jul 31, 2026·

  ·
  1

  4

• perf(tokenizers): reuse one BPE merge queue per Encode call (#413)

  #419·jamesburton opened Jul 31, 2026·

  ·
  1

  4

• feat(eval): shared perplexity harness with llama.cpp-comparable sliding-window mode

  #418·jamesburton opened Jul 30, 2026·

  ·
  8

• fix(tokenizers): pre-tokenize raw text before byte-encoding; add StarCoder/SmolLM pre-type family

  #417·jamesburton opened Jul 30, 2026·

  ·
  4

• fix(gguf): parse v2 counts/lengths as uint64 (spec-correct)

  #365·jamesburton opened Jun 30, 2026·

  ·
  1

  3

• feat(cuda): optional arch-tiered PTX dispatch with compute_61 fallback

  #333·jamesburton opened Jun 19, 2026·

  ·

• feat: Microsoft.Extensions.AI IChatClient adapter (DotLLM.Extensions.AI) (#327)

  #328·jamesburton opened Jun 15, 2026·

  ·
  1

  9

• server: Anthropic-compatible Messages API (/v1/messages) (#325)

  #326·jamesburton opened Jun 15, 2026·

  ·
  1

  8

• Feature - Metal GPU backend for macOS (Apple Silicon)

  #324·rflechner opened Jun 14, 2026·

  ·

• perf(cpu/matmul): Q8_0 outer-product GEMM via AVX2 VNNI (VPDPBUSD-256) — unblock #61 register pressure

  #323·jamesburton opened Jun 9, 2026·

  ·
  1

  1

• models(architecture): add Gemma 4 family — Architecture enum + HF detection + forward wiring (Phase 8 Step 57)

  #320·jamesburton opened Jun 8, 2026·5 tasks done·

  ·
  1

• models(transformer): wire Gemma 3 per-layer SW + attn/final softcaps + QPAS

  #318·jamesburton opened Jun 8, 2026·4 tasks done·

  ·
  1

  1

• models(safetensors): SmolLM3 + Gemma 3 architecture detection (rebased onto #314)

  #317·jamesburton opened Jun 8, 2026·4 tasks done·

  ·

• diagnostics(sae): SparseAutoencoder hook + top-K feature extraction

  #316·jamesburton opened Jun 8, 2026·

  ·
  1

• kernels(cpu)(matmul): F32 outer-product tiled GEMM kernel for prefill

  #315·jamesburton opened Jun 8, 2026·

  ·
  1

  6

• loader(safetensors): port MLA / Mixtral / Qwen-MoE loaders onto ISafetensorsTensorSource

  #314·jamesburton opened Jun 8, 2026·

  ·

• models(transformer-perf): direct-to-cache K/V projection via TryReserveSlot

  #310·jamesburton opened Jun 8, 2026·

  ·

• engine(kv-cache): IKvCache.TryReserveSlot — write-into-cache primitive

  #309·jamesburton opened Jun 8, 2026·

  ·
  1

  7

• tokenizers: TryDecode(Span<char>) zero-allocation overload + IncrementalDetokenizer adoption

  #308·jamesburton opened Jun 8, 2026·

  ·
  8