All pull requests

Pull requests

• feat(ascend): add experimental single-card dense GGUF backend

  #448·Cheongroach opened Oct 28, 2026·

  ·
  1

• perf(server): avoid empty stop lists and reserve appended-stop capacity

  #446·Cheongroach opened Oct 24, 2026·

  ·
  1

• fix(server): preserve default n and reject unsupported choice counts

  #445·Cheongroach opened Oct 24, 2026·

  ·
  1

• fix(engine): validate SimpleKvCache geometry before allocating

  #444·Cheongroach opened Oct 24, 2026·

  ·
  1

• fix(cpu): honor caller-only decode thread configuration

  #440·Cheongroach opened Oct 24, 2026·

  ·
  1

• fix(server): derive consistent model IDs from local GGUF paths

  #438·Cheongroach opened Oct 18, 2026·

  ·
  1

• docs(kv-cache): correct memory sizing, CPU layout, and decode costs

  #436·Cheongroach opened Oct 18, 2026·

  ·
  1

  1

• 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

• 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·

  ·
  2

• 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·

  ·