Issues

• llama.cpp refresh: v0.7.0

  Nihmar/omphalos#538 · Filter by author github-actions opened on Oct 19, 2026
  1

• perf: prefill attention in fp8/iu8 at long context (57 % of a 100k prefill at ~15 TFLOPS)

  Nihmar/omphalos#480 · Filter by author Nihmar opened on Oct 14, 2026
  1

• feat: KV streaming — quantized KV in pinned host RAM, paged through a bounded device working set (262k in 16 GB, tracking)

  enhancement
  Nihmar/omphalos#409 · Filter by author Nihmar opened on Oct 13, 2026
  7

• Re-measure the llama.cpp baselines at the first pinned commit (v0.6.0 changes for Qwen3.8-27B on gfx1200)

  Nihmar/omphalos#396 · Filter by author Nihmar opened on Oct 8, 2026
  4

• test: does the KV precision change repetition on 64k+ agentic contexts? (K8 vs the mixed K4/K8 vs K4 everywhere)

  Nihmar/omphalos#297 · Filter by author Nihmar opened on Oct 4, 2026
  4

• perf: fewer launches per step: sibling GEMV groups for 1..16 tokens (#203 D1)

  Nihmar/omphalos#255 · Filter by author Nihmar opened on Oct 3, 2026
  2

• Time to first token through the servers: omphalos vs llama.cpp, K4/V4 and K8/V4

  Nihmar/omphalos#233 · Filter by author Nihmar opened on Oct 3, 2026
  1

• Prefill speed curve: omphalos vs llama.cpp, K4/V4 and K8/V4, with and without MTP

  Nihmar/omphalos#232 · Filter by author Nihmar opened on Oct 3, 2026
  2

• perf: prefill GEMM against the practical WMMA ceiling

  Nihmar/omphalos#213 · Filter by author Nihmar opened on Oct 2, 2026
  5

• perf: decode and prefill toward the practical ceiling (profile + plan)

  Nihmar/omphalos#203 · Filter by author Nihmar opened on Oct 2, 2026
  7

• perf: prefill attention register pressure (compute-bound after #173)

  Nihmar/omphalos#191 · Filter by author Nihmar opened on Oct 2, 2026
  5