Issues

• epic: Strata vs fork gap analysis — why Strata is 2.1x decode / 3.0x prefill on the 3060 (concept + implementation)

  ddvnguyen/hydra_vortex#811 · Filter by author ddvnguyen opened on Oct 2, 2026

• llama-server --decode-overlap crashes on GSQ-RCO IQ3_S (MUL_MAT_ID backend change at graph rebuild, HTTP 500)

  review-finding
  ddvnguyen/hydra_vortex#810 · Filter by author ddvnguyen opened on Oct 2, 2026

• llama-server MoE telemetry: layer 40 never reports routing ('dead layers: 40') on Ornith-1.5-35B

  review-finding
  ddvnguyen/hydra_vortex#809 · Filter by author ddvnguyen opened on Sep 30, 2026

• EAN: ggml_sqr contiguity abort with HYDRA_EAN_STATS=1 + missing POST /atlas/reset?scope=ean (RULING 1+3)

  review-finding
  ddvnguyen/hydra_vortex#806 · Filter by author ddvnguyen opened on Sep 27, 2026

• MoE cache misses/step unmeasurable: slot.moe_cache_enabled false on serving path despite --moe-expert-cache-size

  review-finding
  ddvnguyen/hydra_vortex#804 · Filter by author ddvnguyen opened on Sep 25, 2026

• CI: wire atlas-web proxy smoke test into Build & Test

  review-finding
  ddvnguyen/hydra_vortex#802 · Filter by author ddvnguyen opened on Sep 24, 2026

• Stale moe-35b-pd references in comments + dead InitializeFallback model synthesis (PR #800 review finding S4)

  review-finding
  ddvnguyen/hydra_vortex#801 · Filter by author ddvnguyen opened on Sep 24, 2026

• Monitoring stack down since 2026-09-14: Loki, Prometheus failed, Langfuse never started

  ddvnguyen/hydra_vortex#799 · Filter by author ddvnguyen opened on Sep 24, 2026

• moe-cache.cu version guards claim CUDA >=12.8 but some calls need CUDA 13; no CUDA 12.x CI build on this lineage

  review-finding
  ddvnguyen/hydra_vortex#798 · Filter by author ddvnguyen opened on Sep 24, 2026

• deploy-heads rewrites sm60 tags to llama-engine; teach the resolver the llama-server lineage

  p2-lowreview-finding
  ddvnguyen/hydra_vortex#797 · Filter by author ddvnguyen opened on Sep 24, 2026

• CI sm60/sm86 images of flash-next lineage hardcode GGML_CUDA_FORCE_CUBLAS=ON — do NOT use for rig measurements

  review-finding
  ddvnguyen/hydra_vortex#796 · Filter by author ddvnguyen opened on Sep 24, 2026

• P100 main test env: Ornith-1.5-35B-A3B on fork 1f3f099cb + atlas telemetry, Bifrost/OMP integration (P0-P4)

  priority
  ddvnguyen/hydra_vortex#795 · Filter by author ddvnguyen opened on Sep 24, 2026

• MoE expert cache h2d/phase counters undercount real PCIe traffic by 3-4 orders of magnitude (both GPUs)

  review-finding
  ddvnguyen/hydra_vortex#794 · Filter by author ddvnguyen opened on Sep 23, 2026

• Startup tripwire: warn/refuse when MoE cache + same-type gate/up would hit #791's grouped-dispatch defect

  review-finding
  ddvnguyen/hydra_vortex#793 · Filter by author ddvnguyen opened on Sep 22, 2026

• CI: ddvnguyen/llama.cpp#120 cuda build fails under -Werror (3 trivial fixes, unrelated to the MoE cache defect)

  review-finding
  ddvnguyen/hydra_vortex#792 · Filter by author ddvnguyen opened on Sep 22, 2026

• IQ3_S grouped-dispatch (MoE cache) defect: 512/512 elements wrong, confirmed live in production serving

  review-finding
  ddvnguyen/hydra_vortex#791 · Filter by author ddvnguyen opened on Sep 22, 2026
  1

• Consolidated status: expert ranking for decode speed — measured results, the CPU-vs-GPU expert-compute contradiction, and what #771 inherits

  llama-forkM-Perftask
  ddvnguyen/hydra_vortex#790 · Filter by author ddvnguyen opened on Sep 20, 2026
  2

• Epic: per-turn / per-prompt telemetry with turn(step) selection + delta — revive Profiling, port upstream Atlas galaxy tab

  ddvnguyen/hydra_vortex#787 · Filter by author ddvnguyen opened on Sep 20, 2026
  1

• Missing telemetry: wire REAP saliency + Edge0 prerouter observers (Brain tabs currently honest not-wired)

  ddvnguyen/hydra_vortex#786 · Filter by author ddvnguyen opened on Sep 20, 2026
  3

• Atlas follow-ups: design-doc §3 readout names wrong, lop_validated gate never set by analyze.py, null_tested provenance unverified

  ddvnguyen/hydra_vortex#784 · Filter by author ddvnguyen opened on Sep 19, 2026

• Follow-up profile: moe-flashnext-solo (D10a peer-free placement, measured once 24.7432 tok/s)

  review-finding
  ddvnguyen/hydra_vortex#782 · Filter by author ddvnguyen opened on Sep 19, 2026

• EngineConfig.cs newline-joins override_tensors — silently breaks multi-entry arrays against the engine's comma wire shape

  review-finding
  ddvnguyen/hydra_vortex#781 · Filter by author ddvnguyen opened on Sep 19, 2026

• atlas-web: assess and port Colibri Profiling view

  ddvnguyen/hydra_vortex#777 · Filter by author ddvnguyen opened on Sep 19, 2026
  2

• atlas-web: hook Colibri chat UI to llama.cpp engine via server-side proxy

  ddvnguyen/hydra_vortex#776 · Filter by author ddvnguyen opened on Sep 19, 2026
  3

• [Stage-1] CPU/GPU arithmetic split carries measured quality cost (+1.57% PPL at h=0.146)

  ddvnguyen/hydra_vortex#775 · Filter by author ddvnguyen opened on Sep 18, 2026