Skip to content

SeisoLocalAI v0.2.0

Choose a tag to compare

View all tags
@Legendarylibr Legendarylibr released this 15 Jul 12:58
· 399 commits to main since this release
v0.2.0
8558bfc
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

SeisoLocalAI v0.2.0

Second release since v0.1.0 (2026-06-19). ~700 commits and 90+ merged PRs focused on making native Linux NVIDIA chat stable, long replies finish reliably, training/SLIME work, and the product story clearer.

Highlights

Long-form chat (multi-pass, OOM-safe)

  • Dynamic multi-pass auto-continue so long answers (songs, papers, essays) finish without one huge completion that can OOM (#340 and earlier long-reply work: #323–#326).
  • Per-pass completion stays OOM-clamped (often ~512–768 tokens on native Linux NVIDIA); total reply budget scales with request, long-form intent, and free VRAM.
  • Linear-decay context packing for continue passes: free headroom grows as the draft grows; recent assistant tail is preferred over a full-context pack.
  • Continue on mid-sentence EOS (not only finish_reason=length) — fixes live unfinished cuts like after “again” / “longer”.
  • Empty continue pass: one strong-cue retry.
  • Smoother Forge UI streaming for multi-pass: adaptive reveal, continue-gap cue, live caret (#340).

Native Linux NVIDIA inference

  • Extensive OOM and hang guards for llama/Ollama sidecar paths (context clamp, batch caps, VRAM budgeting, fail-closed GGUF isolation).
  • Ollama-first / sidecar routing for GGUF chat; improved VRAM planning and tok/s without relaxing safety guards.
  • Native Linux KV pipeline and TTFT improvements; pin chat context after preload so multi-turn reuses KV size.
  • Packing and load hot-path performance for large GGUF / train startup.

System monitoring (native Linux)

  • Correct CPU brand from /proc/cpuinfo (no more arch-only x86_64 label).
  • AMD k10temp / broader thermal sensors for live CPU temperature (#339).

Training / SLIME

  • Distributed SLIME training wired through Forge UI (#335).
  • SLIME sampling aligned with upstream design; optional per-token loss (#336, #337).
  • Train tab no longer runs dataset prep on open (#327).
  • Train/inference hot-path optimizations.

Security & reliability

  • RAG trust-boundary and audit correlation hardening (#332).
  • Inference generation ownership, preload locks, job runtime / kernel boundaries.
  • Local sign-in reset for forgotten passwords (#291).
  • CI: setuptools ≥83 for pip-audit, assorted lint cleanups.

Docs & product framing

  • Clearer project description: local-first platform for chat, fine-tune, quantize, compress, and export (README + package metadata, #341 / #334).
  • Install locks, platform docs, and citations updates.

Upgrade notes

  • Prefer a fresh git pull (or reinstall via start) on native Linux + NVIDIA.
  • Long replies use multi-pass continues; you may see “Reply incomplete — continuing (N/M)…” during generation — that is expected, not a failure.
  • Optional env (defaults are usually fine):
    • SEISO_CHAT_AUTO_CONTINUE_MAX — max extra passes (-1 = auto)
    • SEISO_CHAT_AUTO_CONTINUE_TOTAL_TOKENS — cumulative output ceiling (default 32768)

Full change range

Assets 2
Loading