Skip to content

SeisoLocalAI v0.4.0

Choose a tag to compare

View all tags
@Legendarylibr Legendarylibr released this 24 Jul 01:41
v0.4.0
1e05293
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

SeisoLocalAI v0.4.0

Since v0.3.0 (2026-07-21). 7 PRs focused on NeMo RL integration, slime multi-GPU / GRPO correctness, training-config floors, and Forge red-team hardening.

Highlights

NVIDIA NeMo RL

  • New external training method method: nemo_rl (no vendoring): launches NVIDIA-NeMo/RL via uv run with recipes grpo, dpo, distillation, smoke (#387).
  • CLI (seiso nemo-rl), Forge Train Studio picker, example/smoke configs, and SEISO_NEMO_RL_ROOT setup docs (#387).

Slime / GRPO / multi-GPU correctness

  • DDP no_sync during grad accumulation with flush/average before mid-window optimizer steps; round-robin SGLang/vLLM generate across engine URLs (#390).
  • Drop vacuous GRPO groups (<2 non-truncated rollouts); refuse biased vllm_weight_mode=full + LoRA; require vllm_base_url for cloud GPU + slime + vLLM (#390).
  • Example microbatches kept as multiples of rollouts_per_prompt so GRPO groups stay intact (#388).
  • HTTP rollouts honor finish_reason and server token ids; managed vLLM liveness/model probing hardened (#391).

Training-config invariants

  • Validate-time product slime floors (data_gen_count, held-out auto-split), refuse empty/off data_gen, require local JSONL (or SEISO_ALLOW_TEMPLATE_SLIME), ban CI fixtures on NeMo product configs (#389).
  • Held-out eval uses greedy HF generate; NeMo GRPO num_generations_per_prompt always overridden (default 4) (#389).

Forge security (red-team)

  • Non-overridable security system prompt with tools; quarantine instruction-like KB chunks; Compat inference API keys are chat-only (tools need session JWT) (#392).
  • Refuse remote + code-exec entirely (no ACK override); redacted chat/tool audit hashes (#392).

Dependencies & audit sweep

  • pyasn1 ≥ 0.6.4 for REAL/OID DoS advisories; lock refresh (#386).
  • Broader Critical/High audit fixes across compress/export/distill-RL/DPO/kernels/MoE DDP (#391).

Upgrade notes

  • NeMo RL requires a recursive NeMo RL checkout and SEISO_NEMO_RL_ROOT (plus uv); not bundled in this repo.
  • Product slime configs without SEISO_ALLOW_TINY_RL enforce real-data floors; template/code examples may need operator JSONL or SEISO_ALLOW_TEMPLATE_SLIME.
  • Remote Forge + SEISO_ALLOW_CODE_EXEC together refuse startup; Compat /v1 tools require a session JWT (inference API keys stay chat-only).
  • Multi-engine slime HTTP rollouts round-robin generate; weight sync still fans out to all engines.

Full changelog: v0.3.0...v0.4.0

Assets 2
Loading