[michaellchung]

Summary

Adds end-to-end Kimi K3.1 async GRPO training to yeto:

• Miles adapters for the K3.1 lineage: ports island, sampler island, buffer reader with short-group carry-over and loss-masked padding, padding-aware metrics and metadata checks, release plugin that restores the base checkpoint metadata, ledger set-aside when a trainer restarts from the base.
• Engine samplers (SGLang and vLLM servers): R3 routed experts, top-K log-probs, fp32 lm-head, in-place MXFP4 expert reload served with the base tokenizer, and a bound on how long an agentic round may defer a release (reload_max_defer_s, default 600 s).
• National Compute sampler pool: elastic join/leave, automatic replacement on eviction and before lease expiry, per-node cost accounting, cost guard and resource ledger, yeto cost.
• yeto run up: one-config launch with preflight, budget / wall-clock / silence backstops; task dispenser; run health breakers; remote sandbox hosts for agentic tasks with orphan sweeps.
• Trainer robustness: non-finite reward groups, unconvertible groups, sampler-wait timeout measured on progress, early config refusals, randomized fuzz of the buffer read and train chain.
• Dashboard views for async GRPO runs; docker/miles-k31 image build; examples/k31 specs and launch scripts; sampler node scripts and scripts/prune_layers.py.
• Docs: customer guide (docs/KIMI_K31_ASYNC_RL_GUIDE.md), 262k full-parameter layout, NC pool, cost guard, dispenser, breaker, sandbox hosts, Miles RL and CLI reference.
• Fix: codepack keeps the index mtime on its temporary index copy (a same-size edit within the second after a commit was not seen as dirty).

Validation

• Kimi K3.1, 64-expert full-parameter, 2x8 B300 trainer + NC MI355X vLLM sampler: 4 healthy optimizer steps (grad_norm 0.012-0.033, KL <= 0.07), a release published every step (LATEST committed, config identical to the base, MXFP4 expert blocks changed), in-place hot reload of v1 on the sampler with a matching policy hash.
• NC elasticity: eviction fenced and replaced automatically; the other island was unaffected; no duplicate dispatch.
• Qwen3-0.6B single-GPU rehearsal: 10 steps without error, 6 sampler hot reloads, ledger set-aside on restart, short-group padding and carry-over, stale groups.
• Qwen3-0.6B trainer + 2 NC MI355X agentic sampler islands: 6 forced hot reloads under always-in-flight rounds (sampler followed the trainer to v10, KL 0.0013-0.0027); hard eviction fenced the island and released its open dispense claims within about a minute; the replacement started on LATEST and resumed the predecessor's position; orphan trial containers swept on same-shard restart and, with the new predecessor sweep, after a cross-shard replacement.
• Fixes from that run: a replacement sampler sweeps its fenced predecessor's trial containers; a row dispensed twice in one round gets its own prompt id; a recorder session collision loses one trial instead of the island.
• Tests: full local suite 10161 passed, 0 failed (189 skipped, 7 xfailed); after the reload-bound commits, the sampler / release / agentic / island / policy subset 2395 passed, 0 failed; after the NC fixes, the sampler / agentic / sandbox / recorder / dispense subset 2667 passed, 0 failed.

🤖 Generated with Claude Code