Frank
a34322aeb0
The managed-skill library was regenerated: 94 narrow per-incident skills
replaced by 25 broader ones (radiance-*, llamacpp-*, homepage-*, fleet/
endpoint ops). Git renders this as mass delete + mass add because whole
directories changed name; old content stays in history at 12c5e88.
Surface check: the 25 on disk are exactly the skills the harness advertises
this session.
2.8 KiB
name, description
| name | description |
|---|---|
| radiance-quantize-and-score | Design a rad-convert recipe for a radiance (StillDeadcode) MoE checkpoint, calibrate it, and score it with the KL mode — the procedure plus the servable-format law. |
Radiance: quantize, calibrate, score
For a HuggingFace checkpoint → .rad on the radiance engine (Codeberg StillDeadcode/radiance).
1. Know the servable formats before writing a rule
The block/plugin serves a fixed list; the converter refuses an encoding no kernel reads (docs/TOOLS.md:184). For qwen4exp MoE experts: fp8a8, w4a8, w4a8h, bf16, w4nla8h, w4nl64a8h, w4nl32a8h, w5nl64a8h (arch/common/rad_block_moe_fp8.h kServed); dense: fp8a8, i8a8, fp8a16, mxfp4a8, bf16, plus the u2 draft head. fp6_*/u6 encode but never serve. Read kServed in source — never trust a code-name list from docs prose.
2. Recipe file
Header # lines are stored in the container; first matching rule wins, so put cast dtype=bf16 protected-expert rules and projection-specific rules ahead of the general blk.*.ffn_*_exps.*.weight rule. Unmatched weights are served as the checkpoint holds them (bf16). Verify names against radiance's internal stems (ffn_gate_up_exps, ffn_down_exps, attn_qg, ssm_inz, ple_ngram, output.weight), not the HF names.
3. Calibrate (GPTQ/awq/autoround/paroquant need $CALIB)
Bootstrap: rad-convert with plain rtn → serve with RADIANCE_CALIB_DIR (+RADIANCE_CALIB_TOKENS) and drive prefill → re-convert with --recipe … --reuse <bootstrap.rad> (--reuse copies every weight whose encoding did not change). A BF16 checkpoint may be calibrated as it ships: convert with no recipe + --max-ctx 4096, serve --placement expert_tiered --weights-disk-tier, MNBT large (a step costs the same at 1 token as at 32K). Corpus must look like production traffic; score on a held-out slice.
4. Score
radiance --model BF16.rad … --kld-record DIR --kld-corpus C.jsonl (exact wire, bf16 KV) once, then --kld-ref DIR --kld-out R.json per candidate. Report = mean/median/p99/p99.9/p99.99 KL, top-1 agreement, both perplexities, per source — label the corpus's source field to get per-turn-type scopes. KLD corpus disjoint from calibration. Keep --max-num-batched-tokens ≤ 8192 (the logits buffer is ~2 GiB/rank per 4096 tokens).
5. Set the gate from the published ladder, not from a wish
Compare against docs/MOE-W4.md's table for the arch (served tier / heavier tier / bf16 floor). A threshold below the bf16 floor rejects every quantized build by construction. Escalation ladder reuses the same Grams (Hessians are grid-independent) → escalate by re-converting only.
6. Verify before serving
rad-tune at the served shape (TP, spec depth) with a persistent cache mount; read the boot ledger for residency; run the fleet probes.