command-center/agent/managed-skills/radiance-model-feasibility-scout/SKILL.md

name, description

name	description
radiance-model-feasibility-scout	Scout whether a new GGUF model can run on radiance (R9700 fleet): GGUF header census, REUSE/ADAPT/NEW gap matrix vs radiance docs, ledgers (traffic/KV/RAM/disk), KLD-gate substitution, HTML report delivery.

Radiance Model Feasibility Scout

Answer "can model X serve on radiance (dankserver R9700 TP=2), and what's the bill?" Produce an HTML report. Complements gguf-header-vram-fit-analysis (fit math), radiance-w4-convert-playbook (conversion mechanics), radiance-quantize-and-score.

1. GGUF header census (trust header, not the card)

• Range-parse the header: tensor count, per-type census (type → tensor count / params / bytes), layer pattern from tensor names, general.architecture, llm_type, rope dim, expert counts, mtp declaration.
• MTP: presence is decided by the tensor-name census (nextn.*/mtp tensors), never by metadata.
• Confirm tokenizer + chat_template KV entries ship in the file; mmproj = separate BF16 tower file.
• Cross-check total bytes vs the published file size.

2. Gap matrix vs radiance docs

Read PLUGIN.md (arch-plugin contract, quantizer registry, core/mm image-pipeline law), OPS.md (op catalog: attention tiles h256 gqa2/4/8 + h512 gqa8, GDN, QSA, MoE, MTP), QSA.md (indexer decomposition bounds), GUIDE.md + gguf.cpp (dtype table 0–44: all standard GGUF grids readable — quantization happens in rad-convert, not the loader). Classify every component REUSE / ADAPT / NEW with the specific op named. The live qwen4exp plugin (~2k lines, QSA+MoE+mHC+GDN) is the template and the shared-header reuse list.

3. Laws checklist (recurring traps)

• GGUF ssm_a = −exp(A_log); undo at convert/plugin time.
• One-table-one-encoding: a routed-expert table takes ONE rad encoding; mixed GGUF grids (q2/q3/q4 per layer) re-quantize to w4 GPTQ + whole-expert BF16 protected set.
• int8 trunk + int8 lm_head laws; swiglu clamp carries from the arch family; rope-free layers need no RoPE path.
• KDA ≠ GDN: per-channel decay + 3 convs + L2-normed q/k + lower-bound gate → new kernel family (FLA chunk_kda is the oracle); kkt_solve/chunk_scan is the critical path.
• MLA: latent-cache absorbed form (KV = latent width × dtype), needs its own attention tile (e.g. h512 gqa-64) + per-head v bmm.
• llama.cpp merged PR for the arch = the ground-truth formula oracle (sign conventions, APE layout, Sinkhorn input).

4. Ledgers (all in the report)

• Per-token weight traffic GB (trunk int8 + active experts w4 + lm_head) → decode anchored on frank-archon's measured 99 t/s @ ~6.3 GB/token; state [INFERENCE].
• KV: per-token bytes per pool + per-seq state terms; cross-check vs any prior fleet runbook that served the same model.
• Pinned RAM: host pool + prefix vs LimitMEMLOCK/ttm-preflight.sh law.
• Disk peak = GGUF + .rad (~187 GiB-class for a 300B MoE, no PLE) vs free space; options: retire rollback payload / external convert host / --in-place.

5. KLD gate feasibility

BF16 teacher = GGUF params × 2 B; if unservable (626 GB-class), the substitute gate = llama.cpp (arch PR merged in master) serving the GGUF on the Dankclaw iGPU — GTT aperture ~114 GiB — stop+mask frank-dragoon for the window, KL + top-1 census vs radiance .rad. Document as a gate deviation.

6. Deliver as self-contained HTML

• Single file, inline CSS with prefers-color-scheme vars, section anchors + TOC, census/gap/risk/plan tables, [INFERENCE] labels on arithmetic.
• Write in chunks: one write ≈ ≤8 KB (output-token limit), end each chunk with <!--MORE-->, then edit with PUT N.=N: where N = marker line — prefix every body row with + (bare rows get auto-prefixed; register lines like @name land as literal content).
• Validate: python html.parser tag-balance script + grep for stray markers; xdg-open the file.