RESUME-RUNBOOK.md

GLM-5.3-BF16 ShapleyMCG — 3.25 bpw resume runbook

State at archive (2026-08-30): all 75 main layers hold sealed K3+K4 candidate pairs; MTP78 has a sealed uniform-K4 inventory (plus a decode-verified K5 side store); the uniform-K3 draft is published with reconciled KLD evidence. The only unfinished chain is: five-node Aumann-Shapley attribution -> exact global 3.25-bpw allocation -> mixed materialization -> KLD replay -> publication.

0. Inputs (all public/HF)

• This dataset: tars/candidates-layer-*.tar (75), tars/capture-*.tar (4 roles), tars/mtp78.tar, tars/panel.tar, tars/evidence.tar, tars/attribution-final.tar, candidates/main-inventory.json, code/glm53wt.bundle, code/campaign-scripts.tar
• Teacher logits: brandonmusic/GLM-5.3-BF16-full-logits -> bf16-nonfinal/ (fit 32 / conditional_fit 16 / selection 16 / confirmation 16, fp32 full vocab)
• Weights: zai-org/GLM-5.3-BF16 at revision 304b8051cfb2b260b61ce0cbe330e02a98e73639
• Transformers checkout at 805a9e939fa8c1bff8d8ffdf041c051b71a914aa (PYTHONPATH-bound)
• exllamav3 upstream @c5d9c65, setup.py build_ext --inplace ONLY (pip -e tries flash_attn)

1. Restore layout (any 8-GPU node with >=268 GiB/GPU; B300-class for TP8 BF16)

/root/glm53wt            <- git clone from code/glm53wt.bundle (ALL node-local
                            commits: no-intra-LDLQ env gate, cold-expert
                            fallbacks, attribution data_ptr/__func__ fixes)
/root/campaign           <- untar code/campaign-scripts.tar (runners incl. all
                            13-attempt fixes already applied)
/root/glm53-shapleymcg   <- untar: panel, evidence, attribution-final,
                            candidates (75 layer tars + main-inventory.json),
                            capture/mtp78 not needed; mtp78.tar -> mtp78/
/root/logits/bf16-nonfinal <- hf download the 80 windows
/root/models/GLM-5.3-BF16  <- hf download at the pinned revision

Then reseal nothing — the shipped evidence/ closures match the bundle tree. Env: GLM53_NO_INTRA_LDLQ=1, venv torch 2.12.1+cu132.

2. Attribution (the one unfinished computation)

Known-good up to the OOM: bash /root/campaign/run_attribution.sh (fires glm53_attribution_runner_draft.py --mode attribute, TP8, path-nodes 5, fisher-rank 32, probes 2/window; endpoint receipt already in attribution-final.tar — or regenerate with --mode seal-endpoint, ~25 min CPU).

Blocker to fix first (documented in ATTRIBUTION-STATUS.md): OOM at 264/267.7 GiB in the path-node phase. Fix in campaign/glm53_attribution_runner_draft.py near the node loop (search retain_graph=True): a. preferred: drop retain_graph=True; recompute the blended forward for the Fisher-probe backward passes (second forward per node; ~2x node compute, ~halves peak graph memory), or b. wrap the blended forward in torch.utils.checkpoint.checkpoint (use_reentrant=False) per decoder layer, or c. stream the K4 residency per layer instead of fully resident (GLM53TP8PackedK4Residency). All alpha-0/alpha-1 exactness gates stay armed; the f32 softmax path is already in the shipped runner (alpha-0 KL is exactly zero by construction).

3. Allocation -> materialization -> KLD -> publish (all staged, tested paths)

PYTHONPATH=src python scripts/allocate_glm53_mcg_exact_3p25.py \
  --scope main --candidate-inventory .../candidates/main-inventory.json \
  --attribution <attribution output> \
  --output .../allocation/main-3p25.json --execute

# materialize: /root/campaign/post_attribution.sh already does everything below
#  - 3-layer-wide pool, per-rank materialize_glm53_selected_packed --allocation
#  - MTP78: --uniform-bits 4 with mtp78/candidate-inventory.json
#  - KLD replay fires in parallel from the CANDIDATE store (no materialization
#    dependency): kld_student_capture.py --allocation ... then kld_compare.py
#    + kld_axes.py on the 16 confirmation windows (teacher = bf16-nonfinal)
# pack for HF with pack_layers.py (HF hard limit: 20,000 files/repo; one
# consolidated safetensors per layer; xet dedups re-uploads)

Expected: uniform-K3 measured 0.0979 (new panel) / 0.0375 (published windows); the allocated 3.25 should land near ~0.03 on the new panel given the davidsyoung 3.42-mixed reference at 0.0240 on the mild published windows.

4. Operational traps (all hit once already)

• Per-process GPU isolation for 8-wide encode/decode workers: CUDA_VISIBLE_DEVICES=$R ... --device cuda:0 (8 procs with --device cuda:$R = illegal-access collisions on GPU0).
• Kill stale rank processes by exact PID from nvidia-smi query-compute-apps before any TP8 relaunch (orphans hold ~190 GiB/GPU).
• Write-once artifacts: rm stale attribution-run-plan.json (keep uniform-k4-byte-verification.json) before every attribution relaunch; the endpoint receipt binds the code-closure sha — reseal endpoint after any sealed-repo edit.
• Every sealed-repo edit => reseal 3 code closures + capture-verification (fix chain exists: campaign/fix_reseal_attr.sh).
• lm_head is UNTIED (tie_word_embeddings: false); MTP hnorm consumes the RAW pre-model.norm hidden (vLLM Glm4MoeMultiTokenPredictorLayer semantics).