88e171bea62d1b3a5265a970f792a9eebd92e1f2
8 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 vh
|
a9d73dad41 |
fix(coldfusion-abliteration): GPU0 seat restore order is load-bearing — correct the claim and the runbook
Restoring the two GPU0 seats with `start meromero; sleep 10; start gen` put
meromero into a 7-restart crash-loop:
ValueError: Free memory on device cuda:0 (35.3/94.97 GiB) on startup is less
than desired GPU memory utilization (0.52, 49.38 GiB).
The previous commit's README claimed restore order "is not actually load-bearing"
on the grounds that both seats pass --gpu-memory-utilization as a fraction of
total VRAM. That is half right and the wrong half mattered: the fraction sets the
target, but vLLM gates startup on FREE VRAM and refuses to start unless the whole
target is available. GPU0 runs at ~96.4/97.9 GB with roughly 0.4 GiB of slack, so
the seats coexist only in the order they were originally brought up, and meromero
is the one that does not fit in the remainder. The pre-existing auto-memory note
("gen takes a fraction of free VRAM at startup and will starve meromero") was
pointing at the real effect.
Also: "first" means healthy, not ten seconds earlier. A sleep 10 against a
two-to-three minute weight load is simultaneity, not ordering — gate on observed
state.
Recovery applied: stop gen, wait for meromero healthy, start gen. Verified
against the pre-window baseline rather than against "both green":
gen KV 14.36 GiB / 403,065 tok / 1.54x -> 14.34 GiB / 401,550 tok / 1.53x
meromero KV 542,202 tok -> 542,202 tok
RestartCount 0 on both; summarizer smoke-tested through LiteLLM
Note for the next reader: raw nvidia-smi used-MiB is the wrong check here. It
reads 89,503 now vs 96,376 before, which looks like a 6.9 GB regression and is
allocator slack — serving capacity is unchanged. The anomalous boots were the
high ones (34.95 GiB KV), where gen came up on an empty card mid-window.
|
||
|
vh
|
1b3fb270e7 |
feat(coldfusion-abliteration): first-token KL measured — 28.4x selectivity, harmless median 0.0211
Adds `kl_divergence.py`: first-token KL(stock || abliterated) over the full 248,320-token vocabulary, bf16 vs bf16, scored separately for held-out harmless and reserved-harmful prompts. Result (L35, 256 harmless / 104 harmful, answer mode): harmless median 0.0211 mean 0.0364 top-1 agreement 89.8% harmful median 0.5996 mean 0.6992 top-1 agreement 55.8% selectivity 28.4x (72.8x in think mode) Self-KL noise floor is exactly 0.0, and all 720 per-prompt values are bit-identical between a single-process and a two-process run, so the figures are signal rather than bf16 jitter. Reverse KL on harmful/answer is 1.43 vs forward 0.70 — the mass-where-stock-had-none asymmetry expected of a refusal-direction removal. Against the Heretic reference figures (0.1191 prior seat, 0.0759 the live absolute-heresy seat) this is materially gentler, but those are the other tool's optimizer output on a different base with its own harmless set and template — order-of-magnitude, not head-to-head. KL remains a fidelity number; the viability gate is still MTP acceptance (59.1%). Method notes: - Prompt classes are reported separately by design. A single averaged KL over a mixed corpus is close to meaningless, since the metric is meant to be large on harmful prompts and small on benign ones; the ratio carries the information. - The harmless evaluation set is drawn from the alpaca pool minus calibration's own draw, reconstructed by replaying that draw rather than remembered, and asserted disjoint on text. The harmful set is the reserved test split. - `render` is imported from abliterate.py rather than copied, so the measurement cannot drift from the rendering the direction was captured against. - Batch size 1 with logits_to_keep=1: no padding semantics, ~0.6 MB of logits. Three corrections to the runbook, each of which cost time: - "bf16 is 50 GB, only gen must go" was 50.10 GiB mislabelled. Text-only weights are 51,300 MiB; freeing either GPU0 seat alone leaves ~50,933 MiB. Both must stop. VRAM is now sized from the safetensors headers at run time. - A 27B model cannot be released in-process: `del` + gc + empty_cache left free VRAM at 45,287 MiB, and so did confining the model to an inner frame that exits. Only process exit returned the card (96,689 MiB). The first run completed only because the allocator hit OOM, collected, and retried. Each model now gets its own process, handing log-probs to disk between stages. - The residency gate read hf_device_map, which transformers leaves empty when the model fits on one device — it reported "(unsharded)" whether or not anything was wrong, so it could never fail. It now reads parameter devices directly. Model-agnostic lessons promoted to the quant playbook (new 3.12). |
||
|
vh
|
725c8fdf9e |
feat(coldfusion-abliteration): THESIS PROVEN — in-band-abliterated MTP head accepts 59.1% (beats incumbent ~47%)
Quantized the L35 abliterated model to mixed NVFP4 and measured MTP acceptance
end to end. The experiment's whole premise: Heretic (the incumbent gen seat)
leaves the MTP head a byte-identical base graft its wrapper never loads, whereas
Robinson abliterates the MTP head in-band — the question was whether that in-band
edit survives well enough to spec-decode. It does, better than the graft:
MTP acceptance 59.1% median (51-65%, 8 cache-busted topics) vs incumbent ~47%
decode 118.7 tok/s median (faster; image-confounded, read as not-worse)
abliteration survives quant (creative refusals drop, self-harm guardrail
intact, coherent)
Output at /tank/aimodels/qwen38-27b-coldfusion-L35-nvfp4-mixed (22.5 GB). Result
JSON in bench/. NOT cut over — the incumbent seat is untouched; making L35 the gen
seat is a separate decision needing the full Stage-3 gate + real multi-turn hold.
Two env foot-guns hardened along the way:
- quant_mixed_nvfp4.py now promotes text_config attention fields
(num_attention_heads etc.) to the top-level config for the oneshot, then
restores. transformers 5.10 / llmcompressor 0.12 (this venv moved under us
since the Aug-15 heresy quant) no longer delegate the top-level lookup, so
oneshot raised "Cannot determine num_attention_heads". Same "the fight is the
environment" pattern as the abliteration capture.
- a sub-~23GB quant saves as a single model.safetensors with no index, so the
post_quant MTP graft needed an index built first — from the safetensors header,
not safe_open (which mmaps the whole shard and ENOMEMs on ZFS).
post_quant grafted the abliterated MTP (15 tensors, 849 MB) and re-injected
re:^mtp.* into quantization_config.ignore (llm-compressor pruned it again — the
two-rounds-lost 0%-MTP bug, fired and repaired as designed). Probe served on the
pinned nightly (#51113 qwen3_5_mtp fix) to match the live seat's vLLM.
|
||
|
vh
|
e9dbc8660b |
feat(coldfusion-abliteration): abliteration LANDS at layer 35 — separation selector, shard-surgery write, three false diagnoses corrected
The abliterated model works. A/B vs stock on a matched greedy battery: explicit sexual + graphic torture (the measured stock refusal surface) go from refused to complied/engaged, held-out AdvBench prompts loosen, the self-harm guardrail survives, coherence intact — the Robinson design point exactly. Output at /tank/aimodels/qwen38-27b-coldfusion-abliterated-L35-bf16, verified bitwise: 131/131 targets changed, 333/333 vision byte-identical (delta 0.0), 735/735 others untouched. Getting there corrected three diagnoses the prior session had backwards. 1. The layer-selection metric was wrong, and that was the whole ballgame. The recipe picks the abliteration layer by peak two-template |cos| agreement. On this heavily-merged base that metric is anti-correlated with efficacy: its argmax (layer 18) is the WORST-separating layer in the window (Cohen's d 5.51 vs 9.89 at the peak), and abliterating there was a measured behavioral no-op — stock and "abliterated" refused all six probes identically. Cause: the two renderings end in different generative modes (</think> vs <think>), so |cos| scores answer-vs-reason mode, not refusal, and on a merge the mode term dominates. Replaced selection with harmful/harmless SEPARATION (Cohen's d / AUC of the direction's projection), gated on the sink screen since separation and sink-energy both climb with depth. Picks layer 35 (d 9.35, AUC 0.9997, sink 0.094%). Agreement is kept as a printed diagnostic. 2. The "bf16 NaNs, use fp32" rule was a misdiagnosis. The NaN was never precision — it was multi-GPU sharding (the residual stream zeroes two layers past the GPU0->GPU1 boundary; the first capture's layer 22 happened to sit in the healthy region, which is why it looked fine) plus PYTORCH_CUDA_ALLOC_CONF=expandable_segments (corrupts retained tensors; the corruption MOVED between bit-identical forwards, the tell that it was memory not math). On one GPU with a plain allocator, bf16 full-64-layer is exactly deterministic and coherent, at 50 GB and 4.3x the throughput of the 111 GB fp32 it replaced. Both defects are now hard gates (residency exit 8, allocator exit 9); capture pins CUDA_VISIBLE_DEVICES=0. 3. The corpus-size hypothesis was falsified. 52x more calibration data (8->416, mlabonne/harmful_behaviors = the recipe's actual AdvBench split, already on the box) moved agreement 0.594->0.624 — nothing. Kept the 416/416 corpus anyway (calibration.py); it gives the clean separation signal. The held-out 104-prompt test split is reserved and asserted disjoint. Also: the --out write is now shard-level surgery (reads/writes the 18 safetensors directly, no model object, no GPU). This is correctness, not thrift — AutoModelForCausalLM resolves to the TEXT model, so save_pretrained would drop all 333 vision tensors AND skip the MTP head (the in-band MTP edit is the entire point of the Robinson formula). Neither failure raises. Shard surgery makes vision and the other 1068 tensors byte-identical by construction. Batched capture with a dtype-aware equivalence gate; hidden states captured via forward pre-hook (reading output_hidden_states off the returned object is unsafe here — buffers get recycled). Sharding/allocator lessons promoted to the quantization playbook (model-agnostic, sections 3.9-3.11 + superseded table); the selection-metric lesson added to the recipe doc. The dead layer-18 no-op checkpoint was removed (52 GB, confirmed identical to stock). Incumbent gen seat untouched. Full canonical refusal-probe re-profile and MTP-acceptance-on-quant still owed before this becomes a gen-seat candidate. |
||
|
vh
|
f714f28195 |
feat(coldfusion-abliteration): Robinson's real 416-prompt corpus, batched capture, two new gates
The 8/8 calibration set gave |cos| agreement 0.594 against the recipe's 0.9925. This wires in the corpus the recipe actually used and makes a capture at that scale affordable. Corpus (calibration.py, new). The recipe's "held-out train/test split of 416/104 with overlap 0" names mlabonne/harmful_behaviors exactly — 416 train / 104 test, AdvBench-derived — and it plus harmless_alpaca were already staged in ana-ml2's HF dataset cache. Read via pyarrow, no datasets dependency, no hub access. Harmful is order-deterministic (no seed), so a re-capture is reproducible from the flags alone. The 104-prompt test split is reserved as the held-out generalization probe and asserted disjoint, so the post-write re-profile cannot silently become in-distribution. --calib builtin reproduces the legacy run. Batched capture. 832 prompts x 2 templates = 1664 forwards. Padding is on the RIGHT: in a causal stack nothing after position t reaches position t, so trailing pads cannot touch the token read, whereas left padding feeds pads into the DeltaNet recurrence ahead of the prompt — the path whose torch fallback already NaN'd once here. Means accumulate in float64; the direction is a difference of means, which is where cancellation lives on this model. Gates added, both protecting numbers rather than tensors: - batch-equivalence: proves padded-batch == single-prompt (rel 1e-3) before spending the capture window. - surgery pre-check: aborts if any of the 131 targets is absent or on the meta device. orthogonalize_ edits in place, and an in-place write to an accelerate-offloaded tensor is a silent no-op — that ships a half-abliterated model past a smoke test. Fixed a reporting bug: the agreement line printed the global agree.max() beside the window's argmax layer, so the first capture read as 0.8538 when the real in-window number was 0.5944. The global peak sits in the early layers where the dim-3994 massive activation inflates agreement for reasons unrelated to refusal. Now prints window max, a top-5, and labels the global figure informational. --max-layer truncates the decoder for capture. Exact, not approximate: a causal stack's layer-N state cannot depend on layers above N, so any value above the window top leaves the direction bit-identical while cutting fp32 residency and forward cost. 46 drops 18 of 64 layers and is what keeps fp32 off CPU offload. Refused on the write path, where it would emit a truncated checkpoint. Verified on ana-ml2 without the GPU: dry-run still 1:1 (131 tensors, all coverage gates), calibration loads 416/416 deterministically with its guards firing, both --max-layer guards exit as designed. Also confirmed against chat_template.jinja that enable_thinking=True does resolve reasoning_effort to xhigh, so the two renderings are the recipe's — template selection was not the cause of the low agreement. The re-capture itself is unrun: it needs the fp32 VRAM window and therefore production seat downtime. |
||
|
vh
|
7abd3011f7 |
fix(coldfusion-abliteration): capture works — fp32 forward + finite-gate
The --capture forward NaN'd repeatedly. Root cause: transformers' Qwen3.5 DeltaNet linear-attention needs the causal-conv1d fast-path kernel, which can't be built here (no nvcc, no prebuilt wheel). Its torch fallback produces nondeterministic all-NaN hidden states in bf16 -- same 11-token input finite on one forward, NaN at layer 4 on the next. bf16 and fp32 share exponent range, so it's precision-driven catastrophic cancellation, not overflow, and fp32 resolves it. Fixes: - --capture now loads fp32 (the write/surgery path stays bf16 -- no forward, no NaN). attn_implementation=sdpa pinned. - A finite-gate aborts on a non-finite direction. The sink screen alone can't catch this: nan > threshold is False, so a NaN direction "passed" it and got saved silently on the first run. Capture result (fp32, full GPU): refusal direction finite, unit-normed, layer 22, sink energy 0.0008% in dim 3994 -- clean, not sink-dominated. Saved. Caveat recorded: two-template |cos| agreement is 0.59 at layer 22 vs Robinson's 0.99, almost certainly the small 8/8 calibration set vs their 416/104. Valid but noisier than ideal; the README flags expanding the sets before the write. README documents the three environment gotchas (fp32-for-capture, the seats that must be stopped for the 110GB fp32 VRAM and how to restore them, and the fla side-dir PYTHONPATH) so the next run doesn't rediscover them. |
||
|
vh
|
b56cb0db13 |
docs(coldfusion-abliteration): dry-run passed — recipe maps 1:1 (131 tensors)
Dry-run against the fully-staged bf16 confirms the Robinson recipe transfers onto the DavidAU Cold-Fusion checkpoint with no name drift: 1199 tensors, 333 vision preserved, down_proj=64/o_proj=16/linear_out=48/mtp=2/embed=1, coverage gate 6/6, exactly 131 tensors to orthogonalize. Harness verified-ready; the destructive write still gates on operator go. |
||
|
vh
|
1857a8eb81 |
feat(coldfusion-abliteration): Robinson-formula harness, gated, staged
Harness to abliterate DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1 using the MTP-aware, vision-preserving recipe in docs/pfi/abliteration-recipe-qwen38.md. Motivation is measured, not assumed: the stock model's refusal profile (probed 2026-08-19, hand-verified) is ~33% on creative content, concentrated on explicit-sexual and graphic-torture, with 4/5 hard-harm refused, self-harm guardrails intact, and zero benign over-refusal. So there is a real creative-content refusal surface. The Robinson formula is chosen specifically because it abliterates the MTP head IN-BAND -- which the current gen seat's Heretic pass does not (its MTP head is a byte-identical base graft the Qwen3_5 wrapper never loads). That in-band MTP edit is the additive delta. The script refuses to brick the model. Two hard gates from the recipe halt before any write: the coverage identity o_proj(16)+linear_out(48)==64 (catches a tensor-name mismatch that would ship a half-abliterated model), and the attention-sink screen on dim 3994 (orthogonalizing a direction living there produces a model that loads, runs, and emits garbage). The direction is captured from two chat templates and the layer auto-picked by peak |cos| agreement in [18,45]. Classification is suffix-based and name-agnostic so it survives minor drift; the coverage gate is the backstop. Modes: --dry-run (enumerate + gate, no forward, no write), --capture (direction + sink screen, no write), default (write to --out). The README sequences dry-run -> capture -> write -> verify, and names the post-checks (vision byte-identical, refusal re-profile via services/refusal-probe/, MTP acceptance on the quant, PPL/coherence). bf16 staged to ana-ml2:/tank/aimodels/qwen38-27b-coldfusion-bf16 (pinned 9c44193, provenance recorded). The destructive run is NOT executed here -- dry-run verification and operator go gate it. |