1a36e60d3ad10cd5f0536cd8b8f3fe52b6e61692
12 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 vh
|
1a36e60d3a |
docs(quant-playbook): §3.7's APC-off mitigation was reverted nine days ago and the section never said so
Found while answering a question from the operator, relayed via brokkr-smithy-dev, about whether a recorded Qwen3.8 degeneracy at ~1,700 tokens relates to a length sensitivity just measured on the tuned Gemma-4. The record is §3.7 and the number is ~2,000 -- but reading it to answer that question surfaced that the section is stale. §3.7 presented "disable prefix caching, keep MTP" as THE MITIGATION, resolved 2026-08-17, and stated the gen seat runs that config. It does not and has not since that same day: APC-off passed a synthetic 7-turn probe and the operator still saw severe degeneration in real use, so it was reverted. The multi-day hunt resolved to the AEON W4A4 quant being defective, with MTP / prefix-caching / gateway merely amplifying it (§3.8 records the corrected causal story; §3.7 was never updated to match). Verified against the live container rather than against the compose file alone: vllm-gen runs --enable-prefix-caching with qwen3_5_mtp / num_speculative_tokens 3. stacks/gen-seat/compose.yaml carries the full corrected history inline and is the current authority. §3.7's superseded text is kept and fenced rather than deleted -- it is the history of a mitigation that looked right and was not. Added a dated row to §7 per the standing rule that a wrong playbook claim gets a superseded-claims entry, not just a fix. The lesson inside the lesson is worth more than the correction: §3.7's own standing rule is "gate MTP on a multi-turn coherence probe, not just single-shot acceptance." The APC-off mitigation was gated on exactly that probe, passed it, and still failed in real use -- the multi-turn probe was itself too small to gate on. A passing probe is not sufficient evidence at any size that has not been calibrated against real use. |
||
|
vh
|
6a8582936e |
feat(erp-tune): NVFP4A16 serving pipeline, and the MoE landmine it uncovered
Merge + quantize path for turning the Gemma-4 26B-A4B ERP/RP LoRA into a
servable NVFP4A16 seat, plus a playbook entry for the defect found while
validating it.
The landmine (playbook §3.15): a `targets=["Linear"]` NVFP4 recipe silently
misses every MoE expert on this architecture. Gemma-4 stores each layer's 128
experts as two fused 3-D nn.Parameter tensors, not nn.Linear modules, so the
recipe resolves 205 of 427 modules and ZERO experts — 22.84 B params, 88.5% of
the model, left in BF16 with no warning. This is the same blind spot that
killed QLoRA here via bitsandbytes; the tool changed, the checkpoint layout did
not.
before linearize_moe: 427 Linears, 205 targeted, experts 0
after linearize_moe: 11,947 Linears, 11,725 targeted, experts 11,520
(30 layers x 128 experts x 3 projections)
llmcompressor's linearize_moe unfuses them; no registration needed because
Gemma-4 satisfies FusedExpertsProtocol structurally. Caught by an §4.1 dry run
that asserts the expert count before any GPU spend, which is now the documented
requirement rather than an optional step.
Scheme is NVFP4A16, deviating from the playbook's mixed-W4A4 default on
measured grounds: brokkr-smithy-dev benched the W4A4 quant of this checkpoint
at 12% on contradiction detection with CoT off against gen's 81%, the signature
of 4-bit input activations on a reasoning-dense task, and W4A4 KLD degrades
2-4x past ~10k ctx on sm_120. This is a 16,384-ctx RP seat. Marlin's prefill
cost is accepted.
Two further silent-failure guards, both from prior hard-won lessons:
- the merged model ships the UPSTREAM chat template, not the trainee base's
stale 365-line one, because training rendered through upstream and the
mismatch would present as a tuning failure
- calibration reads the run's own encode cache rather than re-tokenizing, which
sidesteps §3.14 (a fast tokenizer mutated by truncation=True and persisted by
save_pretrained clamps every prompt forever)
Merge-then-quantize rather than LoRA hot-swap, since hot-swap onto NVFP4 was a
silent no-op on vLLM 0.24.0 (#47639). merge_lora.py asserts sampled target
weights actually changed, so an inert adapter cannot ship as a tune.
|
||
|
vh
|
ab3a0ca5bc |
docs(quant-playbook): acceptance is not throughput -- always run the depth control
Measured 2026-08-22 on one target with one instrument: raising MTP num_speculative_tokens from 3 to 7 improved accepted length from 2.753 to 3.041 per forward pass while throughput fell from 114.9 to 74.0 tok/s. Reporting acceptance alone would have recommended a 36% regression. The cause is architectural rather than model-specific. A single-module MTP head has no depth of its own, so vLLM runs it autoregressively and k draft tokens cost k sequential forward passes. Past a shallow depth the drafting cost exceeds what the extra accepted tokens save. Records the comparison rule that follows: match k when comparing two speculative methods, or the measurement is of depth rather than method. A parallel-drafting drafter at k=7 against an autoregressive MTP at k=3 is not a method comparison. In the case that produced this, the depth control showed most of the apparent acceptance advantage was depth, while the throughput advantage was real and came from parallel drafting -- our MTP was better at position 0 and still lost overall. Only the measured, model-agnostic result is recorded here. The DFlash2-specific findings, the hypotheses that remain unproven, and the wrong turns taken along the way live in persistent-memory.d/2026-08-22-dflash2-spec-decode.md with explicit epistemic labels, deliberately kept out of the playbook. |
||
|
vh
|
0755ba7d00 |
fix(quant): stop baking the calibration truncation cap into the shipped tokenizer
load_calib tokenizes with tok(..., truncation=True, max_length=seqlen). For a
fast tokenizer that mutates the Rust backend's truncation state in place, and
the subsequent tok.save_pretrained() persisted it, so every mixed-NVFP4 build
shipped a tokenizer.json carrying
"truncation": {"direction": "Right", "max_length": 2048, ...}
against a source whose value is null. Every prompt was clamped at the
calibration length, permanently.
It hid because older transformers does not enforce the text-vs-ids count
check. On a newer one the seat dies at startup with a message that names
images and never mentions tokenizers:
ValueError: Mismatch in `image` token count between text and `input_ids`.
Got ids=[2047] and text=[16384].
The cap also silently limited image resolution well before it killed
anything -- at 2048 the largest servable image is about 1448x1448, since
(edge/patch)^2 / merge^2 image tokens have to fit under it.
Fix saves a pristine tokenizer re-read from the source rather than the
mutated calibration object, and then asserts truncation is null so the
defect fails the build instead of shipping again.
Playbook gains section 3.14 with the symptom, the cause, the audit one-liner
and a table of which builds were affected, plus a fourth mandatory post-step.
The transferable lesson is called out: this is the third case of an artifact
carrying config authored against an older transformers that a newer one
begins enforcing, so an image bump is a config-compatibility event rather
than just a version change.
|
||
|
vh
|
f90a5025de |
feat(coldfusion-abliteration): Heretic-300 — 8/100 refusals at KL 0.0136, beats the heresy bar 3.6x
Ran Heretic v1.4.0's 300-trial TPE search on Cold-Fusion. Best trial scores 8/100 refusals at KL 0.0136 against a 98/100 base, versus absolute-heresy at 29/100 and our hand-tuned Robinson L35 at 72/100 / KL 0.0116 — i.e. 64 fewer refusals for the same damage. Hand-verified coherent: correct arithmetic with shown working, clean code, 66-167 word prose across nine probes. Durable findings: - direction_scope=0 (single shared direction) is decisive on this merged base: n=129, best 8/100. Per-layer directions n=131 never beat 52/100 despite a better median. Points against the multi-direction intuition for a diffuse direction (our two-template |cos| is 0.62 vs Robinson's 0.99 on stock). - Aggression is not the lever. r(KL, refusals) = -0.561 over 261 trials; the KL<0.02 band contains both the worst results (median 87/100) and the single best. A KL 0.3554 trial scored worse than one at 0.0193. - PR #317 confirmed: Heretic silently drops the MTP head on save. Source 1199 tensors -> export 1184, all 15 mtp.* gone, vision 333/333 intact, exit 0, no warning. This is also why absolute-heresy ships a byte-identical MTP head — a bug, not a design choice. Always diff tensor keys after a Heretic export. - Heretic's recovered direction carries 6.18% of its energy in sink dim 3994, versus 0.094% for our L35 and 1.97% for the L39 we rejected as brick-inducing. It survives that only because of magnitude-preserving ablation (row_normalization=FULL); our plain projection has no such protection, so the sink screen correctly refused the in-band MTP graft. Same direction, different operation. MPOA is the prerequisite for in-band MTP on a Heretic trunk. - Heretic's edit is recoverable from weights: delta is rank-1 (s2/s1 ~ 0.010), SVD gives the direction, norms give per-layer weights (1.08 -> 1.34, i.e. over-projection). Cross-layer |cos| agreement 0.9903 independently confirms the single-direction result. New tooling in services/coldfusion-abliteration/: kl_divergence.py first-token KL, class-split, zero noise floor catatonia_gate.py 12 probes x 220 tokens, prints every completion heretic_export.py PTY driver; selects by measured value, never by menu position — Heretic's resume prompt puts "delete the checkpoint and all results" one arrow-key from the target graft_mtp.py recovers the trunk direction by SVD; --pristine for the safe path when the sink screen refuses Also adds quant playbook 3.13: the NVFP4 recipe sets observer="imatrix_mse" but llm-compressor has always silently fallen back to uniform MSE for want of importance data — on this build and on the incumbent. Existing A/B comparisons stay valid since every build shares the fallback. Parked as id 42. Guardrail note: this build has lost the self-harm guardrail that the Robinson L35 build retained. Restoration is the operator's own work item. |
||
|
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
|
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
|
2f2bbce73d |
docs(quant): correct 3.8 — TWO real causes, not a lone defective quant
Operator correction to the prior 3.8 framing (
|
||
|
vh
|
d28a371049 |
fix(gen-seat): AEON W4A4 was the defect — purged; mixed FP8-attn build is primary gen
Root cause of the multi-day degeneration hunt, operator-confirmed: the AEON NVFP4 W4A4 quant (sakamakismile/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-NVFP4, full W4A4 incl. attention) went degenerate ~15-20% of generations in real multi-turn use and forced regenerates. MTP, prefix-caching, and the gateway all merely AMPLIFIED it, which is why MTP-off, APC-off, and the vLLM #51113 fix each 'helped' a synthetic probe without fixing it -- three plausible false root-causes, each passing one clean run then failing in real use. The fix was the WEIGHTS: the in-house JonathanColetti/Heretic mixed NVFP4+FP8 build (qwen38-27b-uncensored-nvfp4-mixed, FP8 attention not W4A4, same base, same MTP) is coherent through long multi-turn with MTP ON. W4A4 *attention* was the defect; FP8 attention is not. This commit: - GEN_MODEL -> the mixed FP8-attn build (primary gen until DavidAU 3.8 lands) - GEN_IMAGE pinned to vllm/vllm-openai:nightly-311b3513... (v0.27.2rc1.dev150, carries #51113; pinned by sha so it does not drift on the next pull) - AEON weights PURGED from /tank (no-good), safety-checked not-in-use first - playbook 3.8: the stochastic-W4A4-degeneration lesson + isolate-weights-early + do-not-declare-a-fix-from-one-probe (it validated three non-fixes) AEON is re-pullable from HF if ever needed, but the operator ruled it no-good. |
||
|
vh
|
63a3cb2d86 |
fix(gen-seat): MTP mitigation — disable prefix caching, keep MTP (speed restored, multi-turn clean)
The qwen3_5_mtp corruption (playbook 3.7) is gated on MTP x prefix-caching
TOGETHER (vllm#43559 / #47194), per both cross-frontier peers. Disabling
prefix caching (--no-enable-prefix-caching; vLLM V1 defaults it ON, so the
explicit --no- form is required) forces the GDN cache into a mode where the
partial-accept align-path bug is inert, so MTP can stay on.
Verified on our stack (AEON W4A4): MTP on + prefix-caching off -> the 7-turn
varied series stays coherent through 3.9k tokens, zero cross-turn bleed, at
104.6 tok/s / 53.6% acceptance -- the FULL MTP speedup restored (vs ~half
with MTP off), losing only prefix-cache reuse. All 7 aliases route.
Ruled out on the way: num_speculative_tokens=1 (corruption is
depth-independent, n=1 and n=2 both corrupt); switching to SGLang (vLLM /
SGLang / llama.cpp mainline all share the architectural GDN-rollback bug).
Proper upstream fix (#51113) is in main / v0.27.2rc0 only, not stable, so we
hold at APC-off rather than jump the fleet gateway to an RC.
Supersedes the MTP-off config from
|
||
|
vh
|
a8ed6e7428 |
docs(quant): record the MTP-corrupts-Qwen3.8-multi-turn lesson (playbook 3.7)
The single hardest bug of the night, and invisible to the existing acceptance gate: a LOADED, healthy-accepting MTP head still corrupts Qwen3.8-27B multi-turn output past ~2k cumulative tokens (length collapse + cross-turn content bleed), while single-turn is perfect. Model-independent across all three of our Qwen3.8 quants; Qwen3.6 on the same qwen3_5_mtp method is clean; disabling MTP fixes it. New rule: gate MTP on a multi-turn coherence probe, not just single-shot acceptance. |
||
|
vh
|
a91cc3fb38 |
docs(quant): consolidate quantization lessons into a durable playbook
Quants are hard-fought and we keep re-paying for the same lessons. A survey found quant knowledge scattered across 18 files in four trees, with three documents having independently discovered and recorded overlapping "landmines" sections — and one of them now actively misleading. Adds docs/pfi/model-quantization-playbook.md as the single home for the TRANSFERABLE lessons, with per-model artifacts demoted to worked examples that link up to it. Contents: - scheme decision table, incl. that a literal "W4A8" NVFP4 checkpoint is unservable on vLLM (two legal activation settings, FP8 is not one) - the reference mixed-precision recipe and the three parts of it that are load-bearing and easy to drop - the recurring landmines, ordered by cost: the loader-class trap (rediscovered THREE times), the three separate ways to lose the MTP head, toolchain deadlocks, vision configs, memory/device placement - pipeline shape: prove targets before spending GPU time; mandatory post-steps that verify rather than assume - the acceptance gate, and the three ways measurement has lied to us — prefix caching faking both speed metrics, prompt_logprobs going uniform under speculative decoding, and a 0600 .env making compose silently no-op - hardware/co-residency, including that a SMALLER model can starve its neighbour because gpu-memory-utilization is a fraction of the whole card - a superseded-claims table, and measured negatives not to re-chase The superseded table earns its place immediately: the heretic2 runbook tells readers to use modelopt because "compressed-tensors can't load the BF16 MTP head, 0% acceptance". That symptom was real but the cause was not the format -- it was the missing re:^mtp.* ignore entry. compressed-tensors gives 47.7-83.2% acceptance in production. A fresh session following that doc would be sent down the modelopt path that current memory calls dependency hell, so the runbook now carries a stale-warning header pointing here. Wires discovery: an orientation.md "Where to look for what" row, pointers from the gen-seat / heretic2 / mistral artifacts, and a CLAUDE.md maintenance rule so the playbook gets fed instead of going stale -- model-agnostic lessons land in the playbook, model-specific ones stay put, and a wrong claim earns a dated superseded row rather than a silent edit. Motivated by Qwen3.8 having just released: the next model swap will need a requant, and this is what that session should read first. |