b8dbe71a1ca26f32992ce01c868877dbad9f52ad
15 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 vh
|
36c173c6a1 |
feat(mog-sec): quant + serve M.O.G.-SEC pen-test seat; PPL on gen; retire fable
Autonomous overnight run under the operator's full-autonomy grant. End state:
fleet up, gen seat untouched, a new verified pen-test seat serving where fable was.
PPL on the orcarouter gen seat (fable downed to free GPU1 for a nospec probe,
probe torn down after): mean 7.07 / median 5.76, within noise of heresy 6.910 /
5.625 and identical to our recipe's usual 7.059. The gen-seat search is settled.
M.O.G.-SEC: chose Blackfrost-Research/M.O.G.-SEC-27B-1M-CTX-BF16 (rev deede677)
over the pre-made ModelOpt NVFP4, which was disqualified on W4A4 4-bit activations
(the AEON degradation mode, catastrophic on a 1M-context model), zero MTP tensors,
and ModelOpt format. Pulled, format-screened (P(<think>) 1.11e-05, clean), quanted
in-house to mixed NVFP4+FP8 (23.4 GB, MTP + vision preserved), and served in the
retired fable slot.
stacks/mog-sec ana-ml2 GPU1 :8019, KV 418,218 tok / 1.60x @ 262K
aliases mog-sec (non-thinking), mog-sec-reasoning (thinking)
gates surface 6/6, MTP 55.3%, format 0/15 leak, vision 7/3/1,
capability 4/4 (delivers offensive-security content)
Served at native 262K, NOT the card's 1M -- the 1M needs YaRN (absent from the
weights' config) plus the SGLang/DFlash2 path the repo ships a deployment kit for,
neither of which is our vLLM surface. A real 1M seat is a separate SGLang project.
Retired char-rp-reasoning + char-rp-fable (zero traffic, pointed at the downed
fable :8019; now 404 cleanly, not repointed -- a security model is not an RP model).
char-rp (meromero) untouched. Vision preprocessor built from the model's own
image_processor block, same trick as the MeroMero seat.
GPU0 seats (gen, meromero) were untouched and healthy throughout. The quant ran in
GPU1 free space with no production seat stopped except fable, which was replaced.
|
||
|
vh
|
e4576f0989 |
test(gen-seat): PPL on orcarouter — mean 7.07 / median 5.76, within noise of heresy
Measured against a spec-decode-free probe on GPU1 (fable downed to free the VRAM, probe torn down after). eval_quality.py aborts PPL under --speculative-config, so a nospec probe is the only way to read it. orcarouter mean 7.0655 median 5.758 heresy mean 6.910 median 5.625 (+2.2% mean) our recipe's usual mixed-quant PPL: 7.059 -- orcarouter is identical to it So orcarouter matches heresy on fidelity and wins on every other axis: MTP acceptance 58.4% vs 47.2%, zero think-leak, vision 7/8. The gen-seat search that ran through Cold-Fusion, heresy, and preetpatel is settled on orcarouter. |
||
|
vh
|
ce09ac4fa6 |
test(gen-seat): add a real vision battery — orcarouter scores 7/8
surface_test.py's vision check is one image and one word. It proves the tower loads; it does not prove the tower works. This battery uses generated images with known ground truth so every answer is objectively gradeable. Against orcarouter NVFP4-mixed on the `gen` alias: T1 OCR, 5 lines incl. one at 18px PASS all 5 exact T2 counting + attribute binding PASS 7 circles / 3 triangles / 1 square T3 bar chart, 6 values + max/min PASS 6/6 exact T4b occlusion, star behind rectangle PASS T4c aspect ratio of a 160x140 rectangle FAIL called it taller than wide T5 two images, which has text PASS T6 four images, the seat's cap PASS all four named T7 five images, one over the cap PASS rejected with HTTP 400 No <think> leak on any vision call. The single miss is fine-grained relative-dimension estimation on a near-square shape, and it reproduced across two runs (the longer T4 called the same rectangle "equal width and height"). Counting, OCR, chart values and occlusion ordering are all solid, so this is a precise-geometry weakness, not a broken tower. Recorded so nobody builds a feature on this model judging relative sizes. T7 earns its place separately: it confirms the per-prompt image cap fails loudly with a 400 rather than silently dropping the extra image. |
||
|
vh
|
f85d102813 |
test(gen-seat): orcarouter passes every gate — in-band MTP head delivers +11 points
Gates run against the live seat while the operator tested in parallel. <think> leak (n=30, 4 prompt types + multi-turn) 0/30, 0 empty MTP acceptance 58.4% @ 117.11 tok/s median surface 6/6 abliteration survival 4/4 compliance deterministic quality gens coherent and correct PPL still blocked For scale on the leak gate, the abandoned h300 build scored 8/30 on this exact instrument, and its abliteration-survival samples had 2 of 4 open with "<think>Ok, let's figure this out:". Orcarouter has none. The headline is MTP acceptance. 58.4% against heresy's byte-identical base head at 47.2% is +11 points, and it sits level with our own in-band L35 at 59.1%. That is the additive in-band-vs-graft delta the entire Cold-Fusion experiment was built to measure and never cleanly delivered -- orcarouter handed it over for free because the author had already done the Robinson edit on the head. Surface 6/6 covers plain chat, vision, tool calling, the thinking split, a 36,042-token long-context retrieval, and streaming. PPL remains blocked on a spec-decode-free probe seat: it needs ~22 GB and GPU1 has ~16 GB free. Comparison target is heresy at 6.910. |
||
|
vh
|
c8f128bdff |
feat(gen-seat): quant orcarouter — its MTP head is already Robinson-abliterated in-band
Pulled orcarouter/Qwen3.8-27B-Uncensored at rev 9878936b (55.5 GB, gated, our token has access) and built /tank/aimodels/qwen38-27b-orcarouter-nvfp4-mixed (23.4 GB, mixed NVFP4+FP8). Verified, not yet cut over. The operator asked whether we could apply the Robinson path to the MTP head. We cannot, because the author already did. compare_mtp_head.py against the verbatim base graft: 13 of 15 tensors byte-identical, exactly 2 differ -- mtp.layers.0.self_attn.o_proj.weight and mtp.layers.0.mlp.down_proj.weight, which are precisely the two residual writers our own abliterate.py targets (EXPECT_MTP_WRITERS = 2). Reverse-engineered the edit from the weights alone (mtp_delta.py, added here): sigma2/sigma1 = 0.0164 on BOTH tensors rank-1, a single-direction projection |cos| between the two recovered dirs = 1.0000 ONE shared direction ||delta||/||W|| = 1.42% and 1.41% a gentle, consistent projection sink energy dim 3994 = 0.0000% sink-clean; Heretic's was 6.18% That is the Robinson in-band MTP abliteration, already applied, with a direction that passes our sink screen outright. Nothing to do but preserve it, and the quant carries it byte-identically. This is the configuration the entire Cold-Fusion experiment was designed to test and never cleanly delivered. The new format screen paid for itself on its first real use: think_prior.py on the bf16 BEFORE any GPU time gave P(<think>) = 1.23e-06 at rank 52, against Cold-Fusion stock 0.1850 and h300 0.2216. Roughly 150,000x cleaner. Two durable findings about the pipeline itself: The quant needs ~17 GB, not a whole card. It ran entirely in GPU1's spare 16 GB with ZERO production seats stopped -- the h300 run's "stop BOTH GPU0 seats" was never necessary, it simply had a free card by coincidence. The first attempt OOM'd by 2.37 GiB at layer 64 of 65 with 3.57 GiB reserved-but-unallocated, which is fragmentation, and PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True closed it. post_quant.py now builds a missing output index from the safetensors headers. A sub-23 GB quant saves one bare shard with no index, and post_quant needs one; this has broken three separate rounds and been hand-fixed every time. The header is read by struct-unpacking the u64 length and parsing the JSON -- never safe_open, which mmaps the whole 22 GB shard and ENOMEMs on ZFS. Artifact verified: mixed-precision, 1968 tensors, 15 mtp, 333 visual, re:^mtp.* present in the ignore list (llm-compressor pruned it as always), preproc restored. Imatrix deferred per operator; the log confirms the usual uniform-MSE fallback, so this build stays apples-to-apples with heresy's PPL 6.910. |
||
|
vh
|
37e9e1ca7f |
revert(gen-seat): abandon Cold-Fusion, roll back to heresy — the leak is in the base
Operator directive, given before the result was in: if it's the base, abandon h300 and the base too. The dose-response said base (18.5% of 22.2%), so it fired. Live gen seat is /tank/aimodels/qwen38-27b-heresy-nvfp4-mixed again, restored from .env.bak-coldfusion-L35-20260820. The h300 env is preserved at .env.bak-h300-abandoned-20260821. The clincher, same probe pointed at heresy: Cold-Fusion stock P(<think>) 0.1850 Cold-Fusion L35 0.2048 Cold-Fusion h300 0.2216 heresy (restored) not in the top 20, <0.002 A >100x gap between the families, which is why no rollback inside Cold-Fusion would have helped -- stock and L35 leak at nearly the h300 rate. Verified after rollback: 0/30 leaks and 0 empty on the same instrument that scored h300 at 8/30, with the EXISTING enable_thinking:false config; KV pool 403,065 tok / 1.54x, heresy's exact documented baseline; 7/7 aliases; vision intact. No LiteLLM change was needed, so the chat_template_kwargs fix is left unapplied -- it worked, but it was a workaround for a base we no longer serve. Cost, stated plainly: 8/100 refusals becomes 29/100, a 3.6x regression on the axis the whole Heretic-300 run existed to move. Accepted deliberately. What carries forward is the methodology, none of which lived in the Cold-Fusion weights: direction_scope=0 beating per-layer on a merged base, aggression not being the lever, PR #317 silently dropping the MTP head on save, the MPOA and sink-screen reasoning, the graft/KL/catatonia/export harnesses, and the finding that a pristine MTP graft accepts as well as an in-band edit. New acceptance gate earned here: run think_prior.py on a candidate's STOCK weights before committing GPU time. It is a ~10s CPU measurement and it would have disqualified Cold-Fusion before the 300-trial study ever started. Heretic's objective has no format-compliance term at all -- the same blindness that removed the self-harm guardrail. Nothing deleted. Every Cold-Fusion artifact, the 300-trial Optuna journal and catatonia-T260.json remain on disk. Abandon means stop serving, not rm. |
||
|
vh
|
5ee2325820 |
feat(coldfusion-abliteration): dose-response settles the <think> leak — base 83%, our abliteration 17%
Answers "how likely is it that our abliteration caused this?" with a measurement instead of a prior. P(<think>) at the first generated token, template rendered enable_thinking=false so the prompt already carries a CLOSED think pair -- the exact event behind the leak. Raw softmax, bf16, CPU-only, one process per model. Deterministic: stock reproduced to 17 significant figures across two runs. coldfusion-bf16 none (stock) 0.1850 rank 3 coldfusion-abliterated-L35-bf16 Robinson L35, mild 0.2048 rank 2 coldfusion-h300-mtp-bf16 Heretic-300, heavy 0.2216 rank 2 The stock, untouched base already puts 18.5% of first-token mass on opening a think block the template had closed. Abliteration adds a real, monotonic, dose-dependent +3.7 points -- a nudge on a pre-existing base, not the cause. Cold-Fusion is a reasoning-token-compression finetune, i.e. a model trained to think briefly, and the leak's text shape agrees: a compact correct trace with a trained transition marker, which is trained behavior rather than damage. This changes the options. Rolling back to L35 or stock does NOT fix the leak -- at 18.5% under temp 0.7 / top_p 0.8 they leak at nearly the h300 rate. Only leaving the Cold-Fusion family escapes it, at the cost of the 8/100 refusal result. The chat_template_kwargs fix is the correct lever. Durable methodology point: a forward-KL budget cannot catch this. Heretic minimizes forward KL(stock||abliterated), which is near-blind to the model putting new mass on tokens stock barely used -- that is reverse KL's job, and we measured exactly that asymmetry on L35 (reverse 1.43 vs forward 0.70). h300's KL of 0.0136 is not evidence of innocence. For any "did the abliteration break behavior X" question, measure P(token) directly. Ran CPU-only deliberately: 96 EPYC cores and 265 GB of RAM make a 27B forward pass cheap, so this cost no GPU window and no seat downtime, where the obvious route was stopping both GPU0 seats. Also normalizes two more abliteration output dirs from root-owned 0600 to llmuser 0664. The unreadable-model failure surfaces as FileNotFoundError rather than a permission error, which is worth knowing before it wastes a run. |
||
|
vh
|
91f4cf22e1 |
fix(gen-seat): diagnose the unterminated-<think> leak — model defect, temp-triggered
Operator reported the new Heretic-300 gen seat "sends CoT but never completes
the turn" through Lobe. Diagnosed; not yet fixed (the fix changes gen's
semantics, so it is the operator's call).
The Qwen3.8 chat template appends a pre-closed <think>\n\n</think>\n\n when
enable_thinking is false. The h300 model opens a fresh <think> anyway and never
closes it. Because the prompt already closed the block, vLLM's qwen3 reasoning
parser is not in reasoning state, so the tag passes through as ordinary text --
reasoning_content empty, reasoning_tokens 0, and the whole reasoning-plus-answer
blob lands in content. Lobe then correctly treats the unterminated tag as
still-thinking and renders no answer. The client and the serving stack are both
behaving correctly; the model is not.
The trigger is TEMPERATURE, not presence_penalty (n=12 per arm):
temp 0.7, pp 1.5 (current gen) 4/12
temp 0.7, pp 0.0 4/12
temp 0.7, pp 0.5 3/12
temp 0, pp 1.5 0/12
That falsifies the standing hypothesis, recorded in the litellm config comment
and in the operator's own 2026-08-16 note, that presence_penalty 1.5 is the
first dial to move. It is not this bug's cause.
It also explains the blast radius: only the two temp-0.7 aliases leak, `gen`
and `summarizer-large`. summarizer, classifier, image-judge and qwen-image-bench
all run at temp 0 and are clean, so nevermore's summarizer path is unaffected.
Candidate fix, validated n=30 over 4 prompt types plus a 3-turn conversation:
chat_template_kwargs {enable_thinking: true, reasoning_effort: low} takes 8/30
leaks to 0/30, at ~+27% completion tokens and a ~3% empty-content residual.
The tell appears in eval_coldfusion_h300.json and in none of the aeon, heresy,
mixed or w4a16 evals, so it is new with this build -- but L35 was never evaled,
so this does not separate a Cold-Fusion base trait from a Heretic-300
abliteration artifact.
Reproducers and the full method land in bench/think-leak/. Note in particular
that the 7/7 alias smoke test run at cutover structurally could not catch this:
trivial prompts never invite reasoning, so they never sample the leaking token.
|
||
|
vh
|
e3ce713f7f |
feat(gen-seat): cut over to Heretic-300 — 7/7 aliases, vision intact, MTP 59.7%
Live GEN_MODEL is now qwen38-27b-coldfusion-h300-nvfp4-mixed (ana-ml2 GPU0 :8015). Served-name left unchanged so all 7 LiteLLM aliases route without a gateway edit. Verification: KV pool 401,550 tok / 1.53x (baseline 403k / 1.54x) LiteLLM aliases 7/7 green vision 3/3 shapes, colour+form+position correct MTP acceptance 59.7% median @ 118.37 tok/s quality gens 4/4 correct abliteration 4/4 compliance PPL NOT measured (see below) The roadmap predicted ~47% acceptance for a pristine MTP graft versus L35's 59.1% in-band edit. Measured 59.7% on the same harness: there is no acceptance penalty, which removes the throughput argument for reimplementing MPOA. A single long-prose generation read 47.5% off the same counters -- below the 8-run minimum of 49.0% -- and would have "confirmed" the prediction by coincidence. Acceptance must be read from quickbench.py, never one sample. PPL is blocked on VRAM, not on the model: eval_quality.py aborts with "prompt_logprobs look uniform" under --speculative-config, and the probe-seat workaround needs ~22 GB while both cards sit at ~96% committed. Also normalizes the quant dir from root:0600 to llmuser:llmuser 0664 to match every other model dir, and records that config.json sha256 is byte-identical across the h300 and L35 quants and is therefore useless for confirming which weights are mounted (mtime and a head-hash are the discriminating views). Rollback is one line to .env.bak-pre-h300-20260820. |
||
|
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
|
11b688ff68 |
feat(gen-seat): promote absolute-heresy to the live gen seat
MuXodious/Qwen3.8-27B-absolute-heresy (Heretic v1.4.0 + SOMPOA, trial T377, pin c2374593) quantized through our mixed NVFP4+FP8 recipe and promoted after passing the full gate on the probe port. Gate vs incumbent -- MTP acceptance 47.2% (48.2%), decode 103.5 tok/s (96.4), prefill 6618/5403 at 6.7k/27k (6334/5085), TTFT 27k 5.00s (5.31s), perplexity 6.910 (7.059, 2.1% better), surface 6/6, abliteration compliance 4/4. On our battery-instruct arm -- the framing that actually elicits refusals -- 0/55 with zero EMPTY, so no catatonia at the hard edge. Speed deltas are image-confounded: the probe ran the seat's pinned nightly while the incumbent's stored numbers came from an earlier image. Read as not worse. Acceptance, perplexity, surface and refusal are apples-to-apples. All 7 LiteLLM aliases verified end-to-end. GPU0 at 91.3/97.9 GB with meromero healthy -- more headroom than the previous build. Incumbent weights untouched and .env.bak-heresy-20260817 in place for rollback. Candidate is a 2-day-old RC1 with ~348 downloads; watch real multi-turn use. |
||
|
vh
|
b0c2d3d1c4 |
fix(bench): serve_probe must mirror the live seat -- image, parsers, context
Three defects, each of which produced a false read on the candidate: 1. Hardcoded vllm/vllm-openai:latest. The Qwen3.8-27B gen seat is pinned to a nightly carrying the #51113 qwen3_5_mtp x GDN fix; probing on :latest reproduces the multi-turn corruption we already diagnosed and reads as a candidate failure. Now PROBE_IMAGE, defaulting to :latest for older seats. 2. --speculative-config JSON died twice on quoting. The inner double quotes are stripped by the outer double-quoted ssh string, and then bash BRACE EXPANSION splits {"a":1,"b":2} on the comma. Needs escaped quotes AND remote-side single quotes; both traps documented inline. 3. No --tool-call-parser/--enable-auto-tool-choice/--reasoning-parser. Without them surface_test reported tool calling as a 400 and measured a thinking split of reasoning=0ch -- both probe-config artifacts, not model defects. Re-running with the seat's flags took the candidate from 5/6 to 6/6. Also adds PROBE_MAXLEN; the hardcoded 32768 rejected prefill_bench's ~27k prompt. |
||
|
vh
|
d47dd10795 |
feat(gen-seat): swap to AEON-ULTIMATE NVFP4 W4A4; default reasoning_effort=medium
Operator call: the incumbent abliterated model was the first one we could
find, not an optimised pick. sakamakismile/Qwen3.8-27B-AEON-ULTIMATE-
UNCENSORED-NVFP4 (base AEON-7 BF16, abliterix-abliterated, Apache-2.0),
byte-verified at /tank/aimodels/qwen38-27b-aeon-ultimate-nvfp4.
Measured on the same harness, same GPU, cache-busted per playbook 5.
Baseline was RE-measured live before the swap rather than trusted:
incumbent (W4A4+FP8 mixed) AEON (W4A4)
decode bs=1 94.09 tok/s 104.22 +10.8%
MTP acceptance 47.7% 52.3% +4.6pp
abliteration 4/4 4/4
surface 6/6 6/6
weights 22.5 GB 20.6 GB -8.4%
AEON concurrency: conc=1 98.48 tok/s aggregate; conc=6 381.29 aggregate /
63.55 per-stream, MTP holding 50.6% under load.
Surface 6/6 includes vision (image-judge rides this seat) and a 36k-token
needle retrieval, which was the specific risk in going full-W4A4 -- the
packager only validated 32k, and W4A4 long-context collapse is in our own
notes from the Granite work. It held.
reasoning_effort: the AEON template defaults to xhigh (template line 47),
and at xhigh this model can spend its entire budget inside <think> and
emit no answer -- a silent-empty-response hazard for the automated
summarizer/classifier consumers. Seat now pins the default to medium via
--default-chat-template-kwargs, per-request overridable. Override PROVEN
live: chat_template_kwargs.reasoning_effort=bogus returns HTTP 400
carrying the template's own exception text, so caller values genuinely
reach the template and invalid ones fail loudly rather than silently
falling back. Empty GEN_REASONING_EFFORT omits the flag for models that do
not read the kwarg -- the Qwen3.6 line ignores it entirely, where setting
it would be a false lever.
All 7 aliases verified routing. Rollback is one .env line; the previous
build is untouched at /tank/aimodels/qwen38-27b-uncensored-nvfp4-mixed.
TWO GAPS, declared:
- Incumbent concurrency was never captured before the swap (I baselined
bs=1 only), so the conc=1/6 figures have no same-hardware comparator.
- Perplexity NOT measured. eval_quality correctly refused it: under
--speculative-config prompt_logprobs come back ~uniform (median rank
~130k), playbook trap 2. A real PPL number needs both seats served
without spec-decode.
Adds concbench.py (concurrent throughput; wall-clock aggregate, not
sum-of-rates, and delta-based MTP accounting).
|
||
|
vh
|
4a5c3fcccf |
perf(gen-seat): record prefill measurements — roughly doubled
Closes the one axis of the original premise left unverified. Measured cold (cache-busted) on both builds under matching serve configs: ~6.7k-token prompt 3,206 -> 6,334 tok/s prefill (+98%) ~27k-token prompt 2,862 -> 5,085 tok/s prefill (+78%) TTFT on a ~27k doc 9.43 -> 5.31 s (-44%) Prefill gains far exceed the +18% decode gain, and that ordering is the expected one: decode at bs=1 is memory-bandwidth-bound and the weights are 4-bit under either scheme, so little changes; prefill is compute-bound, which is where native Blackwell FP4 tensor cores replace the Marlin dequant-to-BF16 path. The summarizer aliases are the consumers that feel this. Adds bench/prefill_bench.py plus the raw JSON. The harness deliberately uses SystemRandom: a seeded nonce regenerates the previous run's prompts verbatim, prefix caching then serves them, and the first attempt read ~41k tok/s of cache-hit rather than ~5k of actual prefill. |
||
|
vh
|
74f596b1d3 |
feat(gen-seat): mixed NVFP4+FP8 requant — +18% decode at equal MTP acceptance
Re-quantizes the fleet `gen` seat from weight-only NVFP4A16 to a mixed-precision build: NVFP4 W4A4 for layers 0-55 MLPs, FP8 W8A8 for the attention projections / linear_attn / lm_head / layers 56-63 MLPs, FP8 KV cache. Replicates the scheme of unsloth/Qwen3.8-27B-NVFP4 on the abliterated weights. The queued task named this "W4A8" (NVFP4 weights + FP8 activations). That checkpoint cannot be served: vLLM 0.24's compressed-tensors dispatcher (compressed_tensors.py:704-713) accepts NVFP4 weights with either no input quantization (W4A16, which forces the Marlin kernel) or NVFP4 input quantization (W4A4) -- anything else, FP8 included, raises ValueError at load. CompressedTensorsW4A8Fp8 is INT4 weights gated on an exact-sm90 check, so it is closed on Blackwell twice over. The ~20% intuition was correct; the scheme name was not. Getting FP8 into the mix has to be done per-layer-group. Established the gain before spending GPU time: unsloth's build was already on-box, so serving it as a probe measured +19.1% over our seat at identical MTP acceptance -- a kernel-level result, no requant needed to learn it. Measured, cache-busted, bs=1: decode 80.12 -> 94.53 tok/s (+18.0%) MTP acceptance 47.8% -> 47.7% (unchanged) perplexity (n=6) 6.941 -> 7.059 (+1.7%) abliteration 4/4 -> 4/4 (preserved) weights on disk 27.7 -> 22.5 GB (-19%) Surface test green on the live seat: plain chat, vision, tool calling, thinking split, 36K-token needle retrieval, streaming. All 7 LiteLLM aliases verified routing. GEN_GPU_MEM_UTIL 0.45 -> 0.43: the new weights are 5.2 GB smaller, and at 0.45 the seat absorbed that slack as KV, leaving meromero-charrp 0.18 GiB short of its budget on the shared GPU0 -- it crash-looped. Handing the space back leaves gen 422K tokens of KV (1.6x its 262K context) and both seats co-resident at 89.8/97.9 GB. Also records two measured negatives so they are not re-chased: GEN_SPEC_TOKENS is already optimal at 3 (swept 2/3/4/5 -> 77.1/80.1/78.7/ 75.9 tok/s), and vLLM's prompt_logprobs are ~uniform while speculative decoding is on, so perplexity must be measured with spec off. Pipeline, acceptance harness and raw measurements land in services/gen-seat-mixed-quant/. Rollback is one .env line; the previous build is untouched at /tank/aimodels/qwen38-27b-uncensored-nvfp4. |