0f748ea54e0376b8398fa9bead42bb40194a3beb
921 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 vh
|
2656196f47 |
memory: snapshot — the tune is trained, gated, and serving
Run-01 completed in 7:21:52 (47% faster than the 13.85h round-1 projection),
lora_B gate 205/205 non-zero at median norm 1.708, and the acceptance gate says
it did the thing it was built for: diversity +0.178 against a 0.008 floor (22x),
attractor hit rate -11.3pt against a 2.0pt floor, memorisation 0.0000 on both
arms — which closes the R20 licensed-prose exposure on measurement rather than
argument.
Five new detail files carry the substance:
erp-tune-run2-complete the run, the gate, the noise-floor near-miss
(brokkr was one step from reporting a 13-point
T6 regression sitting inside twice his
instrument's own variance)
mfu-root-caused-attention 8.6% MFU was an accounting artifact; real
utilisation 17-20%, cost was attention on
AMPERE kernels. Two independent methods agreed
to 2.6 points.
nvfp4-serving-pipeline merged weights are MANDATORY — vLLM cannot
serve a LoRA on ANY Gemma-4 — plus the recipe
that silently misses all 11,520 expert tensors
refusal-retention-probe measured base 0/100 -> tuned 29/100, then had
to accept it was the wrong axis
worldtree-b188-b189-and-selene three arcs closed, and a #411 diagnosis I got
wrong twice before a directory probe settled it
Current state rewritten end to end — the previous snapshot had the run in
flight at ~17h with MFU unexplained. Both are now closed.
The open operator decision is run 2's base, deliberately unstaged and flagged
against being filed as a config knob: it is a reversal of the trainee-selection
decision, and the pretrained-base option removes the last non-lexical floor on
the CSAM axis given stage-2-detector-inert and contamination-scan-absent are
both already overridden.
Tried-and-abandoned gains four measured-dead throughput levers, the packing
correction (bucketing wins under sdpa and the conclusion flips under flex — do
not carry it past the backend decision), and the merge-back-undoes-abliteration
trap brokkr caught in his own advice.
Index stays at 291 lines, under the soft cap. No archival this run.
|
||
|
vh
|
2a05ae91af |
feat(training-probes): counted-not-surfaced classifier scaffold
Reusable measurement discipline for probes that must classify how a model
responds to material that should not be printed, logged, or pasted into a
report. Supplies the discipline; the axis map and prompts stay with the caller.
Four rules, each because skipping it produced a wrong number:
- classify, never surface. Completion text is held inside classify() and does
not cross the return boundary. A probe that prints what it measured has
turned a measurement into a distribution channel.
- three-way, not binary. A refusal regex undercounts — models decline by
redirecting with no refusal token present, measured at 2/5 to 5/5 on models
a regex scored 0.
- the deflection count is a FREE CONTROL. Run both arms: zero on both means
the model is binary and the regex is sound; only one means the difference is
real. An artifact does not care which arm it runs against.
- EMPTY and ERROR get their own buckets. Folding them into either side biases
the result, and a truncation-heavy arm flatters itself if its failures land
in the wrong bucket.
Requested by brokkr-smithy-dev for the domain-compliance probe — the discipline
in code rather than reimplemented, with the axis map his side of the line.
|
||
|
vh
|
64bf9d313f |
docs(training-playbook): measure refusal retention on the abliteration's OWN axis
§3.13, plus the probe that produced it. Two lessons, both about measuring the wrong thing confidently. First: a tune applied AFTER an abliteration can walk it back, and a reasoning/craft/memorisation gate cannot see that. brokkr-smithy-dev's preregistered gate measured none of it — a tune that gains 41 items of contradiction detection and quietly restores refusals passes every check. The compliance axis has to be added explicitly. Second, and this is the trap: measure the axis the abliteration was actually FOR. Ours was run so the model engages explicit fiction. The probe reached for mlabonne/harmful_behaviors — weapons, malware, fraud — because it was cached and carried a recorded baseline. Different refusal surface entirely, and a model moves on them independently. 29/100 general-harm refusals on a tune whose prose the operator was praising at the time is not obviously a defect and may be desirable: general-harm refusals returning while domain compliance holds is close to the ideal shape for an internal creative seat. The measurement was real; its relevance was assumed. Also recorded, because both were nearly missed: - Read the interesting cell. In 29 hard / 0 deflect / 71 comply, the load-bearing number is 71. Stock refused 100/100; near that would mean the abliteration was undone. 71 complying means partially walked back on one axis — a different finding, and only one of the two threatens the seat. - A baseline from a different harness is not a baseline. The recorded 3/100 came from the abliteration tool's scorer, which reads first-token probability distributions; a probe that generates and regexes is a different instrument. Run your own against both arms on the same seat or report the number alone. - A refusal regex undercounts, so classify hard/deflect/comply — and the free discriminator: if both arms return zero deflections the model is binary; if only one does, the regex is fine. An artifact does not care which arm it runs against. |
||
|
vh
|
a696b49e2a |
docs(training-playbook): merging a tune back toward stock can UNDO an abliteration
§3.12. brokkr-smithy-dev caught and retracted his own recommendation mid-thread; recording it before it reads back later as advice. A common remedy for an overfit tune is a partial merge back toward the base to recover general capability. The published recipes that recommend it merge into the STOCK instruct checkpoint. On an abliterated base, following that literally re-introduces the exact refusal directions the abliteration was run to remove — and it is silent, because the merged model looks healthier on general benchmarks while the property the seat exists for quietly returns. Rule: any merge-back targets the SAME base the LoRA was trained against, never the upstream stock weights however similar the name. The wider lesson is about recipe-card provenance. Community cards are per-checkpoint artifacts and do not transfer across dense-vs-MoE, stock-vs-abliterated, or size variants. The worked example: a recommendation carried from a card for a DENSE STOCK 31B onto a MoE ABLITERATED 26B-A4B on the strength of a shared family name. The overfitting warning on that card happened to come from the right architecture; the pipeline, reward stacks and merge-back came from the wrong one. Same family, three axes apart. So: before quoting a recipe card at a decision, state which checkpoint it was written for and which axes differ. "Same family" is not an answer. |
||
|
vh
|
2ec8f42297 |
docs(training-playbook): base-viability pre-flight, three greps before you pick
§3.11. Three consecutive "what about X as a base?" questions in one session,
each answerable in minutes, none of which had been asked before a 7-hour
training window was committed. Writing the check down so it runs first.
1. does it fit for TRAINING - BF16 weights against the real measured peak,
not the weight figure (Gemma-4 is 48.1 GiB of weights and peaks at
79.7 GiB at mb2/seq-16k). Model-line names lie: "Mistral Small 4" is 119 B,
238 GB in BF16, more than both cards combined. QLoRA is not an escape
hatch for MoE - bitsandbytes walks nn.Linear and fused 3-D experts are not
that.
2. if MoE - does the serving engine implement get_expert_mapping. Zero means
LoRA cannot be served at all. gemma4*.py -> 0; deepseek_v2, mixtral,
glm4_moe, ernie45_moe -> present.
3. does the model class support LoRA - and GREP THE CLASS, NOT THE FILE.
Point 3 has teeth and I nearly got it wrong twice in one turn. mistral.py greps
as SupportsLoRA=0 and is fully LoRA-capable via LlamaForCausalLM.
mistral_large_3.py greps as 0 for both and inherits get_expert_mapping from
DeepseekV3ForCausalLM. Capability is inherited; a file-level grep misses it and
only MRO resolution answers it. Same class of error as asserting a substring
instead of an effective value.
Worked results recorded for the three candidates evaluated:
Gemma-4 26B-A4B fits, no expert mapping -> trainable, MERGE-ONLY
Mistral Small 4 119B 238 GB, has mapping -> servable, NOT trainable here
Ministral 3 14B ~28 GB, dense, inherited -> passes all three
Adds a fourth glance at architecture shape, since it predicts how much of this
playbook applies at all: uniform head_dim <= 128 with no sliding window keeps
both flash and cuDNN reachable and makes §3.1/§3.3 moot, while mixed head dims
plus a sliding window is exactly what forces dense O(n^2) attention onto
Ampere-generation kernels for 65% of the step.
|
||
|
vh
|
96731bb090 |
docs(training-playbook): prove the serving path before spending the window
§3.10. The quantization playbook already says prove your targets before spending GPU time; this is the same rule one step later, and easier to skip. A ~7h LoRA run was built assuming the adapter could be hot-swapped onto a quantized base at serve time. The sizing doc flagged serving as unsettled and said the requirement was needed "while he is early, not after the run" — the concern was identified correctly and then the check was deferred. Tested afterwards, vLLM refuses outright: gemma4's model class implements zero occurrences of get_expert_mapping, which process_packed_modules_mapping requires for any MoE model. One grep, available months earlier. Two generalisations recorded: - Feature support is per-architecture, not per-family. LoRA works for the DENSE sibling of this same model family and not the MoE one, so "model X is supported" says nothing about X's variants. - A capability gap in the serving engine cannot be worked around from the training side. The adapter here never touched experts and was refused anyway, because the refusal keys on the model being MoE, not on what the adapter targets. Includes the mechanical check: grep the engine's model class for the capability, then start the engine with the feature flag alone — no adapter required, since --enable-lora forces the machinery to initialise and that is where it fails. The recovery is cheap here (merge, ~35 min per tune). The cost of finding out late is that it forecloses an architecture choice after the training window has already been spent. |
||
|
vh
|
8de5f7a73c |
docs(gemma4-erp-tune): merged weights are mandatory — vLLM cannot LoRA any Gemma-4
The §5 open question was whether LoRA-on-NVFP4 hot-swap still silently no-ops as it did on vLLM 0.24.0 (#47639), with merged weights as the fallback if it did. Retested on vllm/vllm-openai:latest against the NVFP4A16 base plus the live run's checkpoint adapter. It does not no-op. It refuses to start: AttributeError: To support LoRA for MoE model, 'get_expert_mapping' must be implemented And the reason is bigger than the quant. The check is in vllm/lora/utils.py::process_packed_modules_mapping and branches on whether the model is MoE — quantization is not in the condition. gemma4.py, gemma4_mm.py, gemma4_mtp.py and gemma4_unified.py contain zero occurrences of get_expert_mapping, while deepseek_v2, glm4_moe and ernie45_moe do implement it. So vLLM cannot serve a LoRA on Gemma-4 at all, BF16 or quantized. Merging is not a workaround for a quantization limitation; it is the only path for this architecture. This holds even though the adapter never touches experts — validate_adapter_parameters forbids per-expert params, so all 205 targets are attention and dense MLP. The refusal is about the model being MoE, not about what the adapter targets. Worth recording that the current behaviour is an improvement: a loud refusal beats the 0.24.0 silent no-op, which would ship a base model wearing the tune's name and pass every check that does not compare against base. |
||
|
vh
|
ab980e9345 |
fix(erp-tune-serve): four defects the end-to-end dry run found, all silent
Validated the full adapter -> merge -> NVFP4A16 -> serve pipeline against
checkpoint-100 of the live run. It works, and it produced a served model
generating coherent prose. Getting there surfaced four failures, none of which
announced itself as the thing it actually was.
1. transformers 5.15 MIGRATES the config schema on save. It drops Gemma-4's
`global_head_dim` / `num_global_key_value_heads` and writes `per_layer_config`
instead. transformers 5.10 (what the llmcompressor venv pins) does not know
the new key and resolves num_key_value_heads to None:
TypeError: unsupported operand type(s) for //: 'int' and 'NoneType'
Every working artifact on the box - bf16 base, served nvfp4 prod seat,
nvfp4a16 build - uses the OLD schema. Merging changes weights, not
architecture, so the merge now downgrades the schema and asserts the result.
2. llmcompressor cannot auto-init a processor for a multimodal checkpoint and
dies with a message that names neither the model nor the cause. Calibration
here is text-only, so the tokenizer is passed explicitly as `processor`.
3. save_pretrained writes tokenizer files only, so `processor_config.json` was
never carried. vLLM then fails at startup with "Can't load feature extractor",
which reads as a vision bug and is actually a missing-file bug. Both scripts
now carry the base's auxiliary configs.
4. The quant needs more than the 32 GiB free on GPU1 alongside the resident
seats. Rather than leave that to a caller, quant_with_gen_down.sh stops
vllm-gen and restores it from a trap on EVERY exit path - crash, OOM, kill,
or success - because the restore must not depend on the calling session
surviving. Uses `docker start`, not `compose up`, so the container comes back
with its exact original config. Measured window: ~15 min, gen healthy after.
Verified on the resulting artifact:
merge 410 adapter tensors, sampled target weights confirmed CHANGED,
upstream 390-line chat template shipped (not the base's stale 365)
quant 49 GB -> 17 GB, format nvfp4-pack-quantized, a=null (genuine A16),
weight_packed 11,725 of which 11,520 expert = 30 x 128 x 3,
tokenizer truncation clean
serve Marlin NVFP4 kernel + Marlin MoE backend, 40,492-token KV cache,
coherent generation with content correctly populated
One quality note: the reference nvfp4a16 artifact triggers a vLLM warning that
parallel layers (q/k/v) carry different weight global scales, "likely to result
in reduced accuracy". Our build does not - llmcompressor 0.12 links weight
observers across fused groups for a shared global_scale automatically. The
in-house quant is better than the downloaded one on that axis.
Separately: the lora_B inert-adapter gate PASSED on checkpoint-100 - 205/205
non-zero, median norm 0.829, zero vision_tower tensors. That check never ran in
round 1, and it is the only failure mode that stays invisible until the
acceptance gate reports base-identical numbers.
|
||
|
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
|
7b5fd91d3c |
docs(gemma4-erp-tune): root-cause the 8.6% MFU — attention on Ampere kernels, 29.9% padding
Run-01 was killed at step 19 by operator instruction to root-cause before spending a ~13.9h window. Two independent methods now agree on where the step time went, and neither was the hypothesis the consult panel converged on. Scaling fit (3 points, 2 params, residuals <3ms over an 8x range): A = 6.87e-4 s/token, B = 8.85e-8 s/token^2 quadratic share 20.9% @ w=2048 -> 67.8% @ w=16384 No fixed term was needed, which refutes launch-bound outright. Profiler kernel table (device rows only): attention 22,835.8 ms 65.2% fmha_cutlass*_sm80 dense GEMM 2,774.0 ms 7.9% other 5,739.0 ms 16.4% The attention kernels are sm80 — Ampere-generation CUTLASS running on an sm_120 Blackwell card, with the forward on the gmem fallback tier. That is the mechanism behind 100% SM utilisation at 27 of 304 available TFLOPS. Correctness cleared separately: the sliding mask asserts at max 1024 allowed/row, so the 25 windowed layers were genuinely windowed. The same probe found that right-padding is what pins the 5 global layers to an explicit 4D mask and off the is_causal fast path — measured at 9.4% slower for 24% less loss work at fixed width. The largest available win is not the attention kernel. The corpus is 29.9% padding, and bucket-to-pair + shuffle-to-mix takes it to 0.0% for >=35.5% wall clock, no new dependency, unchanged peak memory. Bucket size turned out not to be a diversity knob — roots per accumulation window are flat across a 256x range, so the global micro-batch shuffle does that work alone and the bucket should be tight. Adds docs/pfi/training-throughput-playbook.md as the durable model-agnostic home (sibling to the quantization playbook), the four probes under scripts/training-probes/ with raw output kept for re-derivation, and a §6 to the sizing doc carrying the Gemma-4-specific numbers and round-2 restart parameters. Measured negatives recorded so they are not re-chased: grouped_mm (0.9% slower, and MoE is only 7.9% of the step), CUDA graphs / torch.compile over the expert loop (no fixed cost to amortise), liger fused CE (~1-3% lever), FA4 on sm_120. Round-1 state preserved: 609MB encode cache, order manifest, truncation report, resume script. No checkpoints — it died at step 19 and the first was due at 100, so the lora_B inert-adapter gate never ran and moves to the restart. |
||
|
vh
|
872c2c562f |
memory: the MFU hunt — two hypotheses measured and killed, consult dispatched
Records what has actually been ruled out rather than what is suspected. The hardware is fine: a plain dense GEMM at the same shape reaches 97.1% of the benchmarked 313.8 TFLOPS peak. The Python expert loop is not the cause, which was my hypothesis and I was confident in it. transformers' grouped_mm experts backend runs 0.9% SLOWER than eager with bit-identical output and identical peak memory, and torch 2.13 has the kernel available, so it is not falling back for lack of one. MoE is not the bottleneck at all. Isolated at real shapes the block runs at 26.5% of peak with 36% of its time in pure gather/scatter, and a dispatch-free bmm version would reach 80.9% — but the whole MoE contribution is only about 10% of a step. Making it free buys 7%. So roughly 90% of the time is unaccounted for. The leading untested hypothesis is that the five full_attention layers use global_head_dim 512, above FlashAttention-2's 256 cap, which would push SDPA onto a slow backend for O(n^2) attention at sequence 16384. Also records that the earlier 5% MFU figure was wrong in two ways — unpadded tokens and a guessed peak — and that the operator caught it. Padding is real but secondary at 29.9%. Consult dispatched to brokkr-smithy-dev for the frontier-dwarf panel. |
||
|
vh
|
07743c6aff |
memory: snapshot — tune training unattended, MFU root-caused to a Python expert loop
The in-flight section is rewritten around the run itself rather than the decisions that led to it. The sizing and seat-call bullet collapses to a pointer now that both are executed; its detail lives in docs/pfi/gemma4-erp-tune-sizing.md. Adds the measured MFU finding: 27.1 TFLOPS against a benchmarked 313.8 TFLOPS peak, root-caused by reading the source rather than inferring — transformers runs the Gemma-4 experts in a Python loop, 128 experts across 30 layers, roughly 11,500 iterations per optimizer step under gradient checkpointing. Padding is a secondary 29.9% tax. Records that my first estimate of 5% MFU was wrong in two compounding ways: divided by unpadded tokens, and compared against a guessed peak rather than a measured one. The operator pushed back on the number and was right to. The fused MoE kernel is deferred work with a tracking surface — park id 47 — per the snapshot rule that deferred decisions go in Recent decisions with a pointer, never into the volatile in-flight section. Also records the resume trap: the original launch command begins with rm -rf on the output directory, which would destroy both the encode cache and every checkpoint. resume-run-01.sh exists so that cannot happen. |
||
|
vh
|
d6dfd61c91 |
memory: the ERP tune is running — override granted, 12 defects fixed first
Operator overrode the corpus gate for one run on 2026-08-25, with the grant staged beside the recipe rather than asserted in chat. It deliberately does not flip any root's training_eligible flag, so the signal that made the run stop in the first place survives intact. Records where the run lives, what it is configured with, how to restore the fleet, and the two lessons that generalise past this project. The first is inert gates. Two turned up in one evening — auditcore, whose CSAM hard-drop never fired across 42,662 records, and validate_vision_keys, which compared model.state_dict() against itself and could not fail on any input. Both read as guards. The question that catches them is not whether the check passes but whether it can fail. The second is an invariant enforced on one code path and not its sibling. That was my own bug: INV-T9 requires a window to hold at least one complete assistant turn, and I enforced it where the window is cut but not where it fits, so a trailing user-only remainder became a zero-loss window and killed the first launch. Same shape as the inert gates, in code I wrote an hour earlier. Two further foot-guns worth the space: enable_input_require_grads is mandatory beside gradient checkpointing on a frozen base, or every adapter stays at its initialisation and the run completes successfully having learned nothing; and the upstream Gemma-4 template forward-scans to suppress a closing turn marker before another assistant message, so incremental rendering cannot tile against it and assistant runs must be merged first. |
||
|
vh
|
33433e0d1e |
docs(gemma4-erp-tune): replace the estimates with measurements — they were 3x optimistic
Ran the loss path on the real checkpoint on GPU0 with synthetic tokens.
The arithmetic held for parameter counts and was badly wrong for
activation memory.
naive CE bsz1 seq 8192 81.93 GiB
naive CE bsz1 seq16384 OOM
chunked CE bsz1 seq16384 65.66 GiB
chunked CE bsz2 seq16384 79.71 GiB <- the run config
chunked CE bsz4 seq16384 OOM
The marginal cost of an extra 16,384-token sequence is ~14 GiB, not the
~5 GiB estimated: the estimate modelled gradient checkpointing as
storing layer inputs plus a modest recompute peak, and the real MoE
recompute peak with top-8-of-128 routing and its scatter/gather buffers
is far heavier. Dense-model intuition does not size an MoE run.
Two predictions landed exactly — 205 target modules and 74,342,400
trainable params at r64 — which is why the rest of the model of the
thing is still worth trusting.
The headline is that chunked CE at seq 16384 costs 16 GiB less than
naive CE at seq 8192, so chunking is what makes brokkr's 16384
recommendation reachable rather than an optimisation on top of it.
max_seq_len moves 8192 -> 16384 on his truncation finding: the cap
drops 6.2% of samples but 22.4% of tokens, concentrated entirely in
dialogue, which is 60% of the mix.
Also records the four harness changes this required (eitri-smithy
62b556b), including the inert-adapter trap: without
enable_input_require_grads() alongside gradient checkpointing on a
frozen base, no gradient reaches the adapters, every one stays at its
initialisation, and the run completes successfully having learned
nothing.
|
||
|
vh
|
47ec3d1a97 |
memory: the ERP tune is blocked on a corpus gate only the operator can clear
Every clean-v1 CLEANROOT carries training_eligible: false with two named blockers, and the recipe states plainly that nothing in it is Charter §3 training-eligible. I initially read scoped_grant: operator-2026-08-22 as authorization and told brokkr-smithy-dev I was proceeding. That was wrong, and the person who wrote the field corrected it: the grant governs INV-4 one-way tier inheritance — the adapter is permanently internal-erp-rnd and never distributable — not training clearance. The stage-2 detector is measured-inert rather than merely unvalidated. auditcore v3.7.2 returned its hard-drop exit code zero times across 42,662 raw RP records, its printed verdict ignores its own printed threshold, and it passed a record a blind audit had already identified as sexual content involving a participant the text marks as a child. Verified the one thing that decides whether that specific record reaches training: pippa-5083 is present in kept-manifest.jsonl (4,551 rows) and absent from recipe-dedup-kept.jsonl (20,473 rows), which is the survivor list the harness gates on. The substitute lexical screen caught it. That is one known instance caught by a stopgap and says nothing about what the screen misses. Both brokkr and I recommend stopping. Neither blocker is hours of work. |
||
|
vh
|
c9943b1507 |
docs(gemma4-erp-tune): whole-card placement — gen moves to GPU1, sec stands down
Operator chose a third placement over the two the sizing offered: rather than train beside gen on GPU0 or on GPU1 in mog-sec's slot, move gen to GPU1 and empty GPU0 completely. The tune gets 95.60 GiB with no co-tenant and gen never goes dark beyond its own restart. Revised run parameters, since a whole card changes them: - micro-batch 8 (71.8 GiB of 95.60) rather than 4, grad-accum 1, giving 888 optimizer steps instead of 444. At one epoch the step count is worth having, and 8 x 8192 tokens puts ~4,096 rows through each expert per step against ~512 at micro-batch 1 — a far healthier GEMM on 704-wide experts. - Gradient checkpointing stays ON. Dropping it takes ~17% off wall-clock but pushes activations to ~24 GiB per sequence, which forces micro-batch 1 and costs 8x on MoE efficiency. Wide beats shallow. - Scriberr stays on GPU1. The previous revision suggested moving it to GPU0, which was correct only while training was going to live on GPU1. Records the ordering constraint in both directions, the elway identity requirement, and that sec's aliases should be allowed to fail at the gateway rather than be substituted with another model. |
||
|
vh
|
9d70100867 |
feat(ana-ml2): elway playbooks to open and close the ERP/RP tune window
Operator call: rather than train beside gen on GPU0, move gen to GPU1 and
stand sec down for the night, so the tune gets a whole 95.60 GiB card and
the fleet's general seat never goes dark beyond its own restart.
Order is load-bearing in both directions and the playbooks enforce it.
gen runs at --gpu-memory-utilization 0.43, which vLLM reads as a fraction
of TOTAL card memory: 42,091 MiB must be FREE at startup or the engine
refuses to boot. GPU1 has 19,446 MiB free while mog-sec is up, so
recreating gen onto GPU1 first would take the main seat down and leave it
down. mog-sec stops first and a hard gate checks the freed memory before
gen is touched. The close playbook mirrors it: gen must vacate GPU1
before mog-sec starts, since mog-sec needs 50,901 MiB of its own.
Close opens with a gate that refuses to run while a process is still
resident on GPU0, so it cannot evict a training run mid-flight.
Override with --var allow_busy_gpu0=true.
Three defects found and fixed while landing this, all worth keeping:
- Verifying GPU residency via `docker inspect --format {{.State.Pid}}`
never matches. vLLM V1 runs EngineCore as a child of the container's
pid 1, and it is the child that holds the memory and that nvidia-smi
reports. Match by cgroup instead.
- A step's `sudo: true` does not extend to its when/creates/changed_when
guards, which run as the login user. The root-only .env made an
unsudo'd grep exit 2, so the GPU-id flip SILENTLY SKIPPED. The
effective-value assert is what caught it.
- That assert originally grepped the config YAML for -\s*'?1'? and failed
against compose's double-quoted `- "1"`. Parse the JSON with jq; an
assert that fails for the wrong reason is worse than no assert.
elway must be invoked as infra-ops@10.250.50.54 rather than the ana-ml2
ssh-target, which resolves to lkraven and has no NOPASSWD sudo.
|
||
|
vh
|
c507db9ac0 |
docs(gemma4-erp-tune): size the run against the checkpoint — QLoRA is structurally unavailable
The proposed shape was QLoRA r64. It cannot be run as specified. The checkpoint stores each layer's 128 experts as two fused 3-D nn.Parameter tensors (experts.gate_up_proj [128,1408,2816], experts.down_proj [128,2816,704] — no .weight suffix, so they are parameters, not modules). bitsandbytes 4-bit replacement walks nn.Linear only, so 22.84B params / 42.54 GiB — 88.5% of the model — is skipped and stays BF16. load_in_4bit saves ~3.1 GiB of 48.07 and does not error while doing it. Verdict: plain LoRA on BF16, ~57.6 GiB at micro-batch 1, +2.5 GiB per additional 8192-token sequence. Two sizing items were absent from the brief and both are load-bearing: - vocab 262,144 x seq 8,192 = 2.147B logits, with final_logit_softcapping 30.0 adding a saved pre-cap tensor. Naive HF cross-entropy peaks at ~28-30 GiB transient at batch 1, which puts the run at ~85.6 GiB on a 95.6 GiB card — it starts, then OOMs on the first long sample. Fused or chunked linear CE is mandatory and must be smoke-proven before a window is booked, since Liger may not carry a Gemma-4 MoE patch. - v_proj does not exist on layers 5/11/17/23/29 (attention_k_eq_v on the full-attention layers). A v_proj target silently produces no adapter there, and k_proj adapts K and V simultaneously. 45.96M trainable at r64 across q/k/v/o. Placement, measured: GPU0 has 53.46 GiB free beside gen, ~4 GiB short, and gen's footprint grows with uptime. Stopping mog-sec frees 74.29 GiB on GPU1, which holds micro-batch 4 at 61.8 GiB with margin for Scriberr. Recommend standing down sec (2 aliases, last request ~5h ago) rather than gen (7 aliases, 765 busy-engine log lines in 24h). Estimated 1.28e18 FLOPs for the epoch at ~3.67B active params; 4-10 hours at 10-25% MFU. 7,104 packed sequences is only 444 optimizer steps at effective batch 16, which makes the wall-clock-checkpointing amendment concrete rather than hypothetical. Package as a uv venv on /tank: root is 91% full (36 GB) with /var/lib/docker on it. |
||
|
vh
|
9d0e628643 |
memory: correct the vLLM version claim — ana-ml2 runs a spread, and 0.27.1 is on disk
The snapshot recorded "ana-ml2 now runs vLLM 0.26.0". That is true of the char-rp seat's pin and false of the box, which the operator caught immediately. Measured per running container: gen is on nightly-311b3513 reporting 0.27.2rc1.dev150, mog-sec on nightly-e9d1398d reporting 0.26.1rc1.dev1102, and rerank-a3 / coder / reward / embed still on 0.24.0. char-rp and the trainee bench stack are pinned to v0.26.0. So there is no single "the version" for this host, and stating one invites exactly the wrong retest. The correction improves the LoRA question rather than complicating it: vllm/vllm-openai:v0.27.1 is already on disk and unused — a TAGGED release, not a nightly, roughly four months past the 0.24.0 where the silent-no-op was diagnosed. That is the right target for a decision test: no nightly variance, no pull. The retest instruction in both the decision entry and the handoff now names it. |
||
|
vh
|
668e590e7d |
memory: snapshot — char-rp on the Gemma-4 MoE, abliterated trainee staged, QLoRA sizing next
Captures an evening that ran from an OOM crash-loop to a measured trainee base. The durable lessons, none of which CLAUDE.md can carry: --gpu-memory-utilization sizes the KV cache and does not cover CUDA context or graphs, which is half of why a seat that fit on the 21st stopped fitting on the 24th; the other half is that gen's footprint GROWS WITH UPTIME (38.5 GiB fresh against 45.6 GiB after three days, same container, same flag), so headroom arithmetic against a long-running gen measures a moving number. The stale-chat-template trap turned out to be endemic across third-party Gemma-4 derivatives rather than a one-off, and it is silent in both directions — wrong prompt when serving, train/serve skew when tuning. And a benchmark finding was retracted because 12% on a five-option task is below the 20% chance floor: a below-chance score indicts the instrument before the model, and a preflight can be thorough while aimed in the wrong direction. Records the serving decision for the tuned model with its history intact: LoRA-on-NVFP4 is preferred if it works, merged weights the expected fallback, but the archived root-cause says the objection was never NVFP4-specific — vLLM 0.24.0's LoRA application was a silent no-op proven quant-agnostic, and ana-ml2 now runs 0.26.0. Retest before designing around merge; the answer changes what Eitri's harness must emit, and he is still early. Auto-archival moved 5 entries (Recent decisions) to archival-memory.md; the guards held back the rest of the 78 age-eligible candidates because their bodies carry open deferred-work language, per the keep-when-unsure rule. Index sits at 286 lines, above the ~250 target and reported rather than forced. |
||
|
vh
|
5415fd4b30 |
docs(gemma4-charrp): abliteration measured in isolation — close to free, but it MOVES capability
Second bench window, operator-authorised after an initial decline and reversal. Stock BF16 against the llmfan46 abliterated BF16: same precision, same pinned upstream template, same 192 items, CoT off. Abliteration was the only axis that moved, which is what the previous run could not claim. Net core cost is 0.6 points — but the headline understates what happened. Capability MOVED rather than degraded: five items lost on contradiction detection, four gained on spatial composition, nearly cancelling. A gain was not predicted by anyone, least of all on that axis. The decision this was authorised to settle: llmfan46 stands as the trainee base. No case for re-staging on TrevorJS at KL 0.09 over 0.6 points — the KL gap between the builds is smaller than the gap this measurement failed to find. Both limits recorded rather than buried, per brokkr-smithy-dev: the swings are ~5 and ~4 items at n=32, so the -15.6/+12.5 percentages read more precisely than the measurement supports and only marginals were run; and this says nothing about quantization, because the stock-NVFP4 T2 figure came from n=16 against n=32 here — different item counts mean different item sets, so that comparison is n-confounded and is not being made. Turnaround was five minutes rather than fifteen because the gemma4-trainee-bench stack already existed — itself the residue of debugging a 35-restart crash-loop caused by the production compose hardcoding --quantization compressed-tensors. The fix outlasted the incident. gen restored and verified through the gateway; char-rp remains down deliberately; bench stack env reset to the heretic base for the post-tune gate. |
||
|
vh
|
019ccff7e8 |
feat(gemma4-trainee-bench): BF16 bench stack; record that gen's footprint grows with uptime
Adds an ephemeral stack for serving the BF16 trainee base on :8016 under the char-rp aliases, so the abliterated base can be measured on the same battery and the same gateway routes as the served seat with no harness edit. It is a separate stack rather than another variable on gemma4-charrp because that compose hardcodes `--quantization compressed-tensors` for the NVFP4 build. Pointing it at unquantized BF16 weights crash-loops immediately — `TypeError: CompressedTensorsConfig.__init__() missing 3 required positional arguments: 'target_scheme_map', 'ignore', 'quant_format'` — vLLM trying to read a quantization config out of a checkpoint that has none. 35 restarts before it was caught. `restart: "no"` here so a bench seat cannot resurrect itself and block gen's restore, and no homepage labels so it leaves no permanently-offline dashboard card. It cannot coexist with gen and says so: 48.07 GiB of BF16 weights plus gen's footprint exceeds the 94.97 GiB card before any KV cache. Running it means gen is stopped. THE MORE USEFUL FINDING is in the meromero env note: gen's memory footprint GROWS WITH UPTIME. Measured today at 46,726 MiB (45.6 GiB) after ~3 days up, and 39,424 MiB (38.5 GiB) immediately after a restart — same container, same --gpu-memory-utilization 0.43, ~7 GiB apart. That is the missing half of this afternoon's crash-loop: the char-rp seat "fit on the 21st and stopped fitting on the 24th" because nothing about char-rp changed and gen crept up underneath it. Headroom arithmetic done against a long-running gen is measuring a moving number, so the note now says to measure against a freshly-restarted one. Operator's requested end state reached and verified through the gateway: gen and summarizer both 200, char-rp down deliberately to hold GPU0 headroom for the upcoming trainee run, bench seat stopped. |
||
|
vh
|
14ff4a3f57 |
docs(gemma4-charrp): stage two abliterated trainee bases; record the endemic stale-template trap
The operator directed that the ERP/RP trainee base be a low-damage abliterated instruct build rather than the stock checkpoint. Two are now staged under /tank/aimodels/, both BF16, both unquantized, both matching upstream's 51.61 GB / 25.8B shape with only transformers_version differing in config: gemma4-26b-a4b-it-heretic-bf16 llmfan46, Heretic v1.2.0 ARA, KL 0.1237, refusals 3/100 gemma4-26b-a4b-it-abliterated-bf16 TrevorJS, KL 0.09, 1/100 effective and 5/686 cross-dataset "Low damage" was treated as a measurable claim rather than a description: the field spreads from KL 0.09 to 0.4118 and the table is in the README so the next choice is made on numbers. Fleet anchor for reading them — our own abliteration work found Heretic at KL 0.12 preserved the MTP head at 83.7% acceptance, so both staged builds sit inside an already-validated band rather than past it. huihui-ai is rejected despite its reputation: no published metrics, its own card calls the method a crude proof-of-concept, it abliterates both thinking and non-thinking modes, and its parameter count runs ~738M over upstream. The operator's independent read matched. The more durable finding is the chat template. NOT ONE third-party Gemma-4 derivative pulled here ships upstream's — three independent repos carry the identical stale 266-line file (sha 58c66fdee4afa297), llmfan46 carries a third 365-line variant, and only the RedHatAI NVFP4 build matches upstream's 6a1015c47ccfcfa6. It propagated through the ecosystem rather than one packager slipping, and it is now recorded as a class rather than as the single incident that surfaced it during the A16 control staging. That matters twice over and silently both times: serving a mismatched template renders a different prompt, which is why production pins it; and training through `base/chat_template.jinja` means training on a different prompt format than production serves — train/serve skew with no error, presenting as a tuning failure. brokkr-smithy-dev has been warned on the training side while the harness contract is still early enough to amend. |
||
|
vh
|
8d6a9390de |
docs(gemma4-charrp): RETRACT the contradiction-deficit claim — the item was ill-posed
Supersedes what commit
|
||
|
vh
|
3446367d5e |
feat(gemma4-charrp): pin the chat template; A16 control run executed and reverted
The template is now passed explicitly, defaulting to the A4 build's chat_template.jinja. That is a no-op for what is served — the A4 build ships that exact file, byte-identical to upstream google/gemma-4-26B-A4B-it once trailing newlines are normalised — and it permanently closes the class of bug found while staging the control: the A16 build ships a stale 266-line template against upstream's 390, with the thinking path built differently and no `thinking` property in its tokenizer_config response_schema. Serving each build with its own template would have moved a second axis. The control ran on the operator's greenlight and has been reverted. Seat is back on the W4A4 build, healthy, RestartCount 0, both aliases verified through the gateway — char-rp returns content with reasoning_content empty, char-rp-reasoning returns both. Result, since it settles a question this repo's config now encodes: activation precision does NOT explain the contradiction-detection deficit. Contradiction detection moved 12% -> 19% between W4A4 and W4A16, which at n=16 is 2/16 -> 3/16 — one item — against gen's 81% on identical items. Every other task is identical across the two builds and the core difference is 2.6 points carried almost entirely by two single items. brokkr-smithy-dev pre-registered that a null result would be the robust branch, because a hidden third axis would tend to create a delta rather than suppress one, so the conclusion survives the residual doubt neither side could close without a dequantization pass. The practical upshot for future scheme choices: W4A4 costs less on this workload than the caution warranted. The caution was still correct to have. Displaced production for 3.7 seconds of measurement plus two container recreates. The A16 build and the BF16 tuning base both stay on disk with the runbook in the stack README, so re-running is a two-minute flip. |
||
|
vh
|
1bd90eaacc |
docs(gemma4-charrp): stack README — the three model dirs, the A16 control runbook
The stack had no README and now carries three model directories that look interchangeable and are not: the BF16 QLoRA base that cannot be served here, the W4A4 quant that is served, and the W4A16 build that exists solely as an activation-axis control. Writing down which is which, and why, before someone "simplifies" the compose to the BF16 path and rediscovers the OOM. Also captures the A16 control procedure end to end, including the two confounds found while staging it — the two Hub repos named NVFP4A16 that declare 4-bit activations, and the stale chat template the real one ships — and the fact that overriding the template is safe because the tokenizers are identical. Both sides have now cross-checked this: brokkr independently diffed every non-quantization config field of both builds against the upstream BF16 and found only transformers_version differing. Residual risk recorded rather than hidden: config identity is not weight identity and nobody has done a dequantization pass. The Gemma-4 flags are documented as architecture-level rather than checkpoint-level, since that is why they survived the seat swap unchanged, and the enable_thinking:false pin is called out as mandatory rather than stylistic — without it every plain prose turn lands in reasoning_content with a null content and every consumer breaks. Notes the non-termination defect with thinking on (32 of 96 calls truncating at 12k tokens, all 16 constraint items among them, reasoning sound right up to the point it fails to stop) and why VLLM_USE_V2_MODEL_RUNNER=0 is deliberately not applied to a seat whose production mode is thinking-off. No live change: the seat is still serving the W4A4 build. Displacing it for the control run is an operator decision and is still open. |
||
|
vh
|
24e8826219 |
docs(gemma4-charrp): the A16 control needs a chat-template override, not just a path swap
Pre-flighting the staged A16 build before handing it to brokkr-smithy-dev's
battery found a second axis hiding inside what was supposed to be a
single-variable control.
The A16 build ships a STALE chat template. Verified by hash against the upstream
weights on the same disk: google/gemma-4-26B-A4B-it is 390 lines, the RedHatAI
A4 build's is 389 and byte-identical to upstream once trailing newlines are
normalised, and the prithivMLmods A16 build's is 266 and is not. The delta is
not cosmetic — upstream and A4 open the thinking path with
`{%- set enable_thinking = enable_thinking | default(false) -%}` and branch off
it, while the A16 template has no such set and guards with
`enable_thinking is defined and enable_thinking` instead. tokenizer_config.json
corroborates: A4's response_schema carries a `thinking` property, A16's has only
role and content. That build was quantized from an older revision of the
checkpoint.
Served with its own template, the A16 arm would render a different prompt for
identical messages, and a contradiction-detection delta could be attributed to
activation precision when it was the template. That is the same failure class as
the misnamed-A16 repos — a field nobody validated, believed because the name
looked right — one layer further down, and it would have produced a result that
looked like a finding.
Overriding is safe because the tokenizers agree: vocab identical at 262,144
entries, added_tokens identical, so the same template over the same vocab
renders the same token ids. Everything else pre-flights clean — both artifacts
complete with no missing shards, generation_config.json byte-identical.
Seat NOT flipped; displacing production for the bench window is the operator's
call and is still open.
|
||
|
vh
|
f509668e45 |
docs(gemma4-charrp): record the A16 activation-axis control and how to run it
brokkr-smithy-dev's first battery on the new seat scored 12% on contradiction detection with CoT off, against gen's 81%, while state tracking, deterministic constraint following, long-context state and the confabulation control all sat at 100%. That is not general degradation — it is the shape 4-bit input activations produce on the most reasoning-dense task, which is exactly the confound flagged when the W4A4 quant was chosen. They have recorded the finding as CONFOUNDED rather than reporting it as a property of Gemma, and asked for an A16 build to separate the two. No quant run was needed: a genuine A16 build of the same checkpoint already exists on the Hub and is now at /tank/aimodels/gemma4-26b-a4b-it-nvfp4a16 (prithivMLmods, compressed-tensors, nvfp4-pack-quantized, input_activations null, 17 GB). Same weights, same loader as the live seat, one axis moved. ⚠ Two other repos would have answered the question wrongly and the note in the env template says so: bg-digitalservices and ManniX-ITA both publish "Gemma-4-26B-A4B-it-NVFP4A16" whose config declares input_activations num_bits 4 — W4A4 under an A16 name, via modelopt, both 16.46 GB against the real one's 17.93. Using either and seeing T2 stay at 12% would have wrongly convicted the model. Running it is a one-line GEMMA4_MODEL swap plus a recreate, because there is no room for a concurrent seat: GPU0 has 3.5 GiB free with gen and the A4 seat on it, and GPU1's 19.4 GiB against ~18 GiB of weights is the same sub-GiB headroom that crash-looped the predecessor this morning. Port and both aliases are unchanged either way, so no consumer config moves. |
||
|
vh
|
27155c0f3b |
feat(char-rp): swap the seat to the Gemma-4 26B-A4B MoE, NVFP4, same port
Straight-across replacement of the dense G4-MeroMero-v2-31B-NVFP4A16 seat with google/gemma-4-26B-A4B-it on ana-ml2 GPU0. Port, served-model-names and every gateway route are unchanged, so no consumer sees a difference in addressing: `char-rp` -> hosted_vllm/char-rp and `char-rp-reasoning` -> hosted_vllm/char-rp-thinking, both still :8016. The seat's requirements now include chain-of-thought, which makes throughput more critical rather than less — the user waits through the whole reasoning block before the first visible token, and the MoE measures ~114 tok/s @32K against the dense 31B's ~40.7. Both artifacts are on disk and they are NOT interchangeable. The BF16 weights (/tank/aimodels/gemma4-26b-a4b-it-bf16, 49 GB) are the QLoRA tuning base, since QLoRA does its own quantization. They CANNOT be served here: 48.10 GiB of weights against ~49 GiB of free GPU0 leaves nothing for KV cache, and the engine would die at allocation exactly the way the predecessor did this afternoon. The serving copy is RedHatAI/gemma-4-26B-A4B-it-NVFP4 (16 GB), chosen over the other -it quants because it is compressed-tensors (nvfp4-pack-quantized) — the same loader path the outgoing seat used — from the llm-compressor team at 357k downloads. The nvidia/ repo is the base rather than -it, and the thinking channel lives in the instruction-tuned weights. Smaller weights at the same 0.47 memory budget buy a much larger KV pool: 27.37 GiB and 1,724,110 tokens, against the predecessor's 371,023 at the same budget. That is 6.5 full-length 262K sequences concurrent rather than 1.4. The gemma4 tool-call parser, reasoning parser and the enable_thinking:false default all carry over unchanged — they are architecture-level, not checkpoint-level. The --chat-template override does NOT carry over: MeroMero pointed at a jinja hand-patched against that checkpoint, and this model ships its own. Verified that dropping it did not reintroduce the failure that flag existed to prevent — non-thinking prose lands in content with reasoning_content empty, and the thinking alias populates reasoning_content with content carrying the answer. ⚠ Scheme differs from the incumbent and the bench should say so: this quant declares 4-bit input activations (W4A4) where the outgoing seat was NVFP4A16. Faster, and not like-for-like on the activation axis. meromero-charrp is retained stopped in `created` state and relabelled to AI - Dormant, per the house rollback pattern. Both stacks want :8016, so rolling back means stopping the gemma4 seat first. |
||
|
vh
|
850e0c3351 |
fix(meromero-charrp): drop GPU0 budget to 0.47 — the seat was OOM crash-looping
`vllm-meromero-rp` had been restarting since 2026-08-24 18:2x, 13 times by the time it was looked at, taking both the `char-rp` and `char-rp-reasoning` gateway aliases down with it (they resolve to the same seat on :8016 — hosted_vllm/char-rp and hosted_vllm/char-rp-thinking). Root cause is CUDA OOM on ana-ml2 GPU0, which the startup logs hide well: the engine gets through weights, torch.compile and CUDA-graph capture looking entirely healthy, then dies at KV-cache allocation with `torch.OutOfMemoryError: ... 195.19 MiB is free`. GPU0 is shared with `vllm-gen`. gen is configured at 0.43 but actually holds ~45.6 GiB of the 94.97 GiB card, because --gpu-memory-utilization sizes the KV cache and does not account for CUDA context, graphs and non-torch overhead. This seat was at 0.51, so the pair was committed to 0.94 of the card with about 0.6 GiB of real headroom. That fit on 08-21 and stopped fitting today. 0.47 restores ~4.8 GiB of margin and costs nothing usable: KV cache 27.36 -> 23.56 GiB, 430,825 -> 371,023 tokens against a max-model-len of 262,144, so the pool still holds 1.4x a full-length sequence. What is lost is concurrent long requests, not context. Verified through the gateway rather than at the container: char-rp returns 200 with content, char-rp-reasoning returns 200 with both content and reasoning_content populated. Seat is healthy with RestartCount 0. The arithmetic and the "check used_memory, not the flag" warning are written into the env template, because the next person to raise either budget needs to lower the other in the same change. |
||
|
vh
|
35adc4a043 |
feat(homepage): rebuild on Australis Skyfall — dual theme, light mode shipped
The board was on the Australis TERMINAL palette, which is dark-only by design
("Always dark first. No light mode in this system"). Skyfall is the dual-theme
web derivative of the same science, and its bundle turned out to be sitting in
this repo's own git history: a predecessor vendored it on 2026-08-19 and a
later commit deleted it. `git show 45c1995:...` returns colors.css with both
`:root` (dark) and `[data-theme="light"]` (Skyfall Day) intact, plus the
calm-depth layout tokens, the typography scale and Supreme 400/500/700. So the
light ramp is canonical rather than derived, which was the entire objection to
building one.
The visual language moves with the palette. Depth is now the recipe and not a
choice — every elevated surface carries a 1px hairline AND a two-layer shadow,
never one without the other. Radii move to Skyfall's scale, cards at
--radius-lg. Widget stat values move from the display face to mono, because
Skyfall is explicit that numbers and telemetry are always --font-mono. The
full-width aurora ribbon under the tab bar is gone: Skyfall sanctions exactly
two accent expressions, the active rail and hero-only glows, and a decorative
gradient across the chrome is neither — so the colour it carried now lands on
the active tab as a 2px accent bar plus an --accent-soft fill, which is the
rail. Every binding is written against the semantic layer; there are no raw
family tokens and no colour literals left in our own file.
build.py now guards the vendoring instead of advising it. The three token files
are hashed and a mismatch FAILS the build — a vendored file is either
byte-identical to the bundle or it is a fork wearing the bundle's name, and the
theme this one replaces had to be torn out twice for exactly that.
⚠ Homepage's own theme toggle is unreachable, and reaching for it breaks the
dashboard. It renders only when settings.yaml leaves `theme:` unpinned, and
with the key absent the page's data loader throws and its catch branch serves
`initialSettings: {}` — no tab bar, no layout, no i18n. Six force-recreates
over seven minutes all came up empty; restoring `theme: dark` rendered
correctly on the next recreate in 12 seconds, while /api/services returned 200
with fully correct content the whole time. That is the first confirmed cause of
the long-running "tab bar goes missing after a recreate" symptom, and it also
retires the homepage.log-size lead recorded earlier today: rolling the log
aside did nothing during this episode, so that coincidence was intermittency.
So the toggle is ours. conf/custom.js renders it and stores the choice;
build.py re-emits each vendored light block twice, once for an explicit
`data-theme` and once inside a prefers-color-scheme media query scoped to
`html:not([data-theme="dark"]):not([data-theme="light"])` — that :not() pair is
what lets a stored dark choice survive a light-mode OS. Verified against both
OS preferences: load, click, click back, reload, all four correct. `data-theme`
is the control surface; Homepage's own `dark` class stays on <html> and does
not fight, because our rules carry !important on the surfaces Tailwind's
`dark:` variants would otherwise claim.
Two font substitutions, both documented rather than silent: Space Grotesk for
Bespoke Sans and JetBrains Mono for Victor Mono. Only Supreme was ever vendored
here and Skyfall's own notes call Victor Mono user-supplied, so this is a
two-line swap when the real faces arrive.
Dark and light, all four tabs: http://10.100.10.50:8090/b/homepage-skyfall/
|
||
|
vh
|
39da1d4a97 |
feat(homepage): recategorise on "do I open this?", collapse the API groups
The board mixed tools with endpoints. A vLLM seat whose href is a /docs page
sat in the same band as ComfyUI; the MQTT broker and the RustDesk relay, which
have no page at all, sat in Apps; and `Service Networking` was thirteen members
spanning three AdGuards, five Dockges, two Traefiks and four headless agents.
Every group is now one of two kinds and they never mix. TOOLS are expanded and
sit at the top of their tab. ENDPOINTS — an API, a broker, a background agent,
an href that is /docs or /ping or nothing — carry `initiallyCollapsed: true`
and sit at the bottom. Collapsed is not hidden: the eyebrow and its rule still
render, so the tab still says the thing exists and one click expands it.
A second rule fell out of the same pass and now shapes the group boundaries: a
group's members should all carry a widget or none should. A stat strip makes a
card ~50px taller, so one widget card in a row of plain ones opens a void under
the plain ones. That is why AdGuard and Traefik get their own groups rather
than sharing one with Dockge, and it is most of why the old Service Networking
band looked broken. AdGuard (ANA) was the last short card in its row and now
carries the same query/blocked/latency strip as its two siblings — one
infra-ops AdGuard login authenticates against all three instances, verified
against each; it lives in that stack's .env on the host and is vaulted.
The sixteen GPU-backed model seats were deliberately NOT relabelled.
`homepage.group` is read at container creation, so clearer names for
`AI - Inference` and friends would have cost a recreate on six vLLM seats, four
eval seats and four TTS engines — multi-minute model reloads on endpoints peers
reach through the gateway. Order plus `initiallyCollapsed` buys the same
separation for nothing, so those names stay as they are on purpose.
28 containers that ARE cheap to bounce were relabelled, across five hosts, via
rerunnable elway playbooks. Their label steps are gated on the old value still
being present, so a second run reports skipped rather than churning. Two verify
steps were wrong on first contact and are fixed with the reason recorded: the
traefik check raced its own recreate, and asserting a model seat is "running"
cannot answer "did I bounce it" when a seat may be legitimately stopped —
container age can, and now does.
The canonical stacks/ tree was synced to the deployed labels afterwards, so
intent and reality agree again on all fourteen tracked stacks.
Also documents the real nature of the post-recreate blank dashboard, which cost
~25 minutes here and an hour on 2026-08-19. `initialSettings":{}` in the served
HTML is the catch branch of the page's data loader, not a warm-up and not a
cache — and the error can vanish entirely, because the logger is assigned inside
the same try and the catch only logs if the logger exists. Ruled out by
measurement this time: all four API routes return 200 with correct content while
the page serves {}, and the previous known-good settings.yaml reproduces it
identically. The README now carries the one-command test and the next lead.
Before/after, all four tabs: http://10.100.10.50:8090/b/homepage-relayout/
|
||
|
vh
|
f6f2f69649 |
fix(homepage): uniform 4-column grid, hold the status gutter, unleak AI Systems
The board's card width changed at every group boundary because `columns:` was being tuned per group under the 2026-08-18 "columns = member count" rule. That rule is retired: it sets `lg:grid-cols-N` for one group, so it fixes that group's CARD WIDTH, not its density. Measured on the live board, Notes rendered a single 1464px card, News and Media 728px, Eval & Retrieval 286px, everything else 360px. All 20 groups are now `columns: 4` and every card renders at 360. `.service-name` reserved a 78px status gutter with `padding-right` and relied on `overflow: hidden` to hold it, but overflow clips at the PADDING box — the gutter was spill room the title printed straight through. Six cards on the AI tab rendered their name underneath their own status pill, measured by testing the title text node's box against the status cluster's. The intended ellipsis never fired either: it is painted by whichever block's own line overflows, and that is the anonymous box around the bare title text node, which does not carry `overflow`. The gutter holds by wrapping now, and the description opts back out of it with a negative margin since the pill only ever covers the first line. Scriberr's `homepage.group=AI Systems` named a group absent from `layout:`, so it had no `tab:` and rendered on all four tabs — the same defect as UltraSeedbox in 2026-08-18, arriving this time from a container label. Relabelled to `AI - Audio Tools`, where the other ASR seats already live. The `homepage.group=AI Systems` sample in the repo-root CLAUDE.md was the source of the name and now carries the constraint, plus the fact that a labelled container is discovered from any of the five configured engines and must not also be listed in services.yaml. Also: descriptions clamped to three lines so a five-line outlier stops dragging its row 50px taller than its neighbours; icon ramp overridden off slate-400 -> slate-700, which was sinking the bottom half of every glyph into the card fill; bookmark groups and Jellyfin's trailing stream rows brought into the card and eyebrow vocabulary; group gap 10px -> 22px now that width no longer separates them. The icon override sits on `html[class]`, not `:root`: Homepage sets the same variables on `.theme-slate`, which is on the <html> element, and a class beats `:root` on the same element. Verified with Playwright against the live board — per-group card width, card height spread, and a geometric title-vs-status collision check, before and after. Before/after captures: http://10.100.10.50:8090/b/homepage-relayout/ |
||
|
vh
|
32349b7653 |
memory: snapshot — Anaheim tunnels on AES-128, ana-gw admin closed, Scriberr live, ESH DNS fixed
Rewrites the in-flight section, which had gone stale in one place (speaches is stopped, not live) and did not reflect a session in which everything opened was also closed. Three detail files carry the bulk: the FortiGate's public surface reduced to nothing including the ACME listener, Scriberr's deployment and the three upstream defects it required working around, and the ESH DNS repair that also made the IPv6 naming scheme real on three hosts. The tunnel entry in the decisions log described a cipher change as proposed and pending; it has since been attempted, found impossible in the form intended, and resolved differently, so the line now points at the outcome rather than the intention. New decisions cover the cipher adoption and the finding that the per-flow ceiling belongs to the UniFi gateway's software cipher rather than the firewall, the administrative closure, the transcription deployment, the DNS repair, the retirement of an ASR service whose only consumer was abandoned, and a database VM that until now had no fleet identity and no vaulted credential at all. Three approaches are recorded as abandoned: an AEAD cipher the far end cannot express, an upstream DNS setting the gateway accepts and ignores, and a scheme to claim unused delegated prefixes that founders not on the prefixes but on having to rebuild a firewall policy to use them. |
||
|
vh
|
d419b11d43 |
docs(ipv6): close the NH3 multi-prefix question by operator ruling
The gateway has no IP-passthrough mode, confirmed by the operator with admin access to it, so the inexpensive path of letting the UDM take the delegation whole and carve it natively is unavailable at this site. What remains is a separate DHCPv6 client presenting several identities, which needs recabling to reach the gateway, splits routing so that one device handles v4 and another v6, and above all moves IPv6 off the UDM's zone firewall, leaving the entire policy to be rebuilt elsewhere before any host could safely hold a globally routable address. The operator has declined it and the LANs stay without v6. The index entry also still asserted the superseded single-delegation conclusion as fact; it now carries the corrected reading alongside the ruling, so the finding and the decision not to act on it arrive together rather than the first inviting a retry of the second. |
||
|
vh
|
bd209951ac |
docs(ipv6): correct the NH3 delegation finding — eight /64s, not one
The note concluded AT&T delegates a single /64 and that reading the address pattern as a /60 was a mistake. A later finding in the same session overturned that by reading the BGW's own statistics page, and that correction was never written down; it survived only in the transcript and surfaced again while assessing whether more prefixes could be claimed. The gateway holds the /60 and rations it, keeping half for itself and delegating the top eight /64s one at a time. Both observations agree: a prefix-id only carves within a delegation already held, so the earlier test could not have moved a lone /64 regardless. The limit is that UniFi solicits once, which makes the requester the ceiling rather than the carrier. |
||
|
vh
|
d127e29fac |
docs(ipv6): retire the next-candidates line now that all three are done
Replaces it with why the remaining segments have no eligible hosts: two are appliance-only and three have no IPv6 enabled yet, pending the firewall-policy pass that SLAAC on a client segment would require. |
||
|
vh
|
4e83395ddf |
docs(ipv6): all three esh-server hosts now carry the segment name
esh-pve-nas and esh-vm-db join esh-docker-vm on 4411:b105, at :50:55 and :50:60 respectively, each applied by the same prefix-deriving if-up.d hook so the last two groups read straight off the IPv4 address. Two obstacles are recorded because both will recur. The Proxmox node had link-local only despite every relevant sysctl appearing correct, because its bridge carries per-interface forwarding and the kernel ignores router advertisements on a forwarding interface unless accept_ra is explicitly two rather than one. The fix takes the advertised prefix while declining the default route, so the hypervisor gains an address without any change to how it routes; this was verified after applying, with the v6 default route count still at zero. The database VM refuses key authentication for the privileged accounts and its unprivileged login cannot escalate without a password, so the hook went in through the QEMU guest agent from the hypervisor, which executes as root inside the guest. The document notes the base64 indirection needed to get a multi-line script through intact. |
||
|
vh
|
ffb7fba346 |
docs: give the ESH IPv6 naming scheme a home, and make it real on one host
The scheme has existed since August as a single line of persistent memory, which a snapshot then deleted. It is a naming convention rather than temporal state, so it now lives in docs/pfi as a proper document, and the memory entry is reduced to a pointer at it. The document carries the full table, the address structure, the reasoning about which slots can and cannot hold a name, and the recipe for applying one to a host. It also corrects the conclusion the original note ended on. That note held that these names could never appear on the wire, which is true of everything UniFi is able to assign but not of what a host can assign to itself, and the distinction is the whole difference between a joke and an address. AdGuard on esh-docker-vm now holds the esh-server name, at 2607:73c0:402:1d02:4411:b105:50:45, where the segment identity and the IPv4 address are both legible. It is applied by an if-up.d hook that derives the prefix at runtime rather than hardcoding it, backgrounds itself with a retry so it cannot stall interface bring-up, and adds nothing to the existing interface configuration. This is load-bearing rather than decorative. The gateway advertises an IPv6 resolver to clients, macOS prefers it over the IPv4 one, and it previously pointed at an address derived from that host's MAC. |
||
|
vh
|
41091eef8f |
docs: restore the ESH IPv6 naming scheme, dropped by a snapshot
The six ESH LAN hexspeak names were recorded in |
||
|
vh
|
6217d3993e |
fix(scriberr): force uv to copy rather than reflink when building envs
Scriberr builds each model backend's Python environment with uv at container start. uv's default link mode reflinks or hardlinks out of its cache, which fails on this overlayfs over ZFS combination and reports it as a failure to clone a metadata file with errno 11, resource temporarily unavailable. The wording points nowhere near the cause. The damage was partial and therefore easy to miss: WhisperX and PyAnnote came up fine and the application served normally, while the Parakeet and Sortformer backends were quietly absent. Setting UV_LINK_MODE to copy trades a little disk and build time for environments that actually materialise, and the occurrence count for that error is zero on the following start. The knob is exposed through the environment so it can be moved back to the default if a future host does support reflinks. |
||
|
vh
|
efddb4e511 |
feat(scriberr): stand up transcription on ana-ml2, pinned to GPU1
Scriberr transcribes audio and video locally with WhisperX and speaker diarization, and it lands on ana-ml2 rather than ana-docker because the work is GPU-shaped: ana-docker offers eight cores already shared with fifty containers and thirty-seven gigabytes of disk, against ninety-six cores, terabytes on /tank and idle capacity on GPU1. The reservation names device 1 explicitly, since GPU0 is fully committed to the gen seat, and the container is confirmed to see that card alone. The image is built from source, which is not a preference. These are Blackwell cards at sm_120; the published CUDA image covers Pascal through Ada only, and the blackwell image the upstream README documents has never been published at all. The path upstream actually ships for sm_120 is Dockerfile.cuda.12.9, carrying CUDA 12.9 and cu128 torch, so that is what gets built. The compose header says so, because the obvious cleanup is to swap in the published image and that would silently drop the deployment to CPU. Two configuration details are load-bearing and documented where someone would go to change them. The application runs as uid 10001 rather than the usual 1000: that Dockerfile moves its user aside for Ubuntu 24.04's own uid-1000 account and chowns /app accordingly, while the entrypoint's remapping covers only the data directories, so at 1000 the process cannot open its database and restarts forever behind a SQLite error that reads as though the machine were out of memory. Secure cookies stay off while the service is reached over plain HTTP, or sessions are dropped by the browser and login appears to loop for no visible reason. Storage is bind-mounted onto /tank because model weights run to several gigabytes and the root pool on that host is nearly full. Also adds the scriberr service alias to internal DNS, following the existing alias convention so consumers name the service rather than the box. |
||
|
vh
|
22ae9cd480 |
ops(ana-gw): disable ACME and retract the all-port VIP claim
The ACME client is unbound from wan1, which removes the HTTP-01 challenge listener that had been holding port 80 open irrespective of allowaccess. An external sweep of fifty-five ports against the WAN address now finds nothing open at all, while the internal interface still serves the GUI and SSH and retains its certificate, which remains valid until late October and simply stops renewing. The previous note's claim that four virtual IPs were unrestricted all-port static NATs is withdrawn. A FortiOS virtual IP can be scoped either by an explicit forwarded port or by a service binding on the object, and the earlier parse inspected only the former. All four use the latter, and the custom services behind them are narrow. None of the fourteen is unrestricted. Ground truth from outside is recorded in place of further config reading, listing what each public address actually exposes. Three configured mappings answer nothing at all and are noted as tidy-up candidates for the migration rather than as exposure. |
||
|
vh
|
f39b66d2e1 |
ops(ana-gw): port 80 on the WAN address is the FortiOS ACME listener
The previous note attributed the open port to an ISP transparent proxy. That was wrong. It terminates on the FortiGate: system acme is bound to wan1, and FortiOS opens port 80 there to answer HTTP-01 challenges regardless of what allowaccess permits, which is why the port stays open with the interface set to ping only. Every non-challenge request returns a fixed 403 whose body reads ACME Access Only. No DNAT is involved; of the fourteen virtual IPs only two land on that address, neither on port 80. The wrong conclusion came from a sniffer filtered on dst host, which matches inbound packets alone and so excluded the replies being looked for. Filtered bidirectionally the box is plainly seen emitting SYN-ACK on port 80. The note records the rule. Two consequences follow. The earlier warning that certificate renewal would fail without http in allowaccess is retracted, since FortiOS opens the challenge port itself. And the listener is not an administrative surface, though its value is now marginal with WAN administration closed, so the note records how to remove it and leaves that decision open. Also captured: four virtual IPs are all-port static NAT and map every port of their external address. |
||
|
vh
|
7bc9754e40 |
ops: adopt AES-128 on both Anaheim tunnels and close public admin
Both tunnels now negotiate AES-128 for ESP, applied make-before-break so neither dropped waiting on its far end: the FortiGate was widened to accept the new cipher alongside the old one first, then each UniFi gateway was flipped. Single-stream throughput moves from 245 to 270 on the NH3 tunnel and from 268 to 304 on the ESH tunnel. Both network objects were diffed against pre-change snapshots and the only field that moved on either is the ESP cipher. The proposal lists are left accepting AES-256 as well. The peers offer only AES-128 so the extra entries are inert, and retaining them means a gateway reverting cannot strand a tunnel. With that up, the WAN administrative surfaces are closed. The interface is back to permitting only ping, and the infra-ops account is again restricted to RFC1918 space. Ports 443 and 22 were confirmed closed from two separate sites and management over the tunnel still works. The close was issued over the tunnel rather than over the WAN, since withdrawing SSH from the interface while connected through it would sever the session mid-command. The box now has no out-of-band path, which the memory records explicitly. Also captured: the two UniFi vault items have different shapes, one a bare key and one a documentation note requiring extraction, which produces an opaque nginx rejection if missed, and the ESH key's first confirmed write. |
||
|
vh
|
6edebe4864 |
ops: forwarding through a downstream WireGuard terminator is free
FortiOS has no WireGuard, so any WireGuard site-to-site has to terminate behind the edge, and ana-wg already fills that role. The earlier 767 Mbit/s figure was taken with traffic terminating on that box, which left the forwarding case unmeasured and overstated what a real deployment would see. Measured properly, transit from NH3 through the tunnel and onward to ana-docker returns 763.8 Mbit/s on one stream and 790.4 across eight, so the forwarding hop costs nothing and the full threefold gain over IPsec survives. The container uses around a fifth of four cores at that rate, against the UniFi gateway spending a third of its own four to move a third as much. The note records what still needs deciding rather than presenting this as ready to build: Anaheim hosts have to route to the terminator directly or pay a hairpin through the edge, which was not measured; the terminator introduces a failure mode the edge did not have, with a fallback route as the mitigation; and the NH3 end needs its own terminator, where a Linux host matches the measured figure while the existing UniFi WireGuard server would likely land lower. The test tunnel, its peer and the temporary route were all removed. |
||
|
vh
|
062215e81a |
ops: the tunnel's per-flow limit is the UDM's software AES-CBC
Varying only the cipher settles what inference could not. AES-128 has ten rounds against AES-256's fourteen, so a software-bound path must speed up when the cipher gets cheaper while a hardware-offloaded one will not move. Run as A/B/A on a single stream, AES-256 returns 232, 245 and 243 Mbit/s and AES-128 returns 282 and 275, a gain of about a fifth at identical CPU. The FortiGate's offload engine is therefore not the constraint, and the operator's reading of the UniFi side as software crypto without acceleration is correct. Two earlier claims of mine were wrong and are corrected in place. The observation that the UDM sat at seven percent CPU was a sampling artifact, since UniFi refreshes that statistic on the device report interval and the windows used were four seconds; a sustained run reads thirty-five percent, around one and a half of four cores. The per-session offload hypothesis is refuted rather than merely unproven, because it predicts no response to a cipher change. The remaining figures follow from this. One stream is one core, eight streams engage about three, and aggregate stops responding to cipher choice once several cores are working because the path bounds it. CBC chaining is what makes this expensive, as it serialises blocks and prevents the AES instructions from pipelining, which is also why the same gateway manages far more over WireGuard. The cipher was restored and the network object verified unchanged against its pre-test snapshot. |
||
|
vh
|
8ecffa1fab |
ops: WireGuard over the same path carries one stream at 767 Mbit/s
The operator asked for a WireGuard test between ana-wg and NH3 across the public internet, which is the arrangement that separates the path from the crypto since neither gateway performs encryption in it. A single stream returns 767 Mbit/s against 245 for IPsec over the same wire, and eight streams return 763, so one flow already saturates the path and there is no per-flow penalty whatsoever. That settles several things at once. The limit is not the circuit, the NH3 uplink, the ISP or the physical path, all of which carry 767 on one flow. Ranking the implementations, Linux WireGuard shows no per-flow penalty, UniFi's WireGuard shows roughly 1.4x, and IPsec on this pair shows 2.8x. Latency under a single bulk stream tells the same story, rising to 12.7 ms on WireGuard against 102 ms on IPsec. This overturns the earlier conclusion that changing transport was not worth pursuing, which compared eight-stream figures and so understated the gap for single-stream work by a factor of three. A WireGuard site-to-site terminated on ana-wg now looks worth considering, and the note records the open questions around failover and policy. Attribution between the FortiGate and the UniFi gateway remains unresolved, since both perform IPsec in the slow measurements, but the remedy does not depend on which one is responsible. The test tunnel was removed, ana-wg is back to its original three peers and the generated keys were shredded. |
||
|
vh
|
d42e9d8712 |
ops(ana-gw): establish the per-flow cap is IPsec-specific, not capacity
The operator asked whether the 80F is simply out of capacity. It is not. Routing a single flow between two Anaheim VLANs through the same box, with no tunnel involved, sustains 940 Mbit/s, which saturates the link, and eight flows over that path return the same figure. There is therefore no per-session ceiling in the plain forwarding path and the roughly 250 Mbit/s per-flow limit belongs to the IPsec datapath alone. Measuring the second tunnel with the same probe supports this. The ESH side, terminating on different gateway hardware at nearly half the round-trip time, returns 268 Mbit/s on one stream against NH3's 245, and 715 against 692 across eight. A window-limited path would have been substantially faster at the shorter round-trip, so the agreement between two dissimilar far ends is further evidence of a rate cap. This also corrects the previous note, which named the FortiGate on reasoning that could not separate the box from the protocol, since every slow path was both. That separation is still not established: both tunnels terminate on UniFi gateways running common firmware, so the cap could belong to either side. The note records the tunnel topology that would decide it and flags the bearing on the pending replacement. |
||
|
vh
|
cf0cb2cbb3 |
ops(ana-gw): pin down what limits a single stream across the tunnel
The earlier note described the constraint as per-flow serialisation without evidence. Pinning SO_SNDBUF and sweeping it shows what it actually is: throughput holds flat between 224 and 247 Mbit/s across a thirteenfold range of in-flight data while round-trip time scales with the buffer instead, from 7.8 ms up to 107 ms. Retransmissions are absent throughout. That is a fixed service rate with a standing queue ahead of it, so socket buffer, window scaling and congestion control are not the lever and should not be pursued. The same measurements surface something with wider consequences than throughput: a single bulk stream lifts tunnel latency from 6.9 ms to 102 ms average, 136 ms peak, with no loss. Interactive traffic sharing the Anaheim link degrades sharply whenever anything moves bulk data, and parallelising transfers makes that worse rather than better. Localisation rests on the FortiGate appearing in both slow paths and in neither fast one, with aggregate throughput over the same security association reaching 692 Mbit/s. Per-session NPU offload fits the shape but was not confirmed on the device; the note records the test that would settle it and flags the bearing this has on the pending FortiGate replacement. |