v1.34.0 - fboudra's code-quality pass, and the security fix that was never published

Thanks to @fboudra for PR #6: ruff format/lint at line-length 100, pre-commit, a CI checks job, bandit security scanning, and three genuine file-handle leaks closed (json.load(open(f)) in recipes.py leaked a descriptor on every recipe load).

⚠️ Breaking: requires-python is now >=3.11 (was >=3.9). Python 3.10 reaches EOL in October 2026; Debian 12 oldstable already ships 3.11. On an older interpreter pip resolves to v1.33.x.

This release also carries v1.33.0 and v1.33.1, which were changelogged but never tagged or published. Upgrading from 1.32.0 delivers these for the first time:

• Security: a spawned llama-server is pinned to --host 127.0.0.1 instead of inheriting llama.cpp's default. llama.cpp's server starts with CORS * and no API key, so an inherited 0.0.0.0 would put an unauthenticated LLM endpoint on the user's LAN. An explicit --extra "--host 0.0.0.0" still wins.
• probe --eval defaults to auto and fetches WikiText-2 itself, instead of requiring a corpus the README never located.

Full notes: CHANGELOG.md

v1.28.2 - hotfix: 1.28.1 was broken on GPU-less machines

If you installed 1.28.1, upgrade. pip install -U quantprobe

1.28.1's AMD refactor wrapped the non-NVIDIA detection branches in a guard with no else, so a machine with no GPU at all fell through every arm: detect() returned hardware with no vram/vram_bw keys and printed no GPU: line, and None propagated downstream — the same failure mode issue #1 reported. Every CPU-only user of that release was affected; no GPU user could be.

• detect() now always settles vram/vram_bw and always emits exactly one GPU: line. The empty case falls through the same else as every other dead end, which is what the pre-refactor code did.
• The regression test stubs all three probes empty in-process, so the no-GPU path is exercised on GPU machines too — and it was verified against the actual released 1.28.1 file rather than an imitation of the bug: it fails there, passes here.

How it got out. The full suite ran green locally before the release, and this development box has an NVIDIA card, so the broken branch was unreachable here. CI — which runs on GPU-less runners — caught it on four tests within minutes, but the tag and the PyPI upload had already gone out in the same sequence as the push. The rule that follows, and that this release is the first to obey: CI green is a release gate, not a formality that runs after the upload.

None of this is @fboudra's doing — the AMD detection in 1.28.1 is good work and stays exactly as it is. The guard was a merge-time hole that our review and our local suite both missed, and the fix belongs to us.

v1.28.1 - AMD GPU detection (first external contribution)

AMD GPUs are detected, not guessed at — and the patch came from outside the project.

@fboudra built the rocm-smi path end to end and validated it on an RX 9070 XT: PR #4, closing issue #2. First external code contribution quantprobe has received.

• rocm-smi detection (Linux, amdgpu): name, VRAM, sclk/mclk, temperature and the supported-frequency ceiling, through a pure parser that is testable without a GPU — the same convention the Windows registry parser follows. detect() tries NVIDIA first, so no existing machine changes behaviour.
• calibrate works on AMD: gpu_state_amd() returns the identical dict shape, so ClockSampler, the boost verdict and the calibration state id needed no changes at all.
• audit-ollama can verify a clean GPU on AMD, and the stuck-boost advice now names rocm-smi --showclocks alongside the NVIDIA command.
• Five parser tests shipped with the patch; full suite green on the merged tree before merge.

One follow-up of ours, disclosed: the AMD fallback collapsed unload()'s two failure modes into one return, which made the "ollama is still holding the GPU" refusal unreachable — a working nvidia-smi plus a squatting ollama would have reported an unreadable GPU and sent the user to fix the wrong thing. Contention reporting is a measurement-discipline surface here, so the tristate is restored and pinned by a test that fails if the distinction collapses again (mutation-checked).

pip install -U quantprobe

Running an AMD card? quantprobe bench --gguf <model> --contribute would put the first AMD row in the hardware atlas — and a miss on RDNA4 is worth more to us than a hit, because misses are published at the same size as hits.

v1.28.0 - hybrid linear-attention priced correctly, Qwen3.8 in the atlas, Linux CI

The planner now prices hybrid linear-attention models correctly - two days after the class shipped.

Qwen3.8-27B landed 2026-08-14 with 48 of its 64 layers as linear attention; every prior model this tool had met was full-attention throughout.

• KV priced on full-attention layers only (U-51, staked as prereg #101 P-5 before measuring). The old formula assumed every layer caches K+V with position - a measured 4x over-estimate on the hybrid (260 KB/pos read, ~64 real). The count now comes from the file: a block carrying an attn_k projection caches K+V. Full-attention models byte-identical (regression-pinned); hybrids print KV 68 KB/pos (hybrid: 17 of 65 layers cache KV).
• Recipes atlas: Qwen3.8-27B added - the first measured depth-aware band published for it, probed launch-day+1 (fragile band 51-64, back-fragile).
• Linux end-to-end CI - a GPU-less ubuntu runner asserts calibrate exits cleanly with no GPU (the exact issue #2 failure class), then runs a real llama-bench and plan --gguf. Green on first run. AMD/Vulkan detection lands as v1.28.1 with the contributed patch.
• The receipt: predicted 1.80 tok/s for Qwen3.8-27B on a 2016 GTX 1060 before the model generated a token; measured 2.04 +/- 0.02. Both logs committed.
• Also since 1.27: the quant-quality campaign page (docs/QUANT_QUALITY.md) and prereg #102's honest negative (effective rank does NOT predict the fragile band - the probe stays).

pip install -U quantprobe

🤖 Generated with Claude Code

v1.26.2 - install this one

Fixes bench --contribute sending the machine as None under auto-detect - the payload now carries the same resolved hardware the prediction used, guarded with the exact failing shape. Found by a pre-launch gauntlet walking the exact commands a new user runs, on the clean index-installed package. All v1.26 features: see v1.26.1 notes.

v1.26.1 - the batching release

Install: pip install -U quantprobe (1.26.1 supersedes 1.26.0, whose wheel misreports its version string - functionally identical, see CHANGELOG.)

Since 1.25.0, all measured on the reference box and staked before measurement:

• C-06 overturned: the "2x batching ceiling" was our own sweep stopping at 8 streams. A GTX 1060 serves 32 concurrent 7B streams at 219 aggregate tok/s (9.5x single-stream); the jump sits exactly at llama.cpp's mat-vec width cap.
• U-39 confirmed as staked: batching does NOT survive expert offload (2.0x cap) - so the best model inverts with user count: 1 user -> 30B MoE; 32 users -> dense 7B, 219 vs 40 aggregate. plan now prints both halves as placement-gated serving advice.
• X-1: speculation's draft length is a kernel decision - drafts of 4-7 are strictly dominated on pre-Ampere (48-51 tok/s) vs m>=8 (88.5-132.1). Up to 5.8x single-stream on copy-heavy work, byte-identical output. The advisory ships in plan.
• Business validation: the recommended 2.5-bit 30B scored 40/40 machine-checked tasks against a pre-staked bar; a four-model capability ladder (52 executable predicates, T4 designed to fail today's models) ships in the repo.
• A2A, staked: depth-aware vs uniform Q2_K at equal bytes (+0.48%): -13.2% perplexity, -39.5% median KLD, +5.1 pts same-token, +6.6% tok/s - with the speed-invariance miss published as a miss.
• Two corrections of our own published numbers against primary sources (E-12 byte model; the Gemma-26B "C-02 violation" dissolved by its own GGUF header), prereg #92 killed by its own gates, probe workdir/exit-code fixes, 154 tests.

Every claim links to a pre-registration with its verdict - hits and misses at the same size.

v1.24.0 — three defects that were live, and a disk tier we measured and missed

Three defects were live in 1.23.0. One of them could only ever make the tool look cautious.

pip install --upgrade quantprobe

The one worth upgrading for

upgrade_advisor built its counterfactual by copying the baseline and overriding one resource — but it dropped n_layer and true_size_gb on the way in. evaluate() re-derived both from presets, so the thing being priced was not your model with more RAM. It was a different model with more RAM.

The error is one-directional: it could only ever suppress good upgrades, never invent bad ones. So every symptom looked like the tool being conservative. If it told you "more RAM won't help," it may simply have been mis-pricing the question. 31 of 340 grid cells.

Also fixed:

• The speculation block printed a multiplier the row could not reach — 17 of 127 cells, true bound 1.025–1.229× against a printed 2.10×.
• A +73% prefill figure was quoted flat on every row, including all-in-VRAM rows where the same flag measured −39%. And it ignored which ubatch was actually emitted, so a 3 GB card sized to -ub 1024 (measured +38.7%) was told +73%.
• Partial calibration printed the same "calibration applied" line as a complete one — while measuring 12.5% median error against 8.8% for the fully-preset baseline. Worse than no calibration, presented identically.
• The disk probe took one sample and asked you to spot the bad ones. Its docstring said "treat a number above ~1.5 GB/s as evidence of a warm cache." Measured: with 73% of a file warmed, 6 of 8 draws returned RAM speed (max 2.854 GB/s — a 6.3× error), and 1 in 8 even after deliberate eviction. It now takes the minimum over N disjoint regions.

The disk tier: measured for the first time, and we missed

Every previous release validated against 14 ladder rows, all VRAM- or RAM-resident. Zero touched the disk tier. A component no validation row reads can stay wrong indefinitely.

So we staked a prediction and measured it:

predicted	0.332 tok/s
measured	0.476 tok/s
error	−30.4% — outside our own band

The law is pessimistic about streaming from disk. We have not repriced it, because one datapoint is not a coefficient.

What the miss cost to chase, in order: the access-pattern explanation was refuted (scattered 0.452 vs contiguous 0.459 — and the same experiment independently confirmed the corrected disk calibration). The byte deficit is real but over-explains the speed-up by 1.84×, so two errors are partially cancelling. A per-fault latency mechanism was staked, appeared to confirm at r²=0.9999, and was withdrawn — fault counts were derived from byte counts, making "per-fault latency" and "lower throughput" the same equation in different units.

The disk tier ships disclosed as 30% pessimistic rather than quietly fitted.

What is new

• Layer-by-layer streaming placement — one layer resident, the rest streaming, so VRAM scales with layer size instead of model size. Shown but never selected as the winner: its number is an upper bound with two named unpriced terms.
• KL divergence as a quality delta. Perplexity called a Q2_K regression "23% worse." The model picks a different token on 27.4% of positions. llama.cpp was already computing this on the passes we ran; we were discarding it.
• Every cited evidence file is now in the repository, and a gate enforces it. The register cited 109 paths; 37 were missing. For a project claiming every number is re-derivable, an unverifiable citation is the worst defect available — it looks like rigour and isn't.

Refutations this cycle

Eight, five of them our own claims, including two sub-noise-floor "improvements" I announced and retracted, a per-fault mechanism I staked and withdrew, and a "damage hides in the tail" framing the data corrected — the tail wasn't special; KL beats perplexity at every percentile.

Full ledger: FINDINGS.md · preregistrations · CHANGELOG

Still not validated

The disk tier's 30% pessimism, the 1.82× streaming-efficiency gap between llama.cpp and raw reads, expert-usage skew (inferred by elimination, never observed), and one ladder row — gemma4-12B — that returned 13.23 / 12.25 / 15.62 across three passes and should not be cited until diagnosed.

All stated in the changelog at the same size as the wins.

v1.23.0 — the feature we tested and did not ship

A feature we promised, tested, and did not ship - plus the finding that replaced it.

We told an external reporter (E-08, RTX 5070) that this release would turn the all-in-VRAM
FLOOR into a point estimate, using the 61 public benchmarks we had collected. Pre-registration
#72 staked that with their measurement as the holdout. It failed its own gate and does not
ship. On one basis (efficiency = tok/s x file bytes / spec bandwidth) the public corpus spans
0.512-0.812 across 8 GPU architectures - and our 2016 Pascal card measures 0.509-0.577, INSIDE
that range. Their box sits at 0.810: the population maximum. Fitting a constant until their
number landed in band would have been the exact failure C-02 taught us to refuse.

• L-18 (new law): the all-in-VRAM spread is a property of the POPULATION, not of our old
  GPU and not of any architecture. 1.6x wide, irreducible from public data, collapsed only by
  measuring the specific machine. The floor advice now cites those 61 benchmarks as evidence
  and points at calibrate as the concrete fix instead of apologising.
• U-23 (shipped, with its limit stated): --no-mmap is no longer unconditional on expert
  splits. Above 75% of the usable RAM pool the emit keeps mmap so pages stay evictable, and
  says why - naming the +13.7% decode it is giving up. Honest limit, in the register: the gate
  reasons about model share, so it does NOT catch the reporter's own case (their 330 MiB free
  came from other processes and the mapped file). Reading actual available memory is U-24.
• The mark: quantprobe has an icon - old-style pixel art, one 12x12 bitmap that generates
  both the SVG and the terminal banner. Bare quantprobe now prints it (with plain-text
  fallbacks for pipes, NO_COLOR, and legacy Windows consoles).
• Register at 92 entries: E-08 (the first out-of-sample validation on hardware we do not own -
  58.1 predicted vs 54-57 measured), E-09 (rabbit: disk-tier bandwidth-indifference), L-18,
  U-23 closed, U-24 opened.

v1.22.0 — the IQ ladder priced in, and native MTP breaks the MoE-speculation drought

The IQ ladder priced in, and the speculation map's first version qualifier.

• IQ formats are measured, not guessed (prereg #70): FORMAT_EBW gains IQ2_XS 51.1 / IQ3_S
  61.1 / IQ3_XXS 68.3 / IQ4_NL 117.0 - the divide is CODEBOOK vs not, and IQ4_NL's kernel is
  Q4_0-class. The format advice now offers IQ4_NL (+14% over Q4_K_M at the same size class WITH
  imatrix quality) and warns off codebook IQ for VRAM decode on pre-Ampere. The staked split-arm
  retrodiction MISSED and is published with its diagnosis (U-19 refuted as hypothesized; the
  residual is structural, U-20 - which now also carries public out-of-sample support from a
  four-point residency curve on RDNA3).
• Native MTP is the first positive MoE-split speculation (prereg #71): models with a
  trained-in MTP head (Qwen3.6 class) measure +11.4% at --spec-draft-n-max 2 (93.2%
  acceptance) on the expert-split placement where every external draft loses. K MUST stay at 2:
  K=4 measured 0.61x on the same pair - the expert-union tax is alive, the head just out-accepts
  it at short drafts. speculation_advice prints the exception with its limits.
• Register at 86 entries: E-07 (colibri cross-engine confirmations), the 61-entry public
  benchmark corpus + audit script ship in-tree (weights/public_bench_audit.py), and the
  draft-note %% display fix.

v1.21.0 — the dense-split release: +33.5% from a CPU-resident draft, three defects closed

The dense-split release: speculation's best cell shipped as advice, and three registered defects closed.

The headline is pre-registration #69: on a dense model split across GPU/CPU, a small same-family
draft model is the BEST speculation cell this box has measured - +33.5% decode at K=2
(5.54 -> 7.40 tok/s on a 14B, 76% acceptance, novel code) with the draft running on CPU
(-ngld 0), so it costs zero VRAM. The mechanism is the mirror of the MoE union tax: a K+1-token
verify batch reads each CPU-resident layer once, so the CPU share of every token amortizes. The
staked K-shift MISSED and is published as such: every K>=3 lands at or below baseline (llama.cpp's
default 3 is a measured LOSS here), because the CPU draft spends the same RAM bandwidth the
amortization saves. plan now prints the whole recipe on dense-split placements
(dense_draft_note), completing the speculation map: dense split 1.335x > dense all-in-VRAM
1.11x > MoE split 0.74-0.81x (never pays).

Three queued defects closed, each with its test:

• C-11 - the dense split budgeted vc*0.9 with no desktop reserve and no compute buffer, so
  at 16k depth it emitted a config that overcommitted VRAM and measured -58% (driver memory
  fallback). The budget now subtracts both (the reserve, and the #23 measured 0.5874 MiB/ub-token
  slope at the default -ub), counted once, same discipline as the MoE path: a 14B emits 21/48
  layers at ctx 0 and 16/48 at 16384 instead of a flat 28/48. Disclosed cost: ~5 layers more
  conservative at shallow ctx than a config that measured healthy; the cliff config is gone.
• U-17 - the IQ-on-CPU warning is now priced, not prose: every RAM weight read is discounted
  by a per-IQ-byte penalty on the file's measured IQ share. The retrodiction gate rewrote the
  plan: the hypothesized 2.7x would have overshot ~7x; the calibrated e2e value is k = 0.242
  (pure-CPU IQ arm now exact: 11.4 vs 11.44 measured). The split IQ arm improves (28.0 -> 26.9 vs
  23.60) but stays over - its GPU share is still priced at K-quant eta, registered as U-19 with
  V-11's 1.55x pointing the way.
• U-18 - fetch no longer treats "a file with that name exists" as "the download is
  complete": it compares local size against the remote and fails loudly on mismatch, naming the
  collision; --force replaces the file. This is the exact failure that cost pre-registration
  #69 three crashed runs (a June-era 0.5B under the target's filename fed llama-speculative an
  incompatible tokenizer).

Also in this release: the U-16 split-ub gate (split placements emit -ub 1024, recovering the
+29% prefill the flat cap left on the table), run/bench/plan flag unification for --threads and
-ub (no command drops what another prints), and the residency buffer term measured rather than
estimated (0.9 GB = #23 buffer + margin, emitting 15 residents where the sweep measured the
optimum). Pre-registrations #67-#69 and their raw logs ship in-tree.