The NPU prefill's expert arena read every expert of every layer ahead of its
routing. It now reads, ahead of a layer's routing, the experts the previous
graph routed there, and at the routing node whatever the routing adds. The
matmul reads only routed experts, so the output is bit for bit the same. A layer
routing more than --prefill-routed-full (0.85) of its experts gets the next one
read whole; --no-prefill-routed restores whole layers everywhere.
Phone, Hexagon v81 NPU, top-4, same session, every answer identical:
Qwen3.6-35B-A3B Q4_0 7.68 -> 4.16 s, Q4_K_M 9.95 -> 5.37 s, Gemma 4 26B-A4B
Q4_K_M 6.69 -> 3.70 s, Nemotron 3.5 30B-A3B Q4_0 7.42 -> 6.81 s.
Also: --decide-probe (experimental per-decision expert usage and layer-exit
answers), BMOE_DECIDE prefill_dev_* counters, gates G17f/G17g, app 0.28.0.
A session opened with --decide answers which of a list of choices the model would pick, read from
the next-token distribution after one prefill, with no decode. The state after a shared prefix is
kept and restored when the next prefix extends it. Android app: a Choose from options switch.
With --prefill-device, a decision is prefilled by the chat turn's placement rule and keeps no
prefix state: llama.cpp saves a sequence through KV views that do not follow the moved model state
(gate G18g). Also fixes the engine version, stuck at 0.23.0 since 0.24.0. App 0.27.0 (42).
Wide prefill graphs run on the Hexagon NPU through a two-layer arena streamed from flash; decode stays on the CPU. The app offers it in its own NPU section, off by default, Snapdragon only. A missing or unopenable device leaves the run on the CPU. release-apk builds the Hexagon backend and a skel per NPU generation in a separate, secret-free job. Bundles the llama.cpp bump to bmoe/expert-ready-hook-2609 (K-quants on the NPU). App 0.26.0 (41).
nemotron_h_moe is the third expert layout: gate-less. Each expert is
up, ReLU^2, down, so the registry row names ffn_up_exps and
ffn_down_exps and leaves the tail slot empty, as the fused gemma4 row
does. The Mamba2/attention blocks, the shared expert, the optional
latent projections and the MTP block all stay on the resident side of
the seam. No llama.cpp change and no submodule bump: the pinned tree
already builds nemotron_h_moe.
make-tiny-moe.py learns the whole shape in miniature (hybrid stack,
latent projections, biased sigmoid router, shared expert, a trailing
MTP block that is never loaded), and it runs as a third byte-identity
gate. Every identity gate passes on it.
The architecture never puts two MoE blocks next to each other, so the
forward predictors (predict-prefetch, route-ahead, the stale half of
predict-log) have no next layer to target. The gate reads that from the
file and reports those checks N/A instead of failing or passing them
vacuously; an unreadable file keeps them strict.
Ornith-1.5-35B-A3B is qwen35moe and needs no engine change. Both models
join the Android catalog at Q4_K_M; neither has device numbers yet.
Also releases 0.25.0: versionCode 40, versionName 0.25.0, dated changelog.
llama.cpp maps every gguf it loads and keeps the mapping for the model's lifetime. On
Windows that is expensive in a way nothing had attributed: while a section of a file is
alive, NTFS serialises concurrent unbuffered reads on that file, and a lane opened while
the section existed keeps serialising against it after the section is gone. Four I/O lanes
therefore delivered exactly one lane's throughput, which is why lanes and threads have
always measured dead on the desktop host and why the engine read at about a third of what
the drive can serve.
--release-mmap hands the mapping back after load: unmap the file, close its section, reopen
the reader lanes. Both halves are needed. Whether it is safe is decided by looking rather
than by reasoning, the engine asks the OS whether any weight the capture pass observed still
points inside a mapping of the model files, and declines if any does. Off by default,
because the check answers for the pointers the capture saw and for no others.
Host A/B on Qwen3.6-35B-A3B Q4_K_M: 3.16 to 4.63 tok/s (+46%), flash stall per token 0.182
to 0.074, with bytes read, hit rate, evictions and re-reads identical to the digit and the
generated text byte-identical. On the phone the read path is flat (f2fs does not serialise)
but CPU per token falls about 9%; that cell is two short runs per variant and is recorded as
a direction, not a number.
Also fixes a bug this uncovered, independent of the flag: when a gguf carries no
output.weight, llama.cpp builds the output head from the token embedding table and the model
holds two identically named tensors over the same bytes. The capture pass keyed its map by
name, so --dense-weights anon and ahwb rebound one and left the twin reading the mmap for
the whole run, which on a model past RAM means the output projection served by page faults
from flash. The capture now records every distinct leaf object by address and the dense
policy rebinds every tensor over one file range onto the same buffer.
Adds bmoe-iobench --mmap / --reopen-lanes / --range-mb / --fresh, the cells that isolate the
mechanism, a mapping_release unit test on both platforms, the app switch "Release the model
mapping", and the bench findings. README, architecture, AGENTS and roadmap updated, the last
correcting a diagnosis this refutes.
The refreshed clip went in unlabelled, so the frame said nothing about which
model was running while the DeepSeek hero next to it names its own. Same
treatment now: a black band over the app title bar carrying the model and
quantization, "running on a phone, airplane mode" under it, and a second band
over the navigation bar with the real-time note and the run's tok/s.
Typography is measured off the DeepSeek asset rather than guessed - Arial 14 for
the subtitle (the same string lands on the same pixels, x 72 to 287), Arial Bold
16 for the model name, Arial 15 for the caption, and its purple. The palette is
built with stats_mode=full: on static overlay text the diff mode spends no
colours and quantizes white down to 249.
The caption reads "real time" rather than the DeepSeek clip's "from here on -
real time", which was there because the master it was cut from had sped-up
stretches ahead of that point. This clip is real time end to end.
The clip in the README was the UD-IQ3_XXS file at 2.03 tok/s. This is a new
in-app run on the 12 GB test phone: 82 tokens at 3.48 tok/s, real time, with
prefill 6.87 s (3.9 tok/s), 13487 MB streamed, a 1000/1000 MiB expert cache at
56 % hit, 70 major faults per token, pinned dense weights, overlap on, four
lanes, cold experts dropped at 100 %.
The file is the DevQuasar Q2_K build the catalog gained in #189, not the one the
old clip used, so the caption names the quantization and the size follows it:
six shards, 80,447,449,856 bytes. The two builds are the same size on disk to
within 2 %, because the 28.8 GB n-gram table dominates both and stays at IQ4_NL
in each. What differs is the dense side — 2.4 GB pinned against 4.3 — which is
the cache room the app README already describes.
The architecture table carried two stale claims: a single ~2 tok/s figure that
belonged to UD-IQ3_XXS alone, and a note that the model was pinned to an
unmerged upstream PR. Upstream support has been in since b10666.
On a 12 GB phone this model is capped by its dense side: 4.3 GB pinned
and walked every token in the UD-IQ3_XXS, which leaves the expert cache
1000-1500 MiB. Every published dynamic quant keeps that side at 5-8 bit
whatever its overall size. DevQuasar's plain Q2_K is the one file with
the dense side at 2-4 bit (2.4 GB pinned), at the price of coarser
experts (no importance matrix). The entry sits next to the UD file, its
blurb states the trade, and the docs that list the catalog follow.
A first command with nothing but -m, -p and -t ran plain llama.cpp on mmap: streaming
off, cache off, and a dense policy that only applies once streaming is on. Nothing in
the report said so, because the moe-stream: block only prints when streaming is enabled,
so a baseline run read as a measurement of this engine and got reported as one (#186).
A mode: line is now printed on every run. With streaming off on a MoE architecture the
build has a recipe for it says the run is a baseline and names the flag; on any other
model it says the architecture is not one this build streams. RunSummary carries the
model's arch so the CLI can tell those apart.
--cache-mb defaults to auto whenever --moe-stream is on. The previous default of 0 meant
the cache was off, which re-reads every routed expert from flash every token. On a 16 GB
host streaming Qwen3.6-35B-A3B Q4_K_M, 63 tokens, -t 8 --overlap: 1.164 to 2.351 tok/s
and 585 to 238 MiB per token, same output. The default is resolved in the CLI, not in
the library, so an embedder passing 0 still means no cache; an explicit --cache-mb or
BMOE_CACHE_MB still wins, including an explicit 0.
Reported by @eiffel31.
Three lines still described the state before #182 and #184.
- README said macOS is not exercised by CI. It is now compile-checked on every
pull request and release tag, alongside Windows.
- benchmark-method told a Mac reader that o_direct is a no-op there, which would
have them ignore the one field #182 made truthful.
- The platform-caveat paragraph still said macOS does not bypass the page cache.
limitations.md and community-benchmarks.md were already correct; these three were
missed in the same pass.
A direct request on Apple opens normally and applies fcntl(F_NOCACHE, 1) to the descriptor instead of silently returning a buffered fd. F_NOCACHE is a caching hint, not an I/O mode: no alignment contract, no DMA promise, so a direct reader on Apple keeps plain pread semantics (pio::direct_needs_alignment() splits "uncached descriptor" from "alignment-constrained reads") and skips the O_DIRECT bounce path.
Independently, the o_direct telemetry (CSV preamble, decode-trace header, streaming banner) now reports what the shard opens achieved, the AND across shards after every platform refusal and open-time downgrade, instead of the requested configuration, on every platform.
Measured on a 16 GB Apple-silicon Mac, model on an external volume, 256-token protocol, interleaved A B B A A B: decode 0.94 vs 0.61 tok/s (+56%, non-overlapping), byte stream identical between arms, cache-hit equal; the gain is the buffered arm's page-cache pollution doubling the compute residual while stall stays flat.
Addresses the macOS half of #179.
Experimental, off by default. Before a decode routing is committed, every
expert already in the LRU cache gets its score raised by L times the
token's score range and the top-k is taken again, so a near-tie goes to
the expert already in RAM (Skliar et al., arXiv:2412.00099). The same
number of experts runs; fewer are read from flash. Scores are read from
the tensor the graph itself sorted, exact for any gating function.
Desktop, Qwen3.6-35B Q4_K_M at L=0.15: 258 to 119 MiB of flash per
token, 2.37 to 3.84 tok/s, perplexity +1 to 4 %, tinyMMLU 88 to 84/100,
HumanEval-50 42 = 42. The on-device A/B is still owed, hence experimental.
Also: --ppl / --ppl-step / --ppl-list / --ppl-choices (teacher-forced
perplexity, one token per decode so cache-dependent policies are priced
where they act), scripts/tinymmlu-bench.py, scripts/humaneval-bench.py,
gates G8d/G8e, app switch "Prefer cached experts" under Experimental,
docs/cache-aware-substitution.md.
Dense tables the graph only gathers rows from (the token embedding, on
most models) are bound to reserved address space and fetched in 16 KiB
slabs inside a bounded LRU window, instead of being read whole and kept
resident. Which tables qualify is decided from the captured graph, not
from a name list. Byte-identical to the resident reference; -497 MiB
pinned on Qwen3.8-Flash-Next and -515 MiB on Qwen3.6-35B on the 12 GB
test phone, throughput neutral, off by default. Gates G15a/G15b.
App 0.24.0 (unreleased), new page docs/row-gathered-tables.md.
Split the prompt phase into the same wall-additive terms decode already reports:
prefill_cpu_s / prefill_read_mib / prefill_io_s / prefill_stall_s / prefill_mgmt_s,
as session-level deltas of the streamer's cumulative counters across the prefill
chunks. Session layer only; the streamer is untouched. The keys ride both
BMOE_DONE and the CSV `# summary` trailer, appended so existing readers ignore them.
Closes#173.
The benchmark call is pinned, and both pages were written for the maintainer
rather than for the people landing on them. community-benchmarks.md opened with
a list of hardware we want, which reads as an entry requirement, and neither
page ever answered the first question a contributor has: what do I set?
community-benchmarks.md:
- a "start here" for the three cases someone is actually in (PC or laptop,
Android phone, Apple hardware), each with the command and what to paste back
- the settings-override table, which used to be one buried sentence
- an explicit adb protocol for phones
- "hardware we want to see" moved to the end as open questions: any hardware is
a useful row, the list is what we cannot answer ourselves
benchmark-method.md:
- reference device out of the opening, named once at the end as the provenance
of the published numbers
- new "choosing the parameters": lossless, lossy and experimental knobs kept in
separate tables, each with its default and the telemetry field that says
whether moving it worked
- the hard-won rules kept as method rather than as the story of one session
The community protocol now pins --ubatch 512, which the app has always done and
bench-report.sh never did: prefill width costs resident memory the expert cache
would otherwise get, so a host row was running a different configuration from
the app it is compared against. UBATCH= overrides it.
Two platform limits documented for the first time: macOS has no O_DIRECT and
the engine does not call the F_NOCACHE equivalent, so expert reads there go
through the page cache while the metrics still report o_direct=1; and there is
no iOS target at all. Both were already true.
The dense policy assumed the dense set fits in RAM. Qwen3.8-Flash-Next breaks
that with its 51B n-gram table (per_layer_token_embd, ~28.8 GB at IQ4_NL),
which every mode failed on in its own way. A dense tensor larger than the
kernel's MemAvailable is now held back under every mode: it stays mmap'd
with MADV_RANDOM, leaves the warm sweep, the residency sensor and the auto
cache budget, and one stderr line names it. The bound is size, not access
shape: a row-gathered table that fits keeps its mode. Inert on every other
supported model (Qwen3.6-35B matches its baselines to the decimal under
anon, warm, mmap and auto); gates pass; the guard fires on device and pinned
dense weights survive load.
Qwen3.8-Flash-Next (qwen4exp): 125B total, ~6B active, 512 routed experts at
top-10 plus one shared, 48 hybrid gated-delta SSM / sparse attention layers,
and a 51B n-gram embedding table. One registry row streams the experts; a
dense-policy guard keeps the n-gram table (larger than any phone's RAM)
mmap'd under every mode so pinned and anonymous dense weights survive load.
Runs on the 12 GB test phone at ~2 tok/s with pinned dense weights, compute-
bound, and sits in the app catalog as a three-shard download. Submodule
pinned to upstream master b10666, the first with the architecture merged,
with the expert-ready hook on top. README hero clip, changelog and docs
updated. App 0.22.0 (versionCode 37).
A benchmark contributor should be able to download one file and run
scripts/bench-report.sh without a toolchain. release-host builds a static,
portable CLI (GGML_NATIVE=OFF; x86_64 assumes AVX2, aarch64 armv8.2-a+dotprod)
from a clean checkout of the tag for linux-x86_64, linux-aarch64, macos-arm64
and windows-x86_64, and attaches the archives to the release. workflow_dispatch
with upload=false keeps them as artifacts instead, for trying the workflow on a
branch without touching a release.
Every published number comes from one phone and one laptop. This adds what a
contributor needs to add a row from hardware we do not own, without reading code:
- scripts/bench-report.sh runs the fixed README protocol on any Linux/macOS host
(256 greedy tokens, the reference prompt, auto cache, 4 lanes, overlap, dense
weights out of the page cache), records CPU / RAM / drive and the drive's
measured O_DIRECT rate at 512 KiB requests from the model file itself, and
prints the two markdown tables a report needs. Every figure is read from the
CSV `# summary` trailer by key name.
- .github/ISSUE_TEMPLATE/benchmark-report.yml collects hardware, model, engine
version and the pasted tables; tok/s alone is not accepted as a row.
- docs/community-benchmarks.md holds the protocol, the hardware wanted and why,
the reference models in the catalog quants, the meaning of each column, and
the results table seeded with the README rows.
- README, CONTRIBUTING and the docs index point at it.
Overlap stall is the union of stalled intervals, not the per-thread mean; the app panel draws compute / flash wait / cache mgmt / unattributed from measured terms. Fixes#98.
Below one token's worst-case routed bytes, global LRU evicts precisely
what it is about to read: the hit rate is 0% while the run still pays
the cache's management time and its RAM. The only protection so far was
cache_min_mb, a fixed floor that happens to sit above the cycle for the
shipped models at their default top-k and stops holding the moment
--n-expert-used widens the routing without touching the budget.
The cycle is priced at init from the model's shape alone — every bound
layer's entry_bytes times min(top_k, n_expert), at the top-k the run
actually applies — and recorded as cache_cycle_mb in the metrics
preamble next to the budget it should be judged against, so a committed
CSV answers on its own whether its cache could ever have hit. The engine
prints one stderr line when the resolved budget falls under it. It warns
rather than refuses: the budget is legal and the output byte-identical,
and the engine states the fact and leaves the choice — the argument that
--cache-mb 0 is strictly better below the cliff stays in
docs/cache-sizing.md, not in the engine's output.
Docs updated in the same commit: cache-sizing.md (the guard section,
past tense), roadmap.md (the item moves from still-worth-doing to
shipped) and telemetry.md (the new preamble key).
* feat(app): settings grouped by purpose, and four defects an audit found
Settings now show the recommended configuration first and fold everything else
into a collapsed Experimental group per category. That is a statement about
evidence, not about how finished the code is: inside are the levers measured on
one device, measured once, or still owed a measurement. They stay in the
release build, because testing them on other hardware is what this app is for
and a lever nobody can reach is a lever nobody can refute. The caveat is stated
once in the group header instead of leaking into some descriptions and not
others.
Every description was rewritten to say what the setting does for the person
reading it. Out went the measured figures, which need the device, the model and
the day beside them to mean anything and have none of that room under a switch,
and out went the implementation names: O_DIRECT, top-k, dma-buf, mmap and KV
cache are not what someone deciding whether to turn something on needs to know.
The metrics screen keeps the flag names, deliberately: there the reader is
matching the UI against a CSV column and the technical name IS the vocabulary.
Four defects, all found by auditing rather than by anything failing:
The session signature is now derived from the argv instead of being a
hand-written list beside it. Those two had to be kept in step with nothing
enforcing it, and forgetting a field is a silent bug: the setting appears to
change while the engine keeps running the old configuration. Three of four
rebases this week collided on exactly that list.
A malformed end-of-turn summary no longer strands the UI. The whole handler sat
inside a catch with no failure branch, so a parse error left the state in
GENERATING with no turn committed and nothing said. The streamed answer is now
kept, the reason is shown, and the state returns to READY.
MainActivity drops from about 1050 lines to under 700: the model download and
import UI moves to ModelPickerUi.kt, which shares nothing with the chat screen.
No logic moved, only its address.
Dead code removed: a field whose own comment described a use it did not have,
two functions nobody called, and five string resources describing a UI two
rewrites ago.
* docs: record the settings regrouping and the signature fix
Rule 6: the changelog and the docs a change invalidates ship with it. The app README described the settings screen as it was before the regrouping, and explained one experimental lever in terms of predictor accuracy percentages that the UI no longer shows.
* docs: the improved DeepSeek hero recording
* feat(app): keep the flag vocabulary in Settings, and make Experimental read as a boundary
The first pass at rewriting the descriptions went too far: it renamed the controls into consumer phrasing and lost the vocabulary that lets a setting here be matched against the CLI, the CSV preamble and the docs. Labels are back to the flag's own names; the descriptions are shorter than the originals rather than longer, and still carry no measured figures. The Experimental group gets a divider and a tonal bar: collapsed, it is the only thing between the recommended configuration and the levers that can change the reply, so it has to look like a boundary rather than one more row.
The flagship demo is now the 284B model generating on a 12 GB phone at about
1 tok/s, with its own recording, instead of gpt-oss carrying that slot. The
quoted 0.94 tok/s is the app's own reading and the sentence next to it says
what produced it: cold-expert dropping at full strength, which trades quality.
gpt-oss keeps its lossless and knob-on figures one paragraph down, and the
three-model clip stays where it was.
The model is named by its exact release, 0731, everywhere it appears rather
than only in the opening line. Anyone reproducing this needs to know which
DeepSeek V4 Flash it was.
The list of what ctest gates claimed the LRU cache, evictions, overlap,
temporal prefetch, the dense rebind and multi-turn sessions. It omitted
predictive prefetch and the split multi-shard model, both of which are gated,
and it did not mention that a lossy knob only gets machinery gates. Fixed to
match what the suite actually runs today.
Also: docs/ngram.md was missing from the documentation index despite being a
250-line document for a shipped flag, and the methodology caveat carried an
editorialising aside about future storage in the one paragraph that has to
stay dry.
* feat(engine): --route-ahead N — commit decode routing to the N-layers-early prediction
Every prefetch lives under the same ceiling: layer L's routing needs layer L-1's
output, so any predictor working earlier is approximate and every speculated read
can miss. This inverts the bet. The expert selection of decode layer L is REPLACED
by the ranking layer L's own gate matrix produced on the hidden state N layers back
in the same forward pass, so the selection is known N layers early and cannot miss.
With the cache on, those reads are issued the moment the selection is fixed.
Lossy by construction: it changes the output, and roughly a fifth of slots route to
a different expert than the router chose at N=1. Off by default, mutually exclusive
with both prefetchers and with the prediction probe, since each would speculate on a
future this policy has already decided.
Quality is measured rather than assumed: the committed generations match the
baseline on a four-prompt objective battery, on a long essay, and on a second model
of a different generation and quantization; output stays deterministic across
repeated runs, which expert dropping does not. See docs/route-ahead.md.
Squashed from the seventeen commits of exp/route-ahead: the branch predated the
multi-shard and speculative-decoding work, and replaying it commit by commit meant
resolving the same two collisions seventeen times over. The history is preserved on
the pull request; what lands here is what a squash-merge would have produced anyway.
* fix(engine): refuse route-ahead alongside self-speculation, and say why
Running the two together on a real model committed NOTHING: 0 routings taken,
249 passed through. A verify decode is several positions wide and the policy
correctly declines each one. But it still charged for itself — the prediction
GEMVs ran (2.6 ms/token of worker CPU) and its early reads degraded into
ordinary speculation, falling from 100% useful to 81%.
Cost with no commitment is worse than either feature alone, and nothing told
the user. validate() now rejects the pair the way it already rejects
route-ahead beside the two prefetchers, the app stops emitting the flag and
greys the row out, and the config test covers both draft sources.
Making the combination work is a different change: commit the whole verify
batch to one selection. That is written and measured, and it costs draft
acceptance (70% to 53%) while route-ahead alone still won, so the exclusion
is the honest state today rather than a limitation to be worked around.
Found by the desktop smoke run, not by the gates.
* feat(engine): MTP self-speculative decoding for Qwen3.5/3.6 (proposal)
Qwen3.5/3.6 ship a trained multi-token-prediction block inside the gguf. With
--mtp that head drafts --mtp-draft continuation tokens and the target verifies
all of them in one wider decode, confirming the longest prefix whose argmax
equals what the target itself would have produced. Nothing is approximated and
no weight is skipped, so the quality is the full model's — but it is NOT
byte-identical the way --overlap and --prefetch are, and must not be used in a
byte-identity gate: a verify pass evaluates 1+N positions in one batch, and a
batched matmul is not bit-identical to N single-token ones, so a near-tie can
flip. Off by default.
The prize is that a decode's dominant cost, moving the dense weights and the
routed expert slices, is paid once per group instead of once per token. The
counterweight is that the verify positions route independently, so a layer's
read set widens toward N*k wherever adjacent tokens disagree, and the draft
pass routes through the MTP block's own expert layer on top. Measured on the
desktop host (DRAM-bandwidth-bound, model streamed at ~1.4x RAM): +15.1% at
draft 3 with the host's best recipe (7.12 -> 8.19 tok/s), +29% without the
lossy drop knob, acceptance falling from 71% at draft 2 to 52% at draft 4, and
flash bytes per token rising 19.7 -> 33.7 MiB as the widening predicts. Draft 3
is the optimum here; 4 is worse than 2. On a flash-I/O-bound phone that balance
can invert, so the flag ships off pending the device A/B.
The orchestration is llama.cpp's own (common/speculative.h, public headers
only): no fork, no patch, no submodule bump. Self-speculation is one model with
two contexts over it — the target, created with n_rs_seq so a rejected tail is
rewound from a bounded snapshot rather than replayed, and a draft context
created with ctx_type = LLAMA_CONTEXT_TYPE_MTP. The engine builds the draft
context itself rather than through common_speculative_init_from_params because
the eval callback is per-context: the streamer only sees the MTP block's expert
layer if the draft context carries the same cb_eval.
The MTP block is streamed like any other layer. It sits at layer index n_layer,
contiguous with the trunk and using the same tensor naming, so the hook and the
expert source are sized n_layer + n_layer_nextn; left at n_layer its experts
stay silently mmap-resident. Two consequences that are easy to get wrong: the
capture warm-up has to run on the draft context too (the MTP graph is built
nowhere else), and prefill is fed through the driver so the draft context's KV
reaches the last prompt position.
The loop accepts BEFORE catching the draft context up, so the catch-up runs on
the accepted prefix instead of the whole verify batch. Acceptance depends only
on the target's logits, which are already in hand once the decode returns, and
the rejected tail was being computed only to be deleted a few statements later.
The resulting state is identical — the driver seeds from row
min(n_accepted, n_rows-1), the same row under either batch, and the surviving
KV is exactly the range the rollback used to carve out — while skipping
n_draft - n_accepted positions through the MTP block per group. Since that
block carries its own MoE FFN, on a streamed device those are expert reads that
no longer happen. It also removes the draft context's rollback entirely: it is
never given a tail to drop.
Requires an MTP-converted gguf (most quantisations strip the nextn tensors) and
greedy decoding; both are rejected at load with a message rather than silently
ignored, as is a n_ubatch narrower than the verify batch, which would split the
graph back into single-token passes and spend the draft for nothing.
Telemetry: an "mtp:" summary line, an mtp_batch per-token CSV column (a verify
decode's whole cost is charged to its group's first row, the rest carry zeros),
mtp_drafted / mtp_accepted / mtp_decodes in the CSV trailer and in BMOE_DONE,
and mtp / mtp_draft_max in the CSV preamble. The Android app exposes the flag
and the draft width, off by default.
Host gates pass. Validated on Qwen3.6-35B-A3B-MXFP4 with the streamed recipe:
draft 1 and draft 3 produce identical text, which is the invariant a broken
accept/rollback path would violate. Device A/B still owed.
* perf(mtp): shrink the draft context, make its cost measurable, record the device verdict
The first on-device A/B says MTP loses at every draft width, and the counters
say why. Same gguf with the flag on and off, shipping recipe (overlap, 3000 MiB
cache, pinned dense, drop 0.75), Qwen3.6-35B-A3B-Q4_K_M streamed:
off 5.82 / 6.14 tok/s 69.3 MiB/tok 69-109 majflt/tok
--mtp-draft 2 5.59 93.8 230
--mtp-draft 3 4.38 106.6 633
Speculation is working - 2.35-2.52 tokens per verify decode, 52-69% acceptance
- and still losing, because the prize does not exist in this regime.
stall_s/tok is 0.025-0.027 in every one of those runs, MTP on or off: 11-16% of
the token. This configuration is compute-bound, and what MTP amortises is weight
movement. The costs meanwhile are real and monotonic in the draft width: the read
set widens (+35%, +54% flash bytes per token), CPU per token rises (+28%, +67%),
and the draft context's memory tips the device into a fault storm.
Two things follow, and both are engine bugs rather than facts of nature.
The draft context's graph width drops from 256 to 32. Compute buffers are
reserved for the widest ubatch and the dominant term scales with
ubatch x vocabulary; on device that reservation measured 493 MiB - for a context
that evaluates ONE token per draft step and is handed at most 1 + draft_max
positions by the catch-up, with no logits asked for. Only prefill ever feeds it a
wide batch, and that is one layer, so splitting it costs very little. On this
engine memory is never free: it is the expert cache's, and the cache is what
decides whether the widened verify read set is a hit or a flash read.
And the cost of speculation is now measured instead of inferred. Drafting happens
between decodes, so it never entered wall_ms and tok/s never included it - a
speculated run could report a rate the user was not experiencing. New
mtp_draft_ms per-token column (a slice of loop_overhead_ms, not an addition),
mtp_draft_s/tok in the CSV trailer, mtp_draft_s_tok and loop_overhead_s_tok in
BMOE_DONE, and a second "mtp:" summary line printing the effective rate next to
the reported one.
Adds --mtp-p-min F, which stops drafting once the head's confidence in what it is
proposing falls below F. The draft loop already had this floor and the engine was
passing 0, so it always drafted the full width however unsure the head was - with
roughly half the drafts rejected at draft 3, that is the cheapest waste available
to cut. On a streamed device it pays twice: a draft not made is a pass through the
MTP block (which carries its own MoE FFN, so its own expert reads) that never
happens, AND one fewer independently routed position in the verify batch. Default
0, the setting the host numbers were measured at; the useful value is a property
of a device's balance between drafting cost and acceptance, so it is a knob to
measure rather than a constant to guess.
The Android app now reads the mtp_* keys it was already being sent: acceptance,
tokens per pass, and the effective rate. Before this the UI could not tell whether
speculation had run at all - only the session CSV could - which made the A/B this
commit reports impossible to run from the phone.
Neither mitigation changes the regime. The honest expectation is nearer
break-even, not a win, and the flag stays off by default.
Host gates pass. Note the noise floor: the two off runs did byte-identical work
and still differ by 5.6% in tok/s, and the runs were back-to-back without thermal
gating - the mechanism counters are the trustworthy part, not the exact deltas.
* perf(mtp): split the drafting flash cost from the widened verify batch
A speculated run streams more bytes per token for two unrelated reasons: the
MTP block carries its own MoE FFN, so every draft pass routes experts of its
own, and the verify batch widens the trunk's read set wherever adjacent
positions disagree. They need opposite fixes -- a narrower draft attacks the
first, only better agreement attacks the second -- and the route trace can
separate neither, since its framing brackets the target decode while the head
only ever runs in the draft context.
Measure the head's share directly by bracketing both drafting passes with the
expert source's byte counter, and report it as a third mtp: summary line.
Also record the branch-deletion rule in AGENTS.md: a branch list should only
show work in flight, and a rejected PR loses nothing.
* feat(engine): n-gram prompt-lookup draft source, and the measurement that closes it
The flash split added last commit said where MTP's cost actually is: at draft 3 on
the host, the head's own routing was 2.9% of the extra bytes a speculated run
streams and the widened verify batch was the other 97.1%. So a cheaper draft
producer is worth almost nothing, and the only property that could matter is one
the head does not have -- the ability to decline to draft at zero cost.
--ngram is that source. It takes the last few tokens, finds where that run occurred
before in the prompt or in what has been generated, and proposes whatever followed.
No head, no draft context, no decode, no expert read, and it works on any gguf
including the ones --mtp refuses for want of a nextn block. Below --ngram-min-match
it proposes nothing and the step falls through to a plain single-token decode.
Measured on the host, Qwen3.6-35B-A3B-MXFP4 streamed, 256 greedy tokens, cells
back-to-back with off run twice:
prose off 5.80 / 6.59 mtp3 7.32 eff ngram3 6.51 (cov 7.4%)
copy-heavy off 5.45 / 5.65 mtp3 6.43 eff ngram3 5.24 (cov 15%)
The zero-cost claim holds exactly -- mtp_draft_s/tok reads 0.0000 in every n-gram
cell, against 0.020-0.023 for the head plus the ~500 MiB of expert cache its draft
context takes. But the floor turns out to be per STEP, not per run: the 15% of steps
that did draft widened the read set to 67.2 MiB/token against 48-58 at baseline and,
at 44% acceptance, bought 1.20 tokens per decode. That is not enough to earn the
widening back, and a modest fraction of such steps sinks the run.
A --ngram-min-match sweep settles it rather than leaving it open. Raising the gate
3 -> 5 -> 8 lifts acceptance 44% -> 75% while coverage collapses 15% -> 3.4%, and
narrowing to --draft 1 reaches 82.6% -- the head's own figure on this prompt. Every
cell climbs toward baseline from BELOW and none crosses it; the best configuration
found lands on the floor. A knob whose optimum is its own disablement is not a
tuning problem. Acceptance, not drafting cost, is what pays for a widened batch, and
what a trained head buys is being right often enough to justify a batch that has
already been widened.
--ngram ships off. It is kept because it is the only speculation available on a
model with no head, because the per-step floor is real, and because the counters it
adds make the next speculation claim falsifiable.
Wiring. MtpConfig became SpecConfig with DraftSource {none, mtp, ngram}, and
--mtp-draft became --draft: the width belongs to the verify batch, not to whoever
filled it. --mtp and --ngram are rejected together rather than resolved by flag
order. In the session the gate split in two -- spec_on (wide batch, acceptance,
rollback: both sources) against mtp_on (draft context, common/speculative.h, the
catch-up: the head only) -- which is what lets the n-gram source reuse the whole
verify half while allocating nothing.
A step that drafts nothing now takes the plain path: llama_batch_get_one with a
logits row of -1, byte for byte the unspeculated decode. It used to build the wide
batch anyway. Required for --ngram, and it tightens --mtp-p-min's zero-draft steps
for free.
The matcher is pure policy over token ids with no llama.cpp at all -- not even
llama.h, since llama_token is int32_t -- so it sits on the clean side of the seam,
adds no dependency on the common layer, and is unit-tested with no model
(tests/ngram_test.cpp covers tie-breaks, clipping, self-match exclusion and the gate
boundary). Telemetry: spec= / spec_draft_max= / ngram_min_match= in the CSV
preamble, a new drafted_steps key in the trailer and BMOE_DONE, and an ngram: line
reporting coverage -- without which a delta cannot be divided by the fraction of the
run it applies to. The per-token and trailer counters keep their mtp_ names: they
always described the loop rather than a source, spec_* already means the temporal
prefetch in that trailer, and renaming would break every CSV already holding a
measurement. The Android setting became a three-way picker, migrating the old
boolean preference.
The device A/B agrees and adds a cost the host could not show. Thermally gated cells
(a 120 s settle, then a battery-temperature gate, so all six start between 35.3 and
36.4 C): prose 4.90 inside a 4.59-5.17 band, copy-heavy 3.14 against 4.43 -- a 29%
loss, worse than MTP's 18%. Major faults per token go 126 -> 1427 for a source that
allocates no draft context at all, and that is the rollback snapshots: n_rs_seq =
draft_max is asked for by ANY speculation, since rejecting a draft means rewinding the
KV, and on a hybrid attention/SSM model that snapshot is a real allocation scaling with
the context. The n-gram source escapes MTP's draft context but not the loop's own
memory, and on device that memory is the expert cache's.
The same run re-measured MTP with the thermal confound removed -- 3.64 effective
against 4.43, so the earlier device verdict was not an artefact of benching without a
cooldown gate -- and reproduced the flash split at 3.7% head against 96.3% widened
verify batch, matching the host's 2.9-3.0%.
Byte-identity gates pass; speculation stays out of them for the reason docs/mtp.md
gives.
The app's CSV configuration surface follows: the three new preamble keys get their own
glossary entries rather than falling through to the unexplained-key renderer, and the
draft source joins the short run label. A speculated run is not the same KIND of run --
under speculation a decode confirms a whole group, so its per-token rows are not even
accounted the same way -- and two compare legends differing by it must not read alike.
* feat(moe): stream split multi-shard ggufs natively + DeepSeek V4 Flash recipe
Hugging Face rejects single files above 50 GB, so every large model ships as
-00001-of-0000N.gguf shards; until now the streamer assumed one file, forcing
a merge with double the disk. gguf_offsets now fans the first shard out to the
whole set and resolves every tensor to (shard, offset); the expert streamer
and the dense loader open one positioned reader per shard and route each read
by the tensor's shard index. Pass the first shard, exactly as llama.cpp takes
it; a missing sibling fails the load with the shard named.
Add the deepseek4 recipe row: V3.2-style routing (256 routed experts, a
per-expert bias like lfm2moe, an always-on shared expert that stays resident)
over the standard split expert suffixes. The V4 compressed-attention machinery
is dense-side llama.cpp code, invisible to the streaming seam.
The byte-identity gates gain a 4-shard qwen3moe fixture (metadata-only first
shard, the layout large quants actually use); make-tiny-moe.py learns
--split-max-tensors. All gates pass, split included.
* fix(moe): cache auto must budget for the anon dense conversion
The auto budget read MemAvailable while the dense weights were still reclaimable
page cache, then dense-weights=anon converted them into buffers the kernel cannot
take back: the same bytes planned twice. Latent since the anon policy shipped
(dense sets were 2-3 GiB and explicit budgets were the benched path); DeepSeek V4
Flash's 6.5 GiB dense set turned it into a device-taking overcommit on first load.
The budget now deducts the pending conversion and says so in the log.
* fix(moe): review pass on the multi-shard path
Three defects the split rewrite introduced, none of which the gates could see:
- The shard index rode in an int8_t, so a model past 127 shards wrapped to a
negative index into the reader vector. The bounds check could never catch it:
it validated the untruncated value. Widened to int16_t, which covers the whole
-%05d-of-%05d filename space.
- DenseWeights::warm() reused one flag as both the inner loop condition and the
partial-warm report, so the first shard that failed to open silently skipped
the warm-up of every later shard. Per-shard condition, sticky report.
- The dense readers stayed allocated for the session after read_anonymous had
copied and rebound every tensor: fds and a per-lane bounce buffer per shard,
sitting next to a cache counting every MiB. Released at the end of init.
Also: the streaming banner read O_DIRECT off shard 0, which under the
small-first-shard layout is metadata only and too short to verify, so it could
claim a mode the shards carrying experts had not got. It now reports the weakest
of the readers.
* build: the engine version says 0.19.0, like the changelog does
The version is declared in CMakeLists.txt and reported by `--version` and by the
run-parameter preamble of every metrics CSV, so a committed benchmark file names
the engine that produced it. This release section landed while the number stayed
at 0.18.0, which would have stamped the wrong engine on every CSV this branch
produces, defeating the one purpose the string has.
* docs(readme): reposition around large MoE models in general, add logo and TOC
The front page led with gpt-oss-120b and read as a single-stunt repo. It now
leads with what the engine is for: running MoE models past RAM on phones and
PCs, on llama.cpp's public API, with every llama.cpp quantization coming for
free. The 120B stays as the most extreme proof, not the pitch.
Structure follows the usual professional layout: logo (chip-buddy, light and
dark variants under docs/assets/logo/), badges, table of contents, a "Why this
exists" section that covers all three regimes (far past RAM, just past RAM,
and models that barely fit, which streaming keeps inside a chosen budget),
and Features grouped the way the app's Settings groups them. Benchmark tables
and the evidence sections are unchanged in substance; the test device is
phrased generically.
* docs(agents): make AGENTS.md the canonical agent guide, CLAUDE.md points to it
The agent guide lived in CLAUDE.md with AGENTS.md as a stub pointing at it,
which is backwards: AGENTS.md is the cross-tool convention (Codex, Copilot,
Cursor and others read it), CLAUDE.md is one tool's name for the same file.
The full guide now lives in AGENTS.md and CLAUDE.md is a one-line import.
While moving it, the guide gains the release rules that were only tribal
knowledge: release APKs come from the release-apk workflow, never a local
build; every released feature bumps versionCode/versionName in the same PR;
release titles are the bare version; on-device validation before a release.
The style section now names the CI clang-format version (18), since a
mismatched local formatter passes locally and fails the check.
* docs(readme): state per-platform host status honestly
Linux is CI-verified, Windows is where the desktop bench ran (but needs CMake
directly, not the bash script, and MSVC's Release output path), macOS builds
from the same sources but is unvalidated and has no O_DIRECT.
A cold layer's batch became visible to the I/O lanes only after every
miss took its page commits - up to three vm_commit syscalls per cold
expert of bookkeeping sitting in front of the first byte of I/O, which
is the latency-to-first-slice the sidecar refutation (PR #90) identified
as the binding constraint. (#118)
With the flag on, only the first present projection - the one
mul_mat_id blocks on first - is committed up front; its jobs publish
and wake the lanes immediately, and the remaining projections are
committed and appended while the lanes already read.
The drain protocol grew the one thing this needs:
- io_drain copies each job out under the lock, so jobs_ growing (and
possibly reallocating) mid-batch cannot leave a worker holding a
dangling reference;
- the worker wait predicate admits next_idx_ < batch_njobs_, so a
worker that drained wave one and left comes back for a batch that
grew in the SAME generation - the gen comparison alone never would;
- a wave-two commit failure goes fatal and wakes the ready waiters,
because wave one already published flags this batch will never flip.
Batch completion cannot fire between the waves: the only thread that
waits on done_cnt_ == batch_njobs_ is the eval thread, and it is the
one appending wave two.
Overlap + LRU cache only (validate() enforces both); recorded in the
CSV preamble as io_two_wave. Default off: the win is bounded by the
commit cost per cold layer, and the failure mode of a drain-protocol
bug is a hang the host cannot reproduce - so a new gate (G4d) holds
two-wave output byte-identical to serial streaming, and the flag stays
off until the on-device A/B (#120) says the win is real.
Every per-token line repeated the whole answer and reasoning so far, so
a generation of n tokens wrote, JSON-escaped and made the app parse
O(n^2) bytes - megabytes of pipe traffic to deliver a few kilobytes of
text on a reasoning model (#119).
The line now carries delta_reasoning/delta_text - the tail since the
previous line - and the reader appends. A pure append-only protocol
cannot express the one thing common_chat_parse does retroactively:
when a closing think tag arrives, text already reported as answer
becomes reasoning. That case falls back to a full snapshot with
"reset":1, and the reader replaces instead of appending.
Both emitters (one-shot --progress and --session) share the single
format string, so they changed together; the app's TelemetryParser
accumulates in StringBuilders (appending to a String re-copied the
whole answer per token) and resets them with the generation. The full
final text still travels in BMOE_DONE, untouched.
The engine-side re-parse per token remains (common_chat_parse cannot
resume); this removes the pipe, escape and app-parse cost, which is
what loop_overhead_ms can now see. On-device numbers are part of the
#120 A/B.
Closes#119.
The audit flagged the bounce buffer as the largest avoidable CPU cost on the
read path: every streamed byte is pulled into a per-lane staging buffer and
then memcpy'd to its cache slot, so the CPU touches 100-200 MB per token twice,
on the I/O lanes, alongside a decode that is already compute-bound.
Skipping the copy needs the read to be block-aligned at offset, length AND
destination. Measured against the three shipped models: bytes-per-expert IS
4096-aligned, but the tensor's file offset is not — 512, 1152 and 2272 bytes
past a block on Qwen3-30B, Qwen3.6-35B and gpt-oss-120b respectively, because
gguf aligns tensor data to general.alignment (32), not to a device block. So an
O_DIRECT window always rounds outward and always spills one block into the
neighbouring experts' bytes.
That spill writes the file's own bytes into buffer positions that hold exactly
those bytes, so it is content-identical. It is still not safe: the pages belong
to other cache entries, which may have been evicted, and touching them faults a
page back with nothing accounting for it — the residency budget stops matching
what is resident, and that budget is what keeps a >RAM model running. Offsetting
each buffer's base by file_off % align, the trick that makes the destination
inherit the file's misalignment, shrinks the spill but does not remove it.
An implementation of the safe subset — take the direct path only when all three
are already aligned — was written and thrown away: it is dead code on every
model in the catalog, and a branch on the hot read path that never fires, with a
startup line that would always report the fallback, is worse than the honest
note.
No code change; the finding is the deliverable.
Two gaps, both about a benchmark file being unable to explain itself.
The CSV preamble had eighteen keys and had fallen behind several releases.
Missing: n_ubatch, which sets the compute-buffer reservation and therefore
moves the very memory columns printed underneath it; predict_log,
predict_spec_max and prefetch_sync, so a probed run was indistinguishable from
a benchmark run the docs explicitly say it is not; drop_renorm and
drop_prefill, which change how much mass dropping discards; cache_floor_mb, the
input behind an auto-sized cache_mb; load_all, whose read_bytes mean something
else entirely; every sampling parameter, so a stochastic run read as a greedy
one; and compute_trace_layers. All are recorded now, under a bumped
"# bmoe_metrics v2" banner.
`think` stays out on purpose: it is a property of a request, not of a session,
so one value in a session-wide preamble would be wrong for every turn that
asked for the other.
The file also now names the build that wrote it. There was no version string
anywhere in the engine — the CMake project had none — so a committed CSV could
only be dated by the commit that copied it in. Added as a project VERSION, a
BMOE_VERSION define, bmoe::version(), `bmoe-cli --version`, and engine= in the
preamble. It sits on its own line so the model= line still BEGINS with model=,
which is how the app's CSV reader finds a run's name; the app's lookup is made
order-independent too, so the next key to be appended cannot break it again.
Second gap: wall_ms brackets llama_decode and nothing else. That is what makes
compute_ms a clean residual, and it also means sampling, detokenization,
rendering and the sink writes are outside wall_ms, outside gen_seconds and
outside the reported tok/s. Work moved into or out of that region was
unmeasurable by the number the project optimizes. loop_overhead_ms now reports
it per token, and loop_overhead_s/tok in the summary closes the accounting with
the tail after the last token that no row can carry.
docs/telemetry.md documents the preamble — it specified every other # block but
not this one — the new column, and the stall_ms divisor: stall is a per-thread
mean, so a stall that is not simultaneous across threads is under-stated and
compute_ms absorbs the difference.
Gates 7/7; preamble, column and --version verified against the tiny gate model.
Both readers checked: scripts/bench-analyze.py reads columns by name and skips
unknown # lines; the app's Csv.read keeps unknown keys and picks the new column
up automatically.
Producing TokenMetrics::text means parsing the entire generation so far —
common_chat_parse takes the whole string and cannot resume — so the cost is
O(n) per token and O(n-squared) across a turn, growing as the answer does. Off
the chat path it was not free either: shown_view returns the raw string, so
every token copied everything generated so far into the metrics struct.
It ran for every token of every run. The plain CLI path writes m.piece and
never looks at m.text; a benchmark run reads neither. Only the BMOE_PROGRESS
line protocol actually renders it.
GenerateRequest::render_text now carries that decision. The session loop sets
it (the protocol puts the parsed answer on every line); run() ties it to
--progress; it defaults to true so an embedder that has never seen the flag
behaves exactly as before. Note this cost sat OUTSIDE the wall bracket that
feeds gen_seconds, so it never showed up in the reported tok/s while being paid
on every benchmark run.
The end-of-turn parse was also done twice — once for RunResult::generated_text,
once to commit the assistant message to history — over the same string with the
same parameters. Parse once, use twice; the prefilled-turn and parse-failure
fallbacks are unchanged and now stated in one place.
Also: three back-to-back stats() snapshots to read three fields become one, and
--compute-trace / --io-trace are actually passed to the session loop. The CLI
parsed those flags, opened the files, wrote their headers and then dropped both
sinks when entering --session, so the user got an empty trace and no
explanation.
Gates 7/7. Both CLI output paths verified by hand on the tiny gate model:
--progress still accumulates the parsed text per line, the plain path still
streams pieces.
* feat(moe): --predict-log, measure how predictable expert routing is
Temporal prefetch was built on a predictor nobody had priced. The
previous-token bet turned out to be right ~38% of the time on Qwen and
~18% on gpt-oss, which cannot pay for the reads it speculates. The
lesson was not that prefetch is impossible but that a predictor should
be measured before it is wired into anything. This adds the instrument,
not a policy.
--predict-log ranks each layer's experts a layer early, by running the
NEXT layer's router matrix on the CURRENT layer's gate input. The
residual stream barely moves between layers, so the stale input ranks
nearly as the real one will -- the mechanism FATE (arXiv 2502.12224)
reports 78.8% for. It is training-free and changes no model: the router
matrix is a dense weight already resident, and the prediction is one
GEMV per layer. The matrix is learned from the graph (the gate matmul's
first source) rather than looked up by tensor name, so no architecture
is named anywhere in the path; being a weight leaf, its pointer stays
valid into layers the current token has not reached.
Three predictors are scored against the routing the router actually
produced, so they are comparable on one run: the stale gate, the
previous-token bet --prefetch already places, and a zero-staleness
control. The control is the load-bearing part. It shares every line of
code with the prediction under test and differs only in using the
layer's own matrix, so it must reproduce the selection llama.cpp
computes from those same two tensors. A transposed matrix, a mis-strided
row or the wrong token of the batch collapses it toward chance while the
stale figure would stay superficially plausible; an architecture that
selects by something other than raw-logit ranking (an additive bias,
group-limited routing) puts it below 100% and by that much the stale
figure understates the method. The CLI says so rather than letting the
gap be blamed on staleness.
Reported per layer as well as in aggregate, because an aggregate
flatters a prefetch: what a prefetch costs is set by the layers it gets
wrong, and a MoE model's first layers route far less predictably than
its last. Denominators are printed per predictor -- the stale gate
structurally cannot rank layer 0 (nothing precedes it) or the first
token of a run, and those routings are counted as unscored rather than
folded in, since a routing that was not ranked is not a wrong guess. A
predictor with no routings at a layer prints "-", never 0.0.
Diagnostics only: nothing it computes reaches load_layer, the cache or
the graph, so a probed run reads exactly the bytes an unprobed one does.
G9a gates that byte identity and G9b gates the control, which reads
100.0% on the tiny model. It is not free -- one isolated node and two
GEMVs per layer on the eval thread -- so a probed run is not a benchmark
run, and it requires --moe-stream since routing does not depend on how
the weights reached memory.
docs/expert-prediction.md also records the caveat the number will need:
a high score would say the routing is knowable earlier, not that knowing
it earlier makes decode faster. On a flash already saturated, starting a
read sooner adds no bandwidth -- which is why prefetch, layer-LFU and
the expert sidecar all lost despite improving the metric each was
designed around.
* feat(moe): --predict-prefetch, speculate on the stale-gate prediction
The probe said the routing is knowable a layer early (~89% of routed
slots on a 128-expert model, vs ~43% for the previous-token bet the
temporal prefetch acts on). This wires that prediction into the existing
speculative read path -- same cache buffers, same accounting, same
settle, same moe-prefetch summary line (tagged [stale-gate]) -- so the
only thing that changes is which guess rides the idle lanes.
Two decisions carry the design:
Speculation is issued AFTER the current layer's load, not at prediction
time. Every load path begins by quiescing speculation, and a layer's own
load sits a few graph nodes after its gate matmul -- reads queued at
prediction time would be cancelled before a lane picked them up. Each of
the three load sites (plain topk, deferred drop, drop fallback) issues
the pending next-layer prediction right after its load_layer, restoring
the same read-ahead window the temporal prefetch gets. On the tiny-model
gates this is the difference between 0% and 27% of speculated experts
proving useful -- the latter matching the probe's measured accuracy on
that model, which is the accounting agreeing with itself.
It is drop-aware. With --drop-cold-experts armed, a predicted expert
whose predicted routing weight (softmax over the predicted top-k)
falls below the drop threshold is not speculated: if it misses, the
policy discards it unread, so reading it ahead would spend the exact
I/O the policy exists to save. The top prediction is always kept,
mirroring the policy's own pin of the top-weighted expert. The known
interplay is inherited from the temporal prefetch and deliberate: a
correct guess un-drops an expert, buying quality at the same threshold
rather than speed.
The routing width the prefetch predicts at is learned from the topk
node, not read from config, so an --n-expert-used override stays honest
with no extra plumbing. Mutually exclusive with --prefetch (two
predictors would double-speculate the same future); requires the LRU
cache; decode only. The control GEMV remains probe-only, so the
production path costs one gate GEMV per MoE layer per token.
Gates: G10a proves byte-identity through the speculative path
(prefetch-sync, forced small cache, hits and evictions both occur);
G10b proves the run actually speculated and that useful-hit accounting
tracks the probe's accuracy. Off by default, pending an on-device A/B.
* feat(moe): cap predictive speculation at the top 2 predicted misses per layer
Speculating the whole predicted routing was measured on device at -38%
against its own baseline despite every intermediate metric improving
(hit 77.6->88.9%, stall 32->9 ms, 84% of speculations useful): the +33%
flash bytes and the vm commits fighting a full cache (major faults x3)
cost several times the stall removed. The stall a prefetch can remove is
head-of-line only -- overlap already hides the tail behind the expert
matmul -- so the cap keeps the part of the bet that can pay and drops
the part that provably cannot. Residency-aware: a predicted expert
already in cache does not burn a slot of the cap.
* perf(moe): rebuild the predictive prefetch around its measured costs
The observer-tax run priced the naive implementation: ~35-45 ms per
GEMV pass (ggml_fp16_to_fp32 is a function call per weight element --
21M calls/token) and ~20 ms/token for the extra isolated node, against
a speculation machinery that costs ~15-25. The GEMV and the barrier
were the feature; this commit removes both.
- gate_scores converts F16 natively on aarch64 (one instruction,
vectorizable) instead of a function call per element.
- The prefetch no longer isolates the gate matmul: the ask pass hands
over its source pointers for free, and the gate-input row is read at
the topk callback with no barrier of its own. A sampled watchdog
(the zero-staleness control, every 512 routings) validates the
barrier-less read and disarms the prefetch out loud if the memory
planner ever reuses that buffer -- without it, a future llama.cpp
bump could silently turn the predictor into a noise generator.
- The GEMV runs on a dedicated worker at a TWO-layer horizon: one
layer ahead has no landing spot (the callbacks between a layer's
gate and its own load are microseconds apart), while at l+2 the
worker has a whole layer for a ~0.5M-MAC job and the result inherits
the same post-load issue window as before. The probe now also scores
stale-2, so the extra layer of staleness is priced per model rather
than assumed.
- The prediction's residents are RETAINED (new IExpertSource::retain:
move-to-MRU, deliberately not a cache hit so the hit-rate metric
stays honest) -- protecting a predicted expert costs zero bytes,
unlike prefetching it. Only predicted misses are speculated, still
capped at 2.
The probe path keeps its barrier and both GEMVs: a probed run is
diagnostics, and its job is to be right, not fast.
* feat(moe): --predict-spec-max N — how much flash the prediction may spend (0 = retention only)
The cap was a constant; the retention-only point (0) is the config the
whole experiment now hinges on -- the prediction protecting predicted
residents from eviction while spending no flash at all -- and a
measured constant that cannot be varied is not a mechanism. Validated
[0, 8]; retention happens at every value because it is free.
* docs(predict): record the 2026-07-23 campaign — accuracy confirmed, throughput verdict open
Accuracy on device: stale-gate 88.6% (Qwen3-30B) / 80.7% (Qwen3.6),
control exactly 100.0% on every layer of both, prev-token 43/35% --
corroborating the offline route-trace estimates.
Throughput: the day's C-vs-B losses are recorded WITH their
invalidation. Re-running the reference on the by-then-hot device gave
3.93 tok/s against the cool morning's 6.54 with byte-identical I/O,
hit and drop counts -- the engine is deterministic, the -40% was
silent thermal capping, and every variant had been compared against
the cool number. The one thermally matched pair that was measured
(speculation on 8 io lanes, -28%, effective flash bandwidth 585->392
MiB/s) kills the more-lanes hypothesis specifically; spec-max 2 and
retention-only still owe a matched cool pair.
What did survive: the observer-tax decomposition (109 ms/token: the
per-element exported-function F16 conversion at 21M calls/token, plus
the barrier), and retention moving hit rate 0.1pp on a 3000 MiB cache
-- the offline replay bound confirmed from inside the engine.
* feat(app): expose the predictive prefetch as an experimental Streaming toggle
Off by default. Gated on streaming + a live cache like the temporal
prefetch, and the two settings disable each other in the UI -- the
engine refuses the pair, and a control the engine will reject is worse
than one that cannot be set. The spec-max rung selector (0/1/2/4)
surfaces the retention-only point, which is the configuration the open
throughput question most needs measured from the app. Session
signature includes both fields so flipping them reopens the process.
No versionCode bump: this is a PR-branch test build, not a release.
* docs(predict): record the 2026-07-24 matched pairs — read-ahead refuted, retention hit-neutral
A four-cell session (B, retention-only, spec-2, B sentinel) run at fixed 30 s
spacing re-proved the thermal-contamination mechanism (sentinel −17%, clusters
silently capped from cell 2) and yielded one genuinely matched pair: spec-2 vs
the B sentinel at the same caps and battery temperature, 3.14 vs 3.96 tok/s
(−21%) with hit rate up 4.3pp and 79% of speculations useful. Speculation
improves every metric it owns and still loses the wall clock — the flash has no
spare bandwidth to spend. Retention-only again moved the hit rate by nothing
(77.2% vs 77.6%), as the offline replay bound predicted.
* feat(app): contrast the two prefetch predictors in the UI, default spec-max to 0
The temporal and predictive toggles now say what actually differs — the bet
("repeats the previous token", ~40%) vs the question ("ask the next router a
layer early", ~85%) — instead of describing mechanisms side by side. Spec-max
defaults to 0 (retention only): the matched-pair A/B showed the read-ahead
losing −21% on a saturated flash, so 0 is the only rung the measurements did
not refute, and the helper text says so.
* chore(predict): file the changelog under Unreleased, drop a dead member, record the verdict
Three loose ends found reviewing the branch for merge:
- The changelog entries had been appended to the already-released 0.15.1 section;
they belong under [Unreleased], where the release commit carves them out.
- nu_hint_ was written on every routing and read nowhere: a leftover of the
synchronous first design, whose successor passes the routing width straight into
the prediction job.
- docs/roadmap.md still closed the routing-prediction question on the 2026-07-12
removal. It now records what reopening it with a training-free predictor found:
the accuracy is real and the throughput is not, for the same reason more lanes
and the sidecar lost.
Qwen3.6-35B (22.3 GB, ~1.5x RAM) on a Windows x86 laptop, engine
unmodified: baseline 4.78 tok/s, best recipe --overlap --cache-mb auto
--drop-cold-experts 0.75 = 7.33 (+55%), above the phone's 5.0-5.8 on
the same model.
Streamed decode is DRAM-bandwidth-bound there, the opposite of the
phone: compute sits at ~0.11 s/token in every cell (a ~9 tok/s ceiling
at zero I/O), doubling compute threads buys 2 ms, and io8 = io4 to the
millisecond once the cache is fixed — round 1's apparent +17% from
lanes was entirely a --cache-mb auto budget confound (5.3-7.4 GiB
across cells), recorded with its resolution. Aggregate flash read
stays ~900 MiB/s on a ~3 GB/s NVMe regardless of lanes, and with
overlap on it drops to ~330 MiB/s: async reads compete with FFN
compute for the same DRAM bandwidth.
README gets a dedicated Desktop section (table + the flip) replacing
the old one-line quick check; raw CSVs and findings.md land under
docs/bench-data/2026-07-24-desktop-qwen36/.
The threshold is a fraction of the uniform share 1/top-k, so what it
removes scales with how wide the routing is. At top-k 8 -- where every
number in docs/expert-dropping.md was collected -- 0.75 means "below 9.4%
of the routing", a tail trim. At top-k 4 it means "below 18.8%", and at
top-k 2 "below 37.5%", which on a miss discards the whole minority
expert: closer to halving the routing than trimming it, and unmeasured.
The engine now says so once at load when dropping is armed and the
effective top-k is 4 or fewer, quoting the actual share for the model in
hand. It warns rather than clamping or refusing: it cannot know whether
that trade is acceptable for a given model and task, and silently
adjusting a number the caller chose would be worse than a loud caveat.
MoeStreamConfig::drop_low_topk_warn is documented as an EVIDENCE
boundary, not a physical one -- nothing in the streaming path reads it.
The app shows the same caveat inline under the setting, in the error
colour, computed from the width the loaded model reports rather than
assumed -- and only once a model is loaded, since guessing would be worse
than staying quiet. gpt-oss is the case this exists for: it routes 4 of
128, and the app default is 75%.
To make that possible, BMOE_READY gains n_expert_used (the effective
width after any override, 0 on a non-MoE model) and Session exposes
n_expert_used() for embedders. Additive: older consumers ignore it.
Turbo top-k and cache-aware dropping made the same point in three places:
two long Features bullets and a top-k-only benchmark section that ended by
introducing the other knob. Now it is one three-line Features bullet and
one section, "Trading quality for speed", that states both knobs' numbers,
their determinism difference, and the GSM8K caveat once.
Also drops a near-verbatim duplicated paragraph in expert-dropping.md and
repoints its README anchor at the renamed section.
Closes the open item the throughput measurement left behind. 15 questions
taken verbatim from the GSM8K test split, same model and config as the
throughput A/B (Qwen3.6-35B-A3B, cache 3000, 4 lanes, overlap, ahwb,
no-think), four cells differing only in the threshold:
off 12/15 0 routings dropped
0.50 13/15 ~3%
0.75 13/15 ~14%
1.00 13/15 ~28%
Twelve of the fifteen questions give an identical final answer in all four
cells. All variation sits on two questions, and it flips in BOTH
directions as the threshold rises rather than worsening with it -- the
signature of a perturbation on problems already at the edge of the model's
competence, not of accumulating damage. Reply length is flat and no cell
produced a missing #### marker or an empty reply.
Decoding is greedy, so there is no sampling noise: every difference
between cells is caused by the policy. That makes the twelve identical
answers a real statement rather than a coincidence. It does not make 13
against 12 an improvement -- perturbing a question the model already got
wrong can land either side of the right answer, and at 6.7 points per
question this sample cannot establish the sign of the effect. It rules out
a collapse, not a subtle cost, and the write-up says so.
The grading rule was fixed before any output was read (number after the
last ####, falling back to the last number in the reply, applied
identically to every cell) and answers.csv carries every reply in full so
it can be audited. The prompt set, driver and grader are committed with
the results.
README gains a section stating how the three kinds of evidence are kept
apart -- gates assert correctness, device runs measure speed, a public
benchmark checks quality -- and where each comes from, including the
GSM8K source and licence.
* feat(moe): --drop-cold-experts, spend quality only where it buys I/O
Turbo top-k drops the tail of a routing whether or not those experts were
already in RAM. A resident expert costs no flash read, so that trade pays
quality for nothing on the ~80% of decode routings that are cache hits.
This skips a routed expert only when it is a cache MISS and the router
weighted it below frac x (1/top-k). Replayed over the committed route
traces at frac 1.0, decode phase, that avoids 66% of flash reads for 9.5%
of the router's weight mass, where --n-expert-used 5 avoids 23% for a
comparable 10.6% -- about 3x the reads at the same quality cost.
Implementation. The decision needs the FINAL router weights, which arrive
several nodes after the topk where the streamer normally loads, so
load_layer() is deferred to the terminal node of the layer's weight chain.
Which node that is depends on the model's gating, so the hook learns it
from the graph rather than carrying an architecture table; if it fails to
arrive the hook forgets it and re-learns rather than re-betting. A dropped
slot has its weight zeroed and its expert id repointed at the routing's
top-weighted expert: an unread expert can sit in reserved-but-uncommitted
VM and mul_mat_id would touch it anyway, so the kernel is given memory
that is certainly resident and multiplies it by exactly zero.
Requires the LRU cache -- with --cache-mb 0 residency reads all-miss and
the policy would silently degenerate into an unconditional weight cut.
Prefill is excluded by default. The top expert is always pinned, so no
routing can be emptied at any threshold.
Gates: G8a/G8a' prove the deferral and the learned terminal node are
transparent (byte-identical output, zero drops, at a threshold below any
producible weight); G8b that full strength against a constantly-evicting
cache never reaches an unloaded slot; G8c that at top-k 1 dropping is a
no-op, pinning both the top-expert guarantee and the threshold tracking
the effective top-k.
Three existing metrics shift meaning under dropping and the docs now say
so: cache_hit_pct rises without the cache serving more (a dropped routing
is a miss that is never looked up), and token/layer_demand measure what
was staged rather than routed. prefetch.md's "cannot change output" is
scoped, limitations.md gains the non-reproducibility entry, and
benchmark-method.md warns that reversing the run order cannot distinguish
a moved drop rate from a contaminated cell.
Off by default in the CLI and in the app. The output is not reproducible
-- what gets dropped depends on what the cache held -- so it carries no
rows in the README tables, and switching it on by default waits on a
published on-device A/B rather than on the replay argument alone.
* feat(app): default cache-aware dropping to 75%, measured on device
Qwen3.6-35B-A3B (top-k 8 of 256), in-app, cache 3000, one variable
changed: 2.549 tok/s off, 3.938 at F=0.75 (+55%), 4.702 at F=1.0 (+84%),
with flash reads falling 248 -> 163 -> 48 GiB. Per-token bootstrap
intervals separate every pair except off vs 0.50, which overlaps -- at
half the uniform share the policy drops 2.7% of routings and buys
nothing, which doubles as a negative control that the machinery is free
when it does not fire.
Run order was 1.0, off, 0.5, 0.75, so the two fastest cells are the first
and the LAST; thermal drift would have made the last the worst. The
mechanism orders by threshold even though the run order does not.
The replay turned out conservative rather than optimistic. It is
documented as an upper bound because it cannot model the cache changing
in response to dropping: at F=0.75 it was accurate (37% predicted, 34%
measured), at F=1.0 it understated (66% predicted, 81% measured). Avoided
reads free cache capacity, which raises the hit rate, which leaves fewer
misses to drop.
75% rather than 100% is deliberate: it takes the larger part of the win
for half the discarded routings (14% against 28%). Quality is still
unquantified -- no perplexity number and no side-by-side exists -- so the
conservative end of a measured range is the defensible default. The CLI
stays off; the byte-identity gates need a deterministic default.
Also records cache_hit_pct rising 67.8 -> 90.7% as the documented
accounting artefact rather than the cache serving more, and majflt/token
as dominated by each run's starting memory state, not by the threshold.
This reverts 45a90a2.
The feature merged before the evidence for its shipping default did. The
replay numbers argue the shape of the trade is favourable, but no on-device
A/B is published in this repository, and the app default it landed with
(75%) changes model output for every user of the demo app -- and changes it
non-reproducibly, which no other setting in this engine does.
Nothing was found wrong with the code. This is a sequencing decision: the
work returns as a pull request, with the app default back to off, so the
measurement lands before the default does.
Reverted rather than force-pushed: main is public and this commit was
already pushed, so the history stays honest about what happened.
* feat(moe): --drop-cold-experts, spend quality only where it buys I/O
Turbo top-k drops the tail of a routing whether or not those experts were
already in RAM. A resident expert costs no flash read, so that trade pays
quality for nothing on the ~80% of decode routings that are cache hits.
This adds the cache-aware version: skip a routed expert only when it is a
cache MISS and the router weighted it below frac x (1/top-k). Replayed over
the committed route traces at frac 1.0, decode phase, that avoids 66% of
flash reads for 9.5% of the router's weight mass, against 59%/37% for
--n-expert-used 3 — about 3x the reads avoided at a comparable cost. The
threshold is a curve, not a switch: 0.75 trades 4.4% of the mass for 37% of
the reads, better than --n-expert-used 5 on both axes.
Implementation. The decision needs the FINAL router weights, which arrive
several nodes after the topk where the streamer normally loads, so with the
policy armed load_layer() is deferred to the terminal node of the layer's
weight chain. Which node that is depends on the model's gating, so the hook
learns it from the graph rather than carrying an architecture table; until
it is known a layer loads at its topk node undropped. A dropped slot has its
weight zeroed and its expert id repointed at the routing's top-weighted
expert: an expert we decline to read may sit in reserved-but-uncommitted VM
and mul_mat_id would still touch it, so the kernel is given memory that is
certainly resident and multiplies it by exactly zero. Survivors are rescaled
by default, since a systematically shrunk expert output perturbs the residual
stream more than the missing contribution does.
Prefill is excluded by default (cold cache, ~4x the weight mass discarded,
and compute-bound anyway). The largest weight in a routing is always at least
the uniform share, so frac <= 1 can never empty a layer; validate() enforces
the bound and the top expert is pinned regardless.
Gates: G8a proves the deferral and the learned terminal node are transparent
(a threshold below any producible weight leaves the output byte-identical),
G8b that full strength with the cache off never reaches an unloaded expert.
Unlike every other knob this one is state-dependent: what gets dropped
depends on what the cache held, so output is not reproducible across runs.
Off by default, not in the app's settings, and NOT yet measured on device —
the numbers above are a static replay and an upper bound. docs/expert-
dropping.md states what is owed before it is recommended anywhere.
* feat(app): expose cache-aware expert dropping in Settings
Speed / quality -> Drop cold experts, as a percentage of the uniform
share (off / 50 / 75 / 100). The engine takes a fraction; the app stores
integer rungs, so the setting divides by 100 on the way to the flag.
Disabled in mmap mode: the policy asks the expert source what is resident,
and there is no expert source without the streamer. Included in the session
signature, so changing it reopens the session rather than being ignored by
a process already loaded.
Off by default. This exists so the A/B can be run where the engine actually
ships -- through the app, not a pushed CLI binary.
* fix(moe): require the cache for dropping, and correct what it reports
Review of the first two commits found the policy could be armed in a
configuration where it is not cache-aware at all, and that two of the
numbers it reports were wrong.
- Require the LRU cache. With --cache-mb 0 query_residency answers
all-miss, so the policy silently degenerated into an unconditional
weight cut -- exactly what --n-expert-used already does, under a flag
claiming to consult residency. validate() now rejects it, as it already
did for --prefetch, and the app gates the setting on the same condition.
- Fix experts_routed. It was incremented inside apply_drop, so it counted
what the policy examined rather than what the router selected: layers
before the terminal weight node is learned, and every un-armed phase,
were missing from the denominator. The reported drop rate was a fraction
of the wrong thing.
- Re-learn instead of re-betting. If the node learned as terminal does not
arrive, the deferral now also forgets it, so the next graph loads at the
topk node while it re-learns. Deferring again on a stale guess would
repeat the fault every token against a graph that had moved.
- Point the gates at a real cache. G8a/G8b ran with the cache off, where
the shared-slot path has no reserved-but-uncommitted memory -- so the id
repointing, which is the design's whole safety argument, was never
exercised. They now run against a constantly-evicting budget. Adds G8a'
(asserts routings were examined and none dropped, so an inert-threshold
flake fails legibly) and G8c (at top-k 1 dropping is a proven no-op,
pinning both the top-expert guarantee and the threshold being taken
against the effective top-k).
Docs: three metrics change meaning under dropping and none of them said
so. A dropped routing is a miss that is never looked up, so cache_hit_pct
rises without the cache serving more, and token/layer_demand measure what
was staged rather than routed -- documented in telemetry.md, pressure.md
(size the cache with dropping off, then turn it on) and metrics.h.
prefetch.md's "cannot change output" is scoped: under dropping a correct
guess un-drops an expert. limitations.md gains the non-reproducibility
entry, benchmark-method.md the axis plus a warning that reversing the run
order cannot distinguish a moved drop rate from a contaminated cell, and
architecture.md/runtime.h no longer claim unconditional determinism.
Fixes two anchors the README rename broke, and a changelog sentence that
quoted the equal-I/O row while drawing the equal-quality conclusion.
App: Drop cold experts defaults to 75%. The default is a product decision
taken on the maintainer's device; no benchmark for it is published here,
and docs/expert-dropping.md says that plainly instead of implying a
measured figure. The CLI stays off by default -- the byte-identity gates
need a deterministic default.
* feat(dense): add --dense-weights ahwb, dense weights in reclaim-exempt memory
The dense weights must stay resident — every token touches them — and
docs/android-memory.md finds every lever for holding them there closed: mlock is
capped at 64 KiB by the vendor, the cgroup protections are v2-only, MGLRU is
disabled at runtime, and MADV_COLD only redirects reclaim. The exception measured
in 0.13.5 is dma-buf: its pages stay pinned for the buffer's lifetime because a
device may DMA from them, an unprivileged app can allocate one through
AHardwareBuffer, and it reads at exactly anonymous-memory speed.
This wires that allocation into the dense-weights policy. `ahwb` is `anon` with a
single substitution — pio::pinned_alloc instead of the heap — leaving the O_DIRECT
read, the tensor rebind and the mmap handback identical, so an A/B between the two
moves one variable rather than comparing two code paths. Allocation is per tensor,
which keeps it well under the 2047 MiB lock ceiling; a tensor that did exceed it
fails the run instead of quietly taking an anon buffer, since a silent mix would
corrupt the comparison in the direction that flatters the feature. For the same
reason the mode refuses to start on platforms without such an allocation rather
than falling back, which would let an A/B become a mode against itself.
dense_resident_frac keeps working under it — mincore does report on the dma-buf
mapping, which was not obvious — and there it doubles as the falsification test:
pinned pages that fall below 1.0 disprove reclaim-exemption directly.
Exposed as a Dense weights -> Pinned (experimental) setting in the example app,
default off.
What is NOT established is that any of this helps, and the mode should not be
turned on because the reasoning is good. Reclaim-exempt memory does not create
memory: under a >RAM model the RAM the dense weights stop yielding is taken from
the expert cache or from the page cache feeding the stream. That is the trade that
already refuted the bulk restore (#28) and the per-layer LFU cap, both of which
delivered exactly the local gain they predicted and lost throughput anyway. The
deciding A/B is owed and must be run in the app, not over adb: single-shot adb runs
never idle, and this class of bug lives in the reclaim the app's engine suffers
while it sits.
Verified: all 7 host gates pass; on device the three modes generate identical text
on the tiny MoE model, ahwb allocates its 39 pinned buffers, and dense_resident_frac
reaches 1.000 under it.
* test(dense): measure --dense-weights ahwb at +17.9%, and correct the mechanism
In-app on a 1354-token generation (Qwen3.6-35B-A3B, k=8, cache 3000, same session
and binary as its control): 2.588 -> 3.053 tok/s, bootstrap intervals disjoint.
dense_resident_frac reads exactly 1.000, minimum included, in every pinned run, so
reclaim-exemption is now measured rather than inferred.
The mechanism is not the predicted one, and that correction is worth more than the
number. Major faults are EQUAL between the modes (265 vs 257): anon already keeps
the dense weights off the flash. What it does not prevent is the kernel taking ~15%
of them into zram, where a later touch costs a minor fault plus a decompression —
a cost that appears in no I/O counter and no fault counter, so it lands in
compute_ms, which is a residual rather than a measurement of arithmetic. The entire
delta shows up there (298 -> 241 ms) while io_ms, stall_ms and cache hit rate stay
within 1%, and swap falls 562 -> 294 MiB. anon protects the dense weights from
flash; ahwb also protects them from zram.
The trade this was expected to lose on does not appear: the expert cache is
untouched, hit rate identical to the decimal, because the dense set (~1.6 GiB) is
small next to a 3000 MiB cache budget. That is also why it should NOT be extended to
the cache without sizing the prize first — only ~294 MiB of that budget sits in zram,
against a cost of 3 GiB of rigid LMK-accounted memory and the loss of the
reserve/commit/evict elasticity the cache is built on.
Default stays anon. In the decisive pair ahwb ran first and an order effect cannot be
excluded — the reversed pair is owed — and this is one device, one model, one config.
Two negative results are committed alongside so the reasoning is checkable: three
67-74 token pairs that are ALL inconclusive (per-token CV 33-71%, every interval
overlapping), because reclaim accumulates and short turns never build up enough of
it; and a cross-day pair reading +63.6% that is not usable, since anon alone moved
+38.8% between the two days.
Transferable: compute_ms has been absorbing zram decompression all along, so earlier
"this regime is compute-bound" conclusions deserve re-examination.
* feat(tools): add bmoe-membench, a read-bandwidth probe for pinned memory
The dense weights are the one part of the model that must stay resident, and no
lever documented in docs/android-memory.md can keep them there: RLIMIT_MEMLOCK is
capped at 64 KiB by the vendor, the cgroup knobs are v2-only, and MADV_COLD only
redirects reclaim rather than preventing it. One allocation an unprivileged app can
make is exempt by construction — a dma-buf, whose pages stay pinned because a device
may DMA from them at any time. Userspace reaches it through AHardwareBuffer.
Whether that is usable hinges on a property nobody publishes: gralloc decides per
allocation whether a buffer is CPU-cacheable, and uncached memory loses the cache
line, the prefetcher and most memory-level parallelism. A dense matmul streams
weights, so it would pay that in full. On a device whose flash serves 1.3-2.5 GB/s,
uncached pinned weights would read SLOWER than refaulting them off storage — so the
idea is gated on a bandwidth measurement, not on a design argument.
bmoe-membench measures exactly that: the same sequential read kernel over an
anonymous mapping (what --dense-weights anon already allocates) and over a locked
AHardwareBuffer BLOB, reporting the ratio. The two outcomes are an order of
magnitude apart, so the tool is built to be unambiguous rather than precise.
--probe-max additionally reports the largest BLOB the driver will hand over, since
no limit is documented and the dense working set is multiple GiB.
It links nothing beyond the stdlib, so it cross-compiles in seconds and also builds
on the host, where the anon row still runs. No engine, CLI or app code is touched.
* test(memory): measure reclaim-exempt memory — bandwidth gate passes, size capped at 2047 MiB
A locked AHardwareBuffer BLOB reads at exactly anonymous-memory speed: within 0.5% on
both CPU clusters single-threaded (30.4k and 45.2k MiB/s) and at 4 threads (59.0k),
and the CPU_READ_OFTEN hint makes no difference to it. The way this idea could have
died on arrival was an uncached mapping — gralloc chooses cacheability per allocation,
and uncached memory would read at or below the flash bandwidth it is meant to save,
making pinned dense weights slower than refaulting them. It does not. The gate passes.
The binding constraint turned out to be size, in a place the documentation does not
point at. Allocation succeeds up to the 4 GiB format cap implied by the 32-bit BLOB
width, but AHardwareBuffer_lock — which is what yields a usable CPU pointer — refuses
at exactly 2048 MiB with EINVAL while 2047 MiB succeeds. A ceiling landing precisely
on 2^31 is a signed 32-bit type in the lock path, not memory exhaustion. The usable
unit is therefore 2047 MiB and anything larger must span several buffers, which this
engine can do since dense_weights already allocates per tensor.
Two corrections to the tool, both cases of it producing a confident wrong number:
- --probe-max probed allocation only, and so reported roughly double the usable size.
It now locks every candidate, and refines to 1 MiB so a ceiling sitting on a power
of two is visible as evidence about its cause.
- --modes ran anon first whatever order was asked for, and a single pass turned out to
rank CPU clusters rather than allocators: the first, un-pinned run reported a clean
0.67x that inverted when the modes were swapped, because the scheduler's choice of
core moves this number 1.5x. Modes now run in the order given, --repeat interleaves
them, and the header says to pin with taskset.
What this does not show is that pinning helps. Reclaim-exempt memory does not create
memory: under a >RAM model the RAM the dense weights stop yielding has to come from
the expert cache or from the page cache feeding the stream, which is the same trade
that refuted the bulk restore and the per-layer LFU cap. Nor did any run here put the
device under the pressure a >RAM decode creates, so reclaim-exemption itself remains
an inference from how dma-buf works. Recorded as an open, unproven lever.
docs/android-memory.md gains the dma-buf row its lever table was missing, the roadmap
records the open question, and the raw runs are committed alongside the analysis.
The branch carrying the experiment is going away; a tag preserves the
implementation without leaving a stale branch behind, and keeps the verdict note's
reference resolvable.
Adds the two instruments the measurements were taken with, the data, and fixes to
the maintained docs that turned out to assert things that do not hold. No engine
code changes: the one candidate that was implemented is a measured regression and
stays on its branch.
tools/bmoe-iobench (new, BMOE_BUILD_TOOLS=OFF by default) sweeps flash read
bandwidth against lane count and read size. It drives bmoe::FileReader -- the
engine's own read path, so alignment, the bounce buffer and the O_DIRECT
verify/fallback are part of what is measured -- but links nothing else, since the
I/O layer has no llama.cpp dependency. It therefore cross-compiles in seconds and
cannot perturb the streamer. --compute-load adds CPU contention, because the
streamer reads while ggml's threads spin and an idle-CPU number is not the
condition it operates under.
scripts/route-replay.py (new, stdlib only) replays the committed route traces
through hypothetical cache policies at zero device cost. It reproduces the
recorded on-device hit rate to the decimal on all three captures and independently
predicts a historical budget-shrink measurement it was not calibrated against.
What they found, and what it invalidated:
- roadmap.md opened its read-bandwidth theme on the premise that effective
O_DIRECT bandwidth sits far below the drive's ceiling because routed slices are
scattered. Measured, reads saturate at 2 lanes and are flat above 256 KiB, so
scatter is cheap here. Runtime read coalescing is retired (0.6-4 % of a layer's
routed experts are id-adjacent); the expert-contiguous repack survives but needs
a different justification.
- prefetch.md opened on "MoE routing has strong temporal locality". It is 17.9 %
on gpt-oss, beaten there by a static hot list, and --prefetch 1 is a 2x
slowdown. The mechanism and its correctness argument stand; the bet is annotated
with what it returns and with the case still open (top-6 models).
- cache-sizing.md said a too-small budget makes the hit rate "collapse". It makes
it exactly 0.0 %, and the boundary is one token cycle -- not cache_min_mb, which
is a different quantity roughly n_layer smaller and protects today's models only
by coincidence. Reproduced on device at a budget the CLI accepts.
bench-data/2026-07-20-cache-replay/ carries the three notes, the curves and every
run CSV, indexed by a README stating the six verdicts.
think_ctl reported none for any model whose template ignores enable_thinking while its handler
declares reasoning tags. But handlers publish that tag pair for a whole family, not per model: the
non-reasoning members advertise a <think> they never emit. LFM2-8B-A1B and LFM2.5-Instruct came out
uncontrollable, so the app disabled their Thinking switch and explained it with "this model always
reasons" — about models that never reason at all. A control that was merely moot became a control
that was taken away with a false reason.
none now requires positive evidence of both halves: the model declares a reasoning span AND its
template actually uses it. That second test is the one llama.cpp itself applies before wiring up
reasoning extraction for the family, so it costs nothing and stays model-supplied.
A second path had the same flaw: a template where the continuation hook does nothing also fell into
none, which caught any plain non-reasoning template. Reordered so the span question is asked first
and none is reachable only through it. Everything with nothing to suppress reports template — what
these models did before any of this existed.
Verified: LFM2.5-8B-A1B still none on-device. Host gates pin LFM2-8B-A1B, LFM2.5-Instruct, a plain
template and Gemma 4 (which reads the flag and never reaches the tag test) as template.
The prefill landed on every model whose template ignores enable_thinking. On-device that made
LFM2.5 strictly worse: handed a pre-closed empty reasoning span it reasons straight past it and
emits the reasoning UNTAGGED into the answer, where before it at least parsed into a collapsible
block. 500 tokens without reaching an answer.
The two families were never the same case. gpt-oss declares no reasoning tags because reasoning is
a channel the format separates structurally, so starting the turn past it is not something the
model can decline. LFM2.5 declares <think>/</think>: the span is the model's own to open and close,
so handing it an empty closed one is a suggestion, and this model was not trained on that
convention. Liquid ships a separate non-reasoning checkpoint rather than an off switch.
So the probe now reads that distinction off the tags the model itself declares — no model names,
no template string matching. Tags declared and the flag inert means the request cannot be honoured:
report none, disable the control, say why. No tags means the prefill binds.
Measured on-device, three models:
LFM2.5-8B-A1B none reasoning tagged into reasoning_content, answer clean ("4")
gpt-oss-120b prefill no reasoning, direct answer (the retired hardcoded path's behaviour)
Qwen3.6-35B-A3B template unchanged, no reasoning, direct answer
