get_block_mask_input takes only the mask now, so the helper pads to 4d itself, and the standalone nesting filter is inlined into prune_configs. The raise for a tile list that cannot nest moves into the prune_configs row, pinned with from_small so the fallback to the small configs cannot refill the list before the filter runs.
The option defaulted to 0 and a property turned that into 8192, which hid
the real default behind a sentinel and left no way to sparsify every
sequence. The default is now 8192 in the settings registry, 0 means no
floor, and the slider tops out at 32768 rather than a length nothing
reaches. A negative value is left unclaimed for a derived floor.
install_layout_hook took transformer or unet and stopped, so a pipeline
with a second denoiser left that one unhooked and its forwards ran against
whatever the first had published. The slot list follows the denoiser names
in sd_offload_state; the aux components on that list pack no attention
sequence, and a publication from one would clear the layout.
Fourteen axes drive the attention subsystem from the grid: the diffusers
method, the sdp override chain, the dispatcher kernel, five sdnq attention
knobs and six sparse settings. An axis writes shared.opts.data and rebuilds
the chain itself, since the onchange for these settings runs through the
queue lock a grid cell already holds. SharedSettingsStackHelper saves and
restores every attention setting around the grid, keys an axis introduced
included, and an axis for a setting a backend owns warns when that backend
is not in the active chain.
The classic tick test used a bare object as the pipeline's transformer,
which context.begin hands to register_forward_pre_hook whenever sparse
attention is enabled in the config.
Only the modular path had a per-forward kwargs hook, so a classic pipeline had
no layout and the selector sparsified its whole sequence, prompt conditioning
included. A denoiser pre-hook installed per generation now reads whichever
convention the model uses: the *_indices tensors, or the txt_ids and img_ids a
joint transformer is given.
Packing order is not derivable from the call and differs between architectures,
so only classes verified to pack text first publish; anything else falls back
as before rather than pinning the wrong half of the sequence dense.
The transformer is the only site that knows every segment length, since the
pad to a multiple of 256 happens there and the text stream is padded to a
fixed 512 whatever the prompt. It now publishes the live runs of the mask, so
the selector pins the conditioning dense and drops the padding instead of
sparsifying the whole sequence. At 2048x2048 with a short prompt that is 6 of
264 key blocks per query row moved from pinned to dropped.
The sdnq Triton kernel gains a block mask input (on the submodule's feat/block-mask branch; the pointer here is unchanged), so the backend advertises the block_mask cap and passes the router's selection as block_mask plus its block sizes. Prepare probes the installed kernel's signature and narrows the cap when the input is missing, warning when sparse attention is on, so an older submodule stays dense instead of failing. PlanEntry carries the narrowed caps, and the router, the sparse stage and the report read those rather than the declaration.
- benchmark: int8-sparse100/50/30/15 and int8-radial30 rows feed the same producer as the flex rows into the quantized kernel, skipped on builds without block_mask; a --configs filter selects attention rows the way --block-configs selects block rows; sparse rows of either backend are held out of the settings advice
- test/test-attention-sdnq-sparse.py: the kernel contract against the token-expanded mask, the nesting filter, the public entry on the quantized paths, the backward against the same selection and against fp32 autograd, and the flex consumer on one selection, pinned to one tile per run with a --tiles sweep
- locale: the sparse attention hint names both capable backends
Selecting blocks synchronized with the accelerator three times per
attention call: once to size the top-k, once to decide whether the
budget covered every candidate, and once to compute the density for
reporting. At the sdxl shape those stalls cost 1.6 ms against a 1.2 ms
attention, so the selector lost to dense at every budget.
The parts that depend only on geometry and layout are now built once and
cached per layout, ranking replaces top-k so no host side k is needed,
and density became a method that reads back only when something asks.
The block mask also stops sorting a mask of zeros to fill partial slots
it leaves empty. Selection overhead at sdxl drops from 1.84 ms to
0.23 ms, and a 30 percent budget moves from 0.36x of dense to 1.19x.
A sequence below the minimum length left the feature enabled and idle
with nothing in the log. The stage now records why it declined and says
so once per length, so an enabled setting that cannot act is visible
rather than quiet.
Sparse attention is a stage over the chain rather than a member of it:
one switch, and the router hands the selection to whichever active
backend advertises that it consumes a block mask, currently flex. A
backend declares that through a capability set, so the quantized kernel
joins later without touching the router.
The stage gates on the component role, self attention, a minimum
sequence length defaulting to the measured 8192 token crossover, and the
absence of a token mask or causal flag, which flex cannot combine with a
block only mask. Budgets follow a precomputed per step schedule with at
most two distinct values. Enabling the feature with no capable backend
in the chain warns and leaves attention dense rather than doing nothing
quietly.
The modular pre-forward hook now receives kwargs and publishes whatever
token layout the pipeline passes by name, so a packed sequence gets its
conditioning pinned without any model specific code. Without a layout
the whole sequence is sparsified and that is logged once per length.
The selector mean-pools query and key tiles, scores the tile pairs and
keeps the highest scoring key tiles per query tile within a budget
expressed as a fraction of the sparsifiable candidates. No scale and no
softmax, since top-k is invariant under both. Scoring runs on query
heads so grouped attention needs no key expansion, the diagonal is
always kept so no query row is left empty, and a budget that covers
every candidate reports dense instead of building a full mask.
The layout says what a packed sequence holds. Only video and image
tokens are sparsifiable; text, conditioning, audio and anything
unrecognized pin their rows and columns dense, padding is dropped, and a
tile straddling a boundary pins. Layouts come from the *_indices tensors
a pipeline passes its transformer by name, from ordered segments where a
transformer packs the sequence itself, or from a leading prefix as a
fallback.
The flex consumer builds a BlockMask with every selected tile in the
full slots, so mask_mod is never invoked and no dense mask is
materialized, and calls flex_attention compiled: called eagerly it reads
mask_mod rather than the block lists, so a block only mask attends
densely and silently. test/test-attention-sparse.py covers this with a
row that fails if the selection stops changing the output, alongside
tile equivalence against sdpa fed the same tiles, measured against the
flex kernel floor rather than an absolute tolerance.
Hosted truncation kept the full rank cap even when the spectrum ends in
numerical zeros, padding exactly low-rank deltas (low-rank LyCORIS,
full-family diffs) up to the cap. Slice the kept factors where
cumulative capture reaches 1 - 1e-6 of the sketch total, and trim
trailing all-zero columns when attaching cache entries written before
the slice, so they collapse the same way without a format bump. The
hosted log line reports the realized rank spread when it sits below
the cap.
- flat spectra keep the cap; a rank-8 delta under cap 256 stores 8 ranks
- select segments follow the effective rank
- suite pins the collapse, the flat-spectrum guard and the padded-entry trim
A cache hit still paid full calc_updown per layer, the dominant cost of a
replayed apply. The walk now probes the pass's cache entry first through
apply_cached, which evaluates the requantize routing rule from the delta
rms stored in the entry and attaches the cached factors without assembling
the delta; layers the rule declines fall through with the assembled delta
reused for the requantize path instead of recomputing it. The entry format
gains the rms and bumps to fmt 5, so older files recompute once and
rewrite. Hit accounting stays single-count through a plain lookup plus an
explicit hit note.
- replayed krea2 LoKR apply drops calc from 9.1s to 0.2s and activate
from 39.5s fresh to 4.8s replayed, hosted/routed split identical
- three cache tests pin the calc skip on plain, mixed and dense-pair sets
Hosting truncates every non-factorable set on sub-8-bit layers, but a
delta large against the grid step whose truncation capture is low is
retained better by the grid than by the rank cap. apply_hosted now
returns such layers to the requantize path when rms(delta)/mean(step)
exceeds 0.30 and the sketch capture falls below 0.90, thresholds sized
on 610 calibrated modules across krea2 and anima. Both terms must
agree: a thin delta rounds away on the grid however low its capture,
and a low-rank delta hosts exactly however fat it is.
- scoped to pure non-factorable sets; factorable members, dense-combined
deltas and svd-channel checkpoints keep hosting
- factor cache entries memoize the decision through their stored
capture, so replays route without re-running the sketch
- routed layers log as info apart from the forced-fallback warning
- the fidelity CLI applies the same rule so its reports track the loader
- seven routing tests, constants module-level and test-overridable
Persist now fires either when every hooked layer reaches the token quota
or at a fixed denoiser-forward deadline; layers under a small token floor
are omitted and stay on plain truncation. The all-done barrier alone
never fired on models whose modulation or pooled projections see a few
tokens per forward, so hooks stayed registered forever and statistics
recollected every session. eligible_modules walks the transformer or the
unet, so unet checkpoints collect statistics at all.
- lora_sdnq: drop the unused use_previous parameter; the factor cache
store now always runs inside a begin_pass
- tests: deadline persist, token-floor omission, unet root walk
The compiled dequant specializes per factor rank, so each distinct lora
set shape paid a compile stall on switch. Pad appended factors to a
power-of-two rank ladder (multiples of 64 past the hosted cap) with zero
fill: switches inside a bucket reuse the compiled graph, and common
trained ranks land on their bucket exactly so padding is usually a
no-op. Regression tests pin the factor add inside the single compiled
graph and the bucket reuse.
Rowwise int8 with fp32 scales halves cache entries; measured in output
space on real hosted deltas the roundtrip is fidelity-free (within
0.0003 of fp32 factors, bf16 storage likewise). Factors are quantized
before first use and the dequantized roundtrip is what the apply
attaches, so a fresh compute and a later cache hit stay bit-identical;
the entry format is versioned and pre-int8 entries reload as misses.
- lora_factor_cache: quantize_rowwise/dequantize_rowwise, store returns
the applied pair, fmt guard on read
- test/test-sdnq-lora-factors.py: int8 quantization test, entry-size
assertion in the roundtrip test
Hosting a non-factorable set costs one truncated svd per layer on every
fresh apply. The factors are deterministic in the checkpoint, the loaded
set, the host rank and the calibration statistics, so they persist under
data/lora-factor-cache keyed by that identity and replay bit-identically
on later applies. lora_sdnq_host_cache sets the disk budget in GB
(0 disables), least-recently-used entries are evicted past it. With the
cost paid once per configuration, svd subspace iterations rise from 2 to
4: about two thirds of the captured-energy gap to an exact decomposition
for +45% one-time compute; an exact svd measures 200-400x slower at
these shapes and is not viable.
- modules/lora/lora_factor_cache.py: signature, store, flush, eviction
- lora_sdnq.apply_hosted: fetch before computing, store after, hits in
the load summary
- cli/lora-quant-fidelity.py: matching niter for the hosted mirror
- test/test-sdnq-lora-factors.py: factor-cache category, 3 tests
New lora_sdnq_apply radio (exact, requantize) in the lora settings.
requantize keeps the previous behavior: every quantized layer takes the
dequantize-add-requantize path, with factor attach and svd hosting gated
off. A settings-only flip re-applies loaded networks: the mechanism
rides a per-module apply stamp and the network-changed signature, and
the activate fallthrough strips factors a closed gate leaves attached.
Requantize chosen by the setting logs as info instead of the
reduced-fidelity warning.
- locale hint covers fidelity and memory tradeoffs of both methods
- suite: gate, legacy routing and flip-transition tests
Plain svd truncation of hosted deltas is optimal in weight space but not
in output space: activations concentrate energy in a few input channels,
so scaling the delta by per-channel input RMS before the svd spends the
rank budget on output error instead. Statistics stream from the model's
own forwards on sub-8-bit SDNQ checkpoints and cache per checkpoint;
measured on real LoKR files this raises output-delta retention by ~0.05
at rank 256 and ~0.09 at rank 64, most on MLP down projections.
- modules/lora/lora_calib.py: capture hooks, per-checkpoint cache under
data/sdnq-calib, statistics land on layers as sdnq_calib_rms; gated by
lora_sdnq_host_calib, skipped when the model is compiled
- lora_sdnq.apply_hosted: weighted truncation when statistics exist,
calib count in the load summary
- cli/sdnq-calibrate.py: complete calibration now against a live server
- cli/lora-quant-fidelity.py --calib: hosted rho scored in the
activation-weighted norm
- test/test-sdnq-lora-factors.py: calibration category, 5 tests
Non-additive families (lokr, loha, oft, dora, full) merged into the
quantized weight and lost most of their delta on low-bit formats. On
sub-8-bit layers the set's calc_updown delta now rides the svd
side-channel as its top singular directions instead: factorable members
are subtracted out and appended exactly, so only the non-factorable
remainder is truncated. Truncation keeps the dominant part of the
effect and drops an orthogonal residual, where requantize keeps the
grid extrema and adds grid-shift noise of the delta's own magnitude;
on real lokr files retention rises from 0.04 to about 0.5 at the
default rank.
Hosted layers take no weight backup and unload bit-exactly. The svd
runs under a forked rng so generation seeds are unaffected. At 8 bits
and above requantize retains most of the delta and remains the path.
lora_sdnq_host_rank caps the hosted rank; 0 disables hosting.
- restore stashed svd factors onto the layer's current device; the
stash tuple does not follow module device moves, so an offload
between apply and remove left restored factors on a stale device
- recheck factor shapes for layers already in factor mode, so a
malformed stacked network downgrades to the legacy path instead of
raising in the concat
- clear the fallback log at activate entry so a raise mid-pass cannot
leak stale entries into the next report
- pin both behaviors in the suite and state the compute-dtype fidelity
floor in the module docstring
Checkpoints quantized without hadamard must attach factors unrotated;
checkpoints carrying their own svd correction must keep it under apply
and get the original factors back on remove. Both pinned in both svd
layouts.
Baking a lora into a quantized weight requantizes it, and on low-bit
formats round-to-nearest erases sub-step deltas (uint4 retains roughly
2/group_size of the signal). Plain lora deltas now ride the sdnq svd
side-channel: factors append to svd_up/svd_down with the down factor
hadamard-rotated, applied by the dequantizer at full precision in every
forward mode. Apply and remove are exact and take no weight backup.
- non-factorable families (dora, lokr, loha, oft, cp mid, dense bias)
fall back to requantize with a per-pass summary warning
- native fuse now honors the quantized-model guard; fuse requantized in
place on every network swap and accumulated drift
- layers that fell back on a mixed set restore from backup before
re-entering the factor path; untargeted quantized layers are no
longer flagged
- test/test-sdnq-lora-factors.py pins the erasure law, factor-path
exactness, memory accounting and set transitions
Batch matrix-matrix and Dynamic Attention BMM applied a legacy Attention
processor to pipe.unet, which a diffusion transformer does not have, so
they served unet models alone and said nothing elsewhere. The choices, the
processor and its slice helper are removed, an unrecognized method now
warns rather than selecting nothing, and a stored value is rewritten to
Scaled-Dot-Product on load.
attention_slicing holds one of Default, Enabled or Disabled, so testing the
string for truth sent Disabled down the enable branch and left the disable
call unreachable, while the log line below it reported the choice rather
than the action taken.
bypass_sdpa_hijacks and llm_context restore the pinned original over the
router and put the router back, including when the body raises. Captioners
and detailers run inside them, so the router must be removable.
The sdnq backend read six settings on every call; it now captures them
when the chain is built. Each backend declares the settings its call
captures, and webui registers one onchange over those names plus the
override set and the torch kernel flags, so a change rebuilds the chain
between jobs. When a compiled model is resident the rebuild also resets
dynamo, since its graphs hold the previous router.
SD_ATTN_DEBUG logs each distinct route once: backend, component role,
step, shapes, dtype and mask presence. The router takes an optional
observer for it, so the clean path carries one pointer check. report()
returns the active chain and generation context, and torch_info records
the whole chain as one string instead of the last prepared backend.
A module-level context tells attention consumers what is running: the
component role (transformer, text encoder, vae), the index of the
denoiser forward about to run, the pass length, and the model. It is
opened and closed around process_images, reset per denoising pass beside
the callback setup, and advanced by both step sources: the classic
callback passes the completed step plus one, the modular pre-forward
hook counts forwards. Roles come from the existing text encoder and vae
hijacks and the modular phase hooks. The step also lives in a device
scalar updated in place, so a compiled reader keeps its graph across
steps.
The flex backend never called the sdpa it replaced, so any backend
stacked before it was unreachable and every call it could not serve,
cpu or 3d inputs included, failed inside flex_attention. It is now an
ordinary entry gated on what flex_attention accepts: 4d tensors on one
non-cpu device. The mask path drops the 2d special case, which indexed
attn_mask.size and reshaped the mask onto the wrong axis; expanding to
(batch, heads, q, kv) already follows sdpa broadcast semantics.
Replace the six closure hijacks stacked in devices.set_sdpa_params with
a registry of declarative backends and one router installed in their
place. Each backend declares the constraints its closure carried as a
predicate, a priority matching its old stacking position, and a prepare
step that imports and configures the implementation; the router walks
the prepared entries by priority and hands declined calls to the
terminal backend (dynamic, flex) or the original sdpa, so fallback is
the router's job rather than each closure's.
- parity held: gates transcribed literally, the same kernel kwargs,
enable_gqa passed to the original only when set, torch_info keeps the
last prepared backend, the dynamic pin still set
- a backend enabled on a platform without it warns instead of silently
doing nothing
- the legacy set_* entry points are gone; devices.py installs the router
- test/test-attention-router.py checks every override subset against the
old stacking order, gate parity over 16,000 shape cases, dispatch,
terminal handoff and prepare isolation, offline
A delta that does not fit its target module cannot apply, and applying only
the layers that do fit leaves the model in a state nothing was trained for,
so try_load_chain drops the whole file when any family reports a mismatch.
Bias deltas were never checked against the target bias and could only surface
at apply time; a module with no bias stays a non-mismatch, since whole
architectures are built bias=False.
- check bias deltas against the module bias in the lora, norm and full loaders
- carry the mismatch count on the network so the chain can refuse the file
- record refused writes in the infotext so a partial apply is not read as clean
- point the krea2 full-diff test at a module that has a bias
network_add_weights defaulted its base tensor to self.weight for the bias
delta as well, so in fuse mode a diff_b was added to the weight matrix and
the result written into the bias. Layers where in and out differ threw a
shape error and had the weight matrix installed as their bias, square layers
broadcast silently, and either way the summary still counted the delta as
applied.
- pick the base tensor from the bias flag
- name the layer, target and both shapes in the mismatch error
- return which of (weight, bias) took a write, count the rest as refused
- report refused= on partially applied and partially removed networks
- cover both apply paths in test/test-lora-apply.py
run_ltx reported failure by yielding a string, which is why LTX had no API.
run() is the core underneath: keyword arguments named as video_run.run names
them, a VideoResult back, VideoError out with 499 for an interrupt. The lock,
progress and summary stay in the adapter, whose signature is unchanged since
callers bind to it by keyword. Failure now closes the processing object and
deactivates networks, which abort never did.
Pins the sentinel contract, row uniqueness, and the equivalence between
dispatch_mode and the ladder it replaced. A row that declares neither a name
marker nor a mapped pipeline class now fails here instead of reaching a runner
that would generate it as text to video.
Pins the cases that separate a complete slice from a truncated one, including
the zero-argument script whose empty slice is complete, and asserts no hook
runner slices the vector on its own.
The tab held the only code that built video and audio references, sniffed the
file type itself, and dropped anything it did not recognize: an unknown
extension, a file that had gone missing, and any decode failure were all skipped
without a word, leaving a request that generated from fewer references than were
uploaded.
Reference marshalling now goes through the same funnel the api path uses, and
runs before the load, so a rejected file costs nothing and says which file and
why. The workflow comes from the registry row, which is where the loader reads
it from as well.
- a rejected input returns its reason to the output box, since the general
handler only reaches the log
- references uploaded against a keyframe workflow warn instead of vanishing:
the accordion hides on a row change but the files it held do not
- guard p.close() in the finally, which the model-not-loaded return has always
reached before p exists
parse_options reads : and , as separators and = as the only assignment, so a
segment without = becomes a valueless flag set to 1. The shipped strings used
ffmpeg command line spelling, which parses without error into other values.
- crf:16 parsed to {'crf': '1', '16': '1'}, encoding every api, framepack and
seedvr video near lossless rather than at crf 16
- crf=23:b:v=0 pinned the generic bitrate option to 1 bit per second on vp8 and
vp9, collapsing their output
- qscale:v=3 reached mpeg4 and mjpeg as nothing at all, replaced by an explicit
quantizer range
- test-video-codecs.py asserts every preset segment carries an assignment
The core took reference images only, so no api caller could send the video and
audio references the ref2va workflow conditions on, and the marshalling that
handles them existed solely in the MiniMax tab.
validate_references now gates on the workflow and hands the entries to the
architecture that owns them, which accepts decoded images and local file paths
in any mix and preserves their order, since order fixes the labels a prompt
addresses. reference_caps exposes the same limits the validation enforces, so a
client reads them instead of mirroring the numbers.
- MAX_IMAGE_REFERENCES is gone: the limits now cover all three kinds and a total
- the run body no longer builds reference objects or knows their class
- an image is converted where it is built rather than at the call site, so a
reference decoded from a file and one posted as base64 arrive the same way
- pipeline args summarize a reference list by kind, since a decoded video would
otherwise print its frames into the per-generation log line
- the video endpoint documents what it actually accepts: images alone, because
video and audio decode from files rather than from the wire, and an upload
reference only where an extension provides the store that resolves one
MiniMax-H3 conditions on image, video and audio references, and the limits it
enforces on them are constructor defaults on a block class the package does not
re-export, so they cannot be imported and are mirrored here instead.
The resolver takes decoded images and local file paths and returns the reference
objects the pipeline reads, checking cheapest first: classification and counts
open no files, container headers are read without decoding, and only then is the
media decoded. Everything runs before the model load, so a rejected request costs
nothing.
- reference limits as a frozen dataclass, keyed by workflow since the rows that
carry ref2va differ only in which repo they load
- media classification and container probing as generic video helpers
- a url is refused before construction: the reference classes fetch one and
decode whatever comes back
- a video is bounded by duration and by what it decodes to, since the pipeline
truncates it to the generated length and the decode is held across the load
- the frame floor is counted on the decoded video at the rate it resamples to,
which is what the conditioner measures
- torchaudio and av are checked up front rather than surfacing as an import
failure once the weights are resident
Static inventory of every registered pipeline: component classes come from
the init type annotations, the component specs on modular pipelines, and
the video model definitions, so the real role function runs against real
classes with no weights loaded. The audit prints the role of every
component, lists which components turn resident under the shipped 22 GB
never-offload default, names the custom pipelines it cannot see, and
asserts that no component carries an undecorated entry point, that
opted-out classes route on-demand, and that every pipeline places a
per-step component.
CalamitousFelicitousness
Vladimir Mandic