A checkpoint-name key survives unload, so a reload of the same checkpoint
never printed the per-module stats block again, even with a different
quantization. Each component now carries its own reported stamp: a task
switch rebuilds the pipe around the same modules and stays quiet, while a
reload or a component swap brings new modules and reports them.
The group offload functions live in sd_offload_group, sd_offload_utils and
sd_offload_state rather than modules.sd_offload, so the tests and their
monkeypatches now import and patch those modules directly.
The parser no longer takes reference-name prefixes to tell bare
reference keys from bare diffusers keys. Any bare key carries the
sentinel and the arch resolver renames what it knows and passes the rest
through. Flux2 keeps its list for file-format detection only.
A bare key that matches no known prefix is parsed with the
bare-diffusers sentinel and handed to the resolver instead of being
dropped at parse time. The per-arch lists of bare diffusers prefixes are
gone, and a path that names no live module counts as unmapped.
A ChunkSpec can reorder equal row blocks of the rows it selects, so an
arch declares a swapped SwiGLU projection on the target instead of
permuting the state dict first. Only the LoRA family applies it; the
others skip a reordered target with a warning.
The reference fc1 is a fused [gate; value] SwiGLU projection and the
diffusers port stores [value; gate]. The native mapping did not swap the
halves, so gate and value deltas landed on each other's rows. The
mapping now also renames the standalone projections, reads a
metadata-only alpha, and accepts the musubi, peft dit and diffusers-named
layouts.
SDNQ codebook layers keep their Lloyd levels in the scale slot, so reading
scale.mean() as the grid step returned the levels' near-zero mean and sent
sub-step deltas to requantize, where the grid erases them. grid_step returns
the mean adjacent-level gap for those layers and the plain scale mean otherwise.
Anima 2.9B interleaves twelve new blocks among the 28 of Anima 1.0, so a block
index trained against 1.0 names a different block on the expanded model. Every
such key still resolves, since blocks 0 to 27 exist either way, so the
mismatch was silent. The Anima loader now shifts base-depth indices onto the
blocks that carry those weights, keyed by (base depth, expanded depth) and
applied only when the transformer is expanded and the LoRA stays inside the
base depth. Transformer keys move; llm_adapter and text encoder keys keep
their own numbering.
Anima-2.9B is a depth-expanded finetune of Anima 1.0 Base carrying 40
transformer blocks against the base repo's 28. The reference entry points at
the Diffusers conversion. Single-file releases load through the native loader:
TransformerSpec gains an infer_config hook, the Anima spec uses it to size
num_layers to the block indices in the file, and model_anima routes a
checkpoint-selected safetensors through the loader with the remaining
components from the base repo.
Two tables decide the native path: one says which architectures may take
it, the other says which loader they get. An entry in the second without
one in the first is a loader nothing can reach, and nothing checked that.
The error limiter halts a pass by raising, and nothing between the raise
and the caller put the model back. A halted pass left group offload hooks
stripped from every component the walk had reached, left a sequential
model on the cpu with offload disabled, and left the counters other
modules read describing the pass before it.
The epilogue moves into finish_pass under a finally, so the model returns
to its offload mode and the counters describe the pass that just ran. The
abort still reaches the caller.
Pass state is reset in one place in lora_sdnq now. Two of the six
accumulators were not being cleared at the start of a pass, and a stale
routed layer suppresses the fallback count for that layer next time.
The per-module walk carried four mechanisms inline, each repeating the
same tail: count the layer, stamp the pair that marks it current, advance
the bar, continue. Five copies of that tail and three of the backup probe
put the deepest arm nine levels in.
Each mechanism is now a function that either takes the layer or declines
to the next, and the walk reads as the four of them in order. The pass
state they share moves onto one object built before the walk starts, with
the accept tail, the stamp and the bar tick as its methods. That takes
network_activate from 218 lines to 55, none of it deeper than the module
loop.
Two shapes are deliberately not folded into that tail: the weight path
counts weights and bias separately and tracks what the module refused,
and the factor-strip restore stamps without counting. Hosting hands a
declined delta back rather than leaving it in a flag, so a pair of Nones
still reads as assembled and no layer is calculated twice.
The keys that mark a degradation as reported lived for the life of the
process, so a user who saw "flip=skipped weight=offloaded", changed the
offload mode and hit the same wall again was told nothing the second
time. Tie the set to the settings the warnings speak about: the stack
signature, the offload mode, the host rank and the checkpoint. Repeating
under one context still says it once.
The generic loader never offered a file to the boft type, so butterfly
OFT adapters reached the oft type instead, which claims any oft_blocks
key without checking its rank and then reads the block count as the lora
dim. Register boft ahead of oft; files with 3-d blocks still land on oft.
The nunchaku path replaced the loaded network list with the on-disk
entries it composed from, so reading a loaded network back hit an object
without the fields it expects: choosing the reported method reads
len(net.modules) and raised on every set change, costing that generation
its infotext and trigger tags. The adapter was already composed by then,
so the image was unaffected. Wrap the composed set in Network objects
and mutate the list in place.
The mechanism gate tests assert select_candidate declines under
requantize, and the apply-method hint names the cache option among
those the requantize choice disables.
The in-place weight installs, the promote-after-deactivate fuse
ordering, and the dynamo reset at model unload live in the shared
loader code; their regression pins belong in the campaign suite beside
the paths they protect.
<lora:name:1.0:lbw=VALUE> scales each targeted layer's delta by a slot of a
per-architecture block vector. VALUE is a preset name, a scalar, or a comma
vector; presets stretch onto the block count of the current model and the
a1111 17-slot and 12-slot layouts are accepted on sd and sdxl. The factor
enters through the module multiplier, so every apply path carries it: the
exact factor channel, hosting, requantize routing, dense stack combines and
select scoring.
- modules/lora/lora_blocks.py: slot classification from network_layer_mapping
(namespace-first, anchored chain prefixes), preset resolution reusing the
merge block-weight tables with BASE forced neutral, generated classic
segment names plus DOUBLE/SINGLE chain names, per-model memoization
- the raw spec stages through pending_config and promotes with the other
multipliers, keeping fuse removal consistent
- block weights join the activation signature, the per-module apply stamp
and the factor cache identity; entries without block weights keep their
existing signature bytes
- non-native load methods warn once and ignore the argument
The weight-kind schedule reset ran each winner's calc_updown on the target
weight's device, which on a block-swapped denoiser is the cpu; at hundreds
of layers per pass the cpu matmuls dominated every select generation. The
delta now computes on the accelerator and moves back, matching the activate
walk's convention.
- reset and flip execution log a debug timing line (materialize, select
loop, move/calc/apply split); the reset runs outside the activate walk,
so its cost was invisible to the load timers
Select scoring staged full fp32 copies, an abs copy and top-k workspace per
layer (hundreds of MB of transients that collide with block swapping on
offloaded denoisers) and recomputed scores from freshly assembled deltas on
every apply, which kept select modes out of the factor-cache fast path.
- score_pair: row-chunked fp32 interiors, fp64 accumulators, one device sync
- select scores persist in the factor cache as additive per-layer records
under the existing configuration signature
- apply_select_cached and register_weight_pair_cached replay a pair without
assembling deltas; the weight-kind winner is still computed at schedule
time
The klora and estlora balance factors accumulated across registrations
without ever resetting, so a multiplier change or pair swap blended the
previous registration into every later schedule. Balances now sum over
the live entries at finalize time, which keeps drop and re-register
consistent by construction.
- key flips one step early: step callbacks fire after the denoise, so
the winner is now live during the crossover step forward and a
final-step crossover engages instead of expiring
- include the calibration toggle in the factor cache signature so a hit
never replays factors computed under the other setting
- fall back to summation with a warning when hosting is disabled on a
quantized model instead of registering schedules that cannot flip
- drop the unused score_topk helper
EST-LoRA scores each layer by squared Frobenius energy, so a magnitude gap
between the two networks enters squared and the louder network wins nearly
every layer, starving the quieter one. The style side is now scaled by the
total-energy ratio (mirroring klora's gamma), making selection scale-invariant
so a network cannot take layers on magnitude alone. On the krea2 subject+style
pair this lifts the style network from 18% to 65% of the layer-step budget.
- lora_stack: accumulate per-mode energy totals, apply the balance in the est ramp
- test: content-louder est pair now hands over mid-schedule where raw scoring never would
- locale: note the est magnitude balance, and that a select mode gives each layer
to one network so both can be under-applied, while dense modes blend more fully
Requantizing a dense-combined delta into 8-bit weights is checkpoint-fragile:
on some checkpoints the round trip visibly damages the render while the same
combination hosted on the svd channel is clean. Dense-mode sets with two or
more contributing networks on a layer now ride the hosted path regardless of
bit width; single-set behavior at 8 bits and above is unchanged.
- host_candidate: dense multi-net layers qualify at any width
- suite: dense pair at int8 hosts; single non-factorable set at int8 keeps
the requantize fallback
Select pairs now ride the svd side channel on any SDNQ linear, not
only sub-8-bit ones: quantized backups are packed tensors, so the
weight rewrite path cannot recompute a winner from them and left
layers stripped mid-requantize. The sub-8-bit gate stays for dense
hosting, where requantize retains the delta at 8 bits and above.
Weight selection now only serves unquantized modules: finalize
iterates a snapshot so dead entries drop cleanly, materializes
balanced-offload modules before rewriting weights and skips modules
with stripped or quantized weights instead of corrupting the layer.
A select mode forced backup mode whenever it was merely set, so a
leftover setting changed behavior for ordinary single-network loads.
The fuse gate now engages only when the loaded set could actually
select (exactly two networks, compile permitting) or while selection
segments are still live on model layers. Re-application drops any
stale per-layer schedule so a later pass reset can never replay an
old winner over freshly applied weights. Fallback notices log per
activation instead of once per session; only the in-loop flip notice
stays latched.
The exact factor path was the only apply route with no log line; its
success read as silence. Track layers taking it beside the hosted and
fallback lists and report all three as key=value apply lines
(apply=exact/hosted/requantize); the stack fallback notices use the
same form. The suite pins its stack-mode baseline to sum so a mode
left set in user config cannot reroute tests that assume plain
summation.
Two-network subject+style sets select a winner per layer instead of
summing: scores are top-K magnitude sums (klora) or Frobenius energies
(estlora), and a timestep ramp shifts layers from the subject network
toward the style network across sampling, reduced to at most one
precomputed flip per layer per pass. On sub-8-bit SDNQ the pair rides
the side-channel as separate segments flipped in place; other layers
recompute the winner from the pristine backup, so select modes force
backup mode. Selection resets per pass from the callback setup and is
gated off under model compile. estlora's measured style-discrepancy
term is exposed as an option. Adds XYZ axes for the stack settings.
Add lora_stack_mode with ties, dare_ties, dare_linear and
magnitude_prune combination of per-network deltas when several loaded
networks target one layer; sum stays the default and the exact factor
path. Combined deltas ride the existing tail: hosted svd on sub-8-bit
SDNQ, requantize at int8 and above, direct add elsewhere. Text-encoder
layers and single-network sets keep plain summation. DARE masks draw
from per-layer sha256 seeds so re-applies and cache entries stay
deterministic; the stack settings join the activation and factor-cache
signatures so settings changes re-apply without a reload.
sparse_attention_exclude is a comma separated denylist matched case insensitively against the architecture, the pipeline class and the denoiser class, so one entry works whichever name is to hand. It resolves once per model rather than per call and declines with a log line. Seeded with CosmosTransformer3DModel, the transformer Anima runs, which returns banded noise at every budget tested against a sound dense baseline; listing the class rather than the architecture covers the other models built on it, none of which have been checked.
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
CalamitousFelicitousness
nan
Vladimir Mandic