This website requires JavaScript.
28ec4e746e
fix: make no-op sparsification scoring consistent
Ethanfel
2026-09-16 19:45:56 +02:00
2334459577
fix: reduce H3 merge and autotuner RAM retention
Ethanfel
2026-09-16 18:21:20 +02:00
a844802476
fix: bound QKV reassembly memory (1.8.7)
Ethanfel
2026-09-10 17:21:46 +02:00
27482bb81a
fix: restore inline loader parity and retire legacy nodes (1.8.6)
Ethanfel
2026-09-09 12:45:10 +02:00
011c915fa7
release: ship H3 correctness fixes and research tools in 1.8.5
Ethanfel
2026-09-09 12:10:02 +02:00
cb92bd4f6e
fix: improve H3 merge fidelity and autotuner correctness
Ethanfel
2026-09-09 12:08:39 +02:00
481197701c
chore: bump version to 1.8.4
Ethanfel
2026-09-06 19:04:13 +02:00
2e5bf042d8
feat: add opt-in NP-LoRA and CT-Merging experiments
Ethanfel
2026-09-06 13:20:10 +02:00
5bc297010c
fix: harden H3 LoRA merge and export correctness
Ethanfel
2026-09-06 12:02:53 +02:00
a5e80a9601
fix: preserve MiniMax H3 merge fidelity
Ethanfel
2026-08-24 14:06:06 +02:00
72911cb7f7
feat: add MiniMax H3 support
Ethanfel
2026-08-16 09:53:55 +02:00
f57f4649da
chore: bump version to 1.8.2
Ethanfel
2026-07-19 19:21:35 +02:00
0cbc2631a8
fix(inline): preserve additive merge fidelity
Ethanfel
2026-07-19 19:21:27 +02:00
7c68c73748
chore: bump version to 1.8.1
Ethanfel and Claude Fable 5
2026-07-07 20:46:44 +02:00
925495930a
perf(inline): low-rank fast path for captured LoRAAdapter chains (10x AutoTuner sweep, identical output)
Ethanfel and Claude Fable 5
2026-07-07 20:21:06 +02:00
52a7a9ecfc
perf(inline): expand captured diffs on the compute device (was CPU matmul + dense transfer)
Ethanfel and Claude Fable 5
2026-07-07 20:13:47 +02:00
0ff74be52f
fix(inline): thread model-class arch hint into auto_tune so captured chains score under the right preset
Ethanfel and Claude Fable 5
2026-07-07 19:49:26 +02:00
46bf1a066c
fix(inline): key whole-stack analysis cache on captured-content identity so it hits cross-session
Ethanfel and Claude Fable 5
2026-07-07 19:34:33 +02:00
3609de6a6b
fix(inline): detect architecture from the model class for captured chains (resolves Qwen/ACE-Step ambiguity)
Ethanfel and Claude Fable 5
2026-07-07 19:29:56 +02:00
ca2b0e9125
fix(inline): conservative order-independent stamp matching to prevent dataset pollution
Ethanfel and Claude Fable 5
2026-07-07 19:02:10 +02:00
7136054af0
fix(inline): harden the global LoraLoader stamp wrapper (0/0 guard, signature-agnostic, chain idempotency)
Ethanfel and Claude Fable 5
2026-07-07 19:00:23 +02:00
48d726d1df
docs(inline): document filename stamping + community reconciliation for stock/rgthree loaders
Ethanfel and Claude Fable 5
2026-07-07 18:36:24 +02:00
ea6561feb2
feat(inline): use stamped filenames for real names, exact attribution, and file-based community reconciliation
Ethanfel and Claude Fable 5
2026-07-07 18:35:21 +02:00
106f036da6
feat(inline): stamp LoRA filenames onto the model via a LoraLoader wrapper (stock + rgthree Power Loader)
Ethanfel and Claude Fable 5
2026-07-07 18:28:50 +02:00
0b961f25b2
fix(inline): pin cache_patches off for the AutoTuner delegate + honest memory/community docs
Ethanfel and Claude Fable 5
2026-07-07 18:00:39 +02:00
1e1cce54b0
fix(inline): key memory + analysis caches on captured-content hash so persistence works across sessions
Ethanfel and Claude Fable 5
2026-07-07 17:59:40 +02:00
9ce78a8aa0
feat(inline): run the AutoTuner (search/memory/community) on captured chains
Ethanfel and Claude Fable 5
2026-07-07 17:34:58 +02:00
c2392aa216
feat(inline): stable content hash for captured adapters (enables inline memory/community identity)
Ethanfel and Claude Fable 5
2026-07-07 17:28:57 +02:00
cbe1a0e76f
fix(inline): read true LoRA rank from captured adapters in analysis (was always 1, over-compressed high-rank inline merges)
Ethanfel and Claude Fable 5
2026-07-07 17:18:01 +02:00
ee30b826db
chore: revert version to 1.8.0 (drop the extra bump from the refactor)
Ethanfel and Claude Fable 5
2026-07-07 16:39:30 +02:00
7115ce94a8
chore(inline): drop vestigial MAX_LORAS from slimmed inline node
Ethanfel and Claude Fable 5
2026-07-07 16:35:14 +02:00
93ba7b03c1
fix(inline): detect architecture from captured model-space keys (LTX/DiT no longer mis-detected as sd_unet)
Ethanfel and Claude Fable 5
2026-07-07 16:22:31 +02:00
7e6077897f
docs(inline): document the chain_options side node + version bump 1.8.0 -> 1.9.0
Ethanfel and Claude Fable 5
2026-07-07 16:07:04 +02:00
aa994e8ead
refactor(inline): move slot-visibility JS to the LoRAInlineChainOptions node
Ethanfel and Claude Fable 5
2026-07-07 16:06:03 +02:00
45189919e6
refactor(inline): move per-LoRA option widgets to the chain_options input
Ethanfel and Claude Fable 5
2026-07-07 16:05:07 +02:00
293f2d2b1a
fix(inline): treat alpha=None adapters (PEFT/diffusers LoRAs) as scale 1.0 in _expand_patch_to_diff
Ethanfel and Claude Fable 5
2026-07-07 15:59:56 +02:00
a93aa6a131
feat(inline): add LoRAInlineChainOptions side node
Ethanfel and Claude Fable 5
2026-07-07 15:56:00 +02:00
771193693e
fix(inline): pin cache_patches off (cache can never hit for inline) + clip patches guard
Ethanfel and Claude Fable 5
2026-07-06 23:10:55 +02:00
ee1d803e2e
docs(inline): README section for the inline chain optimizer + version bump
Ethanfel and Claude Fable 5
2026-07-06 22:59:20 +02:00
8f923b17c0
test(inline): harden e2e assertions (pre-run snapshot, merge direction, output_strength separation)
Ethanfel and Claude Fable 5
2026-07-06 22:50:15 +02:00
731c2ab0c9
feat(inline): dynamic slot visibility JS for the inline chain node
Ethanfel and Claude Fable 5
2026-07-06 22:47:39 +02:00
8176112751
feat(inline): register LoRAOptimizerInline node
Ethanfel and Claude Fable 5
2026-07-06 22:30:05 +02:00
0518924077
test(inline): end-to-end multi-LoRA chain capture through real optimize_merge
Ethanfel and Claude Fable 5
2026-07-06 22:29:11 +02:00
40f58c9681
refactor: share advanced-settings kwarg mapping between Simple and Inline optimizers
Ethanfel and Claude Fable 5
2026-07-06 22:15:38 +02:00
637dccfea7
fix(inline): dedicated IS_CHANGED + honest passthrough/arch reporting
Ethanfel and Claude Fable 5
2026-07-06 22:14:31 +02:00
28f2ba7a00
fix(inline): route single-LoRA virtual stacks through the full merge pipeline
Ethanfel and Claude Fable 5
2026-07-06 22:10:25 +02:00
75aafa9319
feat(inline): LoRAOptimizerInline node — chain capture, strip, merge, re-apply
Ethanfel and Claude Fable 5
2026-07-06 20:19:53 +02:00
a45421aa87
fix(inline): keep clip-side chain contributions alive through the zero-strength filter
Ethanfel and Claude Fable 5
2026-07-06 20:06:26 +02:00
2e8a4efd37
feat(inline): build virtual stack items from chain groups with slot options
Ethanfel and Claude Fable 5
2026-07-06 19:49:15 +02:00
40f30b3303
fix(inline): harden chain-capture classification and collision repair
Ethanfel and Claude Fable 5
2026-07-06 19:42:17 +02:00
1aaa3dedff
feat(inline): strip captured LoRA entries from patcher clones
Ethanfel and Claude Fable 5
2026-07-06 19:21:21 +02:00
cc9bf7485d
feat(inline): reconstruct ordered per-LoRA groups from ModelPatcher patches
Ethanfel and Claude Fable 5
2026-07-06 19:19:50 +02:00
dadb629ea2
feat(inline): classify capturable LoRA-family patch entries
Ethanfel and Claude Fable 5
2026-07-06 19:17:52 +02:00
194e34f5b8
docs(credits): credit CoreyCorza's comfyui-lora-loader for the cleaning feature
Ethanfel and Claude Opus 4.8
2026-07-04 21:13:07 +02:00
7154f986f5
feat(ui): expose per-LoRA cleaning knobs (star_eta / tame_layers / tame_threshold)
Ethanfel and Claude Opus 4.8
2026-07-04 21:08:56 +02:00
6b70c81574
feat(merge): per-LoRA cleaning engine — STAR spectral truncate+rescale + base-norm taming
Ethanfel and Claude Opus 4.8
2026-07-04 21:01:02 +02:00
3bb407efb0
fix(detect): stop krea2 from hijacking LTX-2 (to_gate_logits false-match)
Ethanfel and Claude Opus 4.8
2026-06-27 19:11:39 +02:00
4a474c80aa
feat(merge): extend shape-incompatibility report to LoKr/LoHa concepts
Ethanfel and Claude Opus 4.8
2026-06-27 12:24:30 +02:00
c03be771c2
chore(merge): bump AUTOTUNER_ALGO_VERSION to 1.12.0 for SLERP-imbalance gate
Ethanfel and Claude Opus 4.8
2026-06-27 01:28:19 +02:00
ee32e1a9e1
feat(merge): gate SLERP on per-prefix imbalance + clear shape-mismatch report
Ethanfel and Claude Opus 4.8
2026-06-27 01:19:30 +02:00
7f8e5035e9
perf(merge): skip dense+compression SVD for single-LoRA groups
Ethanfel and Claude Opus 4.8
2026-06-24 01:41:25 +02:00
d4ef3c75d2
fix(merge): serialize GPU compression SVD to stop async-race abort
Ethanfel and Claude Opus 4.8
2026-06-24 01:32:00 +02:00
b160b02c51
diag(merge): per-group 'processing <target_key>' log under profiler sentinel
Ethanfel and Claude Opus 4.8
2026-06-24 01:16:10 +02:00
9cd56c5c83
fix(krea2): normalize LoRA keys to model-native form (fixes silent drop + likely the crash)
Ethanfel and Claude Opus 4.8
2026-06-24 00:51:32 +02:00
21c3829a3b
Revert SVD change on main; keep it isolated on the krea2 branch
Ethanfel and Claude Opus 4.8
2026-06-23 22:55:36 +02:00
249664561e
perf(svd): Gram+eigh truncated SVD (fast + GPU-stable, more accurate)
Ethanfel and Claude Opus 4.8
2026-06-23 22:53:15 +02:00
cb1ac2dacc
fix(svd): replace torch.svd_lowrank with torch.linalg.svd (GPU-stable)
Ethanfel and Claude Opus 4.8
2026-06-23 22:27:08 +02:00
726d3cf06f
fix(arch): detect all Krea 2 LoRA forms via the attention gate marker
Ethanfel and Claude Opus 4.8
2026-06-23 22:02:34 +02:00
0750ff023c
docs(workflows): add LoRa Optimizer 1.7 example workflow
Ethanfel and Claude Opus 4.8
2026-06-23 21:41:45 +02:00
26828757eb
feat(arch): add Krea 2 detection + dit preset
Ethanfel and Claude Opus 4.8
2026-06-23 21:36:36 +02:00
f073ff87c1
docs(ui): clarify floor=0 removes the shrink limit
Ethanfel
2026-06-22 23:42:30 +02:00
79b001b757
docs(ui): fix off-switch tooltip — AutoTuner has no auto_strength toggle
Ethanfel
2026-06-22 23:27:44 +02:00
ed22e84954
docs(ui): point auto_strength_floor_mode tooltip at the real off switch
Ethanfel
2026-06-22 23:19:37 +02:00
c7b2be5043
feat(ui): auto-strength floor switch + corrected tooltips
Ethanfel and Claude Opus 4.8
2026-06-22 23:17:09 +02:00
89927a18c7
feat(profiling): auto-flag dominant overhead phases + 'other' row
Ethanfel and Claude Opus 4.8
2026-06-22 22:47:11 +02:00
138de7fb0e
chore: registry 1.7.8 (patch, not minor — reserve 1.8.0 for a major feature)
Ethanfel
2026-06-22 22:20:28 +02:00
f415ca7338
perf(diff-cache): fp32 storage + recompute past RAM budget (no disk spill)
Ethanfel and Claude Opus 4.8
2026-06-22 22:18:59 +02:00
95693ff96c
perf(merge): bypass diff cache for low-rank diffs (recompute instead)
Ethanfel and Claude Opus 4.8
2026-06-22 21:54:48 +02:00
8ca838e848
feat(profiling): add sentinel-file toggle for merge profiling
Ethanfel and Claude Opus 4.8
2026-06-22 21:05:50 +02:00
c7c1db0d84
feat(profiling): opt-in Pass-2 merge time breakdown by phase
Ethanfel and Claude Opus 4.8
2026-06-22 19:02:19 +02:00
18073fff66
fix(kernel): stop routing single-matrix SVD through the Triton kernel
Ethanfel and Claude Opus 4.8
2026-06-22 18:52:08 +02:00
f8931849ca
feat(kernel): add LORA_OPTIMIZER_DISABLE_TRITON escape hatch
Ethanfel and Claude Opus 4.8
2026-06-22 18:46:12 +02:00
76df75e4fb
feat(community-cache): hash fileless LoRAs from in-memory weights
Ethanfel and Claude Opus 4.8
2026-06-22 17:03:43 +02:00
562e62d013
fix(cache-key/editor): handle mixed tuple/dict stacks in cache key + conflict editor
Ethanfel and Claude Opus 4.8
2026-06-22 16:48:19 +02:00
149897e5da
fix(stack): support mixed tuple/dict stacks so dynamic stacker chains with extractor
Ethanfel and Claude Opus 4.8
2026-06-22 16:37:59 +02:00
9b6849076f
feat(extract): add optional lora_stack passthrough to chain extracted LoRA
Ethanfel and Claude Opus 4.8
2026-06-22 16:16:30 +02:00
4d6eb98200
perf(extract): run SVD on GPU instead of CPU
Ethanfel and Claude Opus 4.8
2026-06-22 16:03:50 +02:00
ac0bf9dd1d
fix(extract): align cross-device weights + robust SVD in LoRAExtractFromModel
Ethanfel and Claude Opus 4.8
2026-06-22 15:51:38 +02:00
588d6b0243
chore: untrack pre-existing plan docs swept in by git add -A
Ethanfel and Claude Opus 4.8
2026-06-22 12:51:08 +02:00
34ac112041
chore: registry 1.7.6 (patch), algo 1.11.1 (cache-clear bump)
Ethanfel and Claude Opus 4.8
2026-06-22 12:46:16 +02:00
12ba77a63a
fix(autotuner): community_cache + clear_and_run ignored by in-node cache
Ethanfel and Claude Opus 4.8
2026-06-22 12:34:02 +02:00
7eef6d6000
fix(autotuner): make all caches preserve-aware (toggling preserve now re-tunes)
Ethanfel and Claude Opus 4.8
2026-06-22 12:16:14 +02:00
785983d13f
fix(merge): revert automatic sum_preserve; style preservation is preserve-flag only
Ethanfel and Claude Opus 4.8
2026-06-22 11:53:36 +02:00
9bec8adc39
fix(report): correct stale 'upgraded to SLERP' reasoning for orthogonal groups
Ethanfel and Claude Opus 4.8
2026-06-22 11:35:37 +02:00
044f680ef7
fix(merge): use sum_preserve for orthogonal groups in ALL strategy sets
Ethanfel and Claude Opus 4.8
2026-06-22 11:27:47 +02:00
54fb127f37
docs+chore: document sum_preserve + preserve flag; bump algo 1.11.0, registry 1.8.0
Ethanfel and Claude Opus 4.8
2026-06-22 11:08:20 +02:00
cdb14d6a01
feat(merge): per-LoRA preserve flag to protect tagged style LoRAs in conflict merges
Ethanfel and Claude Opus 4.8
2026-06-22 11:05:34 +02:00
418be13847
feat(merge): bounded-additive sum_preserve mode to keep style LoRAs from washing out
Ethanfel and Claude Opus 4.8
2026-06-22 10:52:22 +02:00
ebff0615b8
chore: bump registry version to 1.7.5
Ethanfel and Claude Opus 4.8
2026-06-22 02:13:08 +02:00
d51c3918dc
feat(arch): add Anima (CircleStone Labs / Cosmos-Predict2 DiT) support
Ethanfel and Claude Opus 4.8
2026-06-22 02:11:45 +02:00
Ethanfel