Commit History
-
Sync the bundled nodes, workflows and docs with GitHub master (56adad3)
Verified signatureThe mirrored copy had drifted to roughly the 5f40849 era, before the all-in-one work existed. That left the all-in-one checkpoint in this repo unloadable by the code shipped next to it -- Krea2SVDQuantCheckpointLoader was not in it -- and kept the dynamic-VRAM pin that made iteration time jump to 30-100 s past ~12 GB. Updates 3 files and adds 0, including build_all_in_one.py and the 12 workflow JSONs that were missing.4be7e0c -
Sync the bundled nodes, workflows and docs with GitHub master (7d6a043)
Verified signatureThe mirrored copy had drifted to roughly the 5f40849 era, before the all-in-one work existed. That left the all-in-one checkpoint in this repo unloadable by the code shipped next to it -- Krea2SVDQuantCheckpointLoader was not in it -- and kept the dynamic-VRAM pin that made iteration time jump to 30-100 s past ~12 GB. Updates 4 files and adds 0, including build_all_in_one.py and the 12 workflow JSONs that were missing.1de67d5 -
Sync the bundled nodes, workflows and docs with GitHub master (099a051)
Verified signatureThe mirrored copy had drifted to roughly the 5f40849 era, before the all-in-one work existed. That left the all-in-one checkpoint in this repo unloadable by the code shipped next to it -- Krea2SVDQuantCheckpointLoader was not in it -- and kept the dynamic-VRAM pin that made iteration time jump to 30-100 s past ~12 GB. Updates 12 files and adds 15, including build_all_in_one.py and the 12 workflow JSONs that were missing.06a0bf4
-
Sync the node pack: compile-friendly kernel op, faster conversion, honest int8 numbers
Verified signatureNo checkpoint changed. This is the code and documentation catching up with a round of measurement, mirrored from the GitHub repo. Speed, per sampling step, RTX 3090 at 1024x1024, one ComfyUI session: int8 convrot 0.993 0.886 compiled svdq r256-actaware 0.845 0.718 compiled W4A4 noLowRank 0.678 crashes compiled (known, documented) So 1.18x over int8, 1.23x with both compiled -- not the 2-3x the old "vs BF16" column implied, and the README now carries an int8 row it never had. The decisive win over int8 is VRAM: 8.50 GiB resident against 12.84 GiB. Code: * krea2::w4a4_linear, an opaque torch.library custom op around the comfy_kitchen kernel. The old graph-break workaround cost 448 breaks per step across the 224 blocks; a compiled step is now one graph. 0.730 -> 0.696 s/step. Falls back automatically, KREA2_W4A4_OP=0 forces the old path. * Memoized backend resolution -- kitchen re-resolved it 1792 times per image. 0.846 -> 0.833. * Low-rank factors store…b0956bc
-
Sync docs and code: gallery, troubleshooting, cleanup pass
Verified signatured50078f -
Add activation-aware low-rank objective (--act-stats) + Test 4 benchmark
Verified signature6f63bf4