Add Ideogram support and improve BF16 dequantization handling - #459

Open

molbal wants to merge 75 commits into

city96:mainfrom

Commits on Jun 12, 2026

.
molbal
committed
dffa75b

Commits on Jun 24, 2026

Add Krea-2 model support and batch converter for GGUF files
molbal
committed
16ddaa4

Commits on Jun 25, 2026

Update README and loader.py for Krea-2 and Ideogram 4 GGUF support
molbal
committed
cd2c9bd
Update README to include new pre-quantized models and clarify quantization support limitations
molbal
committed
e532b7d
fix typo
molbal
committed
f65d0d9

Commits on Jun 27, 2026

Enhance tensor handling in convert.py to drop incompatible FP8 scale tensors and skip 0-dim scalars during GGUF conversion
molbal
committed
73439d2

Commits on Jul 26, 2026

Adds Q8_CR quant type
molbal
committed
3551793
Q8_CR follow-up work
molbal
committed
89f20c8
Adds docs
molbal
committed
2976720
Merge pull request #3 from molbal/feature/molbal-dynamic-gguf
molbal
authored
288327f

Commits on Jul 29, 2026

Add Dynamic VRAM support for GGUF loaders
molbal
committed
bcaa754
Merge pull request #4 from molbal/feature/molbal-dynamic-gguf
molbal
authored
39cf309

Commits on Jul 31, 2026

Update tensor handling for Krea2 architecture
boessu
authored
340da95
Enhance tensor reading and comparison functionality
boessu
authored
4534f96

Commits on Aug 1, 2026

Merge pull request #6 from boessu/patch-2
molbal
authored
26dff74
Merge pull request #5 from boessu/patch-1
molbal
authored
e2dc203

Commits on Aug 2, 2026

Implement key_matches function for tensor name validation
boessu
authored
937051b
Merge pull request #7 from boessu/patch-3
molbal
authored
e157a2d
Add experimental INT4 LoRA support
molbal
committed
90260fe

Commits on Aug 3, 2026

Update README and pyproject.toml for repository details; modify quantization defaults in conversion script
molbal
committed
a731dc3
Add icon and update pyproject Icon URL
molbal
committed
c7bd78f
Add targeted GGUF quantization and progress
molbal
committed
f010b7d
Merge pull request #8 from molbal/feature/targeted-quantization
molbal
authored
8f03ac0
Add Minimax H3 support
molbal
committed
17901fa
Merge pull request #9 from molbal/feature/add-minimax-h3-support
molbal
authored
21cdd03

Commits on Aug 4, 2026

Add Qwen3-VL detection marker injector and tests
molbal
committed
783ca38
I always forget
molbal
committed
0625a03

Commits on Aug 10, 2026

Commits on Aug 11, 2026

Commits on Aug 12, 2026

Commits on Aug 13, 2026

Fix GGMLTensor.dtype crash on inference-mode tensors

Danielkgr
and
claude
committed
07e73da
Add local GGUF conversion dashboard
molbal
committed
c608205
Add LTX 2.5 latent upscaler support
molbal
committed
e530862
Add MiniMax Music3 support and docs
molbal
committed
d37bf4f
Fix missing import
molbal
committed
8cd821d
Update pyproject.toml
molbal
authored
f89200e
Merge pull request #12 from Danielkgr/fix/ggmltensor-dtype-inference-mode
molbal
authored
9c5e485

Commits on Aug 14, 2026

Ensure force_patch_weights is set when partially loading/unloading models

blepping
authored and
molbal
committed
f65d77e
Merge pull request #15 from molbal/fix_partially_load_unload
molbal
authored
ab74c65
Add IQ2/IQ3 dequant support and tests
molbal
committed
b1bd745
Merge branch 'main' of https://github.com/molbal/ComfyUI-GGUF
molbal
committed
c368529
bump version
molbal
committed
3246b55
Load and map Qwen3-VL mmproj vision tower
molbal
committed
9f2f166
Defer deepstack check; add mmproj patch test
molbal
committed
1288606
Support misspelled 'v.deepstast' prefix for Qwen3-VL
molbal
committed
b601643

Commits on Aug 24, 2026

Add Qwen3.5 support
blazewicz
committed
f77606e

Commits on Aug 27, 2026

Merge pull request #16 from blazewicz/add-qwen35-support
molbal
authored
0befefd
Update issue templates
molbal
authored
4fb1ec6

Commits on Aug 30, 2026

Add support for Q4_CR quantization format (INT4 W4A16)
molbal
committed
94c519c

Commits on Aug 31, 2026

Commits on Sep 1, 2026

Commits on Sep 3, 2026

Restore native INT4 LoRA residual path
molbal
committed
a13ea9b

Commits on Sep 4, 2026

Restore native INT4 LoRA residual path
molbal
committed
a553b8a

Commits on Sep 5, 2026

Refactor legacy INT4 LoRA path handling to prevent corruption of packed weights during fusion. (Make LoRA/LoKR work with Q4_CR format with legacy loader)
molbal
committed
967b26c

Commits on Sep 6, 2026