MiniMax H3 mmproj support for Qwen3 VL 32B GGUFs - #473

Open

Nif00 wants to merge 39 commits into

city96:mainfrom

Nif00:minimax-h3-mmproj

Commits on Jan 16, 2026

fix: correct assignment operator and method signature
radicazz
committed
ba6409d
refactor: improve path resolution and code quality
radicazz
committed
bc9af6a
feat(nodes): add dropdown configuration helpers for loaders
radicazz
committed
a23a2dd

Commits on Mar 5, 2026

quant_ops: Implement GGUF as a QT
rattus128
committed
5cea0fa
loader changes for Dynamic VRAM
rattus128
committed
1ef0118
nodes: UNET: reconstructable ans support dynamic mode
rattus128
committed
6d95238
nodes: CLIP: make reconstructable and support dynamic mode
rattus128
committed
68b2b05
nodes: factor out clone() core of GGUFModelPatcher
rattus128
committed
5e905a1
GGUFModelPatcherDynamic
rattus128
committed
4c76351

Commits on Apr 6, 2026

feat: add torch dequantization for IQ1_S, IQ1_M, IQ2_XXS, IQ2_S, IQ3_XXS, IQ3_S
octo-patch
committed
7de582a

Commits on May 21, 2026

nodes: fix gguf loading
rattus128
committed
effce96

Commits on Jun 12, 2026

Add Ideogram 4 architecture support for GGUF loading

DanrisiUA
and
cursoragent
committed
5c4b048

Commits on Jun 16, 2026

Handle WeightAdapters in ops.move_patch_to_device
blepping
committed
3224f5c

Commits on Jun 17, 2026

Merge pull request #1 from radicazz/pr/upstream-cherrypicks-fixes
m8rr
authored
c04b0a2
Merge pull request #2 from octo-patch/feat/iq-lower-quant-dequant
m8rr
authored
8660a00
Merge pull request #3 from DanrisiUA/add-ideogram4-support
m8rr
authored
da823ec
Merge pull request #4 from blepping/fix_move_patch_to_device
m8rr
authored
6afc310
Merge Dynamic VRAM PR #427
m8rr
committed
54c5950
Apply torch.compile to dequant, up to 25% speedup depending on conditions
m8rr
committed
7c3cee5

Commits on Jun 18, 2026

refactor: Apply torch.compile to dequant
m8rr
committed
a7c141b

Commits on Jun 20, 2026

cherry-pick BF16 fix from commit a5d4ce5
m8rr
committed
b3abe74
Add workaround for Qwen3-VL text encoder
m8rr
committed
a963418
Added Vision Support in Qwen3-VL
m8rr
committed
249624c
Allow Qwen3-VL to run without mmproj
m8rr
committed
6a495b2

Commits on Jun 22, 2026

fix: add Dynamic-VRAM to mmproj loader
m8rr
committed
b7d7413
Draft: Gemma 4 vision support, DVRAM+BF16
m8rr
committed
00edba0
Refactor: Gemma4 now works with --disable-dynamic-vram
m8rr
committed
949a9cb

Commits on Jun 23, 2026

Generate tokenizer from Gemma4 GGUF metadata
m8rr
committed
5a6b742
Qwen3VL-4B support
m8rr
committed
c7fb703
refactor: change tokenizer return
m8rr
committed
a1641cb
Reuse GGUFReader for tokenizer loader
m8rr
committed
7187555

Commits on Jun 24, 2026

Stole: KREA-2 support (RealRebelAI/ComfyUI-GGUF_KREA-2)
m8rr
committed
5b2d245
Improve GGUFReader loading performance
m8rr
committed
736fc9c

Commits on Jun 25, 2026

Refactor: logging message handling, Feat. 19e396a
m8rr
committed
9a19c7d
fix: resolve torch.compile detection issues
m8rr
committed
1c113ea

Commits on Jun 29, 2026

gemma4 E4B runs, but unknown what side effects there might be
m8rr
committed
5e98d9c
Refactor: Remove unused argument
m8rr
committed
ccfb08e

Commits on Aug 4, 2026

Support Qwen3 32B
m8rr
committed
e71c2f6

Commits on Aug 5, 2026

Fix MiniMax H3 Qwen3-VL 32B mmproj loading
Nif00
committed
853daf6