Sync with upstream master, including Qwen-Image-2.1 support - #15

Merged

danielhanchen merged 76 commits into

sync-upstream-20260921

Sep 21, 2026

Commits on Aug 11, 2026

fix: preserve "token_refiner" token for MiniMax H3 LoRAs (#1864)
stduhpf
authored
b4f1fd6
fix: fail with a message when MiniMax-H3 is run in img_gen mode (#1863)
danielhanchen
authored
487de75
feat: support INT8 ConvRot safetensors (#1857)
leejet
authored
bcc7e29

Commits on Aug 12, 2026

fix: replace free_compute_buffer with runner_done in vae (#1872)
LostRuins
authored
06c359f
sync: update ggml (#1873)
leejet
authored
fabe481
fix(ci): trigger builds for ggml updates
leejet
committed
de298c2

Commits on Aug 19, 2026

feat: add taeh3 support (#1874)
stduhpf
authored
6100d83
fix: prevent gallocr hash overflow in tiny graph-cut segments (#1880)
fszontagh
authored
58b6cb6
fix: re-clamp streaming VRAM budget to currently free memory (#1878)
fszontagh
authored
1706b32
fix: mark graph cuts with both a prefix and a suffix (#1883)
wbruna
authored
88b044b
fix: make max_order of lms sampler configurable (#1885)
vmobilis
authored
760717a
fix: guard against missing sampler/scheduler names (#1887)
vmobilis
authored
16304cc
chore: format code
leejet
committed
97d2990

Commits on Aug 25, 2026

Commits on Aug 27, 2026

feat: load scaled FP8 weights without upfront conversion (#1913)
leejet
authored
be0e344

Commits on Aug 30, 2026

fix: match exact weights in LLM config detection (#1923)
leejet
authored
2c92949
feat: add LTX-2.5 support (#1893)

pwilkin
and
leejet
authored
afd5306
fix: correct MiniMax H3 reference audio encoding (#1886)
jk212h20
authored
c797899
fix: correct MiniMax H3 audio Euler steps (#1908)
jk212h20
authored
dc4000d
feat: use backend-native FP8 matmul when supported (#1916)
leejet
authored
2540a4f
sync: update ggml
leejet
committed
134c821
feat: additional `--preview-interval` values (#1915)

vmobilis
and
leejet
authored
d9b6e27
feat: support numbering for preview images (#1895)
vmobilis
authored
9029655
fix: use carrier sampling for MiniMax H3 audio (#1924)
leejet
authored
40e605f
feat: generalize temporal tiling across video VAEs (#1926)
leejet
authored
6b3edaa

Commits on Sep 7, 2026

refactor: split ggml extensions and move implementations to cpp files (#1945)
leejet
authored
31ab2b2
fix: preserve K-quantized embedding weights (#1936)
nan
authored
9cdb6b6
fix: correct SDXL embeddings loading (#1939)
wbruna
authored
d04e895

Commits on Sep 11, 2026

Commits on Sep 13, 2026

Commits on Sep 16, 2026

Commits on Sep 19, 2026

Commits on Sep 20, 2026

feat: add Qwen Image 2.1 support (#1994)
leejet
authored
137f740
feat: restore legacy fp8 handling when building with upstream ggml (#2001)
wbruna
authored
008ca5b
fix: avoid passing ggml logs as format strings (#2002)
wbruna
authored
b8248a8
feat: add LLaDA-Image support (#1968)

fszontagh
and
leejet
authored
15f335d
feat: add native CUDA SageAttention support (#2005)
leejet
authored
187b256
fix: avoid narrowing conversion in SigVQ patch embedding and format code
leejet
committed
b56c686
docs: update CONTRIBUTING.md
leejet
committed
c678dfe

Commits on Sep 21, 2026

Merge upstream master (Qwen-Image-2.1, pipeline split)
danielhanchen
committed
8609f11