Skip to content

Sync upstream master-1079-a3c49b1 - #10

Closed
Vect0rM wants to merge 49 commits into
atomicfrom
upstream-sync/master-1079-a3c49b1
Closed

Sync upstream master-1079-a3c49b1#10
Vect0rM wants to merge 49 commits into
atomicfrom
upstream-sync/master-1079-a3c49b1

Conversation

@Vect0rM

@Vect0rM Vect0rM commented Nov 10, 2026

Copy link
Copy Markdown
Member

Upstream master-1079-a3c49b1 (leejet/stable-diffusion.cpp) for stable-diffusion.cpp.

Opened by the engine update pipeline. It merges this PR itself if the core tests and the analysis come out green; otherwise it leaves the verdict here.

leejet and others added 30 commits November 1, 2026 21:18
@leejet
refactor: move model-specific operators and fusions into backend exte…
45a63ac 
…nsions (leejet#2253)
@leejet
fix: chunk MiniMax-H3 MLP to respect backend buffer limits (leejet#2257)
5356982
@leejet
fix: update CUTLASS license paths in Docker and Windows CUDA packaging
df0fa26
@leejet
fix: retry Vulkan capacity checks after memory reclamation (leejet#2256)
fe2682b
@leejet
fix: include missing cassert headers in SageAttention CUDA code
62c5443
@leejet
ci: trigger builds on CUDA header changes
5d33b1b
@leejet
fix: prefer discrete GPUs in auto-fit device selection (leejet#2259)
133087d
@leejet
feat: add custom modulate operator and ConvRot fusion (leejet#2258)
290148d
@leejet
ci: cache CUDA compilation for Windows and Docker builds
1350cf5
@cumal
fix: layer split across cuda (leejet#2262)
250ea6e
@leejet
fix: support integer LoRA scalar conversion to FP32 (leejet#2263)
b4d6b83
@leejet
ci: use ccache to fix Windows CUDA command parsing failures
3be2378
@mirek190
perf: use native SiTU for MOSS with composite fallback (leejet#2210)
f8ea2e1
@pwilkin
perf: view qkv slices for moss instead of slice+reshape (leejet#2232)
6225bf8
@leejet
perf: fuse Z-Image GPU operations and optimize BF16 GEMM (leejet#2265)
d308544
@leejet @mirek190
perf(cuda): fuse MOSS modulation and direct RoPE F16 output (leejet#2269
984eb06 
)

Co-authored-by: mirek190 <mirek190@gmail.com>
@leejet
feat: add P-H3-Image generation and editing support (leejet#2272)
3e92bf9
@farawayso
feat: MiniMax-H3 latent hires second pass (leejet#2249)
d135894
@leejet @mirek190
perf(vulkan): accelerate MOSS normalization and SiTU staging (leejet#…
58bb3c1 
…2273)

Co-authored-by: mirek190 <mirek190@gmail.com>
@leejet @noctrex
feat: support packed INT4 convrot models (leejet#2276)
7952693 
Co-authored-by: noctrex <noctrex@gmail.com>
@leejet
perf: accelerate packed INT4 CUDA inference with CUTLASS (leejet#2278)
f5df5f8
@leejet
perf(vulkan): accelerate packed INT4 with cooperative matrices (leeje…
5a1ba5d 
…t#2280)
@daniandtheweb
feat: ControlNet Union support for SDXL (leejet#2284)
fdfc913
@leejet
fix: chunk MiniMax-H3 QKV projections within backend buffer limits (l…
b1f5a51 
…eejet#2286)
@leejet
feat: add RelayVSR video upscaling with CUDA window attention (leejet…
9234ad8 
…#2285)
@leejet
perf: reduce fused compute buffers (leejet#2290)
5f8c19f
@LWWZH
fix: clean LoRA multiplier and prompt spacing in image metadata (leej…
7bcd4ce 
…et#2283)
@leejet
feat: add configurable attention chunk budget for MiniMax-H3 (leejet#…
de00baf 
…2292)
@leejet
fix: avoid CPU fallback based on backend allocation chunk size (leeje…
0b3b98d 
…t#2294)
@leejet
fix(metal): handle all batch elements in group normalization (leejet#…
f61053c 
…2296)
DrMkdaddy and others added 19 commits November 8, 2026 21:37
@DrMkdaddy
fix: use patch token count for the Flux scheduler (leejet#2293)
b43af40
@leejet
feat: add Metal INT8 ConvRot support (leejet#2297)
3d315c4
@leejet Video Rebirth
feat: support HyperFlow H3 Flash (leejet#2299)
80c4707 
Co-authored-by: Video Rebirth <opensource@videorebirth.com>
@leejet
fix(metal): reclaim completed transfer resources without blocking
abc6718
@leejet
feat: add SeedVR2 image and video upscaling (leejet#2300)
cd5594e
@leejet
fix: preserve Krea2 text-fusion names in Kohya LoRAs (leejet#2301)
b43ce4b
@leejet @maximedrn
feat(server): report generation stages independently of previews (lee…
fc350be 
…jet#2303)

Co-authored-by: Maxime Dréan <maximedrean@proton.me>
@wbruna
fix: disable batch_cfg for Metal when building with upstream ggml (le…
f977add 
…ejet#2311)
@Windyskr
perf: cache the device type in sd_backend_is_cpu (leejet#2307)
d22dcda
@Windyskr
fix: merge LoRAs in F32 when convert changes the tensor type (leejet#…
cfdf3cf 
…2302)
@Windyskr
fix: apply MiniMax-H3 Diffusers LoRA token_refiner weights (leejet#2306)
008b9a2
@leejet
fix: limit default MSVC compilation parallelism
991f825
@d3Ita
docs: add Prisme to the UIs list (leejet#2274)
fbb381d
@0xConflict
docs: add Quo to UIs (leejet#2279)
c44cbec
@Windyskr
fix: size layer-split ranges in proportion to device capacity (leejet…
46da352 
…#2309)
@georgi
docs: add NodeTool link to README (leejet#2277)
d36f72b
@leejet
feat: update server web UI with streaming prompt enhancement (leejet#…
73fef08 
…2312)
@leejet
refactor: unify ggml operation helpers around runner context (leejet#…
a3c49b1 
…2314)
@Vect0rM
Merge upstream master-1079-a3c49b1
b5ed0c4
@Vect0rM

Vect0rM commented Nov 11, 2026

Copy link
Copy Markdown
Member Author

Superseded by #11.

@Vect0rM Vect0rM closed this Nov 11, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.