Repository navigation

Stable Diffusion 1.5 - #779

Open

nshejwalkar wants to merge 34 commits into

cactus-compute:mainfrom

nshejwalkar:neel-diffusion

Open

Stable Diffusion 1.5#779
nshejwalkar wants to merge 34 commits into
cactus-compute:mainfrom
nshejwalkar:neel-diffusion

Commits on Aug 2, 2026
- add upsample_nearest2d op and TAESD VAE decoder test
  nshejwalkar
  committed
  5fc9b2c
Commits on Aug 3, 2026
- add simdgroup-MMA Metal fast path for conv2d_k3s1p1
  nshejwalkar
  committed
  8fa51be
Commits on Aug 7, 2026
- convert TAESD through the standard weights pipeline
  nshejwalkar
  committed
  780027c
Commits on Aug 8, 2026
- prepare VAE test weights through cactus test
  nshejwalkar
  committed
  4bb0426
- test upsample_nearest2d against a hand-computed reference
  nshejwalkar
  committed
  d4b8543
- inline the diffusers family check
  nshejwalkar
  committed
  bf4e6d2
- try every checkpoint stem when one fails
  nshejwalkar
  committed
  4e483fa
- reject colliding tensor names before writing, not after
  nshejwalkar
  committed
  c17def5
- pass the VAE weights through the test environment
  nshejwalkar
  committed
  8e69c98
- transpile nearest-neighbour upsampling
  nshejwalkar
  committed
  ba6db4c
- drop the hand-built VAE test for kernel benchmarks
  nshejwalkar
  committed
  216f911
- persist the upsample scale factor across save/load
  nshejwalkar
  committed
  e1c8303
- benchmark conv2d and upsample at graph level
  nshejwalkar
  committed
  a94815b
- test the transpiled decoder end to end
  nshejwalkar
  committed
  44f210f
- inline the checkpoint stem loop
  nshejwalkar
  committed
  a44b4cb
- drop the unused importorskip binding
  nshejwalkar
  committed
  4dd57e0
Commits on Aug 9, 2026
- restore the py.typed marker the packaging test expects
  nshejwalkar
  committed
  417ecb4
- tolerate group_norm exports that omit eps
  nshejwalkar
  committed
  6951845
- test the transpiled CLIP text encoder end to end
  nshejwalkar
  committed
  f78c6d6
- test the transpiled text-to-image pipeline end to end
  nshejwalkar
  committed
  dde57ef
- convert the SD text encoder and UNet through family adapters
  nshejwalkar
  committed
  1bb943f
- lower upsample_nearest2d in the replacement transpiler
  nshejwalkar
  committed
  82e01c0
- distinguish conv1d and conv2d fusions by kernel rank
  nshejwalkar
  committed
  92aa1fc
- wire SD text-to-image through the replacement transpiler
  nshejwalkar
  committed
  aa931c9
- materialize diffusion runtime assets and fix weight-name aliasing
  nshejwalkar
  committed
  6ed7fcb
- run text-to-image bundles in the engine
  nshejwalkar
  committed
  ca410ab
- keep declarations grouped by kind in the profile modules
  nshejwalkar
  committed
  08c6893
Commits on Aug 12, 2026
- reuse the Converter registries for pipeline components
  nshejwalkar
  committed
  4c0a701
- give the diffusion scheduler its own translation unit and test
  nshejwalkar
  committed
  e6a59b9
- drop comments the surrounding files do not carry
  nshejwalkar
  committed
  5b18b72
- add a generate-image command
  nshejwalkar
  committed
  cc4cb46
- keep attention whole through export so the fusion can match it
  nshejwalkar
  committed
  2ea22e8
- normalize the attention mask rank so masked attention can fuse
  nshejwalkar
  committed
  c6d1e5b
Commits on Aug 13, 2026
- pin the preserved-attention contract and tighten the conv discriminator
  nshejwalkar
  committed
  2ba9f46

Commits

Commits on Aug 2, 2026

Commits on Aug 3, 2026

Commits on Aug 7, 2026

Commits on Aug 8, 2026

Commits on Aug 9, 2026

Commits on Aug 12, 2026

Commits on Aug 13, 2026