Repository navigation

FLUX.2-klein PR 1: Scaffold for flow matching shared by several models, first of which is flux2-klein - #286

Merged

stephen-dwq merged 63 commits into

mainfrom

engine/dit-scaffold

Oct 2, 2026

Merged

FLUX.2-klein PR 1: Scaffold for flow matching shared by several models, first of which is flux2-klein#286
stephen-dwq merged 63 commits into
mainfrom
engine/dit-scaffold

Commits on Sep 27, 2026
- Add the DiT scaffold package for diffusion models
  merceod
  committed
  22dbcf4
- Add flow-matching schedules and the Euler step
  merceod
  committed
  44dbe7d
- Add multi-axis rotary embeddings for FLUX-family DiTs
  merceod
  committed
  6e00245
- Add joint attention over packed DiT tokens
  merceod
  committed
  1e93597
- Add latent packing and PNG helpers for image DiTs
  merceod
  committed
  e7aa1f1
- Add a native Qwen3 hidden-state text encoder
  merceod
  committed
  dedf868
- Add the generic denoise-loop submodule
  merceod
  committed
  6208799
- Add the KL autoencoder port to the DiT scaffold
  merceod
  committed
  5de1bfe
- Expose fused-linear shard slices for in-place updates
  merceod
  committed
  3042cc5
- Add static LoRA merging to the DiT scaffold
  merceod
  committed
  4b81515
- Test the denoise-loop scaffold on CPU
  merceod
  committed
  1deecf0
- Add a generic image adapter for diffusion models
  merceod
  committed
  b683390
- Add edit_image to the SDK client
  merceod
  committed
  0f6b87f
- Document edit_image in the SDK client table
  merceod
  committed
  4b41de5
- Test static LoRA merging on plain and fused linears
  merceod
  committed
  8ff1929
- Add an SDK image edit example
  merceod
  committed
  509a8d2
- Document the diffusion scaffold in the adding-models guide
  merceod
  committed
  a0246d8
- Let a ragged attention config declare its activation dtype
  merceod
  committed
  20b9ff1
- Build the ragged manager from the config dtype, KV dtype as fallback
  merceod
  committed
  ce13e60
- Test the ragged attention dtype resolution
  merceod
  committed
  a3672d0
- Declare every attention span of a denoise step with its own label
  merceod
  committed
  f667d12
- Test multi-span step declarations in the denoise loop
  merceod
  committed
  031e883
- Match the HF RMSNorm rounding order in the Qwen3 text encoder
  merceod
  committed
  75b1ced
- Quantise pixels in the image processor's dtype order
  merceod
  committed
  b545a10
- Add the Z-Image pipeline's torch.linspace base grid
  merceod
  committed
  d2e6e66
- Test the scaffold's norm, quantisation and grid numerics
  merceod
  committed
  3fca8d7
- Document attention spans in the diffusion scaffold guide
  merceod
  committed
  dbe8557
- Hand the compiled DiT one ragged callable per label
  merceod
  committed
  1bb86f8
- Test that ragged callables are stable across steps
  merceod
  committed
  20150e9
- Build the Qwen3 rotary tables exactly as HF does
  merceod
  committed
  209ca2e
- Test the Qwen3 rotary tables against HF
  merceod
  committed
  3dedc9f
- Mirror HF's mask-free attention path for unpadded batches
  merceod
  committed
  fdce9c9
- Encode PNG with parallel deflate and add the OpenAI output-format knobs
  merceod
  committed
  54c1c0a
- Test the PNG writers, parallel deflate and output formats
  merceod
  committed
  c971456
- Broadcast a per-request Euler dt over latents of any rank
  merceod
  committed
  9c2ed2f
- Test the batched Euler step on token and latent layouts
  merceod
  committed
  9b20840
- Address the VAE activations as modules
  merceod
  committed
  8931fd0
- Accept a single filename-bytes pair as one SDK upload
  merceod
  committed
  d0bb6d3
- Test the SDK's single filename-bytes upload
  merceod
  committed
  253217c
- Forward the edit size field and answer validation errors with 400
  merceod
  committed
  1cf1286
- Test the edit size passthrough and the 400 on validation errors
  merceod
  committed
  aa46539
- Let a graph config declare its static inputs' token axis
  merceod
  committed
  cb3962d
- Intern static buffers along the declared token axis
  merceod
  committed
  56bcf4f
- Declare the denoise loop's static inputs as row-leading
  merceod
  committed
  c11e5de
- Test static buffer interning with a declared token axis
  merceod
  committed
  4a2ee49
- Test the denoise loop's static token-axis declaration
  merceod
  committed
  e3f2ebf
- Take LoRA scaling from the adapter's sidecar config like PEFT
  merceod
  committed
  f0eb890
- Test the sidecar-config LoRA scaling
  merceod
  committed
  523b66f
- Say which engine dtype a cacheless attention inherits
  merceod
  committed
  a58c629
- Count a loop's re-injected inputs as ready for its next iteration
  merceod
  committed
  e2697c0
- Run speculative loop iterations under their own loop counters
  merceod
  committed
  5e3f552
- Test speculation readiness with a loop's re-injected inputs
  merceod
  committed
  7dff73e
- Test the loop counters a speculative batch runs under
  merceod
  committed
  ac351c5
- Stage persisted outputs for a send only when peer workers exist
  merceod
  committed
  6d4deb5
- Test when persisted outputs are staged for a send
  merceod
  committed
  e2d0d63
- Take the ragged dtype comment wording from #291
  merceod
  committed
  4d5bf44
Commits on Oct 2, 2026
- Merge main into engine/dit-scaffold
  
  NSagan271
  and
  claude
  committed
  7699dcc
- Add persist check to the rust runtime
  NSagan271
  committed
  462c697
- Add persist check to Python runtime
  NSagan271
  committed
  d628695
- make LoraAdapter a NamedTuple
  NSagan271
  committed
  710d5a5
- Make the denoise harness more generalizable
  NSagan271
  committed
  bcbf44b
- Merge main into engine/dit-scaffold
  
  NSagan271
  and
  claude
  committed
  95fea94
- raise on latent failure post step0 denoise rather than reseed
  stephen-dwq
  committed
  5ffb22c

Commits

Commits on Sep 27, 2026

Commits on Oct 2, 2026