Release v26.9.1

New Models & Model Updates

Delivered via the xDiT bump (2f74acb → 3aa7b34) and supporting
diffusers (f6e1c4d → 9f7aee4) and DistVAE (6d8025c → dce4841)
bumps:

• LTX-2.5 (full and distilled), with xDiT configs now in the image
  (#48, xDiT #754)
• MiniMax-H3, with xDiT configs now in the image
  (#16, xDiT #762)
• DistVAE tiled / tile-parallel VAE execution and extended model support
  (DistVAE #14, xDiT #750)

Performance Improvements

• Attention backends moved from Triton / Sage-style paths onto AITER MHAv4.
  Quantized and sparse configs now use aiter_i8fp8 / aiter_fp8 /
  aiter_fp8_sparge instead of aiter_sage and the older Sparge names
  (#41, AITER #4627, xDiT #752 / #758).
• Fused QK-RMSNorm + RoPE (and related packing) for FLUX.1, FLUX.2, Qwen Image,
  Wan2.X, and Z-Image (xDiT #755, #756, #760, #767, #763).
• LTX-2.5 parallel VAE decoding and AITER RMSNorm (xDiT #759).
• Memory-efficient model loading (xDiT #751).

Docker & Environment

• Dockerfile stages were rewired so build_python and base share a common
  base_os layer, and license evidence is copied into core instead of
  final, which restores cache hits when only later stages change
  (#43).
• Meson is pinned to 1.7.0 to avoid a ROCm link-arg duplication regression
  in meson 1.12.0
  (#19).

Benchmark Configurations

• Quantized-attention configs renamed to quantattn and switched to
  aiter_fp8; HunyuanVideo FP8/Sage and several Wan2.2 quantized/Sage
  variants were dropped
  (#50).
• Distilled and distilled-sparse HunyuanVideo 1.5 configs were removed
  (#29).

Other Changes

• AITER bumped 3a52614 → 0c2b0f77
  (#31).
• xDiT stack bump (xDiT, diffusers, DistVAE)
  (#32).
• MiniMax-H3 torch.compile no longer recompiles
  (#57).

Release v26.8

New Models & Model Updates

Delivered via the xDiT bump (c302663 → 2f74acb, 16 commits) and the
supporting diffusers bump (447e571 → f6e1c4d):

• Krea 2 (Raw and Turbo) text-to-image support
• Ideogram 4 support
• MiniMax-H3 support
• LingBot-Video support (MoE 30B and Dense 1.3B)
• FLUX.2 PipeFusion support, and PipeFusion enabled for FLUX.1-dev low-VRAM inference

Performance Improvements

• New aiter_sparge_fp8 attention backend built on a hand-written gfx950 FP8
  block-sparse assembly kernel (fwd_hd128_fp8_sparse.co). Q/K/V are quantized
  per-tensor to E4M3 and dispatched on an LPT-sorted work table so the heaviest
  sparse tiles launch first, collapsing one-workgroup-per-tile tail latency. The
  wan2_2.quantgemm_spargeattn.gfx950 benchmark now uses it in place of the
  hybrid aiter_sparge_v2 / aiter_sage schedule
  (#5).
• W8A8 INT8 quantization support (xDiT #729).
• AITER CK VSA for Wan (xDiT #744).
• AITER FP8 attention exposed as a plain function rather than a custom op, so it
  no longer blocks torch.compile (xDiT #749).
• QwenImage attention mask corrected under Ulysses sequence parallelism
  (diffusers #13756).

Docker & Environment

• The ffmpeg_builder stage now roots on base instead of ubuntu:24.04.
  Building from a prebuilt core image corrupted the separately-rooted stage and
  caused apt/GPG failures
  (#10).
• Patches and GPU code objects are now bind-mounted into the build instead of
  being copied in, so the explicit /patches and /code_objects cleanup layers
  are gone
  (e38f9f0).
• Image assets consolidated: assets/data now provides /app/data and
  assets/docs/README.md provides /app/README.md; patches moved from
  src/patches/ to patches/; the references/ quality-check payload is no
  longer added to the image
  (e38f9f0).