Skip to content
Permalink

Comparing changes

Choose two branches to see what’s changed or to start a new pull request. If you need to, you can also or learn more about diff comparisons.

Open a pull request

Create a new pull request by comparing changes across two branches. If you need to, you can also . Learn more about diff comparisons here.
base repository: AI-Hypercomputer/maxdiffusion
Failed to load repositories. Confirm that selected base ref is valid, then try again.
Loading
base: main
Choose a base ref
...
head repository: AI-Hypercomputer/maxdiffusion
Failed to load repositories. Confirm that selected head ref is valid, then try again.
Loading
compare: kernel_improv
Choose a head ref
Checking mergeability… Don’t worry, you can still create the pull request.
  • 1 commit
  • 19 files changed
  • 1 contributor

Commits on Sep 13, 2026

  1. feat(attention,wan): fixed-m splash and 2D-ring attention with Wan2.2…

    … TPU optimizations
    
    Implements exact fixed-m splash attention with Global Virtual K-Centering,
    2D Ulysses+Ring attention, and Wan2.2 fast serving optimizations for Cloud TPUs.
    
    Key Features & Safety:
    1. Fixed-m Splash Attention:
       - Dynamic C(N) headroom constants guaranteeing FP32 accumulator safety.
       - Fixed-m dtype safety checks (rejecting FP16/FP8 exponent overflow).
       - Global V-magnitude bounds (|V| <= 256) with safe online softmax fallback.
    2. 2D Ulysses + Ring Attention:
       - Global Virtual K-Centering via cross-ring pmean(k).
       - Cross-ring v_ok reduction via jax.lax.pmin preventing unsafe hop-level fixed-m.
       - Per-Q-block fixed-m accumulation with tuned tile size (block_q=9472).
       - Full GQA support across arbitrary Ulysses and Ring splits.
    3. Wan2.2 Fast Serving:
       - Persistent AOT compilation caching and zero-exec warmup.
       - Tuned TPU inference recipes in run_wan_fast_inference.sh.
    
    Hardware Verification (Cloud TPU v6e-8):
    - 44 tests passed in ring_fixed_m_test and custom_splash_fixed_m_test.
    - 25-prompt VBench evaluation: zero quality regression vs main.
    - Step latency: 3.445 s/step (tile 9472) vs 3.526 s/step (main), delivering a +2.32% denoise speedup.
    Perseus14 committed Sep 13, 2026
    Configuration menu
    Copy the full SHA
    66c8eb6 View commit details
    Browse the repository at this point in the history
Loading