Add MagCache inference acceleration for Wan2.2 (T2V + I2V)

[HadarIngonyama]

Add MagCache inference acceleration for Wan2.2 (T2V + I2V)

Summary

This PR adds MagCache support to the Wan2.2 dual-transformer pipelines (both T2V and I2V), extending the existing Wan2.1 T2V MagCache support. MagCache skips the transformer blocks and reuses the cached block residual when the accumulated magnitude-ratio error stays below a threshold, using a precalibrated per-step mag_ratios_base curve so the skip schedule is deterministic (no data-dependent control flow, TPU/JIT friendly).

Measured speedups vs the dense render: ~1.82× for T2V and ~1.75× for I2V, with visually near-indistinguishable output.

What's included

• Wan2.2 T2V (wan_pipeline_2_2.py): MagCache skip path for the dual transformer — a single interleaved mag_ratios_base curve spanning both the high-noise and low-noise phases, a per-phase forced-compute (retention) zone, and an explicit cached-residual reset at the high→low transformer boundary.
• Wan2.2 I2V (wan_pipeline_i2v_2p2.py): the same skip path adapted for the image-conditioned pipeline (image condition concatenated with the latents, with the required BFHWC↔BCFHW transposes).
• generate_wan.py: threads use_magcache / magcache_thresh / magcache_K / retention_ratio through to both 2.2 pipelines.
• Configs:
  • base_wan_27b.yml (T2V): MagCache params + official mag_ratios_base, and flow_shift defaulted to 12.0 (see note below).
  • base_wan_i2v_27b.yml (I2V): MagCache params + official I2V-A14B mag_ratios_base, with boundary_ratio=0.900 to align the high→low switch with the curve (flow_shift stays at the I2V default of 5.0).
• Tests (wan2_2_magcache_test.py): host-side validation/schedule/core tests plus a TPU-only end-to-end smoke test.
• README: documents MagCache for Wan2.2 T2V and I2V, including the support matrix, config flags, sampling-shift requirement, and benchmark results.

Important: flow_shift alignment

mag_ratios_base is calibrated against where the high→low noise boundary lands, which flow_shift controls. Wan2.2 T2V requires flow_shift=12.0 (the official A14B sampling shift) — the previous default of 5.0 moved the boundary several steps out of phase, so MagCache skipped at the wrong steps and quality dropped. This PR sets the correct default, which also fixes the off-spec dense baseline. For I2V the official shift is 5.0, paired with boundary_ratio=0.900.

Results

Measured on a v7x (720×1280, 81 frames, 40 steps), reference = dense (use_magcache=False) render with the same seed/config:

Model	Settings	Speedup	Steps skipped	SSIM	PSNR
Wan2.2 T2V	flow_shift=12.0, thresh=0.04, K=2	~1.82×	18/40 (360s→198s)	≈0.72	≈21.8 dB
Wan2.2 I2V	flow_shift=5.0, boundary_ratio=0.900, thresh=0.06, K=2	~1.75×	17/40 (6.30→3.61 s/step)	≈0.91	≈25.4 dB

The reference-based metrics mostly reflect trajectory divergence — caching nudges the sampler onto a different but equally plausible sample — rather than visible degradation; cached clips are visually hard to tell apart from dense. I2V scores higher because the image conditioning anchors the trajectory. Recalibrating mag_ratios_base for a specific dtype/attention kernel can tighten the metric gap further.

Usage

MagCache is one of several mutually-exclusive caching strategies (CFG Cache, SenCache, MagCache) — enable only one at a time.

# Wan2.2 T2V
python src/maxdiffusion/generate_wan.py \
  src/maxdiffusion/configs/base_wan_27b.yml \
  use_magcache=True magcache_thresh=0.04 magcache_K=2 ...

# Wan2.2 I2V
python src/maxdiffusion/generate_wan.py \
  src/maxdiffusion/configs/base_wan_i2v_27b.yml \
  use_magcache=True magcache_thresh=0.06 magcache_K=2 ...

Testing

• wan2_2_magcache_test.py host-side tests pass (schedule/core logic).
• End-to-end T2V and I2V runs validated on a v7x TPU; speedup and SSIM/PSNR numbers above were collected from those runs.

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.