Qwen image latent pack unpack

[nodeeeeee]

Add Qwen-Image latent pack/unpack kernels

Adds CUDA, Triton, and PyTorch backends with forward/backward support for FP32, FP16, and BF16. Pack converts spatial latents into tokens by grouping 2x2 regions across channels; unpack restores the spatial layout. Both preserve values and bit patterns exactly.

CUDA uses a two-stage shared-memory transpose with a 32x32 data tile and 32x8 threads. Autograd executes the inverse kernel for backward. Backend registration preserves existing dispatch behavior.

Validation passed 241 tests. Numerical correctness was verified against the PyTorch reference for pack/unpack outputs and gradients using zero tolerances and byte-for-byte equality. Batch invariance was tested by comparing the first sample processed alone with the same sample processed in batches of 1, 2, 4, and 8. Tests also cover round trips, edge shapes, special values, CUDA streams, fallback, and build configuration. Recorded sanitizer runs reported zero memory errors and race hazards; three launch configurations passed 90 cases.

H100 measurements cover 36 configurations: C=16, three image sizes, batches 1/2/4/8, and three dtypes. The table shows geometric-mean speedups over PyTorch; values above 1 are faster. CUDA Graph ranges show minimum-to-maximum speedups. Eager backward remains slightly slower. Measurements predate the readability refactor and do not represent end-to-end training speed.

Operation	CUDA Graph	CUDA Graph range	CUDA eager events	CUDA eager wall time	Triton Graph
Pack forward	1.337x	1.050-1.609x	1.092x	1.091x	0.543x
Unpack forward	1.646x	1.294-1.975x	1.096x	1.095x	0.958x
Pack backward	1.644x	1.299-1.972x	0.945x	0.945x	0.957x
Unpack backward	1.337x	1.046-1.606x	0.956x	0.956x	0.543x

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: a2a32221-2e05-476e-98d9-7daa267dd8cb

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}