Cache ModelMixin.dtype to avoid named_parameters walk per access

[akshan-main]

What does this PR do?

Addresses #13401

ModelMixin.dtype calls get_parameter_dtype(), which walks named_parameters() on every access. Pipelines read self.transformer.dtype / text_encoder.dtype / vae.dtype inside denoise loops, so the walk fires every step.

This PR caches dtype on first access and invalidates in _apply, which .to(), .cpu(), .cuda(), .half(), and .bfloat16() all route through. Generation outputs are bit-identical. A microbench on AutoencoderKL drops cached .dtype access from ~88us to ~0.1us (~1000x).

device is intentionally not cached: with group offloading the effective device changes per-forward, so a cache there would be wrong.

Same shape as the cache_context._set_context cache in #13356

Profiling: 10 pipelines (eager, 2 inference steps, A100)

pipeline             |  dtype calls  | dtype total (ms)  | step gap (ms) | pipeline_call (ms) 
                     | before  after |   before    after | before  after |    before     after
----------------------------------------------------------------------------------------------
flux2                |      2      0 |     6.04     0.00 |   3.83   0.85 |    384.78    369.71
qwenimage(fd)        |      1      0 |     2.46     0.00 |   0.29   0.29 |   2148.46   2147.69
qwenimage_edit(fd)   |      1      0 |     2.44     0.00 |   0.30   0.31 |   6250.67   6237.57
z_image              |      2      0 |    10.52     0.00 |  10.63   5.40 |   1394.03   1391.21
chroma               |      0      0 |     0.00     0.00 |   0.16   0.17 |   2163.04   2162.70
sdxl(fd)             |      2      0 |     4.46     0.00 |   0.00   0.00 |    752.58    758.52
sana                 |      1      0 |     3.68     0.00 |   1.69   1.57 |    387.98    388.07
hunyuanv15(video)    |      5      0 |    82.59     0.00 |   0.19   0.18 |  16970.52  16881.84
wan2.2(video)        |      1      0 |     6.60     0.00 |   0.15   0.16 |   3015.05   3045.30
ltx2(video)          |      0      0 |     0.00     0.00 |   8.75   8.79 |   5240.18   5228.16

The fix removes the walk where it appears. The largest single saving is hunyuanv15 at 82.59 ms over 2 inference steps, scaling linearly to ~2.1 s at a typical 50 steps. Pipelines without the walk (chroma, ltx2) are unaffected.

Reproduction notebook (Colab)

Before submitting

• This PR fixes a typo or improves the docs.
• Did you read the contributor guideline?
• Did you read our philosophy doc?
• Was this discussed/approved via a GitHub issue or the forum? (Addresses Help us profile important pipelines and improve if needed #13401)
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

@sayakpaul @dg845

[akshan-main]

Profiled SD3 too (eager + compile, RTX PRO 6000 Blackwell, 2 steps) following the profiling guide.
Notebook: https://colab.research.google.com/gist/akshan-main/cb9ee83575806704e93e03496ba0d940/sd3_profiling.ipynb

Denoising loop is clean. 0 syncs in in_loop_body, in_transformer_forward, or in_scheduler_step after set_begin_index(0).

Pre-loop has 2x ~10ms aten::copy_ from scheduler.set_timesteps (numpy to GPU sigmas) and _get_clip_prompt_embeds (tokenizer ids to GPU). One 62ms aten::nonzero in the first _init_step_index call which set_begin_index(0) eliminates.

Tested adding set_begin_index(0) (matches Flux/Wan/Flux2). Trace sync drops from 62ms to 0 but wall-clock is within noise:

Mode	Before	After	Delta
Eager	233.0 ± 0.6 ms	231.9 ± 1.0 ms	-1.1 ms
Compile	200.2 ± 0.3 ms	199.9 ± 0.3 ms	-0.3 ms

The sync was queue-drain. GPU has to do that work anyway, CPU just doesn't wait for it. Unlike Z-Image #13461, no per-step .item()/.cpu() to chase here. Remaining pre-loop syncs are legitimate one-time copies. Not opening a PR for SD3 profile.

@sayakpaul @dg845

[DN6]

Proposal looks good to me. But we also need to account for enable_layerwise_casting.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[DN6]

@akshan-main Could you take a look at the CI failures please.

[akshan-main]

@DN6 test_to_dtype broke because StableUnCLIP's custom .to() bypasses _apply; I added register_parameter invalidation. That's a third hook though, if you'd rather not chase every mutation path, I'm happy to instead fix the normalizer's .to() or drop the cache. wdyt.