downcast MiniMax-H3 position_ids at the device transfer on fp64-less backends

[SammyTourani]

Fixes #14639

One file, 9 added / 3 removed lines. Both .to(device) sites now pass the dtype through
maybe_adjust_dtype_for_device from diffusers.utils.torch_utils:

position_ids_dtype = maybe_adjust_dtype_for_device(position_ids.dtype, device)
block_state.position_ids = position_ids.to(device, position_ids_dtype)

Why this shape rather than the torch.device(device).type == "mps" check the issue proposes:

• It is the codebase's own idiom for exactly this problem — transformer_flux.py:512,
  transformer_anyflow_far.py:48, unet_2d_condition.py:857, pipeline_pixart_alpha.py:922,
  controlnet_sparsectrl.py:608, auraflow_transformer_2d.py:438, pipeline_flux2_klein.py:408.
• The repo's own agent guide prescribes it by name: .ai/references/models.md, gotcha Add UNet for Latent Diffusion #5
  ("Only if float32 visibly degrades output, use the maybe_adjust_dtype_for_device
  helper ... Never leave an unconditional torch.float64 in the model").
• It covers NPU and Neuron too (_FP64_UNSUPPORTED_DEVICES = {"mps", "npu", "neuron"}),
  which a hand-rolled MPS check does not.

The float64 construction is untouched, and no device's numerics change — including MPS.
The issue's precision table (worst case 1.6e-4 rad) is pessimistic: the grid's fp64
precision never reaches the model at all, because MiniMaxH3RotaryPosEmbed.forward
(transformer_minimax_h3.py:95) casts position_ids to float32 as its first statement, on
every device. position_ids has no other consumer — I grepped the whole minimax_h3
package; outside before_denoise.py it only appears in docstrings. So the downcast just
moves a cast that already happens one step earlier. Measured on the largest realistic grid
(64×64 latent frame at patch 2, 102 latent frames, 4000 audio latents, max |coord| 5023):

cos bit-identical: True     max abs diff: 0.0
sin bit-identical: True     max abs diff: 0.0

And on cpu/cuda maybe_adjust_dtype_for_device(torch.float64, device) returns torch.float64,
so those paths are byte-for-byte the code that was there before.

I added no test. Reasoning, which a reviewer may disagree with (see below): the existing
suite already covers this on MPS, because the shared mixins build the pipeline with
.to(torch_device) and tests/testing_utils.py:116 resolves torch_device to "mps"
automatically on Apple silicon — and .github/workflows/push_tests_mps.yml runs pytest tests/
on a macOS runner. The before/after numbers below are that suite. A test that only ass

Verification

Environment: .venv with torch 2.14.0 (MPS available), transformers 5.17.0, torchvision
0.29.0, torchaudio 2.14.0, accelerate, peft. torchvision and torchaudio were missing at
first and caused unrelated failures; both are needed for this suite to run at all.

1. Full suite on MPS, without the patch (baseline, git stashed):

Command:

python -m pytest tests/modular_pipelines/minimax_h3/ -q -p no:randomly -n 4 --timeout=1800

Result:

11 failed, 90 passed, 10 skipped, 225 warnings, 23 errors in 299.12s (0:04:59)

286 occurrences of Cannot convert a MPS Tensor to float64 in that log. Broken tests include
TestMiniMaxH3ModularPipelineFast::test_inference_is_not_nan,
TestMiniMaxH3Ref2VAModularPipelineFast::test_inference_is_not_nan,
::test_float16_inference on both, both Loading::test_save_from_pretrained, and most of
the LoRA class.

2. Same command, same machine, with the patch:

Command:

python -m pytest tests/modular_pipelines/minimax_h3/ -q -p no:randomly -n 4 --timeout=1800

Result:

4 failed, 116 passed, 14 skipped, 249 warnings in 535.19s (0:08:55)

The 4 remaining failures are all Memory::test_*_auto_cpu_offload* and are an unrelated,
pre-existing MPS gap — components_manager.py:749 raises
NotImplementedError: enable_auto_cpu_offload() relies on the mem_get_info() method. It's not implemented for mps. because torch.mps has no mem_get_info. They fail the same way
with and without this patch and have nothing to do with position_ids.

3. Full suite on CPU with the patch (no regression on the default path):

Command:

DIFFUSERS_TEST_DEVICE=cpu python -m pytest tests/modular_pipelines/minimax_h3/ -q -p no:randomly -n 6 --timeout=1800

Result:

108 passed, 26 skipped, 228 warnings in 266.66s (0:04:26)

4. All three workflows end to end on MPS (scratch script, tiny checkpoint, 124 frames,
32×32, 2 steps) — fails at before_denoise.py:444 / :768 before, passes after:

OK t2va:  video=(1, 124, 3, 32, 32) audio=(1, 2, 828)
OK fl2va: video=(1, 124, 3, 32, 32) audio=(1, 2, 828)
OK ref2va: video=(1, 124, 3, 32, 32) audio=(1, 2, 828)

5. Style / consistency:

ruff 0.9.10 check           -> All checks passed!
ruff format --check         -> 1 file already formatted
python utils/check_copies.py  -> clean
python utils/check_dummies.py -> clean
python utils/check_ai.py      -> clean
doc-builder style ... --max_len 119 --check_only -> clean